初创公司尽调
尽调报告 AI / infrastructure Seed-stage private lab 2026-07-05

Core Automation

前沿 AI 实验室,商业验证薄弱

Core Automation 拥有顶级前沿 AI 人才和可信的研究野心;但产品、客户和收入的公开证据仍太薄,难以支撑对传闻估值目标做高置信度承保。

封面要素

初始融资 01
$100M [CO021]
初始估值 02
$1B [CO021]
后续融资目标 03
$300M-$500M at ~$4B [CO022]
成立时间 04
March 2026 [CO003]
总部 05
San Francisco, CA [CO002]
商业化状态 06
Pre-product / pre-revenue [CO029, CI009]

公司概况

Core Automation 是一家总部位于 San Francisco 的前沿 AI 研究实验室,由前 OpenAI 研究副总裁 Jerry Tworek 于 2026 年 3 月与联合创始人 Mark Saroufim 创建,早期团队来自 OpenAI、DeepMind 和 Anthropic。公司称自己在打造全球自动化程度最高的 AI 实验室,产品方向指向持续学习;据报道,其核心模型 Ceres 可在生产环境中更新,同时避免灾难性遗忘。公开证据支撑其强研究履历和异常迅速的资本形成,但还不能证明成熟商业产品、具名客户或已披露财务运营指标。

官网
www.coreauto.com
成立时间
2026-03-01
创始人
Jerry Tworek, Mark Saroufim
创立地点
San Francisco, California, USA
总部
San Francisco, California, USA
产品
Core Automation 正在打造以实验室为起点的 AI 系统,目标是自动化研究与系统工作;据报道,其中包括名为 Ceres 的持续学习模型方向,设计目标是在生产环境中更新,相比标准大模型流程减少遗忘、减少训练数据需求。
客户
目标客户包括前沿 AI 实验室、企业 AI 平台团队,以及有技术复杂度高的研究、工程和模型生命周期工作流的设计伙伴型组织。
商业模式
尚未公开披露;现有来源暗示,未来可能是与研究自动化和持续学习基础设施绑定的 B2B 模型/API 或企业软件订阅,而不是当前已有的自助式产品。
阶段
Private, pre-product, pre-revenue
融资情况
据报道,初始轮融资 $100M、估值约 $1B;另有未经确认的 2026 年后续融资报道,目标融资 $300M-$500M、目标估值约 $4B。
[CO003, CO005, CO011, CO021, CO022, CO029, CO032, CE002]

执行摘要

主要优势

  • Jerry Tworek 的 OpenAI 研究领导经历,加上 Mark Saroufim 深厚的 ML 系统背景,带来极强的创始人-市场匹配。
  • 技术逻辑清晰,聚焦持续学习、后 Transformer 架构和研究工作流自动化,而不是泛泛包一层 AI 应用。
  • Strong early capital access, with a reported $100M initial raise only months after founding. 创立数月后据报道即完成 $100M 初始融资,早期融资通道很强。
  • 早期高密度招入 OpenAI、DeepMind 和 Anthropic 人才,让一家如此年轻的公司具备少见的技术可选性。

主要风险

  • 公开层面没有可验证客户、试点、定价或产品支持界面,商业需求很大程度上仍是假设。
  • 传闻中的估值上调建立在很薄的公开证据和未确认融资细节上,一旦市场情绪降温,下行风险会被放大。
  • 与企业买家的预期相比,产品成熟度、信任、隐私和合规界面的说明仍不足。
  • 前沿 AI 的算力和薪酬强度很高,在可复制收入模型跑通前,种子资金可能被快速吃掉。
  • 既有平台和开源栈已经覆盖大量工作流、治理和基础设施界面,Core 想占据的空间并不空白。
  • 公开叙事高度围绕少数可见领导者展开,关键人和治理集中度仍然很高。

未决问题

  • 公开信息没有披露收入、毛利率、烧钱速度、现金余额或现金跑道。
  • 是否存在 Form D 或同等融资通知、何时提交以及条款如何,公开记录中仍未解决。
  • 公开层面无法验证任何具名客户、试点、设计伙伴或留存指标。
  • 除少量公开文章和发布报道外,产品架构、部署工作流、信任控制和基准测试证据都很稀疏。
  • 股权结构细节、领投方身份、董事会构成和任何结构化融资条款仍未公开。

目录

Chapter 01

01公司概览

1.1 身份、产品命题与公司阶段

Core Automation 是一家私营、尚未产生收入的人工智能研究实验室,总部位于 California 州 San Francisco。公司官网和 X 账号把使命描述为打造「全球自动化程度最高的 AI 实验室」,并明确押注:AI 能力的下一次跃迁,不会只是用更多数据训练更大模型,而会来自新的学习算法,取代大规模预训练和强化学习,再加上比 Transformer 更易扩展的架构。公司旗舰研究方向在媒体报道中被称为「Ceres」,目标是在生产环境中持续学习,用比当前前沿模型少约 100x 的训练数据完成更新;但这仍是未经验证、产品前的研究赌注,不是已经交付的能力。截至本次运行日期,Core Automation 自己的网站只露出首页、博客、损坏的团队页,以及联系 / 招聘链接——没有产品、定价或注册入口——这与第三方报道(Sacra)一致:公司没有公开 API、定价页或商业产品。公开报道把公司成立时间放在 2026 年 3 月,但至少有一个说法称融资活动在「1 月底成立后不久」就开始了;这一公开记录冲突会在下文里程碑时间线进一步讨论。公司的阶段最适合描述为私营且未产生收入;目前没有公开披露正式融资轮次标签(例如「Series A」)。[CO001, CO002, CO003, CO004, CO029, CO030]

FO002: 公司快照逻辑

Core Automation 的身份、团队、产品、资金和依赖如何连接。

[CO001, CO020, CO029, CO021, CO032]

1.2 创始人、领导层与团队

Jerry Tworek 是 Core Automation 的 CEO 兼联合创始人,在 OpenAI 工作近七年:2019 年加入时公司约有 30 名员工,后来升任研究副总裁。他主导了 o1 和 o3 推理模型开发,参与 GPT-4 后训练和 2025 年 GPT-5 部署,也做过 Codex 代码生成线和早期机器人强化学习。2026 年 1 月 5 日,他私下告诉 OpenAI 同事自己打算离开;几天后,离职消息在一波 OpenAI 高管出走中公开。他称离开是为了追求在 OpenAI 内部「很难做」的研究。Mark Saroufim 是以 PyTorch 和 GPU MODE 社区工作知名的系统工程师,公开身份为联合创始人,并撰写了公司第一篇详细技术博客。公司和第三方来源还提到多名加入创始团队的研究人员:Rohan Anil 与 Anmol Gulati(均曾在 Google DeepMind,Anil 还曾在 Anthropic)、Joanne Jang(前 OpenAI 总经理)、Ehsan Amid 与 Avery Lamp(前 DeepMind)、Julia Villagra(前 OpenAI 人才负责人),以及 Sai Surya Duvvuri(前 Google/Meta 研究实习生)。但核查时,其中几人的公开 X 账号要么零发帖,要么私密 / 不可用;公司也没有公开披露董事会或治理结构,关键人物风险和治理问题仍未解决。[CO005, CO006, CO007, CO008, CO009, CO011]

领导层和创始人表
人员角色背景创始人-市场匹配 / 职能覆盖关键人物依赖
Jerry TworekCEO 与联合创始人OpenAI 前研究副总裁(2019–2026);领导 o1/o3 推理模型、GPT-4 后训练、Codex、机器人强化学习深厚的前沿推理研究履历,与公司的后 Transformer 论点直接契合极高——公开叙事、融资和技术方向都围绕他展开
Mark Saroufim联合创始人系统 / PyTorch 工程师;运营 GPU MODE 社区;撰写公司第一篇公开技术博客覆盖「自动化系统代码」论点所需的底层系统 / GPU 工程深度中——可见技术发言人,但不是唯一公开面孔
Rohan Anil联合创始人(报道)前 Google DeepMind 和 Anthropic 研究员跨实验室优化 / 训练研究经验中——身份无法通过已检查的公开 X 账号独立验证
Anmol Gulati研究团队(报道)前 Google DeepMind 研究员,曾参与 Gemini模型架构 / 训练经验,与后 Transformer 论点契合低-中——多名报道中招聘人员之一,未确认为公司高管
Joanne Jang研究团队(报道)OpenAI 前总经理(2021 年 12 月–2026 年 4 月);参与 GPT-4o 工作模型行为和产品塑造经验低-中
Ehsan Amid研究团队(报道)前 Google DeepMind 研究员优化 / 学习算法研究背景低——身份无法通过已检查的公开 X 账号验证
Avery Lamp研究团队(报道)前 Google DeepMind 研究员公开报道未说明
Julia Villagra运营(报道)OpenAI 前人力负责人面向快速扩张研究组织的人才 / 运营职能
Sai Surya Duvvuri研究团队(报道)前 Google 和 Meta 研究实习生早期研究贡献者低——身份无法通过已检查的公开 X 账号验证

名单汇总自公司 X 帖文和第三方报道(BigGo、ai2.work、Let's Data Science);不是公司确认的完整员工名单,且多个账号无法验证。

[CO005, CO006, CO011, CO012, CO013, CO014]

1.3 融资、估值与投资者

Core Automation 已报道的资本化历史,把异常迅速的估值跃升压缩进了短短几个月。分析数据提供商 Sacra 报道,公司初始轮融资 $100M,估值约 $1B,Nvidia、Spark Capital 和 Accel 参投,但没有公开披露领投方。到 2026 年 5 月初,The Information——经 Techmeme、Intellectia 和 SiliconReport 转述——报道称 Core Automation 正在寻求 $300M 到 $500M 新资本,目标估值约 $4B,不到一个季度就翻了约四倍。AI CERTs 对同一融资的报道明确指出,Bloomberg 和 Reuters 尚未证实这些数字;Sacra 同样无法确认首轮领投方。尚无二级股票出售或债务 / 信贷融资报道。作为参照,这一节奏大体符合其他 2025-2026 年「neo-lab」超大种子轮(尽管绝对规模偏小):Thinking Machines Lab 在 2025 年中以 $10B 估值完成 $2B 种子轮;Humans& 在 2026 年 1 月宣布以 $4.48B 估值完成 $480M 种子轮;Safe Superintelligence 则在 2024 年融资 $1B。由于 Nvidia 既是据报道的投资者,又是 Core Automation 算力所依赖的主导 GPU 供应商,它的参与也引出一个需要单独尽调的关联方问题。[CO021, CO022, CO023, CO024, CO025, CO026]

利益相关方或投资人地图
利益相关方角色控制权 / 经济重要性尽调要求
Jerry Tworek创始人、CEO设定研究议程和融资条款;推定持有最大创始人股权确认股权结构表中的持股比例,以及任何特殊投票 / 控制安排
Nvidia据报道的首轮融资投资人(据 Sacra)少数财务持股;可能与算力供应形成协同确认参与情况、支票规模和任何算力供应附带协议
Spark Capital据报道的首轮融资投资人(据 Sacra)少数财务持股确认参投情况和任何董事 / 观察员权利
Accel据报道的首轮融资投资人(据 Sacra)少数财务持股确认参投情况和条款
未具名的潜在 B 轮规模投资人The Information 称正在洽谈(2026 年 5 月),金额 $300M–$500M,估值 $4B潜在较大少数股权和新控制条款识别领投方,并确认条款书 / 关闭状态
创始研究团队(Anil、Gulati、Jang、Amid、Lamp、Villagra、Duvvuri)持股关键员工约十余人团队的留任和关键人物风险确认归属时间表和留任条款
GPU/云算力供应商算力密集型研究的关键资源供应方算力可用性直接卡住研究和产品速度核实算力合同、容量承诺和定价条款

投资方和利益相关方身份来自 Sacra 的专有报道以及 The Information(经 Techmeme/SiliconReport/Intellectia 转引);截至运行日,未取得一手股权结构文件。

[CO015, CO020, CO022, CO024, CO025, CO040]

1.4 里程碑与时间线

公开时间线从 Core Automation 存在之前就开始了:Tworek 2019 年至 2026 年 1 月离开 OpenAI 的经历,构成公司融资叙事所依赖的创始人履历。报道把 Core Automation 的正式成立时间放在 2026 年 3 月,但 BigGo 2026 年 4 月 24 日的说法称,公司「1 月底成立后不久」就已开始融资——这是公开记录中的未解冲突。公司第一次公开亮相是 2026 年 4 月 21 日在 X 上发帖,宣称正在「打造全球最自动化的 AI 实验室」;多名资深研究人员在该帖发布后一天内确认加入,报道把这描述为从 Anthropic、Google DeepMind 和 OpenAI 协调式「nerdsniping」人才。融资里程碑随后迅速到来:先是 $100M 初始轮、估值 $1B,随后在 2026 年 5 月 7-8 日被报道正推动融资 $300M-$500M,目标估值 $4B。公司第一次有实质内容的公开技术披露发生在 2026 年 5 月 28 日,当时联合创始人 Mark Saroufim 发布关于 AI 研究系统代码自动化的博客——这是迄今为止超出融资和招聘新闻之外、最清晰的实际研究产出证据。到那一刻,公司仍未披露任何商业产品、定价或客户。[CO003, CO004, CO006, CO008, CO009, CO010]

里程碑表
日期事件类型金额 / 估值 / 状态参与方影响
2019Jerry Tworek 以研究员身份加入 OpenAI创立n/aJerry Tworek在 Core Automation 成立数年前,创始人已积累前沿研究履历
2025Tworek 在 OpenAI 牵头 o1/o3 推理模型项目、GPT-4 后训练、Codex 以及机器人 RL 研究产品n/aJerry Tworek、OpenAI为 Core Automation 的融资叙事提供技术履历支撑
2026-01-05Tworek 私下告诉 OpenAI 同事自己打算离职治理n/aJerry Tworek触发创办 Core Automation 的内部节点
2026-01-07在 OpenAI 高层密集离职潮中,Tworek 离开 OpenAI 的消息公开负面n/aJerry Tworek、OpenAI显示人才战背景,也把 Tworek 的信誉带到 Core Automation
2026-03据 The Information(经 Techmeme/Intellectia 转引),Core Automation 已创立 / 注册;其他报道暗示公司在 2026 年 1 月成立创立n/aJerry Tworek公司正式成形;各来源对准确日期说法不一
2026-04-21Core Automation 通过首条 X 帖文公开亮相,称正在「打造世界上自动化程度最高的 AI 实验室」产品n/aCore Automation公开亮相节点,媒体报道也由此开始
2026-04-22多名资深研究员(Rohan Anil、Anmol Gulati、Joanne Jang、Ehsan Amid、Avery Lamp、Julia Villagra、Sai Surya Duvvuri)公开确认加入,来自 Anthropic、Google DeepMind 和 OpenAI扩张n/a已披露姓名的团队成员迅速从顶级前沿实验室拼出团队
2026(4-5 月报道)Core Automation 在首轮融资中融得 $100M,估值约 $1B融资$100M @ $1BNvidia、Spark Capital、Accel(据 Sacra;其他来源未确认)确立初始资本规模;投资方身份尚未获独立交叉验证
2026-05-08The Information 报道,Core Automation 正寻求以 $4B 估值融资 $300M-$500M融资$300M-$500M @ $4B(目标)未具名潜在投资方显示公司创立一个季度内估值快速跳升
2026-05-28联合创始人 Mark Saroufim 发布公司首篇详细公开技术博客「When AI Starts Writing Systems Code」产品n/aMark Saroufim除融资和团队新闻外,首次披露有实质内容的公开技术细节
2026-05(截至)公司没有公开 API、定价页、注册流程或商业化产品负面n/aCore Automation确认公司披露画像仍处于收入前、商业化前阶段

时间线由本章综合公司 X/博客帖文和第三方报道(The Decoder、BigGo、AI CERTs、SiliconReport、Techmeme、Intellectia、Yahoo Finance)整理,口径截至运行日;更早的私有里程碑(如准确注册文件)未获独立确认。

[CO006, CO007, CO008, CO009, CO003, CO004]
FO001: 公司里程碑时间线

从 Tworek 在 OpenAI 任职到 Core Automation 第一篇公开技术博客的有日期里程碑。

日期结合了明确发布日期和不同的报道事件日期;保留 1 月与 3 月成立时间冲突,不加调和。

[CO006, CO008, CO003, CO004, CO010, CO012]

1.5 快照指标与尽调展望

把身份、团队、融资和里程碑证据放在一起看,Core Automation 更像一个履历很强、但极早期且披露很薄的研究赌注。已确认事实很窄:$100M 融资、$1B 估值、San Francisco 总部、约十几名公开团队成员,以及零披露客户或收入。其余内容——$4B 目标估值、投资者身份、准确员工数、董事会构成,以及 1 月还是 3 月成立——要么来自没有被顶级通讯社证实的报道,要么来自单一专有数据源。高报道估值、极少的公开产品证据,再叠加对 Jerry Tworek 的关键人物依赖,构成本章的核心尽调张力:后续关于市场规模、竞争和财务的章节,应把这里每个封面指标都视为暂定,并直接向公司或其投资者确认,而不是只依赖二级媒体报道。[CO020, CO021, CO022, CO023, CO028, CO029]

快照 KPI 表
指标数值 / 状态日期置信度尽调缺口
报道的估值(首轮融资)$1B2026(报道为 4–5 月)领投方和融资关闭日期未确认
报道的估值目标(后续融资)$4B(目标,洽谈中)2026-05Bloomberg/Reuters 尚未交叉确认
已确认融资额$100M(首轮融资)2026一级与二级交易占比未披露
目标新增资本$300M-$500M2026-05条款书 / 关闭状态未知
员工数~12 名公开成员(估计)2026-04没有官方员工数披露
客户 / 收入0 / 收入前2026-05Sacra 确认商业化前状态
总部San Francisco, CA2026-01n/a
成立日期2026 年 3 月(存在争议;其他报道暗示为 2026 年 1 月)2026未审阅公司注册记录

汇总截至运行日的公司官方信号(X 简介)、分析师市场数据(Sacra)和新闻线报道(Techmeme/Intellectia/SiliconReport/AI CERTs);置信度和尽调缺口两列反映交叉验证强度,而不是底层事实的确定性。

[CO002, CO003, CO004, CO021, CO022, CO023]
FO003: 快照 KPI

截至运行日期的核心成熟度、牵引和披露风险指标。

[CO021, CO022, CO028, CO029, CO034, CO023]

1.6 图表与证据

Chapter 02

02市场分析

2.1 市场边界:Core Automation 实际卖向哪里

Core Automation 还没有产品,因此本章定义的是它瞄准的市场,而不是它当前服务的市场。四类需求锚定这个目标:前沿 AI 实验室研究自动化工具(最接近公司「自动化 AI 实验室」的自我表述)、企业知识工作和报告自动化、持续学习与后 Transformer 模型研究,以及 AI for science 发现自动化。本章审阅的分析机构没有一家把这四类作为单一条目跟踪;即便有测算,也被放进更宽的「智能体 AI」「AI 软件」或「机器人流程自动化」桶里(Table TM001)。还有两类与边界保持距离,而不是落在边界内:传统规则型 RPA,它自动化的是重复任务,不是开放式认知工作;以及工业 / 物理 AI 自动化,World Economic Forum 和 BCG 将其描述为建立在硬件具身机器人上的平行「自主工作」命题,而不是纯软件研究工具(CM037、CM038、CM039)。因此,Core Automation 所述赌注的现状替代品不是单一竞争者类别,而是人类研究工程师、现有巨头的智能体式代码 / 研究助手,以及传统 RPA 的组合——这些今天已经拿走企业自动化预算。由于 Core Automation 没有披露产品、定价或客户,本节把四个边界内类别当作后续章节的方向性框架,而不是公司已经验证的可服务市场。[CM037, CM038, CM039]

市场定义表
细分市场 / 类别计入支出排除支出买方 / 付款方与 Core Automation 的关系
前沿 AI 实验室研究自动化工具前沿实验室自建或采购的内部算力、研究工程工具、实验编排软件通用模型训练算力本身;消费级 AI 产品前沿实验室 CTO / 研究负责人(由实验室资本自行出资)直接目标:最接近 Core Automation 所称建设方向的类比
企业级智能体 AI / 知识工作自动化软件面向研究、报告和分析工作负载的独立智能体平台,以及带智能体能力的 SaaS 功能多数供应商会单列核心 LLM API 支出;基于规则的 RPA企业 CIO / COO / 职能 VP相邻市场:如果 Core Automation 最终销售企业工具,这是最近且有融资、可衡量的市场
持续学习与后 Transformer 架构研究学术和产业实验室 R&D 预算、算力资助、公开基准产品收入(这是商业化前研究,不是已变现市场)研究出资方:实验室、学术资助、企业研究预算Core Automation 所称技术论点;本身不是已测算规模的商业市场
科学发现 / AI 科研自动化用于假设生成、实验设计、论文撰写的供应商工具和实验室内部智能体湿实验室设备和物理科学仪器高校 PI 预算、生物技术 / 制药 R&D、企业科学实验室相邻的概念验证类别,显示买方愿为自动化研究付费
传统机器人流程自动化(RPA)面向重复、基于规则的企业流程的软件许可证和服务高度依赖判断或开放式认知任务企业 IT / 运营预算传统相邻市场:显示企业愿在另一套技术底座上为自动化付费
工业 / 物理 AI 自动化(机器人)工厂车间和物流场景中的感知—推理—行动机器人系统纯软件 / 云端 AI 服务工厂运营 / COO 资本开支与运营开支预算关联较松;认知 / 办公场景之外的平行「自主工作」论点
AI 基础设施与算力(云、GPU、数据中心)超大规模云厂商和 GPU 供应商收入应用层软件支出采购算力的前沿实验室和企业上述所有细分市场的上游成本驱动和容量约束,本身不是 Core Automation 的产品

边界由本章作者基于 claimRefs 中引用来源划定,并非任何单一出版方给出的分类;查阅到的分析机构也没有把这七类作为统一分类法发布,因此计入 / 排除两列应视为方向性框架,而非经审计的市场切分。

[CM037, CM038, CM039]

2.2 相邻市场测算:多重视角,没有单一 TAM

没有发布方专门给前沿实验室研究自动化或持续学习工具拆出美元规模,因此本节叠加现有相邻视角,而不是断言一个数字。最宽的一层,Gartner 预计 2026 年全球 AI 总支出为 $2.59T,同比增长 47%(CM002);它标为「智能体 AI」能力支出的较窄切片在 2026 年达到 $201.9B(CM003),约为总额的 8%。再窄一层,四家分析机构(Fortune Business Insights、Precedence Research、MarketsandMarkets,以及 Deloitte 的 TMT Predictions)把独立 AI 智能体软件市场在 2025-2026 年测算为 $7.0B-$8.5B,但到各自预测终年时,多年预测相差近 30x;Gartner 自己宽口径与窄口径的智能体 AI 数字在同一时间点也相差约 25x(CM004、CM048)。同样的测量问题在相邻一层也出现:Grand View Research 估算全球 RPA / 知识自动化市场 2025 年为 $4.68B,2033 年增长至 $35.84B;Precedence Research 则把同一名义类别估算为 2025 年 $28.31B、2035 年 $247.34B——仅基准年就相差约六倍(CM006、CM007)。IDC 的替代框架完全绕开美元规模,预测到 2030 年 45% 的组织会「规模化」编排 AI 智能体,这是采用指标,不是市场规模(CM008)。Gartner 自己的预测在八个月内移动了约 $500B(CM047),进一步说明这些标题数字仍有很大漂移。Figure FM001 从最宽到最窄堆叠这些视角;Figure FM002 展示四个独立市场预测即便名义上描述「同一」类别,也会相差多远。[CM002, CM003, CM004, CM006, CM007, CM008]

TAM/SAM/SOM 或规模测算视角表
发布方年份地区数值CAGR方法可信度局限
Gartner2026全球$2.59T AI 总支出(基础设施 + 软件 + 服务)同比 47%覆盖硬件、软件、服务的自上而下供应商 / 企业 AI 支出模型整体 AI 市场数字;未拆出研究自动化或持续学习支出
Gartner2026全球$201.9B「智能体 AI」能力嵌入式支出n/a(总量子集)在企业软件类别中按能力打标签低-中只要嵌入智能体功能就计入;比独立智能体供应商估算高约 25 倍,显示定义口径很敏感
Fortune Business Insights / Precedence Research / MarketsandMarkets / Deloitte TMT(由 Axis Intelligence / SoftwareStrategiesBlog 汇总)2025-2034全球独立 AI 智能体软件市场 $7.0-8.5B(2025-2026),到 2032-2034 年随机构不同升至 $93.2B-$199.05B随机构不同,CAGR 为 40.5%-44.6%汇总 4 家分析机构供应商收入的自下而上口径各机构在各自终期年份的估算相差近 30 倍;「智能体软件」类别边界并不统一
Grand View Research2025-2033全球RPA / 知识自动化市场从 $4.68B(2025)增至 $35.84B(2033)29.0% CAGR(2026-2033)供应商收入市场测算,软件 + 服务低-中比 Precedence Research 对相近命名类别的估算小约 6 倍,显示类别边界很敏感
Precedence Research2025-2035全球RPA 市场 $28.31B(2025)、$35.27B(2026)、$247.34B(2035)24.2% CAGR(2026-2035)供应商收入市场测算,软件 + 服务低-中与 Grand View Research 对同一名义类别的估算明显分叉;底层细分定义披露不够细,难以调和
IDC(FutureScape 2026)2030(预测)全球45% 的组织「规模化」编排 AI 智能体(采用率%,非美元值)n/a分析师调查 / 技术采用预测仅为采用指标;若不额外假设每个组织支出,无法换算成美元 TAM

本表没有任何发布方测算 Core Automation 的具体细分市场(前沿实验室研究自动化或持续学习工具);每行都是更宽泛的相邻类别估算,用来给问题设边界,并非公司专属 TAM。各行差异刻意保留,没有平均成一个虚假的单一数字。

[CM002, CM003, CM004, CM006, CM007, CM008]
FM001: 市场规模测算视角

三层相邻市场视角,从最广义 AI 支出下探到独立智能体软件市场,并不暗示存在干净的 Core Automation 专属 TAM。

最窄层为可读性使用 10.9-11.8B 区间中点。任何已审阅来源都没有单独测算 Core Automation 具体的前沿实验室研究自动化细分市场,因此本报告有意省略,而不是自行估算;见 TM002 及其 evidenceGap。

[CM002, CM003, CM005]
FM002: 市场估计区间

四家分析机构对同一名义独立 AI 智能体软件类别从 2025/2026 年至终端年的预测,保留为区间,而不是平均成一个数字。

四行均将单位固定为独立 AI 智能体软件类别的十亿美元;预测期不同(2030/2032/2034),并在每个标签中注明,未做标准化,因为标准化本身会引入无支撑的精度。

[CM004, CM022]

2.3 买方、用户与从试点走向生产的路径

在四个边界内细分市场里,买方、用户和付款方的拆分方式不同。前沿实验室内部,研究自动化买方是实验室自己的 CTO 或研究负责人,资金来自实验室算力 / 运营费用预算,研究工程师是直接用户(Table TM003)。企业内部则延续更熟悉的模式:CIO、COO 或职能副总裁用现有 IT 或运营预算资助试点,而不是专设「AI 智能体」预算线;分析师和普通员工是日常用户(CM040)。买方意向很广:MuleSoft/Deloitte Digital Connectivity Benchmark 调研中,93% 的 IT 领导者计划在两年内引入自主智能体(CM041);Deloitte 也单独把研发列为企业选择智能体 AI 的顶部用例之一,与客户支持、供应链和网络安全并列(CM046)。但意向不等于 Core Automation 具体赌注的已验证买方:最接近 AI 驱动研究自动化的证明点——Anthropic 的 Claude Science 和 Sakana AI 的 AI Scientist——来自已经占优、资本充足的既有玩家,且建立在标准 Transformer 模型上,不是等待新进入者填补的空白市场(CM042)。Figure FM003 描绘从研究命题到企业级工作流嵌入的通用采用路径;Figure FM004 则从广泛的智能体实验收窄到近乎为零的、已部署且专门验证持续学习研究自动化产品的基数,这一基数才会验证 Core Automation 的特定命题。[CM040, CM041, CM042, CM046]

细分市场 / 买方地图
细分市场买方用户付款方工作流预算负责人采用触发点
前沿 AI 实验室(研究自动化)实验室 CTO / 研究负责人研究工程师实验室自身(算力 / 运营开支预算)实验设计、执行和论文撰写实验室 CTO / CFO算力成本压力和研究员稀缺
企业 R&D / 创新团队CIO / 首席创新官科学家 / 分析师企业 IT / R&D 预算文献综述、假设检验、报告CIO加快创新周期的竞争压力
企业知识工作 / 报告自动化COO / 职能 VP分析师、助理部门运营开支预算研究、分析、报告生成部门 VP劳动力成本 / 人头约束,以及既有 RPA 先例
科学 / 学术发现自动化高校 PI、生物技术 / 制药 R&D 负责人科学家课题资助 / 生物技术 R&D 预算假设生成、实验设计、同行评审辅助PI / 资助办公室发表产出压力和湿实验室实验成本
工业 / 物理自动化相邻市场工厂运营高管操作员 / 技术员制造业资本开支 / 运营开支预算车间里的感知—推理—行动闭环COO / 工厂经理劳动力短缺和供应链波动

各行描述引用来源中记录的通用买方 / 用户 / 付款方模式,并非已确认的 Core Automation 客户;截至运行日,公司未披露产品或客户。

[CM040, CM041, CM042, CM046]
FM003: 买方 / 细分地图

研究和知识工作自动化工具的通用采用路径,从实验室或企业的初始论点一路到企业级工作流嵌入。

[CM024, CM040, CM043]
FM004: 采用漏斗或价值链地图

从广泛企业智能体实验,收窄到接近零的部署基数;后者类似 Core Automation 特定持续学习研究自动化论点。

最后一行(value=0)表示未找到证据,并不等于声称这类部署不可能存在;见相关证据缺口。

[CM024, CM036, CM023, CM045]

2.4 增长驱动因素与采用约束

三股力量把认知密集型自动化需求往上推。第一,能力正在迅速追近人类表现:在 OSWorld 基准上,智能体系统完成真实世界电脑任务的比例约 18 个月内从约 12% 升至约 66%,但仍约每三次尝试失败一次(CM021)。第二,资本正集中流入这一领域:仅 2026 年第一季度,AI 初创公司就拿走了全球约 $300B 风险投资中的 80-81%(CM028、CM029);Gartner 预计到 2026 年底,40% 的企业应用会嵌入任务专用 AI 智能体(CM025)。第三,Thinking Machines Lab 等资本充足的前沿实验室正在锁定数十亿美元算力交易和吉瓦级 Nvidia 硬件承诺(CM030),说明严肃资本相信 Core Automation 自动化研究命题的某个版本具备可投资性。三股力量反向拉扯。信任与治理落后于能力:只有 21% 的组织报告已有成熟的自主智能体治理模型;Gartner 另预计,到 2027 年底,40% 的智能体 AI 项目会因成本、ROI 和风险控制顾虑被取消(CM043)。即便组织开始实验,真正规模化者也很少:2025 年 11 月的一项调查发现,只有 23% 在单一职能内规模化智能体,跨多个职能规模化的比例低于 10%(CM036)。算力本身也是瓶颈:受 TSMC 封装和 HBM 内存约束推动,H100/H200 交付周期达 36-52 周,这把资本充足的既有玩家推向多年预留合同,而 Core Automation 这样规模更小、更新的进入者很难匹配(CM032、CM033)。[CM021, CM025, CM028, CM029, CM030, CM032]

增长驱动与约束表
驱动因素 / 约束方向时间影响尽调问题
智能体能力跃升(OSWorld 任务成功率约从 12% 到 66%)驱动已经可见(2024-2026)复杂认知工作流自动化的可信度提高对比基准表现与真实部署任务成功率,而不是只看实验室条件
企业智能体 AI 支出增长(2026 年 $10.9-11.8B,至约 2030-2034 年超过 $50B)驱动中期(2026-2030)显示智能体工具总体可触达预算在扩张其中多少会流向研究自动化专门厂商,而非既有平台
前沿实验室资本集中与算力稀缺约束眼前GPU/HBM 短缺推高训练成本,并拖慢包括新进入者在内的非巨头实验室Core Automation 的算力合同、分配额度和供应商关系
治理 / 信任缺口(21% 具备成熟智能体治理)约束近期即便能力存在,也会放慢自主部署Core Automation 是否已有或计划提供合规 / 评估框架
试点到生产失败率(95% GenAI 试点没有 ROI)约束当前买方保持怀疑;预算扩大前必须证明 ROICore Automation 有产品后,核查任何试点数据、客户推荐或案例
AI 泡沫 / 资本开支怀疑(Burry 等)约束宏观 / 时间不确定可能压缩产品前研究实验室的融资可得性跟踪资本市场情绪和 Core Automation 现金续航期
持续学习技术突破(Nested Learning、Mamba-3)驱动研究阶段,1-3 年窗口一旦商业化,可能显著降低训练数据和算力需求,与公司所称论点一致将 Core Automation 任何模型与这些已发表技术独立对标
科学研究自动化验证点(Sakana AI Scientist、Anthropic Claude Science)驱动已经上线证明买方愿意尝试 AI 研究自动化,但验证来自资本雄厚的在位者Core Automation 相比 Anthropic、Google 和 Sakana AI 的差异化

方向和时间是本章对引用来源的综合判断,并非某一家发布方框架;多行会随 Core Automation 出货速度不同而双向影响。

[CM021, CM023, CM025, CM026, CM032, CM034]

2.5 持续学习与后 Transformer 研究为何可能具备经济意义

Core Automation 的技术赌注是:持续学习——让模型不必从零重训也能保持最新——以及比 Transformer 更易扩展的架构,最终可能降低前沿 AI 的训练数据和算力成本,直接抵消本章其他部分记录的 GPU 稀缺(CM044)。这个研究方向真实且活跃,并不边缘:Google Research 的「Nested Learning」范式及其「Hope」架构,把模型重构为嵌套、自我修改的优化问题,目标是保留长周期记忆(CM010、CM011);Mamba-3 是 2026 年获 ICLR oral 接收的状态空间模型架构,声称以约一半推理成本达到或超过 Transformer 基线的困惑度表现(CM012、CM013)。截至 2026 年中,这些都还不构成经过验证、可在前沿规模商业替代 Transformer 模型的方案(CM014);本次审阅也没有任何来源发布这一研究类别的美元市场规模(CM009、CM045)。证明自动化研究本身可行,比证明持续学习本身可行更容易:Sakana AI 的 AI Scientist 产出了一篇发表于 Nature 的论文,描述一个能够提出假设、运行实验并撰写稿件的智能体(CM015);更早版本通过了真实同行评审(CM016),代码也已开源(CM017)。Anthropic 的 Claude Science 为科学家推进类似的工作流自动化命题,但没有交付新模型(CM018);Anthropic 还单独报告称,其生产代码中超过 80% 现在由 Claude 编写,相比 2021-2025 年基线提高 8x(CM019)——这一点应谨慎看待,因为它是公司自报、未经审计的数据(CM020)。合起来看,证据更强地支持「自动化研究是可行的」,远不如支持「Core Automation 的特定持续学习方法已经被证明」。[CM009, CM010, CM011, CM012, CM013, CM014]

2.6 不利证据与尽调缺口

企业采用的标题数字掩盖了从采用到影响之间的巨大缺口,任何给 Core Automation 做市场规模测算都应对此打折。Stanford 2026 AI Index 报告称,组织 AI 采用率为 88%,生成式 AI 使用率三年内达到 53%——扩散速度快于 PC 或互联网(CM022);但 Deloitte 发现,只有 23% 的组织哪怕中等程度使用智能体 AI,且只有 21% 拥有成熟治理模型(CM023)。Axis Intelligence 交叉引用的「Deployment Gap Index」用另一种方式量化同一模式:93% 的 IT 领导者打算在两年内部署自主智能体,但只有 23% 哪怕在单一职能中规模化部署,声明意向与生产现实之间相差 70 个百分点(CM024)。最常被引用的警示数据点是 MIT 2025 年「GenAI Divide」研究,覆盖 300 多个企业部署,发现 95% 的组织从生成式 AI 试点中没有获得可衡量的 P&L 回报;失败原因归于集成浅、各自为政,而不是模型质量或监管(CM026、CM027)——这一警示至少同样适用于产品前的研究赌注,也适用于已经交付的生成式 AI 工具。宏观层面,投资者 Michael Burry 公开把 2025-2026 年 AI 投资周期类比为 1999-2000 年互联网泡沫,认为 hyperscaler 的折旧会计低估了 AI 基础设施支出的真实成本(CM034、CM035)。这些都不能证明 Core Automation 的具体命题会失败,但意味着它押注的需求环境,是 AI 行业迄今最资本密集、也最怀疑评估结果的环境。[CM022, CM023, CM024, CM026, CM027, CM034]

2.7 图表与证据

Chapter 03

03竞争对手

3.1 竞争格局框架:分层,而不是一张同行名单

由于 Core Automation 没有披露产品、定价或客户入口,不能把它简化为单一竞争者短名单。Radical Ventures 2026 年「NeoLab」分类明确把 Core Automation 放进「持续学习」桶,与 Adaption Labs 并列,区别于「世界模型」实验室(AMI Labs、World Labs、Decart)、强化学习实验室(Reflection AI、Ineffable Intelligence),以及扩散或能量模型实验室——这证实即便是成熟的 AI 投资者,也把持续学习视为几个相互竞争的后 Transformer 赌注中的一条窄赛道,而不是一个天然可防守的独特类别。neolab 上层,是 OpenAI、Anthropic、Google DeepMind 等前沿既有玩家,它们已经在生产中运行内部自动化研究智能体,并交付能产生收入的产品;下层,则是 FutureHouse、Glean、Hebbia、Manus 等产品优先公司,它们已经把自动化研究或知识工作工具卖给付费客户,不需要等待持续学习突破。Radical Ventures 还把算力获取,而非资本,框定为整个格局的硬约束:战略算力伙伴关系(hyperscaler 承诺、Nvidia 配额协议)已成为股权结构中的标准配置,这意味着 Core Automation 仅有的已报道 Nvidia 投资人关系,比下文每一层多数同行的算力位置都薄得多。[CP028, CP029, CP030, CP046, CP010]

竞争对手画像表
竞争对手类别规模 / 融资目标细分市场差异化局限
OpenAI前沿在位者~$852B 估值(2026 年 3 月轮次);该轮融资约 $122B消费者 + 企业 + 开发者 API分发最广(ChatGPT、API、Codex);GPT-5.5 瞄准智能体编程和「早期科学研究」2025 年泄露财务数据显示,收入约 $13B、经营亏损约 $21B;严重依赖 Oracle / Stargate 算力
Anthropic前沿在位者$965B 投后估值(2026 年 5 月 Series H);年化收入约 $47B企业(Claude Code/Cowork)+ 开发者 API多云算力(AWS/Google/SpaceX);在前沿实验室中企业收入增长最快与 OpenAI 一样,面临看空分析师指出的补贴使用和泡沫风险敞口
Google DeepMind前沿在位者由 Alphabet 资产负债表供资;未披露独立估值消费者(Gemini)+ 企业 + 研究Gemini Deep Think 已经自动化研究智能体工作流(「Aletheia」),并产出同行评审成果研究自动化只是众多产品线之一,不是 Core Automation 那样押注到公司层面的专门赌注
Thinking Machines Lab新实验室 / 论点相邻$2B 种子轮,估值 $12B(2025)研究员 / 开发者(开放权重微调)已发布 Tinker(2025 年 10 月),进入收入轨道;创始人 Mira Murati 曾任 OpenAI CTO重点是微调基础设施,不是持续学习或自动化内部研究
Safe Superintelligence新实验室 / 论点相邻已融资 ~$1B+;报道估值约 $32B未披露(无产品)创始人 Ilya Sutskever 的 OpenAI 联合创始人 / 首席科学家履历,让投资人给出超常信任无公开路线图或产品;单一目标使命不如 Core Automation 所称论点具体
AMI Labs新实验室 / 论点相邻(世界模型)$1.03B 种子轮,投前估值 $3.5B(2026 年 3 月)医疗(Nabla)优先,之后机器人 / 科学Yann LeCun 基于 JEPA 的「世界模型」论点,是资金充足且不同于 LLM 扩展和持续学习的另一条路线未披露收入计划;商业化应用前仍有多年研究窗口
Adaption Labs新实验室 / 论点相邻(持续学习)$50M 种子轮(2026 年初)需要即时模型适配的企业Sara Hooker 的「无梯度」持续学习论点几乎与 Core Automation 的核心押注相同,但披露估值低得多资本基础远小于 Core Automation 报道中的 $1B-$4B 区间;规模化未验证
Reflection AI新实验室 / 论点相邻(开放前沿模型)$8B 估值(2025 年末);洽谈 $25B 投前估值融资(2026 年 3 月)企业 + 政府(主权 AI)Nvidia 支持;明确采用开放权重策略,瞄准「西方替代方案」叙事截至运行日尚未发布模型;变现计划未验证
Sakana AI相邻区域 / 效率型玩家$2.65B 估值;累计融资约 $379M(2025 年 11 月)日本企业(金融、工业、政府)为语言 / 文化适配的高效小数据模型,而非前沿规模路线地域和技术范围都比 Core Automation 的全球前沿雄心更窄
FutureHouse相邻 AI 科研同行非营利;慈善资金支持学术界生物 / 科学研究人员已推出多个研究智能体系统(Robin、DISCO、OXtal),并发布可复现成果非营利结构,且聚焦范围更窄的生物 / 科学研究;相比之下,Core Automation 的自动化研究雄心更宽
Glean产品层 / 企业知识工作自动化$7.2B 估值;ARR 约 $300M(2026 年估算)企业 IT / 知识工作者100+ 个 SaaS 集成、权限感知搜索、已付费的 Fortune 500 客户群横向聚合器,不是前沿模型开发商;竞争点在工作流层,不在底层研究
Hebbia产品层 / 企业知识工作自动化$700M 估值;披露 ARR 约 $13M(2024 年)金融服务、法律(资产管理公司、银行)垂直尽调 / 文档分析产品,已被约 30% 资产管理公司使用垂直范围窄;不是通用自动化研究平台
Manus产品层 / 智能体工作流自动化Meta 约 $2B 收购存在争议;据报 ARR 约 $125M(2026 年 1 月,第三方估算)消费者 + 企业(智能体任务自动化)基于第三方模型(Claude、Qwen)的编排层;消费者 / 企业产品已上线所有权状态存在争议(中国国家发改委要求 Meta 撤销交易,2026 年 6 月);自身不开发前沿模型

规模 / 融资数字混合了已披露轮次、第三方估值估算以及不同时间点(2024-2026 年)的分析师 ARR 估算;应视为方向性证据,而非审计数据。覆盖范围是截至本轮报告日期,与 Core Automation 所述论点和产品层最相关的 AI 研究 / 智能体实验室与知识工作自动化厂商代表性样本,不是所有 AI 公司的穷尽清单。

[CP001, CP002, CP004, CP005, CP006, CP008]
FP001: 竞争定位图

按产品 / 分发成熟度,以及与 Core Automation 持续学习论点的贴近度,对 Core Automation 和 13 家竞争者作序位定位。

坐标轴是有证据支撑的序位分数(1-5),不是来源逐字报告的数字。X =「产品和分发成熟度」(1 = 尚无产品,5 = 已有广泛商业分发)。Y =「与 Core Automation 持续学习押注的贴近度」(1 = 技术路径无关,5 = 公开论点几乎相同)。定位是本章为便于比较,对所引主张的综合整理;不是精确测量,也不是独立基准测试。

[CP001, CP005, CP008, CP011, CP014, CP017]

3.2 前沿既有实验室:OpenAI、Anthropic 与 Google DeepMind

三大前沿实验室已经达到 Core Automation 尚未接近的规模和产品成熟度。OpenAI 在 2026 年 4 月发布 GPT-5.5,明确面向智能体编码、电脑使用和「早期科学研究」营销;其官方研究页仍以产品前实验室无法匹配的节奏发布应用成果,例如 2026 年 6 月的基因组学基准 GeneBench-Pro。Anthropic 在 2026 年 5 月完成 $65B Series H,投后估值 $965B,年化收入超过 $47B,已承诺多云算力覆盖 Amazon、Google/Broadcom TPU 和 SpaceX 的 Colossus 数据中心,使 Claude 成为首个同时在三大主要云上运行的前沿模型。Google DeepMind 的 Gemini Deep Think 模式已经运行一个内部自动化数学和科学研究智能体(代号「Aletheia」),可自主生成、验证并修订研究级证明,成果已提交同行评审场所——功能上就是 Core Automation 所称「自动化研究过程」的同一野心,只是它已经在生产中运行并有公开成果。三者都不像 Core Automation 那样依赖单一已报道投资者来获得算力。[CP001, CP002, CP003, CP004, CP044, CP045]

3.3 直接命题竞争者:持续学习与后 Transformer Neolab 浪潮

一批资金充足、由研究人员带队的初创公司,正在追逐与 Core Automation 自身高度重叠的技术赌注。Adaption Labs 由前 Cohere 高管 Sara Hooker 和 Sudip Roy 创立,2026 年初完成 $50M 种子轮,商业化「无梯度」持续学习,使已部署模型无需完整重训即可适配——这对「冻结模型」经济性的批评,几乎与 Core Automation 创始人对静态预训练的批评相同。Yann LeCun 的 AMI Labs 在 2026 年 3 月以 $3.5B 投前估值融资 $1.03B,追求基于 Joint Embedding Predictive Architecture 的「世界模型」,这是不同但相邻的后 LLM 范式;其 CEO 表示,达到商业应用可能需要数年。Reflection AI 由前 DeepMind 研究人员创立,2025 年末估值达到 $8B,据报道到 2026 年 3 月正洽谈一轮 $25B 融资,押注开放权重前沿模型而不是持续学习。前 OpenAI CTO Mira Murati 创办的 Thinking Machines Lab 已在 2025 年 10 月发布商业产品 Tinker,比 Core Automation 披露任何公开内容早了数月。Sakana AI 估值 $2.65B,追求高效、日本本地化模型——这是地理和技术上更窄的赌注。按 Radical Ventures 统计,2026 年前的三年里,40 多家此类 neolabs 合计融资超过 $40B。[CP005, CP006, CP007, CP011, CP012, CP013]

3.4 产品层竞争者:研究智能体与企业知识工作自动化

多家组织已经在销售或运营 Core Automation 希望打造的自动化研究和知识工作工具,而且不需要等待持续学习突破。FutureHouse 是非营利 AI for science 实验室,2026 年 5 月发布用于端到端生物发现的多智能体系统「Robin」,此前还发布过 DISCO、OXtal——其慈善资金支持结构与 Core Automation 的风投支持、营利模式根本不同,尽管自动化目标相似。Glean 估值 $7.2B,2026 年 ARR 估计约 $300M,已经通过 100 多个 SaaS 集成和在线智能体平台服务 Fortune 500 企业客户。Hebbia 披露 ARR 为 $13M、估值 $700M,约 30% 的资产管理机构用它做财务尽调;其官网还称客户 AUM 约 $30T、日均提示词 200,000 次。Manus 是一个智能体工作流平台,Meta 于 2025 年 12 月宣布以约 $2B 收购;Manus 自己的网站仍写着「Manus is now part of Meta」,但中国 NDRC 在 2026 年 4 月要求交易撤销,Meta 到 2026 年 6 月已开始运营层面拆分——数据防火墙、访问撤销、禁止内部使用。Manus 当前所有权仍有未解冲突,本章将其视为开放矛盾,而非既定事实。[CP014, CP015, CP016, CP021, CP022, CP023]

3.5 比较分析:能力、定价、分发与信任姿态

在能力、定价和分发上,既有玩家和产品层竞争者拥有结构性优势,Core Automation 需要多年才能追上。下方能力矩阵显示,每个前沿既有玩家和多数产品层竞争者都已有已发布产品和在线企业分发,而 Core Automation 与多数 neolabs 还没有。定价上,OpenAI、Anthropic 和 Google DeepMind 已经把「智能体」和「研究」工作负载的用量计费和企业合同定价常态化;Glean、Hebbia 和 Manus 则直接变现工作流层——这意味着 Glean 自身架构(明确可与「来自 OpenAI、Google、Amazon、Meta 和 Anthropic 的领先 LLM」互操作)能够吸收 Core Automation 未来交付的任何模型,同时不放弃客户关系。这种多模型并用格局限制了一个新模型提供商单独能捕获多少分发价值。在信任与监管姿态上,Manus 的 Meta 收购争议显示,一家中国背景 AI 公司的跨境所有权,即便迁移总部后也可能被监管逆转;Reflection AI 明确的「主权 AI」和政府定位也显示,竞争者已在争取同样对国家安全敏感的客户,而 Nvidia 作为关联方进入 Core Automation 自身股权结构,可能让这一客户群更复杂。[CP046, CP047, CP048, CP049, CP021, CP024]

功能 / 能力矩阵
采购标准Core AutomationOpenAIAnthropicGoogle DeepMind新兴实验室(Thinking Machines/SSI/AMI/Adaption/Reflection/Sakana)企业智能体(Glean/Hebbia/Manus)
已上线商业产品否(产品前阶段)是(ChatGPT、API、Codex)是(Claude、Claude Code)是(Gemini、Deep Think)部分(Tinker 已上线;SSI/AMI/Adaption/Reflection 尚未上线)是(Glean、Hebbia、Manus 均已上线)
持续学习 / 部署后学习主张是(核心论点)不明确 / 非主要重点不明确 / 非主要重点不明确 / 非主要重点是(仅 Adaption Labs)否(通过第三方 API 使用静态模型)
已披露前沿规模算力承诺否(Nvidia 仅为投资方)是(Oracle Stargate 约 $300B)是(AWS/Google/SpaceX 多吉瓦)是(Alphabet TPU 集群)部分(AMI/Reflection 获 Nvidia 支持;其他未披露)否(算力是模型供应商的问题,不是它们的问题)
独立 / 同行评审研究输出否(未验证)部分(GPT-5.5 系统卡)部分是(Gemini Deep Think 论文、会议投稿)部分(因实验室而异)否(产品公司,不是研究实验室)
企业分销 / 渠道已上线否(多数新兴实验室)是(100+ 集成、具名 Fortune 500 客户)
已披露融资 / 估值是(约 $1B,目标估值约 $4B)是(约 $852B)是(约 $965B)N/A(Alphabet 业务单元)是,区间不一($50M-$25B)是($700M-$7.2B)
创始人前沿实验室出身是(OpenAI 前研究副总裁)N/AN/AN/A是(OpenAI 前 CTO、OpenAI 前首席科学家、Meta 前首席科学家、Cohere 前员工、DeepMind 前员工)混合(Hebbia/Glean 创始人并非前沿实验室出身;Manus 创始人来自企业市场)
具名付费企业 / 政府客户否(多数)是(Booking.com、资产管理公司、Fortune 500)

标为「不明确」或「N/A」的单元格反映截至本轮报告日期缺少公开披露,并非确认负面;若干单元格把一组公司(新兴实验室和企业智能体列)合并成一个定性判断,而不是逐家公司逐事实验证。

[CP044, CP001, CP008, CP037, CP021, CP024]
定价 / 打包方式对比
公司定价模式标价 / 单位包含能力折扣 / 未知项对 Core Automation 的含义
OpenAI用量计费 + 订阅(ChatGPT Plus/Pro/Business/Enterprise + 计量 API)消费者分层 + 按 API token 计量;所审来源未披露 2026 年确切费率GPT-5.5 智能体编程、计算机使用、研究任务企业 / API 批量折扣未披露把「开箱即用」的智能体价格锚点立住,Core Automation 要么定价更低,要么给出清晰差异化
Anthropic订阅 + 计量 API + 企业合同Claude Code/Cowork 企业交易;所审来源未披露 API token 定价Claude Opus 4.8 编程 / 智能体任务企业批量条款未披露其 $47B 年化运行率表明企业已经大规模为自动化知识工作付费,正是 Core Automation 要争取的同一批买方
Google DeepMind (Gemini)通过 Google Cloud 按用量计费 + 消费者订阅计量 API + 消费者分层;报道称每 token 定价具竞争力Gemini 3.5 Flash、Deep Think 模式企业折扣具体条款未披露Google 可以把研究自动化功能打包进现有 Cloud/Workspace 合同,边际增量成本很低
Thinking Machines Lab (Tinker)按用量计费(免费 beta 后)私测 beta 免费;发布帖称「未来几周」推出按用量计价托管微调 API、基于 LoRA 的算力共享发布时未披露具体用量费率说明新兴实验室即使尚未解出更难的研究命题,也能先靠基础设施 / 工具变现
Glean企业席位 + 按智能体动作授权企业合同;标价未披露搜索、Glean Agents、Glean Protect 安全层折扣 / 企业条款未披露买方已经为知识工作自动化任务向 Glean 付费,Core Automation 必须从这里赢单
Hebbia企业合同 / 按 AUM 分层企业合同;标价未披露文档分析(Matrix)、尽调工作流定价随客户 AUM / 席位扩展;具体数字未披露说明金融 / 法律等垂直领域愿意为研究自动化支付远高于通用 SaaS 的价格
Manus消费者 / 企业订阅 + API + 团队套餐Web 应用、移动端、API、团队套餐分层;所审来源未披露费率幻灯片、网站、浏览器操作器、「Wide Research」具体分层定价未披露证明同一类「帮我自动做事」的消费者 / 企业智能体产品已经上线并定价
Sakana AI企业合作 / 定制交易未公开列价;披露了与 MUFG、Daiwa 的交易面向日本企业工作流适配的小型高效模型定价未披露;逐单谈判展示了区域 / 垂直定制合作的定价打法;Core Automation 尚未披露是否推进这条路

多数标价和折扣条款未公开披露;单元格明确说明这一点,而不是估算数字。数字混合 2024-2026 年不同时间点,应视为变现路径的方向性证据,不是当前价目表。

[CP044, CP001, CP021, CP015, CP016, CP035]
FP002: 功能广度 / 能力图

围绕五项能力标准,为 Core Automation 和九家代表性竞争者打分。

「不明」表示截至运行日未见公开披露,不代表已确认的负面结论;单元格概括本章来源中的定性证据,不是标准化评分表。

[CP044, CP008, CP037, CP021, CP046, CP035]

3.6 护城河耐久性、快速跟进风险与不利证据

独立证据确实让 Core Automation 任何护城河的耐久性存疑。Radical Ventures 把蒸馏和开放权重商品化标为每个 NeoLab 的长期风险;其熊市情形——团队还没交付任何东西,人才就为既有玩家 10x 薪酬离开——对 Core Automation 是直接风险,因为其名册本就是从资本更充足的 Anthropic 和 DeepMind 招来。TechSpot 报道称,全球具备前沿能力的 AI 研究人员约 2,000 人,Meta 开出的签约奖金最高达 $100M,这意味着 Core Automation 刚招来的人仍可被更高报价挖走。关于 Core Automation 自身发布的独立评论明确持怀疑态度,称行业已经「听过自动化发现这个故事十几次了」,并称自主研究系统可能「过拟合自己的噪声」。另据 Forbes 披露的泄露财务数据,OpenAI 2025 年营业亏损约 $21B,收入 $13B;Palantir CEO 称 AI token 商业模式「疯狂」;AMI Labs 创始人 Yann LeCun 警告,如果不削减成本或涨价,前沿实验室可能面临「大泡沫爆炸」——这些不利信号意味着,一旦更广泛的 AI 融资回调,Core Automation 这种未产生收入的新进入者,受到的冲击可能大于本章提到的任何能产生现金流的既有玩家。[CP028, CP029, CP030, CP031, CP032, CP033]

护城河耐久性 / 竞争风险登记表
护城河主张威胁严重性缓释措施 / 尽调问题
「训练数据少约 100x」的持续学习突破(Ceres)Adaption Labs 和 AMI Labs 也在用独立资金和团队推进高度相近的持续学习 / 世界模型命题,因此差异化并不独有索取任何将 Ceres 与 Adaption Labs 无梯度方法或 AMI Labs 基于 JEPA 的结果对比的内部基准
创始人 / 团队履历吸引力Meta/OpenAI 级别签约奖金(最高 $100M)和总薪酬($3M-$10M+)可能把已从 Anthropic/DeepMind 招来的人才再挖走索取留任协议、归属时间表,以及创立以来任何离职情况
通过 Nvidia 获得算力据报 Nvidia 是投资方,不是像 Anthropic 的 AWS/Google/SpaceX 交易或 OpenAI 约 $300B Oracle Stargate 合同那样承诺多吉瓦供应的供应商;股权关系不保证算力分配索取任何已签署、区别于股权投资的算力供应协议(期限、容量、价格)
「自动化 AI 研究」的先发主张Google DeepMind 已经运行内部自动化数学 / 科学研究智能体(Aletheia),并产出同行评审成果;OpenAI/Anthropic 也以生产级规模发布前沿研究索取技术备忘录,说明 Core Automation 的方法与 DeepMind 的 Gemini Deep Think 研究智能体工作有何不同
无已上线产品 / 无分销Glean、Hebbia、Manus 已有付费企业客户,并在同一自动化知识工作领域上线集成;Core Automation 尚未发货,工作流层可能已被封住索取 Core Automation 的商业化落地计划和首个付费试点时间表
开放研究 / 社区可信度Thinking Machines Lab 和 AMI Labs 都计划或已经实践开放发表与开放权重发布;一旦算法优势被证明,可能被商品化询问 Core Automation 是否打算发表或申请专利保护其持续学习方法
资本续航与泡沫风险可信批评者(泄露的 OpenAI 财务数据、Palantir 的 Alex Karp、Yann LeCun)认为前沿实验室经济性靠补贴支撑且脆弱;融资市场回撤对未产生收入的新进入者打击可能比对能造现金的现任厂商更大索取资金续航、烧钱率,以及下一轮延迟或缩水时的应急计划
人才池稀缺全球估计约 2,000 人被认为能构建前沿 AI 系统,而本章几乎每个具名竞争者都在争夺同一批人索取团队稳定性数据,以及初始创始团队之外的招聘管线深度

严重性评级是作者基于证据的定性判断(高 / 中),不是标准化评分模型;每行都引用用于评级的具体竞争者证据。

[CP037, CP005, CP033, CP034, CP001, CP047]
FP003: 护城河 / 就绪度 KPI

用紧凑数字概括本章提到的竞争耐久性信号。

计数来自本章对 TP001 中具名公司的统计:哪些公司符合各项标准;这不是穷尽式市场普查。估值占比 KPI 用 Core Automation 据报 ~$4B 的目标估值,对比 Anthropic $965B 的投后估值(~0.41%,四舍五入为 0.4%)。

[CP005, CP037, CP001, CP021, CP025, CP028]

3.7 图表与证据

Chapter 04

04财务

4.1 已有记录:注册与已报道融资

Core Automation 唯一可独立核验的政府记录,是 California Secretary of State 的一份备案:显示「Core Automation (de), Inc.」为 Delaware 设立的股份公司,正式备案日期为 2026 年 3 月 24 日,文件号 B20260125942,Jerry(Jaroslaw)Tworek 列为注册代理人。该备案披露了实体类型、备案日期和注册代理人地址,但没有任何财务数字——没有授权股份数、没有已融资本、没有资金用途说明。对 SEC EDGAR 全文检索限制在 2026 年 1 月 1 日至 7 月 5 日之间的 Form D 备案,搜索「Core Automation」返回零结果;这意味着截至本次运行日期,公司尚未提交豁免证券发行的公开通知,尽管多家媒体报道其初始融资约 $100M、估值 $1B,随后又推动 $300M-$500M 融资、目标估值接近 $4B。已报道但未独立确认的初始轮投资者包括 Nvidia、Spark Capital 和 Accel。由于 Core Automation 没有申请 IPO 或发行公开债务,它没有义务发布审计财务,因此本章每个美元数字都追溯到 The Information 通过匿名人士得到的媒体报道,而不是备案文件——这一证据基础明显弱于招股书或 10-K。[CI001, CI002, CI003, CI004, CI005, CI006]

资本充足性表
项目数值来源状态尽调问题
账面现金所审来源均未披露索取最新银行账户或股权结构表证明
月度烧钱所审来源均未披露索取过去三个月按类别拆分的烧钱额
资金续航(月)没有现金和烧钱数字,无法计算仅凭公开数据无法推导
已确认融资$100M(初始轮)多家媒体报道,可追溯至 The Information;未找到 SEC 文件通过已签署融资文件确认
目标新融资(洽谈中)$300M-$500M,估值约 $4B截至 2026 年 5 月已有报道;本轮报告日期尚未确认是否交割确认交割状态和最终条款

现金、烧钱额和资金续航为空,因为没有公开来源披露;所示融资数字为媒体报道金额,不是备案或审计数字。

[CI001, CI002, CI003, CI005, CI007]

4.2 无收入、无产品:商业披露缺口

分析机构 Sacra 报告称,截至 2026 年 5 月,Core Automation 没有公开 API、定价页、注册流程或商业产品,并把公司描述为「未产生收入且未商业化」,成本结构由前沿 AI 研究人才和算力主导,没有客户收入抵消。Sacra 将公司当前唯一「产品」描述为对自身内部研究流程的自动化——实验室既是自身自动化栈的建设者,也是第一个客户,并不是向外部买方销售的供应商。Core Automation 自己的首页也强化了这一点:它把使命框定为打造「全球自动化程度最高的 AI 实验室」,可抓取文本中没有任何定价、产品目录或收入表述。技术博客「When AI Starts Writing Systems Code」由客户端渲染,抓取时几乎不返回静态文本,进一步限制了除标题之外任何业务细节的独立核验。公司网站的 /careers 路径也返回 404,因此没有可见的公开招聘入口,无法显示除已报道研究招聘之外还有财务、法务、销售或运营招聘活动。Sacra 提出几个可能的未来变现路径——B2B 模型或 API 访问、面向特定领域自动化的企业订阅,以及与自主任务或算力绑定的用量计费——但把它们都框定为猜测,且未获公司确认。本次审阅没有任何来源识别出具名付费客户、已签合同或任何形式的已披露收入数字。[CI008, CI009, CI010, CI011, CI012, CI013]

收入来源表
收入来源机制单位当前数值 / 状态证据质量尽调问题
商业产品销售未披露n/a截至 2026 年 5 月,无公开 API、定价页或注册流程(Sacra)有记录的缺失;独立分析师确认是否存在任何私有试点或设计伙伴收入
模型 / API 授权(未来)Sacra 推测的 B2B 模型或 API 访问n/a路线图愿景,尚未上线分析师推断,公司未确认索取产品路线图和目标发布日期
企业订阅(未来)Sacra 推测的特定领域自动化订阅n/a路线图愿景,尚未上线分析师推断,公司未确认索取任何已签署意向书或设计伙伴协议
按用量 / 算力计价(未来)Sacra 推测的、绑定自主任务 / 算力的定价n/a路线图愿景,尚未上线分析师推断,公司未确认如有,索取定价模型测试结果或试点条款

四行全部描述已记录的缺失,或第三方对未来变现的推测;截至本轮报告日期,公司未确认任何收入来源。

[CI008, CI010, CI011, CI015]
定价 / 变现表
模式标价与实际成交价包含能力折扣 / 未知项来源
公开定价页不存在n/a不清楚是否存在内部定价实验Sacra;coreauto.com 主页
B2B 模型 / API 访问(假设)未发布标价不清楚——尚未产品化完全是推测Sacra 分析师推断
企业订阅(假设)未发布标价不清楚——尚未产品化完全是推测Sacra 分析师推断
按用量 / 算力绑定定价(假设)未发布标价不清楚——尚未产品化完全是推测Sacra 分析师推断

每个假设行都是第三方分析师推测,不是公司发布的价目表;仅可作为方向性参考。

[CI008, CI011, CI012]
FI001: 收入模型桥接图

从 Core Automation 内部研究自动化栈走向潜在未来收入的假设路径;该路径尚未确认。

这是一座由第三方分析师推测(Sacra)拼出的假设桥;Core Automation 尚未确认任何路径或发布日期。

[CI010, CI011]

4.3 成本结构来自推断而非披露:人才和算力可能是主要驱动

由于 Core Automation 没有披露员工数、薪酬或算力支出,本章用可比前沿实验室来基准化可能的成本驱动,而不是编造公司特定数字。行业薪酬报道显示,2026 年资深前沿 AI 研究工程师总薪酬(基本工资、奖金和股权)约为每年 $500,000 到 $1.5M;顶级研究人员的离群套餐据报道最高可达四年 $300M,最大实验室单年签约奖金最高达 $100M——这是前沿研究人才成本有多昂贵的极端例子。算力方面,OpenAI 报告 2026 年第一季度经营现金消耗约 $3.7B,同时计划 2026 年训练和算力支出约 $32B;据报道,Anthropic 到 2026 年 4 月年化收入约达 $30B,但每年仍在算力上支出约 $6B-$10B,每月现金消耗约 $80M,其中超过 60% 用于云基础设施。Deloitte 报告称,单位 AI 推理成本两年内约下降 280 倍,但企业 AI 总支出仍继续上升,因为使用量增长在全行业超过了效率提升——这一动态很可能同样适用于任何算力密集型研究实验室,包括 Core Automation。鉴于这种人才和算力密集、未产生收入的运营模式与这些同行相似,Core Automation 自身 burn 方向上可能落在可比区间;但公司没有披露特定数字,本次审阅也没有来源说明 Core Automation 当前员工数或总体薪酬年化支出。[CI016, CI017, CI018, CI019, CI020, CI021]

单位经济性表
指标数值置信度重要性尽调问题
毛利率unknown没有收入,无法计算毛利率产品上线后索取服务成本假设
CAC / 回本周期unknown未披露客户或销售动作索取 GTM 计划和任何试点层面的经济性
前沿 AI 研究员总薪酬(基准)$500K-$1.5M 常见;异常值据报可达 $300M 多年薪酬包小型研究密集团队里,这很可能是最大成本项索取实际薪资年化支出和按职能划分的员工数
可比实验室算力支出(基准)OpenAI 计划 2026 年训练 / 算力支出约 $32B;Anthropic 约 $6-10B/年框定前沿实验室维持的算力支出规模;为建模 Core Automation 自身烧钱规模提供方向性上限 / 下限索取 GPU / 云合同承诺和单位算力成本
推理成本下降与总支出两年内单位成本下降约 280x;AI 基础设施总支出仍在上升(2026 年 $300B+)效率提升被全行业用量增长追上甚至超过,Core Automation 很可能复刻这一模式可获得后,索取 Core Automation 自身算力利用率和单次实验成本趋势

第 3-5 行是行业 / 同行基准,用作一家未披露自身单位经济性的公司的方向性代理;没有一项是 Core Automation 专属数据。

[CI016, CI018, CI019, CI022]
FI002: 单位经济模型桥接图

Core Automation 未披露自身单位经济模型,因此用同业实验室和行业基准作方向性代理,搭出定性的成本到烧钱桥接。

节点只把同业实验室和行业基准区间作为方向性代理;没有一个数字是 Core Automation 专属数据。

[CI016, CI018, CI019, CI023]

4.4 Neolab 类别的资本密集度与现金跑道敏感性

放到 2024-2026 年可比「neolab」超大种子轮旁边看,Core Automation 已报道的估值轨迹并不异常。前 OpenAI CTO Mira Murati 创办的 Thinking Machines Lab,在交付任何商业产品之前,于 2025 年 7 月以 $12B 估值完成 $2B 种子轮。前 OpenAI 首席科学家 Ilya Sutskever 共同创办的 Safe Superintelligence,在 2024 年 9 月以约 $5B 估值完成 $1B 种子轮。Humans& 由前 Anthropic、xAI 和 Google 研究人员在 2025 年 9 月创立,成立三个月后的 2026 年 1 月以 $4.48B 估值完成 $480M 种子轮。受超大 AI 实验室融资强力推动,2026 年第一季度全球初创公司融资创下约 $300B 的记录,覆盖约 6,000 家初创公司。在这一背景下,Core Automation 已报道的 $100M 到 $1B、再到 $300M-$500M 和 $4B 的推进,落在同行已经建立的区间之内,且绝对美元规模低于同行。这说明,产品前公司超过十亿美元估值已经是类别常态,而不是 Core Automation 特有的离群值。Sacra 明确把公司的核心商业模式风险放在现金跑道上:内部自动化「飞轮」能否在已融资本耗尽前足够快地复利,产出可分发产品。这一框架对尽调很重要,因为没有来源披露 Core Automation 手头现金、月度 burn 或现金跑道(月数);也没有来源披露已确认或目标融资中算力、招聘和设施的资金用途拆分。[CI025, CI026, CI027, CI028, CI029, CI030]

FI003: 财务估算区间

将 Core Automation 据报估值区间,对照 2024-2026 年新型 AI 实验室种子阶段可比公司,以及 AI 基础设施资本开支与收入之间的缺口。

除非报道给出区间,新型 AI 实验室估值均把单一报道数字显示为点区间(low=high);Cahn 框架是第三方分析结构,不是 Core Automation 专属估算。

[CI003, CI005, CI025, CI026, CI027, CI033]

4.5 融资市场风险:AI 融资狂热与算力稀缺的不利证据

几个彼此独立的警示信号,直接关系到 Core Automation 未来融资环境是否还能保持友好。Sequoia Capital 合伙人 David Cahn 的 AI 基础设施收入缺口框架,从 2023–2024 年估算所需年化 AI 收入 $200 billion,上升到 2026 年约 $600 billion;原因是超大规模云厂商的 AI 资本开支增长快过已经兑现的 AI 收入。评论者越来越常把这个缺口类比为电信产能泡沫:供应商融资推高建设,真实需求不足时迅速回调。另一路证据来自 MIT NANDA 项目报告 "The GenAI Divide: State of AI in Business 2025"。该报告基于 300 个 AI 部署和 150 次高管访谈,发现 95% 的企业生成式 AI 试点未能交付可衡量的财务 ROI;这对任何最终商业化依赖企业看到可量化价值的研究自动化公司都是警示。供给侧,Nvidia Blackwell 芯片的 GPU 租赁价格据称在 2026 年 4 月升至每小时 $4.08,60 天上涨 48%;同期算力短缺据报道已导致 Anthropic 宕机,并迫使 OpenAI 取消部分产品计划,Bank of America 还预计需求将到 2029 年前持续超过供给。Nvidia 本身——另有报道显示它参与了 Core Automation 的首轮融资——在 2026 年承诺超过 $40 billion 的 AI 股权投资,其中包括对 OpenAI 约 $30 billion 的持股;同时还有算力额度和收入分成交易,Goldman Sachs 分析师将其标记为可能夸大真实终端需求的「循环收入」风险。这些都不能证明 Core Automation 本身参与了循环融资安排,但提供了一条合理传导路径:一旦 Nvidia 的 AI 股权策略或算力价格转向,公司的融资可得性或成本基础可能受到影响。[CI031, CI032, CI033, CI034, CI035, CI036]

公开财务缺口表
缺失的私有指标影响尽调路径
账面现金 / 资金续航如果后续轮次推迟、缩水或未能交割,无法评估生存能力向管理层索取最新资产负债表或银行对账单
员工数和薪资年化支出最可能的最大成本项完全没有量化按职能提供当前组织架构和薪酬区间
计算 / 云合同承诺可能是第二大成本项,也是现金跑道敏感性的关键变量提供 GPU / 云供应商协议和承诺支出
股权结构、所有权和控制条款无法评估稀释、创始人持股集中度或投资人控制权提供股权结构表及任何特殊投票权 / 清算条款
已报道 / 目标轮次的资金用途无法核验资金规模是否足以支撑公司声称的研究议程提供董事会批准的预算或资金用途备忘录

每一行都是截至本次报告日尚未公开披露的指标;这里列的是直接向管理层提出的尽调请求, 而非估算。

[CI024, CI031, CI032, CI045]
FI004: 资本强度 / 现金流图

展示 Core Automation 披露了哪些资本充足性维度,以及哪些维度只有同业 / 行业基准信号。

是 / 否反映是否有公开来源专门披露 Core Automation 的该项信息;基准列引用同业 / 行业数字,不是公司数据。

[CI020, CI021, CI024, CI031, CI036, CI037]

4.6 财务判断:公开证据无法支撑什么

把披露缺口和基准背景合在一起看,截至运行日,外部没有任何可独立核验的 Core Automation 收入、ARR、毛利率、月度烧钱、在手现金或资金续航 数据。公开记录里唯一量化的资本事实——$100 million 首轮融资,以及洽谈中的 $300–500 million 后续融资——都来自媒体报道,并追溯到 The Information 援引的匿名人士,而不是备案文件;所谓 $100 million 以上融资没有任何 SEC Form D 备案,本身也有信息量,意味着该发行可能尚未触发公开通知备案,可能使用了另一条豁免路径,也可能是报道中的交易条款仍处初步阶段。在一个资本密集、尚无收入的商业模式里,同类公司在产品发布前就已投入 $1 billion 或更多,外部投资人不能只靠公开证据承销 Core Automation 的财务画像。若要显著提高投资判断的置信度,管理层需要提供数据:签署版股权结构表和融资条款书、GPU/云算力合同与已承诺支出、按职能拆分的当前人数和薪酬支出速度,以及报道所称 $4 billion 融资前任何未偿可转债、SAFE 或债务工具的披露。已审阅来源没有披露 Core Automation 的股权结构表、持股比例、董事会构成或债务/可转工具,因此治理和稀释风险与烧钱速度一样,仍无法量化。[CI040, CI041, CI042, CI043, CI044, CI045]

4.7 附录

Chapter 05

05产品与技术

5.1 产品看起来是什么,真实成熟度到哪一步

Core Automation 的公开材料把使命说得很清楚,却没有同样清楚地定义产品。官网称公司在打造「世界上最自动化的 AI 实验室」,希望系统能优化并自动化工作,并从研究本身开始。这个表述很关键,因为它把当前交付物更像放在内部研究操作系统的位置,而不是已经对外发布的应用。Nextomoro 2026 年 5 月的画像强化了同一判断:这是一家聚焦自动化前沿 AI 研究工作流、并试图用更自适应学习算法替代大规模预训练的实验室。同样重要的是缺失项:没有公开 API 文档页、没有产品定价、没有状态页面、没有模型卡,也没有 Ceres 或任何其他具名模型的基准测试表。因此,本产品章节必须把概念资产和已验证资产拆开。公开层面,真正的资产是论点、人才画像和系统叙事;外部产品界面仍像是内部使用、发布前状态,或尚未披露。[CE001, CE002, CE007, CE008, CE027, CE035]

产品模块 / 资产矩阵
模块或资产主要用户公开状态差异化假设尽调缺口
Ceres / 已命名的持续学习模型概念先面向内部研究员;未来可能面向外部用户概念阶段 / 对外未披露可能让模型改进持续发生,而不再受完整重训周期束缚无公开模型卡、基准或演示
内部研究智能体工作流Core Automation 研究团队由使命陈述推断可能压缩实验设计和迭代循环无工作流截图或量化生产力提升
系统代码自动化层研究工程师 / 基础设施工程师由博客和 Saroufim 相关足迹推断可能降低底层工程瓶颈无公开代码库或实现细节
评估与基准栈研究团队和安全评审人员未公开披露可能决定持续学习主张是否成立无公开评估方法或结果
未来对外接口(API / 工作流软件)企业开发者或研究机构尚未公开发布可能是最直接的商业化路径无文档、定价、SLA 或试点证据

这张表把真实公开资产与推断的未来产品界面拆开;公司对外成熟度仍显著低于其论题成熟度。

[CE001, CE002, CE007, CE008, CE027, CE035]
工作流 / 用例表
用户任务当前工作流Core Automation 解决方案假设若属实可衡量的收益当前限制
研究构思人来界定问题并选择实验智能体帮助提出并打磨下一批实验从想法到测试的迭代更快无公开前后对比指标
系统优化工程师手工调优内核和基础设施系统代码自动化起草或优化内核降低基础设施瓶颈,加快训练无公开实现证明
模型适配静态模型需要完整重训周期持续学习栈逐步适配降低数据和重训负担灾难性遗忘仍未解决
科学发现研究人员手工串起假设、代码和分析研究智能体协调多步骤工作流探索性研究吞吐更高无公开 Core Automation 案例
未来企业自动化开发者或分析师使用割裂工具可能在研究栈之上做出工作流产品可能把研究自动化转化为可销售软件无已发布产品界面

这些行描述的是公开论题暗示的工作流,不是已完全验证的生产部署。

[CE002, CE006, CE018, CE025, CE029]
FE002: 客户工作流 / 运营流程

在任何外部产品界面出现之前,该论点先从内部研究员工作流起步。

流程反映使命表述和相邻系统证据;目前没有公开的 Core Automation 工作流图。

[CE002, CE006]

5.2 架构论点:持续学习、后 Transformer 设计与系统代码自动化

对一家整体低调的公司来说,其技术论点异常明确。Core Automation 表示,它不认为 AI 的下一次阶跃来自更大的模型、更多数据和静态部署;它指向新的学习算法、优于 Transformer 的架构,以及围绕高能力智能体搭建的实验室。这与当前研究前沿一致。2026 年面向 LLM 的持续学习综述,把该领域定义为让模型适应不断演化的知识,同时限制灾难性遗忘;更广泛的灾难性遗忘文献也说得很清楚:这仍是艰难且未解决的系统问题,不是已经跑通的产品配方。后 Transformer 综述又加了一层:研究者正积极探索替代路线,因为 Transformer 依然强大但有缺陷。换句话说,Core Automation 瞄准的是一个真实技术问题,但不是无人竞争的问题。这个领域充满论文、基准和实现,公司最终的差异化必须来自可测能力和运行效率,而不能只靠品类标签。[CE003, CE004, CE005, CE009, CE010, CE011]

技术 / 运营架构表
层级或组件作用关键依赖风险
持续学习方法层让模型适配新数据或新任务限制遗忘的算法可能无法保留既有知识
后 Transformer 架构选择在标准 Transformer 之外提升效率或能力新型架构研究可能跑不赢成熟 Transformer 栈
研究智能体编排协调编码、评估和迭代循环强智能体规划和工具集成可能产生隐性失败链或噪声输出
GPU 内核 / 系统优化提升吞吐和成本效率顶尖系统工程师和硬件资源专长集中且与硬件强耦合
评估与基准测试衡量收益和退化私有数据集和严谨测试框架没有公开证明,外部无法验证
安全 / 治理层约束行动并监控故障人工评审、政策和工具目前未公开记录

该架构来自官方公司表述、围绕 Saroufim 的开发者信号以及相邻的已发表系统工作综合推断, 并非 Core Automation 发布的架构图。

[CE004, CE005, CE009, CE011, CE018, CE029]
FE001: 产品架构图

可见技术栈从新型学习方法一路延伸到系统优化和治理。

这套栈来自公开表述和相邻开发者信号的推断,不是公司发布的架构图。

[CE004, CE018]
FE004: 产品成熟度 / 能力图

该论点在愿景和人才上成熟,但公开证明和控制仍不成熟。

矩阵分数是基于外部证据质量的定性判断,不是公司内部评分卡。

[CE027, CE033]

5.3 依赖、开发者信号与隐含路线图

本章最强的外部证据来自开发者信号,而不是公司的自有代码库或文档。Mark Saroufim 的公开作品集和 GitHub 足迹,正好展示了 Core Automation 若想把 kernel、训练工作流和基础设施纳入研究循环自动化,会需要的底层系统经验。GPU MODE 的课程、YouTube 频道和相关材料,围绕这些技能形成了一个异常可见的实践者社区;PyTorch 的 KernelAgent 文章也显示,相邻领域已经在用多智能体工作流,结合真实硬件反馈优化 GPU kernel。联合创始人兼 CEO Jerry Tworek 则提供了第二个、可独立交叉验证的可信度信号:行业画像称他曾任 OpenAI 研究副总裁,领导 o1 和 o3 推理模型项目,并参与 Codex 与 GPT-3/GPT-4 系列,让创始团队在公司声称要超越的推理和后训练研究上,具备直接的前沿实验室经验。换句话说,Core Automation 公开故事所暗示的系统栈可信,是因为团队过去的公开工作可信。但可信度不等于产品成熟。路线图仍只能推断:先自动化内部研究任务,证明这些智能体能提升生产力,之后也许再对外提供工具、API 或模型访问。在此期间,关键依赖——算力、评测工具、顶尖系统人才和私有研究反馈回路——大多仍不在公开视野内。[CE006, CE016, CE017, CE018, CE019, CE020]

路线图 / 发布 / 开发阶段表
日期或阶段功能或里程碑状态含义来源
2026 年公开发布状态自动化 AI 实验室论题公布仅公开论题产品化之前,使命已经可见Core Automation 主页 / Nextomoro
当前公开状态无 API 文档、定价或信任页面对外未披露对外产品成熟度仍低已审阅的 Core Automation 材料
2025-10 同行基准Thinking Machines 发布 Tinker已上线产品同行已推出面向开发者的工具Thinking Machines
2026-05 同行基准FutureHouse 公开 Robin研究系统已公开上线研究自动化可以对外展示FutureHouse
2026-02 同行基准Gemini Deep Think 研究智能体获介绍竞争对手已有上线研究工作流自动化科学的竞争门槛已在抬升Google DeepMind
推断的下一步从内部研究自动化走向最终对外产品化推断可能路径是先证明内部 ROI,再对外发布公开证据综合

路线图一部分来自观察,一部分来自推断,因为 Core Automation 尚未发布正式发布计划。

[CE007, CE008, CE024, CE025, CE026, CE035]
FE003: 核心依赖图

Core Automation 暗含的技术栈依赖稀缺人才、硬件和私有评估循环。

公开证据只能间接看见这些依赖,因此本图是综合判断,不是公司披露。

[CE017, CE037]

5.4 信任、质量与就绪缺口

产品成熟度不只看模型质量,也看买方能不能在部署中信任系统。Core Automation 的公开面在这里最薄。本章审阅的材料没有呈现信任中心、安全架构、正常运行时间承诺、合规认证或公开支持模式。行业背景让这个缺口更严重,而不是更轻。Info-Tech、Google 和其他 2026 年治理报告都把智能体驱动自动化描述为正在走出实验阶段,进入自适应治理、控制界面和负责任 AI 流程成为基本要求的阶段。Google 自己的进展报告展示了成熟组织现在会围绕原则和生命周期控制发布什么。相比之下,Core Automation 的公开证据仍停在论点和团队。这不意味着公司没有内部控制,但意味着外部尽调无法核验。除非出现模型卡、安全文档、事故流程或生产可靠性材料,否则这项技术应被视为有前景但运营上尚不成熟。[CE030, CE031, CE032, CE033, CE034, CE039]

信任 / 质量 / 合规表
控制或质量信号公开状态范围缺口
模型卡 / 评估报告未公开出现可说明性能和限制无公开基准或评估材料
安全 / 信任中心未公开出现可说明控制措施和架构未发现公开安全态势页面
状态 / 可用性页面未公开出现可支持可靠性审查未发现公开服务运营信号
负责任 AI 流程有一般行业背景可参考Google 等同行已公开生命周期控制做法未看到 Core Automation 对应披露
合规认证未公开出现可帮助企业采购未发现 SOC 2、ISO 或类似证据
支持 / 事件流程未公开出现可显示运营就绪度未发现公开升级或响应承诺

这里的缺口表示在已审阅材料中没有浮现公开证据,并不证明内部控制不存在。

[CE030, CE031, CE032, CE033, CE034]

5.5 附录

Chapter 06

06客户

6.1 可能的目标客户:研究团队、R&D 买方与知识工作运营方

Core Automation 的公开材料没有点名买方,但强烈暗示了买方是谁。公司把自己描述为在建设自动化 AI 实验室,并从研究本身开始创建能优化和自动化工作的系统。这套语言指向的不是消费者分发,而是研究密集型组织、企业 R&D 团队、前沿实验室,以及已经在实验、代码和信息检索上重金投入的高价值知识工作团队。同类公开证据从另一个角度说明了同一点。Glean 的横向企业采用横跨 IT、支持、知识管理和业务运营;Hebbia 的证据则集中在金融、法律和类似高风险文档工作流。如果 Core Automation 最终把内部技术栈产品化,最早一批合理客户很可能坐在这些世界的交汇处:技术或分析团队,人力工作流昂贵,数据资产庞大,并且有理由为更快迭代付费。但这个分群仍是推断,因为 Core Automation 还没有发布客户名称、垂直行业或买方画像。[CU001, CU005, CU006, CU007, CU010, CU025]

客户细分表
细分市场买方 / 用户 / 付费方核心用例战略价值当前缺口
前沿 AI 实验室或研究机构研究负责人 / 基础设施负责人 / CTO自动化实验设计、编码、评估和迭代与论题最匹配,也最愿意测试前沿工具尚无公开客户背书
企业 R&D / 创新团队R&D 负责人 / 技术赞助人 / 预算负责人加快内部模型研究和应用实验有预算和数据的早期设计伙伴候选无公开产品包装
知识工作平台团队IT、知识或运营负责人自动化文档搜索、综合和工作流路由若产品化,存在大席位或工作流机会Glean 类在位者已很拥挤
金融 / 法律研究团队董事总经理 / 业务负责人 / 运营高风险文档分析和决策支持质量足够高时,ROI 和痛点都清晰Hebbia 类垂直专家已存在
开发者平台或基础设施团队工程领导层自动化系统代码和内部工具工作匹配 Saroufim 式系统能力无公开证明表明外部模块存在

这些细分市场依据使命、相邻客户验证以及公开可比公司已经变现的工作流类型推断而来。

[CU001, CU006, CU007, CU010, CU025, CU026]
可比公司客户牵引力基准
可比对象公开客户证据买家类型为什么重要局限
Glean具名公开客户证言和客户案例横向企业职能说明广泛内部工作流有需求未披露准确账户数
HebbiaOak Hill 具名证言及规模指标垂直金融 / 法律 / 高风险分析说明专业研究工作流愿意为 AI 工具付费大部分证据由公司披露
Anthropic 企业智能体调研调研了 500+ 名技术负责人对智能体部署的看法跨行业技术买家说明企业多步骤智能体工作流有需求调研结果,不是具体产品客户数
Microsoft Copilot / Power Automate公开的企业自动化平台叙事知识工作者和流程负责人表明大型既有厂商正在培养买家对智能体工作流的预期平台采用不等于 Core Automation 适配
Core Automation尚无公开客户证据未知 / 推测为研究和研发买家凸显公司在商业化证据上仍处早期公开证据缺失不等于私下尽调也没有证据

这张额外基准表让 Core Automation 的客户证据缺口保持可见,同时用公开可比对象锚定品类。

[CU005, CU007, CU016, CU030, CU031, CU039]
FU001: 客户旅程图

Core Automation 可能需要由内部发起人牵头的企业旅程,而不是自助式采用。

旅程来自公开使命表述和 2026 年企业采购信号的推断,不是 Core Automation 已发布的销售流程。

[CU029, CU034]

6.2 公开证据:同类需求强,Core Automation 直接证据尚未出现

客户证据的不对称,是本章的核心事实。公开层面,Glean 和 Hebbia 已经提供了具名或至少按角色区分的证据,说明组织在近似生产环境中使用其平台。Glean 的客户故事页展示了跨多个企业职能的公开推荐,并点名一个公司案例;Hebbia 官网则同时提供具名推荐,以及提示词、处理页数和 AUM 覆盖等大规模使用主张。这些信号不能直接证明 Core Automation 的需求,却能证明相邻品类里存在对工作流自动化和研究辅助的买方胃口。相比之下,Core Automation 的公开材料仍没有客户标识、没有试点公告、没有设计伙伴推荐,也没有部署结果。这意味着市场里的具名客户证据几乎全部属于可比公司,而不是 Core Automation 自身。做尽调时,这个缺口比品类热度更重要,因为客户证据质量通常是产品能否从研究雄心走向可重复采用的第一个硬信号。[CU002, CU004, CU005, CU008, CU009, CU011]

客户增长 / 采用轨迹表
指标日期来源置信度含义缺失分母
Core Automation 具名客户公开披露为 02026-07-05在已审阅公开材料中观察到尚无直接采用证明可能不同于私下真实情况
Hebbia 覆盖 AUM使用 Hebbia 的机构覆盖 $30T AUM2026Hebbia 主页显示其在垂直研究工作流中真正打入企业未披露账户数
Hebbia 每日提示词数日均 200k 条提示2026Hebbia 主页表明有重复使用,不只是攒 Logo未披露留存数据
Hebbia 处理页面数已处理 1.5B 页2026Hebbia 主页显示其在文档密集型工作流中的规模未披露客户结构
Glean 估算收入2026 年收入 $300M2026-07-03GetLatka 估算表明横向企业需求可能显著放大估算值,未审计
企业 AI 智能体生产环境采用54% 企业已将 AI 智能体集成进核心运营2026 年中Ampcome市场对智能体工具的准备度正在提高非研究自动化专属
企业应用嵌入预计到 2026 年底,40% 企业应用将集成智能体2026 年预测Reinventing AI / Gartner 引用更广泛的买方熟悉度应会提升预测值,不是已观察到的 Core Automation 需求

这张表把 Core Automation 尚无公开客户证明的直接观察,与可比公司的客户指标和市场采用代理指标放在一起。

[CU008, CU011, CU019, CU021, CU027, CU038]
具名客户证明表
客户或证明材料细分市场部署 / 用例生产 vs 试点结果限制
Core Automation(未披露)Unknown未公开披露具名部署或试点Unknown无直接证明缺少证明本身就是发现
McCarthy Holdings / Glean 客户故事页企业运营AI 驱动的知识访问和内部工作搜索公开案例研究材料显示 Glean 有具名公开客户背书已审阅摘录中的详细指标有限
Glean 具名实践者(Michael Bassani、Kathleen Cauley、Elizabeth Vaggelatos、Manu Narayan)IT 运营 / 知识 / 业务运营更快查找文档、更快解决事件、部署生成式 AI 应用按证言看更偏生产环境跨职能多个角色验证了横向使用证言未披露合同金额或留存
Oak Hill Advisors 对 Hebbia 的证言金融 / 资产管理分析师提效和想法生成按证言看更偏生产环境具名用户称 Hebbia 影响了投资决策单条证言无法展示完整客户群

这是已审阅材料中可见具名公开证明材料的部分枚举,不是该类别所有可比客户的完整地图。

[CU002, CU005, CU009, CU027, CU039]
FU003: 客户证明矩阵

相比相邻 AI 工作流供应商,Core Automation 的公开客户证明少得多。

矩阵取值是基于公开材料对证据质量的定性判断,不是公司内部评分。

[CU005, CU008]

6.3 从试点到生产:市场时机有利,采购摩擦仍重要

更大的市场背景支持,但并不宽容。Google、Microsoft、Anthropic、Deloitte、Ampcome 和其他 2026 年调查给出的信息一致:企业正把智能体系统从实验带入多步骤生产工作流。这个背景有利,因为 Core Automation 不必从零教育市场;买方越来越熟悉智能体式自动化,也想要可衡量结果。难点在好奇心之后的一切。采购、治理和信任审查正成为购买流程的核心部分,尤其是系统能够规划、执行并触达敏感数据时。Infosys 和聚焦采购的来源强调护栏、韧性和自主决策约束,而不是单纯的新颖性。对 Core Automation 来说,这很可能意味着第一波商业化动作会是高接触、由赞助人推动:研究或创新负责人支持一个试点,买方进行技术和治理审查,只有当供应商证明了可衡量的工作流收益,才会进一步扩大部署。这条路径可行,但比自助式采用更慢,也更集中。[CU013, CU014, CU015, CU016, CU017, CU018]

扩张与集中度风险表
扩张驱动因素集中度或阻力风险影响尽调路径
研究工作流试点成功少数早期设计伙伴可能主导路线图,并形成议价杠杆按账户审查 ARR 预测和产品优先级图
满足治理要求的部署采购和信任审查可能明显拖慢扩张收集所需安全、审计和模型治理材料
工作流集成集成不足可能卡住更大范围推出中高梳理所需连接器和实施负担
可量化 ROI 证明如果看不到时间或质量收益,续约风险会迅速上升要求提供试点 KPI 看板和商业论证材料
先落地再扩张打法目前没有公开证据证明 Core Automation 能从一个团队扩到多个团队要求提供管线、转化和扩张假设
预算竞争既有厂商可能已经吃掉买家用于 AI 搜索或智能体自动化的预算中高了解它与 Glean、Hebbia、OpenAI 和 Microsoft 预算项的品类重叠

这些风险来自直接客户证据缺失,以及 2026 年智能体 AI 市场的采购和项目取消信号。

[CU022, CU023, CU024, CU029, CU033, CU035]
FU002: 采用 / 部署漏斗

市场支持智能体采用,但 Core Automation 仍必须跨过企业证明和治理关口。

由于没有公开管道数据,这是概念漏斗,不是经实测的 Core Automation 转化漏斗。

[CU019, CU023]

6.4 留存、扩张与客户判断

因此,客户判断明显分裂。正面看,相邻市场清楚显示,企业会购买能改善研究、检索、分析和工作流自动化的工具,也越来越期待智能体系统承担更多工作。负面看,本章审阅的公开证据没有显示 Core Automation 自身跨过了哪怕第一个持久客户门槛。没有披露的参考账户、没有续约或留存指标、没有满意度评分,也没有客户集中度数据。这里的负面市场证据同样重要:如果相当一部分企业智能体项目会因执行风险而被取消或放慢,那么一家尚未有产品、也没有参考客户的供应商,起点不是白纸,而是信任赤字。实际含义很简单。Core Automation 未来可能有真实客户,但今天的客户故事仍是投资判断假设。公司需要参考账户、试点结果和留存信号,之后采用才可以被当作资产,而不是预期。[CU022, CU028, CU035, CU036, CU037, CU038]

留存 / 重复使用 / 满意度表
指标细分市场置信度尽调请求
Core Automation NRR / GRRUnknown按账户提供续约、扩张和流失数据
Core Automation 合同期限Unknown提供试点、年度或多年合同结构
Core Automation 客户背书Unknown要求提供客户访谈或书面案例研究
Hebbia 复用代理指标每天 200k 条提示词金融 / 法律研究用户核验活跃用户、席位数和续约画像
Glean 满意度代理指标多个跨职能公开客户证言横向企业用户核验这些证言是否对应部署扩张
企业购买准备度治理和采购审查在加码广泛的企业 AI 买家梳理试点所需控制项和采购障碍

Core Automation 的空值是刻意保留的,因为公开资料没有留存或满意度证据。

[CU008, CU028, CU033, CU037]

6.5 附录

Chapter 07

07风险

7.1 按严重程度排序的风险概览

Core Automation 的核心风险不是单一法律事件或技术缺陷,而是节奏排序。公开证据仍把它描述为一家实验室优先的研究项目,想先自动化自己的工作;它还不是有客户、有信任材料、有外部买方可测试产品界面的商业平台。与此同时,第三方报道已经把公司放进前沿实验室级融资预期和估值叙事里。两者叠加,让产品、客户和披露证据的缺失,比普通种子期创业公司更关键。因此,公开风险图谱集中在三个相互关联的问题上:团队能否在资本预期跑得太远之前,把研究论点转化为买方愿意接受的产品;能否发布足够的信任和合规证据,缩短企业尽调;以及当在位厂商 和开源已经大规模提供智能体、编码和基础设施界面时,公司能否守住差异化工作流。摘要风险登记把产品未成熟、客户证据缺失、融资依赖和商品化列为最高剩余敞口,因为如果证据不能快速改善,每一项都会放大下一项。[CR005, CR011, CR015, CR031, CR045, CR046]

按严重程度排序的风险摘要
风险发生可能性影响缓释成熟度剩余风险敞口
产品未成熟 / 无公开产品界面
无公开客户证据
商业化前依赖融资
商品化 / 被既有厂商挤出
信任 / 合规缺口
关键人 / 治理集中度
算力和人才密集度
证据质量 / 披露缺口

风险评级是作者对公开证据的综合判断,反映当前可见度,而非公司内部控制水平。

[CR011, CR015, CR021, CR037, CR045, CR047]
FR001: 风险热力图

覆盖公开记录可见的主要风险类别:可能性、影响和剩余暴露。

单元格是基于按严重度排序的风险清单和当前公开证据,综合得出的定性评分。

[CR011, CR015, CR021, CR037, CR047]

7.2 监管、法律与信任风险,目前主要是公开面缺失问题

Core 目前没有可见诉讼或执法行动,但公开记录仍留下了有意义且未缓释的法律与信任风险。欧洲通用 AI 合规时间线已经启动,FTC 执法清楚表明,缺乏支撑的自动化主张可能变成法律责任;围绕训练数据、输出和透明度的版权法不确定性仍在发酵。这些问题不会自动击垮公司,但公司尚未发布信任、隐私或合规公开面,客户无从理解 Core 准备如何管理这些问题,因此它们更重要。实际采购中,企业买方会把这种缺席,与已经开放信任门户、审计资源、认证清单和共同责任表述的供应商相比。因此,监管问题与其说是「Core 已经陷入麻烦」,不如说是「Core 还没有给外部一个基础,让人相信它能通过希望取代的供应商同样需要通过的信任审查」。除非公司能展示具体政策、面向客户的法律条款和模型治理映射,否则这会保持为高剩余风险,而不是理论上的未来事项。[CR021, CR026, CR027, CR028, CR029, CR031]

监管 / 法律风险台账
风险当前信号发生可能性严重性缓释状态剩余风险敞口尽调路径
EU AI Act 合规GPAI 规则自 2025 年 8 月生效,更广泛执法从 2026 年 8 月开始公开未见证据要求提供 EU 产品范围界定、模型角色映射和拟定合规控制项
缺少支撑的自动化或效果声明FTC 称 AI 没有免责,并已处罚缺少支撑的 AI 服务声明未见公开支撑框架要求提供产品声明审查流程和法律签批工作流
训练数据和输出版权风险数十起 AI 版权案件仍在审理,拟议透明度规则也未落定未公开披露数据集或授权立场要求提供训练数据来源、许可证和赔偿立场
企业隐私和安全尽调既有厂商发布信任门户,Core 没有可比公开界面公开缺失要求提供 DPA、留存政策、安全架构和事件处理流程
治理标准漂移NIST AI RMF 仍在修订,治理预期在抬升Unknown将控制项映射到当前 NIST 和客户保障预期
跨境销售准备度主要厂商公开 GDPR、FedRAMP、HIPAA 和区域合规资源;Core 没有Unknown要求提供目标市场合规路线图和审计计划

这份部分台账覆盖最显性的公开监管、法律和信任风险,并非逐司法辖区穷尽审查。

[CR021, CR026, CR027, CR028, CR029, CR031]

7.3 产品、客户与运营就绪风险仍明显未解决

最具体的公开尽调发现仍是缺席。官网和博客让研究论点容易理解,但同一个公开面没有让买方评估定价、部署、上线引导 或产品范围。Sacra 明确表示没有公开 API、注册流程或商业产品,公司也被描述为先内部使用自有技术栈,再进行任何公开外部发布。这是一条合理的早期路径,但也意味着没有公开客户证据能把概念强度和市场需求区分开。甚至公司外围运营信息也很薄:团队、联系方式和招聘页面在运行日都返回 404 错误,而首页仍在暗示招聘。在许多 AI 副驾驶和智能体试点已经难以进入规模化生产的企业环境里,这种公开面缺失很关键。买方找不到信任页面、产品工作流或设计伙伴参考,就必须从零完成全部尽调。这会拉长每一次未来客户对话,使其更依赖参考客户,也更容易受到「等等看」的怀疑回应影响。[CR003, CR004, CR005, CR006, CR007, CR014]

运营 / 产品 / 客户证据风险台账
风险公开证据发生可能性严重性缓释成熟度剩余风险敞口
无公开产品界面无公开 API、定价、注册流程或商业产品
无公开客户证据唯一公开证据是实验室内部使用自有工具
运营界面不成熟team、contact 和 careers 页面返回 404
信任审查失败风险未见公开的信任、隐私或合规界面
技术逻辑执行滑坡持续学习和 post-transformer 主张尚未在公开部署中验证
商业验证延迟市场数据显示,许多 AI copilot 仍卡在试点与规模化生产之间

这份台账聚焦可从公开产品、客户和运营信号观察到的风险,而非内部路线图细节。

[CR003, CR004, CR005, CR006, CR007, CR014]
FR002: 风险传导图

投资论点层面的缺口如何传导为商业化放慢、融资承压和估值脆弱。

箭头显示因果风险传导,而非实测弹性。

[CR005, CR011, CR021, CR031, CR045, CR047]

7.4 依赖、融资与竞争风险相互强化

Core 的财务和战略依赖异常交织。公司尚无收入,公开商业模式看起来也更像资本密集型研究机构,而不是软件公司;因此额外融资不是增长加速器,而是证明论点的前提。如果公司在空白市场建设,这会更容易承销。但事实相反,企业 AI 采购正在收紧,在位厂商和开源已经把相邻界面产品化。Microsoft 公开主张,模型优势会压缩进工作流和数据集成;Google 销售受治理的企业智能体平台;AWS 销售智能体式编码和开发工作流;GitHub 销售企业政策和审计功能;开源项目现在也在宣传低成本、OpenAI 兼容的大规模服务能力。风险不在于 Core 没有技术洞察,而在于公开差异化故事可能比公司商业化更快收窄。如果买方已经能从在位厂商 技术栈获得智能体式工作流价值、安全材料和采购舒适度,Core 就需要比「未来自动化实验室优势」更尖锐的证据,来同时支撑采用和估值。[CR008, CR009, CR010, CR011, CR023, CR024]

伙伴 / 依赖 / 融资风险台账
依赖重要性失败情景发生可能性严重性剩余风险敞口
持续外部资本Core 尚未产生收入,仍在客户证据出现前为产品化融资下一轮融资延迟、重新定价,或结构上不可得
前沿人才市场投资逻辑依赖留住异常集中的研究团队人员流失会拖慢路线图,或迫使公司高价补招
算力可得性和成本该运营模式在收入出现前就高度消耗算力GPU 获取或成本压力会压缩现金跑道
既有平台AWS、Google、GitHub、OpenAI 和 Anthropic 已经推出重叠的工作流和信任界面买家选择现有供应商堆栈,而不是等待 Core
开源推理服务层vLLM 和 SGLang 降低了复现基础设施原语的成本Core 的基础设施层变成入场标配
采购纪律新 AI 供应商要接受基准测试、安全、托管和 ROI 审查定制化 Core 产品难以跨过评估门槛

这份依赖台账把外部交易对手、市场结构和融资依赖放在一起,因为三者都会传导到同一条商业化时序风险上。

[CR008, CR009, CR010, CR011, CR023, CR024]
FR003: 依赖关系图

Core 同时依赖资本、人才、算力和既有生态的回应。

图强调公开证据中可见的依赖循环,而非内部运营责任归属。

[CR011, CR039, CR041, CR042, CR043, CR044]

7.5 治理集中与证据质量缺口,让缓释措施大多仍停留在愿景层面

公开的缓释故事仍很薄,因为投资人通常用来下调风险的证据大多根本不存在。公开来源显示,公司高度围绕 Jerry Tworek 这样的创始人兼研究者画像集中,但没有披露董事会构成、投资人控制条款、继任规划或运营授权。因此,公司是在要求市场相信,一个异常小且技术精英化的团队,既能把研究论点转化为可重复商业执行,也能搭建企业买方期待的信任和治理外壳。这可能发生,但尚未被证据证明。实际结论 是,尽调应少看抽象热情,多看证据转化:具名试点、买方参考、安全和隐私材料、外部人员可测试的产品工作流、算力和烧钱披露,以及能够说明公司在规模扩大后不再只是单一创始人叙事的治理材料。在这些材料出现前,缓释方向合理但仍偏愿景,剩余敞口保持高位,因为本报告承销的是未来披露,而不是当前披露。[CR013, CR048, CR051, CR052, CR053, CR054]

人员 / 执行 / 治理风险台账
风险公开信号发生可能性严重性缓释状态尽调路径
Jerry Tworek 关键人风险创始人身份主导公开叙事和可信度Unknown要求提供接班计划和运营授权图
小团队执行集中度小团队逻辑意味着管理冗余有限Unknown要求提供按职能划分的组织架构图和单线程负责人图
治理披露缺口未公开披露董事会构成或投资人控制条款公开缺失要求提供董事会材料、投票权和主要投资人条款
缓释落地缺口公开缓释措施大多是逻辑陈述,不是商业控制项要求提供经营 KPI,证明缓释措施已在使用
尽调负担集中Core 要求投资人押注未来证据,而不是当前证明承诺投资前,坚持拿到具名试点、客户证明和技术评审
剩余治理依赖公开监督图景仍无法与创始人判断区分开Unknown索取会议纪要、委员会结构和独立顾问参与情况

治理行基于公开披露;结论并不预设治理薄弱,但确实显示,风险图景仍有多少无法从公司外部核验。

[CR013, CR051, CR052, CR053, CR054]
缓释措施与论点破裂触发表
风险可监控触发项阈值 / 事件行动含义
产品未成熟公开产品露出下一轮融资里程碑前仍没有公开 API、定价或可试点工作流将论点视为研究可选性,而非近期软件落地能力
客户证明缺口具名客户推荐或试点融资估值上台阶后仍没有具名设计伙伴或部署案例在出现客户验证前暂停形成确信
信任 / 合规缺口公开信任文档尽调中未共享信任中心、DPA、留存政策或安全材料假设企业销售周期仍会受阻或拉长
融资依赖融资条款相对证明证明集实质扩大前,融资规模已大幅扩大将估值视为叙事驱动,而非证据驱动
商品化在位者或 OSS 覆盖类似工作流在位者或 OSS 已覆盖同一工作流,并已带有治理和分发下调纯基础设施或封装层经济性的权重
治理集中董事会与控制权披露NDA 尽调下仍看不到董事会、控制权或继任安排投资前要求更强治理条件

触发项是投资案例的尽调阈值,不是公司发布的经营指引。

[CR047, CR048, CR051, CR052, CR053, CR054]
Chapter 08

08估值

8.1 投资论点与反论点

Core Automation 的多头逻辑,建立在一个拥有可信前沿研究履历的创始团队上;团队追求持续学习和智能体系统自动化论点,而现有大型实验室尚未在公开层面解决这个问题。更广泛的 2024–2026 年模式也支撑这一论点:Safe Superintelligence、Thinking Machines Lab、World Labs 和 Periodic Labs 等可比的产品发布前实验室,都能在创立数月内以数十亿美元估值融资,说明即便没有已发布产品,资本也愿意投向这个精确品类 (CV012, CV015, CV018, CV020, CV026, CV050)。如果得到确认,在种子轮后仅数周就有 $300–500 million 后续融资兴趣,将进一步证明投资人需求仍在 (CV007)。反论点至少同样强。Core Automation 没有提交 SEC Form D,加州公司注册 文件只确认实体存在,并不能确认任一轮融资条款 (CV009)。没有产品、基准、客户或收入证据公开,因此估值无法接受除媒体报道和可比公司模式匹配之外的检验 (CV008, CV034)。同时,2026 年 AI 风险资本异常集中于少数前沿实验室,并出现了有记录的 ARR 膨胀和泡沫动态,可能压缩可比新型实验室目前享有的倍数 (CV029, CV030)。至少两家资金充足、论点相邻的竞争对手也在独立追逐重叠赌注,这削弱了 Core Automation 是明确品类赢家、而非多个类似定位进入者之一的判断 (CV017)。[CV012, CV015, CV018, CV020, CV026, CV050]

正向论点与反论点
立场论据改变判断的证据
正向论点创始团队履历和持续学习论点,指向在位者尚未解决的真实前沿研究缺口。经验证的基准测试或模型发布,证明该论点能在实践中跑通。
正向论点SSI、Thinking Machines、World Labs、Periodic Labs 等尚未有产品的同类实验室,成立不久后都以数十亿美元估值融资,说明该类别仍拿得到资金。同类新型实验室若出现降价轮或后续融资失败,会削弱这一比较。
正向论点即使尚未确认,据报道 $300M-$500M 后续融资兴趣表明,种子轮后仅数月投资人需求仍在。确认后续轮搁浅,或定价显著低于报道的 $4B 目标。
反论点公司没有提交 SEC Form D,也未披露产品、客户或收入,外部尽调无法核验估值。确认的 Form D 备案,或公开产品 / 基准测试披露。
反论点2026 年 AI 风险投资异常集中在少数前沿实验室,并出现 ARR 通胀和泡沫动态迹象,可能压缩可比倍数。证据显示 AI 领域 VC 集中度和倍数在正常化,而不是泡沫。
反论点至少两家资金充足、论点相邻的竞争者(Adaption Labs 和 AMI Labs)在押注重叠方向,提高了 Core Automation 不是品类赢家的风险。相对这些对手,出现清晰技术或商业差异化证据。

论据综合来自本章引用的声明和来源;不是单一一手来源表。

[CV017, CV026, CV034, CV009, CV029, CV050]

8.2 建议、置信度、风险与估值立场

本章建议是跟踪,不买入。推理链从 Core Automation 的产品发布前、收入前状态和未经确认的融资记录出发,经过其可比公司集中体现的关键人物和市场时点风险,最后落到「偏高到昂贵」的估值立场 (CV001, CV043)。这一判断的置信度为中等:四家独立媒体交叉印证了所报道的种子轮和后续融资数字,但没有任何报道获得监管备案确认,全文 EDGAR 搜索截至运行日也没有返回公司的 Form D 结果 (CV002, CV009)。风险评级为高。最接近的公开可比公司 Safe Superintelligence 是一家产品发布前的前沿实验室,其联合创始人在据报道出现收购兴趣后转投竞争实验室,说明即便估值最高的产品发布前 AI 实验室,也暴露于关键人物风险;Core Automation 这种小而履历集中的团队也会承担同类风险 (CV003, CV028)。估值立场为偏高到昂贵,因为报道中的估值在数周内从约 $1 billion 爬升到约 $4 billion,却没有披露产品、客户或收入锚点;尽管类似的快速重估模式在至少五家可比新型实验室中反复出现 (CV004, CV026, CV027)。按市场、证据、护城河、经济性、风险、估值和证据质量打分,Core Automation 最强的是市场规模和团队履历,最弱的是证据和经济性披露,这种失衡更符合跟踪而不是买入 (CV047)。实际含义是,在公司披露足以检验报道价格的产品、收入或备案证据之前,暂不投入新资本 (CV005)。[CV001, CV043, CV002, CV009, CV003, CV028]

建议摘要
维度评估理由
建议跟踪(暂不买入)公司尚未推出产品、尚无收入,融资也没有任何监管确认,现在承诺资金为时过早。
信心融资数字已有四家独立媒体相互佐证,但没有任何监管备案确认。
风险评级关键人、融资、执行和市场时点风险集中叠加,放大了未经验证的论点。
估值立场偏高至昂贵据报道,从种子轮到后续轮,估值数周内从约 $1B 升至约 $4B,但没有披露的产品或收入支撑。
决策含义里程碑披露前不投入新资金任何投资都须以尽调要求表列出的产品、收入、备案和股权结构证据为前置条件。

报告作者基于本章和前文审核证据形成的判断综合;不是公司披露数字。

[CV001, CV002, CV003, CV004, CV005]
FV001: 建议逻辑

建议取决于两个变量:披露缺口仍未解决,而入场价格偏高,还要面对高风险、 高不确定性的可比样本。

[CV001, CV004, CV009, CV026, CV043]
FV004: 投资 KPI

Core Automation 在市场规模和团队履历上得分较高,但产品证明、经济性和证据质量偏弱。

评分是面向 IC 讨论、基于公开证据集综合形成的 0-10 顺序判断,不是公司披露的评分卡。

[CV003, CV004, CV009, CV026, CV047]

8.3 融资背景、入场纪律与稀释悬置

媒体报道称,Core Automation 在 2026 年启动后数周内,以接近 $1 billion 估值完成约 $100 million 首轮融资,并正洽谈以约 $4 billion 为目标的 $300–500 million 后续融资 (CV006, CV007)。合在一起,这些数字意味着公司在数周内寻求约四倍加价;即便按 2026 年 AI 实验室融资标准,这个间隔也异常短 (CV011)。没有来源披露任何一项数字的收入、ARR 或使用量基础,因此无法计算标准倍数;也没有公开模型卡、基准或付费试点,能把价格锚定在团队履历和叙事之外的任何东西上 (CV008, CV052)。因此入场纪律应严格:这种从种子轮到后续融资的快速加价,通常会伴随更重的清算优先权和董事会保护给新投资人,但没有 股权结构表、条款书或优先权栈细节公开可用于为 Core Automation 确认这一点,使任何新资本的有效所有权和下行保护都不清楚 (CV054)。本章审阅的最新融资报道日期为 2026 年 5 月;截至 2026 年 7 月运行日,本轮没有发现任何来源报道后续融资已关闭、估值已修订或融资已放弃,因此融资叙事应在刷新前视为暂定 (CV053)。Nvidia 自己把对 OpenAI 的拟议 $100 billion 承诺重校准为 OpenAI 接近 IPO 时的 $30 billion 实际持股,显示即便最大战略投资人也在 2026 年收紧 AI 实验室资本承诺规模;这也可能收紧 Core Automation 这类规模更小、披露更少赌注的资金可得性 (CV039)。[CV006, CV007, CV011, CV008, CV052, CV054]

8.4 乐观、基准与悲观情景

乐观情景下,Core Automation 在约 12–18 个月内发布经过基准验证的持续学习或智能体研究能力,按目标附近的 $4 billion 完成报道中的后续融资,并像 World Labs 与 Autodesk 那样吸引一个战略算力或云合作伙伴,估值重估至 Thinking Machines Lab 和 World Labs 在类似阶段达到的 $8–12 billion 区间 (CV041, CV018, CV015)。基准情景下,公司继续靠团队履历和叙事融资,而没有公开产品;后续融资在报道目标附近或以下关闭,并附带更重的投资人保护,符合本章审阅的 2024–2026 年新型实验室集合中的主流模式 (CV042, CV017)。悲观情景下,随着 AI 板块风险资本进一步向最大实验室集中、投资人要求收入证据,后续融资停滞或以下降轮定价;AI 包装层 创业公司中已经可见的全行业清算是最清楚的书面先例,截至 2026 年中,至少 118 家公司倒闭、约 $49.9 billion 资本被毁,显示被炒作定价的 AI 公司可以多快失去价值;一份专门分析智能体 AI 估值的报告也发现,种子前 估值已经下降,且预计到 2027 年超过 40% 的智能体 AI 项目会被取消 (CV037, CV038, CV040)。合理估值区间为:悲观情景约 $0.5–1 billion,基准情景 $2–4 billion,乐观情景 $8–12 billion;这个区间锚定于情景逻辑和可比公司集合,而不是披露的财务模型 (CV048)。[CV041, CV018, CV015, CV042, CV017, CV037]

牛市、基准与熊市场景
场景假设估值 / 回报逻辑关键风险概率信号
牛市持续学习论点在 12-18 个月内产出有基准验证的能力;后续轮接近报道的 $4B 目标完成,并由战略算力或云合作伙伴锚定。估值可能向 Thinking Machines 和 World Labs 在类似阶段达到的 $8-12B 区间重估,为种子投资人带来数倍账面升值。技术论点无法与在位者内部自动化工具拉开差异;算力成本超过已募资金承受能力。低至中:同类实验室达到过类似估值,但尚无一家围绕 Core Automation 的精确论点推出商业产品。
基准公司继续靠团队履历和叙事融资,没有公开产品;后续轮在报道的 $4B 目标附近或以下完成,同时投资人保护更重。估值保持在 $2-4B 区间,有账面升值,但在最终推出产品或被收购前多年无法实现流动性。执行滑坡、关键研究员离职(如 SSI),或预期后续轮推迟或缩水。中等:与本章审视的 2024-2026 年新型实验室主流模式一致。
熊市后续轮搁浅或以降价轮定价;AI 领域 VC 集中度收紧到最大实验室周围,投资人要求收入证明;人才流向资金更充足的实验室。估值压缩至约 $1B 种子轮估值或更低;早期支持者面临重大减记,符合 2026 年次前沿 AI 初创在二级市场折价的记录。算力或人才成本结构迫使公司折价出售或关闭,符合 2026 年 AI 封装层和新型实验室失败模式。考虑到 2026 年智能体式 AI 和 AI 封装层初创公司已出现出清迹象,概率中至高。

假设和逻辑是报告作者基于可比样本和 2026 年市场证据形成的场景综合,不是公司指引。

[CV041, CV042, CV040, CV037, CV038]
FV002: 估值敏感性

Core Automation 的隐含加价倍数会大幅摆动,取决于产品证明能否落地、行业融资环境如何, 以及融资是否卡住。

这些数值是不同驱动情景下、相对于报道的 ~$1 billion 种子轮估值的示意性倍数;由于缺少财务披露, 并非来自已披露收入或 DCF 模型。

[CV011, CV035, CV040, CV048]
FV003: 估值 / 回报区间

Core Automation 的可行估值结果跨度约 20x,取决于披露和产品证明能否落地。

区间为示意性判断,锚定乐观 / 基准 / 悲观情景表和可比样本结果;由于缺少财务披露, 并非正式估值模型。

[CV041, CV042, CV040, CV048]

8.5 可比估值图谱

最相关的可比公司是 Safe Superintelligence。它在 2025 年以 $32 billion 估值融资约 $2 billion,尽管没有面向公众的产品;此前它在 2024 年以 $5 billion 估值融资 $1 billion,意味着一支约 20 人团队不到一年估值增长超过六倍 (CV012, CV013)。Thinking Machines Lab 的种子轮估值据报道为 $12 billion,后来通过一项新的数十亿美元交易加深了与 Google 的联系;Core Automation 还没有披露这种战略验证 (CV015, CV016)。World Labs 和 Periodic Labs 都说明产品发布前相邻实验室可以多快重估:World Labs 从 2024 年 $1 billion 种子轮估值,到 2026 年 2 月发布商业产品后瞄准据报道约 $5 billion;Periodic Labs 则在不到八个月内,从 $1.3 billion 走到据报道 $7.5 billion 的洽谈 (CV018, CV019, CV020, CV021)。Mistral AI 和 xAI 等后期实验室显示,即使公司已有发布产品、披露使用量和收入,重估仍会继续;但这也让它们对 Core Automation 这类披露前公司来说,直接可比性更弱 (CV022, CV023)。相比之下,产品层竞争对手 Hebbia 和 Glean 的定价基于披露的盈利收入或具名企业客户基础,而不是单靠履历,这凸显出 Core Automation 目前披露少得多 (CV024, CV025)。在整个集合里,产品发布前的前沿研究实验室估值主要由创始人履历、算力合作和叙事动能驱动,而不是由披露指标驱动;这让 Core Automation 报道中的加价方向与同类公司一致,却无法独立对照验证 (CV026, CV027, CV051, CV049)。[CV012, CV013, CV015, CV016, CV018, CV019]

可比估值表
可比对象指标估值 / 状态与 Core Automation 的相关性局限
Core Automation报道的 $100M 种子轮;$300-500M 后续轮讨论据报道种子轮估值约 $1B;后续轮目标约 $4B,均未经 SEC 确认标的公司两个数字均无独立确认;没有收入或产品基础
Safe Superintelligence (SSI)累计融资约 $3B$32B(2026),高于 $5B 种子轮(2024)最接近的类比:没有公开产品,估值由创始人履历和算力合作驱动不到一年约 6x 升值,且没有已发布产品;近期有联合创始人离职
Thinking Machines Lab种子轮$12B(2025 年种子轮)创始人履历驱动的新型实验室可比对象;后来加深与 Google 的关系背后有后来扩大的 Google 战略交易支撑,而 Core Automation 尚未披露类似安排
AMI Labs融资 $1.03B除融资规模外未披露论点相邻的直接竞争者(世界模型)估值倍数未公开
World Labs$230M 种子轮(2024),随后 $1B 轮(2026 年 2 月)约 $1B(2024)升至报道的约 $5B 目标(2026)可比轨迹:约 18 个月内快速重估至数十亿美元已有商业产品 Marble 和具名战略投资人,不同于 Core Automation
Periodic Labs$300M 种子轮(2025 年 9 月),随后洽谈 $500M(2026 年 5 月)$1.3B 升至报道的 $7.5B 目标面向科学 AI 的新型实验室可比对象,8 个月内约 6x 升值截至来源日期,交易仍在洽谈,尚未完成
Mistral AI迄今融资约 $4B€11.7B(2025 年 9 月)升至报道约 €20B 目标(2026)后期可比对象,显示已有产品和收入的实验室仍可继续重估已有产品和收入;不是干净的无产品阶段可比对象
xAI$20B Series E 轮(2026 年 1 月)~$230-250B说明一个披露使用量(600M MAU)的前沿实验室,相对披露前的新型实验室,规模能走多远规模和披露使用量使其很难直接类比尚未有产品的公司
Hebbia融资 $130M$13M 盈利性收入支撑 $700M 估值对照案例:产品层竞争者按已披露收入定价,而非仅靠履历不同商业模式(企业软件 vs. 研究实验室)

估值均为媒体报道,未经审计或监管确认;进行中轮次(Core Automation、Periodic Labs、Mistral)的数字反映报道目标,而非已完成条款。

[CV006, CV007, CV012, CV015, CV017, CV018]

8.6 退出准备度、论点失效触发器与最终尽调要求

OpenAI 和 Anthropic 同步走向 2026 年末 IPO,估值各自接近 $1 trillion,说明前沿实验室退出窗口正在打开;但这个窗口首先为规模最大、披露最多的实验室打开,Core Automation 自身没有披露近期退出路径 (CV046)。最关键的论点失效触发器包括:后续融资停滞或低于目标,出现 SSI 曾经历的关键研究员离职,十二个月过去仍没有产品或基准披露,或者监管备案与媒体报道条款相矛盾 (CV044, CV028, CV009)。如果更广泛的 AI 实验室融资收缩波及两个或更多可比新型实验室,将成为专属于 Core Automation 所在层级的明确悲观情景确认 (CV037)。在任何资本承诺可以被负责任承销前,优先级最高的未满足要求是:两轮融资条款的监管确认、任何产品或基准证据、客户或设计伙伴证据、股权结构表与优先权细节、已披露烧钱和算力承诺,以及创始人和关键研究员的留任条款;截至运行日,这些都不在公开记录中 (CV045, CV052, CV054)。本次审阅期间无法访问覆盖 Core Automation 或其最接近可比公司的持牌二级市场或私募股权定价数据,因此仍无法判断 2026 年广泛记录的次前沿 AI 初创公司二级市场折价是否也适用于这里 (CV055)。[CV046, CV044, CV028, CV009, CV037, CV045]

论点破裂与终止触发项
触发项阈值对论点的传导行动含义
确认降价轮或后续轮搁浅后续轮以低于约 $2B 完成,或报道洽谈后 6 个月内未完成表明投资人需求弱于媒体报道暗示将估值立场下调为昂贵;暂停进一步尽调
关键研究员离职创始人或具名前三研究员转投竞争者或对手实验室呼应 SSI 联合创始人离职,释放团队风险信号重新评估团队履历溢价;视为红旗
12 个月内没有产品或基准测试披露到 2027 年中仍没有公开模型卡、基准测试或付费试点确认公司未向可交付产品收敛如无新证据,建议回避
SEC 或监管备案披露实质不同条款Form D 或其他备案显示估值、结构或投资人条款与媒体报道不一致削弱本章所有媒体来源估值主张的可靠性基于备案条款重新承做投资测算,而非媒体报道
更广泛 AI 实验室融资收缩两家或更多可比新型实验室遭遇降价轮或关闭确认 2026 年来源记录的全行业清算已传导至新型实验室层将其视为 Core Automation 的熊市触发项

阈值是报告作者用于监控的判断,不是公司披露的契约条款。

[CV044, CV028, CV009, CV037]
最终尽调要求
主题缺失证据为什么重要尽调路径
融资监管确认种子轮和后续轮的 Form D 或最终轮次文件~$1B 和 ~$4B 数字只来自媒体报道直接向公司或其律师索取股权结构表和轮次文件
产品或基准测试证据任何模型卡、基准测试结果或可运行演示截至运行日期,没有公开产品技术证据要求安排技术尽调会议,提供可运行演示或基准测试
客户或设计伙伴证据任何具名试点、设计伙伴或意向书没有客户证据支撑上市路径论点要求与任何已披露设计伙伴进行推荐访谈
股权结构表和优先权条款清算优先权栈、董事会组成、反稀释条款决定新资金的下行保护和有效所有权索取报道中后续轮的股权结构表和条款清单
烧钱速度和算力承诺披露的现金消耗、跑道和算力合同判断报道融资规模是否符合前沿实验室资金密集度常模索取财务报表或数据室
人才留存和关键人风险具名研究员的股权归属安排和留存条款SSI 联合创始人离职表明,履历驱动估值暴露于关键人风险索取创始人和顶尖研究员的留存 / 归属条款

尽调要求按最能改变建议或可接受入场价格的因素排序,而非按收集难易度排序。

[CV045, CV052, CV054, CV028]

8.7 附录

免责声明

本报告是 AI 辅助生成的尽调材料,基于截至 2026-07-05 的公开信息。 非上市公司的经营指标、融资条款和客户证据可能与公开可见信息不同。 本报告仅供研究使用,不构成投资建议。

证据索引

结论
编号陈述可信度来源
CO001 Core Automation's official website describes its mission as building "the world's most automated AI lab," pursuing new learning algorithms that supersede large-scale pretraining and reinforcement learning and architectures designed to scale better than transformers. SO001, SO017
CO002 Core Automation is headquartered in San Francisco, California, per its official X account bio and independent company-data coverage. SO005, SO021
CO003 Core Automation was founded in March 2026 by Jerry Tworek, according to The Information as relayed by Techmeme and Intellectia. SO019, SO025
CO004 BigGo's April 24, 2026 reporting states Core Automation had already initiated fundraising negotiations "shortly after its founding in late January" 2026, conflicting with the March 2026 founding date reported by The Information via Techmeme and Intellectia. SO023
CO005 Jerry Tworek identifies himself as CEO and co-founder of Core Automation on his X profile, corroborated by third-party reporting. SO006, SO024
CO006 Jerry Tworek was OpenAI's vice president of research and led development of the o1 and o3 reasoning models before departing after nearly seven years. SO026, SO029
CO007 Tworek joined OpenAI in 2019 when the company had roughly 30 employees, and was involved in GPT-4 post-training, the 2025 deployment of GPT-5, and the Codex code-generation model. SO023, SO022
CO008 Tworek informed OpenAI colleagues of his intent to leave on January 5, 2026, and his departure became public around January 7-8, 2026. SO027, SO028
CO009 In his farewell note, Tworek said he was "leaving to try and explore types of research that are hard to do at OpenAI," and separately told colleagues he believed foundational deep-learning research "is done" inside OpenAI. SO029
CO010 Core Automation publicly launched via its first X post on April 21, 2026, stating it is "building the most automated AI lab in the world." SO023, SO031
CO011 Mark Saroufim describes himself as a co-founder of Core Automation on his X profile and previously worked on PyTorch and the GPU MODE community. SO007, SO024
CO012 Rohan Anil, a researcher who previously worked at Google DeepMind and Anthropic, said Jerry Tworek "nerdsniped" him into co-founding Core Automation after leaving Anthropic in early 2026. SO023, SO030
CO013 Anmol Gulati, a Google DeepMind research scientist who worked on Gemini, publicly confirmed joining Core Automation, citing skepticism that scaling models and static deployment alone will reach the "final goal." SO023, SO030
CO014 Joanne Jang, who served as an OpenAI general manager involved in GPT-4o development from December 2021 to April 2026, joined Core Automation and describes herself on X as "trying to automate my work @coreautoai." SO008, SO023
CO015 Other reported team members include former Google DeepMind researchers Ehsan Amid and Avery Lamp, former OpenAI head of people Julia Villagra, and Sai Surya Duvvuri, a former Google and Meta research intern. SO023
CO016 Public X handles retrieved for several reported non-founder team members (Rohan Anil, Anmol Gulati, Ehsan Amid, Sai Surya Duvvuri) either show zero posts, do not match the biography claimed in press coverage, or are private/unavailable, preventing direct primary-source verification of their affiliation. SO011, SO012, SO013, SO014, SO015
CO017 Core Automation's public website team page (coreauto.com/team) returned a not-found error during evidence collection on the run date, indicating the company does not currently maintain a live public team roster page. SO004
CO018 Jerry Tworek's personal homepage still describes him as "a research lead at OpenAI" and does not mention Core Automation, indicating the page has not been updated since his departure and the founding of the new company. SO016
CO019 No public source reviewed discloses whether Core Automation has a formally constituted board of directors or who holds any board seats.
CO020 Core Automation's public credibility and fundraising narrative are structurally concentrated in Jerry Tworek's individual track record as principal architect of OpenAI's reasoning-model program, creating material key-person dependence for a pre-product company. SO022, SO021
CO021 Core Automation raised $100 million in an initial round at approximately a $1 billion valuation, as reported by Sacra and corroborated by Techmeme and Intellectia citing The Information. SO021, SO019
CO022 As of May 7-8, 2026, Core Automation was reportedly seeking $300 million to $500 million in new capital at a target valuation of approximately $4 billion, a roughly fourfold step-up from its initial $1 billion valuation in under three months. SO020, SO025
CO023 AI CERTs explicitly reported that Bloomberg and Reuters had not corroborated Core Automation's reported fundraising figures as of its May 2026 coverage. SO018
CO024 Sacra reports that Core Automation's initial $100 million round included participation from Nvidia, Spark Capital, and Accel, with no publicly disclosed lead investor. SO021
CO025 No source reviewed independently confirms Core Automation's initial-round investor identities beyond Sacra's proprietary reporting, and the broader financial terms remain uncorroborated by Bloomberg or Reuters per AI CERTs. SO021, SO018
CO026 No public source reviewed discloses any secondary share sale or debt/credit financing involving Core Automation as of the run date.
CO027 Core Automation's reported fundraising pace compares to peer neo-labs Thinking Machines Lab ($2B seed at a $10B valuation, closed mid-2025), Humans& ($480M seed at a $4.48B valuation, announced January 2026), and Safe Superintelligence ($1B series-seed in 2024), per AI CERTs. SO018
CO028 BigGo estimated Core Automation had "roughly a dozen public members" as of its April 24, 2026 report, versus thousands of researchers at OpenAI and Google DeepMind. SO023
CO029 As of May 2026, Core Automation has no public API, pricing page, signup flow, or disclosed commercial product, per Sacra's company profile. SO021
CO030 Core Automation's official website navigation, as of the run date, offers only Home, Blog, X/Twitter, Contact, and Join Us -- with no product, pricing, or signup pages -- consistent with a pre-product research lab. SO001, SO002
CO031 Sacra characterizes Core Automation's business model as "lab-first": raising substantial capital to build proprietary learning systems, using them to automate the lab's own research, and later commercializing the resulting capabilities for external B2B customers. SO021
CO032 Core Automation's primary research project, internally named "Ceres," is described as a single model capable of continual learning in production, targeting roughly 100x less training data than current state-of-the-art models while enabling weight updates during deployment. SO018, SO023
CO033 Reported details of Ceres include revisiting optimization methods "up to and including gradient descent" and biologically inspired synaptic-consolidation techniques intended to counter catastrophic forgetting, per The Information as relayed by AI CERTs. SO018
CO034 On May 28, 2026, Core Automation published its first detailed public technical writing, a blog post by co-founder Mark Saroufim titled "When AI Starts Writing Systems Code," discussing systems-code automation for AI research. SO003, SO002
CO035 Tworek's OpenAI exit was one of roughly a dozen senior departures from OpenAI in the prior year, following the 2025 exits of CTO Mira Murati, chief research officer Bob McGrew, and VP of research Barret Zoph. SO029
CO036 No lawsuit, regulatory action, sanction, or formal governance controversy directly naming Core Automation or Jerry Tworek was identified in the sources reviewed as of the run date.
CO037 Sacra identifies OpenAI, Anthropic, and Google DeepMind as Core Automation's most direct strategic threats, alongside smaller thesis-aligned competitors Sakana AI and Reflection AI, and enterprise research-agent products Hebbia, Manus, Glean, and FutureHouse. SO021
CO038 Public disclosure about Core Automation's product roadmap, cap table, and headcount remains substantially thinner than typical late-stage private disclosures, though comparable peer neo-labs (Thinking Machines Lab, Safe Superintelligence) followed a similar pre-product, high-valuation disclosure pattern before shipping products. SO021, SO018
CO039 SiliconReport calculated that Core Automation's reported valuation step-up from $1 billion to $4 billion would add roughly $3 billion in paper value in under three months, an unusually rapid re-rating for a company without disclosed revenue. SO020
CO040 Confirming Core Automation's actual capitalization table, lead-investor identity, and round-close terms requires direct diligence access to company counsel or data-room documents; no public substitute exists as of the run date. SO021
CO041 As of the run date, Core Automation is characterized as a private, pre-revenue company with no formally named funding-round stage (e.g., "Series A") disclosed in any source reviewed; coverage instead describes discrete "initial" and "follow-on" raises. SO021, SO022
CO042 No source reviewed reports that any of Core Automation's founding-team members have left the company since its April 2026 public launch.
CO043 This chapter's snapshot KPI table records at least four cover metrics (exact headcount, initial-round lead investor, exact founding date, and Bloomberg/Reuters corroboration of financial figures) as unsupported by primary sources, each paired with an explicit diligence gap rather than a fabricated figure. SO021, SO018
CM001 Analyst coverage of AI agents treats standalone agent-software spend, embedded agentic-capability spend, and total AI infrastructure/software/services spend as three distinct, non-nested measurement scopes that differ by roughly 25x at the same point in time. SM004, SM022
CM002 Gartner forecasts total worldwide AI spending (infrastructure, software, and services) will reach $2.59 trillion in 2026, a 47% increase year-over-year. SM004, SM022
CM003 Gartner's 2026 estimate of enterprise 'agentic AI' capability-embedded spending is $201.9 billion, roughly 7.8% of its total 2026 AI spending figure. SM004, SM022, SM023
CM004 Four analyst firms (Fortune Business Insights, Precedence Research, MarketsandMarkets, and Deloitte's TMT Predictions) size the standalone AI-agent software market at $7.0-8.5 billion in 2025-2026, but their forecasts diverge by nearly 30x by their respective terminal years (2030-2034). SM022
CM005 Axis Intelligence's cross-firm aggregation puts the standalone global AI-agent market at $7.9-8.0 billion in 2025, rising to $10.9-11.8 billion in 2026, a 44-47% compound annual growth rate through 2030. SM023
CM006 Grand View Research estimates the global robotic process automation (RPA) market at $4.68 billion in 2025, reaching $35.84 billion by 2033 at a 29.0% CAGR. SM021
CM007 Precedence Research estimates the same nominal RPA market at $28.31 billion in 2025 and $35.27 billion in 2026, reaching $247.34 billion by 2035 at a 24.2% CAGR -- roughly six times Grand View Research's 2025 baseline for an ostensibly similar category. SM025
CM008 IDC's FutureScape 2026 research projects that 45% of organizations will orchestrate AI agents 'at scale' by 2030, an adoption metric rather than a dollar-denominated market size. SM006
CM009 No analyst report reviewed publishes a distinct dollar-denominated market size for frontier-AI-lab research-automation tooling or continual-learning model research specifically; every available estimate covers a broader enterprise-agent or AI-infrastructure category instead.
CM010 Continual learning -- a model's ability to keep acquiring new knowledge and skills without forgetting prior ones -- remains an unsolved, actively researched problem in large language models as of 2026, according to Google Research. SM008
CM011 Google Research's 'Nested Learning' paradigm, embodied in its 'Hope' architecture, reframes models as nested, self-modifying optimization problems intended to retain long-horizon memory without overwriting previously learned knowledge. SM008
CM012 Mamba-3, a 2026 state-space-model architecture, is presented as a post-transformer design that replaces the Transformer's quadratic-compute attention mechanism with linear-time sequence processing while matching or beating Transformer baselines on perplexity at roughly half the inference cost. SM011, SM028
CM013 Mamba-3 was accepted as an oral presentation at ICLR 2026, one signal of continued peer-reviewed research momentum behind non-Transformer sequence architectures. SM028
CM014 Post-transformer and continual-learning research remains at the architecture/paper stage rather than a validated, production-scale replacement for Transformer-based frontier models as of mid-2026; no source reviewed reports a shipped frontier-scale commercial model built on these alternatives. SM008, SM011, SM028
CM015 Sakana AI's 'AI Scientist' system, an agent that autonomously formulates hypotheses, runs experiments, and authors machine-learning research papers, had a paper describing its methodology published in Nature in March 2026. SM009
CM016 An earlier AI Scientist-v2 paper produced the first entirely AI-generated manuscript to pass a genuine human peer-review process at a workshop track, using an agentic tree-search method that removed reliance on human-authored code templates. SM010
CM017 Sakana AI has open-sourced both AI Scientist versions on GitHub, lowering the barrier for other teams to replicate or extend automated-research-agent techniques. SM026
CM018 Anthropic launched 'Claude Science,' a research-automation product line explicitly positioned around workflow integration for scientists rather than a new underlying model. SM012
CM019 Anthropic reported that more than 80% of the code merged into its own production codebase in May 2026 was authored by its Claude model rather than human engineers, alongside an 8x increase in code shipped per engineer versus its 2021-2025 baseline. SM013
CM020 Anthropic's self-reported code-automation figures are being described in press coverage as an early, unaudited signal of 'recursive self-improvement' inside a frontier lab, the same category of research automation Core Automation says it is pursuing. SM013
CM021 Stanford HAI's 2026 AI Index reports that AI agents completed real-world computer tasks (OSWorld benchmark) at roughly 66% success as of March 2026, up from about 12% roughly 18 months earlier, while still failing about one-third of attempts. SM001, SM002
CM022 Organizational AI adoption reached 88% in the 2026 AI Index, with generative AI adoption reaching 53% within three years of ChatGPT's release, both cited as outpacing the historical diffusion rates of the PC and the internet. SM001, SM027
CM023 Deloitte's 2026 State of AI in the Enterprise survey found about 23% of organizations using agentic AI at least moderately, with 74% planning to implement it within two years, even though only 21% report a mature governance model for autonomous agents. SM003, SM023
CM024 Axis Intelligence's cross-referenced 'AI Agents Deployment Gap Index' finds 93% of IT leaders plan to introduce autonomous agents within two years, but only 23% have scaled deployment in even one business function -- a 70-percentage-point gap between stated intent and production reality as of Q2 2026. SM023
CM025 Gartner predicts 40% of enterprise applications will embed task-specific AI agents by the end of 2026, up from under 5% in 2025, and separately predicts that 40% of agentic AI projects will be canceled by the end of 2027 over cost, ROI, and risk-control concerns. SM005, SM023
CM026 MIT's 2025 'GenAI Divide' study of 300-plus enterprise generative-AI deployments found 95% of organizations captured zero measurable P&L return, with only about 5% of integrated pilots extracting significant value. SM007, SM024
CM027 MIT's study attributes generative-AI pilot failure primarily to shallow, siloed tool deployment and a lack of workflow-level integration and organizational learning, rather than to model quality or regulation. SM007
CM028 Q1 2026 global venture investment reached roughly $300 billion across about 6,000 startups, an all-time quarterly high driven disproportionately by a handful of massive AI funding rounds. SM018
CM029 AI startups captured roughly 80-81% of all global venture capital deployed in Q1 2026, up from about 55% a year earlier, concentrating capital in a small number of frontier labs. SM018
CM030 Thinking Machines Lab, a 2025-founded frontier AI lab, closed a $2 billion seed round at a $12 billion valuation in mid-2025 and subsequently secured a multi-billion-dollar Google Cloud compute partnership plus a roughly one-gigawatt Nvidia hardware commitment. SM015, SM016
CM031 Reported funding scale among 2025-2026 frontier AI labs varies enormously, from Safe Superintelligence's roughly $1 billion 2024 raise up to OpenAI, Anthropic, and xAI rounds valued in the hundreds of billions of dollars, illustrating the capital intensity a new entrant like Core Automation is competing against. SM029
CM032 GPU compute for frontier-scale model training faces structural constraints in 2026, with H100/H200 lead times of 36-52 weeks driven by TSMC CoWoS packaging capacity and HBM memory supply bottlenecks, pushing well-capitalized labs toward multi-year reserved-capacity contracts. SM017
CM033 Compute scarcity and reserved-capacity contracting favor hyperscaler-backed frontier labs, since buyers without 2025-era procurement commitments face queued training jobs and rising costs -- a structural disadvantage for a smaller, newer entrant. SM017
CM034 Investor Michael Burry has publicly compared the 2025-2026 AI investment cycle to the 1999-2000 dot-com bubble, arguing hyperscaler depreciation accounting understates the true cost of AI hardware and predicting a 'Panic of 2026 or 2027.' SM014
CM035 Burry's critique centers on an accounting argument that spreading GPU depreciation over long useful-life assumptions could mask tens of billions of dollars in real costs at major AI infrastructure spenders, a critique aimed at hyperscalers rather than at early-stage research labs directly. SM014
CM036 McKinsey's November 2025 State of AI survey, as aggregated by Axis Intelligence, found that while 62% of organizations experiment with AI agents, only 23% scale them in at least one business function and fewer than 10% scale across multiple functions. SM023
CM037 The World Economic Forum describes 'physical AI' -- robotic systems capable of perception, reasoning, and autonomous action -- as an emerging complement to rule-based industrial automation, driven by labor shortages and supply-chain volatility. SM019
CM038 BCG describes physical AI as changing automation economics by letting manufacturers retrain existing hardware with new 'brains' rather than replacing production lines, potentially lowering the capital intensity of adopting autonomous systems on the factory floor. SM020
CM039 Industrial/physical AI automation is a structurally adjacent but distinct market from Core Automation's cognitive-research-automation thesis: both share an 'autonomous work' framing, but physical AI requires hardware/robotics investment that a software-and-model-only thesis does not. SM019, SM020
CM040 Enterprise buyers for agentic/knowledge-work automation are typically CIOs, COOs, or functional VPs funding pilots from existing IT or operations budgets rather than a newly created 'AI agent' budget line. SM003, SM023
CM041 The MuleSoft/Deloitte Digital Connectivity Benchmark found 93% of IT leaders plan to introduce autonomous agents within a two-year window, indicating broad buyer-side intent even where production deployment lags. SM023
CM042 Scientific and research-automation buyers (university PIs, biotech/pharma R&D heads) are already being courted directly by well-capitalized incumbents such as Anthropic (Claude Science) and Sakana AI (AI Scientist), rather than this segment being open white space. SM012, SM009
CM043 Trust, governance, and evaluation difficulty -- not raw model capability -- are the most commonly cited blockers to scaling agentic AI in enterprise settings across Deloitte, Gartner, and Axis Intelligence reporting. SM003, SM005, SM023
CM044 The mechanism by which continual learning research could matter economically is reducing the training-data and compute cost of keeping frontier models current, which -- if achieved -- would directly offset the GPU/compute scarcity documented in 2026 supply-chain reporting. SM008, SM017
CM045 No enterprise deployment of a frontier-lab-style, continual-learning-based research-automation product resembling Core Automation's stated thesis was identified in any source reviewed as of the run date; comparable proof points (Sakana AI Scientist, Anthropic Claude Science) come from better-capitalized incumbents building on standard Transformer-based models, not on a validated continual-learning replacement.
CM046 Deloitte's 2026 enterprise AI survey identifies research and development as one of the top enterprise use cases named for agentic AI, alongside customer support, supply chain management, and cybersecurity. SM003
CM047 Gartner's worldwide AI-spending forecast was revised upward by roughly $500 billion within about eight months (from just above $2 trillion to $2.52-2.59 trillion for 2026), illustrating how quickly headline AI market estimates move and how little precision they can offer for sizing a narrow pre-product sub-segment. SM022
CM048 Analyst estimates of the 'agentic AI' category span at least a 25x range at the same point in time depending on whether embedded capability spend or standalone agent-vendor revenue is counted, meaning any single-point TAM claim for Core Automation's segment would overstate precision the underlying data does not support. SM004, SM022, SM023
CP001 Anthropic closed a $65 billion Series H round in May 2026 at a $965 billion post-money valuation, co-led by Altimeter Capital, Dragoneer, Greenoaks, and Sequoia Capital. SP001, SP002
CP002 Anthropic's run-rate revenue crossed $47 billion by May 2026, up from its prior Series G round closed in February 2026. SP001, SP002
CP003 Anthropic's Series H round included committed multi-cloud compute capacity -- five gigawatts from Amazon, five gigawatts of next-generation TPU capacity from Google/Broadcom, and GPU capacity from SpaceX's Colossus data centers -- making Claude the first frontier model available on AWS, Google Cloud, and Azure simultaneously. SP001, SP002
CP004 Anthropic's $965 billion post-money valuation reported in May 2026 exceeded OpenAI's most recently reported $852 billion valuation, per TechCrunch. SP001, SP002
CP005 Yann LeCun's AMI Labs raised a $1.03 billion seed round in March 2026 at a $3.5 billion pre-money valuation to build 'world models' based on LeCun's Joint Embedding Predictive Architecture, a distinct post-LLM technical bet from Core Automation's continual-learning thesis. SP003
CP006 AMI Labs' investor group includes Nvidia, Samsung, and Eric Schmidt, overlapping with the investor pool reported across other 2025-2026 frontier-adjacent funding rounds. SP003
CP007 AMI Labs CEO Alexandre LeBrun stated the company does not plan to generate revenue in the near term and expects it could take years for world models to reach commercial application, prioritizing published, open research instead. SP003
CP008 Google DeepMind's Gemini Deep Think mode powers an internal research agent (codenamed 'Aletheia') that autonomously generates, verifies, and revises solutions to research-level mathematics, physics, and computer science problems, with results submitted to peer-reviewed venues as of February 2026. SP004
CP009 Gemini Deep Think progressed from International Mathematics Olympiad gold-medal-standard performance in 2025 to scoring up to 90% on the IMO-ProofBench Advanced benchmark by February 2026, per Google DeepMind. SP004
CP010 Google DeepMind's Gemini Deep Think research-agent work directly overlaps with Core Automation's stated ambition to automate parts of the AI research process, except DeepMind's version is already in production with published outputs. SP004
CP011 Sakana AI raised a $135 million Series B in November 2025 at a $2.65 billion post-money valuation, bringing its total disclosed funding to roughly $379 million. SP005
CP012 Sakana AI, founded in 2023 by former Google researchers David Ha, Llion Jones, and Ren Ito, focuses on efficient, smaller models optimized for the Japanese language, culture, and enterprise sectors (finance, industrial, government) rather than frontier-scale general models. SP005
CP013 Sakana AI's own site describes its mission as 'Building Frontier AI in Japan,' positioning it as a geography- and efficiency-focused alternative to U.S. frontier labs rather than a continual-learning research-automation peer. SP006
CP014 FutureHouse is a non-profit lab building AI agents to automate scientific discovery in biology and other complex sciences, pairing early-career researchers with AI tools and academic co-advisors through its AI-for-Science Postdoctoral Fellowship. SP007
CP015 FutureHouse published 'Robin,' a multi-agent system demonstrating end-to-end scientific discovery in biology, in May 2026, following earlier releases including DISCO (enzyme design) and OXtal (molecular crystal structure prediction). SP008
CP016 FutureHouse operates as a non-profit lab, in contrast to Core Automation's for-profit, venture-funded structure, even though both target automating scientific or research work. SP007
CP017 Reflection AI, founded in March 2024 by former Google DeepMind researchers Misha Laskin and Ioannis Antonoglou, raised a $2 billion round led by Nvidia at an $8 billion valuation in late 2025, up from a $545 million valuation seven months earlier. SP009
CP018 By March 2026, Reflection AI was reportedly in talks to raise $2.5 billion at a $25 billion pre-money valuation, with JPMorgan considering participation through its Security and Resiliency Initiative. SP010
CP019 Reflection AI's strategy centers on open-weight frontier models pitched as a Western/U.S. alternative to closed labs (OpenAI, Anthropic) and Chinese models (DeepSeek), monetizing via enterprise and government deployments rather than a direct consumer product. SP009
CP020 As of the run date, Reflection AI's official site describes its mission simply as building 'open models that let anyone control their intelligence,' with no public product, pricing, or model release disclosed. SP011
CP021 Glean raised a $150 million Series F round in mid-2025 at a $7.2 billion valuation led by Wellington Management, and by mid-2026 reported roughly $300 million in estimated ARR with over 850 employees and a Glean Agents platform reported to power more than 100 million agent actions annually. SP012, SP013
CP022 Glean's enterprise customers include Fortune 500 organizations, and its platform integrates with more than 100 SaaS applications while enforcing per-user data permissions, directly competing for the enterprise research/knowledge-work automation budget Core Automation would need to enter if it commercializes. SP012
CP023 Hebbia raised a $130 million Series B in mid-2024 at a $700 million valuation (about 54x its reported $13 million ARR), with backers including Andreessen Horowitz, Index Ventures, Google Ventures, and Peter Thiel. SP014
CP024 Hebbia's product was used by roughly 30% of asset managers as of its 2024 disclosure, and the company's own site reports over $30 trillion in client AUM and roughly 200,000 average prompts processed per day, indicating an established, revenue-generating enterprise research-automation position well ahead of Core Automation's pre-product stage. SP014, SP015
CP025 Manus, an AI agent platform originally developed by Singapore-based Butterfly Effect, was the subject of a roughly $2 billion Meta acquisition announced in December 2025, and Manus's own website states as of the run date that 'Manus is now part of Meta.' SP016
CP026 In April 2026, China's National Development and Reform Commission ordered Meta to unwind its Manus acquisition, and by June 2026 Meta had begun an operational unwind (data firewalls, access revocation, internal-use prohibition) while Manus's founders sought roughly $1 billion to buy the company back at its original valuation. SP017
CP027 Manus's own site (still stating 'Manus is now part of Meta') conflicts with contemporaneous reporting that the Meta acquisition is being unwound under Chinese regulatory order, making Manus's actual current ownership status an unresolved discrepancy as of the run date. SP016, SP017
CP028 More than 40 'NeoLabs' -- research-led AI startups founded by alumni of frontier labs -- raised a combined $40+ billion in the three years before 2026, with billion-dollar first rounds common, per Radical Ventures' 2026 analysis. SP018
CP029 Radical Ventures explicitly classifies Core Automation, alongside Adaption Labs, as pursuing the 'continual learning' paradigm among NeoLabs, distinguishing it from world-model labs (AMI Labs, World Labs, Decart), reinforcement-learning labs (Reflection AI, Ineffable Intelligence), and diffusion or energy-based-model labs. SP018
CP030 Radical Ventures identifies compute access, not capital, as the binding constraint for NeoLabs, noting that strategic compute partnerships (hyperscaler commitments, Nvidia allocation agreements) have become standard cap-table features and that Nvidia is the single most active strategic investor across the model-provider landscape. SP018
CP031 Radical Ventures' bear-case scenario for NeoLabs is a 'wind-down / fire sale / zombie' outcome in which talent leaves for incumbents offering 10x compensation before a team can build a sustainable business. SP018
CP032 Radical Ventures flags distillation and open-weights commoditization as a key risk category for NeoLabs generally, meaning capabilities can commoditize faster than a research-stage team can build a durable commercial moat. SP018
CP033 Industry estimates cited by TechSpot put the global pool of people capable of building frontier AI models at roughly 2,000 as of 2026, with Meta offering signing bonuses as high as $100 million and senior AI research compensation packages now ranging $3 million to $10 million-plus annually. SP019
CP034 OpenAI's Chief Research Officer publicly described losing researchers to Meta's recruiting push as feeling like 'someone has broken into our home,' illustrating how intense the competition for the same narrow researcher pool Core Automation has already recruited from remains. SP019
CP035 Thinking Machines Lab (founded by former OpenAI CTO Mira Murati) shipped its first commercial product, Tinker -- a managed API for fine-tuning open-weight language models using LoRA -- in October 2025, roughly six months before Core Automation had disclosed any product surface. SP020
CP036 Tinker was adopted in private beta by research groups at Princeton, Stanford, Berkeley, and Redwood Research within its first weeks, evidencing early developer/research-market traction Core Automation, as pre-product, cannot yet claim. SP020
CP037 Adaption Labs, founded by former Cohere executives Sara Hooker and Sudip Roy, raised a $50 million seed round led by Emergence Capital in early 2026 to commercialize 'gradient-free' continual-learning technology that lets deployed models adapt without full retraining -- a technical thesis that overlaps directly with Core Automation's continual-learning bet. SP021
CP038 Adaption Labs CEO Sara Hooker argues the 'frozen model' paradigm of retraining from scratch whenever facts change is 'economically unsustainable and scientifically inelegant,' echoing the same critique of static pretraining that Core Automation's founders have made, indicating at least two well-funded teams are pursuing near-identical theses independently. SP021
CP039 Leaked, subsequently audited financials reported by Forbes show OpenAI posted a $20.9 billion operating loss on $13.07 billion of 2025 revenue (total costs near $34 billion), with net loss widening to $38.5 billion after a one-time charge tied to its for-profit conversion. SP022
CP040 Palantir CEO Alex Karp publicly called the AI token business model 'insane' in July 2026, arguing enterprise customers gain little value from frontier-model subscriptions while surrendering competitive data advantage to model providers. SP022
CP041 Yann LeCun (founder of AMI Labs) warned in June 2026 that frontier labs including OpenAI and Anthropic risk a 'big bubble explosion' unless they cut costs or raise prices, because current usage is subsidized by investor capital rather than paying customers. SP023
CP042 LeCun separately called Elon Musk's xAI 'kind of a failure' due to co-founder departures, illustrating that even well-capitalized, frontier-scale entrants can struggle to sustain a research team and roadmap. SP023
CP043 Independent commentary on Core Automation's launch is explicitly skeptical, noting the industry has 'heard this story about automated discovery a dozen times before' and that autonomous research systems risk 'overfitting its own noise,' concluding the bet remains an unproven gamble rather than a demonstrated capability. SP024
CP044 OpenAI released GPT-5.5 in April 2026, explicitly marketed for agentic coding, computer use, 'knowledge work,' and 'early scientific research,' positioning OpenAI's shipped, revenue-generating product line directly against the research-automation and knowledge-work use cases Core Automation's thesis targets. SP025
CP045 OpenAI's official research page lists shipped outputs such as 'GeneBench-Pro' (a genomics/biology AI benchmark, June 2026) and a next-generation model preview, showing OpenAI continuing to publish applied research at a cadence Core Automation, as a pre-product lab, has not yet matched. SP026
CP046 Unlike Core Automation, which as of the run date has no disclosed pricing or product surface, Glean, Hebbia, and Manus each monetize through named commercial pricing motions (enterprise seat/agent-action licensing for Glean, AUM/enterprise contract pricing for Hebbia, and consumer/business subscription plus API/team plans for Manus), giving each a live go-to-market engine Core Automation would have to build from scratch. SP012, SP014, SP015, SP016
CP047 Anthropic and OpenAI each operate multi-gigawatt, multi-vendor compute supply agreements (Anthropic's AWS/Google/SpaceX commitments; OpenAI's roughly $300 billion Oracle Stargate contract), while Core Automation's only disclosed compute-relevant relationship is Nvidia's reported participation as an investor in its seed round -- a materially smaller and less diversified compute position. SP001, SP022
CP048 Glean's platform explicitly avoids proprietary lock-in via open APIs and interoperates with 'leading LLMs from OpenAI, Google, Amazon, Meta, and Anthropic,' meaning a horizontal aggregator like Glean can multi-home any new model Core Automation might eventually ship, capturing the workflow layer regardless of which lab wins the underlying model race. SP012
CP049 Hebbia's official site reports approximately $30 trillion in assets under management among client institutions and roughly 200,000 average prompts processed per day, establishing a usage scale no disclosed Core Automation metric currently approaches. SP015
CP050 Manus's positioning as an orchestration layer built on third-party foundation models (Claude, Qwen) rather than its own frontier model illustrates a structurally different, lower-capex competitive path into the same automated-knowledge-work space Core Automation says it wants to enter, one that does not depend on a continual-learning research breakthrough. SP016
CP051 Core Automation's stated differentiation -- continual learning with roughly 100x less training data than current frontier models -- has not been independently verified or demonstrated in any public benchmark as of the run date, unlike Google DeepMind's Gemini Deep Think results, which are documented in submitted or published papers. SP004
CP052 Because frontier labs (OpenAI, Anthropic, Google DeepMind) and well-funded neolabs (AMI Labs, Reflection AI, Thinking Machines Lab) are all simultaneously recruiting from the same narrow pool of frontier researchers Core Automation has already hired from, Core Automation faces meaningfully elevated key-person retention risk relative to incumbents with larger balance sheets and more diversified research staffs. SP019, SP018
CI001 California Secretary of State filing records show "Core Automation (de), Inc." as a Delaware-formed stock corporation officially filed in California on March 24, 2026 under document number B20260125942, with Jerry (Jaroslaw) Tworek listed as registered agent. SI001, SI004
CI002 A SEC EDGAR full-text search for "Core Automation" restricted to Form D filings between 2026-01-01 and 2026-07-05 returned zero results, indicating no public Form D notice of an exempt securities offering has been filed for the company as of the run date. SI002
CI003 Multiple outlets reported that Core Automation closed an initial funding round of approximately $100 million at a roughly $1 billion valuation within weeks of its 2026 founding. SI003, SI004, SI006, SI008
CI004 Reported, but not independently confirmed, participants in Core Automation's initial funding round include Nvidia, Spark Capital, and Accel; no lead investor has been publicly confirmed. SI003, SI004
CI005 As of May 2026, Core Automation was reported to be in early discussions to raise $300 million to $500 million in new capital at a target valuation of approximately $4 billion, roughly a fourfold step-up from its initial round. SI004, SI005, SI006, SI008
CI006 As a private company that has not filed for an IPO or issued public debt, Core Automation is not required to disclose audited financial statements, making independent verification of any revenue, margin, or cash position impossible from public filings alone. SI002
CI007 The only independently verifiable government record identified for Core Automation is its California Secretary of State filing, which discloses entity type, filing date, and registered agent but no financial figures such as authorized shares, capital raised, or use of proceeds. SI001, SI002
CI008 Analyst firm Sacra reported that as of May 2026, Core Automation has no public API, pricing page, signup flow, or commercial product. SI008
CI009 Sacra characterizes Core Automation as "pre-revenue and pre-commercial," with a cost structure dominated by frontier AI research talent and compute and no offsetting customer revenue. SI008
CI010 Core Automation's current product is described by Sacra as the automation of its own internal research process -- the lab is both builder and first customer of its own automation stack, rather than selling to external customers. SI008, SI009
CI011 Sacra identifies plausible future monetization paths for Core Automation as B2B model or API access, enterprise software subscriptions for domain-specific automation, and usage-based pricing tied to autonomous tasks or compute -- all explicitly speculative and unconfirmed by the company. SI008
CI012 Core Automation's official homepage describes its mission as building "the world's most automated AI lab" and frames its objective around automating research itself, without referencing pricing, a product catalog, or revenue. SI009
CI013 Core Automation's technical blog post "When AI Starts Writing Systems Code" is rendered client-side and returns minimal static text on fetch, limiting independent verification of any business or cost detail it may contain beyond the headline framing. SI010
CI014 Core Automation's public website returns a 404 for a /careers path, providing no visible public job-listing page that would signal finance, legal, sales, or operations hiring activity. SI011
CI015 No independent source reviewed identifies a named paying customer, signed contract, or disclosed revenue figure for Core Automation as of the run date. SI008, SI009
CI016 Frontier AI research engineers in 2026 commonly receive total compensation, including base, bonus, and equity, of $500,000 to $1.5 million per year at senior levels, per industry compensation coverage. SI021
CI017 Reported outlier pay packages for elite AI researchers reached as high as $300 million over four years at large labs in 2025-2026, including signing bonuses reported as high as $100 million for a single year, illustrating extreme upside cost exposure in frontier-lab hiring. SI021
CI018 OpenAI reported approximately $3.7 billion in operating cash burn in Q1 2026 alongside a planned roughly $32 billion in 2026 model-training and compute spending, illustrating the scale of compute-driven cost at a frontier lab. SI023
CI019 Anthropic was reported to have reached roughly $30 billion in annualized run-rate revenue by April 2026 while spending on the order of $6 billion to $10 billion per year on compute and roughly $80 million per month in cash burn, with more than 60% of that spend allocated to cloud compute providers. SI025
CI020 Global AI infrastructure spending was projected to exceed $300 billion in 2026, with energy representing 30% to 40% of data-center operating costs. SI019, SI020, SI022
CI021 The four largest US hyperscalers -- Amazon, Alphabet, Meta, and Microsoft -- planned combined AI infrastructure capital expenditure of approximately $725 billion in 2026, a 77% increase over roughly $410 billion in 2025. SI022
CI022 Deloitte reports that per-unit AI inference costs fell roughly 280-fold over two years, yet total enterprise AI spending kept rising because usage growth has outpaced those efficiency gains -- a dynamic that would plausibly apply to any compute-intensive research lab, including Core Automation. SI020
CI023 Given Core Automation's talent- and compute-heavy, pre-revenue operating model and its similarity to peer neolabs that report monthly compute/payroll burn in the tens of millions to low hundreds of millions of dollars, its own burn rate is plausibly in a comparable range, though no company-specific figure has been disclosed. SI008, SI018, SI019
CI024 No source reviewed discloses Core Automation's current employee headcount or an aggregate payroll run-rate figure. SI008, SI009
CI025 Thinking Machines Lab, founded by former OpenAI CTO Mira Murati, closed a $2 billion seed round in July 2025 led by Andreessen Horowitz at a $12 billion valuation, with Nvidia, Accel, ServiceNow, Cisco, AMD, and Jane Street participating, before shipping a commercial product. SI012
CI026 Safe Superintelligence, co-founded by former OpenAI chief scientist Ilya Sutskever, raised a $1 billion seed round in September 2024 at approximately a $5 billion valuation, led by Andreessen Horowitz, Sequoia Capital, DST Global, SV Angel, and NFDG. SI013
CI027 Humans&, founded in September 2025 by former Anthropic, xAI, and Google researchers, raised a $480 million seed round in January 2026 at a $4.48 billion valuation, with Nvidia, Jeff Bezos, SV Angel, GV, and Emerson Collective participating -- one of the largest seed rounds in venture history. SI014
CI028 Global startup funding reached a record approximately $300 billion across roughly 6,000 startups in Q1 2026, driven heavily by outsized AI-lab funding rounds. SI015
CI029 Core Automation's reported $100 million-to-$1 billion, then $300-500 million-to-$4 billion valuation progression sits within, and in absolute dollar terms below, the range of comparable 2024-2026 neolab megaseed rounds, suggesting billion-dollar-plus pre-product valuations are now a category norm rather than a company-specific outlier. SI012, SI013, SI014, SI003
CI030 Sacra frames Core Automation's central business-model risk as whether its internal automation 'flywheel' can compound fast enough to produce a distributable product before its raised capital runs out -- an explicit third-party runway-risk framing rather than a company disclosure. SI008
CI031 No source reviewed discloses Core Automation's cash on hand, monthly cash burn, or a runway-in-months figure as of the run date. SI002, SI008
CI032 No source reviewed discloses a use-of-funds breakdown, such as compute versus hiring versus facilities, for Core Automation's confirmed or targeted financing rounds. SI002, SI008
CI033 Sequoia Capital partner David Cahn's AI infrastructure revenue-gap framework escalated from an estimated $200 billion in required annual AI revenue in 2023-2024 to roughly $600 billion by 2026, as hyperscaler AI capital expenditure grew faster than realized AI revenue. SI016
CI034 An MIT NANDA project report, "The GenAI Divide: State of AI in Business 2025," based on 300 AI deployments and 150 executive interviews, found that 95% of enterprise generative AI pilots fail to deliver measurable financial return on investment. SI017
CI035 GPU rental prices for Nvidia's Blackwell chips reportedly rose to $4.08 per hour in April 2026, up 48% in 60 days, amid a compute shortage reported to be causing outages at Anthropic and forcing OpenAI to cancel some product plans, with Bank of America projecting demand will outstrip supply through 2029. SI018
CI036 Nvidia committed more than $40 billion in AI equity investments in 2026, including a roughly $30 billion stake in OpenAI, alongside compute-credit and revenue-sharing arrangements with AI startups. SI027
CI037 Goldman Sachs analysts characterized Nvidia's dual role as supplier and equity investor in AI startups as a "circular revenue" risk, estimating that a large share of Nvidia's AI-related equity financing flows back to Nvidia itself as hardware purchases, which can overstate genuine end-user demand. SI026
CI038 Nvidia is separately reported as a participant in Core Automation's initial funding round, so any change in Nvidia's broader AI equity or compute-pricing strategy is a plausible, though not directly evidenced, channel through which Core Automation's own capital access or compute costs could be affected. SI003, SI004, SI027
CI039 Analysts and commentators increasingly compare the scale of 2026 AI infrastructure capital expenditure to prior vendor-financed infrastructure buildouts, such as the late-1990s telecom capacity boom, that preceded a sharp correction when realized demand fell short of built capacity. SI016, SI018
CI040 As of the run date, no independently verifiable figure exists for Core Automation's revenue, annual recurring revenue, gross margin, monthly burn, cash on hand, or runway. SI002, SI008, SI009
CI041 The only quantified capital facts on Core Automation's public record are press-reported, not regulator-confirmed, figures -- a $100 million initial raise and an in-talks $300-500 million follow-on -- both ultimately sourced to The Information via unnamed parties rather than to a filed document. SI003, SI004, SI006
CI042 Given a capital-intensive, revenue-free business model in a category where peers already deploy $1 billion or more before shipping a product, an outside investor cannot underwrite Core Automation's financial profile from public evidence alone and would need management-provided data such as a cap table, signed term sheet, compute contracts, and a payroll run-rate. SI008, SI012, SI013, SI014
CI043 The absence of any SEC Form D filing for a reported $100 million-plus raise as of the run date is itself informative: it implies either the offering has not yet triggered a public notice filing, uses a different exemption pathway, or that reported deal terms remain preliminary and unclosed. SI002
CI044 Concrete diligence requests that would materially improve underwriting confidence include a signed cap table and round term sheet, GPU/cloud compute contracts and committed spend, current headcount and payroll run-rate by function, and any convertible note, SAFE, or debt instrument outstanding ahead of the reported $4 billion round. SI001, SI002, SI008
CI045 No source reviewed discloses Core Automation's cap table, ownership percentages, board composition, or any debt or convertible financing instruments. SI001, SI002
CE001 Core Automation says it is building the world’s most automated AI lab. SE001
CE002 Core Automation says its objective is to build systems that optimize and automate work, starting with research itself. SE001
CE003 Core Automation says the next step change in AI will not come from larger models, more data, and static deployment. SE001
CE004 Core Automation says it is pursuing new learning algorithms beyond large-scale pretraining and reinforcement learning. SE001
CE005 Core Automation says it wants architectures that scale better than transformers. SE001
CE006 Core Automation says it is building the lab around small teams with highly capable agents. SE001
CE007 Nextomoro characterizes Core Automation as focusing on automating frontier-AI research workflows and on learning algorithms designed to replace large-scale pretraining. SE003
CE008 The public company surface reviewed here shows no external API documentation, no pricing, and no customer-facing product support surface. SE001, SE002
CE009 The 2026 continual-learning survey on arXiv says continual learning for LLMs aims dynamic adaptation to evolving knowledge and sequential tasks while mitigating catastrophic forgetting. SE004
CE010 The same survey frames continual learning methods across core training stages such as continual pre-training, continual fine-tuning, and post-training adaptation. SE004
CE011 The catastrophic-forgetting survey says neural networks tend to quickly forget prior knowledge when learning new information from a non-stationary stream. SE005
CE012 The post-transformer survey says transformers remain dominant despite shortcomings that include energy inefficiency and hallucinations. SE006
CE013 The post-transformer survey says active research is exploring alternative architectures, layers, objectives, and optimization techniques beyond transformers. SE006
CE014 ContinualAI’s public paper list organizes hundreds of continual-learning resources across architectural methods, benchmarks, and catastrophic-forgetting studies. SE008
CE015 The Awesome Incremental Learning repository lists multiple 2024-2025 surveys on continual and class-incremental learning. SE023
CE016 Mark Saroufim’s portfolio says he is a Core Automation cofounder, a PyTorch maintainer, and a cofounder of GPU MODE. SE011
CE017 The GPU MODE lectures repository provides public materials on CUDA, Triton, and PyTorch kernel integration. SE009
CE018 PyTorch’s KernelAgent post describes a hardware-guided multi-agent workflow for optimizing Triton kernels using real GPU performance signals. SE010
CE019 The same KernelAgent post says the system achieved 100 percent correctness across 250 KernelBench tasks. SE010
CE020 Mark Saroufim’s GitHub profile publicly shows major PyTorch and GPU MODE projects, including the gpu-mode lectures repository with thousands of stars. SE022
CE021 The GPU MODE YouTube channel and website show an active practitioner community around low-level GPU optimization. SE012, SE025
CE022 OpenAI says GPT-5.5 is available in the API and excels at writing and debugging code, researching online, analyzing data, and creating documents. SE013
CE023 OpenAI’s research index shows a continuing cadence of product, safety, and research releases in mid-2026. SE014
CE024 Thinking Machines launched Tinker as a fine-tuning API for researchers and hackers in 2025. SE015
CE025 FutureHouse says Robin integrates hypothesis generation with experimental data analysis in one continuous workflow for biological discovery. SE016
CE026 Google DeepMind says Gemini Deep Think is solving professional research problems across mathematics, physics, and computer science under expert direction. SE017
CE027 Core Automation’s public materials do not expose a model card, benchmark table, eval report, or public demo for Ceres or any named model. SE001, SE002
CE028 Because no benchmark or model card is public, differentiation is still thesis-led rather than performance-led in external evidence. SE001, SE002, SE003
CE029 The technical stack implied by public evidence centers on research agents, systems-code automation, continual-learning methods, and GPU-level optimization rather than a launched application SKU. SE001, SE002, SE010, SE011
CE030 Info-Tech says organizations are moving beyond experimentation into adaptive governance and agent-driven automation. SE020
CE031 Google says it applies its AI principles to product and research development through a Responsible AI progress process. SE018
CE032 The Blockchain Council 2026 state-of-AI report frames enterprise readiness and governance as central constraints on AI maturity. SE024
CE033 No public trust center, status page, security whitepaper, or compliance certification page was surfaced in the Core Automation materials reviewed for this chapter. SE001, SE002
CE034 The absence of public deployment, support, or reliability artifacts means external buyers cannot yet evaluate uptime, incident response, or service commitments. SE001, SE002
CE035 Core Automation’s public roadmap is implicit rather than explicit: automate internal research first, then potentially externalize tools or models later. SE001, SE002, SE003
CE036 The existence of extensive public continual-learning surveys, paper lists, and code repositories means the category itself is crowded even if Core Automation’s exact implementation remains private. SE004, SE008, SE023
CE037 Developer-signal sources reduce ambiguity about the team’s low-level systems competence because Saroufim’s public work spans PyTorch, quantization, GPU MODE, and kernel benchmarking. SE009, SE011, SE022
CE038 The same developer-signal sources also imply concentration risk because a meaningful share of the public systems narrative runs through a small number of named technical leaders. SE011, SE022, SE025
CE039 Adjacent organizations already expose public products or research systems for fine-tuning, coding, or scientific discovery, so Core Automation is behind them on external product maturity. SE013, SE015, SE016, SE017
CE040 The combination of agentic orchestration work, research-automation products, and post-transformer research makes Core Automation’s thesis technically plausible but not unique. SE006, SE010, SE016, SE017
CE041 Co-founder and CEO Jerry Tworek led OpenAI's o1 and o3 reasoning-model program and was a principal contributor to Codex and the GPT-3/GPT-4 series before founding Core Automation in 2026. SE026
CU001 Core Automation’s public homepage is framed around automating research work rather than serving a named external customer segment. SU001
CU002 The public company materials reviewed for this chapter show no customer logos, deployment claims, pricing, or sign-up flow. SU001
CU003 Nextomoro describes Core Automation as a 2026 AI research lab whose public story is still pre-product and research-first. SU002
CU004 Sacra’s company profile for Core Automation is still dominated by funding and valuation information rather than customer traction metrics. SU003
CU005 Glean’s customer stories page presents named practitioners and a named company case study as public customer proof. SU004
CU006 Glean’s customer stories include roles in IT operations, knowledge management, customer support, software engineering, and business operations. SU004
CU007 Hebbia’s homepage says the product is trusted by leading investors, bankers, advisors, and Fortune 500 companies for high-stakes decisions. SU005
CU008 Hebbia’s homepage reports 30 trillion dollars of AUM across firms using Hebbia, 200 thousand average prompts per day, and 1.5 billion pages processed. SU005
CU009 A named Oak Hill Advisors testimonial on Hebbia’s homepage says the product accelerated analyst work and created insights that influenced investment decisions. SU005
CU010 Glean’s homepage says leading enterprises use the platform and highlights integrations with Slack, Google Drive, Jira, Confluence, SharePoint, GitHub, and Salesforce. SU007
CU011 Glean’s Series F announcement and GetLatka’s 2026 estimate together imply that broad enterprise knowledge-work buyers can support material revenue at scale. SU008, SU009
CU012 TechCrunch reported that Hebbia had profitable revenue, indicating that high-stakes research automation can monetize with paying customers. SU006
CU013 Google Cloud says AI agents can understand a goal, develop a multi-step plan, and take actions under user guidance and oversight. SU010
CU014 Microsoft says organizations have moved from exploring AI agents to expecting measurable workflow impact from them. SU011
CU015 Microsoft’s Power Automate 2026 wave describes process mining, low-code flows, and robotic process automation as part of a broader automation platform for enterprise workflows. SU012
CU016 Anthropic says enterprises are shifting from simple task automation to complex multi-step workflows that span teams and business processes. SU013
CU017 Unite.AI says 2026 is a turning point in which AI agents move from demos into reliable business tools embedded in daily workflows. SU014
CU018 Deloitte says enterprise AI success in 2026 depends on moving from ambition and pilots to activation and scale. SU017
CU019 Ampcome says 54 percent of enterprises have already integrated AI agents into core operations by mid-2026. SU023
CU020 Joget says AI agents are moving out of labs and into business operations across tasks such as invoice reconciliation and security monitoring. SU024
CU021 Reinventing AI says 40 percent of enterprise applications will integrate task-specific AI agents by the end of 2026. SU025
CU022 The same Reinventing AI article says more than 40 percent of agentic AI projects may be canceled by 2027 because of execution risk. SU025
CU023 Infosys says procurement leaders now expect agentic systems to sense, decide, and act within defined guardrails rather than just provide passive support. SU015
CU024 Focal Point cites a Gartner prediction that by 2028 most B2B buying will be AI-agent intermediated, underscoring how procurement channels themselves are changing. SU016
CU025 Core Automation’s most plausible early buyers are frontier research teams, enterprise R&D groups, and knowledge-work organizations that want research or workflow automation. SU001, SU004, SU007, SU013
CU026 Because no public customer list exists, all current customer-segmentation hypotheses for Core Automation are inferred from mission statements and adjacent comparables rather than from direct evidence. SU001, SU002, SU003
CU027 No public evidence in reviewed materials confirms a Core Automation pilot, production deployment, or design-partner contract. SU001, SU002, SU003
CU028 No public retention metrics such as GRR, NRR, renewal rate, or contract length are disclosed for Core Automation. SU001, SU002, SU003
CU029 Without a public product surface, Core Automation’s likely first go-to-market path is high-touch design partnerships rather than self-serve adoption. SU001, SU002, SU015
CU030 Glean’s public proof supports a broad horizontal buyer set that spans IT operations, support, knowledge management, and business operations. SU004, SU007
CU031 Hebbia’s public proof supports a narrower but more vertical buyer set centered on finance, legal, and other high-stakes document workflows. SU005, SU006
CU032 Anthropic, Microsoft, Google, Deloitte, and Ampcome all describe 2026 as a period when agent systems are moving from pilots toward production use. SU010, SU011, SU013, SU017, SU023
CU033 Enterprise procurement, governance, and guardrail design are likely to be major friction points for any customer considering an agentic research platform. SU015, SU016, SU017, SU019
CU034 The likely customer journey for Core Automation runs from research or innovation sponsorship into technical evaluation, governance review, pilot, and only then broader production rollout. SU001, SU015, SU017
CU035 Because no named customers are public, customer concentration cannot be measured and top-account risk remains unknown. SU001, SU002, SU003
CU036 If early adoption concentrates in a handful of design partners, expansion risk and bargaining power could skew sharply toward those first accounts. SU015, SU016
CU037 Comparable buyers now expect measurable ROI, not just AI novelty, as shown by enterprise case studies and adoption surveys. SU004, SU021, SU023
CU038 AI Monk’s 2025-2026 case-study roundup says organizations report average ROI above traditional automation and cites high-volume production deployments such as JPMorgan and Klarna. SU021
CU039 Core Automation currently has lower external customer-proof quality than Glean or Hebbia because its public evidence stops at thesis rather than deployment. SU001, SU004, SU005
CU040 Until Core Automation can publish reference customers, retention, and deployment outcomes, customer adoption remains an underwriting hypothesis instead of a proven asset. SU001, SU003, SU025
CR001 Core Automation's homepage says the company is building the world's most automated AI lab and starts with automating research itself. SR001
CR002 The homepage says the lab is being built around small teams with highly capable agents and by automating the company's own work first. SR001
CR003 The official public surface currently resolves to a homepage and blog index rather than a product surface with public documentation or product modules. SR001, SR002
CR004 Sacra says Core Automation is both the builder and first customer of its own automation stack. SR006
CR005 Sacra reports that as of May 2026 Core Automation had no public API, pricing page, signup flow, or commercial product. SR006
CR006 Core's team, contact, and careers URLs all returned 404 errors on 2026-07-05. SR003, SR004, SR005
CR007 The homepage invites people to join while the careers page is broken, signaling an immature recruiting and operating surface. SR001, SR005
CR008 Sacra characterizes Core as pre-revenue and pre-commercial, with a cost structure dominated by frontier AI research talent and compute. SR006
CR009 Sacra reports a disclosed $100 million initial raise and later discussions of a roughly $300 million to $500 million follow-on at about a $4 billion valuation. SR006
CR010 Other 2026 coverage described broader fundraising ambitions of roughly $500 million to $1 billion and valuation expectations above $5 billion. SR007, SR008
CR011 Regardless of exact size, multiple 2026 sources show Core pursuing frontier-lab-scale financing within weeks of launch, making the business dependent on continuing capital access before productization. SR006, SR007, SR008
CR012 The Decoder says Core Automation launched to build the most automated AI lab in the world by automating its own research. SR009
CR013 Nextomoro says Core's public credibility centers on Jerry Tworek and a founding team recruited from OpenAI, Anthropic, and Google DeepMind. SR008
CR014 Public evidence still describes a frontier-lab recruiting story more than a customer or product rollout story. SR001, SR006, SR008
CR015 Gartner says generative AI for procurement entered the trough of disillusionment in 2025, with uneven ROI and some deployments falling short of expectations. SR010
CR016 Gartner says at least 30% of generative AI projects will be abandoned after proof of concept because of poor data quality, inadequate risk controls, escalating costs, or unclear business value. SR011
CR017 ISG says only 31% of the AI use cases it studied reached full production in 2025 and expected cost and productivity gains are underdelivering. SR015
CR018 ISG says leading AI copilots use cases are only one-third in production, underscoring how hard it is to scale front-line productivity tools. SR015
CR019 Wharton says enterprise users are optimistic but cautious as generative AI adoption shifts from experimentation toward measurable ROI. SR013
CR020 Deloitte says only 34% of organizations are truly reimagining the business with AI rather than simply optimizing existing processes. SR012
CR021 Deloitte says only one in five companies has a mature governance model for autonomous AI agents. SR012
CR022 Deloitte says 42% of companies feel strategically ready for AI but less prepared on infrastructure, data, risk, and talent. SR012
CR023 A16z reports that enterprise AI procurement now resembles traditional software buying, with rigorous evaluations, hosting choices, and benchmark scrutiny. SR014
CR024 A16z says security and cost have gained ground in model selection and that buyers increasingly use disciplined evaluation frameworks. SR014
CR025 A16z says switching costs rise once teams build guardrails and prompting around agentic workflows. SR014
CR026 OpenAI's trust portal publishes public security and compliance artifacts including SOC 2 Type 2 and ISO 27001, 27017, 27018, and 27701 certifications. SR019
CR027 Microsoft Trust Center exposes GDPR, EU AI Act, NIS2, Zero Trust, audit, privacy, and compliance resources for enterprise buyers. SR028
CR028 AWS says it supports 143 security standards and certifications and provides AWS Artifact plus a shared-responsibility model for customer compliance. SR029
CR029 Google Cloud Trust Center says Google undergoes independent verification and documents ISO, SOC, PCI DSS, FedRAMP, GDPR, and HIPAA-aligned controls. SR030
CR030 Anthropic maintains a public Trust Center, showing leading AI vendors now expose trust resources as part of the enterprise sales motion. SR024
CR031 Core Automation's public site does not surface an equivalent trust, privacy, or compliance resource. SR001, SR002, SR003, SR004, SR005
CR032 The EU AI Act service desk says general-purpose AI rules and governance apply from 2025-08-02 and the majority of rules and enforcement start on 2026-08-02. SR025
CR033 The FTC says there is no AI exemption from existing law and has already brought cases against unsupported AI service claims and AI-enabled deceptive schemes. SR027
CR034 Copyrightlaws says the United States has more than 80 active generative AI copyright cases and highlights training-data, output, and transparency disputes as unresolved legal questions. SR023
CR035 Copyrightlaws says proposed US transparency rules such as CLEAR would require reporting copyrighted training data and could impose civil penalties up to $2.5 million. SR023
CR036 NIST says AI RMF 1.0 is being revised and launched a 2026 profile for trustworthy AI in critical infrastructure, indicating governance expectations are still moving upward. SR022
CR037 Microsoft says LLMs will become quickly commoditized and today's breakthroughs will become tomorrow's table stakes. SR016
CR038 Microsoft says advantage shifts to how vendors integrate models with proprietary data and workflows rather than the models alone. SR016
CR039 vLLM markets cost-efficient LLM serving for everyone, supports open-source models on any hardware, and exposes an OpenAI-compatible API. SR020
CR040 SGLang says it powers production deployments across more than 400,000 GPUs and is designed for low-latency, high-throughput serving from single GPUs to large clusters. SR021
CR041 Google says Gemini Enterprise Agent Platform lets developers build, scale, govern, and optimize enterprise-ready agents grounded in enterprise data. SR026
CR042 Amazon Q Developer says its agentic capabilities can autonomously implement features, document, test, review, refactor code, and perform multistep development tasks. SR017
CR043 GitHub Copilot now offers enterprise license management, policy management, audit logs, and a control plane for agents across enterprise workflows. SR018
CR044 Together, Google, AWS, GitHub, and OpenAI already ship productized agent, coding, governance, and trust surfaces that overlap with Core's eventual workflow-automation pitch. SR017, SR018, SR019, SR026
CR045 The combination of product immaturity, no public customer proof, and tougher enterprise procurement creates a slower path to commercial validation than the funding narrative implies. SR006, SR010, SR011, SR015
CR046 The combination of open-source serving alternatives and incumbent platforms reduces the defensibility of generic automation or infrastructure differentiation. SR016, SR020, SR021, SR026
CR047 Core's highest risk is sequencing: if external proof, trust posture, and productization lag capital consumption, financing risk compounds rather than diversifies. SR006, SR011, SR015
CR048 The fastest risk-reducing mitigation is not more ambition but more evidence: named pilots, trust documentation, compute-governance disclosure, and a publicly testable product surface. SR006, SR019, SR028, SR029, SR030
CR049 No public case study, reference customer, or deployment evidence is visible on the official site as of 2026-07-05. SR001, SR002
CR050 A16z says off-the-shelf AI-native applications are eclipsing custom builds, raising the bar for a pre-product startup to convince buyers to wait for a bespoke future platform. SR014
CR051 Publicly visible mitigations are still thesis-level—small-team automation, continual learning, and hiring ambition—rather than demonstrated commercial controls. SR001, SR006, SR008
CR052 The public evidence lacks named pilots, security documentation, board disclosure, and compute-contract detail, so the near-term mitigation path is further diligence rather than immediate trust in execution. SR006, SR028, SR029, SR030
CR053 Public sources reviewed do not disclose board composition, independent directors, or investor-control terms. SR001, SR002, SR006, SR008
CR054 Because funding terms and oversight are undisclosed, governance resilience cannot be separated from founder judgment in public diligence. SR006, SR008
CV001 Core Automation's evidence set supports a track recommendation rather than buy because the company is pre-product, pre-revenue, and has no SEC-confirmed financing record. SV002, SV001, SV003
CV002 Confidence in this recommendation is medium because financing figures are corroborated by four independent outlets (AI Certs, Intellectia.ai/The Information, Silicon Report, The Decoder) but zero regulatory filings confirm them. SV003, SV004, SV005, SV006
CV003 Core Automation's risk rating is high given compounding key-person, financing, and market-timing risk documented across its comparable set, including a co-founder departure at the most richly valued peer. SV024, SV027
CV004 Valuation stance is stretched to expensive because the reported seed-to-follow-on markup from about $1 billion to about $4 billion within weeks lacks any disclosed product, customer, or revenue support. SV004, SV006
CV005 The practical decision implication is to withhold new capital commitments until Core Automation discloses product, revenue, or regulatory-filing evidence sufficient to test the reported valuation. SV002, SV030
CV006 Core Automation reportedly raised approximately $100 million in an initial round at a valuation near $1 billion within weeks of its 2026 launch. SV003, SV004
CV007 Reporting attributed to The Information via Intellectia.ai, and corroborated by Silicon Report and The Decoder, describes Core Automation as in talks for a $300 million-$500 million follow-on round targeting a valuation of approximately $4 billion. SV004, SV005, SV006
CV008 No source discloses a revenue, ARR, or user-metric basis for either the roughly $1 billion seed mark or the roughly $4 billion follow-on target, so no standard revenue multiple can be computed for Core Automation. SV004, SV006
CV009 A full-text search of SEC EDGAR for Form D filings mentioning "Core Automation" between January and July 2026 returned zero results, and a California business filing confirms only incorporation, not a securities filing, for either round. SV002, SV001
CV010 Core Automation's California corporate filing shows the entity was formally registered on March 24, 2026, shortly before the reported funding announcements. SV001
CV011 Taken together, the reported figures imply the follow-on talks sought roughly a four-fold markup on the seed valuation within a matter of weeks, an unusually short interval even by 2026 AI-lab standards. SV004, SV006
CV012 Safe Superintelligence (SSI), the closest public comparable for a pre-product frontier-research lab, raised roughly $2 billion in 2025 at a $32 billion valuation despite having no public-facing product, after an earlier $1 billion raise at a $5 billion valuation in 2024. SV024, SV008
CV013 SSI's valuation rose more than six-fold in under a year with a team of roughly 20 employees and no shipped product, illustrating how far founder-pedigree-driven valuations can run absent commercial proof. SV024
CV014 SSI co-founder Daniel Gross departed the company for Meta's AI division following reported acquisition interest, illustrating key-person risk even at the most richly valued pre-product AI lab. SV024
CV015 Thinking Machines Lab, another founder-pedigree neolab, raised a seed round reported at a $12 billion valuation in 2025. SV007
CV016 Thinking Machines Lab subsequently deepened its ties to Google through a new multi-billion-dollar deal, a form of strategic validation Core Automation has not disclosed. SV032
CV017 AMI Labs, a directly thesis-adjacent competitor pursuing world models, raised $1.03 billion, though no source discloses its resulting valuation. SV009
CV018 World Labs emerged from stealth in 2024 with a $230 million seed at a $1 billion valuation and, by February 2026, had raised a further $1 billion round including $200 million from Autodesk, with reports describing a target valuation of roughly $5 billion. SV022
CV019 Unlike Core Automation, World Labs had already shipped a commercial product (Marble) and disclosed a named strategic investor and commercial partnership (Autodesk) at the time of its re-rating. SV022
CV020 Periodic Labs, an AI-for-science neolab, raised a $300 million seed at a $1.3 billion valuation in September 2025 and by May 2026 was in talks to raise at least $500 million at a $7.5 billion valuation, a nearly six-fold increase in under eight months. SV023
CV021 The Periodic Labs round was reported as significantly oversubscribed, with discussions already underway for a further round at an even higher valuation. SV023
CV022 Mistral AI, a later-stage lab with shipped products and revenue ambitions, was reportedly in talks in June 2026 to raise about $3.5 billion at a valuation near $23 billion, nearly double its September 2025 mark. SV020
CV023 xAI raised $20 billion in a Series E round in January 2026 at a valuation of roughly $230-250 billion, disclosing about 600 million monthly active users and revenue from subscriptions and API usage, a disclosure profile far more mature than Core Automation. SV021
CV024 Hebbia, a product-layer competitor, raised $130 million at a $700 million valuation backed by $13 million of profitable revenue, illustrating that some AI companies are priced on disclosed unit economics rather than pedigree alone. SV013
CV025 Glean, another product-layer competitor, raised a $150 million Series F at a $7.2 billion valuation with a disclosed enterprise AI product and customer base, a disclosure profile Core Automation does not match. SV014
CV026 Across the reviewed comparable set, valuations for pre-product frontier-research labs are driven primarily by founder pedigree, compute partnerships, and narrative momentum rather than by disclosed product or revenue metrics, making Core Automation's reported markup directionally consistent with, but not independently verified against, its peer set. SV024, SV007, SV022, SV023
CV027 The pattern of rapid, multi-billion-dollar re-ratings for founder-pedigree AI labs recurs across at least five 2024-2026 comparables (SSI, Thinking Machines, AMI Labs, World Labs, Periodic Labs), suggesting Core Automation's reported trajectory follows a recognized category pattern rather than being an outlier. SV024, SV007, SV009, SV022, SV023
CV028 SSI's co-founder departure to a rival lab after reported acquisition interest is a documented precedent for key-person risk among founder-pedigree AI labs valued primarily on team reputation, a risk category that applies directly to Core Automation given its own small, concentrated founding team. SV024
CV029 Global venture capital reached a record $510 billion in the first half of 2026, with OpenAI and Anthropic alone absorbing roughly 43% (about $217 billion) of that total, and the top five AI companies capturing roughly 73% of all US venture deal value in Q1 2026 per PitchBook-NVCA data. SV027, SV019
CV030 AI’s share of global venture capital dollars rose to roughly 70% in the second quarter of 2026, up from about 50% a year earlier, after touching 80% in the first quarter, indicating capital concentration is intensifying rather than normalizing. SV027
CV031 Senior venture investors publicly described 2026 AI-sector concentration as unprecedented at a May 2026 industry panel, citing widespread ARR inflation concerns among AI startups. SV027
CV032 Investor Michael Burry publicly argued in a widely covered analysis that AI-sector spending and valuations show bubble characteristics, and a Financial-Times-corroborated report found OpenAI posted a $20.9 billion operating loss on $13.07 billion of 2025 revenue, intensifying AI-lab bubble criticism through mid-2026. SV015, SV017
CV033 Turing-Award-winning AI researcher Yann LeCun publicly called xAI a 'failure' in June 2026 and warned that AI labs are risking a 'big bubble explosion,' noting most frontier labs are losing money and are effectively funded by investors rather than revenue. SV016
CV034 Analyses of pre-product AI lab valuations in 2026 describe the pricing as reliant on team pedigree and comparable-startup 'hype multipliers' rather than on financial metrics, user growth, or go-to-market validation, criticisms that apply directly to Core Automation's undisclosed product and revenue status. SV025, SV026
CV035 AI startup revenue multiples in 2026 typically run 10x-50x, with a median around 20x-30x and late-stage category leaders occasionally clearing 100x, but these multiples require a disclosed revenue base that Core Automation does not have. SV025
CV036 Nearly 700 AI seed rounds above $10 million priced in 2025 alone, roughly four times the typical seed-round scale of a few years earlier, indicating seed-stage AI valuations broadly have been inflated well beyond historical norms. SV026
CV037 By mid-2026, at least 118 tracked notable startup collapses had destroyed roughly $49.9 billion in capital across the AI sector, with 'AI wrapper' companies lacking proprietary data or workflow moats disproportionately represented among the failures. SV030
CV038 A dedicated analysis of agentic-AI startup valuations found more than 40% of agentic AI projects are forecast to be canceled by 2027 per Gartner, and documented pre-seed valuations already declining from a median of $8.0 million in Q2 2025 to $7.7 million in Q3 2025, evidence of an emerging correction in the same agentic-workflow category Core Automation is pursuing. SV031
CV039 Nvidia scaled back a proposed $100 billion equity commitment to OpenAI to a $30 billion actual stake as OpenAI moved toward an IPO, illustrating that even the largest strategic investors are recalibrating the scale of AI-lab capital commitments in 2026. SV028
CV040 A sector-wide reckoning already visible among AI-wrapper startups and now reaching the agentic-workflow category is the clearest documented downside scenario that could compress Core Automation’s implied valuation multiple toward or below its reported seed mark. SV030, SV031
CV041 For Core Automation's bull case to be justified, the company would need to ship a benchmarked continual-learning or agentic-research capability within roughly 12-18 months and close its reported follow-on near the targeted $4 billion mark, mirroring the pattern by which World Labs and Thinking Machines paired fast re-rates with actual product or partnership disclosure. SV022, SV007
CV042 The base case assumes Core Automation continues to raise on team pedigree and narrative without public product disclosure, closing the follow-on near or below the reported target with stronger investor protections, consistent with the modal pattern across the 2024-2026 neolab comparable set. SV024, SV009
CV043 The chain from evidence to recommendation runs from pre-product/pre-revenue status and unconfirmed financing, through concentrated key-person and market-timing risk, to a stretched-to-expensive valuation stance and a track (not buy) recommendation. SV002, SV027
CV044 The most consequential thesis-break triggers for Core Automation are a stalled or below-target follow-on, a key-researcher departure, absence of product disclosure within 12 months, and a regulatory filing that contradicts the press-reported terms. SV024, SV002
CV045 The highest-priority outstanding diligence items are regulatory confirmation of both rounds’ terms, any product or benchmark evidence, customer or design-partner evidence, and cap-table/preference detail, none of which is available in the public record as of the run date. SV002, SV001
CV046 OpenAI and Anthropic’s move toward late-2026 IPOs at valuations approaching $1 trillion each signals the frontier-lab exit window is opening for the largest labs first, while smaller neolabs like Core Automation have no disclosed near-term exit path. SV028
CV047 Scored across market, proof, moat, economics, risk, valuation, and evidence quality, Core Automation rates strongest on market size and team pedigree and weakest on proof, economics disclosure, and evidence quality, an imbalance consistent with a track rather than buy recommendation. SV002, SV030
CV048 The plausible valuation range for Core Automation spans roughly $0.5-1 billion (bear), $2-4 billion (base), and $8-12 billion (bull), anchored to the scenario table and comparable-set outcomes rather than a disclosed financial model. SV024, SV022
CV049 Anthropic raised further capital in 2026 that brought its valuation near $1 trillion ahead of a planned IPO, illustrating that even top-tier frontier labs command valuations far beyond the neolab tier Core Automation occupies. SV010
CV050 Venture analysis of the "neolab" category explicitly names Core Automation alongside Adaption Labs as continual-learning bets, framing these companies as pursuing a research thesis rather than an immediate product, consistent with the absence of public product evidence found in this review. SV018
CV051 Reflection AI's valuation reportedly soared to $8 billion after a $2 billion round, another example of a frontier-adjacent lab re-rating rapidly amid the 2026 funding environment. SV011
CV052 No public source discloses Core Automation revenue, ARR, paying customers, or a benchmarked model release as of the run date, leaving the reported valuation without an independent financial or product anchor. SV004, SV006
CV053 The most recent financing-related coverage of Core Automation reviewed for this chapter is dated May 2026; no source found during this run reports a closed follow-on round, a revised valuation, or a lapsed/abandoned raise as of the July 2026 run date. SV004, SV006
CV054 A rapid seed-to-follow-on markup of the kind reported for Core Automation typically comes with heavier liquidation preferences and board protections for new investors, but no cap-table, term-sheet, or preference-stack detail is publicly available to confirm this for Core Automation. SV004, SV001
CV055 No licensed secondary-market or private-share pricing dataset covering Core Automation or its closest pre-product comparables was accessible during this review, leaving open whether secondary markdowns documented broadly for sub-frontier AI startups in 2026 apply to this specific comparable set. SV027
来源
编号出版方标题引文
SO001 Core Automation Core Automation -- homepage WE'RE BUILDING THE WORLD'S MOST AUTOMATED AI LAB.
SO002 Core Automation Core Automation Blog -- listing page
SO003 Core Automation When AI Starts Writing Systems Code To automate research, we must automate systems
SO004 Core Automation Core Automation -- team page (not found)
SO005 Core Automation Core Automation (@CoreAutoAI) / X Building systems to automate and optimize / San Francisco, CA / Joined January 2026
SO006 X (Jerry Tworek) Jerry Tworek (@MillionInt) / X CEO and co-founder of Core Automation / former VP of RL @ OpenAI: reasoning models, o3, o1, GPT4, ChatGPT, Codex, RL for robots
SO007 X (Mark Saroufim) Mark Saroufim (@marksaroufim) / X mts & co-founder @coreautoai
SO008 X (Joanne Jang) Joanne Jang (@joannejang) / X trying to automate my work @coreautoai // prev: model behavior & labs @openai
SO009 X (Julia Villagra) julia villagra (@juliavillagra) / X
SO010 X (Avery Lamp) Avery Lamp (@AveryLamp) / X
SO011 X (rohan_anil) ROHAN ANIL (@rohan_anil) / X 0 posts
SO012 X (anmol_gulati) Anmol gulati (@anmol_gulati) / X 0 posts
SO013 X (ehsanamid) Profile / X (account unavailable)
SO014 X (saisurya) Profile / X (account unavailable)
SO015 X (anmolGulatiAI) Profile / X (account unavailable)
SO016 Jerry Tworek (personal site) Jerry Tworek's homepage I'm a research lead at OpenAI, focusing on teaching language models to solve problems...
SO017 The Decoder Ex-OpenAI researcher Jerry Tworek launches Core Automation to build the most automated AI lab in the world unveiled his new AI lab, "Core Automation," with the goal of building "the most automated AI lab in the world"
SO018 AI CERTs AI Startup Funding: Core Automation Seeks $1B Weeks After Launch Bloomberg and Reuters have not yet corroborated the reported figures.
SO019 Intellectia.AI Core Automation, an AI model development company established by Jerry Tworek in March, seeks to secure $300M-$500M at a $4B valuation
SO020 SiliconReport Core Automation reportedly seeks $300M-$500M at a $4B valuation after $100M seed a move from $1 billion to $4 billion would add $3 billion in paper value before the company has been around for a full quarter
SO021 Sacra Core Automation funding, news & analysis As of May 2026, there is no public API, pricing page, signup flow, or commercial product.
SO022 Nextomoro Core Automation
SO023 BigGo Finance Former OpenAI Research Lead Launches AI Startup Core Automation, Poaching Top Talent from Anthropic and DeepMind shortly after its founding in late January, had already initiated negotiations seeking $500 million to $1 billion
SO024 Let's Data Science Core Automation Recruits Researchers From Anthropic, DeepMind
SO025 Techmeme Sources: AI model builder Core Automation, founded in March by Jerry Tworek, aims to raise $300M-$500M at a $4B valuation
SO026 Business Today OpenAI reasoning chief Jerry Tworek quits after 7 years: Here's why
SO027 Moneycontrol OpenAI VP and veteran researcher Jerry Tworek steps down, here's why
SO028 TokenRing AI (FinancialContent) The Reasoning Chief Exits: Jerry Tworek's Departure from OpenAI Marks the End of an Era
SO029 aiHola OpenAI's Reasoning Chief Leaves to Do Research "Hard to Do" at the Company He Helped Build I am leaving to try and explore types of research that are hard to do at OpenAI.
SO030 ai2.work Core Automation Poaches Top Anthropic and DeepMind Talent for AI Lab
SO031 Yahoo Finance New AI lab Core Automation 'nerdsniped' researchers from Anthropic, Google DeepMind Jerry Tworek nerdsniped me into starting this with him and others.
SM001 Stanford Institute for Human-Centered AI (HAI) The 2026 AI Index Report AI agents made a leap from 12% to ~66% task success on OSWorld, which tests agents on real computer tasks across operating systems, though they still fail roughly 1 in 3 attempts on structured benchmarks.
SM002 Forbes Stanford's AI Report Card: Agents Are Ready. Companies Are Not. AI agents failed 88% of real-world computer tasks 18 months ago. As of March 2026, the best models complete such tasks at rates approaching human performance.
SM003 Deloitte The State of AI in the Enterprise, 2026 Only about one in five organizations report having a mature framework for overseeing autonomous agents.
SM004 Gartner Gartner Forecasts Worldwide AI Spending to Grow 47% in 2026 Worldwide spending on AI is forecast to total $2.59 trillion in 2026, a 47% increase year-over-year.
SM005 Gartner Gartner Predicts 40% of Enterprise Apps Will Feature Task-Specific AI Agents by 2026, Up from Less Than 5% in 2025 Forty percent of enterprise applications will be integrated with task-specific AI agents by the end of 2026, up from less than 5% today.
SM006 IDC (via Business Wire) IDC FutureScape 2026 Predictions Reveal the Rise of Agentic AI and a Turning Point in Enterprise Transformation New IDC research outlines how agentic AI will reshape strategy, workforce, and innovation across global enterprises by 2030.
SM007 MIT NANDA (MIT Media Lab Project NANDA) The GenAI Divide: State of AI in Business 2025 Despite $30-40 billion in enterprise investment into GenAI, this report uncovers a surprising result in that 95% of organizations are getting zero return.
SM008 Google Research Introducing Nested Learning: A new ML paradigm for continual learning Despite the success of large language models (LLMs), a few fundamental challenges persist, especially around continual learning, the ability for a model to actively acquire new knowledge and skills over time without forgetting old ones.
SM009 Sakana AI The AI Scientist: Towards Fully Automated AI Research, Now Published in Nature An agent powered by foundation models capable of executing the entire machine learning research lifecycle... a paper describing all of this work and that includes new insights has been published in Nature.
SM010 arXiv (Lu et al.) The AI Scientist-v2: Workshop-Level Automated Scientific Discovery via Agentic Tree Search The AI Scientist-v2... is capable of producing the first entirely AI generated peer-review-accepted workshop paper.
SM011 arXiv (Lahoti et al.) Mamba-3: Improved Sequence Modeling using State Space Principles While the current Transformer-based models deliver strong model quality, their quadratic compute and linear memory make inference expensive. This has spurred the development of sub-quadratic models.
SM012 TechCrunch Anthropic's Claude Science bets on workflow, not a new model, to win over scientists
SM013 VentureBeat Anthropic says 80% of its new production code is now authored by Claude More than 80% of the code merged into Anthropic's production codebase in May wasn't authored by humans, but by its own AI model, Claude... an 8x increase in the volume of code shipped per engineer per quarter.
SM014 CNBC Michael Burry's next 'Big Short': An inside look at his analysis showing AI is a bubble
SM015 TechCrunch Exclusive: Google deepens Thinking Machines Lab ties with new multi-billion-dollar deal
SM016 TechCrunch Mira Murati's Thinking Machines Lab is worth $12B in seed round
SM017 Spheron Network GPU Shortage 2026: How to Secure AI Compute When GPUs Are Sold Out H100 SXM5 nodes are sitting at 36-52 week lead times from resellers right now. That is not a supply blip. It is a structural problem.
SM018 Crunchbase News Q1 2026 Shatters Venture Funding Records As AI Boom Pushes Startup Investment To New Heights Crunchbase data shows investors poured $300 billion into 6,000 startups globally in the quarter, up over 150% quarter over quarter and year over year.
SM019 World Economic Forum Physical AI: Powering the New Age of Industrial Operations
SM020 Boston Consulting Group Physical AI Will Reshape the Economics of Automation
SM021 Grand View Research Robotic Process Automation Market Size, Share Report, 2033 The global robotic process automation market size was estimated at USD 4.68 billion in 2025 and is projected to reach USD 35.84 billion by 2033, growing at a CAGR of 29.0%.
SM022 SoftwareStrategiesBlog Roundup of agentic AI forecasts and market estimates, 2026 Every major forecast agrees on direction. None agrees on scale. The standalone agentic AI market lands between $7 billion and $8.5 billion... That 25x gap is not a contradiction. It is a measurement problem.
SM023 Axis Intelligence Research AI Agents Statistics 2026: Market Size, Adoption, and the Deployment Gap ADGI Score (Q2 2026) = 93% minus 23% = 70 percentage points... roughly 70 have not yet achieved even single-function production scale.
SM024 National CIO Review MIT Finds GenAI Projects Fail ROI in 95% of Companies
SM025 Precedence Research Robotic Process Automation Market Size, AI-Driven Automation Trends The global robotic process automation market size was estimated at USD 28.31 billion in 2025 and is predicted to increase from USD 35.27 billion in 2026 to approximately USD 247.34 billion by 2035.
SM026 GitHub (SakanaAI) AI-Scientist: Towards Fully Automated AI Research
SM027 Axis Intelligence Research AI Adoption Statistics 2026: 88% Use It, 39% Benefit
SM028 Princeton Language and Intelligence (PLI) Mamba-3: Improved Sequence Modeling using State Space Principles (research summary)
SM029 Presenc.ai Research New Frontier Lab Tracker: Thinking Machines, SSI, xAI 2026
SP001 Anthropic Anthropic raises $65B in Series H funding at $965B post-money valuation This latest funding is expected to advance our safety and interpretability research, expand compute to meet growing demand for Claude, and scale the products and partnerships our customers rely on.
SP002 TechCrunch Anthropic raises $65 billion, nears $1T valuation ahead of IPO
SP003 TechCrunch Yann LeCun's AMI Labs raises $1.03B to build world models
SP004 Google DeepMind Gemini Deep Think: Redefining the Future of Scientific Research We built a math research agent (internally codenamed Aletheia), powered by Gemini Deep Think mode... this agent can admit failure to solve a problem, a key feature that improved the efficiency for researchers.
SP005 TechCrunch Sakana AI raises $135M Series B at a $2.65B valuation to continue building AI models for Japan
SP006 Sakana AI Sakana AI (homepage)
SP007 FutureHouse FutureHouse (homepage)
SP008 FutureHouse Research | FutureHouse
SP009 Folio3 AI Pulse Reflection AI Secures $2 Billion in Massive Funding Round, Valuation Soars to $8 Billion
SP010 Invezz Nvidia-backed Reflection AI eyes $25B in massive funding showdown
SP011 Reflection Reflection (homepage)
SP012 Glean Glean raises $150M Series F at $7.2B valuation to transform how companies use AI to accelerate innovation Our platform links seamlessly with more than 100 SaaS applications and enterprise data repositories... customers always retain complete control over their information.
SP013 Latka Glean Revenue 2026: $300M Est. ARR, $7.2B Valuation
SP014 TechCrunch AI startup Hebbia raised $130M at a $700M valuation on $13 million of profitable revenue
SP015 Hebbia Hebbia (homepage)
SP016 Manus Manus: Hands On AI (homepage) Manus is now part of Meta -- bringing AI to businesses worldwide
SP017 andrew.ooo Meta Unwinds $2B Manus AI Acquisition: What Builders Need to Know (June 2026) Meta is dismantling its December 2025 $2 billion acquisition of agentic AI platform Manus after China's National Development and Reform Commission (NDRC) ordered the deal unwound in April 2026.
SP018 Radical Ventures The Rise of NeoLabs Continual learning. Today's frontier models are frozen following training... Examples include Core Automation and Adaption Labs.
SP019 TechSpot In Silicon Valley's AI war, top researchers now command multi-million dollar paychecks The stakes are high because the number of people capable of building foundational AI models is extremely limited, estimated at around 2,000 globally.
SP020 Thinking Machines Lab Announcing Tinker
SP021 Creati.ai Adaption Labs Secures $50M Seed Funding for Adaptive AI Models That Learn On-the-Fly We have spent years optimizing for the training phase, building massive frozen artifacts that stop learning the moment they are deployed... Real intelligence isn't static. It adapts.
SP022 Forbes Credible AI Lab Critics Pile Up As The Bubble Math Worsens Documents obtained by Ed Zitron and confirmed by the Financial Times show OpenAI posting a $20.9 billion operating loss on $13.07 billion of revenue in 2025.
SP023 CNBC Godfather of AI blasts Musk's xAI as 'failure,' says labs are risking a 'big bubble explosion' Those companies are losing money, and basically, the use for most people is funded by the investors. That can't go on for a very long right?
SP024 Singularity Moments Jerry Tworek is betting that humans are the bottleneck in AI research We have heard this story about automated discovery a dozen times before, and it usually ends with a fancy dashboard and no breakthrough models.
SP025 OpenAI Introducing GPT-5.5
SP026 OpenAI OpenAI Research (listing page)
SI001 BizProfile.net Core Automation (de), Inc. San Francisco, CA - filing information Officially filed on March 24, 2026, this corporation is recognized under the document number B20260125942.
SI002 U.S. Securities and Exchange Commission (EDGAR Full-Text Search) EDGAR full-text search results for "Core Automation" Form D filings, Jan-Jul 2026 Zero results returned for a full-text search of Form D filings matching "Core Automation" filed between 2026-01-01 and 2026-07-05.
SI003 AI Certs AI Startup Funding: Core Automation Seeks $1B Weeks After Launch
SI004 Intellectia.ai Core Automation, an AI model development company established by Jerry Tworek in March, seeks to secure $300M-$500M at a $4B valuation
SI005 Silicon Report Core Automation reportedly seeks $300M-$500M at a $4B valuation after $100M seed
SI006 The Decoder AI money keeps flowing as Deepseek plans record raise and Core Automation quadruples valuation in weeks
SI007 AI2.Work Core Automation Poaches Top Anthropic and DeepMind Talent for AI Lab
SI008 Sacra Core Automation funding, news & analysis The company is pre-revenue and pre-commercial. Its current cost structure is dominated by frontier AI research talent and compute, with no offsetting customer revenue.
SI009 Core Automation Core Automation -- homepage
SI010 Core Automation When AI Starts Writing Systems Code
SI011 Core Automation Core Automation -- careers page (404 Not Found)
SI012 TechCrunch Mira Murati's Thinking Machines Lab is worth $12B in seed round
SI013 TechCrunch Ilya Sutskever's startup, Safe Superintelligence, raises $1B
SI014 TechCrunch Humans&, a 'human-centric' AI startup founded by Anthropic, xAI, Google alums, raised $480M seed round
SI015 Crunchbase News Q1 2026 Shatters Venture Funding Records As AI Boom Pushes Startup Investment To $300B
SI016 TechSpot Big Tech needs to generate $600 billion in annual revenue to justify AI hardware expenditure By Q4 2024, Nvidia's data center run-rate revenue forecast is predicted to be $150 billion... the AI revenue required for payback [is] $600 billion.
SI017 The National CIO Review MIT Finds GenAI Projects Fail ROI in 95% of Companies
SI018 Tomasz Tunguz The Beginning of Scarcity in AI GPU rental prices for Nvidia's Blackwell chips hit $4.08 per hour, up 48% in 60 days... Bank of America projects demand will outstrip supply through 2029.
SI019 Sustainable Atlas AI compute infrastructure costs in 2026: energy, chips, and cooling economics
SI020 Deloitte The AI infrastructure reckoning: Optimizing compute strategy in the age of inference economics
SI021 CNBC Behind the AI talent war: Why tech giants are paying millions to top hires
SI022 European Business Magazine Big Tech AI Capex 2026: $725B Spend vs. Nation GDP
SI023 WiseToast OpenAI Reports $3.7 Billion Q1 2026 Operational Cash Burn
SI024 Faceoff Technologies Core Automation Targets Massive Funding for Continual-Learning AI
SI025 SaaStr Anthropic Just Passed OpenAI in Revenue. While Spending 4x Less to Train Their Models
SI026 Benzinga Nvidia Gets A Goldman Sachs Warning: Circular Revenue Is A Risk
SI027 TechCrunch Nvidia has already committed $40B to equity AI deals this year
SE001 Core Automation Core Automation Our objective: systems that optimize and automate work, starting with research itself.
SE002 Core Automation When AI Starts Writing Systems Code | Core Automation
SE003 Nextomoro Core Automation
SE004 arXiv Continual Learning in Large Language Models: Methods, Challenges, and Opportunities
SE005 arXiv Continual Learning and Catastrophic Forgetting
SE006 arXiv What comes after transformers? – A selective survey connecting ideas in deep learningThis is an extended version of the published paper by Johannes Schneider and Michalis Vlachos titled “A survey of deep learning: From activations to transformers” which appeared at the International Conference on Agents and Artificial Intelligence(ICAART) in 2024. It was selected for post-publication.
SE007 Papers with Code Papers with Code
SE008 ContinualAI GitHub - ContinualAI/continual-learning-papers: Continual Learning papers list, curated by ContinualAI
SE009 GPU MODE GitHub - gpu-mode/lectures: Material for gpu-mode lectures
SE010 PyTorch Hardware-Guided GPU Kernel Optimization via Multi-Agent Orchestration – PyTorch
SE011 Mark Saroufim Portfolio · Mark Saroufim
SE012 GPU MODE GPU MODE
SE013 OpenAI Introducing GPT-5.5
SE014 OpenAI OpenAI Research
SE015 Thinking Machines Lab Announcing Tinker
SE016 FutureHouse Research | FutureHouse
SE017 Google DeepMind Gemini Deep Think: Redefining the Future of Scientific Research
SE018 Google Our 2026 Responsible AI Progress Report
SE019 Kearney Kearney AI Trends Report 2026 - Kearney
SE020 Info-Tech Research Group AI Trends 2026 Report: Risk, Agents, and Sovereignty Will Shape the Next Wave of Adoption, Says Info-Tech Research Group
SE021 GitHub GitHub - Paper2Chinese/CVPR-2026-reading-papers-with-code
SE022 GitHub msaroufim - Overview
SE023 GitHub GitHub - xialeiliu/Awesome-Incremental-Learning: Awesome Incremental Learning
SE024 Blockchain Council The State of AI in 2026: Trends, Challenges & Enterprise Readiness
SE025 YouTube GPU MODE
SE026 nextomoro Jerry Tworek At OpenAI he served as vice president of research and led the development of the o1 and o3 reasoning models, with prior contributions across Codex, early reinforcement-learning research applied to robotics, and the GPT-3 and GPT-4 series.
SU001 Core Automation Core Automation
SU002 Nextomoro Core Automation
SU003 Sacra Core Automation funding, news & analysis
SU004 Glean Enterprise AI customer stories | Glean Work AI See how leading companies put Work AI to work with Glean.
SU005 Hebbia Hebbia Hebbia has not only increased the speed at which analysts can perform, but also created insights into our various positions that have influenced our investment process.
SU006 TechCrunch AI startup Hebbia raised $130M at a $700M valuation on $13 million of profitable revenue | TechCrunch
SU007 Glean Glean – Work AI that Works | Agents, Assistant & Search
SU008 Glean Glean raises $150M Series F at $7.2B valuation to transform how companies use AI to accelerate innovation
SU009 GetLatka Glean Revenue 2026: $300M Est. ARR, $7.2B Valuation
SU010 Google Cloud 5 ways AI agents will transform the way we work in 2026
SU011 Microsoft 6 core capabilities to scale agent adoption in 2026 | Microsoft Copilot Blog
SU012 Microsoft Learn Overview of Power Automate 2026 release wave 1
SU013 Anthropic How enterprises are building AI agents in 2026 | Claude by Anthropic
SU014 Unite.AI AI Agents in 2026: How Businesses Will Use Them Differently
SU015 Infosys BPM agentic AI in procurement: a 2026 playbook | Infosys BPM
SU016 Focal Point The Future of Procurement | Trends and Predictions for 2026
SU017 Deloitte The State of AI in the Enterprise - 2026 AI report
SU018 Blockchain Council The State of AI in 2026: Trends, Challenges & Enterprise Readiness
SU019 Info-Tech Research Group AI Trends 2026 Report: Risk, Agents, and Sovereignty Will Shape the Next Wave of Adoption, Says Info-Tech Research Group
SU020 Clear Data Science AI Agents in 2026: From Prototypes to Autonomous Workflow Orchestrators - Clear Data Science Limited
SU021 AI Monk 12 Agentic AI Examples With Measurable ROI: Enterprise Case Studies From 2025-2026 | AI Monk
SU022 OpenAI Introducing GPT-5.5
SU023 Ampcome Enterprise AI Agents 2026: Mid-Year Report on What's Working
SU024 Joget AI Agent Adoption 2026: What the Data Shows | Gartner, IDC
SU025 Reinventing AI Enterprise AI Agents Move From Pilot to Production: What 2026 Data Reveals Gartner predicts that over 40% of agentic AI projects will be canceled by 2027.
SR001 Core Automation Core Automation
SR002 Core Automation Core Automation Blog
SR003 Core Automation 404: NOT_FOUND
SR004 Core Automation 404: NOT_FOUND
SR005 Core Automation 404: NOT_FOUND
SR006 Sacra Core Automation funding, news & analysis
SR007 AI CERTs AI Startup Funding: Core Automation Seeks $1B Weeks After Launch
SR008 Nextomoro Core Automation
SR009 The Decoder Ex-OpenAI researcher Jerry Tworek launches Core Automation to build the most automated AI lab in the world
SR010 Gartner Gartner Says Generative AI for Procurement Has Entered the Trough of Disillusionment
SR011 Gartner Gartner Predicts 30% of Generative AI Projects Will Be Abandoned After Proof of Concept by End of 2025
SR012 Deloitte The State of AI in the Enterprise - 2026 AI report
SR013 Knowledge at Wharton 2025 AI Adoption Report: Gen AI Fast-Tracks Into the Enterprise
SR014 Andreessen Horowitz How 100 Enterprise CIOs Are Building and Buying Gen AI in 2025
SR015 ISG State of Enterprise AI Adoption Report 2025
SR016 Microsoft WorkLab LLMs Are Becoming a Commodity—Now What?
SR017 Amazon Web Services Amazon Q Developer
SR018 GitHub GitHub Copilot · Your AI pair programmer
SR019 OpenAI OpenAI Trust Portal
SR020 vLLM vLLM
SR021 LMSYS SGLang
SR022 National Institute of Standards and Technology AI Risk Management Framework
SR023 Copyrightlaws.com Copyright and Generative AI - 2026 Quarterly Update One
SR024 Anthropic Anthropic Trust Center
SR025 European Commission AI Act Service Desk Timeline for the Implementation of the EU AI Act
SR026 Google Cloud Gemini Enterprise Agent Platform (formerly Vertex AI)
SR027 Federal Trade Commission FTC Announces Crackdown on Deceptive AI Claims and Schemes
SR028 Microsoft Microsoft Trust Center | Data Security, Privacy, and Compliance
SR029 Amazon Web Services Cloud Compliance - AWS
SR030 Google Cloud Trust Center - Security and Compliance | Google Cloud
SV001 BizProfile.net Core Automation (de), Inc. San Francisco, CA - filing information Officially filed on March 24, 2026, this corporation is recognized under the document number B20260125942.
SV002 U.S. Securities and Exchange Commission (EDGAR Full-Text Search) EDGAR full-text search results for "Core Automation" Form D filings, Jan-Jul 2026 Zero results returned for a full-text search of Form D filings matching "Core Automation" filed between 2026-01-01 and 2026-07-05.
SV003 AI Certs AI Startup Funding: Core Automation Seeks $1B Weeks After Launch
SV004 Intellectia.ai Sources: AI model builder Core Automation, founded in March by Jerry Tworek, aims to raise $300M-$500M at a $4B valuation after raising $100M at a $1B valuation (The Information)
SV005 Silicon Report Core Automation reportedly seeks $300M-$500M at a $4B valuation after $100M seed
SV006 The Decoder AI money keeps flowing as DeepSeek plans record raise and Core Automation quadruples valuation in weeks
SV007 TechCrunch Mira Murati's Thinking Machines Lab is worth $12B in seed round
SV008 TechCrunch Ilya Sutskever's startup Safe Superintelligence raises $1B
SV009 TechCrunch Yann LeCun's AMI Labs raises $1.03 billion to build world models
SV010 TechCrunch Anthropic raises $65 billion, nears $1T valuation ahead of IPO
SV011 Folio3.ai Reflection AI secures USD 2 billion in massive funding round, valuation soars to USD 8 billion
SV012 TechCrunch Sakana AI raises $135M Series B at a $2.65B valuation to continue building AI models for Japan
SV013 TechCrunch AI startup Hebbia raised $130M at a $700M valuation on $13 million of profitable revenue
SV014 Glean Glean raises $150M Series F at $7.2B valuation to transform how companies use AI to accelerate innovation
SV015 CNBC Michael Burry's next 'Big Short': An inside look at his analysis showing AI is a bubble
SV016 CNBC Godfather of AI blasts Musk's xAI as 'failure,' says labs are risking a 'big bubble explosion' Those companies are losing money, and basically, the use for most people is funded by the investors. That can't go on for a very long right?
SV017 Forbes Credible AI Lab Critics Pile Up As The Bubble Math Worsens Documents obtained by Ed Zitron and confirmed by the Financial Times show OpenAI posting a $20.9 billion operating loss on $13.07 billion of revenue in 2025.
SV018 Radical Ventures The Rise of NeoLabs
SV019 Crunchbase News Q1 2026 Shatters Venture Funding Records As AI Boom Pushes Startup Investment To $300B Crunchbase data shows investors poured $300 billion into 6,000 startups globally in the quarter, up over 150% quarter over quarter and year over year.
SV020 TechCrunch Mistral is rumored to be raising €3B at €20B valuation
SV021 TechCrunch xAI says it raised $20B in Series E funding
SV022 TechCrunch World Labs lands $1B, with $200M from Autodesk, to bring world models into 3D workflows
SV023 Forbes Former OpenAI Researcher To Raise $500 Million For AI Science Startup Periodic Labs, a startup building an AI scientist that can use automated labs to make discoveries, is in advanced talks to raise at a $7.5 billion valuation... a nearly sixfold jump since it was founded.
SV024 VC Tavern Safe Superintelligence Raises $2 Billion at $32 Billion Valuation Despite No Public Product
SV025 Qubit Capital AI Startup Valuation Multiples: 10x-50x Range (2026)
SV026 Eqvista AI Startup Fundraising Trends 2026 (Seed to Series B)
SV027 Angel Investors Network AI Mega-Rounds Are Making VC a Concentrated Bet Three out of every four dollars deployed by American venture capital in the first quarter of 2026 went to five companies, and every one of them is an AI or AI-infrastructure business.
SV028 Tech Times NVIDIA OpenAI Investment Shrinks From $100B to $30B: Compute Lock-In War Continues
SV029 World Economic Forum How would the bursting of an AI bubble actually play out?
SV030 IdeaProof.io Startup Failures 2026: The Ongoing AI Reckoning Report 118 tracked shutdowns, $49.9B capital destroyed, 8 sectors hit.
SV031 AgentMarketCap The AI Agent Valuation Correction Is Here: Which Startups Are Stalling and Who Gets Acquired
SV032 TechCrunch Exclusive: Google deepens Thinking Machines Lab ties with new multi-billion-dollar deal