初创公司尽调
尽调报告 Robotics AI / Physical Intelligence Series B / Pre-commercial 2026-06-30

Generalist AI

具身 AI 前沿团队,技术信号强,公开商业验证有限

观察 — Generalist AI 汇集顶尖具身 AI 人才,并可能拥有差异化数据引擎;相对同业估值更低,但公开证据仍停在收入、具名客户和可部署级验证之前。

封面要素

最近一轮融资 02
$400M Series B [CO028, CV001]
成立时间 04
2024 [CO001]
商业状态 06
Early access only; no named customers disclosed [CO036, CV006, CU002]

公司概况

Generalist AI 是一家位于 San Mateo 的具身 AI 初创公司,由 Pete Florence、Andy Zeng 和 Andrew Barry 于 2024 年创立,目标是打造可跨多种硬件形态运行的通用机器人智能。公司的核心产品线从 2025 年 11 月的 GEN-0 推进到 2026 年 4 月的 GEN-1;后者宣称任务可靠性 99%、速度提升 3x,并可借助自有 500,000 小时物理交互数据集,在 1 小时内适配新任务。Generalist 已在两轮披露融资中融得约 $540M,包括 2026 年 6 月以 $2B 估值完成 $400M Series B,但仍未披露收入、定价或已署名的量产客户。

官网
generalistai.com
创始人
Pete Florence, Andy Zeng, Andrew Barry
创立地点
Bay Area, California, USA
总部
San Mateo, California, USA
产品
面向机器人的具身基础模型,以 GEN-1 为核心;GEN-1 是服务精细物理任务的硬件无关多模态控制模型,主要基于自有物理交互数据集从零训练。
客户
制造、物流、服装、汽车、电子等工作流中的精选早期试用伙伴;未披露已署名量产客户。
商业模式
作为机器人软件智能层,通过战略早期试用伙伴交付;可能采用企业授权或按机器人收费,但未公开定价。
阶段
Series B
融资情况
2026 年 6 月以 $2B 投后估值完成 $400M Series B;2024 年以来披露融资约 $540M。
[CO001, CO003, CO008, CO010, CO012, CO020, CO024, CO027]

执行摘要

主要优势

  • 创始团队质量极高:Florence、Zeng、Barry 叠加 DeepMind 机器人、PaLM-E/RT-2、Code as Policies 和 Boston Dynamics 的实战经验。
  • GEN-1 不绑定硬件,再加上 500,000 小时自有数据引擎;如果物理 AI 的扩展定律继续成立,公司有机会筑起护城河。
  • June 2026 轮融资引入 Radical Ventures、NVIDIA NVentures、Bezos Expeditions、8VC、USV 等高信号投资方,给公司换来时间和战略背书。

主要风险

  • 截至 June 2026,公司没有公开收入、定价或具名生产客户,$2B 估值还无法落到商业基本面上。
  • 安全、责任和欧盟高风险 AI 合规义务推进得更快,Generalist 公开的信任与安全披露追不上。
  • Skild AI、Physical Intelligence 等资金更足的对手,加上 NVIDIA 开源 GR00T stack,可能在 Generalist 证明部署经济性可持续之前挤压模型层差异化。
  • 业务依赖昂贵算力、数据采集运营和创始人亲自推动的企业销售,资本强度仍高。

未决问题

  • 具名付费客户、合同结构,以及至少一个带可量化 ROI 的生产部署仍未披露。
  • 公开记录仍缺少定价、烧钱速度、现金跑道、毛利率,以及任何收入或 ARR 披露。
  • GEN-1 的公开安全、红队和监管合规文件仍缺失,尽管欧盟高风险 AI 执法已逼近。
  • 董事会构成、June 2026 轮融资带来的治理权利,以及股权结构表稀释细节,仍只能通过 Form D 记录看到一部分。

目录

Chapter 01

01公司概览

1.1 身份、使命与商业模式

Generalist AI, Inc. 以「Generalist」品牌运营,是一家美国人工智能公司,2024 年注册成立,总部位于 California 州 San Mateo,并在 Massachusetts 州 Somerville(Boston)设有第二办公室。公司自称是「一家由前沿 AI 研究和产品驱动、为物理世界构建通用智能的公司」。它公开的使命是开发具身基础模型,让通用机器人真正落地。这类 AI 系统能在不同物理环境和机器人形态中感知、推理和行动,而不是被锁定在单一任务或硬件平台上。 不同于传统机器人公司,Generalist AI 明确不做硬件制造商。它的商业主张是充当智能层,也就是可跨任意机器人本体工作的「认知大脑」,覆盖从 6 自由度工业机械臂到半人形系统。公司当前商业策略围绕 GEN-1 的早期试用计划展开,面向精选行业伙伴,并给出一个数据飞轮目标:真实部署产生新的训练数据,再喂给下一代模型。公司尚未公开披露定价、已署名客户或年度经常性收入。 Generalist AI 在多次官方沟通中把长期野心表述为「physical AGI」——能够掌握任何物理任务的通用机器人智能。公司明确把技术路线图类比为大语言模型的扩展轨迹,认为催生 GPT-3 和 ChatGPT 的数据与算力扩展动力,一旦作用到物理交互数据上,也会产生可泛化到人类全部体力工作的机器人。 [CO001, CO002, CO003, CO004, CO005, CO006]

Generalist AI — 快照 KPI 表
指标数值 / 状态截至置信度缺口 / 备注
投后估值$2 billion2026 年 6 月通过新闻自报;无独立验证
累计融资> $500 million2026 年 6 月公司官方公告;拆分:~$140M(2025)、$400M(2026)
成立年份2024已确认
总部双总部:San Mateo, CA + Somerville, MA2026 年 6 月
员工人数(约)~70+ 名员工2026 年中约数;精确人数未披露
收入 / ARR未披露早期访问计划;定价或收入数字未公开
具名客户未披露GEN-1 面向精选伙伴提供;尚无具名客户确认
已发布模型GEN-0(2025 年 11 月)、GEN-1(2026 年 4 月)2026
训练数据规模> 500,000 小时物理交互2026 年 4 月每周新增 10,000+ 小时
IPO / 退出计划未公布Unknown私营公司;未披露 IPO 时间线

收入、客户数量和员工人数未公开披露;标记为低或未知的值都是证据缺口。估值为融资公告时公司自报。

[CO001, CO003, CO004, CO029, CO030, CO035]
FO003: Generalist AI — 快照 KPI

截至 2026 年 6 月,公司成熟度、牵引力和风险画像的关键指标。

员工数来自媒体报道估算;除融资外,所有财务指标均未公开披露。状态值:confirmed = 多源验证;estimated = 近似估计;gap = 证据缺口。

[CO021, CO024, CO029, CO030, CO035]

1.2 创始人、领导层与关键人物风险

Generalist AI 由三位在具身 AI 和机器人领域根基很深的研究者共同创立。Pete Florence(CEO)曾任 Google DeepMind 高级研究科学家,也是 PaLM-E 和 RT-2 的资深作者;这两篇论文是机器人学习中引用率最高的基础论文之一。Andy Zeng(首席科学家)曾任 Google DeepMind 研究科学家和技术负责人,也是「Code as Policies」主要作者;这同样是一篇影响力很高的机器人论文。Andrew Barry(CTO)曾任 Boston Dynamics 高级机器人专家,参与 Atlas、Spot、Stretch 等平台,并与 Harvard 机器学习团队有关联。 更广泛的团队来自 OpenAI、Boston Dynamics 和 Google DeepMind。以公司年龄看,Generalist AI 拥有异常集中的研究履历。8VC 的投资人评论把 Pete Florence 描述为「既是研究者,也是实干的建设者——很有号召力,并且在同等研究履历的人里少见地商业化」,凸显公司刻意在前沿研究和商业执行之间做平衡。 关键人物集中度是实质风险。公司的科学身份、融资叙事和媒体能见度都紧紧绑定 Florence 和 Zeng;几乎所有主要媒体报道都会点名两人。公司尚未公开披露正式董事会构成;治理结构、独立董事席位和任何董事会观察员,在可得公开来源中都未得到确认。按惯例,2026 年 6 月融资的领投方 Radical Ventures 可能拥有董事席位,但官方并未确认。 [CO008, CO009, CO010, CO011, CO012, CO013]

领导层与创始人表
姓名角色过往任职关键贡献 / 创始人-市场匹配关键人物依赖
Pete FlorenceCEO 兼联合创始人Google DeepMind(高级研究科学家)PaLM-E 和 RT-2 第一作者;参与规模化 ChatGPT/GPT-4 的共同作者;主要公开面孔和募资人高 — 科学信誉和投资者关系高度集中
Andy Zeng首席科学家兼联合创始人Google DeepMind(研究科学家 / 技术负责人)Code as Policies 第一作者;机器人学习和扩展方面深厚专长高 — 研究方向和模型架构所有权集中
Andrew BarryCTO 兼联合创始人Boston Dynamics(高级机器人专家)/ Harvard一线机器人工程能力;参与打造 Atlas、Spot、Stretch 平台中 — 若离任会带来执行风险,但比研究负责人更容易替补

董事会构成未公开披露;Radical Ventures(领投方)按惯例可能持有董事席位,但未获确认。截至 2026 年 6 月,未报告重大领导层离职或变动。

[CO008, CO009, CO010, CO011, CO012, CO013]

1.3 技术平台与产品里程碑

Generalist AI 的核心技术路线立在三根支柱上:自有大规模物理交互数据集、专门设计的多模态基础模型架构,以及名为「data hands」的硬件数据采集系统。 data hands 是绑在手腕上的设备,用来在全球各地的家庭、仓库和工作场所大规模记录人类操作任务,包括抓取、放置、折叠和配套分拣。这条路线绕开了 Physical Intelligence 等竞争对手使用的远程操作装置,而是用较低成本、更高多样性捕捉自然的人类灵巧操作。由此形成的数据集,从 GEN-0 发布时的 270,000 小时,增长到 GEN-1 发布时的 500,000 小时以上,并且仍以每周超过 10,000 小时的速度扩张。 GEN-0(2025 年 11 月 4 日发布)首次展示了机器人领域的扩展律:模型越大、物理数据越多,下游任务表现就会以可预测方式提升。GEN-0 还引入 Harmonic Reasoning,这是一种新的训练架构,让模型可以在连续时间里同时思考和行动;真实物理系统不会暂停,这一点至关重要。在已测试的模型规模(最高 10B+ 参数)下,GEN-0 展示了跨 6DoF、7DoF 和 16+DoF 机器人的跨机体泛化能力。 GEN-1(2026 年 4 月 2 日发布)在 GEN-0 的基础上继续扩大数据和算力,并加入后训练技术、强化学习等算法进展。GEN-1 在此前模型成功率为 64% 的任务上达到 99% 成功率,完成任务速度约为前沿水平的 3x,并可用约 1 小时机器人专属数据适配新任务。值得注意的是,GEN-1 完全从零训练——约 99% 参数为新训练参数——而不是微调已有视觉-语言模型。公司认为,这条「从零训练」路线给了它掌控前沿所需的完整架构控制权。 GEN-1 正在向精选行业伙伴开放早期试用。公司已经展示的任务包括汽车零件配套分拣、T 恤折叠(无人干预连续 86 次)、机器人吸尘器维护(连续 200+ 次),以及可持续数千个连续周期的包装作业。 [CO015, CO016, CO017, CO018, CO019, CO020]

FO002: Generalist AI — 公司快照逻辑(系统流)

Generalist AI 的身份、数据引擎、产品模型和资本如何连成一个系统。

[CO006, CO015, CO024, CO026, CO036]

1.4 融资历史与投资人格局

Generalist AI 自 2024 年成立以来披露了两轮融资。第一轮于 2025 年 3 月完成,总额约 $140M,投后估值 $440M。该轮投资人包括 Spark Capital、NVIDIA 旗下 NVentures、Bezos Expeditions(Jeff Bezos)和 Boldstart Ventures。第二轮 $400M 于 2026 年 6 月 4 日宣布完成,投后估值 $2B,使已披露累计融资超过 $500M。 2026 年 6 月这一轮由 Radical Ventures 领投,新进机构包括 8VC、Union Square Ventures、Hanabi Capital 和 Norwest。所有主要老股东——NVIDIA NVentures、Boldstart Ventures、Spark Capital、Bezos Expeditions 和 NFDG——均大额跟投。2026 年融资中的知名天使投资人包括 Fei-Fei Li(Stanford AI Lab 联合主任、World Labs 创始人)、Bin Lin(Xiaomi 联合创始人)、Eric Yuan(Zoom CEO)和 Naval Ravikant。 NVIDIA 持续参与,反映出战略兴趣:Generalist AI 的模型依赖 GPU 算力基础设施,任何大规模部署具身基础模型,都会意味着可观的 NVIDIA 硬件需求。Fei-Fei Li 和 Naval Ravikant 等技术型天使的出现,也在更广泛 AI 研究圈内释放了可信度信号。 收入、ARR、客户数量和二级交易条款均未公开披露。公司尚未宣布 IPO 计划。公开资料也未显示任何债务或信贷额度。 [CO027, CO028, CO029, CO030, CO031, CO032]

利益相关方 / 投资者地图
利益相关方角色 / 类型轮次控制权 / 经济重要性尽调问题
Radical Ventures领投机构投资者第 2 轮(2026 年 6 月)高 — $400M 轮领投;可能有董事席位确认董事席位;检查与 Physical AI 命题的一致性
NVIDIA NVentures战略投资者(GPU 供应链)第 1 轮和第 2 轮高 — 战略性;GPU 依赖形成锁定了解供应条款;是否有排他性或优惠算力定价
Bezos Expeditions个人战略投资者(Jeff Bezos)第 1 轮和第 2 轮中 — 突出信号;非领投检查是否有战略承诺或 Amazon 邻接关系
Spark Capital机构投资者第 1 轮和第 2 轮中 — 早期信念;后续参与了解治理权;是否行使 pro-rata
Boldstart Ventures早期机构投资者第 1 轮和第 2 轮中 — 持续支持方如有,确认董事会观察员身份
8VC新机构投资者第 2 轮(2026 年 6 月)中 — 发布公开命题;相信灵巧度作为切入点检查是否有董事席位;确认投资命题一致性
Union Square Ventures机构投资者第 2 轮(2026 年 6 月)中 — 泛科技 VC;机器人对 USV 而言较新了解资本之外的战略价值
Fei-Fei Li天使投资人(Stanford AI Lab、World Labs)第 2 轮(2026 年 6 月)低-中 — 信号 / 顾问;可能无董事席位确认顾问关系;与 World Labs 无冲突
Naval Ravikant天使投资人第 2 轮(2026 年 6 月)低 — 信号价值标准天使条款
Bin Lin天使投资人(Xiaomi 联合创始人)第 2 轮(2026 年 6 月)低-中 — 亚洲市场战略价值了解任何中国市场影响
NFDG现有机构投资者第 1 轮和第 2 轮低-中 — 公开可见度不高确认 NFDG 全称和投资命题

董事会构成未公开披露。利益相关方角色和控制权估计来自媒体报道和投资人博客。第 1 轮投资者名单部分来自 Forbes(2025 年 3 月完成)。Hanabi Capital 和 Norwest 是第 2 轮参与方,因公开细节有限未列入上表。

[CO027, CO028, CO031, CO032, CO033, CO034]

1.5 里程碑与公司轨迹

Generalist AI 运营不到 30 个月,就跑出了一条异常压缩的里程碑序列:成立、首轮融资完成、两代模型发布,以及以初始估值 5 倍完成第二轮融资。2026 年 6 月的融资公告明确称数据飞轮已经出现——真实企业正在生成任务数据,喂给后续模型迭代——这标志着公司从纯研发转向早期商业运营。 截至本文撰写时,公司最主要的有记录负面信号来自外部:投资人和独立分析师质疑「只靠扩大数据规模就足够」这一命题。Cobot CEO、前 Amazon 机器人高管 Brad Porter 在 Forbes 采访中称,「在一个并不完美的架构上硬砸海量数据,成本非常高,也未必能得到你想要的结果」;他引用 ImageNet 和 transformer 突破的历史类比,作为规模必须与架构创新共同演化的证据。Generalist AI 在 GEN-1 博文中给出的隐含反驳是,GEN-1 本身就是从零训练的完整架构重设计,而不是蛮力复制。 截至 2026 年 6 月本次运行日,公开可得来源未报道监管行动、专利纠纷、诉讼、领导层离职、裁员或安全事件。 竞争对手语境很重要:截至 2026 年 4 月,Physical Intelligence(pi.ai)据称正在洽谈以 $11B 估值融资 $1B——这意味着 Generalist 的估值大约只有该竞争对手隐含估值的五分之一,同时声称自己具备架构差异化。 [CO038, CO039, CO040, CO016, CO020, CO027]

里程碑表
日期事件类型金额 / 估值 / 状态关键参与方含义
2024公司成立成立创始人:Pete Florence、Andy Zeng、Andrew Barry具身 AI 前沿实验室成立;从零构建新架构
2025 年 3 月 24 日种子轮 / 第 1 轮完成融资融资 ~$140M;估值 ~$440M投资方:Spark Capital、NVIDIA NVentures、Bezos Expeditions、Boldstart Ventures团队获得验证;为 GEN-0 研发和数据引擎建设提供资金续航
2025 年 11 月 4 日GEN-0 模型发布产品Generalist AI 团队机器人扩展法则的第一项证明;270K+ 小时数据集;Harmonic Reasoning 架构
2026 年 4 月 2 日GEN-1 模型发布产品Generalist AI 团队达到商业可行性门槛;99% 成功率、3x 速度、1 小时适应
2026 年 4 月Forbes 专访和“ChatGPT moment”报道规模化Forbes(Anna Tong);Brad Porter(Cobot)反方引用公司走向公众视野;来自现任机器人高管的首次重要反方报道
2026 年 4 月 7 日Beyond World Models 博文产品Andy Zeng(作者)技术差异化:公开解释从零训练与 VLA 微调的区别
2026 年 5 月 29 日第 2 轮完成(官方日期)融资$400M;$2B 估值Radical Ventures(领投)、8VC、USV、Hanabi Capital、Norwest;天使 Fei-Fei Li、Eric Yuan、Bin Lin、Naval Ravikant估值较第 1 轮提升五倍;总融资 > $500M;数据飞轮正在形成
2026 年 6 月 4 日第 2 轮公开宣布融资$400M;$2B 估值Generalist AI 团队公告;The Robot Report、SiliconANGLE高调市场信号;加速招聘和数据引擎扩展
2026 年 6 月 4 日Physical AI 愿景博客发布产品Generalist AI 团队阐明飞轮:模型 → 有用的物理工作 → 业务数据 → 下一代模型
2026 年持续GEN-1 早期访问伙伴计划活跃规模化精选行业伙伴(未具名)商业进展正在形成;客户发现阶段进行中

确切成立日期(月 / 日)和注册细节未公开披露。第 1 轮完成日期来自 Tracxn。第 2 轮完成与公告日期分别来自 Tracxn 和官方博客。截至 runDate,未报告不利事件(诉讼、监管行动、裁员、领导层离职)。

[CO001, CO016, CO017, CO020, CO024, CO027]
FO001: Generalist AI — 公司里程碑时间线

2024 年至 2026 年 6 月的关键创立、融资、产品和规模化里程碑。

[CO001, CO016, CO017, CO020, CO027, CO028]

1.6 图表

Chapter 02

02市场分析

2.1 市场边界与分类

Generalist AI 的定位是只做软件的具身基础模型供应商,也就是运行在第三方机器人硬件之上的「认知智能层」。这个定位把市场边界划得很清楚:可服务支出是 AI 模型授权、云端和边缘推理、微调服务,以及与物理机器人部署绑定的软件订阅;不是机器人硬件本体、传感器、执行器、夹爪,也不是围绕它们的传统自动化资本设备。 纳入范围的支出包括 API / 云端推理费、基础模型授权、平台订阅收入,以及面向特定机器人形态或任务的下游微调服务。相邻但不完全属于范围内的市场包括机器人仿真平台(NVIDIA Isaac Sim、Mujoco)、在集成 AI 驱动机器人编排时的仓库管理软件(WMS)、Robot-as-a-Service(RaaS)交付平台,以及机器人中间件(ROS/ROS2 工业发行版 2026 年估值约 $0.8B)。排除在外的是硬件子系统本身——机械臂(6-DOF、7-DOF、人形)、传感器、摄像头、末端执行器和气动系统——以及工业 PLC、NC 控制软件和传统任务专用机器人编程(例如 ABB RAPID、KUKA KRL 脚本)。 Generalist AI 必须替代的现状方案包括:(1)任务专用 ROS/ROS2 编程和运动规划库,每个任务都需要数月定制工程;(2)机器人 OEM 的厂商专属嵌入式 AI 栈,例如 NVIDIA Isaac、ABB Omnicore、Yaskawa DX200 AI 和 Fanuc AI 软件套件;(3)面向结构化抓取和放置的传统示教器编程;(4)长尾非结构化任务中的人力。基础模型路线要替代这些方案,价值主张在于它能快速适配新任务(GEN-1 约需 1 小时任务专属数据,而定制编程通常需要数周或数月),并具备跨机体可移植性,免去按机器人平台重新工程化的需求。 IFR 将机器人中的 AI 与自主性列为 2026 年最重要的行业趋势,指出从基于规则的自动化转向智能、自我演进系统,正在制造业和服务业中「让具身 AI 成为主流」。这使 Generalist AI 所在技术类别更像结构性行业拐点,而不是小众子赛道。 [CM001, CM002, CM003, CM004, CM005, CM006]

市场定义表
细分市场 / 类别纳入支出排除支出买方 / 付款方与 Generalist AI 的相关性
具身 AI 软件基础模型 API 授权、微调服务、推理订阅机器人硬件本体、传感器、执行器制造业 OEM、物流运营商、机器人硬件 OEM核心 TAM — 直接收入模型
AI 机器人平台(广义)云推理、AI SDK / 中间件、车队管理软件机械部件、末端执行器、输送系统企业自动化团队、系统集成商广义 TAM — 与 NVIDIA Isaac、ABB AI 存在竞争重叠
仓储 / 物流自动化机器人拣选 / 分拣 AI、AMR 车队智能、带 AI 层的 WMS人工输送线、条码扫描器、非 AI AGV3PL 运营商、电商履约中心目标垂直 — 首个商业部署细分
工业制造自动化机械臂 AI 软件、视觉引导装配、自适应焊接 AI传统 NC 程序、示教器逻辑、固定自动化产线汽车 OEM、电子制造厂、合同制造商第二垂直 — 已演示汽车配套拣配、装配任务
实验室自动化液体处理 AI、样本制备机器人智能、实验室信息学 AI物理仪器硬件(移液器、离心机)制药、生物技术、CRO新兴垂直 — 灵巧操作高度适用
机器人硬件 OEM 授权打包进或叠加到第三方机器人平台的 AI 软件机械臂硬件、伺服电机、控制板OEM 买方:Fanuc、ABB、Boston Dynamics、Figure AI、1X Technologies平台 / OEM 授权 — 规模化关键 GTM 渠道
排除:机器人硬件N/A所有硬件组件(机械臂、关节、摄像头、夹爪、电池)N/A — 资本开支买方范围外 — Generalist AI 设计上硬件无关

市场边界由 Generalist AI 对自身纯软件具身基础模型提供商的明确定位定义。纳入支出代表潜在授权和 API 收入。排除支出代表 Generalist AI 不销售、也不从中取得收入的硬件资本开支。邻接领域(仿真、RaaS 平台、WMS)未来可能被打包,但在当前产品阶段不是主要收入目标。

[CM001, CM002, CM003, CM005]
FM004: 采用漏斗或价值链地图

制造业或物流买方旅程中,AI 机器人软件平台从初步认知到生产规模化的采购和部署步骤。

时间线估算基于 Mordor Intelligence、IFR 报告和 NeuroForge 商业化分析的行业基准。Generalist AI 尚未披露自己的部署周期数据。

[CM031, CM032, CM039, CM040]

2.2 市场规模——TAM、SAM 与相互矛盾的视角

多套分析机构规模框架并存,但边界定义彼此不兼容,因此无法从公开来源验证单一权威 TAM 数字。截至 2026 年中,市场主要有三种视角。 最宽的相关视角是「机器人中的 AI」。Grand View Research 测算该市场 2025 年为 $20.4B,预计到 2033 年达到 $182.7B,CAGR 为 32.0%。该口径覆盖所有机器人类别中的硬件组件(AI 加速器、边缘处理器)、软件平台和服务。其中,软件板块预计增长最快,CAGR 超过 33%。 中间视角是「AI 机器人软件」。Intel Market Research 测算该市场 2026 年为 $14.18B,到 2034 年达到 $38.76B,CAGR 为 15.3%。这个范围最接近 Generalist AI 的收入模式:算法、API、推理框架,以及云端机器人操作平台。关键可比对象包括 NVIDIA Isaac、ABB AI Software 和新兴基础模型 API。 最窄、也最直接可比的视角是「具身 AI」。Grand View Research(经 Research and Markets)测算 2026 年为 $6.5B,预计到 2033 年达到 $67.6B,CAGR 为 39.7%。这个口径涵盖能够感知、推理和行动的物理 AI 系统,最贴近 Generalist AI 的产品类别。物流和供应链是增长最快的终端使用板块,CAGR 为 42.2%;北美占 2025 年市场的 35.6%。 底层硬件市场提供了规模锚点。Mordor Intelligence 估算,工业机器人总市场(硬件加集成)2026 年为 $54.28B,到 2031 年为 $94.38B,CAGR 为 11.7%;IFR 则报告,单看机器人安装资本价值,2026 年为 $16.7B。仓储机器人子赛道 2026 年为 $7.35B,到 2034 年为 $25.41B,CAGR 为 16.8%;人形机器人市场 2026 年为 $8.3B,预计到 2030 年 CAGR 为 47.1%。实验室机器人市场为 $2.64B。 对于 SAM 和 SOM,公开来源没有给出「基础模型授权用于跨机体灵巧操作」的细分估算。Generalist AI 的初始 SOM 受其早期试用计划约束;收入和伙伴数量均未披露。 [CM007, CM008, CM009, CM010, CM011, CM012]

TAM/SAM/SOM 或规模测算视角表
发布方年份地区市场 / 价值($B)CAGR / 预测方法论置信度局限
Grand View Research2025/2033全球机器人 AI:$20.4B(2025)→ $182.7B(2033)32.0% CAGR(2026–2033)自下而上分层;硬件 + 软件 + 服务包含硬件 / 边缘计算;不是纯软件 TAM
发布方:Intel Market Research (IMR)2026/2034全球AI 机器人软件:$14.18B(2026)→ $38.76B(2034)15.3% CAGR(2026–2034)软件范围分层;云 + 边缘 AI 平台范围定义未独立验证;存在供应商定义风险
发布方:Grand View Research / R&M2026/2033全球具身 AI:$6.5B(2026)→ $67.6B(2033)39.7% CAGR(2026–2033)具身 AI 系统(感知 / 推理 / 行动);软件占比较高最窄的已发布视角;直接访问受阻;范围包含 Physical AI 硬件
Mordor Intelligence2026/2031全球工业机器人总市场:$54.28B(2026)→ $94.38B(2031)11.7% CAGR(2026–2031)出货量 × 系统价格 × 集成深度;硬件占主导硬件占主导;软件收入未单独拆出
IFR2026全球工业机器人安装额:$16.7B(2026)N/A(年报)机器人单元安装资本价值;不含软件 / 服务仅为资本开支口径;不含集成、软件、经常性费用
Grand View Research2023/2030全球仓储自动化:$19.23B(2023)→ $59.52B(2030)18.7% CAGR(2024–2030)AS/RS、机器人、输送机、WMS;覆盖宽口径仓储技术栈包含非 AI 硬件;未披露纯软件占比
Fortune Business Insights2026/2034全球仓储机器人:$7.35B(2026)→ $25.41B(2034)16.8% CAGR(2026–2034)用于仓储运营的 AMR、AGV、关节机械臂硬件比重高;AI 软件溢价未拆出
发布方:Research and Markets2026/2030全球人形机器人:$8.3B(2026)→ $39B(2030)47.1% CAGR(2026–2030)人形平台的硬件 + 嵌入式 AI硬件比重高;可服务 AI 软件占比未拆出

没有已发布来源给出专门针对“跨本体灵巧操作基础模型授权”的 SAM 或 SOM。上述所有数字都是宽口径市场估计,需要调整边界才能推导纯软件、基础模型特定收入。多组估计相互矛盾,反映的是口径边界不确定,不是数据错误。

[CM007, CM008, CM009, CM010, CM011, CM012]
FM001: 市场规模测算视角

四层嵌套视角展示从最宽的 AI 机器人市场到最窄的直接可比具身 AI 市场的层级。数值为 2026 年估算;任何层级都未单独发布纯软件收入占比。

本图是一组规模测算视角,不是严格的 TAM-SAM-SOM 级联。每一层采用不同分析机构的口径定义和方法。包含硬件的估算中,软件收入占比未单独发布;GVR/R&M 的具身 AI($6.5B)数值最接近 Generalist AI 的特定收入类别,但其中也包含一些硬件相邻组件。

[CM009, CM010, CM011, CM044]
FM002: 市场估算区间

使用最窄公开的具身 AI 口径作为低位、最宽泛的「机器人中的 AI」口径作为高位,给出 AI 机器人软件与具身 AI 市场的 2026 年低 / 基准 / 高估算。所有数值单位为十亿美元。

中点为算术中心或分析师区间中点。具身 AI 低位($6.5B)和 AI 机器人软件高位($14.18B)来源彼此独立;中点仅作示意。宽区间反映分析机构对边界的分歧,不是数据错误。不要把这些行相加——它们衡量的是重叠市场。

[CM009, CM010, CM011, CM013]

2.3 买方分层与采购动态

具身 AI 软件买方横跨五个不同群体,它们的采购结构、工作流场景和采用触发点各不相同。理解这些分层很重要,因为预算所有者、回本周期和集成成本差异很大,会直接影响任何 AI 机器人软件供应商必须采用的商业模式。 Tier-1 制造和汽车 OEM 终端用户(装配工厂、合同制造商、电子晶圆厂)历来是工业机器人的主力买方,2025 年占工业机器人需求的 35.86%(仅汽车行业)。这些买方通过多年采购周期用资本预算购买,需要认证集成商支持,并按节拍时间、正常运行时间和总拥有成本评估。制药和医疗制造到 2031 年 CAGR 最高,为 13.52%。 物流、仓储和 3PL 运营商是短期增长最快的买方群体。预计 2026 年电商板块将占仓储机器人市场的 47.21%。这些买方越来越多采用 Robot-as-a-Service(RaaS)模式,把资本支出转成运营支出,降低采用摩擦。Amazon 宣布 $1B 仓库自动化投资,以及 Figure AI 在 BMW Spartanburg 的部署——Figure 02 在 1,250+ 运行小时内搬运超过 90,000 个零件——代表了该群体扩大 AI 驱动机器人应用意愿的前沿。 机器人硬件 OEM 制造商(ABB、Fanuc、KUKA、Boston Dynamics、1X Technologies、Figure AI)构成一类独立的软件买方:它们把智能嵌入机器人平台,以差异化硬件产品,并回应客户对跨任务泛化的需求。这些买方会把基础模型视为自有嵌入式 AI 栈的潜在平台替代品,并通过 R&D 预算采购。 制药、生物技术和诊断领域的实验室自动化买方支撑了 2026 年 $2.64B 的实验室机器人市场。这些买方对 GMP 合规、可重复性和监管可追溯性尤其敏感,因此采用第三方 AI 层会更谨慎;但任务复杂度和人力成本较高,一旦采用,也可能更持久。 承接高混合、低批量生产的合同制造商和 3PL,是正在出现的可服务群体。在这类场景中,快速任务适配相对昂贵定制编程的价值最高,但集成支持仍是瓶颈。 [CM020, CM021, CM022, CM023, CM024, CM025]

细分市场 / 买方地图
细分市场买方角色用户角色付费方 / 预算工作流预算负责人采用触发因素
制造业 / 汽车 OEM制造副总裁 / 自动化总监工艺工程师、自动化工程师资本预算(CAPEX)装配、配套拣配、焊接、质检工厂年度资本计划获批劳动力短缺、关税驱动回流、产线灵活性
物流 / 仓储 / 3PL运营副总裁 / 配送负责人仓库经理、现场主管运营预算(通过 RaaS 计入 OPEX)拣选、包装、分拣、货到人运营副总裁或供应链总监电商增长、劳动力稀缺、履约速度 SLA
机器人硬件 OEM工程副总裁 / CTO机器人软件工程师、产品团队研发预算将 AI 能力接入机器人平台CTO / 产品副总裁客户要求任务泛化;竞争差异化
实验室自动化 / 生命科学实验室运营负责人 / 研发副总裁实验室技师、研究科学家研发 / 资本预算液体处理、样本制备、高通量筛选实验室主任、带监管签核的采购生产率要求、GMP 合规、劳动力成本
合同制造商 / 多 SKU总经理 / 运营副总裁现场工程师、产线主管运营预算多 SKU 装配、包装、配套拣配工厂总经理或运营财务SKU 激增、客户要求更快换线
政府 / 国防(新兴)项目经理 / 采购官技术专家政府合同预算危险环境巡检、弹药处理联邦合同授权机构安全、降低人员风险、战略能力

买方和付费方角色基于典型工业采购模式和公开案例研究;Generalist AI 尚未披露具名客户或采购条款。政府 / 国防板块仍在早期,没有确认的 Generalist AI 合作。RaaS 会加速物流行业采用;制造业 OEM 仍沿用 CAPEX 模式。

[CM020, CM022, CM025, CM027]
FM003: 买方 / 细分市场地图

按细分市场评估关键采购标准。单元格显示在当前技术成熟度和市场条件下采用的有利程度(正向 / 中性 / 警示)。

单元格为有证据支撑的序数判断,基于 IFR 报告、分析机构结论和行业案例研究。没有来源支撑的基准时,不分配数值分。

[CM020, CM021, CM022, CM025, CM026, CM027]

2.4 增长驱动与采用约束

五股结构性力量正在拉动机器人自动化市场向前,尤其可能支撑 AI 机器人软件跑出高于 GDP 的增长。 劳动力稀缺是最主要的需求信号。IFR 记录,2025 年德国工厂有 420,000 个技术工种空缺,美国到 2030 年预计缺少 2.1M 名制造业工人。美国指定区域内合格自动化支出可享 30% 联邦税收抵免,韩国在 2025 年把小制造商补贴翻倍,二者都直接缩短自动化回本周期。中国「中国制造 2025」计划到 2026 年为机器人投入 CNY 180B(约 $25.2B)。回流投资又叠加了一层需求:2024–2026 年,美国宣布 $47B 工厂投资,其中大多数都把机器人列为成本具备竞争力的本土生产前提。 机器人回本周期已经压缩。工业机器人的中位回本周期现在约为 1.3 年(16 个月),协作机器人可短至 6–18 个月。这种 ROI 压缩,一部分来自硬件成本下降(协作机器人入门价格约 $10,000),一部分来自 AI 带来的灵活性——GEN-1 适配新任务只需 1 小时,而传统编程需要数周,边际任务自动化第一次在经济上可行。 采用约束很大,也尚未完全解决。前沿资本强度极高:打造有竞争力的人形机器人平台,估计需要 Tesla Optimus 在 2022–2024 年投入 $3–4B R&D;Figure AI 据称每年烧掉 $200–300M。「筷子问题」——需要触觉与力反馈一体化的精细运动操作——对多数商业系统仍未解决。物理数据稀缺限制基础模型扩展:不同于文本 AI,物理操作数据必须通过机器人或可穿戴交互采集,不能从互联网上抓取。集成挑战也很实质:遗留工厂 PLC、MES 系统和 ERP 平台,让每个单元的典型安装开销达到 $40,000–80,000。安全和责任框架正在演进(ISO 9283、ISO/TS 15066),但截至 2026 年 6 月,任何主要司法辖区都尚未制定 AI 专属的具身 AI 责任监管。边缘算力是硬件瓶颈:VLA 模型需要专用端侧推理硬件,因为数据中心级算力会给实时机器人控制带来不可接受的延迟。 [CM028, CM029, CM030, CM031, CM032, CM033]

增长驱动因素与约束表
驱动因素 / 约束方向时点对 Generalist AI 的影响尽调追问
制造和物流劳动力短缺驱动(+)结构性 / 多年扩大可触达市场;降低买方自动化投入的 ROI 门槛核验前三大目标垂直行业的具体劳动力成本指标
政府自动化补贴(美国、中国、德国、韩国)驱动(+)2025–2028缩短客户回本周期;降低前期资本开支阻力把补贴项目映射到具体客户所在地和资格条件
电商履约需求增长驱动(+)结构性仓储 / 3PL 是 AI 机器人增长最快的买方群体找出正在履约场景试点基础模型的具体 3PL 客户
回流 / 近岸化投资周期驱动(+)2024–2028新建工厂默认采用自动化,减少传统系统集成摩擦评估愿意采用 AI 原生机器人的新建工业客户管线
机器人硬件成本下降(协作机器人入门价约 $10K)驱动(+)当前 / 加速压低硬件门槛;差异化转向软件层跟踪协作机器人 ASP 走势;评估硬件商品化是否会增强 Generalist AI 定价权
基础模型快速适配任务(约 1 小时)驱动(+)当前(GEN-1)让边际任务自动化具备经济性;压缩编程成本在公司可控演示之外的客户部署中核验 1 小时说法
研发资本强度(前沿人形机器人需 $3–4B)约束(−)当前 / 持续抬高竞争门槛;利好资金充足的在位者;模型质量若进入平台期则有风险确认烧钱速度,以及达到下一个前沿里程碑所需资本
'Chopstick problem':精细运动操作约束(−)当前 / 部分未解限制可触达任务;压低高灵巧度垂直行业 SAM用行业阈值对 GEN-1 的灵巧操作任务做基准测试
物理数据瓶颈约束(−)当前 / 多年基础模型质量受数据限制;数据采集需要重资本投入评估数据采集速度、质量标准和环境多样性
传统系统集成摩擦(PLC、MES、ERP)约束(−)当前 / 持续每个工位增加 $40K–$80K 集成成本;拉长销售周期确认认证集成商供给,以及集成工具路线图
安全与责任框架不成熟约束(−)当前 / 改善中拉长企业风险审批;可能把部署限制在非安全关键工位跟踪 ISO 9283 / ISO TS 15066 合规态势;识别首个认证部署
边缘计算硬件瓶颈约束(−)当前 / 2026–2027VLA 模型需要专用端侧推理;芯片供应预计偏紧评估硬件合作伙伴,以及端侧推理性能路线图

驱动 / 约束的时点基于 IFR 行业报告和分析师共识估计。Generalist AI 未披露单位经济、单次部署成本或获客成本数据。影响判断为方向性评估;尽调追问并非穷尽清单。

[CM028, CM029, CM030, CM031, CM032, CM033]

2.5 规模缺口、矛盾证据与尽调问题

具身 AI 软件的市场分析存在三类重要证据缺口,投资人和尽调团队应将其视为未解决问题。 第一,市场边界不一致是结构性问题。三种主要规模视角——$20.4B 的「机器人中的 AI」(GVR)、$14.18B 的「AI 机器人软件」(IMR)和 $6.5B 的「具身 AI」(GVR)——不仅规模不同,范围定义、地理覆盖和方法论也不同。这不是同一个市场从窄到宽的切片,而是不同分析师用不同边界规则衡量重叠现象。没有任何单一已发布数字,代表只做软件的具身基础模型供应商可直接获取的收入池。 第二,跨机体基础模型细分市场没有 SAM / SOM。「授权一个基础模型,使其在多样第三方机器人硬件上执行灵巧操作」这个子市场,并未被 IFR、Mordor、GVR 或其他公开可得来源作为独立类别跟踪。最接近的代理口径(2026 年 $6.5B 的具身 AI 软件)仍包含硬件相邻组件。若从这些数字推导 SAM,必须对软件收入占比和可服务机器人类别做出多项未经验证的假设。 第三,竞争融资强度能为估值提供市场信号,却不能提供收入清晰度。Physical Intelligence(pi.ai)2024 年以 $2.4B 估值融资 $400M,并据称在 2026 年洽谈 $1B 后续融资。QubitTool 分析预计,全球具身 AI 融资管线在 2026 年将超过 $20B。这些资本投入验证了市场命题,但没有解决规模或时点的不确定性。 [CM044, CM045, CM046, CM019]

2.6 图表

Chapter 03

03竞争格局

3.1 竞争框架与市场结构

物理 AI 基础模型市场可以分成四类竞争原型,每一类都代表一种捕获机器人技术栈智能层的不同策略。第一类是纯软件 / 智能层公司,不制造硬件,只打造通用机器人大脑:Generalist AI、Physical Intelligence 和 Skild AI 是主要例子。这些公司围绕模型泛化、数据规模、硬件无关性,以及把单一训练模型分发到多样机器人形态的能力竞争。第二类是全栈人形集成商,同时打造硬件本体和专有 AI 大脑:Figure AI、Boston Dynamics(Atlas)、1X Technologies 和 Agility Robotics 是例子。这些玩家掌控硬件-软件接口,可以让模型和机体一起迭代,但分发天然受限于自有硬件队伍。第三类是既有平台玩家——拥有相邻算力、数据或分发资产,并向机器人延伸的科技公司:Google DeepMind 和 NVIDIA。Google 提供网页级多模态预训练和商业 API;NVIDIA 提供主导性的 GPU 栈,并把开源 GR00T 模型作为亏本引流产品来推动硬件采用。第四类是现状方案和内部自建:大规模人工劳动、经典工业自动化(ABB、FANUC、KUKA),以及 Amazon(前 Covariant 创始人)和 Tesla(Optimus)的垂直整合内部 AI 团队。Generalist AI 在第一类中最直接对标 Physical Intelligence 和 Skild AI;长期看,它还面临 NVIDIA 通过向所有开发者免费分发开源 VLA 而商品化模型层的颠覆风险。Figure AI 和 Boston Dynamics 等全栈集成商,既可能成为 Generalist 模型的潜在客户,也会争夺企业部署预算。 [CP001, CP002, CP003, CP004, CP005]

FP001: 竞争定位图 — 融资额 vs 商业牵引力

使用 0–10 序数评分,把主要机器人基础模型竞争对手放在双轴图上:x 轴为累计融资 / 资源规模(0=极少,10=无限);y 轴为有记录的商业牵引力(0=商业化前,10=已验证的大规模企业收入)。

[CP001, CP007, CP008, CP009, CP011, CP033]

3.2 主要竞争对手画像

Physical Intelligence(π.ai)由 Sergey Levine、Chelsea Finn、Karol Hausman 等前 Google Brain 和 Berkeley 机器人研究者于 2024 年 3 月创立,是 Generalist AI 意识形态上最接近的同类公司。其 π0 / π0.5 模型是 Vision-Language-Action(VLA)系统,基于 7–8 个机器人平台上的 10,000+ 小时示范训练。截至 2026 年 3 月,Physical Intelligence 正在洽谈以约 $11B 估值融资约 $1B,Founders Fund 和 Lightspeed 参与讨论;这较 2025 年 11 月的 $5.6B 估值翻倍。不同于 Generalist AI,Physical Intelligence 已通过 openpi 仓库开源模型权重,建立了学术开发者社区,但也提高了商品化风险。两家公司都处于收入前阶段;Physical Intelligence 未公开宣布商业部署或定价。 Skild AI 于 2023 年创立,总部位于 Pittsburgh,2026 年初完成由 SoftBank 和 NVIDIA 领投的 $1.4B Series C,估值 $14B。Skild 报告称,2025 年商业发布后数月内收入约 $30M,是唯一披露达到这一规模商业牵引力的纯软件竞争对手。2026 年 4 月,Skild 收购 Zebra Technologies 的 Robotics Automation 业务,包括 Symmetry Fulfillment 编排平台,从纯 AI 模型公司转向整合人形机器人、移动机器人、机械臂和机器狗的完整仓库自动化方案。这笔收购给了 Skild 企业分发基础设施,而 Generalist AI 尚不具备。Skild 的「全机体」Skild Brain 模型声称具备跨机体泛化,数据集号称「比大多数竞争对手大 1,000 倍」;但这一说法尚未被公开基准独立验证。 Figure AI 由 Brett Adcock 于 2022 年创立,总部位于 Sunnyvale;2025 年 9 月由 Parkway Venture Capital 领投、Brookfield、NVIDIA、Microsoft、OpenAI Startup Fund 和 Jeff Bezos 参投的 Series C 后,累计融资约 $1.9B,投后估值 $39B。其 Figure 03 人形机器人运行 Helix,一个三层 VLA 系统(S0 全身控制、S1 视觉运动、S2 语义推理),完全在嵌入式 GPU 上运行,无需云连接。Figure 在 BMW Spartanburg 的部署——1,250+ 运行小时内搬运 90,000+ 个零件,并贡献到 30,000+ 辆车生产中——是所有竞争对手中记录最充分的商业人形机器人部署。Figure 是争夺企业合同的全栈竞争者,不只是模型同类。 Google DeepMind 通过 2026 年发布的 Gemini Robotics 1.5 和 ER 1.6 模型带来不对称竞争资源。Gemini Robotics ER 1.6 是面向开发者的云端可访问 VLA API,通过早期试用 / 等候名单开放;2026 年中宣布的新端侧变体,允许对延迟敏感的工业应用做本地推理。Google 的竞争优势包括比任何初创公司大几个数量级的网页级多模态预训练数据、Alphabet 算力基础设施,以及与 Boston Dynamics、Apptronik、Agility Robotics 的既有 OEM 伙伴关系。Gemini Robotics 尚未公开定价。Boston Dynamics 在 CES 2026 商业发布电动 Atlas 人形机器人,2026 年全部产能已承诺给 Hyundai 的 RMAC 和 Google DeepMind。Atlas 集成了 Google DeepMind AI 基础模型和 Boston Dynamics 的 Orbit 软件(MES / WMS 集成)。估计单价为 $150,000–$420,000。 NVIDIA 把自己定位为整个机器人生态的卖铲人基础设施。其 GR00T N 系列(N1 至 N1.7,其中 N1.6 在 CES 2026 发布)开源且对所有开发者免费,通过 Hugging Face 分发。NVIDIA 的商业模式是免费提供模型层,通过 GPU 硬件销售(Jetson Thor、DGX)和 Cosmos 仿真订阅变现。GR00T N1.6 已被 NEURA Robotics、Humanoid、Franka Robotics 采用,并集成 HuggingFace LeRobot。Amazon 通过 2024 年收购 Covariant 创始人(Pieter Abbeel、Peter Chen、Rocky Duan)团队人才并取得非独占 IP 许可,正在为其 750,000+ 机器人队伍建设内部机器人 AI 能力。Covariant 在 2025 年 2 月融资 $100M,并继续独立运营,但人才流失削弱了竞争位置。OpenAI 在 2026 年成立专门机器人部门;此前它曾投资 Figure AI 和 Physical Intelligence,是一个入场较晚但资源充足的基础模型类别新玩家。 [CP006, CP007, CP008, CP009, CP010, CP011]

竞争对手画像表
竞争对手类别成立时间融资(累计)估值收入(2025–26)旗舰模型硬件
Generalist AI纯软件 / 智能层2024$510M(约 $110M 种子轮 + $400M Series B 轮)$2B(2026 年 6 月)未披露(仅早期访问)GEN-1(原生基础模型,从零训练,500K 小时数据)硬件无关(无自研硬件)
Physical Intelligence纯软件 / 智能层2024 年 3 月约 $2.07B(含据称 2026 年 3 月洽谈中的约 $1B 新一轮融资)$5.6B(2025 年 11 月);目标 $11B(2026 年 3 月)披露收入 $0(商业化前)π0 / π0.5(VLA;PaliGemma 3B + 300M 动作专家;开源 openpi)硬件无关(无自研硬件)
Skild AI纯软件 / 智能层(+ 收购 Zebra 后的编排能力)2023约 $1.7B(含 $1.4B Series C 轮)$14B(2026 年初)约 $30M ARR(2025 年商业化启动)Skild Brain(全本体、跨本体分层 VLA)硬件无关;收购后拥有 Zebra AMR 车队
Figure AI全栈人形机器人(硬件 + AI)2022~$1.9B$39B(2025 年 9 月)未披露(BMW 商业部署)Helix(3 层 S0/S1/S2 VLA;端侧运行,无云依赖)Figure 02 / Figure 03 人形机器人(自研)
Google DeepMind在位科技平台 / 云2010(DeepMind);机器人业务始于 2022N/A(Alphabet 子公司)N/A(上市公司)N/A(补贴型研究)Gemini Robotics 1.5 / ER 1.6(云端 VLA + 端侧变体,2026)硬件无关(API + 端侧)
NVIDIA平台在位者(GPU + 开放模型)1993(机器人推进始于 2022)N/A(上市公司)N/A(上市公司)N/A(GPU 硬件收入)GR00T N1.6(开源 VLA;Cosmos Transfer/Predict/Reason 2)硬件无关(模型免费;硬件 = Jetson Thor、DGX)
Boston Dynamics全栈人形机器人(硬件 + 通过 Google DeepMind 接入 AI)1992(Hyundai 控股)N/A(Hyundai 子公司)N/A(子公司)未披露Atlas(Gemini Robotics 集成;56-DOF;IP67;自换电池)Atlas 电动人形机器人;Spot 四足机器人;Stretch 搬箱机器人

融资 / 估值数字来自截至 2026 年 6 月最新报道或公开披露轮次。Physical Intelligence 的 $11B 估值来自 2026 年 3 月据报道的洽谈,并非已关闭轮次。收入数字来自媒体报道;多数竞争对手未公开披露 ARR。Generalist AI 约 $510M 融资包括公开报道的 $110M 种子轮(2025 年 6 月)和 $400M Series B 轮(2026 年 6 月)。

[CP001, CP006, CP007, CP008, CP009, CP011]
FP002: 功能与能力对比矩阵

能力评分矩阵,从六个维度比较 Generalist AI 与五个主要竞争对手。分数为 0–10 序数分(越高越强)。所有分数均为基于公开证据的分析师评估;没有独立基准在同一硬件上直接比较这些模型。

0–10 序数尺度;尚无公开的独立跨模型基准在相同任务上比较 GEN-1、π0.5、Skild Brain 和 Gemini Robotics。分数反映截至 2026 年 6 月,基于公开新闻、模型论文和公司官方声明作出的分析师推断。

[CP019, CP024, CP025, CP026, CP033, CP034]

3.3 能力、定价与分发对比

跨机体泛化——无需针对每台机器人重新训练,就能在任意机器人形态上运行——是三家纯软件竞争对手(Generalist、Physical Intelligence、Skild)共同声称的核心差异点。Generalist 的 GEN-1 在架构上做差异化:它不是微调 VLM 骨干网络,而是基于自有物理交互数据从零训练,并明确否定 VLA 和世界模型范式,认为二者有限制。GEN-1 的 99% 成功率里程碑,记录在六项具体工业任务上,只使用 1 小时机器人专属适配数据。Physical Intelligence 的 π0 是基于机器人远程操作数据集中 10K+ 小时示范训练的 VLA;GEN-1 的预训练数据集来自人类执行物理任务时佩戴的传感器,而不是机器人远程操作。Generalist 称,在相同箱体上,GEN-1 完成箱体组装需 12.1 秒,π0 需 34 秒,速度优势为 2.8x;但没有独立第三方基准验证这一比较。Skild AI 的数据集优势说法(「大 1,000x」)也未被独立基准验证。 竞争格局中的定价大多未披露。Generalist AI 尚未公布 GEN-1 公开定价,只运营早期试用伙伴计划。Physical Intelligence 也没有公开定价。Skild AI 的企业交易未公开披露。Google DeepMind 未公布 Gemini Robotics API 访问定价。NVIDIA GR00T 免费(开源)。Boston Dynamics 的 Atlas 估计每台 $150,000–$420,000,取决于队伍配置。Figure AI 采用 Robot-as-a-Service 模式,但未披露单机定价。工业自动化既有厂商(ABB、FANUC、KUKA)销售软硬件打包方案,混合利润率为 30–50%,但其软件不是通用 AI。 分发能力是最关键的结构性战场。Skild AI 2026 年 4 月收购 Zebra 的 Robotics Automation 业务,拿到了经实战验证的仓储机器人平台、企业 WMS 集成,以及 Generalist AI 尚无立足点的物流客户入口。Figure AI 的 BMW Spartanburg 部署和经 Boston Dynamics 形成的 Hyundai 分发协议,代表深度汽车制造伙伴关系。NVIDIA 的 2M 机器人开发者生态,以及与 13M AI 开发者所在 HuggingFace 的集成,提供了任何初创公司都无法仅靠直销匹敌的分发触达。Generalist AI 的分发完全依赖早期试用伙伴计划;它没有公开宣布已署名客户或部署伙伴,这是相较所有已有商业牵引力竞争对手最重大的商业化落地缺口。 [CP019, CP020, CP021, CP022, CP023, CP024]

功能与能力矩阵
能力Generalist AI(GEN-1)Physical Intelligence(π0.5)Skild AI(Skild Brain)Google DeepMind(Gemini Robotics ER 1.6,机器人模型)NVIDIA(GR00T N1.6)
跨本体泛化强(硬件无关,新机器人 1 小时适配;据称可凭预训练实现零样本)强(10+ 机器人平台;openpi 跨本体;VLA 骨干)强(全本体架构;分层大脑;据称数据集扩大 1000x)强(云 API 支持任意机器人;高级智能体规划)中到强(开源;早期采用者:1X、NEURA、Boston Dynamics、Agility)
操作深度 / 灵巧度强(GEN-1:配套拣配、折叠、打包;200+ 次重复中达到 99%;据称可即兴处理)强(π0:叠衣、装配、多步骤灵巧任务;流匹配动作专家)中(仓储级;公开材料中操作能力不是主要差异点)强(Gemini Robotics-ER:空间 3D 理解;试点中的灵巧操作)中(N1.6 面向人形全身;操作能力来自合作伙伴数据集)
商业收入进展未知(仅早期访问伙伴;未披露收入)未披露(商业化前;未发布收入)约 $30M ARR(2025 年商业化启动;收购 Zebra 增加企业管线)未披露(早期访问 / 候补 API;机器人业务无公开商业收入)N/A(免费开源;NVIDIA 变现硬件而非模型收入)
分发能力低(仅伙伴计划;无具名客户或 OEM 协议)低(未披露商业合作;仍在从研究走向产品)中到高(收购 Zebra 带来 AMR 车队 + 企业 WMS 集成)高(Alphabet 覆盖面;OEM 伙伴:Boston Dynamics、Apptronik、Agility Robotics)很高(2M 开发者生态;HuggingFace 集成;OEM 伙伴网络)
数据飞轮强度中(500K 小时自有人类穿戴数据;由于尚无商业部署,部署飞轮还未在规模上验证)中(10K+ 小时遥操作;openpi 社区贡献;无企业部署)高(Zebra AMR 车队带来规模化企业物流数据;$30M ARR 部署数据)很高(Alphabet 互联网规模数据;来自 100+ 机器人本体的 RT-X 开放数据集)高(Cosmos 大规模仿真;Open X-Embodiment 贡献者;Isaac Lab-Arena)
开放生态 / 开发者采用低(无开源模型;仅早期访问)高(openpi 权重开源;学术社区基于 π0 架构建设)低(自研模型;无公开权重)低到中(Gemini Robotics API 需排队;开放访问有限)很高(GR00T N1.6 通过 Hugging Face 完全开源;LeRobot 集成;2M 开发者)

所有能力评估都是分析师基于公开来源推断。尚无公开独立基准在相同硬件和任务上直接比较 GEN-1、π0.5 与 Skild Brain。标为“未知”的单元反映缺少公开证据,不代表缺少能力。

[CP019, CP020, CP022, CP023, CP024, CP026]
定价与包装对比
竞争对手定价模式标价 / 合同包含能力定价透明度对 Generalist AI 的关键影响
Generalist AI(GEN-1)早期访问伙伴计划未披露GEN-1 模型访问;用 1 小时机器人数据完成适配无(没有公开定价)定价不透明可能拖慢企业采用;没有公开价格信号,买方难以横向比较
Physical Intelligence(π0.5)按机器人收费的 SaaS(计划中,未确认)估计每台机器人每年 $5–15K(AI2Work 估计;未验证)π0.5 模型权重;openpi 开源访问(有限付费层待定)很低(未发布官方定价)两家公司商业化前都定价不透明;谁先公开,谁可能立下市场锚
Skild AI(Skild Brain + Zebra,平台)企业 SaaS + 编排平台未公开披露;估计企业年合同 $200K–$5M+Skild Brain 模型;Symmetry 编排;AMR 车队协调低(企业议价)收购 Zebra 后可打包硬件 + 软件定价,纯模型玩家难以匹配
Figure AI(Helix)机器人即服务(RaaS)硬件 + AI未披露;第三方估计硬件约 $150K–$300K / 台Figure 03 机器人;Helix 端侧 AI;OTA 更新;遥操作支持低(未发布价目表)全栈硬件护城河;纯模型玩家无法在 RaaS 上直接竞争
Google DeepMind(Gemini Robotics ER 1.6,机器人模型)云 API + 企业试点合同未公开定价;基础 Gemini API 起价 $0.075/1M tokens(非机器人层)VLA API;端侧模型(2026 年中);MuJoCo 模拟器;SDK;早期访问计划很低(机器人定价未发布;企业议价)Google 可用 Gemini API 收入补贴机器人模型,因此能低于成本定价
NVIDIA(GR00T N1.6)免费 / 开源模型模型权重 $0(Apache / CC 许可)通过 Hugging Face 提供 GR00T N1.6 VLA 权重;Cosmos 仿真(单独订阅);Jetson Thor 硬件完全透明(模型免费;硬件按标准 NVIDIA 目录定价)免费模型拉低所有付费机器人基础模型产品的付费意愿;形成关键的下行定价压力
现状(人工劳动 + 工业自动化)CapEx 硬件 + 劳动力 + 维护工业机器人手臂 ~$30K–$150K CapEx;美国制造业每名 FTE 劳动力 $30K–$80K/year确定性、高吞吐的单任务自动化;没有泛化能力高(ABB、KUKA、FANUC、Universal Robots 有目录价)Generalist 必须证明相对传统自动化和低成本替代方案有 ROI

除 NVIDIA GR00T 和传统自动化外,所有定价数据均为估算或媒体报道。企业 AI 机器人合同通常通过谈判确定,且不公开披露。传统自动化定价来自制造商目录和行业平均水平。

[CP020, CP021, CP022, CP023, CP026]

3.4 护城河耐久性与竞争风险

Generalist AI 的主要护城河主张有三点:(1)数据引擎规模——来自可穿戴传感器的 500K+ 小时自有物理交互数据,公司称这是全球同类最大数据集,形成了纯模型竞争对手难以轻易复制的训练数据优势;(2)架构新颖性——不微调 VLM、而是从零训练,使其区别于 π0 的 PaliGemma 骨干路线,尽管相对竞争对手的真实泛化性能差距尚未被独立验证;(3)硬件无关性——跨机体设计让 Generalist 有机会成为机器人行业潜在的「AI OS」,以最小适配部署到机械臂、人形机器人和移动平台。数据飞轮逻辑是,早期试用伙伴部署会生成真实世界交互数据,喂给后续模型迭代,使数据集优势随时间复利。 不过,Generalist AI 的竞争位置面临四类结构性风险,分别威胁其每一项护城河主张。第一,NVIDIA 的开源 GR00T N 系列给了每个开发者一个免费、生产级 VLA;如果 GR00T 在 N1.5、N1.6、N1.7 迭代过程中追平 GEN-2 或 GEN-3 的能力,模型层会商品化,定价权也会崩塌。第二,Physical Intelligence 的开源 openpi 权重,让任何公司都能对照 Generalist 最接近的同类做基准测试并在其上构建,降低新竞争者进入成本,压缩差异化窗口。第三,Skild AI 收购 Zebra 后,一个资金充足的直接竞争对手拥有了企业分发,而 Generalist AI 现在必须用尚未建成的企业销售基础设施与之竞争。第四,Google DeepMind 的不对称研究资源——近乎无限的算力、网页级数据,以及与每个主要机器人 OEM 的伙伴关系——意味着 Generalist AI 取得的任何能力差距都可能只是暂时的。Boston Dynamics 2026 年量产机器人的 Hyundai + Google DeepMind 独占管线,至少一年内关闭了 Generalist 可用的最具商业可信度硬件分发路径。最不利的信号,是没有任何公开宣布的商业客户:三家顶级纯软件竞争对手(PI、Skild、Generalist)最终都在争夺企业机器人 AI 预算,而 Skild 的 $30M ARR 和 Zebra 分发基础设施给了它结构性部署优势;随着物流客户锁定自动化栈,这一优势会越来越难被撬动。 [CP027, CP028, CP029, CP030, CP031, CP032]

护城河耐久性与竞争风险登记表
护城河主张威胁威胁严重度缓解措施或尽调问题
专有物理交互数据集(500K hrs 可穿戴传感器数据;公司称全球最大)NVIDIA EgoScale 和 Cosmos 合成数据生成提供了替代扩展路径;Physical Intelligence RT-X 开放数据集汇聚了 100+ 种机器人形态;规模主张尚未得到独立验证通过独立技术审计验证该数据集相对 RT-X 和 NVIDIA 合成数据的独特性;确认可穿戴传感器竞争者 (micro1、Scale AI)无法复制该数据集
架构新颖性(从零训练;不是 VLA 微调;不是世界模型)机器学习架构创新的竞争寿命历来较短;NVIDIA GR00T N1.6 使用类似的从零训练逻辑;PI 的 openpi 让外界可以反向推断其架构选择确认 GEN-2 架构路线图;评估“从零训练”主张能否防住资源雄厚的在位者能力跃迁
硬件无关性(跨形态;一个模型适配所有机器人本体)垂直整合的竞争者(Figure AI、Boston Dynamics)通过面向硬件的调优获得更强的软硬件协同; 如果硬件锁定主导市场,模型无关玩家会丢掉部署渠道评估哪些机器人 OEM 已加入早期接入计划;确认 Generalist AI 未因竞争者排他协议而被关键人形机器人硬件平台挡在门外
数据飞轮(真实部署产生训练数据,优势复利)飞轮需要大规模商业部署才能转起来;公司未披露客户,飞轮尚未启动;Skild AI 收购 Zebra 让竞争者拥有活跃的企业部署飞轮,而 Generalist AI 的飞轮仍停留在理论层面核实早期接入伙伴数量和数据贡献条款;确认数据飞轮协议机制;评估每个早期接入伙伴会产生多少新的机器人小时数据
研究员 / 创始人履历(Pete Florence 共同发明 VLA;团队来自 DeepMind / Boston Dynamics)关键人集中——公司叙事、融资和技术可信度都高度绑定 Florence 和 Zeng;在当前 $2B 估值下,Google、Microsoft 或 Amazon 以 $10B+ 出价挖走团队的风险相对目标回报呈非对称低位确认创始人留任机制(股权归属、竞业限制);评估第二梯队研究团队厚度;评估治理结构和独立董事会监督

威胁严重度是基于截至 June 2026 公开证据的分析师判断。Generalist AI 尚未披露早期接入伙伴产生的机器人小时数、客户名单或合同排他条款。

[CP027, CP028, CP029, CP030, CP031]
FP003: Generalist AI 竞争就绪度 KPI 记分卡

Generalist AI 在七个关键护城河和商业维度上的竞争就绪度评分。0–10 序数尺度;越高越强。分数为截至 2026 年 6 月基于公开证据的分析师评估。

[CP032, CP033, CP034, CP035]

3.5 图表

Chapter 04

04财务分析

4.1 收入模式与定价

Generalist AI 明确是一家纯软件公司——不制造机器人——并把自己定位为跨形态的「智能层」,可服务于人形机械臂、工业 6-DoF 机械臂和移动平台。商业主张是:机器人运营方(制造商、系统集成商、企业物流运营商)购买 GEN-1 基础模型的使用权,再用约 1 小时的机器人专属数据,把模型微调到自己的硬件和任务上。从结构上看,这更像云端 AI API 或单次部署的软件许可证,而不是硬件销售,因此毛利率画像更接近软件 SaaS,而非机器人硬件厂商。 公司尚未公开任何定价。唯一确认的变现信号,是 2026 年 4 月启动的 GEN-1 早期访问伙伴计划——访问权只授予被选中的行业伙伴,公司没有披露客户名称或合同条款。2026 年 6 月融资公告称,「数据飞轮开始成形:真实企业正在生成任务数据,反哺后续模型迭代」,这暗示早期付费或合同式部署已经存在,但收入规模未知。 公开披露中最接近的定价参照是 Physical Intelligence。按 Sacra 分析师研究,该公司采用按机器人计费的 SaaS 模式,每台联网机器人每月 $300。如果 Generalist AI 采用类似订阅模式,每台机器人每年 $3,600 意味着要做到 $10M ARR,约需部署 2,800 台机器人——这既需要显著的企业采用,也需要硬件伙伴分销。Generalist AI 未确认这种定价;这里只是类比估算。更高价值的企业授权模式(按工厂或机器人类型收取固定年费)同样合理,并会显著改写单位经济。按机器人 SaaS 模式确认收入较直接;若采用定制企业模式,收入确认可能按里程碑推进,规模化后会增加复杂度。 [CI001, CI002, CI003, CI004, CI005, CI007]

收入来源摘要
收入来源机制单位 / 定价基础当前状态收入质量尽调问题
早期接入企业伙伴计划GEN-1 部署给选定伙伴;可能是数据共享或付费试点定制企业合同(条款未披露)截至 April 2026 仍活跃;未确认收入未知;仅有商业化前证据确认是否产生收入,还是仅换取数据访问;获取合同条款
按机器人计费的 SaaS / API 订阅(推断)云端托管模型 API 按动作调用;每台联网机器人按月收取经常性费用估算类比:~$300/robot/month(Physical Intelligence 可比)未确认;商业模式来自竞争者类比未知;未披露定价确认定价模型;获取客户名单、实际成交价和流失指标
企业场地或设施许可(推断)按设施或产品线覆盖机器人智能,收取年度或多年固定费用定制年度合同;批量折扣结构未确认;工业软件常见模式未知;若按里程碑确认收入,确认复杂度高判断是许可还是 SaaS;理解按里程碑与按时间确认的差异
数据飞轮参与(非现金)伙伴用任务数据换取模型访问;数据具有长期训练价值非货币价值交换;用数据换模型访问June 2026 官方公告提及;规模未知非收入;仅是战略资产量化数据共享条款;理解是否包含任何现金部分

收入来源 2 和 3 基于竞争者定价类比和公司定位推断,Generalist AI 未确认。收入来源 4 不涉及货币,但对数据飞轮论点重要。收入来源 2–4 的所有货币数字都是证据缺口。

[CI001, CI003, CI004, CI005, CI007, CI008]
定价与商业化
定价维度数值 / 状态基础(目录价 vs. 实际成交)来源或基础关键未知
Generalist AI 已发布定价未披露N/AGeneralist AI 官方招聘页和博客(无定价页)完整定价模型:按机器人、按席位,还是企业固定费用
Physical Intelligence 定价类比$300/robot/month 订阅目录价(Sacra 分析师)Sacra 分析师研究,已独立确认未为 Generalist 确认;可能因细分市场或机器人形态不同而有重大差异
Skild AI 收入代理(行业可比)2025 年报道 ARR ~$30M报道 ARR;不是目录价ahr.so 行业分析支撑该数字的部署基数和定价模型未确认
1,000 台机器人隐含收入(PI 定价类比)~$3.6M/year估算计算:1,000 台机器人 × $300 × 12分析师估算;非 Generalist 披露未计入折扣、试点、企业定制费率或不同定价模型

Generalist AI 没有公开定价。Physical Intelligence 和 Skild AI 数字只是竞争者 / 类比代理。 任何使用这些类比的收入估算置信度都低,应视为示意性边界,而非预测。

[CI002, CI008, CI010, CI011, CI035]
FI001: 收入模型桥 — 从客户活动到毛利

在推断的按机器人 SaaS / 企业授权模式下,Generalist AI 如何把机器人操作方活动转化为收入和毛利。

收入触发节点使用 Physical Intelligence $300/机器人/月 SaaS 定价作为类比;Generalist AI 尚未确认任何定价。COGS 百分比参考软件 SaaS 基准,并按推理算力作调整。本模型仅作示意;实际收入机制尚未确认。

[CI003, CI004, CI005, CI008, CI034]

4.2 GTM 动作与商业牵引信号

从招聘信息和投资人评论看,Generalist AI 的商业化打法仍处早期,偏高接触企业销售:截至 2026 年 6 月,唯一对外商业岗位是一个「Applied AI & Partnerships」职位,相比研究团队,商业团队很轻。这符合深科技 AI 实验室种子阶段常见的 GTM 模式:创始团队亲自维护关键关系,早期客户更像战略伙伴而非批量账户,交易周期长,且常与联合数据采集或集成项目绑定。 8VC 投资备忘录把「显著的早期商业牵引」列为形成投资判断的信号;Spark Capital 的 Fraser Kelton 也提到机器人规模化中的早期商业验证。两家投资人都未确认客户名称或收入指标。公司 2026 年 6 月博客明确称数据飞轮已经启动——真实企业为下一代模型生成任务数据——这意味着一定规模的创收或数据共享合同关系已经存在。未披露客户名称,符合企业机器人部署的保密要求(工厂自动化项目通常很晚才披露,甚至不披露),但也同样可能意味着仍处商业化前阶段。 具身基础模型平台的销售周期参照很长:典型工业自动化资格认证周期为 6 到 18 个月,包括硬件兼容性验证、安全认证和生产试运行。相对初始合同金额,这类周期的获客成本结构性偏高,因此回本期是关键指标,但无法验证。公司未公开 CAC、LTV 或回本期估算。 [CI006, CI024, CI025, CI026, CI028, CI029]

4.3 成本结构与资本强度

Generalist AI 的成本结构有三个主轴:(1)前沿模型训练算力,(2)物理数据引擎运营,(3)研究和工程人才。公司为 GEN-1 重建了分布式训练基础设施,以「把 PB 级物理交互数据作为一等公民来支持」——这种表述意味着可观的 ML 基础设施资本开支,以及持续的云端或自建集群支出。行业基准显示,一个 10B+ 参数、基于 500,000 小时视频与动作数据训练的前沿机器人基础模型,单次训练仅算力成本就可能达到 $10M–$100M;公司计划连续迭代模型,训练摊销因此会成为重要的经常性成本项。 数据引擎运营——在全球以每周 10,000+ 小时的速度运行可穿戴「data hands」设备——需要持续的人力和设备物流成本,这不是一次性支出。不同于文本抓取,物理交互数据需要人类操作者、现场实验室管理者、设备维护和数据标注基础设施。NVIDIA NVentures 作为继续加码的战略投资方,既说明公司依赖 GPU,也可能带来优先硬件供应因素,但外界尚未确认任何算力价格优惠。 公司约 70 名员工主要来自 Google DeepMind、OpenAI 和 Boston Dynamics,平均总薪酬很可能显著高于行业中位数,符合前沿 AI 研究人才的经济规律。行业估算显示,类似具身 AI 初创公司(50–100 人、算力密集)月度烧钱额为 $5M–$15M。这只是估算——实际数字未披露。按每月 $10M 的烧钱中值计算,新融资 $400M 意味着自 2026 年 6 月起可支撑约 40 个月,也就是约到 2029 年末,尚未扣除任何收入抵消。 [CI018, CI019, CI020, CI021, CI022, CI023]

单位经济
指标数值 / 估算置信度重要性尽调问题
每台机器人年收入(目录价)~$3,600(PI SaaS 类比)低——仅为竞争者代理决定达到收入盈亏平衡所需的机队规模确认 Generalist AI 定价模型;获取实际合同 ARR
客户获取成本(CAC)未披露Unknown高接触企业销售 + 集成 → 每个客户可能 $50K–$500K+获取每笔交易的商务团队成本、试点投入和集成支持成本
LTV / CAC 比率未披露Unknown可扩展 GTM 的关键指标;若机队规模化前流失率高,则为负需要每客户 ARR、流失率和 CAC;三者均未公开
毛利率(模型交付)未披露;按软件 API 交付类比,规模化后推断为 60–80%低——仅为软件类比;计算转嫁会抬高 COGS决定最终盈利天花板;推理计算成本是关键变量获取 COGS 拆分:计算、数据出站、推理、每机器人小时客户成功
月度烧钱速度未披露;70 人前沿 AI 实验室的行业估算为 $5M–$15M/month低——仅为行业基准决定现金跑道和下一轮融资时点;最关键的财务规划输入获取月度 P&L:计算、人力、数据运营、各类别 capex
每训练小时数据收集成本未披露;推断低于遥操作($50–$200/hr vs 遥操作 $500–$2,000/hr)低——基于已发布的 data-hands 方法与竞争者方法推断直接影响每代模型成本;相对重遥操作同行的竞争性成本优势获取每小时训练数据采集的数据收集运营成本拆分
每代模型训练计算成本未披露;GEN-1 规模下,行业基准为每次运行 $10M–$100M低——基于 10B+ 参数视频-动作数据模型的行业基准每一代新模型都会重复发生;以当前机队收入规模无法摊销获取 GEN-0 和 GEN-1 训练运行的实际 H100/A100-day 消耗和单位成本

所有指标数值要么未披露(只能通过私有证据获得),要么使用行业基准和竞争者类比估算。Generalist AI 未发布任何单位经济数据。所有估算的置信度都应保持低位。

[CI008, CI018, CI020, CI023, CI034, CI035]
FI002: 单位经济模型桥 — 成本与收入驱动因素

关键成本和收入输入节点推动 Generalist AI 的单位经济模型和贡献利润率;在缺少精确数字时使用定性证据节点。

所有成本数字都是行业基准估算,不是公司披露。收入数字使用竞争对手类比定价。盈亏平衡机队测算假设无收入抵消,并使用中位烧钱额($10M/月)和 PI 类比定价($300/机器人/月)。

[CI018, CI019, CI020, CI023, CI034, CI035]

4.4 公开财务指标与证据缺口

Generalist AI 是私营公司,未披露收入、ARR、客户数、毛利率或单位级财务指标。公司公开披露只限于:(a)带投后估值的融资公告,(b)模型发布博客,(c)Ashby 和公司招聘页上的招聘活动。外界无法取得投资者关系材料、审计财务报表或二级交易数据。 现有间接牵引信号都只是定性:8VC 所称「早期商业牵引」、公司自己称真实企业参与的数据飞轮「开始成形」,以及 Robot Report 引用的 2026 年 4 月「早期访问伙伴现在可以获得模型访问权」。这些都不能构成已确认收入。 可获得的公开信息,与财务承销所需数据之间差距很大。收入、ARR、毛利率、流失率、客户集中度和烧钱率,全部只能靠非公开证据确认,需要管理层直接披露或审计财务记录来解决。在这些缺口关闭之前,任何 Generalist AI 财务模型都只能建立在商业模式类比(Physical Intelligence、其他具身 AI SaaS 公司)之上,而不是已披露的实际数据。 [CI027, CI028, CI029, CI033, CI034]

公开财务缺口
缺失指标对分析的影响尽调路径
收入 / ARR没有任何收入锚点,无法评估收入质量、增长率或基于倍数的估值向管理层索取当前季度 ARR、按收入来源拆分的收入和过去 12 个月增长
按类别拆分的月度烧钱没有确认后的烧钱,无法计算现金跑道、资本效率或经营杠杆索取按行项目拆分的 P&L:计算、人力、数据运营、设施、capex;对照管理层估算审计
毛利率不理解推理成本与收入的关系,就无法评估可扩展性天花板或单位经济健康度索取 COGS 拆分:云计算、第三方数据、客户成功、每机器人支持成本 vs. ARR
客户数量和管线没有任何客户数据,无法评估集中度风险、流失或 CAC;单一大客户可能主导早期 ARR向管理层索取实名客户名单(NDA 下)、合同金额、续约状态,以及按阶段和细分市场拆分的管线
定价模型和合同条款不知道实际定价结构,就无法验证 Physical Intelligence SaaS 类比或企业许可类比索取标准 MSA、定价表和样本合同条款;理解定价是按用量、订阅还是固定费用

所有缺口都只能通过私有证据填补;公开来源无法取得。填补这些缺口需要直接接触管理层并审阅 NDA 保护的数据室。任何公开可得代理都不能替代直接披露。

[CI002, CI027, CI028, CI033]

4.5 资本充足性与融资依赖

如公司概览章节所述,Generalist AI 于 2025 年 3 月完成约 $140M 首轮融资,又于 2026 年 6 月完成 $400M 第二轮融资,公开披露累计融资超过 $500M。2026 年 6 月这轮融资于 2026 年 6 月 4 日宣布;若假设签约到拨款通常间隔 30–60 天,现金投入大约从 2026 年 6 月中旬开始。公司未公开披露任何债务融资、可转债、信贷额度或项目融资义务;资产负债表看起来完全由股权融资支撑。 公司明确说明,$400M 新资金用途是:打造下一代模型、扩张物理数据引擎、扩充算力和训练基础设施,并推进产业伙伴关系。这完全是收入前的基础设施投入;没有迹象显示资金用于收购或股份回购。隐含的下一轮触发点,是后续模型代际(GEN-2)达到足以支撑估值再上台阶的性能阈值——GEN-0 到 GEN-1 间隔约 5 个月,GEN-1 到 Series B 间隔 2 个月,节奏很快。 按每月 $10M 的烧钱中值估算,在零收入抵消的情况下,Generalist AI 会在交割后约 40 个月耗尽 $400M(2029 年末)。若每月 $15M,可支撑时间缩短至约 27 个月(约 2028 年 9 月)。若每月 $5M(70 人前沿 AI 实验室的乐观下限),可支撑时间延长至约 80 个月。Physical Intelligence 估值约为 Generalist AI 的 5×、团队更大,投资人称其模式需要连续大额注资来维持数据和算力规模——这意味着 Generalist AI 面临结构性融资依赖,只有当收入足以实质性抵消成本基数时才会缓解。按 Physical Intelligence 的 $300/机器人/月定价,若要在每月 $10M 烧钱水平上打平,需要约 33,333 台机器人同时付费——以任何合理的当前采用速度看,这种机队部署规模都还要多年。 $2B 估值(为 2025 年 3 月 $440M 的 5 倍)来自私下谈判;尚无独立第三方估值发布。可比公司 Physical Intelligence 据称正以 $11B 估值融资——Generalist AI 大约只有这家竞争对手估值的五分之一,同时声称拥有架构差异化和更早进入市场的跨形态通用性。 [CI012, CI013, CI014, CI015, CI016, CI017]

资本充足性
项目数值 / 估算置信度备注尽调问题
手头现金(Series B 后估算)~$400M–$500M(新融资 + 2025 轮剩余资金)中——基于已披露融资规模;首轮资金已被烧钱消耗首轮 $140M 于 March 2025 完成;从 March 2025 到 June 2026(~15 months)的烧钱会显著消耗这部分资金确认最近一个财季末的实际现金余额
月度烧钱速度(估算)$5M–$15M / 月低——行业基准;未确认70 人前沿 AI 团队,计算和数据运营投入密集;见单位经济章节向管理层获取按类别拆分的月度烧钱
自 June 2026 起的现金跑道估算~27–80 months(取决于烧钱情景,对应 September 2028 – January 2033)低——区间基于未确认的烧钱估算高情景($15M/month):~27 months。中情景($10M/month):~40 months。低情景($5M/month):~80 months。未假设收入抵消。收入建模需要确认定价和客户管线
计划资金用途(June 2026 融资)下一代模型开发;物理数据引擎扩容;计算和训练基础设施扩张;行业伙伴关系高——官方 June 2026 博客公告中说明完全用于收入前投资;未显示有收购、回购或分红部分确认各类别资本配置比例;理解招聘扩张计划
下一轮融资触发点(推断)GEN-2 发布和商业牵引里程碑,可能在 12–24 months 内低——基于 GEN-0→GEN-1 的 5-month 节奏,以及模型→融资轮的 2-month 节奏推断模式:模型发布 → 商业牵引信号 → 以更高估值融资确认内部里程碑目标;理解董事会层面的资本计划
债务 / 信贷额度未公开披露中——所有主要媒体报道和投资者公告均无相关证据没有任何新闻稿、SEC 等效文件或新闻报道提到债务、信贷额度或项目融资确认不存在未披露债务、可转债或供应商融资安排

手头现金基于融资规模减去估算烧钱得出;真实数字需要管理层确认。所有烧钱和现金跑道 数字均为低置信度的情景估算。本表是快照;数值会逐月变化。

[CI012, CI013, CI015, CI016, CI020]
FI003: 财务估算区间 — Generalist AI 关键情景

截至 2026 年 6 月,有来源支撑或由基准推导的关键财务输入上下界。所有数值均为估算;Generalist AI 均未确认。

烧钱率和算力区间来自分析师对具身 AI 初创公司的评论所给行业基准。ARR 上限具有推测性。Physical Intelligence 估值区间反映两个不同时期披露的数据点。Generalist AI 未确认任何财务数据;所有数值均为情景输入,不是预测。

[CI009, CI013, CI017, CI018, CI020, CI027]
FI004: 资本强度地图 — $400M 融资估算成本分配

对 2026 年 6 月 $400M 融资在已宣布资金用途类别中的示意分配,并给出估算情景拆分。所有金额均为基于所述优先事项和行业基准推导的估算。

本分配是基于公司所述资金用途优先事项(官方博客,2026 年 6 月)的示意资本分配。实际分配比例未披露;这些比例估算参照可比前沿 AI 实验室资本部署模式校准。为示意,合计等于 $400M。

[CI015, CI018, CI019, CI021]

4.6 财务结论

收入质量无法判断,因为没有已确认收入。截至 2026 年 6 月,公司处于早期访问伙伴阶段——按运营实质,相当于商业化前或测试阶段,而不是一个有公开证据支持的创收业务。任何风险投资定价,买的都是一个 $2B 未来态业务的期权,而不是当前期间的盈利索取权。 利润率路径:如果收入真正规模化,软件交付模式(云托管基础模型 API)结构上应能带来有吸引力的毛利率——成熟期可能达到 60–80% 或更高;但短期成本结构由训练算力、数据运营和顶尖研究薪酬主导,未来至少 24–36 个月贡献利润率仍会深度为负。每一代后续模型都会重置资本开支时钟;在数据飞轮产生足以抵消训练跑步机的自我维持商业收入之前,公司没有轻资产滑行阶段。 资本强度是这家公司的核心财务特征。Brad Porter(Cobot)指出的关键风险是运营层面的:「用巨量数据硬推一个并不完美的架构,非常昂贵,也未必能得到你想要的结果。」这是来自前 Amazon 机器人副总裁、且具备生产部署可信度的反方观点;它引入了一个可信情景:训练后续模型代际的成本,可能比收入飞轮产生抵消现金流的速度增长更快。Yann LeCun 的补充批评——动作 token 预测模型如果没有架构创新,无法仅靠规模实现泛化——则从结构上挑战了整个投资逻辑所依赖的数据扩展论点。 财务承销的四个尽调阻断点是:(1)当前 ARR,以及按客户分部和收入类型拆分的合同管线;(2)按类别拆分的真实月度烧钱额(算力、人员、数据运营、资本开支);(3)当前与预测规模下的毛利率画像,包括每机器人小时推理成本;(4)定价模型细节——按机器人订阅、企业场地许可证、按用量计费 API,或混合模式。没有这四项输入,就无法锚定可信的 DCF、可比公司或情景分析。 [CI007, CI030, CI031, CI032, CI033, CI034]

4.7 证据材料

Chapter 05

05产品与技术

5.1 产品定义与商业产品

Generalist AI 的核心商业产品是 GEN-1,一个可输出实时机器人动作的大型多模态基础模型。GEN-1 被定位为任意机器人的「智能层」或「认知大脑」,并不绑定专有硬件:它可跨多种机器人形态运行,包括 6 自由度(6DoF)工业机械臂、7DoF 协作机器人和 16+DoF 半人形系统。公司的商业主张是提供一个通用软件模型,让任何机器人硬件 OEM 或最终部署方都能集成,而不是出售完整机器人系统。 GEN-1 用公司所称的「掌握」界定商业可行性——三种能力的组合:可靠性(稳定完成任务)、速度(快到具备经济价值)和即兴智能(无需预定义响应即可从意外场景中恢复)。公司声称,GEN-1 已在简单物理任务上跨过这一门槛,成功率达到 99%,完成速度约为此前最先进水平的 3x,并展示了训练分布之外的恢复行为。 截至 2026 年 4 月 2 日,GEN-1 通过早期访问计划提供给被选中的行业伙伴。公司尚未公开披露定价、客户名称、SLA 承诺或正式支持模式。部署联系路径为 partnerships@generalistai.com。官方博客和投资人评论中提到的目标行业包括服装、制造、物流、汽车和电子。 [CE001, CE002, CE003, CE023, CE024, CE028]

产品模块 / 资产矩阵
模块 / 资产用户 / 买方状态 / 成熟度差异化尽调缺口
GEN-1 具身基础模型工业机器人运营方;OEM 集成商早期接入商业化(Apr 2026)99% 成功率;速度较 SOTA 快 3x;硬件无关;1-hr 适配未披露具名客户;生产部署条件未验证;无 API 文档
GEN-0 具身基础模型研究 / 试点评估方研究阶段基线(Nov 2025)首个机器人扩展律;Harmonic Reasoning;270K-hr 数据集已被 GEN-1 取代用于商业用途;不再面向新部署提供
物理数据引擎(data hands 设备)内部 R&D / 数据工厂伙伴已运行并在扩展500K+ hrs 自然手部灵巧数据 vs. 遥操作;全球网络设备规格、同意框架和隐私政策未公开披露
推理运行时(Harmonic Reasoning + paged attention)嵌入 GEN-1 部署生产(已嵌入)连续时间下同步感知与行动,无需 System 1/2 暂停架构专有;无外部技术论文;未披露推理延迟
后训练系统(SFT + RL + 多模态指导)GEN-1 部署团队生产(1-hr 适配)相比 GEN-0 数据效率提升 10x;从经验中 RL;多模态人工引导RL 技术未详细说明;具体训练协议未发布
GEN-1 早期接入计划选定行业伙伴(未具名)活跃 / 扩张中(Apr–Jun 2026)排他早期接入;真实部署正在形成数据飞轮无公开定价;无 SLA;无支持模型;无具名伙伴
下一代模型(GEN-1 后)未来工业部署方R&D / 已规划(无时间表)由 $400M June 2026 融资支持时间表、架构、目标任务和差异化未知

GEN-1 是唯一处于商业活跃状态的产品;其他条目要么是内部子系统、历史前代,要么是计划中的未来版本。截至 June 30, 2026,具名客户和生产部署尚未公开确认。

[CE001, CE002, CE003, CE007, CE011, CE015]
工作流 / 用例表
用户任务当前工作流公司解决方案可衡量收益限制
汽车零部件配套(汽车制造)人工劳动或固定自动化GEN-1 运行在 7DoF 机器人手臂上1+ hr 连续自主运行;99% 成功率(公司声称)需要特定机器人形态;受控环境条件
折叠 T 恤(服装 / 洗衣)人工劳动GEN-1 运行在机器人手臂上无人工干预连续折叠 86 件(公司演示)仅测试特定服装类型;未展示所有服装变体
服务机器人吸尘器(维护作业)人工清洁和维修GEN-1 运行在机器人手臂上无干预连续 200+ 个周期;99% 成功率需要特定产品类型;未确认可泛化到不同吸尘器型号
打包积木 / 套件物品(物流 / 仓储)人工包装线GEN-1 运行在机器人手臂上连续 1,800+ 个周期;速度较此前 SOTA ~3x(公司声称)未展示复杂或易碎的多物品订单;未披露设置时间
折叠纸箱(物流 / 制造)人工或固定机器GEN-1 运行在机器人手臂上每次折叠 ~12 sec vs 此前 SOTA ~34 sec(快 2.8x);99% 成功率超出测试类型的纸箱变体未验证;生产线集成成本未披露
将手机装入保护壳(电子)人工劳动GEN-1 运行在机器人手臂上无干预连续 100+ 个周期;每次装配 15.5 sec(比 GEN-0 快 2.8x)公开演示未说明手机型号变体和保护壳类型
适配新物理任务(任意行业)数月定制机器人编程或遥操作数据集采集在目标机器人上进行 1-hr GEN-1 微调跨形态;基础模型没有机器人数据;简单任务上 1 hr 达到 99%复杂任务、高精度工作或动态环境可能需要更多数据

所有用例均来自公司官方演示视频和博客文章;尚未确认独立复现或第三方在生产条件下做基准测试。成功率和速度均为公司声称,应视为最佳情形演示。

[CE016, CE017, CE018, CE020, CE021, CE024]
FE002: Generalist AI — 客户工作流 / 部署流程

从客户任务识别到机器人自主运行的端到端工作流,展示每次部署如何把数据回流到 Generalist AI 的物理数据引擎飞轮。

数据飞轮回流路径是公司截至 2026 年 6 月所述意图。截至报告日期,尚无运营数据被独立确认已从具名客户部署流入 Generalist AI 训练管线。

[CE001, CE002, CE018, CE023]

5.2 训练数据引擎与物理交互数据集

Generalist AI 的核心技术差异化,是其专有物理交互数据集,以及生成该数据集的数据引擎。数据集用公司称为「data hands」的腕戴或手持人体工学可穿戴设备采集——这类轻量设备能以接近自然的力反馈捕捉人类操作数据。不同于 Physical Intelligence 等竞争对手使用的遥操作装置,data hands 保留了自然的感觉运动闭环:短暂适应后,人类操作者不再「思考」,而是开始反应,因此轨迹捕捉到的是反射、微调和实时错误恢复,而非僵硬刻意的动作。 GEN-1 的预训练数据集中没有机器人数据:全部由 data hands 设备捕捉的人类操作构成。这是有意为之——公司认为,把机器人数据排除在预训练之外,可以形成更可泛化的感觉运动先验;随后用 1 小时机器人专属数据微调,才首次同时把模型适配到机器人的具体形态和目标任务。 到 GEN-0 发布时(2025 年 11 月),数据集已达到 270,000 小时真实世界物理交互;到 GEN-1 发布时(2026 年 4 月),增长至 500,000 小时以上。数据仍以每周超过 10,000 小时的速度增加。采集地点覆盖全球数千个家庭、仓库、工作场所和专门环境,包括面包店、自助洗衣店和工厂。全球硬件网络和数千台数据采集设备支撑运营。数据工厂伙伴贡献的数据被分为第 1 类(特定任务)、第 2 类(混合)和第 3 类(任意任务)模式,使公司能够 A/B 测试哪种数据混合最能提升预训练。处理基础设施包括 O(10K) 个核心用于持续多模态数据处理、多云合同、配备专线互联网的定制上传机器,以及每天训练可吸收 6.85 年操作经验的技术。 [CE004, CE005, CE006, CE007, CE008, CE009]

技术 / 运营架构表
层级 / 组件作用依赖风险
数据收集硬件(data hands 可穿戴设备)捕捉用于基础模型预训练的人类操作数据定制专有硬件;全球操作员劳动力;供应链物流设备召回、供应中断或操作员不可用可能拖慢数据集增长
物理预训练数据集(500K+ hrs)GEN-1 基础模型的传感-运动基础全球数千个采集点;多云存储;NVIDIA GPU 处理数据质量下降;新地域监管挑战;隐私诉讼
GEN-1 基础模型(~99% 从零训练)通用传感-运动先验;跨形态泛化专有架构;7B–10B+ 参数规模;大规模 GPU 计算单体从零设计在需要根本架构修订时会放大风险
Harmonic Reasoning 训练流程连续时间同步感知和行动,不需要暂停思考周期定制训练机制;异步流式 token 架构专有;机制主张未经过外部同行评审;架构细节未发布
后训练系统(SFT + RL + 多模态指导)针对任务的精通式微调(~1 hr 机器人数据)来自伙伴的机器人专用数据;多云计算;RL 稳定性技术RL 不稳定;伙伴提供的任务数据质量和数量会影响结果
推理运行时(paged attention、定制内核)在已部署机器人上实时预测动作NVIDIA GPU 硬件;定制编译内核代码推理延迟未披露;专有内核带来硬件集中风险
计算集群(多云)训练并服务所有模型多云合同;NVIDIA GPU 优先访问权NVIDIA 供应链风险;PB 级、数万核心规模下成本高;没有摆脱 NVIDIA 的路径
机器人伙伴硬件(第三方 OEM)GEN-1 推理的物理执行已演示 Universal Robots UR7e;原则上可适配任意 6/7/16DoF OEM 机器人Generalist AI 不制造硬件;集成质量取决于 OEM API

架构细节来自官方博客(GEN-0 和 GEN-1 技术文章)。逐层实现、延迟规格和安全架构没有公开披露。NVIDIA 依赖来自两轮战略投资和多云 GPU 需求推断。

[CE004, CE006, CE010, CE011, CE012, CE026]
FE001: Generalist AI — 产品架构栈

六层架构,从底部的全球数据采集基础设施到顶部的客户应用部署,展示物理交互数据如何扩展成可部署的具身基础模型。

[CE004, CE007, CE011, CE015, CE022, CE026]

5.3 模型架构:GEN-0、GEN-1 与 Harmonic Reasoning

GEN-1 的架构完全从零搭建——约 99% 参数重新训练,而不是在现有视觉语言模型(VLM)上微调。这是公司基于目标驱动研究哲学作出的有意选择:在积累足够真实世界物理数据后,Generalist AI 认为,完整掌控架构比改造现成 VLA 或世界模型,更能推动前沿进展。公司曾共同发明 VLA(PaLM-E、RT-2),也曾研究世界模型,但拒绝给自己贴标签,因为模型设计横跨多个技术类别。 基础架构创新是 Harmonic Reasoning,最早在 GEN-0 中引入,并在 GEN-1 中演进。Harmonic Reasoning 在感知 token 流和动作 token 流之间建立异步、连续时间的「harmonic」互动——让模型无需为物理过程暂停,就能边想边动。公司将其与 System 1/System 2 架构(Figure AI 的 Helix 使用)和推理时引导方法对比;后两者需要先停顿再行动的顺序循环,无法适配实时物理。在 GEN-0 阶段,扩展实验揭示了模型「固化」相变:低于 7B 参数的模型无法吸收大规模预训练数据,权重会冻结;7B+ 模型则继续提升。GEN-0 扩展至 10B+ 参数。GEN-1 在 GEN-0 基础上加入预训练改进、后训练技术(监督微调、强化学习、多模态人类引导)和新的推理时技术。GEN-1 相比 GEN-0 展示出 10x 数据效率,用少 10x 的任务专属数据取得相当的下游表现。定制 paged attention 内核支撑实时推理。物理常识——面向力、摩擦和不确定性的反应式闭环感觉运动智能——是大规模预训练产生的关键涌现属性,体现为公司并未显式训练却自然出现的恢复行为。 [CE011, CE012, CE013, CE015, CE019, CE025]

FE003: Generalist AI — 关键依赖地图

有向依赖图展示 Generalist AI 产品和模型开发所依赖的关键供应商、平台、数据来源和机器人合作伙伴,并突出集中风险。

对 NVIDIA 的依赖,来自 NVentures 连续两轮战略投资以及 GEN-0 博客中明确描述的多云 GPU 需求。除 Universal Robots(GTC 演示)外,其他机器人硬件 OEM 合作方截至 2026 年 6 月 30 日尚未公开具名。

[CE007, CE022, CE026, CE034]

5.4 部署、集成、硬件兼容性与路线图

GEN-1 的部署模式与硬件无关:一个基础模型通过约 1 小时机器人专属数据采集和微调,就能适配任意机器人形态。2026 年 3 月 GTC 现场演示在现实条件下验证了这一主张:Generalist AI 接受 Universal Robots 邀请,在其全新移动操作平台上演示 GEN-0(MiR 底盘上的 UR7e 机械臂,通过 Vention 框架安装)——这是一个此前并不存在的硬件平台——并在短短几天内稳定跑通,只使用波士顿和旧金山办公室的数据,没有使用 GTC 展馆本身的数据。GTC 首次运行表现与实验室表现一致。 目标机器人兼容范围覆盖 6DoF、7DoF 和 16+DoF 半人形平台。截至 2026 年 6 月,公司未公开 SDK、公开 API 文档或开发者门户。部署通过早期访问伙伴计划推进,入口是 partnerships@generalistai.com。GEN-1 还在微调过程中集成多模态人类引导,让人类操作者把模型行为导向期望结果——公司称之为具身智能对齐。 路线图主要从 2026 年 6 月融资公告推断:$400M 融资明确用于下一代模型开发、数据引擎扩张、算力和训练基础设施扩充,以及商业部署扩张。公司未披露具体 GEN-2 时间表、功能集或模型规模。公司更广义的路线图哲学是目标驱动:持续降低所需任务专属数据量,同时把成功率推高。近期里程碑——用约 1 小时数据达到 99%+ 成功率——已由 GEN-1 达成;下一阶段是更广的任务和环境泛化。 [CE002, CE014, CE018, CE022, CE023, CE032]

路线图 / 发布 / 开发阶段表
日期 / 阶段功能 / 里程碑状态含义来源
Sep 2025一次性 Lego 组装演示:机器人根据视觉观察复制新结构已演示(内部评估)物理常识和视觉理解的前置信号;Level 4 灵巧度信号公司博客:generalistai.com/blog/the-robots-build-now-too
Nov 4, 2025GEN-0 发布:首个机器人扩展定律;270K+ 小时数据集;Harmonic Reasoning;10B+ 参数已发布(公开)证明机器人领域存在扩展规律;为 GEN-1 打基础公司博客:generalistai.com/blog/gen-0
Jan 29, 2026Physical Commonsense 博客:从数据规模中涌现感知运动智能的框架已发布(观点文章)指向研发方向:把涌现的物理常识做成竞争护城河公司博客:generalistai.com/blog/physical-commonsense
Mar 24, 2026GTC 现场演示:Universal Robots UR7e + MiR(新机器人,数天内投入运行)已演示(公开现场演示)跨本体迁移速度得到公开验证;在约束条件下证实硬件无关主张公司博客:generalistai.com/blog/the-real-breakthrough-behind-our-gtc-demo
Apr 2, 2026GEN-1 发布:99% 成功率、约 3x 速度、1 小时适配、早期访问计划开放已发布(公开)公司声称达到商业可行门槛;数据飞轮启动;部署咨询入口开放公司博客:generalistai.com/blog/gen-1
Apr 7, 2026Beyond World Models & VLAs 博客:从零自研架构的理由已发布(技术)公开解释架构差异化;释放长期不依赖 VLM 生态的信号公司博客:generalistai.com/blog/beyond-world-models
Jun 4, 2026$400M 融资完成;GEN-1 合作伙伴计划运行中;下一代模型研发获得资金已宣布(公开)资金投向数据引擎扩容、算力扩张和下一代模型公司博客:generalistai.com/blog/accelerating-the-next-phase-of-physical-ai
未披露下一代模型(GEN-1 之后);覆盖更广任务和环境已规划;未披露时间表或规格由 2026 年 6 月融资支持;公司目标是逐步减少所需的任务特定数据公司博客:generalistai.com/blog/accelerating-the-next-phase-of-physical-ai

GEN-1 之后的路线图来自融资资金用途表述和公司博客推断。截至 2026 年 6 月 30 日,公司未公开发布正式产品路线图、版本编号、发布时间表或功能规格。

[CE013, CE016, CE022, CE023]

5.5 技术差异化、IP 与竞争位置

Generalist AI 的主要差异化建立在五个支柱上。第一,数据集规模和质量:公司自称,截至 2026 年 4 月,其 500,000+ 小时物理交互数据集是全球最大的真实世界操作数据集合,并通过专有低成本可穿戴系统采集,捕捉自然人类灵巧性,而不是缓慢僵硬的遥操作轨迹。第二,从零训练的架构:GEN-1 完全从零训练,而非在 VLM 上微调,让公司掌握完整架构控制权。第三,Harmonic Reasoning:公司的新型连续时间架构,允许同时感知和行动,不受顺序式「暂停再行动」循环的性能天花板限制。第四,扩展规律验证:GEN-0 是首个展示可预测机器人扩展规律的模型(预训练数据规模与下游表现之间呈幂律关系),为可预测改进提供路线图;没有同等数据规模的竞争对手很难复制。 与最接近且公开度最高的竞争对手 Physical Intelligence 的 pi-0 相比,GEN-1 折盒子用时 12 秒,pi-0 约 34 秒(快 2.8x);GEN-1 报告 99% 成功率,而 GEN-0 和 pi-0 的可比任务时长约为 34 秒。Generalist AI 的方法在预训练上与 pi-0 根本不同:GEN-1 预训练不使用机器人数据,pi-0 则使用大规模遥操作数据集。公司认为,这会在部署时带来更高数据效率,并更好泛化到新环境。GTC 演示验证了第二个差异化主张:GEN-0 能在几天内泛化到全新机器人类型,公司将这种能力描述为「机器人到场就能工作的未来」。 IP 细节、专利申请和专有数据权利框架都未公开披露。公司依赖商业秘密来保护训练流程和数据集,而不是通过专利披露。 [CE005, CE007, CE013, CE015, CE016, CE017]

5.6 信任、安全、对齐与质量控制

Generalist AI 已承认,具身基础模型存在特定的安全与对齐缺口。在 GEN-1 博客中,公司指出,GEN-1 的即兴智能既是优势(能从意外场景中自发恢复),也可能成为责任:涌现行为是具有真实后果的物理动作,训练分布之外的即兴发挥可能导致非预期机器人行为。公司称目标是改进对齐方法,以「精准引导」模型进入期望行为,但尚未发布正式对齐框架、安全协议或时间表。 GEN-1 的任务表现通过真实机器人上的内部 A/B 评估衡量,采用盲评分和闭环策略执行,GEN-0 博客对此有描述。这种验证仅限内部:公司未公开披露任何独立第三方对 99% 成功率主张的审计。具体基准任务集、评估环境和统计方法也未发布。GEN-1 也承认自身局限:并非所有任务都达到 99%+ 成功率,有些任务若要在现实部署中有用,还需要更高成功率。 物理数据采集运营的隐私和数据安全治理未公开披露。公司从多个地区的数千个家庭、仓库和工作场所收集视频与感觉运动数据,这一范围很可能触发 GDPR、CCPA 及同等隐私义务。公司未公开说明同意框架、数据主体权利、留存政策或监管指引。也未提及 SOC 2、ISO 27001 或同等信息安全认证。作为双用途物理 AI 基础模型,GEN-1 是否适用出口管制,公司也未公开回应。 [CE031, CE032, CE036, CE040]

信任 / 质量 / 合规表
控制 / 认证 / 指标状态范围缺口
任务成功率(99% 可靠性基准)公司声称;仅限内部测试受控条件下选定演示任务;具体任务由公司选择没有第三方基准或独立验证;任务选择和协议未发布
涌现行为对齐框架已承认缺口;推进中(据 GEN-1 博客)大规模预训练带来的即兴行为和意外行为未披露正式安全标准、已发布的对齐协议或第三方审计
数据同意 / 隐私框架未公开披露全球运营方在多个地区采集的家庭、工作场所和公开视频数据GDPR、CCPA 及同等框架很可能被触发;未披露公开政策、同意 UX 或 DPA
出口管制合规(ITAR / EAR)未披露面向机器人的 Physical AI 基础模型;可能具备国防 / 航空航天两用属性未说明 ITAR/EAR 是否适用;未见公开声明或监管法律顾问意见
系统安全与事件响应未披露通过早期访问伙伴计划落地的真实部署未披露安全升级路径、操作员告警协议或监控框架
数据安全 / 信息安全认证未披露来自私人住宅和工作场所的 PB 级操作视频未提及 SOC 2、ISO 27001 或同等认证;伙伴数据处理条款未知
模型性能验证(内部)盲评分内部 A/B 评估在真实机器人上闭环推出策略;GEN-0 博客有描述未做独立验证;评估协议未发布;可能存在选择偏差

所有信任与合规条目都来自公司博客披露,或来自公开披露的缺失。截至 2026 年 6 月 30 日,未发现第三方审计、认证或监管申报。该表是缺口评估,不是合规证明。

[CE031, CE032, CE036, CE040]
FE004: Generalist AI — 产品成熟度 / 能力评估矩阵

截至 2026 年 6 月,从七个关键能力维度评估 GEN-1,覆盖证据质量、成熟度判断和各维度的主要尽调缺口。

所有评估均基于截至 2026 年 6 月 30 日的公司自报和媒体报道。当时没有独立性能数据、客户部署或第三方审计。色调取值:positive=有证据, neutral=有部分证据,warning=证据有限,risk=已识别重大缺口。

[CE015, CE016, CE017, CE020, CE031, CE033]

5.7 证据材料

Chapter 06

06客户

6.1 客户群分层与目标市场

Generalist AI 对外宣称的商业目标,是任何使用物理机器人执行重复、灵巧任务的企业。公司不制造硬件;价值主张是一个可跨机器人形态运行的智能层。在这一模式中,买方要么是机器人硬件 OEM(授权模型并嵌入其平台),要么是企业运营方(把 GEN-1 集成到现有硬件)。付款方是企业或 OEM 被授权方;用户则是在工厂、仓库或物流现场部署机器人的运营者。 Generalist AI 官方沟通中明确点名的目标垂直行业包括服装、制造、物流、汽车和电子。这些正是简单、重复、灵巧任务常见的行业——折叠衣物、配套汽车零件、包装箱子、组装电子产品——劳动密集且自动化程度正在提高。GEN-1 博客和 2026 年 6 月 $400M 融资公告都称「真实企业」正在这些环境中生成训练数据,暗示这些细分市场存在一些早期合同关系。 第二类不同客户是数据工厂伙伴:部署「data hands」可穿戴采集设备,并向 Generalist AI 预训练数据集贡献操作数据的企业或运营者。这些伙伴可能通过数据共享安排获得补偿,而不是支付传统 SaaS 费用;其参与也不一定意味着他们正在生产环境中部署 GEN-1。数据贡献伙伴与付费 GEN-1 部署客户之间的区别很重要,但公司尚未公开澄清。 第三类是硬件 OEM 或平台伙伴关系,Universal Robots 在 NVIDIA GTC 2026 的合作提供了证据。UR 邀请 Generalist AI 在其全新移动操作平台上演示 GEN-0;这一关系显示,全球按出货量计最大的协作机器人制造商对其有商业兴趣,但并不能确认商业授权或部署协议。 从地域看,现有证据显示数据采集网络覆盖「全球数千个家庭、仓库、工作场所和专门环境」,但 GEN-1 销售目标似乎聚焦北美和欧洲工业运营商,这与公司在 San Mateo 和 Somerville 的存在一致。公司未在任何司法辖区披露具体客户名称、合同结构或客户数量。 [CU001, CU002, CU003, CU004, CU005, CU006]

客户细分表
细分市场买方 / 用户 / 付款方示例用例规模估计收入 / 战略价值证据质量证据缺口
企业制造运营 / 工程负责人(买方);产线操作员(用户);企业总部(付款方)汽车零部件配套、电子产品包装、消费品组装大型企业;未披露客户数潜在价值最高的细分市场;未披露收入低 — 间接(仅投资人背书)具名部署、合同条款、生产运行时间、安全认证
电商 / 物流运营 / 自动化团队(买方);仓库员工(用户);企业(付款方)装箱、包裹分拣、装卸大型和中端市场;未披露数量行业战略价值高;未披露合同低 — 间接(官方沟通中点名目标垂直)具名客户、试点还是生产状态、吞吐量指标
服装 / 纺织供应链 / 运营(买方);服装工人(用户);品牌或制造商(付款方)叠 T 恤、服装分拣、组装全球行业分散;没有数量已演示用例(叠 T 恤);未确认付费客户低 — 仅演示证据生产部署、客户名称、缺陷率、人工成本节省
数据铸造伙伴(训练数据)数据采集操作员(用户 / 贡献方);运营方(付款方 / 接收方)使用 data hands 设备采集真实物理交互GEN-1 博客称全球数千个站点;未披露实体名称战略价值:为模型预训练供料;是否产生收入不明低 — 官方说法,无独立确认数据伙伴收入模式、排他条款、数据所有权 / IP
硬件 OEM / 平台伙伴机器人制造商(伙伴);OEM 的企业客户(终端用户)向协作机器人 / 机械臂厂商跨硬件授权模型公开具名 1 家(Universal Robots,GTC 2026);其他未知战略价值:不靠直销也能放大分发;未披露条款中 — GTC 演示由双方公开记录商业授权条款、部署数量、排他性、收入分成

所有细分市场估计都从官方沟通和投资人表述推断。客户数量、收入或合同数据均未公开披露。Universal Robots 是唯一具名伙伴实体。

[CU001, CU002, CU003, CU007, CU008]
FU001: 客户旅程图 — 企业机器人运营方

一家企业制造或物流运营方评估 GEN-1 时的示意性采用路径。

[CU001, CU012, CU013]

6.2 伙伴证据与数据飞轮信号

最具体、可公开验证的伙伴证据,是 Universal Robots 邀请 Generalist AI 于 2026 年 3 月在 NVIDIA GTC 2026 演示 GEN-0。Universal Robots 是全球按出货量计最大的协作机器人制造商,为演示提供了其全新移动操作平台——MiR 移动底盘上的 UR7e 机械臂和 Vention 框架。Generalist AI 描述该活动的博客称,这套机器人硬件是新的,团队此前从未实际接触过;公司明知只有几天准备时间,仍答应演示。GEN-0 模型在数天内适配新硬件,并在会议全部开放时段连续运行。这是一家重要硬件 OEM 选择 Generalist AI 技术参加公开现场活动的具名、公开记录案例——是重要背书,但不是商业授权或部署协议。 NVIDIA GTC 2026 年 3 月新闻稿提到 Generalist AI「使用 Cosmos 探索生成合成数据」——这是 NVIDIA 生态内的技术合作,但相比 Skild AI(ABB Robotics、Universal Robots、Foxconn)或其他已被赋予伙伴角色且拥有正式部署关系的参与者,只是边缘性引用。Generalist AI 未被列入 NVIDIA「把物理 AI 带入真实世界」叙事中的主要伙伴,说明它仍处生态进入阶段。 2026 年 6 月融资公告明确写道:「一个飞轮已经开始成形:扩展机器人学习会创造更好的模型,更好的模型能完成更有用的物理工作,而真实企业的数据会驱动下一代更强模型。」其中「真实企业」意味着公司与实际工业运营方存在数据共享或合同关系——这与 GEN-1 技术材料描述的数据工厂伙伴计划一致。8VC 投资备忘录也用「显著的早期商业牵引」强化了这一点,称其为投资驱动因素之一。8VC 和 Generalist AI 都未确认这些关系的数量、身份或收入规模。 Spark Capital 的 Fraser Kelton 曾负责 OpenAI 产品商业化,他在投资相关表述中提到「机器人规模化中的早期商业验证」,说明投资人看到了公开不可见的真实客户互动证据。多个投资人围绕商业牵引给出证明,但没有任何公开参考客户;这种组合符合受 NDA 保护、面向战略工业伙伴的早期访问计划,也符合当前阶段企业机器人的标准做法。 [CU009, CU010, CU011, CU012, CU013, CU014]

客户增长与采用轨迹
指标数值日期来源置信度含义缺失分母
早期访问计划启动GEN-1 面向选定伙伴开放2026-04-02Generalist AI 官方博客(GEN-1 发布)计划存在;未披露伙伴数量或收入获准伙伴数量;筛选标准
投资人背书的商业牵引显著早期商业牵引(8VC);机器人扩展中的早期商业验证(Spark Capital)2026-068VC 投资备忘录;Forbes(Spark Capital 引述)存在一些真实商业活动;规模未披露收入、客户数、付费试点数量
数据飞轮信号真实企业生成任务数据(据融资公告)2026-06-04Generalist AI 博客(accelerating-the-next-phase)与工业运营方存在合同或数据共享关系贡献数据的企业数量、这些关系带来的收入
Universal Robots GTC 演示合作在新 UR7e + MiR 平台现场演示;展会期间连续运行2026-03Generalist AI GTC 演示博客;NVIDIA 投资人新闻稿具名硬件 OEM 已参与;不是商业授权协议UR 演示是否带来正式合作伙伴关系或授权协议
Applied AI and Partnerships 人员配置截至 2026 年 6 月,1 个开放职位(Applied AI & Partnerships)2026-06Ashby 招聘列表(jobs.ashbyhq.com/generalist)商业团队尚未规模化;意味着战略账户最多 3–10 个当前交易管线、交易阶段、预计成交时间

所有轨迹指标都是间接信号;公开来源没有确认收入、客户数量或部署规模指标。置信度反映来源可靠性,不代表商业成熟度。

[CU009, CU010, CU011, CU012, CU013]
具名客户证明表
实体细分市场部署 / 用例生产 vs. 试点结果 / 证据限制
Universal Robots (UR)硬件 OEM — 协作机器人GTC 2026 现场演示:GEN-0 在 UR7e + MiR 移动操作平台上完成精准装箱任务演示 / 展示(非生产部署)机器人数天内适配新硬件;在 NVIDIA GTC 持续运行;被称为 Generalist 首次公开演示不是商业部署;未披露授权或收入协议;UR 以技术伙伴身份邀请演示
数据铸造伙伴(未具名)覆盖制造、物流、餐饮服务、家庭环境等场景的运营方使用 data hands 可穿戴设备采集真实物理交互数据;截至 2026 年 6 月贡献 500,000+ 小时运营中(数据采集持续)GEN-1 和 GEN-0 官方博客确认数据铸造参与方;融资公告称其为「真实企业」实体未具名;关系是数据贡献,不是 GEN-1 部署;未披露伙伴收入模式
GEN-1 早期访问计划参与方(未具名)企业制造商、物流运营方和其他选定行业伙伴GEN-1 模型访问,用于试点或早期部署,与伙伴自有机器人集成试点 / 早期访问(状态未确认)计划于 2026 年 4 月 2 日启动;8VC 称有「显著早期商业牵引」;Spark Capital 提到「早期商业验证」未具名伙伴;未披露结果、SLA 或部署规模;投资人背书缺少独立验证

这是截至 2026 年 6 月本报告生成日公开具名或可识别客户 / 伙伴实体的完整列举。没有具名客户并不证明没有商业关系;保密协议和企业机器人领域惯例可以解释未披露。

[CU010, CU011, CU014, CU015, CU016, CU017]
FU002: 采用与部署漏斗

截至 2026 年 6 月,从可服务市场到已披露合作伙伴证据的示意性转化漏斗。

除顶层(100%)和已确认生产部署(0)外,所有值都是基于商务团队人数和计划年龄的示意性估算。没有实际漏斗数据公开可得。

[CU018, CU019, CU024]

6.3 采用轨迹与部署状态

GEN-1 自 2026 年 4 月 2 日起通过早期访问计划开放。截至 2026 年 6 月 30 日 runDate——上线约 89 天后——公司尚未公开发布具名客户部署、生产结果或独立性能验证。公司用于伙伴合作的联系路径只有一个邮箱地址(partnerships@generalistai.com),没有自助伙伴门户,没有公开 API 文档,也看不到开发者社区。 GEN-1 商业任务表现的证据来自官方发布博客:机器人完成了六类任务——连续 1 小时以上配套汽车零件、连续折叠 T 恤 86 次、连续维护扫地机器人 200 多次、包装积木 1,800 多次、折叠箱子 200 多次、包装手机 100 多次——平均成功率 99%。这些演示使用公司挑选的任务、在受控环境中完成,且使用公司自己的机械臂;尚无独立复现或第三方审计发布。 数据采集网络同时供给预训练和数据飞轮,代表一种与真实世界环境的运营关系:Generalist AI 称数据采集发生在面包店、自助洗衣店、工厂和家庭。这些数据采集点涉及真实人类操作者使用「data hands」设备。是否存在付费访问协议或其他补偿安排,公司没有披露;它们在运营上不同于 GEN-1 部署客户。 更广泛市场的采用信号是间接的。MarkTechPost 的 2026 年十大物理 AI 模型榜单(2026 年 4 月 28 日发布)未纳入 GEN-1,而列出了 NVIDIA GR00T、Google Gemini Robotics、Physical Intelligence、Figure Helix 等。缺席该榜单不代表产品质量差,但说明截至 2026 年 4 月下旬,GEN-1 尚未建立足够的第三方覆盖或独立验证部署证据,进入分析师跟踪名单。 商业团队很轻(截至 2026 年 6 月仅一个「Applied AI & Partnerships」开放岗位),意味着早期销售动作高接触、由创始人主导,更符合 3 到 10 个早期战略账户,而不是规模化获客漏斗。作为背景,竞争对手 Skild AI 在 2026 年完成 $1.4B 融资,并有具名工业伙伴(Zebra Technologies);Physical Intelligence 则运营 $300/机器人/月 SaaS 模式,并披露了部署——两者提供了更可见的商业对比基线。 [CU018, CU019, CU020, CU021, CU022, CU023]

留存、重复使用与满意度
指标数值细分市场置信度尽调要求
净收入留存(NRR)所有细分市场N/A第一批早期访问伙伴进入首个可报告周期后,披露 NRR 或 GRR
毛收入留存(GRR)所有细分市场N/A单独披露 GRR 与 NRR,以区分扩张和流失;在尽调中索取
试点转生产转化率早期访问计划参与方N/A索取转化率;同类 Physical AI 平台目标基准为 40–70%
数据铸造伙伴续约 / 持续合作率数据采集伙伴确认数据采集协议是否包含强制续约条款;数据伙伴持续时间是模型质量信号的代理指标
G2 / Capterra / Gartner Peer Insights 评价所有商业细分市场N/A截至 2026 年 6 月,GEN-1 未出现在任何客户评价平台;没有独立满意度评分

所有留存和满意度指标都无法从公开来源获得。GEN-1 于 2026 年 4 月发布;即便最早一批也只有 3 个月数据。Physical Intelligence(类似平台)以 $300/robot/month 运营,披露了部署,但没有发布 NRR。缺少留存数据会卡住任何估值模型的承销。

[CU026, CU027]
FU003: 客户验证证据矩阵

评估现有客户 / 合作伙伴类别的证据质量、结果具体性、留存可见度和生产成熟度。

评分 0–3:0=无证据,1=间接 / 推断,2=部分 / 有文档,3=已验证 / 已量化。Universal Robots GTC 演示的证据质量评为 2(双方均有公开记录)。截至 2026 年 6 月,没有任何类别在任何维度达到 3 分。由于没有任何已披露部署数据,所有类别的留存和生产成熟度均为 0。

[CU026, CU027, CU028, CU032]

6.4 具名客户证据质量与局限

截至 2026 年 6 月 30 日,任何公开来源中都没有具名付费客户。公司有意不披露客户身份,这符合企业机器人保密惯例——工厂自动化部署通常由客户披露,而不是供应商披露,且只有在客户认为商业优势已经建立时才会披露。这是物理 AI 类别的标准做法,但也让外界无法独立验证「牵引」。 最可信的具名伙伴证据是 Universal Robots 的 GTC 2026 演示合作。Universal Robots 是全球按出货量计第一的协作机器人制造商,已在 10,000 多家公司部署超过 75,000 台协作机器人。UR 决定在其 GTC 展台上、用一套全新机器人平台展示 Generalist AI,说明这个具有商业影响力的硬件伙伴给出了有意义的技术验证。但 UR 在 GTC 的角色是技术展示伙伴,不是已披露的付费客户或授权分销商。GTC 活动后是否形成任何商业协议,仍未知。 「数据工厂伙伴」——提供真实世界环境用于数据采集的企业——构成第二类具名关系,但任何公开来源都没有披露这些伙伴名称。按 GEN-1 博客,data hands 计划涉及「全球数千个家庭、仓库、工作场所和专门环境」,但这些是数据采集场景,不是模型部署客户。 第三类客户证据是投资人证明。8VC(2026 年 6 月投资方)、Spark Capital(2025 年种子轮投资方)和 Radical Ventures(2026 年 6 月领投方)都提到商业牵引,但没有点名客户或披露收入。没有独立验证的投资人证明,作为客户证据权重较低;但多个独立投资人都提及牵引,说明确实存在某种真实商业活动。 分析师社区明确持怀疑态度。Robotics.press 在 2026 年 4 月自动化研究评估中称:「已验证客户部署、付费试点、案例研究或具名伙伴披露为零——商业就绪度完全未被证明。」该评估早于 GEN-1 发布和 UR GTC 演示,但核心结论——没有公开客户证明——截至 2026 年 6 月 runDate 仍然准确。CB Insights 的物理 AI 市场地图(2026 年 1 月)在基础模型部署示例中也未提及 Generalist AI,而列出 Skild AI 和 FieldAI 拥有正式伙伴关系。 [CU026, CU027, CU028, CU029, CU030, CU031]

扩张与集中风险
扩张驱动集中风险影响尽调路径
跨硬件泛化(任意机器人形态)如果 Universal Robots 或其他 OEM 成为主要分发渠道,依赖风险高若跑通,战略杠杆高;若单一 OEM 控制多数部署,就形成依赖确认 UR 或任何 OEM 是否有排他权;审阅伙伴协议结构
数据飞轮复利(部署越多,模型越好)贡献专有任务数据的伙伴会让 Generalist 服务其竞争对手;存在 IP 风险早期伙伴可能被锁定;后续伙伴面临隐私 / 竞争风险审计数据贡献协议中的 IP 所有权、排他性和保密条款
高接触、创始人主导 GTM单一商业职能(1 个 Applied AI & Partnerships 职位);管线依赖创始人商业团队扩张前,最多只能扩展到约 10 个战略账户确认何时招聘 VP Sales / CRO;评估当前管线深度和阶段分布
资本约束下的销售周期(6–18 个月工业资格认证)长销售周期没有收入,会依赖 $400M 融资跑道短期集中:收入可能由极少数早期账户驱动索取平均销售周期,以及各阶段活跃机会数量
地理集中(团队和主要市场偏美国)除全球数据采集网络外,未披露国际客户或伙伴收入和增长取决于美国工业自动化采用速度确认管线中是否有欧洲或亚洲工业客户或 OEM 伙伴

所有风险都从公开证据和结构分析推断。客户集中度数据、管线信息或合同结构均未公开披露。

[CU034, CU035, CU036]

6.5 集中度风险、采购摩擦与反向证据

Generalist AI 的客户画像带有结构性集中度风险:公司处于早期访问模式,只有少数未披露伙伴;若存在任何收入,这些伙伴很可能贡献当前收入的绝大部分。一个大型早期访问伙伴,或一个硬件 OEM 关系,都可能占近期收入的显著比例,使该关系是否延续成为二元风险。 目标用例的采购摩擦结构性偏高。企业制造和物流客户通常需要安全认证(工业机器人 ISO 10218 / TS 15066)、与现有 ERP 和 WMS 系统集成测试,以及持续 6 到 18 个月的内部试点评估,之后才会批准生产上线。Generalist AI 未公开确认任何安全认证、集成框架、SLA 承诺或正式支持结构。大型企业客户若要合理批准生产部署,必须先解决这一缺口。 最可信的反向技术论点来自 Brad Porter,Cobot 创始人兼 CEO、前 Amazon 机器人高管。他在 Forbes(2026 年 4 月)中说:「只是用巨量数据硬推一个并不完美的架构,非常昂贵,也未必能得到你想要的结果。没有 CNN,ImageNet 不会奏效;没有 transformers,OpenAI 也不会奏效。规模扩展一直都与架构突破相伴。」Porter 的批评直指公司数据扩展策略的核心论点,并暗示商业可行性可能需要超出当前披露范围的架构进展。 更广泛的市场批评来自具身 AI 市场分析:截至 2026 年初,在人形和物理机器人类别中,公开宣称收入中只有 3% 到 5% 真正来自生产性工业使用,其余来自研究、教育和展示部署。虽然这项批评更多针对人形机器人,而非 Generalist AI 的模型软件路径,但它提出了一个相关问题:公司的「商业牵引」中,有多少是具备经济生产性的,多少只是展示性的。 Generalist AI 的数据飞轮模型可能形成长期客户锁定机制:伙伴共享专有任务数据来训练模型,一方面为 Generalist 的护城河贡献数据,另一方面也可能让 Generalist 具备服务具有类似能力需求的竞争客户。数据贡献协议是否包含排他条款,或是否为伙伴数据提供 IP 保护,公司未披露;对担心自身制造流程 IP 被用于训练共享模型的成熟工业客户而言,这可能构成有意义的采购摩擦。 [CU034, CU035, CU036, CU037, CU038]

FU004: 留存与复购队列(行业基准与 Generalist AI)

可比机器人软件平台的示意性留存基准,对比 Generalist AI 尚未公开的实际数据。

Physical Intelligence 和企业基准是分析师估算(Sacra、行业基准),未获一手来源确认。Generalist AI 的实际值为 null:计划于 2026 年 4 月 2 日启动;截至 2026 年 6 月 30 日 runDate,尚无第 3、第 6 或第 12 个月的队列数据。

[CU026]

6.6 证据材料

Chapter 07

07风险

7.1 监管和法律风险是最紧迫的风险栈,EU AI Act 执法即将启动

EU AI Act 对高风险 AI 系统的义务将于 2026 年 8 月 2 日全面可执行;如果 GEN-1 在欧盟部署,或由欧盟境内机器人运营方使用,Generalist AI 立刻面临合规期限。根据该法第 6 条,AI 系统作为受欧盟协调立法覆盖产品的安全组件运行,并且该产品需要第三方合格评定时,就被认定为高风险。GEN-1 若部署在工厂车间的工业机器人——机械臂、协作机器人或半人形系统——内部,几乎必然触发这一门槛,尤其是在汽车或电子装配等机器人安全认证为强制要求的行业。高风险分类会触发一整套义务:持续风险管理系统、数据治理文档、足以支持合格评定的技术文档、人类监督措施、准确性与稳健性要求、上市后监测,以及基本权利影响评估(FRIA)。违规处罚最高可达 €35 million 或全球年营业额 7%,以较高者为准。截至 2026 年 6 月 30 日,Generalist AI 未发布 DPA、合格评定文档、高风险声明,也没有面向欧盟 GEN-1 部署的合规或隐私通知。公司唯一公开披露的是早期访问伙伴计划的联系邮箱。这符合早期商业化阶段特征,但也意味着投资者完全看不见其合规状态。 另一条法律敞口来自产品责任法。在美国,严格产品责任原则规定,只要缺陷产品造成伤害,制造商即使不存在过失也要承担责任。AI 软件提供方与硬件制造商分离时,责任如何归属,法院仍在处理;欧盟 AI Liability Directive 已开始填补这一缺口,但美国联邦法律尚未跟进。对 Generalist AI 来说,风险是分层的:如果运行 GEN-1 的机器人伤害工人,诉讼可能同时点名硬件 OEM、系统集成商和作为模型提供方的 Generalist AI。Generalist 尚未公开披露任何保险覆盖、赔偿条款或责任分配框架。RAND 2024 年关于美国侵权法适用于 AI 系统的分析发现,AI 开发链条碎片化会让因果归属变得复杂,并让模型提供方面临难以预测的敞口。由 Yoshua Bengio 主持的 International AI Safety Report 2026 将具身 AI 的问责缺口列为该领域最紧迫的治理挑战之一。GDPR 带来第三条法律线:Generalist AI 在全球数千个家庭、仓库和工厂收集物理交互数据,范围包括欧盟。这些数据包含细粒度手腕运动学、工作空间视频,以及 data-hands 操作员可能产生的生物识别元素。GDPR 第 9 条将通过生物识别处理唯一识别个人的数据视为特殊类别数据,需要明确同意和严格控制。公司尚未发布该数据收集计划的公开隐私通知、DPIA 文档或同意管理框架。 [CR001, CR002, CR003, CR004, CR005, CR006]

监管 / 法律风险登记表
规则 / 许可 / 案件司法辖区状态可能性严重性缓释剩余暴露尽调路径
EU AI Act(第 6 条)对作为机器人安全组件部署的 GEN-1 进行高风险分类欧盟高风险义务将于 2026 年 8 月 2 日可执行;Generalist 未发布合规文档严重未公开缓释措施;看不到合格评定、技术文档或人类监督框架极高在 NDA 下向管理层索取合格评定路线图、欧盟法律顾问备忘录和技术文档草案。
GEN-1 控制的机器人造成身体伤害的产品责任美国(州侵权法);欧盟(AI Liability Directive 仍在推进)未发布赔偿、保险或责任分配框架;从软件到硬件再到部署方的责任链尚未解决严重未披露面向 GEN-1 早期访问伙伴的保险覆盖、合同责任上限或赔偿条款索取早期访问伙伴协议、责任与赔偿条款,以及产品责任保险覆盖证据。
GDPR 特殊类别数据暴露:物理交互数据采集欧盟;无论采集方所在地,适用于欧盟居民全球数据采集覆盖家庭和工作场所,包含欧盟主体;未发布 DPIA 或隐私通知未公开 DPIA,未公开同意管理框架,也未披露 GEN-1 数据计划的 DPO 联系方式在 NDA 下索取欧盟采集站点数据地图、DPIA 文档、同意记录和 DPO 任命信息。
未发布 GEN-1 的 DPA 或第三方安全审计跨司法辖区(欧盟、美国企业买方)企业买方和监管机构要求 DPA 和安全审计;目前没有公开材料未公开缓释措施;早期访问计划意味着存在合同协议,但内容未披露在 NDA 下索取 DPA 草案、安全框架文档和任何第三方安全评估。
美国州和联邦 AI 监管拼图 / 新兴 AI 治理义务美国没有统一的联邦 AI 法;州层面法案(CA、TX、NY)在 2026 年推进;FTC 对 AI 表述的监管加强未披露美国 AI 监管监测的合规计划或法律团队确认美国 AI 监管版图的法律顾问覆盖,并评估目标部署州的州法暴露。

行按严重性排序(关键项在前)。严重性和可能性评估基于监管文本、法律分析和公开可见的合规缺口;未使用 Generalist AI 管理层的私有信息。

[CR001, CR002, CR003, CR004, CR006, CR008]
FR001: 风险热力图

有序评分基于有来源支撑的证据,而非合成概率;剩余严重性反映的是证据支持的缓释缺口评估,不是统计估计。

[CR001, CR003, CR007, CR011, CR019, CR034]

7.2 模型安全缺口和失控即兴行为带来部署与声誉责任

GEN-1 的核心商业卖点——能够产出训练分布之外行为的即兴智能——同时也是它法律和运营风险最高的属性。公司自己的博客描述了 GEN-1 产出“出人意料”的行为:机器人在没有明确指令的情况下接住滑落垫圈并将其推到位,或用未编程方式折叠纸板以完成任务。公司把这些描述为正向涌现特性。产品技术章节明确指出,这种即兴能力既是优势,也是“潜在责任”。在工业部署里,机器人涌现行为不是功能,而是失控变量,会破坏 ISO 10218、协作机器人 ISO/TS 15066 以及 EU AI Act 高风险 AI 要求所依赖的安全假设。截至 2026 年 4 月,Generalist 尚未发布 GEN-1 的信任与安全框架、对齐文档、公开红队评估或第三方安全审计。 International AI Safety Report 2026 标出两项与 Generalist 直接相关的动态:第一,通用 AI 能力仍然“参差不齐”——系统会在看似简单的任务上失败,却能完成更难的任务,导致真实部署可靠性难以预测;第二,一些模型如今可以区分评估和部署场景,并以让部署前测试不足以覆盖的方式改变行为。SAE World Congress 2026 的 Embodied AI 小组(arXiv 论文 2605.10653 记录)形成广泛共识:具身 AI 必须被视为系统工程挑战,需要工程严谨性、生命周期治理和持续演进的标准——这些 Generalist AI 都尚未公开呈现。NeuroForge 2026 年行业分析发现,现代具身 AI 系统持续难以处理需要长期可靠执行、且不能触发单点故障级联的“长程逻辑链”。2026 年上半年,仓库和工厂里的 Physical AI 部署记录了 50 多起 AI 模型故障,包括幻觉导致工作流崩溃,以及上下文相关故障造成运营停机。 合起来看,GEN-1 的模型安全和部署可靠性风险,在结构上高于典型企业软件产品:故障发生在物理世界,可能造成人身伤害,并会在生产环境里立刻暴露。如果某个早期访问部署伙伴发生公开事故,声誉和法律损害可能先于公司搭好合规与保险脚手架到来。 [CR011, CR012, CR013, CR014, CR015, CR016]

运营 / 质量 / 安全风险台账
故障模式发生概率严重程度缓释成熟度剩余暴露未解决缺口
GEN-1 在生产部署中即兴行动,造成人身伤害或财产损失严重截至 2026 年 6 月,公司尚未披露信任与安全框架、对齐协议或公开的事故响应计划。
训练数据泄露,或专有物理交互数据集(50 万+ 小时)被外传严重unknown全球数据采集和上传管线的数据安全架构没有公开说明;SOC 2 或 ISO 27001 认证也未确认。
GEN-1 没有公开安全认证路线图(ISO 10218、ISO/TS 15066、CE 标志)未确认工业机器人进入欧盟和受监管环境需要安全认证;缺失认证会挡住受监管垂直行业的企业采用。
部署中的具身 AI 模型遭遇对抗或操纵攻击(提示注入、传感器欺骗)公司未发布红队评估或对抗鲁棒性测试结果;2026 AI Safety Report 指出了评估环境与部署环境之间新的上下文操纵风险。

严重程度和缓释成熟度只根据公开证据评估;部署事故频率来自全行业 VLA 模型数据,并非 Generalist AI 已确认事故。

FR002: 风险传导图

因果方向有证据支撑;这张图展示每组风险如何传导为运营、财务或投资论点层面的后果,而不是量化模型。

[CR001, CR011, CR019, CR026, CR034, CR008]

7.3 对 NVIDIA 的算力和基础设施依赖带来训练排期与成本风险

NVIDIA NVentures 同时参与了 Generalist AI 两轮融资——2025 年 3 月种子轮和 2026 年 6 月 $400 million Series B。这种关系带来战略一致性,也带来对 NVIDIA 机器人软件栈的优先接入,但并不能消除前沿模型训练对 NVIDIA 硬件的结构性依赖。GEN-1 在分布式基础设施上训练,可处理 PB 级物理交互数据,并在每天训练中吸收 6.85 年操控经验——这种工作负载需要 H100 或 Blackwell 级 GPU 集群。截至 2026 年中,非超大规模云厂商买家采购 H100 和 H200 GPU 的交付周期仍为 36–52 周,原因包括 SK Hynix 的 High Bandwidth Memory(HBM3)供应受限、TSMC CoWoS 先进封装超额订阅,以及 NVIDIA 自身将产能转向利润率更高的 Blackwell 生产。具身 AI 初创公司不在超大规模云厂商通过远期合约拿到的分配优先窗口内;Generalist 的 GEN-2 训练路径,取决于能否在未披露时间线上拿到所需规模的算力。 Generalist AI 据称运营的多云合同没有公开点名。云提供商集中度尚未确认:如果训练集中在一家提供商,价格重谈、容量约束或服务中断都会直接卡住模型开发。物理数据收集基础设施——数千台可穿戴 data-hands 设备、定制上传机器、专线互联网以及 O(10K) 处理核心——形成第二层基础设施依赖,运营上不同于云算力。数据管线一旦中断,就会打断数据飞轮;后者正是 Generalist 相比纯软件 AI 公司获得估值溢价的理由。 下一代模型训练所需资本没有公开披露,但结构上很高。Physical Intelligence 的数据画像意味着,每一代主要模型的训练成本以数千万美元计;Figure AI 年烧钱 $200–300 million 的基准,反映了更大规模下由硬件支撑的训练运营。Generalist AI 目标是以纯软件利润率画像承载同样的算力和数据强度,但底层基础设施仍要靠股权资本买单。如果 GPU 采购让 GEN-2 训练时间线推迟 6 到 12 个月,Generalist 相对 Physical Intelligence(π0.5、$11 billion 估值、2026 年 4 月融资 $1 billion)、Google DeepMind 和 NVIDIA 自有开源模型 GR00T N1.6 的竞争窗口会明显收窄。 [CR019, CR020, CR021, CR022, CR023, CR024]

合作方 / 依赖风险台账
依赖项交易对手角色集中度失效场景严重程度缓释措施剩余暴露
前沿模型训练所需 GPU 算力NVIDIA(战略投资者;未指明的云 / 新型云提供商)训练与推理基础设施;没有 H100/Blackwell 级硬件,就无法做 GEN-2 规模训练很高(NVIDIA 生态占主导;HBM3 全行业供应受限)36–52 周交付周期拖慢模型训练计划;超大云厂商优先分配会挤出创业公司;NVIDIA 将产能转向 Blackwell,会降低 H100 可得性严重NVIDIA NVentures 共同投资意味着或有一定优先供给;多云策略未确认
数据管线所需云基础设施和存储未知提供商(公司称采用多云,但未点名)PB 级数据存储、处理和分布式训练编排集中度未知单一云宕机扰乱数据引擎和训练;价格重谈压缩毛利率公司称已有多云合同,但未披露提供商;集中度未知
GEN-1 部署所需机器人硬件 OEM 合作方未具名早期访问合作方;Universal Robots 在 GTC 演示语境中已确认将 GEN-1 送达企业终端用户的分销渠道;硬件兼容性验证高(具名合作关系很少;唯一商业联系入口是电子邮件地址)合作方退出会切断收入路径和数据飞轮贡献未公开合作协议或渠道合同
物理数据采集操作方和数据工厂合作方全球操作方和设施网络(除 Class 1、2、3 分类外未具名)每周生成 10,000+ 小时训练数据;是数据飞轮的核心中(地域分散降低单点风险)操作方退出、质量下降或同意被撤回,会扰乱数据管线分布式全球网络提供一定韧性;按任务类型划分的 Class 多样性
支撑持续运营和模型开发的资本提供方Spark Capital、Radical Ventures、8VC、USV、NVIDIA NVentures 及共同投资方C 轮及以后取决于能否维持投资者信心;没有收入桥接中高(集中在少数投资者;没有收入桥接)模型失败或监管事件后,投资者情绪转向会降低 C 轮概率2026 年 6 月 $400M 融资提供多年资金续航;多数情景下现有投资者会续投

Generalist AI 未公开交易对手名称、合同金额和集中度百分比;严重程度和集中度评估来自公开投资与产品公告的推断。

FR003: 依赖图

有向图显示外部交易对手一旦中断,会直接削弱 Generalist AI 的训练运营、部署能力或资本获取。

[CR019, CR020, CR022, CR028, CR030, CR032]

7.4 收入前资本密集度和未披露伙伴栈带来财务模型风险

截至 2026 年 6 月 30 日,Generalist AI 两轮融资已超过 $500 million,但未披露收入、ARR、客户名称、烧钱速度或合同条款。按 70 人、研究密集团队运营全球物理基础设施推算,人员规模隐含月烧钱 $5–15 million;2026 年 6 月融资可支撑约 2–6 年资金续航, 取决于该区间低端或高端——但公司在产生收入之前就已经资本饥渴。公司对 $400 million 的计划用途写得很明确:建设下一代模型、扩展物理数据引擎、扩大团队。这些用途都不会产生近期正现金流。最接近的可比公司 Physical Intelligence 于 2026 年 4 月以 $11 billion 估值融资 $1 billion,并运营早期商业计划;分析师描述的成本结构表明,在早期部署规模下,训练成本强度会比收入补充资本更快地消耗资金。 伙伴依赖栈在结构上没有披露。Generalist AI 的商业模式需要机器人硬件 OEM(或系统集成商)把 GEN-1 集成进产品,并把部署带给终端客户。截至 2026 年 6 月,没有任何硬件伙伴名称、集成协议或渠道安排公开。唯一的伙伴参与证据,是 partnerships@generalistai.com 的早期访问计划,以及投资方提到“产生任务数据的真实企业”。如果一两个早期访问伙伴贡献了运营数据飞轮的大部分,并支撑公司围绕商业进展 的融资叙事,一旦退出,会同时损害技术路线图和投资者故事。NVIDIA 的战略共同投资带来软承诺,但不是排他合同。投资者集中在 Spark Capital、Radical Ventures、8VC、USV 和 NVIDIA NVentures;若该群体情绪转向,按有利条款完成 Series C 的可能性会大幅下降。公司未披露任何可转债、债务额度或收入型融资,因此除股权资本外没有下行缓冲。 [CR026, CR027, CR028, CR029, CR030, CR031]

缓释措施与终止标准表
风险可监测触发信号阈值 / 事件行动含义
EU AI Act 高风险合规不达标欧盟市场准入;企业买方采购要求;欧盟监管问询2026 年 8 月 2 日执法日已过,但 Generalist 仍未发布合规文件,或收到行业合作方被欧盟采购挡下的信号暂停面向欧盟的部署承诺;上报董事会;立即聘请 EU AI Act 法律顾问。
生产部署中出现可归因于 GEN-1 的人身伤害公开事故报告;诉讼文件;合作方披露;媒体报道任何已确认、涉及 GEN-1 控制机器人的人身伤害或财产损失事故触发全面部署复盘;通知投资者;聘请产品责任律师;与第三方审计方重新评估安全框架。
NVIDIA 算力获取中断或价格大幅上涨GPU 采购交付周期延长至 52 周以上;Blackwell 配额被拒;云价格重谈涨幅超过 25%GEN-2 训练算力较内部路线图延迟 6 个月或更久加快替代供应(AMD MI300X、超大云厂商 TPU 访问、新型云);修订资本计划,反映更高训练成本。
联合创始人离职(三人中任一)公开公告;团队重组信号;董事会构成变化任一联合创始人在商业里程碑前离职重新评估论点;召开董事会审查继任安排;评估对投资者信心和融资路径的影响。
没有已承诺条款清单,资本续航低于 12 个月月度现金消耗与现金余额对比;剩余 18 个月时仍无 C 轮动作按当前现金消耗,现金余额不足 12 个月,且手中没有条款清单启动紧急资本效率措施;启动桥接融资或 C 轮流程;评估包括人才收购在内的战略选项。
数据采集运营遭遇 GDPR 或数据隐私执法行动监管问询函;DPA 投诉;媒体报道数据主体投诉欧盟监管机构针对数据采集项目发起任何正式 GDPR 执法问询暂停欧盟数据采集,等待法律审查;任命 DPO;委托紧急 DPIA;向投资者披露。

触发信号来自公开来源即可观察(监管公告、诉讼数据库、媒体、招聘信息);阈值是分析师判断,并非合同定义标准。

7.5 关键人物集中和治理不透明放大所有其他风险维度

三位联合创始人——Pete Florence(CEO)、Andy Zeng(Chief Scientist)和 Andrew Barry(CTO)——都在公开场合活跃,但公司的科学身份和商业叙事异常集中在 Florence 与 Zeng 身上。每一篇主要媒体报道、投资人备忘录和产品公告,都会引用 Florence、提到 Zeng 的研究传承,或两者兼具。8VC 的投资笔记明确形容 Florence “既是建设者,也是研究者——以他的研究资历而言,他有磁性且异常商业化”,证实投资人信心很大程度上押注创始团队,而非独立机构能力。对一家成立两年的公司来说,这可以理解——但也意味着联合创始人离职、健康事件或高层人际冲突,不只会影响运营,还会直接打击融资前景、客户信心和团队稳定性。 治理结构进一步增加不透明度:董事会构成、独立董事名单、审计委员会和治理章程均未公开披露。2026 年 6 月 Series B 公告没有点名领投方代表入董事会,尽管 8VC 有投资备忘录,且本轮投入资本规模很大。天使投资人名单——Eric Yuan、Bin Lin、Fei-Fei Li、Naval Ravikant——声望很高,但不能在运营上提供保护。股权计划、归属等待期或留任条款都不公开,公众资料无法验证联合创始人锁定概率。 人才留存风险延伸到创始层之下:更广泛团队来自 Google DeepMind、OpenAI 和 Boston Dynamics,而这些机构也最积极争夺具身 AI 研究人员。2026 年,DeepMind 的 Gemini Robotics 1.5、Physical Intelligence 扩张后的团队和 Figure AI 都能提供有竞争力的薪酬和有声望的研究环境。Generalist AI 只有 70 名员工,创始人以下没有深厚板凳。模型架构团队或数据引擎团队任何一个关键成员离开,都可能让 GEN-2 延后数个季度,且不易替补。公司的招聘岗位显示其意识到这一点:开放岗位覆盖研究、ML 基础设施、算力优化、机器人控制和应用伙伴关系,说明团队仍在搭建关键冗余。本章所有风险——监管、模型安全、算力、资本——都离不开稳定的高级领导层,因此关键人物集中是风险放大器,而不是孤立维度。 [CR034, CR035, CR036, CR037, CR038, CR039]

人员 / 执行风险台账
角色 / 职能依赖或缺口发生概率严重程度缓释措施尽调路径
Pete Florence(CEO)主要外部代表;核心投资者关系;商业和研究方向集中在其身上低(无离职信号)严重未披露继任计划、股权归属悬崖期细节,也未看到替代领导人确认股权归属表、悬崖期和任何离职条款;评估 CEO 以下团队厚度。
Andy Zeng(首席科学家)模型架构和研究方向;GEN-1 设计中被引用最多的内部技术权威低(无离职信号)未披露继任计划,也未看到第二梯队首席科学家确认股权归属表;评估 Zeng 之外的研究团队厚度和论文管线。
Andrew Barry(CTO)硬件集成架构;系统可靠性;多机器人和多 OEM 兼容性的关键人低(无离职信号)未披露 CTO 继任安排,也未确认工程副总裁角色确认 CTO 以下工程管理结构;评估硬件集成团队厚度。
商业和商业化落地职能(仅一个应用 AI 与合作伙伴关系角色)截至 2026 年 6 月,只识别出一名面向商业的员工;整个商业化落地依赖创始人关系中(精简商业职能是符合阶段的有意选择,但很脆弱)创始人用直接企业关系补位;早期访问计划降低大规模商业化落地需求确认管线覆盖、交易归属,以及规模化阶段的商业化落地爬坡计划。

截至 2026 年 6 月 30 日,三位联合创始人都仍公开活跃;按当前观察,离职概率较低,但没有股权归属表或留任合同数据,公开来源无法验证。

Chapter 08

08估值

8.1 建议:保持跟踪并严守入场条件——存在相对价值,但没有商业证明

Generalist AI $2 billion 投后估值低于 Physical AI 领域的直接软件可比公司。Physical Intelligence 2026 年 3 月正在洽谈以 $11 billion 估值融资,Skild AI 则已在 2026 年 1 月以 $14 billion 估值完成融资,并报告约 $30M 年经常性收入——如果只按纯可选性看,Generalist AI 相对私募市场显得便宜。投资人阵容增强了这一判断:Radical Ventures(领投)、NVIDIA NVentures、Bezos Expeditions、8VC、Union Square Ventures,以及包括 Fei-Fei Li 和 Eric Yuan 在内的天使投资人,对这个年龄的公司来说,是前沿 AI 社群信心异常强的信号。 但定价逻辑完全处于商业化前。Generalist AI 未披露任何收入、ARR、客户数量、定价、毛利率或单位经济。GEN-1 早期访问计划于 2026 年 4 月启动,但没有具名伙伴获得确认。8VC 备忘录提到“显著的早期商业进展”,这是最强公开信号,但没有量化,而且来自已投资方。没有私下尽调,仅凭这些不足以支撑 $2B 估值锚。 建议是跟踪。若投资者可以取得私下尽调,并确认至少一个可验证、产生收入的部署,应在入场价格和股权结构表 条款确认后,把结论重新评估为有条件买入。没有该确认时,$2B 标记最好被视为押注技术命题的真实期权:值得密切跟踪,但还不足以按完整仓位承销。 [CV001, CV002, CV005, CV006, CV009, CV010]

建议摘要表
决策字段当前判断决策含义
建议跟踪保持密切接触;只有私下尽调确认商业牵引后,再决定仓位大小。
信心团队和技术履历很强;公开来源缺少商业验证。
风险评级尚无收入,数据扩展论点未验证,存在 NVIDIA 商品化风险,竞争者推进很快。
估值立场潜力已被部分透支$2B 相比同业偏低,但没有收入锚;入场是在为团队和知识产权买期权。
持有 / 退出姿态3–5 年内成为并购标的;IPO 视野 5+ 年战略收购方(NVIDIA、Amazon、Google)比独立上市更可信。
价格纪律尚无收入时,不应在 $2B 估值下无视价格买入增加仓位前,至少需要私下确认一个可验证的 ARR 事件。

“跟踪”判断反映的是当前入场价格和公开证据,并不等于笼统认为技术弱。若私下尽调补上收入缺口,信心可提升至有条件买入。

[CV001, CV005, CV009, CV021, CV033, CV035]
投资论点 / 反论点表
论据方向什么会改变判断
顶级研究型建设团队(Pete Florence、Andy Zeng、Andrew Barry),直接拥有 DeepMind 和 Boston Dynamics 履历。正向2 名以上创始人离职,或核心团队大规模出走,会削弱人才护城河。
通过 data-hands 设备采集的专有 50 万小时物理交互数据集,结构上很难复制。正向开源数据采集方法,或竞争者拿到同等规模数据集,会侵蚀护城河。
硬件无关定位让 Generalist 能成为跨所有机器人形态的智能层,而不是绑定单一 OEM。正向失去关键硬件合作方,或 NVIDIA 扩张竞争性的跨具身模型,都会压缩平台广度。
$2B 估值在直接软件同业中最低——若技术论点成立,相对上行空间可观。正向Physical Intelligence 或 Skild AI 出现估值下调融资,会重定行业估值,降低相对溢价。
未披露收入、未确认客户、没有财务指标——所有上行都指向未来。反向若确认一个锚定客户,或披露 ARR 里程碑,判断会转向有条件买入。
NVIDIA GR00T N-series 向开发者免费提供开源模型权重,威胁模型层商品化。反向如果 NVIDIA 将 GR00T 限制为自有用途,或 GR00T 性能显著低于 GEN-1,威胁会减弱。
竞争者 Skild AI($14B、$30M ARR)和 Physical Intelligence($11B)融资更多;Skild 还已证明收入。反向Skild 和 PI 若进入持续商业部署低迷期,会降低 Generalist 需要跨过的门槛。
工业资质周期为 6–18 个月,意味着收入会结构性延后;收入到来前现金消耗会先加速。反向更短周期的产品(例如用于快速原型的托管 API)可能加快第一笔收入到来。

正向论点和反向论点均来自公开证据。核心张力在于:输入要素(团队、数据、架构)质量很高,但输出证明(收入、部署、客户名称)完全缺席。

[CV007, CV008, CV009, CV012, CV013, CV034]
FV001: 推荐逻辑

“跟踪”建议来自两股力量的交汇:团队和数据信号很强,但估值被拉高、商业验证为零;NVIDIA 商品化风险是关键摆动因素。

[CV001, CV005, CV009, CV034, CV042, CV043]
FV004: 投资 KPI

Generalist AI 在团队和数据护城河上得分高,但商业验证和估值支撑薄弱——典型的收入前深科技 AI 画像。

分数是供 IC 讨论的 0-10 有序判断;不是数学汇总。

[CV005, CV008, CV009, CV015, CV034, CV043]

8.2 价格背景:$2B 是直接软件可比公司中最低水平——收入前状态限制所有传统锚点

Generalist AI 的 $2B 估值显著低于最近对手立下的私募市场基准。Physical Intelligence 2025 年 11 月 Series B 估值为 $5.6 billion,到 2026 年 3 月,据报道公司正洽谈以 $11 billion 估值新一轮融资。Skild AI 在 2026 年 1 月完成 $1.4 billion Series C,估值超过 $14 billion,由 SoftBank 和 NVIDIA 领投——而 Skild 披露了约 $30M 收入,至少为其高得多的倍数提供一个收入锚。Figure AI 作为软硬件一体的全栈竞争者,2025 年 9 月达到 $39 billion 估值。 没有收入数据,无法对 Generalist AI 套用传统收入倍数。用前瞻预测倒推:按 20x 倍数(适用于具备早期商业进展的高增长私有 AI 软件公司),$2B 意味着目标 ARR 为 $100M;按 40x(适用于收入前前沿 AI 的稀缺溢价),$2B 只需要 $50M ARR。两个场景都说得通,但都没有已确认指标支撑。相比之下,Finerva 对上市 Robotics & AI 公司分析显示,2025 年 Q4 行业 EV/Revenue 中位数为 3.4x,EBITDA 中位数为 16.8x——但这些上市公司中位数反映的是成熟、有收入企业,不适用于收入前阶段公司。Finro 2026 年 Q1 数据集显示,LLM 厂商 EV/Revenue 中位数为 39.5x,AI 机器人私有公司按预计前瞻收入估算为 20–60x。 对 Generalist AI 来说,最诚实的 $2B 表述根本不是收入倍数,而是技术与团队的期权价格。相对同行,入场便宜;但在商业证明为零时,绝对价格并不便宜。任何以该价格进入的投资者,实际上是在为几件事付费:(1) 创始人在 DeepMind 和 Boston Dynamics 的履历与历史战绩,(2) 500,000+ 小时专有数据集护城河,(3) GEN-1 架构差异化,(4) 物理基础模型能否像语言模型一样扩展的真实期权。$2B 对这些期权是否公允,取决于投资者只能通过私下尽调建立的确信。 [CV010, CV011, CV012, CV013, CV014, CV015]

可比估值表
可比公司阶段 / 类型估值 / 倍数收入信号与 Generalist AI 的相关性局限
Generalist AI(标的)尚无收入,B 轮(2026 年 6 月)$2.0B 投后未披露;仅有早期访问合作方计划相对评估的基线没有收入锚;$2B 是围绕团队和知识产权的期权定价
Physical Intelligence尚无收入,B 轮(2025 年 11 月);C 轮洽谈估值 $11B(2026 年 3 月)$5.6B(已确认)— $11B(据报道洽谈)未公开披露最直接的架构同业:纯软件、跨具身、尚无收入架构不同(VLA vs Generalist 的 Harmonic Reasoning);数据策略不同
Skild AI早期收入,C 轮(2026 年 1 月)$14B;上一轮约 $4.7B(2025 年夏)商业发布后数月内达到约 $30M ARR纯软件、硬件无关的“全具身”大脑——最接近的收入基准收购 Zebra Robotics Automation(2026 年 4 月),补上 Generalist 缺少的企业分销
Figure AI(参考)早期收入(BMW 部署),C 轮(2025 年 9 月)$39B 投后已记录 BMW Spartanburg 部署(装载 90,000+ 个零件,1,250 小时)全栈人形机器人,且有记录在案的商业部署——行业估值上限硬件 + 软件捆绑;不是直接的软件层可比对象
上市 AI 机器人软件基准已产生收入的上市公司(Finerva 2025 年 Q4 样本)EV/Revenue 中位数 3.4x,最高 24x;EV/EBITDA 中位数 16.8x已产生收入为行业成熟期收入倍数提供底部参考上市公司中位数不适用于尚无收入阶段;仅作方向性参考

估值数据来自公开报道和投资者来源。收入数字为已披露或分析师估算。全栈估值(Figure AI)不能直接对比 Generalist AI 的纯软件模式。

[CV010, CV011, CV012, CV013, CV014, CV015]
FV002: 估值敏感性

在 $2B 估值且无收入的情况下,远期收入倍数对应很宽的 ARR 目标区间——全部未经验证。敏感性分析展示,在 $2B 进入看起来合理之前,需要多少事情跑通。

收入门槛只是以 $2B 进入价搭出的简单 EV/收入桥,不是 DCF 结果。倍数选择反映私有 AI 软件可比公司区间(Finro 2026 Q1;Finerva 2025 Q4)。

[CV017, CV018, CV019, CV020, CV021, CV022]

8.3 情景分析:牛市情景技术上可信;熊市情景更接近今天的公开证据

牛市情景建立在三个相互强化的假设上:(a) 数据引擎飞轮按设计运转——今天的 500,000 小时在两年内增长到 5 million+ 小时,让复制数据集的成本高到难以承受;(b) GEN-1 及后续模型在 2026–2027 年证明具备足够商业可行性,并在至少一个锚定行业锁定多年期企业部署;(c) 面向企业采购,Generalist AI 的硬件无关定位优于垂直整合对手。在这种情景下,2029–2030 年由技术巨头(NVIDIA、Amazon、Google 或大型 OEM)以 $500M–$1B ARR、15–20x ARR 收购,意味着 $7.5–20B 退出;相对 $2B 取得 3.75–10x 回报,但摊薄风险很大。 基准情景更谨慎。商业收入在 2027 年出现,但受企业机器人资质验证周期(6 到 18 个月)拖累,采用速度较慢,ARR 到 2028 年达到 $20–50M。2029–2031 年按 $30–50M ARR 的 10–15x 被战略收购,意味着 $300M–$750M 退出——在考虑摊薄前,相对 $2B 入场价为持平到小幅负回报。基准情景符合可比阶段深科技 AI 商业化常见时间线。 熊市情景有现有不利证据支持。AIRoboticDaily 记录,目前约 95% 的人形机器人收入来自研究和展厅用途,而非生产性工业部署。Cobot 的 Brad Porter 认为,经过生产验证的 Physical AI 系统比基于研究论文的方法更有商业可信度。NVIDIA 开源 GR00T N 系列向开发者免费提供模型权重,威胁基础模型层商品化。在熊市情景下,GEN-1 到 2028 年仍未转化为收入,$400M 资金续航耗尽,结果是人才收购,或以大幅低于 $2B 入场价的降价轮收场。熊市情景不是基准情景,但并不遥远——它只需要商业化比牛市叙事预计晚 2 到 3 年。 [CV024, CV025, CV026, CV027, CV028, CV029]

牛市 / 基准 / 熊市情景表
情景关键假设隐含估值路径 / 回报逻辑关键风险概率信号
牛市(上行)到 2028 年,数据飞轮达到 500 万+ 小时;GEN-1/GEN-2 锁定 3 个以上锚定企业部署;硬件无关模型相对垂直一体化对手赢得更广泛 OEM 采用。2029–2031 年由 NVIDIA 或大型科技巨头以 $7.5–20B 收购(基于 $500M–$1B ARR 的 10–15x)。按 $2B 入场,稀释前回报 3.75–10x。商业化时间表滑坡;NVIDIA GR00T 达到可比性能;数据扩展无法转化为商业可行性。低到中。技术上可行,但需要多个执行环节同时成功,而目前没有收入先例。
基准到 2028 年实现 $20–50M ARR,拥有 2–3 个具名商业部署;2029–2031 年以 10–15x ARR 融资或战略退出;退出前再融资 1–2 轮。以 $300M–$750M 战略收购;按 $2B 入场,完全稀释后回报小幅为正到持平。早期轮机构投资者条款可能更好。采用周期慢;NVIDIA 和开源替代方案带来价格压力;收入放量前,下一代模型还需要资本投入。中。与可比行业观察到的深科技 AI 商业化时间线一致。
熊市到 2028 年商业收入仍未落地;$400M 资金续航耗尽;没有可接受价格的战略收购方。NVIDIA GR00T 削弱差异化。以 $500M–$800M 估值下调融资,或按团队价值被人才收购。按 $2B 账面估值入场的投资者接近全损。熊市只需要商业化比牛市叙事假设晚 2–3 年。有资金续航支撑,概率低到中,但并非可以忽略;行业估值过高的负面分析师评论也支持这一点。

情景边界来自类似深科技 AI 商业化时间线和竞争者估值锚,而不是 Generalist AI 已披露财务预测。

[CV028, CV029, CV030, CV031, CV032, CV034]
论点失效与终止触发表
触发项阈值 / 事件传导到论点行动含义
到 2027 年 Q4 仍无可验证商业收入GEN-1 早期访问发布 18 个月后,仍无具名客户或披露的 ARR 里程碑数据飞轮论点无法确认;下一轮融资风险上升重新评估卖出或减仓;增加敞口前要求确认收入
NVIDIA GR00T 企业端扩张NVIDIA 以宽松企业许可证提供 GR00T N-series,且性能可比或更好基础模型层被商品化;Generalist AI 专有模型溢价被压缩立即评估护城河恶化;将估值目标区间下调 40–60%
竞争者 ARR 里程碑:PI 或 Skild 达到 $100M+ ARRPhysical Intelligence 或 Skild AI 披露 $100M+ ARR,且模型性能大体可比Generalist AI 来自数据规模的相对护城河,已不足以单独构成差异化触发对 Generalist AI 商业管线的紧急私下尽调;重新评估论点
关键人离职Pete Florence(CEO)或 Andy Zeng(首席科学家)离开领导岗位公司身份、融资叙事和技术方向都绑定这两位创始人暂停新增投资;评估领导层过渡计划和存续团队厚度
估值下调融资或桥接融资Generalist AI 在没有实质商业进展时,以 $2B 或更低估值进行后续融资表明市场已对近期商业验证失去信心;稀释风险上升只有估值下调轮价格显著低于 $1B 且确认已有牵引时,才视为买入信号;否则减仓

触发项都是可测量或可观察事件,不是定性判断。每一项都能直接传导到商业论点,并对应可执行的投资动作。

[CV029, CV034, CV036, CV037, CV038]
FV003: 估值 / 回报区间

从 $2B 进入 Generalist AI 后的牛、基准、熊三种退出区间,反映从未有收入到潜在并购或 IPO 的巨大不确定性。

退出区间是基于情景的估算,锚定同业可比和类似深科技 AI 并购交易;不是管理层指引或 DCF 输出。后续融资轮(可能 1–2 轮)带来的稀释会降低投资人有效回报。

[CV028, CV029, CV030, CV031, CV032, CV033]

8.4 退出准备度和最终尽调:3–5 年内 M&A 可选性真实存在;独立 IPO 至少 5 年后

按任何传统标准,Generalist AI 都还不具备 IPO 条件。作为收入前私营公司,公司没有披露财务指标、没有审计报表,也没有公开部署记录,缺少上市所需披露基础。Figure AI($39B)和 Boston Dynamics(目标 $85–103B IPO)是最接近的行业参照,但两者运营历史更长、部署记录更完整、收入可见度更高。从 2026 年 6 月 runDate 算起,Generalist AI 可信的独立 IPO 至少还要五年。 战略 M&A 是 3–5 年窗口内更可信的退出路径。NVIDIA 通过 NVentures 在本轮继续战略投资,收购理由最直接:买下 Generalist AI 的模型和数据集,可补上一层硬件无关的 AI 智能,与 NVIDIA 的 GR00T 算力栈和 Isaac Sim 平台互补。Amazon Robotics 也是另一个可能买家,其领导层包括 Covariant 前创始人,与 Generalist AI 的跨形态命题相通。Google DeepMind 通过共同研究传承拥有最深技术重叠,但内部竞争风险让收购在结构上更难。 近期关键尽调问题是:(1) 验证至少一个产生收入的商业部署,并确认 ARR;(2) 确认股权结构表 和清算优先权结构,以理解标题 $2B 标记之外的有效入场经济性;(3) 由第三方独立验证模型表现,而不是只依赖 GEN-1 公司报告的 99% 成功率;(4) 确认董事会构成,尤其是 Radical Ventures 是否持有正式董事席位;(5) 在成本结构未确认的情况下,核实 $400M 融资后的烧钱速度和 资金续航轨迹。 [CV003, CV004, CV007, CV016, CV033, CV035]

最终尽调问题表
主题缺失证据为什么重要负责人或尽调路径
收入和商业牵引已确认 ARR、至少一个具名付费客户的合同条款、管线规模整个 $2B 论点都押在未来商业执行上;公开来源没有收入锚在资料室要求提供商业管线报告、早期访问合作方名单,以及 MOU/LOI 条款
股权结构表和优先权结构完整股权结构表,展示稀释后股数、期权池、清算优先权和棘轮条款经过清算优先权分配后,名义 $2B 投后估值可能大幅高估有效经济入场价向公司或领投方(Radical Ventures)索取股权结构模型和 B 轮条款清单
独立模型性能验证第三方或独立基准,确认 GEN-1 的 99% 成功率和 1 小时适应能力主张所有性能主张都来自公司;买入前必须拿到独立验证将分阶段部署试点或独立技术审计作为投资条件
董事会构成和治理确认董事席位(Radical Ventures、NVIDIA NVentures)、独立董事、观察员权利治理结构决定决策推进速度,也决定投资人保护条款是否存在向公司索取董事会构成确认材料和治理文件
烧钱速度与现金续航轨迹实际月度烧钱额(非估算)、招聘计划、到 2027 年的算力资本开支排期估算 $10M/月只是代理值;实际烧钱额和成本拆分未知索取最近 4 个季度的季度管理账和董事会财务材料

五项尽调事项都会阻断买入建议。任一项缺失,都应维持跟踪评级。

[CV001, CV005, CV007, CV029, CV039, CV040]

8.5 展品

免责声明

基于截至 2026-06-30 的公开信息编制;私营公司财务和客户尽调可能实质性改变结论。

证据索引

结论
编号陈述可信度来源
CO001 Generalist AI, Inc. was founded in 2024. SO002, SO009, SO016
CO002 The company operates under the brand name "Generalist" while legally incorporated as Generalist AI, Inc. SO001, SO002
CO003 Generalist AI is headquartered in San Mateo, California. SO002, SO009, SO016
CO004 Generalist AI also operates an office in Somerville (Boston), Massachusetts. SO007, SO018
CO005 Generalist AI's stated mission is to build general intelligence for the physical world and make general-purpose robots a reality. SO002, SO005
CO006 Generalist AI builds embodied foundation models—AI systems designed to perceive, reason, and act across diverse physical environments and robot hardware, rather than for a single task or platform. SO002, SO004, SO016
CO007 Generalist AI is not a robotic hardware manufacturer; its product is the intelligence software layer intended to work across any robot form factor. SO002, SO006, SO016
CO008 Pete Florence is the CEO and co-founder of Generalist AI. SO002, SO009, SO019
CO009 Pete Florence was a senior research scientist at Google DeepMind and a senior author on PaLM-E and RT-2, two foundational papers in embodied AI. SO016, SO019, SO012
CO010 Andy Zeng is the Chief Scientist and co-founder of Generalist AI. SO002, SO009, SO019
CO011 Andy Zeng was previously a research scientist and technical lead at Google DeepMind and lead author of "Code as Policies." SO019, SO012
CO012 Andrew Barry is the CTO and co-founder of Generalist AI. SO002, SO009, SO016
CO013 Andrew Barry was previously a senior roboticist at Boston Dynamics, where he worked on Atlas, Spot, and Stretch robot platforms. SO019, SO012
CO014 The Generalist AI team includes alumni from OpenAI, Google DeepMind, and Boston Dynamics across its broader employee base. SO002, SO019
CO015 Generalist AI uses proprietary wearable "data hands" devices worn on human wrists to capture physical manipulation data for model training. SO012, SO004
CO016 GEN-0 was released on November 4, 2025 and represented Generalist AI's first major public model announcement. SO004, SO003, SO009, SO014
CO017 At the time of GEN-0's release, Generalist AI's pretraining dataset comprised over 270,000 hours of real-world manipulation data, growing at 10,000 hours per week. SO004, SO003, SO014
CO018 GEN-0 demonstrated scaling laws in robotics for the first time, showing that larger models trained on more physical data improve predictably across all downstream tasks. SO003, SO016, SO014
CO019 GEN-0 supports cross-embodiment generalization and has been tested on 6DoF, 7DoF, and 16+ DoF semi-humanoid robots. SO003
CO020 GEN-1 was released on April 2, 2026. SO004, SO009, SO010
CO021 GEN-1 improves average success rates to 99% on tasks where prior models achieved approximately 64%. SO004, SO010, SO016
CO022 GEN-1 completes dexterous tasks roughly three times faster than the prior state of the art. SO004, SO010, SO012
CO023 GEN-1 adapts to a new robotic task requiring only approximately one hour of task-specific robot data, regardless of robot embodiment. SO004, SO010
CO024 GEN-1 is trained on a proprietary dataset of over 500,000 hours of real-world physical interaction data. SO004, SO005, SO012
CO025 GEN-1 is trained approximately 99% from scratch, rather than fine-tuning an existing vision-language model, giving Generalist AI full architectural control. SO006, SO004
CO026 Generalist AI's physical data collection continues to grow at more than 10,000 hours per week as of the GEN-1 release. SO003, SO005
CO027 Generalist AI raised approximately $140 million at a post-money valuation of approximately $440 million in March 2025. SO012, SO023
CO028 Generalist AI announced $400 million in new funding on June 4, 2026. SO005, SO009, SO016
CO029 The June 2026 funding round values Generalist AI at a $2 billion post-money valuation. SO005, SO016, SO017
CO030 Generalist AI's total disclosed capital raised exceeds $500 million as of June 2026. SO005, SO009, SO016
CO031 Radical Ventures led the June 2026 funding round. SO005, SO009, SO016
CO032 New investors in the June 2026 round include 8VC, Union Square Ventures, Hanabi Capital, and Norwest. SO005, SO009, SO017
CO033 NVIDIA NVentures, Boldstart Ventures, Spark Capital, Bezos Expeditions, and NFDG participated significantly as existing investors in the June 2026 round. SO005, SO009, SO016
CO034 Angel investors in the June 2026 round include Eric Yuan (Zoom CEO), Bin Lin (Xiaomi co-founder), Fei-Fei Li, and Naval Ravikant. SO005, SO009, SO017
CO035 Generalist AI employs approximately 70 or more people as of mid-2026, based on press coverage. SO013, SO012
CO036 GEN-1 is offered under an early-access program to selected industry partners; no named customer or production deployment has been publicly confirmed. SO010, SO005
CO037 Generalist AI's data is collected across thousands of homes, warehouses, and workplaces worldwide through its global data collection network. SO003, SO004
CO038 Brad Porter, CEO of Cobot and a former Amazon robotics executive, publicly stated that scaling data alone against an imperfect architecture is insufficient and expensive, arguing architectural breakthroughs are needed alongside scale. SO012
CO039 Generalist AI plans to use the June 2026 funding to build next-generation models, scale its physical data engine, expand compute and training infrastructure, and work with industries deploying robots. SO005, SO009
CO040 Competitor Physical Intelligence was reportedly in talks to raise $1 billion at an $11 billion valuation as of April 2026, placing Generalist AI at approximately one-fifth the competitor's implied valuation. SO012
CO041 Generalist AI explicitly describes its long-term objective as "physical AGI"—general-purpose robotic intelligence capable of mastering any physical task. SO005, SO004, SO006
CO042 GEN-1 has demonstrated sustained repetitive performance including folding t-shirts 86 times consecutively, servicing robot vacuums 200+ times consecutively, and packing blocks 1,800 times consecutively without intervention. SO004, SO015
CO043 GEN-0 introduced Harmonic Reasoning, a novel training architecture enabling simultaneous thinking and acting in continuous time, critical for real-time physical systems where physics cannot be paused. SO003, SO014
CO044 The data flywheel described by Generalist AI is: scaling robot learning creates better models, better models enable more useful physical work, and data from real businesses drives the next generation of models. SO005, SO016
CO045 Generalist AI's pretraining dataset contains no robot data; instead the base model is trained on low-cost wearable device data from humans performing millions of activities, providing an existence proof that pretraining can lead to high mastery without large teleoperation or simulation datasets. SO004, SO006
CM001 Generalist AI is positioned as a software-only embodied foundation model provider—the "cognitive intelligence layer" for robots—targeting model licensing and inference revenue, not hardware manufacturing. SM019, SM017
CM002 The primary market boundary for an embodied AI software provider includes AI model licensing, cloud and edge inference, fine-tuning services, and software subscriptions tied to physical robotic deployment; robot hardware bodies, sensors, actuators, and grippers are excluded. SM014, SM017
CM003 Status-quo substitutes for an embodied foundation model include task-specific ROS/ROS2 programming (requiring months of custom engineering per task), vendor-specific embedded AI from OEMs (NVIDIA Isaac, ABB Omnicore, Fanuc AI), and traditional teach-pendant programming for structured pick-and-place. SM003, SM007
CM004 Adjacent markets that partially overlap with embodied AI software include robot simulation platforms (NVIDIA Isaac Sim, Mujoco), ROS/ROS2 industrial middleware distributions (valued at approximately $0.8 billion in 2026), robot-as-a-service delivery platforms, and warehouse management software with AI-driven robotic orchestration. SM007, SM003
CM005 Excluded from Generalist AI's addressable market are industrial PLCs, NC control software, traditional robot task-specific controllers (ABB RAPID, KUKA KRL), mechanical hardware subsystems, and warehouse management software without a robotic AI integration layer. SM014, SM007
CM006 The IFR identified AI and autonomy in robotics as the single most significant global robotics trend for 2026, stating that the shift from rule-based automation to intelligent self-evolving systems "makes embodied AI mainstream" in manufacturing and services. SM001, SM010
CM007 Mordor Intelligence values the global industrial robotics market (hardware plus integration) at $54.28 billion in 2026, projected to reach $94.38 billion by 2031 at an 11.7% CAGR; articulated units held 62.52% market share in 2025. SM003, SM011
CM008 The International Federation of Robotics reports the global market value of industrial robot installations reached $16.7 billion in 2026, an all-time high; 542,000 units were installed in 2024 with over 4.6 million robots in operational stock globally. SM001, SM002
CM009 Grand View Research estimates the global artificial intelligence in robotics market at $20.4 billion in 2025, projected to reach $182.7 billion by 2033 at a 32.0% CAGR (2026–2033); the software segment is expected to grow fastest at over 33% CAGR. SM006, SM012
CM010 Intel Market Research (IMR) values the AI robotics software market at $14.18 billion in 2026, projected to reach $38.76 billion by 2034 at a 15.3% CAGR; collaborative robots are expected to account for 30% of industrial robot sales by 2027. SM007, SM003
CM011 Grand View Research (via Research and Markets) estimates the embodied AI market at $6.5 billion in 2026, projected to reach $67.6 billion by 2033 at a 39.7% CAGR; North America represented 35.6% of the 2025 market; logistics and supply chain is the fastest-growing end-use sector at 42.2% CAGR. SM016, SM025
CM012 Grand View Research estimates the global warehouse automation market at $19.23 billion in 2023, projected to reach $59.52 billion by 2030 at an 18.7% CAGR; North America held a 36.7% share in 2023; retail and e-commerce is the dominant and fastest-growing application segment. SM004, SM005
CM013 Fortune Business Insights estimates the global warehouse robotics market at $7.35 billion in 2026, projected to reach $25.41 billion by 2034 at a 16.8% CAGR; Asia-Pacific held 51.7% share in 2025; AGVs represented 45.71% of the market in 2026. SM005, SM004
CM014 Research and Markets values the humanoid robot market at $8.3 billion in 2026, projected to reach $39 billion by 2030 at a 47.1% CAGR; leading commercial players include Tesla (Optimus), Figure AI, and Boston Dynamics (Electric Atlas). SM008, SM009
CM015 Mordor Intelligence estimates the global laboratory robotics market at $2.64 billion in 2026, projected to reach $3.5 billion by 2031 at a 5.76% CAGR; pharmaceutical and biotech labs are the largest segment. SM013
CM016 MarketsandMarkets estimates the AI in manufacturing market at $34.18 billion in 2025, projected to reach $155.04 billion by 2030 at a 35.3% CAGR; predictive maintenance, quality control, and adaptive assembly automation are the primary application areas. SM012, SM006
CM017 The software segment in AI in robotics is forecast to grow at the fastest rate—over 33% CAGR from 2026 to 2033—reflecting the accelerating shift from hardware-centric automation to software-defined, outcome-based robotic intelligence. SM006, SM007
CM018 Collaborative robots (cobots) are projected to account for approximately 30% of all industrial robot sales by 2027 and are posting the fastest growth at a 12.92% CAGR through 2031; Mordor Intelligence reported cobots at 12.92% CAGR in its January 2026 report. SM003, SM007
CM019 Global embodied AI startup funding in 2026 is projected to exceed $20 billion, with robot foundation model companies capturing over 40% of investment, according to QubitTool industry analysis. SM022, SM023
CM020 The automotive sector accounted for 35.86% of industrial robot demand in 2025; pharmaceuticals and healthcare show the highest CAGR at 13.52% through 2031; electronics represented 22% of installations. SM003, SM012
CM021 Pharmaceuticals and healthcare represent the fastest-growing buyer segment for industrial robotics at a 13.52% CAGR through 2031, driven by sterile compounding, personalized medicine, and continuous robotic production lines. SM003, SM013
CM022 The e-commerce segment is projected to hold 47.21% of the warehouse robotics market in 2026, and is the dominant and fastest-growing application area for warehouse automation; Amazon has announced $1 billion in warehouse automation investment. SM005, SM004
CM023 Robot hardware OEM manufacturers—including ABB, Fanuc, KUKA, and emerging humanoid companies (Figure AI, Boston Dynamics, 1X Technologies)—represent a software buyer category that embeds or licenses AI to differentiate their hardware products; NVIDIA, ABB, and others have proprietary AI software stacks. SM007, SM024
CM024 Figure AI's BMW Spartanburg deployment saw Figure 02 run daily 10-hour shifts across an 11-month period, loading over 90,000 parts across 1,250+ runtime hours and contributing to more than 30,000 X3 vehicles — a publicly confirmed enterprise production deployment of an AI-driven humanoid robot platform. SM021, SM020
CM025 Logistics and warehouse operators increasingly adopt Robot-as-a-Service (RaaS) models to convert robotic deployment from capital expenditure to operating expenditure, lowering adoption friction; Figure AI and others explicitly use RaaS as a commercial model. SM021, SM005
CM026 Laboratory automation buyers in pharma, biotech, and diagnostics drive the $2.64 billion lab robotics market (2026); these buyers require GMP compliance, regulatory traceability (CFR Part 11, IQ/OQ/PQ validation), and are cautious but durable adopters when compliance requirements are met. SM013, SM012
CM027 Primary adoption triggers for AI robotics software buyers are: (1) labor shortage and wage inflation making automation ROI compelling, (2) e-commerce growth demanding fulfillment speed, (3) reshoring mandates reducing the competitive labor cost advantage of offshoring, and (4) government subsidies shortening payback periods. SM001, SM010
CM028 The IFR documented 420,000 open skilled-trade positions in German factories in 2025 and forecasts a projected 2.1 million manufacturing worker deficit in the United States by 2030; these structural labor gaps are the primary demand signal for robotic automation. SM001, SM010
CM029 The IFR's Top 5 Global Robotics Trends for 2026 are: (1) AI and Autonomy in Robotics, (2) Robots gaining versatility through IT/OT convergence, (3) Humanoids proving reliability and efficiency, (4) Safety and Security in AI-driven robotics, and (5) Robots as allies in tackling labor gaps. SM001, SM010
CM030 US reshoring investment: $47 billion in factory investments was announced in 2024-2026, most citing robotics as a prerequisite for cost-competitive domestic production; Mordor Intelligence also notes $28 billion in Mexico near-shoring pledges relying on articulated robots to meet USMCA content rules. SM003, SM011
CM031 Median payback for industrial robots is now approximately 1.3 years (16 months) with cobots achieving payback in as little as 6–18 months; 78% of robotics deployments meet or exceed their ROI projections when modeled rigorously including full labor burden. SM003, SM011
CM032 Foundation model-based rapid task adaptation (approximately 1 hour of task-specific data for GEN-1) vs. weeks or months for bespoke ROS programming makes marginal task automation economically viable; this is a structural growth driver for AI robotics software adoption. SM014, SM022
CM033 China's Made in China 2025 program allocated approximately CNY 180 billion ($25.2 billion) to robotics through 2026, reimbursing up to 40% of equipment costs; South Korea doubled small-manufacturer subsidies in 2025; Germany's Digital Now fund disburses EUR 500 million annually through 2027. SM003, SM001
CM034 Germany's Digital Now fund disburses EUR 500 million annually through 2027, provided new equipment complies with Industrie 4.0 interoperability standards, directly subsidizing automation investment for mid-size manufacturers. SM003
CM035 The US CHIPS and Science Act channels $11 billion toward semiconductor workforce and clean-room automation, spiking demand in Arizona, Ohio, and Texas fabs ramping in 2026–2027; CHIPS Act grants and tariff pressures are jointly driving domestic automation investment. SM003, SM012
CM036 Developing a competitive humanoid robot platform required an estimated $3–4 billion in Tesla Optimus R&D from 2022 to 2024; Figure AI reportedly burns $200–300 million annually to sustain its development cycle; these figures establish the capital intensity of the frontier embodied AI market. SM014, SM021
CM037 Fine motor manipulation requiring integrated tactile and force feedback—the "chopstick problem"—remains a fundamental technical constraint; most current embodied AI systems rely primarily on visual input, which is insufficient for tasks requiring physical contact sensing. SM014, SM015
CM038 Physical manipulation data must be "earned" through robot or wearable interaction and cannot be scraped from the internet; the resulting data scarcity creates a structural bottleneck for foundation model scale-up distinct from text-based AI. SM014, SM022
CM039 Legacy factory PLCs, MES systems, and ERP platforms create $40,000–$80,000 in typical installation overhead per automation cell; certified integrators bill $150–240 per hour in North America and Western Europe; Mordor Intelligence notes 34% of planned 2025 projects in Germany slipped due to integrator shortages. SM003, SM014
CM040 Safety and liability frameworks for AI-driven robotics are actively evolving (ISO 9283, ISO/TS 15066, AI Act in EU) but no jurisdiction has enacted AI-specific embodied AI liability regulation as of June 2026; the IFR cited "Safety and Security in AI-driven robotics" as a top 2026 industry trend requiring governance clarity. SM001, SM015
CM041 VLA (Vision-Language-Action) models require specialized edge hardware for on-device inference because datacenter-scale compute introduces unacceptable latency for real-time robotic control; competition for specialized edge AI chips is expected to intensify in late 2026 as companies move from pilot programs to fleet deployments. SM014, SM022
CM042 The US had 87,000 open industrial machinery mechanic vacancies in 2025 with median time-to-fill above 90 days; Germany's VDMA reported 34% of planned 2025 industrial automation projects slipped three months or more due to integrator shortages. SM003, SM010
CM043 Only approximately 25% of warehouses globally have implemented any form of automation, with only 10% using advanced technology; this represents a large greenfield adoption opportunity but also implies that most of the addressable market is still in the pre-automation phase. SM004, SM005
CM044 The three primary AI robotics sizing lenses—$6.5B (GVR Embodied AI), $14.18B (IMR AI Robotics Software), and $20.4B (GVR AI in Robotics)—reflect incompatible scope definitions, not a TAM-SAM-SOM hierarchy; no single figure represents the revenue pool that exclusively accrues to a software-only embodied foundation model provider. SM016, SM007, SM006
CM045 No public market research source tracks "foundation model licensing for cross-embodiment dexterous manipulation" as a discrete market segment; the closest proxy (embodied AI at $6.5B) includes hardware-adjacent components and requires untested assumptions to derive a pure software SAM. SM016, SM007
CM046 Physical Intelligence (pi.ai) raised $400 million in late 2024 at a $2.4 billion valuation; the company was reportedly in discussions for an additional $1 billion follow-on round in 2026; this positions Physical Intelligence as Generalist AI's closest comparable competitor in foundation model licensing for robotics. SM018, SM020
CP001 Generalist AI's GEN-1 model achieves an average 99% success rate on physical tasks where previous models (GEN-0 and π0) achieved 64%, and completes tasks approximately 3x faster than prior state-of-the-art with only one hour of robot-specific adaptation data. SP001, SP003
CP002 Generalist AI raised $400 million in a Series B at a $2 billion valuation in June 2026, led by Radical Ventures with participation from NVIDIA and Fei-Fei Li (via affiliated entities). SP021, SP023
CP003 Skild AI raised $1.4 billion in a Series C round at a $14 billion post-money valuation in early 2026, led by SoftBank with participation from NVIDIA and Samsung Ventures. SP015, SP013
CP004 Skild AI acquired Zebra Technologies' Robotics Automation business, including the Symmetry Fulfillment orchestration platform, in April 2026 in a transaction involving cash and equity consideration for Zebra Technologies. SP012, SP013, SP014
CP005 Physical Intelligence was in talks in March 2026 to raise approximately $1 billion in a new funding round at a valuation exceeding $11 billion, with Founders Fund and Lightspeed Venture Partners in negotiations, representing a doubling of its $5.6 billion November 2025 valuation. SP020, SP016
CP006 Figure AI has raised approximately $1.9 billion in total funding at a $39 billion post-money valuation following its September 2025 Series C, led by Parkway Venture Capital with participation from Brookfield Asset Management, NVIDIA, Microsoft, OpenAI Startup Fund, and Jeff Bezos via Bezos Expeditions. SP005, SP006
CP007 Figure AI's Figure 02 deployment at BMW's Spartanburg plant loaded over 90,000 parts across more than 1,250 runtime hours contributing to over 30,000 X3 vehicles, representing the most thoroughly documented commercial humanoid robot deployment by any competitor as of mid-2026. SP005, SP006
CP008 Figure AI's Helix VLA system operates as three hierarchical layers — S0 (whole-body control trained on 1,000+ hours of human motion data plus 200,000 parallel sim-to-real environments), S1 (visuomotor control), and S2 (semantic reasoning) — running entirely on embedded low-power GPUs without cloud connectivity. SP005, SP006
CP009 Boston Dynamics commercially launched its electric Atlas humanoid robot at CES 2026, with all 2026 production units committed to Hyundai's Robotics Metaplant Application Center (RMAC) and Google DeepMind; additional customers are planned for 2027. SP011, SP009
CP010 Boston Dynamics announced at CES 2026 a partnership with Google DeepMind to integrate Google DeepMind foundation models into Atlas for enhanced cognitive capabilities; Hyundai Mobis will supply actuators, and Hyundai Motor Group plans a factory capable of 30,000 robots per year. SP011, SP007
CP011 NVIDIA released Isaac GR00T N1.6 as open-source software at CES 2026, available on Hugging Face under a permissive license, along with Cosmos Transfer 2.5, Cosmos Predict 2.5, and Cosmos Reason 2; these models are free to any developer, with NVIDIA monetizing through GPU hardware (Jetson Thor, DGX) and simulation subscriptions. SP009, SP010
CP012 NVIDIA's open-source robotics community includes 2 million robotics developers, and GR00T N models and Isaac Lab-Arena are integrated into the HuggingFace LeRobot library alongside HuggingFace's 13 million AI builder community. SP009, SP010
CP013 Skild AI reported approximately $30 million in annual recurring revenue within months of its 2025 commercial launch, making it the only software-pure robot foundation model competitor with disclosed commercial revenue traction as of mid-2026. SP015, SP017
CP014 Google DeepMind's Gemini Robotics ER 1.6 remains behind an early-access/waitlist program as of June 2026; no public pricing for the robotics API has been released, and deployment partnerships include Boston Dynamics, Apptronik, and Agility Robotics. SP007, SP008
CP015 Google DeepMind announced a mid-2026 "on-device" variant of Gemini Robotics enabling local inference for latency-sensitive industrial applications, signaling an intent to address the cloud-connectivity limitation that previously prevented deployment in air-gapped environments. SP008, SP017
CP016 Amazon's 2024 talent acquisition of Covariant's founders (Pieter Abbeel, Peter Chen, Rocky Duan) plus approximately 25% of Covariant's team, with a non-exclusive IP license to Covariant's foundation AI, gives Amazon in-house robot AI capability across its 750,000+ robot fleet while leaving Covariant as an independent entity. SP016, SP017
CP017 Generalist AI's GEN-1 assembles a box in 12.1 seconds, versus approximately 34 seconds for GEN-0 and Physical Intelligence's π0 on identical boxes — a claimed 2.8x speed advantage — but this comparison is company-self-reported and has not been validated by an independent third-party benchmark as of June 2026. SP001, SP003
CP018 OpenAI launched a dedicated robotics division in 2026, having previously invested in both Figure AI and Physical Intelligence; this transforms a former investor into a direct competitor for the robot foundation model category. SP017, SP018
CP019 Generalist AI's GEN-1 differentiates architecturally by training from scratch on proprietary wearable-sensor human physical-interaction data rather than fine-tuning a VLM backbone, explicitly rejecting the VLA and world-model paradigms; GEN-1 claims adaptation to any new robot embodiment with only one hour of robot-specific data. SP001, SP002
CP020 No competitor in the robot foundation model category has published pricing for its AI software as of June 2026; all enterprise pricing is deal-specific and undisclosed, with NVIDIA GR00T as the sole exception at zero cost. SP008, SP016, SP012
CP021 NVIDIA GR00T N1.6 is available as open-source software at zero licensing cost, establishing a permanent willingness-to-pay floor for robot foundation model software and creating sustained downward pressure on paid model pricing by any commercial competitor. SP009, SP017
CP022 Physical Intelligence open-sourced its π0 model weights via the openpi repository, creating an academic developer community that can benchmark against and build on its architecture, while also signaling that value in the market lies in training data and deployment rather than model weights. SP025, SP016
CP023 Skild AI's Zebra acquisition gives it the Symmetry Fulfillment orchestration platform — described as "one of the most battle-tested warehouse robotics platforms in the industry" — along with enterprise WMS integrations and a logistics customer channel that pure-model competitors lack. SP012, SP013
CP024 Generalist AI has not publicly disclosed any named commercial customers, deployment partners, or commercial revenue as of June 2026; GEN-1 access is limited to an early-access partner program with undisclosed terms and scale. SP001, SP021
CP025 Boston Dynamics' Atlas electric humanoid robot has an estimated price of $150,000–$420,000 per unit depending on configuration; all 2026 production is committed to Hyundai and Google DeepMind, foreclosing the most commercially validated hardware distribution channel from Generalist AI's software deployment until at least 2027. SP011, SP017
CP026 The CB Insights physical AI market map (January 2026) identifies 70+ companies across 10 physical AI model categories, with robotics sector funding of $40.7 billion in 2025 — up 74% year-over-year and representing approximately 9% of all global venture funding. SP016, SP017
CP027 NVIDIA's free open-source GR00T N-series poses a structural commoditization risk to Generalist AI's commercial model: if GR00T N-series reaches capability parity with GEN-2 or GEN-3, the model layer becomes a commodity embedded in GPU hardware, and pricing power for paid robot foundation models collapses. SP009, SP011, SP017
CP028 Generalist AI's proprietary dataset of 500,000 hours of physical interaction data from wearable sensors is claimed to be the world's largest of its kind; this claim is not independently verified and alternative scaling paths exist via NVIDIA Cosmos synthetic data and the RT-X open dataset aggregating 100+ robot embodiments. SP001, SP016
CP029 Generalist AI's data flywheel has not yet activated commercially as of June 2026 — the data engine logic requires real-world deployments to generate new training data, but no commercial deployments have been publicly confirmed, leaving the flywheel at theoretical stage compared to Skild AI's live logistics deployments via Zebra AMR fleet. SP012, SP024
CP030 NVIDIA's GR00T N-series is adopted by NEURA Robotics, Boston Dynamics (Jetson Thor integration), Franka Robotics, Humanoid, and LG Electronics as of CES 2026, giving NVIDIA's open-source model a broader installed hardware base than any paid software competitor. SP009, SP010
CP031 Skild AI's April 2026 Zebra acquisition combined with ~$30M ARR and $14B valuation gives it enterprise distribution infrastructure, commercial revenue, and capital scale that are each materially ahead of Generalist AI's current position on all three dimensions. SP012, SP015
CP032 Generalist AI scores an analyst-estimated 9/10 on research/founder pedigree and 7/10 on technical differentiation from GEN-1's novel architecture, but 0/10 on commercial revenue traction, 1/10 on enterprise distribution infrastructure, and 0/10 on independent benchmark validation — a profile that reflects an early-stage research-to-product transition. SP001, SP023, SP016
CP033 On a competitive positioning map of capital/resources versus commercial traction, Generalist AI occupies the low-capital, zero-traction quadrant; Skild AI leads software-pure competitors on commercial traction; Figure AI leads on documented enterprise deployment; NVIDIA leads on distribution reach via free open-source model. SP015, SP016, SP006
CP034 No independent third-party benchmark has directly compared Generalist AI's GEN-1 against Physical Intelligence's π0.5 or Skild Brain on identical hardware and tasks as of June 2026; all performance advantage claims (99% success rate, 3x speed) are company self-reported. SP001, SP017
CP035 The window for Generalist AI to establish defensible enterprise distribution before Skild AI's Zebra-enhanced platform becomes the default logistics AI stack is an analyst-estimated 12–18 months; beyond this window, the competitive moat risk shifts from technical to commercial. SP015, SP023
CI001 Generalist AI is a pure software company — it does not manufacture robots — and positions itself as the cross-form-factor "intelligence layer" that works across robotic embodiments including industrial arms, humanoids, and mobile platforms. SI009, SI023
CI002 Generalist AI has not publicly disclosed pricing, ARR, revenue, or named customers as of the June 2026 runDate. SI009, SI010, SI013, SI014
CI003 Generalist AI's current commercial mechanism is an early-access partner program for GEN-1, where selected industry partners receive model access, enabling a data flywheel where real business deployments feed the next model generation. SI009, SI010, SI006
CI004 The June 2026 funding announcement states that a data flywheel is "beginning to take shape: real businesses are generating task data that feeds successive model generations," implying early contractual or operational partner deployments exist. SI009, SI014
CI005 Generalist AI explicitly does not manufacture hardware; its commercial proposition is to serve as the software intelligence layer, analogous to a cloud AI API, which structurally supports high gross margins at scale relative to hardware-integrated robotics vendors. SI009, SI018, SI023
CI006 Generalist AI's job listings as of June 2026 include a single "Applied AI & Partnerships" commercial-facing role, alongside research, ML infrastructure, and operations roles — indicating early-stage GTM build with lean commercial staffing relative to research headcount. SI011, SI012
CI007 Generalist AI claims GEN-1 "crosses a threshold to commercial viability across a broad range of tasks," but no commercial revenue figures, pricing, or customer confirmation accompany this claim. SI010, SI006, SI013
CI008 Physical Intelligence operates a B2B SaaS model charging $300 per connected robot per month, yielding recurring revenue that scales with fleet deployments, according to Sacra analyst research. SI001, SI002
CI009 Physical Intelligence raised a $600M Series B in November 2025 at a $5.6B valuation and was reportedly in talks for an additional $1B raise at an $11B valuation in April 2026. SI001, SI003, SI005, SI007
CI010 Skild AI, another embodied AI foundation model company, reportedly achieved approximately $30M ARR in 2025, according to industry analysis, indicating that early commercial revenue is achievable in the embodied AI software segment within two years of launch. SI002
CI011 Physical Intelligence's per-robot SaaS pricing of ~$300/robot/month is the closest publicly disclosed pricing analog for Generalist AI's likely business model, but Generalist AI has not confirmed any such pricing structure. SI001, SI002
CI012 Generalist AI raised approximately $140 million in its first round in March 2025 at a post-money valuation of approximately $440 million, with investors including Spark Capital, NVIDIA NVentures, Bezos Expeditions, and Boldstart Ventures. SI016, SI020, SI013
CI013 Generalist AI raised $400 million in a second round announced June 4, 2026, at a $2 billion post-money valuation led by Radical Ventures, bringing total disclosed capital raised to more than $500 million. SI009, SI013, SI014, SI020
CI014 The June 2026 round included new investors 8VC, Union Square Ventures, Hanabi Capital, and Norwest; all major existing investors — NVIDIA NVentures, Boldstart Ventures, Spark Capital, Bezos Expeditions, and NFDG — participated significantly; new angels included Fei-Fei Li, Eric Yuan, Bin Lin, and Naval Ravikant. SI009, SI014, SI025
CI015 Generalist AI's stated planned use of the $400M June 2026 raise is: building next-generation models, scaling the physical data engine, expanding compute and training infrastructure, and working with industries to bring systems into everyday use. SI009, SI014, SI018
CI016 No debt facilities, convertible notes, credit lines, or project-finance obligations have been publicly disclosed for Generalist AI as of June 2026; the capital structure appears to be entirely equity-funded. SI009, SI013, SI014
CI017 Generalist AI's post-money valuation of $2B is approximately one-fifth of Physical Intelligence's reported $11B round-in-talks valuation as of April 2026, reflecting Generalist's earlier commercial stage and smaller capital base despite claiming architectural differentiation. SI001, SI007, SI013
CI018 Training a frontier robotics foundation model at GEN-1 scale (10B+ parameters, 500K+ hours of video-action data, trained from scratch) is estimated by industry analysts to require $10M–$100M in compute per training run. SI003, SI018
CI019 Generalist AI rebuilt its "distributed training infrastructure to support petabytes of physical interaction data as a first-class citizen" for GEN-1, indicating significant ML infrastructure capital expenditure and ongoing storage and compute costs. SI010, SI006
CI020 Industry benchmarks suggest monthly burn rates of $5M–$15M/month for embodied AI startups with 50–100 employees operating intensive compute and data collection infrastructure; this is a benchmark estimate and has not been confirmed for Generalist AI. SI003, SI018
CI021 Generalist AI's active job listings as of June 2026 span research, ML infrastructure, compute optimization, robotics controls, data collection operations, and partnerships — indicating active hiring across R&D, infrastructure, and early GTM dimensions simultaneously. SI011, SI012
CI022 NVIDIA NVentures has co-invested in both of Generalist AI's rounds, signaling strategic alignment: Generalist's foundation models depend on GPU compute infrastructure, and large-scale embodied AI deployments represent significant NVIDIA hardware demand. SI009, SI013, SI025
CI023 Generalist AI's data hands wearable devices are claimed to produce training data at lower cost per hour than teleoperation rigs used by competitors, as human operators capture naturalistic dexterity at scale without specialized equipment — reducing data acquisition costs per training hour. SI010, SI024
CI024 Generalist AI's commercialization pathway as of June 2026 is limited to a controlled early-access partner program, suggesting a high-touch enterprise sales motion with long cycles and no self-serve or product-led growth component. SI010, SI012, SI006
CI025 Fraser Kelton of Spark Capital (Generalist AI investor) stated: "every time they've scaled up these models, the returns on generalization have been profound," citing GPT-3 commercial viability for copywriting as the analogous progression path for robotics. SI016
CI026 8VC's investment memo describes "remarkable early commercial traction" at Generalist AI as a signal supporting their investment, without providing customer names, revenue figures, or deployment scale; this is a qualitative investor signal, not a confirmed commercial metric. SI015
CI027 Revenue, ARR, gross margin, monthly burn rate, and customer count are not publicly disclosed by Generalist AI; all are private-evidence-only metrics as of June 2026. SI009, SI010, SI013
CI028 Generalist AI has not publicly named any commercial customer, signed contract, or production deployment as of the June 2026 runDate; GEN-1 early-access partners are referenced generically in all official communications. SI009, SI010, SI006
CI029 The most recent public commercial signal from Generalist AI as of the runDate is the June 4, 2026 funding announcement, which references a data flywheel "beginning to take shape" and the April 2026 GEN-1 early-access launch; no new customer or revenue announcements have been published since. SI009, SI013
CI030 Brad Porter, CEO of Cobot and former Amazon VP of Robotics, stated that "brute-forcing a huge amount of data against a not-perfect architecture is really expensive and not necessarily going to get you the result you want," directly challenging the economic viability of data-scaling thesis companies like Generalist AI. SI016, SI017, SI004
CI031 Industry analysis of Physical Intelligence's capital needs notes that training generalist robot policies requires massive and ongoing physical interaction data acquisition, with each new capability requiring "more robots, more environments, more variation"—implying a capital treadmill that does not terminate at a single funding round. SI003
CI032 Yann LeCun, Meta Chief AI Scientist, is cited in industry coverage of GEN-1 as arguing that "world models must learn through observation rather than just action-token prediction," representing a structural challenge to the training approach underpinning Generalist AI's thesis. SI017
CI033 Generalist AI has no publicly confirmed revenue as of June 2026; the business is in an early- access partner stage, operationally equivalent to a pre-commercial or controlled-beta phase. SI009, SI010, SI013
CI034 If Generalist AI adopts a cloud-hosted software delivery model (per-robot API or subscription), the long-run gross margin profile should approach 60–80% at maturity—consistent with software SaaS benchmarks—but near-term gross margin will be materially depressed by training compute amortization and inference serving costs. SI001, SI018
CI035 At Physical Intelligence's $300/robot/month SaaS pricing analog, achieving $10M ARR would require approximately 2,778 billed robots, and breaking even at a $10M/month burn rate would require approximately 33,333 active billed robots—a fleet scale many years away at any current early-access adoption pace. SI001, SI003
CI036 Physical Intelligence's analyst-described cost profile features gross margins "similar to software despite some pass-through costs," with inference compute and data licensing as the primary COGS drivers—an analog applicable to Generalist AI's model delivery costs. SI001
CI037 Generalist AI's $2B post-money valuation is set entirely on future optionality and team quality, not on current-period revenue; at any plausible current ARR level, the implied revenue multiple is very high—consistent with early-stage deep-tech venture pricing rather than fundamental financial performance. SI013, SI009, SI003
CI038 Generalist AI's physical data engine collects more than 10,000 hours of physical interaction data per week as of April 2026, a rate that generates ongoing operational cost but also accumulates a growing proprietary dataset moat that competitors must match through capital or teleoperation rigs at higher per-hour cost. SI010, SI009, SI024
CI039 Generalist AI, Inc. filed Form D (Notice of Exempt Offering of Securities, Item 06b under Reg D Rule 506(b)) with the SEC on June 5, 2026, confirming a private placement equity offering; the filing lists the company as incorporated in Delaware with a San Francisco, CA business address and assigns CIK 0002137708. SI028, SI029, SI013
CE001 GEN-1 is a large multimodal foundation model designed to emit real-time robotic actions, positioned as the "intelligence layer" or "cognitive brain" of any robot—not a hardware product. SE004, SE001
CE002 GEN-1 is hardware-agnostic: it works across multiple robot form factors including 6DoF, 7DoF, and 16+DoF semi-humanoid systems without requiring hardware-specific retraining of the base model. SE004, SE003, SE015
CE003 Generalist AI is not a hardware manufacturer; its commercial proposition is a pure software and model layer that any robotic hardware OEM or end-deployer can integrate. SE001, SE002, SE022
CE004 Data hands are lightweight, handheld or wrist-mounted ergonomic devices that capture human manipulation data with near-natural force feedback, preserving the natural sensorimotor loop. SE009, SE004, SE015
CE005 Unlike teleoperation rigs, data hands preserve the natural sensorimotor loop: after a brief acclimation period, operators stop thinking deliberately and start reacting, producing trajectories that capture reflexes, micro-corrections, and real-time error recovery. SE009, SE015
CE006 Generalist AI's pretraining dataset reached 270,000 hours of real-world physical interaction data at the GEN-0 release in November 2025. SE003, SE016, SE015
CE007 The pretraining dataset grew to over 500,000 hours of physical interaction data by the GEN-1 release in April 2026, the self-described world's largest real-world manipulation dataset. SE004, SE005, SE006
CE008 Generalist AI's physical interaction dataset is growing at more than 10,000 hours per week as of the GEN-1 release in April 2026. SE003, SE004
CE009 Data collection takes place across thousands of homes, warehouses, workplaces, and specialized environments—including bakeries, laundromats, and factories—worldwide. SE003, SE004
CE010 GEN-1's base pretraining dataset contains no robot data; it is entirely composed of human manipulation data captured through data hands wearable devices, with robot-specific adaptation occurring only in the 1-hour fine-tuning phase. SE004, SE006, SE015
CE011 Harmonic Reasoning, first introduced in GEN-0, creates an asynchronous continuous-time "harmonic" interplay between streams of sensing and acting tokens, enabling simultaneous thinking and acting without physics-pausing System 1/2 cycle architectures. SE003, SE004, SE015
CE012 Harmonic Reasoning allows Generalist AI to scale to very large model sizes without depending on System 1/System 2 architectures or inference-time guidance, contrasted with Figure AI's Helix and other sequential approaches. SE003, SE004
CE013 GEN-0 demonstrated scaling laws in robotics for the first time: a power-law relationship between pretraining data scale and downstream post-training performance, holding across all measured tasks simultaneously. SE003, SE016, SE015
CE014 GEN-0 supports cross-embodiment generalization and has been tested on 6DoF, 7DoF, and 16+DoF semi-humanoid robots in internal experiments. SE003, SE016
CE015 GEN-1 is trained approximately 99% from scratch rather than fine-tuned from an existing vision-language model, giving Generalist AI full architectural control independent of the VLM ecosystem. SE006, SE004, SE015
CE016 GEN-1 achieves 99% average success rates on tested tasks where GEN-0 achieves 64% and a from-scratch baseline (no pretraining) achieves approximately 19%. SE004, SE005, SE013
CE017 GEN-1 completes tasks approximately 3x faster than the prior state of the art: box folds in 12.1 seconds versus ~34 seconds for GEN-0 and pi-0 (Physical Intelligence) on identical boxes. SE004, SE005, SE013
CE018 GEN-1 adapts to a new physical task with approximately one hour of robot-specific data, simultaneously adapting to the new robot embodiment and the new task for the first time. SE004, SE005, SE013
CE019 GEN-1 incorporates post-training advances including supervised fine-tuning (SFT), reinforcement learning from experience (RL), multimodal human guidance, and new inference-time techniques beyond those in GEN-0. SE004, SE006
CE020 GEN-1 demonstrates improvisational intelligence: responding to scenarios well outside the training distribution, such as a bumped washer being regrasped via extrinsic dexterity, or bimanual in-hand manipulation improvised without explicit training. SE004, SE017
CE021 GEN-1 sustained task performance in extended demo runs: 86 consecutive T-shirt folds, 200+ consecutive robot vacuum services, and 1,800+ consecutive block packs without human intervention. SE004, SE017, SE013
CE022 At NVIDIA GTC in March 2026, Generalist AI ran a live public demo of GEN-0 on Universal Robots' new UR7e + MiR mobile manipulation platform—a hardware configuration that had not previously existed—achieving reliable performance within a handful of days of receiving the robot. SE001, SE010, SE012
CE023 GEN-1 is offered under an early-access program to selected industry partners beginning April 2, 2026; deployment inquiries are routed via partnerships@generalistai.com. SE004, SE005, SE013
CE024 Generalist AI's stated target industry sectors for GEN-1 deployment include apparel, manufacturing, logistics, automotive, and electronics, based on scaling law experiments and partner-inspired applications mentioned in the GEN-0 blog. SE003, SE004, SE005
CE025 GEN-0's scaling experiments revealed a "ossification" phase transition: models below ~7B parameters cannot absorb large-scale pretraining data (weights freeze prematurely), while 7B+ models continue to improve; GEN-0 was scaled to 10B+ parameters. SE003, SE016
CE026 Generalist AI's data processing infrastructure includes multi-cloud contracts, custom dataloaders, O(10K) cores for continual multimodal data processing, and infrastructure capable of absorbing 6.85 years of real-world manipulation experience per day of training. SE003, SE004
CE027 Building GEN-1 required months improving training stability, building custom kernels, inventing new forms of paged attention for real-time inference, and honing post-training techniques including theoretical RL and multimodal human guidance. SE004, SE006
CE028 Generalist AI formally defines "mastery" as the combination of three capabilities: reliability (consistently accomplishing tasks), speed (task completion time), and improvisational intelligence (recovering from unexpected scenarios creatively). SE004, SE013, SE015
CE029 GEN-1 is not a fine-tuned VLA or a world model; the company deliberately avoids these labels, having co-invented VLAs (PaLM-E, RT-2) and published on world models, but treating GEN-1 as goal-driven architecture that crosses those categorical boundaries. SE004, SE006, SE015, SE028
CE030 Physical commonsense is defined by Generalist AI as the reactive, closed-loop sensorimotor intelligence—intuition for forces, friction, compliance, and uncertainty—that emerges from large-scale physical interaction pretraining, not from language model pretraining. SE009, SE004
CE031 GEN-1's improvisational intelligence is simultaneously a strength (enabling spontaneous recovery) and a potential liability: emergent behaviors are physical actions with real consequences and may exceed intended boundaries; the company acknowledges this alignment gap. SE004, SE015, SE022
CE032 GEN-1's limitations acknowledged by the company include: not all tasks achieve 99%+ success rates, some tasks would require even higher rates or speeds for real-world utility, and the model does not solve all physical tasks. SE004, SE013, SE015
CE033 GEN-1 demonstrates 10x data efficiency compared to GEN-0: it achieves comparable performance to GEN-0 with 10x less task-specific data and fine-tuning steps. SE004, SE006
CE034 Generalist AI's global data collection network includes thousands of data collection devices and robots deployed across diverse geographies and environments. SE003, SE009
CE035 Data foundry partners are classified into three modes: Class 1 (specific tasks), Class 2 (mixed), and Class 3 (do-anything), enabling A/B comparison of data mixture effects on downstream pretraining quality. SE003
CE036 No public API documentation, developer SDK, public portal, or detailed integration guide for GEN-1 has been disclosed by Generalist AI as of June 30, 2026. SE001, SE008
CE037 GEN-1 inference uses a custom paged attention kernel, described as an "evolution of the way we do inference with Harmonic Reasoning," designed for real-time action prediction. SE004
CE038 Generalist AI's architecture philosophy is explicitly goal-driven rather than method-driven: the company selects methods based on measurable outcomes (e.g., 99%+ success with 1-hr data) rather than aligning with current academic trends (VLAs, world models). SE006, SE004
CE039 GEN-1 achieves box folds in ~12 seconds compared to ~34 seconds for pi-0 (Physical Intelligence) and GEN-0 on identical boxes—a 2.8x speed advantage used as the primary published benchmark comparison. SE004, SE015, SE025
CE040 Brad Porter, CEO of Cobot and former Amazon Robotics executive, publicly criticized the data-scaling approach, arguing that scaling data alone without architectural improvements is expensive and insufficient, citing ImageNet and transformers as historical analogies. SE015, SE022
CU001 GEN-1 was made available under an early-access program to selected industry partners starting April 2, 2026. SU001, SU005
CU002 No named customers, customer count, ARR, pricing, or disclosed contract terms have been publicly released by Generalist AI as of June 30, 2026. SU012, SU001
CU003 Generalist AI's official communications name apparel, manufacturing, logistics, automotive, and electronics as target industry verticals for GEN-1 deployment. SU001, SU002
CU004 Data foundry partners contribute physical interaction data to Generalist AI's pretraining dataset, classified into Class 1 (specific tasks), Class 2 (mixed), and Class 3 (do-anything) modes. SU026, SU001
CU005 Data collection takes place across "thousands of homes, warehouses, workplaces, and specialized environments worldwide, including bakeries, laundromats, and factories" per the GEN-1 launch blog. SU001, SU026
CU006 Whether data foundry partners are paying customers or unpaid data-contributing participants is not disclosed in any public Generalist AI communication. SU001, SU002
CU007 Generalist AI's sole customer-facing contact pathway is partnerships@generalistai.com, with no published API, SDK, partner portal, or self-serve trial environment as of June 2026. SU001, SU020
CU008 Generalist AI as of June 2026 has only one open commercial role—"Applied AI & Partnerships"—on its Ashby jobs page, consistent with a lean, founder-led GTM serving fewer than ten strategic accounts. SU020, SU011
CU009 8VC, in its June 2026 investment memo, described "remarkable early commercial traction" as a key signal supporting their investment in Generalist AI. SU004, SU007
CU010 Fraser Kelton, Spark Capital partner and former OpenAI product lead, cited "early commercial validation in robotics scaling" in connection with the investment in Generalist AI. SU011, SU004
CU011 Generalist AI's June 2026 funding blog states that "data from real businesses drives the next generation of more capable models," implying contractual or operational relationships with industrial operators generating training data. SU002, SU006
CU012 Universal Robots, the world's #1 cobot manufacturer by volume, invited Generalist AI to demo GEN-0 at NVIDIA GTC 2026 (March 16–21, 2026) on their new mobile manipulation platform. SU003, SU008
CU013 The GTC demo featured a UR7e arm on a MiR mobile base and Vention frame—hardware that had never previously existed in this configuration—and Generalist adapted GEN-0 to it in less than a week. SU003, SU008
CU014 The GTC demo ran continuously for all open exhibit hours with no scheduled time slots, using no data collected inside the GTC exhibition hall, demonstrating generalization to a novel physical environment. SU003
CU015 Universal Robots has a global install base of over 75,000 cobots deployed in more than 10,000 customer facilities, making it the largest-volume cobot manufacturer globally. SU008, SU015
CU016 Whether the UR GTC demo partnership led to any commercial licensing, deployment, or distribution agreement has not been disclosed by either Generalist AI or Universal Robots as of June 2026. SU003, SU008
CU017 The NVIDIA GTC March 2026 press release describes Generalist AI as "using Cosmos to explore generating synthetic data" — a peripheral technical collaboration, not a named strategic partnership in the same tier as Skild AI's Foxconn or ABB partnership. SU008, SU015
CU018 No production deployments of GEN-1 have been confirmed by any independent source as of June 30, 2026. SU012, SU016
CU019 Skild AI, a direct competitor, publicly confirmed a commercial deployment of its robotics foundation model with Zebra Technologies in 2026, providing a public customer reference that Generalist AI lacks in the same timeframe. SU024, SU015
CU020 Competitor Cobot (Collaborative Robotics) announced Generation 2 Proxie in June 2026 with 12,627 operating hours in production environments, 40 million pounds moved, and 17 million steps saved—demonstrating the level of production evidence Generalist AI has not yet provided publicly. SU025, SU015
CU021 Physical Intelligence (the closest analog platform) operates a $300/robot/month SaaS model per Sacra research, implying that $10M ARR would require approximately 2,800 deployed robots — a scale Generalist AI has not yet demonstrated publicly. SU013, SU011
CU022 The MarkTechPost April 28, 2026 ranking of "Top 10 Physical AI Models Powering Real-World Robots in 2026" does not include Generalist AI's GEN-1, listing instead NVIDIA GR00T, Gemini Robotics, Physical Intelligence, Figure Helix, and others. SU016, SU012
CU023 Enterprise manufacturing and logistics customers typically require 6–18 months of qualification, safety testing under ISO 10218 / TS 15066, and ERP/WMS integration before production sign-off. SU011, SU015
CU024 Generalist AI has not disclosed any safety certifications, integration frameworks, formal SLAs, or published API documentation that would enable enterprise production qualification as of June 2026. SU012, SU020
CU025 GEN-1 demos showed 99% average success rates across six tasks: kitting auto parts (60+ minutes), folding T-shirts (86 reps), servicing robot vacuums (200+ reps), packing blocks (1,800+ reps), folding boxes (200+ reps), and packing phones (100+ reps) — all in Generalist-controlled settings. SU001, SU005
CU026 No NRR, GRR, churn, cohort, pilot-to-production rate, or customer satisfaction data is available in any public source for Generalist AI's GEN-1 early-access program as of June 30, 2026. SU012, SU002
CU027 No G2, Capterra, Gartner Peer Insights, or similar review platform lists GEN-1 as of June 2026; the product is not in self-service consumption and has no public review trail. SU012, SU001
CU028 Robotics.press analyst assessment (April 2026) explicitly identifies "Zero verified customer deployments, paid pilots, case studies, or named partners disclosed" as a bear case risk. SU012, SU027
CU029 CB Insights' January 2026 physical AI market map identifies Skild AI and FieldAI as having formal partnerships with industrial companies — while Generalist AI is not mentioned in deployment context. SU015, SU012
CU030 No named partner or customer has made a public announcement, press release, conference talk, or social media post confirming use of Generalist AI GEN-1 as of June 30, 2026. SU012, SU002
CU031 The GEN-1 launch blog describes early-access partners as being in "industries that will bring these systems into everyday use" and confirms real-world task data is flowing from businesses back into the model training cycle. SU001, SU002
CU032 Generalist AI's early-access program uses a high-touch direct model; no self-serve, partner portal, or marketplace distribution is available, consistent with a 3–10 strategic account GTM. SU020, SU007
CU033 Investor attestation from 8VC, Spark Capital, and Radical Ventures citing "early commercial traction" converges across three independent investment decisions made across different time periods (2025 seed and June 2026 Series B), suggesting some real commercial activity exists. SU004, SU011, SU014
CU034 Brad Porter (CEO of Cobot, former Amazon robotics executive), quoted in Forbes, argued that data-scale alone without architectural breakthroughs "is really expensive and not necessarily going to get you the result you want," directly challenging Generalist AI's core thesis. SU011, SU025
CU035 Early-access partners who share proprietary task data with Generalist AI are contributing to the company's model training moat; whether data contribution agreements include IP protection or exclusivity terms is not publicly disclosed. SU001, SU002
CU036 The data flywheel model means that early customers' proprietary manufacturing or logistics task data may enable Generalist AI to serve competing customers with similar capabilities, creating a potential IP and competitive concern for sophisticated enterprise buyers. SU001, SU012
CU037 Enterprise customers evaluating physical AI platforms must manage hardware certification, safety compliance (ISO 10218 / TS 15066), integration testing, and internal approval cycles before production sign-off, creating structural procurement friction independent of model quality. SU015, SU011
CU038 No customer complaints, adverse user reviews, failed deployment reports, or negative case studies for GEN-1 appear in any public source as of June 2026, consistent with the product being in controlled early-access rather than broad release. SU012, SU009
CU039 Based on enterprise robotics procurement norms and the single-headcount commercial team, estimated GEN-1 customer acquisition involves 6–18 months of sales cycle, high CAC relative to pre-revenue peers, and founder-led discovery — no confirmed sales cycle data is publicly disclosed by Generalist AI. SU011, SU028
CR001 The EU AI Act's obligations for high-risk AI systems become fully enforceable on August 2, 2026 under Regulation (EU) 2024/1689. SR004, SR005
CR002 Under EU AI Act Article 6, an AI system used as a safety component of a product covered by EU harmonisation legislation and requiring third-party conformity assessment is classified as high-risk. SR001, SR003
CR003 High-risk AI classification under the EU AI Act triggers obligations including: risk management system, data governance documentation, technical documentation, transparency, human-oversight measures, accuracy and robustness requirements, post-market monitoring, and a Fundamental Rights Impact Assessment (FRIA). SR003, SR004
CR004 Penalties for non-compliance with EU AI Act high-risk obligations can reach €35 million or 7% of global annual turnover, whichever is higher. SR003, SR001
CR005 Generalist AI has published no conformity assessment documentation, no DPA, no EU-facing compliance notice, and no high-risk AI registration for GEN-1 as of June 30, 2026. SR015, SR029
CR006 GDPR Article 9 classifies biometric data used to uniquely identify a person as special-category data, the processing of which is prohibited by default absent explicit consent or specific legal bases. SR004, SR005
CR007 Generalist AI's data collection program spans thousands of homes, warehouses, workplaces, and specialized environments globally; collection includes wrist kinematics and workspace video captured by data-hands operators, some of whom may be EU residents. SR017, SR015
CR008 RAND Corporation's analysis found that fragmentation of AI development chains makes causation attribution complex under US tort law, leaving model providers with unpredictable product liability exposure when their AI software controls a physical robot that injures someone. SR006, SR007
CR009 The EU AI Liability Directive is beginning to address liability gaps for AI systems, including shared liability models where software providers, hardware manufacturers, deployers, and operators all share responsibility for injuries caused by autonomous robots. SR008, SR007
CR010 Generalist AI has not publicly disclosed any product liability insurance coverage, indemnification framework, or contractual liability allocation for its early-access GEN-1 deployments. SR015, SR030
CR011 GEN-1's improvisational intelligence—producing behaviors outside the training distribution— is simultaneously the company's core commercial differentiator and, as noted in its own blog, a potential liability in real-world deployment environments. SR015, SR025
CR012 As of April 2026, Generalist AI has published no trust-and-safety framework, no alignment documentation, no red-team evaluation results, and no third-party safety audit for GEN-1. SR015, SR030
CR013 The International AI Safety Report 2026 (chaired by Yoshua Bengio, over 100 international experts) found that general-purpose AI capabilities remain jagged: systems fail at seemingly simple tasks while succeeding at harder ones, creating unpredictable real-world deployment reliability. SR009, SR010
CR014 The International AI Safety Report 2026 found that some AI models can distinguish between evaluation and deployment contexts and alter their behaviour accordingly, creating new challenges around evaluation and safety testing for physical AI systems. SR009, SR011
CR015 The SAE World Congress 2026 panel on Embodied AI reached broad agreement that embodied AI must be treated as a systems challenge requiring engineering rigor, lifecycle governance, and evolving safety standards—none of which Generalist AI has made publicly visible as of June 2026. SR024, SR011
CR016 NeuroForge's 2026 industry analysis found that modern embodied AI systems consistently struggle with long-range logical chains requiring sustained reliable execution without a single failure cascade, a limitation directly relevant to factory and warehouse deployments. SR023, SR019
CR017 Physical AI deployments in industrial settings during H1 2026 documented over 50 incidents of AI model failures including hallucination errors causing workflow breakdowns and context-dependent failures resulting in operational downtime. SR009, SR019
CR018 Generalist AI has published no ISO 10218 or ISO/TS 15066 safety certification documentation for GEN-1, and no conformity assessment roadmap for collaborative robot deployments has been publicly disclosed. SR015, SR024
CR019 NVIDIA NVentures co-invested in both Generalist AI's March 2025 and June 2026 fundraising rounds, creating strategic alignment between Generalist's training infrastructure and NVIDIA's hardware ecosystem. SR027, SR018
CR020 As of mid-2026, procurement lead times for H100 and successor GPU chips from non-hyperscaler buyers remain 36–52 weeks, driven by HBM3 memory supply constraints and TSMC CoWoS advanced packaging oversubscription. SR012, SR013
CR021 NVIDIA is shifting capacity from H100 production to higher-margin Blackwell lines, further reducing H100 availability for non-hyperscaler buyers; embodied AI startups without hyperscaler-tier forward contracts face structural compute access disadvantage. SR013, SR012
CR022 NVIDIA released GR00T N1.6, an open-source vision-language-action model for humanoid robots, available on Hugging Face; this free open model from Generalist AI's strategic investor competes with GEN-1 for adoption by robot hardware OEMs and independent developers. SR018, SR019
CR023 Generalist AI's distributed training infrastructure processes petabytes of physical interaction data and is capable of absorbing 6.85 years of manipulation experience per day of training, requiring H100- or Blackwell-class GPU clusters. SR015, SR025
CR024 Generalist AI's multi-cloud compute contracts are not publicly named; cloud provider concentration is unconfirmed and represents an unresolved operational dependency. SR029, SR020
CR025 Physical Intelligence's $1 billion raise at an $11 billion valuation in April 2026 and its expanding team represent the closest competitive benchmark for Generalist AI's compute and capital intensity requirements at the next model generation stage. SR028, SR021
CR026 Generalist AI has raised more than $500 million across two rounds ($140M March 2025; $400M June 2026) and carries a $2 billion post-money valuation but has disclosed no revenue, ARR, customer names, burn rate, or contract terms as of June 30, 2026. SR027, SR016
CR027 Industry benchmarks suggest monthly burn rates of $5–15 million per month for embodied AI startups with 50–100 employees operating global physical data infrastructure and frontier model training compute. SR023, SR028
CR028 Figure AI is reported to burn $200–300 million annually to sustain development cycles, serving as a high-end capital intensity benchmark for physical AI companies operating at hardware-plus- software model scale. SR023, SR028
CR029 Generalist AI's planned use of the June 2026 $400 million raise is explicitly to build next-generation models, scale the physical data engine, and expand the team—all pre-revenue expenditures that extend the pre-commercial phase. SR029, SR020
CR030 No hardware partner name, integration agreement, channel arrangement, or named enterprise deployment has been publicly disclosed by Generalist AI as of June 30, 2026; the only confirmed commercial contact is partnerships@generalistai.com. SR015, SR030
CR031 Generalist AI has not publicly disclosed any convertible notes, debt facilities, revenue- based financing, or credit lines; there is no downside buffer beyond equity capital if revenues do not materialise on schedule. SR016, SR029
CR032 Generalist AI's investor base—Spark Capital, Radical Ventures, 8VC, Union Square Ventures, NVIDIA NVentures—is concentrated in a small group; a shift in sentiment among this group would substantially reduce the probability of a Series C on favorable terms. SR027, SR014
CR033 If Generalist AI's early-access partner program relies on a small number of data-generating partners, the withdrawal of a single major partner would simultaneously damage the data flywheel and the investor narrative around commercial traction. SR029, SR014
CR034 Pete Florence is the CEO and co-founder of Generalist AI; investor commentary (8VC) describes him as the primary source of conviction in the investment, framing the bet as substantially a bet on Florence's ability to blend frontier research with commercial execution. SR014, SR027
CR035 Andy Zeng is the Chief Scientist and co-founder; his research lineage (Google DeepMind, lead author of Code as Policies) is cited in virtually every press article and investor memo as a central pillar of Generalist's technical credibility. SR014, SR027
CR036 Andrew Barry is the CTO and co-founder; his Boston Dynamics robotics background provides hardware integration expertise that is not redundantly covered elsewhere in Generalist's publicly disclosed team. SR014, SR027
CR037 No board composition, independent director names, audit committee, governance charter, or board seat appointments from the June 2026 Series B have been publicly disclosed by Generalist AI. SR027, SR016
CR038 Generalist AI's broader team draws from Google DeepMind, OpenAI, and Boston Dynamics—the same organisations that compete most actively for embodied AI researchers in 2026 and offer competitive compensation, prestigious publishing environments, and larger research teams. SR014, SR018
CR039 At 70 employees, Generalist AI has no deep bench below the founders; a single departure in the model-architecture team or data engine team could delay GEN-2 by quarters without easy replacement. SR029, SR014
CR040 Generalist AI had a single publicly identified commercial-facing employee as of June 2026 (Applied AI & Partnerships), meaning the entire go-to-market motion depends on founder relationships and the early-access partner contact email. SR015, SR029
CR041 No equity schedule, vesting cliff, or co-founder retention terms have been publicly disclosed for Generalist AI's founding team, leaving the probability of co-founder lock-in unverifiable from public sources. SR016, SR014
CR042 Physical AI startups globally have raised over $3.4 billion by early 2026, reflecting the capital intensity of the sector; industry analysis indicates average R&D burns exceeding $3 billion for full-scale humanoid development. SR023, SR022
CR043 The NIST AI Risk Management Framework (AI RMF 1.0) requires physical AI systems to maintain an AI Bill of Materials (AIBOM) including source of base models, datasets, fine-tuning methods, external dependencies, and versioning; Generalist AI has not published AIBOM or equivalent documentation. SR002, SR024
CR044 Brad Porter, Cobot CEO and former Amazon VP Robotics, cited the gap between lab-demonstrated generalisation and production-tested reliability as the central challenge for embodied AI deployments in 2026, representing a third-party adverse view of the commercialisation risk facing GEN-1-class systems. SR026, SR023
CR045 Morgan Stanley's investment research identifies that broad humanoid and embodied AI adoption is still years away; early deployments catalyse learning flywheels but the economic forces driving adoption must be weighed against the technical challenges and capital intensity required for broad deployment—a risk profile that applies directly to Generalist AI's pre-revenue stage. SR022, SR028
CV001 Generalist AI raised $400 million in a Series B round at a $2 billion post-money valuation in June 2026, led by Radical Ventures, with co-investors including 8VC, Union Square Ventures, Hanabi Capital, Norwest, NVIDIA NVentures, and Bezos Expeditions. SV001, SV003, SV014
CV002 Generalist AI's total funding exceeded $500 million since its founding in 2024, combining the Series A (approximately March 2025) and the June 2026 Series B. SV001, SV021, SV026
CV003 SEC EDGAR records a Form D filing by Generalist AI, Inc. (CIK 0002137708) on 2026-06-05, confirming the company is incorporated in Delaware with a principal office in San Francisco, CA, and lists Pete Florence as Executive Officer, Director, and Promoter; Fraser Kelton and Ellen Chisa are also listed as persons. SV014
CV004 The SEC EDGAR Form D filed 2026-06-05 confirms the June 2026 financing event for Generalist AI, Inc. under the correct CIK. The filing classifies the round under item 06B (equity securities not publicly registered) and lists the business state as CA. SV014
CV005 Generalist AI has not publicly disclosed any revenue, ARR, customer count, pricing model, gross margin, unit economics, or any other financial metrics as of June 30, 2026. SV017, SV018, SV019
CV006 Generalist AI's GEN-1 early-access program launched April 2, 2026, with selected industry partners; no partner names or contract terms have been confirmed in any public source as of June 30, 2026. SV017, SV019
CV007 The Generalist AI June 2026 funding announcement blog states: "A data flywheel has begun to take shape: real businesses are generating task data that feeds the next generation of more capable models" — the company's own strongest commercial signal. SV018
CV008 GEN-1 achieves 99% average success rates on simple tasks versus 64% for GEN-0, completes tasks ~3x faster than the prior state of the art, and requires only one hour of robot-specific data to adapt to a new robot, per Generalist AI's official claims. SV017, SV019, SV005
CV009 8VC's June 2026 investment memo described "remarkable early commercial traction" and "strong sample efficiency" at Generalist AI, and characterized Pete Florence as "as much a builder as a researcher — magnetic and unusually commercial." SV004
CV010 Physical Intelligence closed a $600 million Series B in November 2025 at a $5.6 billion post-money valuation, led by CapitalG. SV006, SV009
CV011 Physical Intelligence was reported in talks to raise approximately $1 billion at an over-$11 billion valuation in March 2026, with Founders Fund and Lightspeed reportedly in discussions — an approximate doubling of its Series B valuation in four months. SV009, SV010
CV012 Skild AI raised a $1.4 billion Series C round at over $14 billion valuation in January 2026, led by SoftBank with participation from NVIDIA Ventures, Macquarie Group, and others. SV008, SV015
CV013 Skild AI reported approximately $30 million in revenue within months of its commercial launch in 2025, per Sacra analyst research — the only software-pure robotics AI peer with disclosed commercial traction at this scale as of mid-2026. SV006, SV008
CV014 Figure AI reached a $39 billion post-money valuation in its Series C in September 2025, supported by documented BMW Spartanburg commercial deployments of 90,000+ parts loaded and 1,250+ runtime hours. SV007, SV015
CV015 CBInsights reported that the robotics sector raised a record $40.7 billion in 2025, up 74% YoY and representing 9% of all venture funding globally — positioning Generalist AI within a structurally well-funded sector. SV015
CV016 At $2B, Generalist AI's valuation is approximately 0.14x Skild AI's $14B and approximately 0.18x Physical Intelligence's $11B reported mark — the lowest confirmed valuation in the direct software-pure robotics AI peer group as of June 2026. SV008, SV009, SV001
CV017 Finerva reported that the median EV/Revenue multiple for public Robotics & AI companies recovered from a 2.5x low in Q1 2025 to 3.4x in Q4 2025, with the max reaching 24x. SV011
CV018 Finerva reported that median EV/EBITDA for public Robotics & AI companies was 16.8x in Q4 2025, with a range from approximately 0.9x to 78.2x across the cohort. SV011
CV019 Finro's Q1 2026 dataset shows LLM Vendors at a median EV/Revenue of 39.5x (average 73.5x) across 27 companies; AI robotics private companies likely fall in the 20–60x range for projected forward revenue premiums. SV012
CV020 At a 20x forward revenue multiple (appropriate for high-growth private AI software with early traction), Generalist AI's $2B valuation implies a target ARR of $100 million. SV011, SV012
CV021 At a 40x forward revenue multiple (appropriate for a frontier AI scarcity premium), Generalist AI's $2B valuation implies a target ARR of only $50 million. SV011, SV012
CV022 Physical Intelligence's SaaS pricing analog of $300 per robot per month ($3,600/year) per Sacra means that reaching $50M ARR at that price point requires approximately 13,889 deployed robots — a fleet scale that requires substantial enterprise adoption. SV006
CV023 Generalist AI at $2B vs. Skild AI at $14B and Physical Intelligence at ~$11B represents a relative discount of 7x to 5.5x respectively in the software-pure embodied AI peer group — suggesting the entry is low relative to category, not just absolute. SV008, SV009, SV001
CV024 AIRoboticDaily reported that approximately 95% of humanoid robot revenue in 2026 comes from research and showroom use, with truly productive industrial deployment accounting for only approximately 3–5% of total sales. SV013
CV025 AIRoboticDaily documented a 4x divergence in robotics market size forecasts for 2030 (lowest ~$4B, highest ~$15B), which the article frames as a signal of speculative heat akin to prior tech hype cycles in autonomous driving and IoT. SV013
CV026 AIRoboticDaily warned that planned production capacity from multiple humanoid robotics manufacturers collectively exceeds the most optimistic 2030 total demand forecasts for individual national markets, signaling structural overcapacity risk. SV013
CV027 Brad Porter, founder and CEO of Cobot (a competing physical AI company), argues that production-tested physical AI with 12,627 operating hours logged in real facilities (hospitals, manufacturing, logistics) has superior commercial credibility to research-driven approaches that have not completed equivalent production cycles. SV029
CV028 Analysts covering embodied AI document that down-round risk is material for pre-revenue companies if commercialization lags behind capital deployment or if public market sentiment toward AI robotics corrects. SV013, SV011
CV029 At an industry-estimated $10 million per month burn rate (consistent with a 70-person frontier AI research team with intensive compute operations), Generalist AI's $400M fresh capital implies approximately 40 months of runway from June 2026, extending through approximately late 2029. SV001, SV018
CV030 The bull case for Generalist AI requires the data-engine flywheel to scale from 500,000 hours (GEN-1) to 5 million+ hours by 2028, making the cost to replicate the dataset prohibitive and driving GEN-2/GEN-3 to superior commercial performance. SV030, SV017
CV031 The base case requires verifiable ARR of $20–50M by 2028, followed by a strategic acquisition by a large technology incumbent at 10–15x ARR in 2029–2031, implying an exit of approximately $300M–$750M — flat to modestly negative on a $2B entry after dilution from at least 1–2 additional fundraising rounds. SV011, SV012, SV008
CV032 The bear case is capital depletion without commercial traction: if the $400M runway does not produce verifiable revenue by 2028, a down-round, acqui-hire at team value, or strategic wind-down is the likely outcome, representing near-total capital loss on a $2B entry. SV013, SV029, SV011
CV033 NVIDIA NVentures and Bezos Expeditions are strategic reinvestors in the June 2026 round, suggesting strategic M&A — particularly an NVIDIA or Amazon acquisition — is a more credible exit path than a standalone IPO. SV001, SV020
CV034 NVIDIA's open-source GR00T N-series foundation model for robotics provides free model weights to developers; if expanded to enterprise use at Generalist AI's performance tier, it represents a direct commoditization threat to Generalist AI's proprietary model value proposition. SV004, SV003
CV035 A standalone Generalist AI IPO is not credible within three years from the June 2026 runDate given the company's pre-revenue status, lack of financial disclosure, and the sector norm that software-only robotics AI companies require 5+ years to reach IPO-ready scale. SV027, SV007
CV036 The primary thesis-break trigger is failure to produce a confirmed revenue-generating commercial deployment by Q4 2027, approximately 18 months after GEN-1's early-access launch; without that milestone, the data-flywheel thesis remains unconfirmed. SV019, SV018
CV037 If Physical Intelligence or Skild AI achieve $100M+ ARR with broadly comparable model performance to GEN-1, Generalist AI's relative data advantage and architectural differentiation are no longer sufficient as primary moats. SV006, SV008, SV013
CV038 If NVIDIA expands GR00T N-series to enterprise users under permissive licensing with performance competitive to GEN-1, the foundation model layer commoditizes and Generalist AI's proprietary model premium compresses substantially. SV004, SV015
CV039 Strategic acquirers with the strongest documented rationale for acquiring Generalist AI are NVIDIA (NVentures strategic reinvestor, GR00T platform complementarity), Amazon Robotics (cross-embodiment thesis alignment), and Google DeepMind (shared research lineage with Florence and Zeng). SV001, SV003, SV004
CV040 SEC EDGAR Form D names Fraser Kelton (Spark Capital) and Ellen Chisa among the persons listed for Generalist AI's June 2026 round; Radical Ventures as lead implies a board seat by conventional VC practice, though this has not been formally confirmed publicly. SV014
CV041 Angel investors in the June 2026 round include Fei-Fei Li (Stanford AI pioneer and former Google Cloud AI chief), Eric Yuan (Zoom founder), Bin Lin (Xiaomi co-founder), and Naval Ravikant — a concentration of frontier AI community validation signals unusual for a pre-revenue company. SV001, SV024
CV042 Pete Florence (CEO) was a senior scientist on DeepMind's robotics team and a senior author on PaLM-E and RT-2, two of the most-cited foundational papers in embodied AI. Andy Zeng (Chief Scientist) was lead author of "Code as Policies." Andrew Barry (CTO) built Atlas, Spot, and Stretch at Boston Dynamics. SV004, SV003
CV043 Generalist AI's data engine collected over 500,000 hours of real-world physical interaction data by the GEN-1 launch (April 2026), up from 270,000 hours at the GEN-0 launch (November 2025), at a rate of 10,000+ hours per week. SV017, SV030
CV044 Generalist AI's software-only business model (no hardware manufacturing) implies potential gross margins near 70–80% at scale, structurally similar to a SaaS company; the actual gross margin has not been disclosed, and Physical Intelligence's $300/robot/month SaaS pricing is the closest public reference point for sector economics. SV006, SV031
CV045 The embodied AI TAM was valued at $6.5 billion in 2026 and is projected to reach $67.6 billion by 2033 at a 39.7% CAGR, with logistics and supply chain as the fastest- growing segment at 42.2% CAGR; North America represented 35.6% of the 2025 market. SV028, SV015
CV046 Goldman Sachs projects the humanoid robotics market will reach $38 billion by 2035; robotics startups raised $13.8 billion globally in 2025 (up from $7.8 billion in 2024), with 2026 already on pace to exceed that. The top 10 humanoid robotics companies have captured nearly 80% of all capital raised in the category since 2022, signaling rapid market concentration that will make it harder for undifferentiated entrants to raise at any price. SV033
CV047 Anthropic's May 2026 Series H at $965 billion post-money valuation — with a run rate revenue of $47 billion — and OpenAI's $852 billion valuation (March 2026) mark the current apex of the private AI company spectrum, both actively moving toward IPO. Generalist AI's $2 billion pre-revenue entry sits at a fundamentally different point on that spectrum, where commercial proof — not capital scale — is the gating factor to any credible IPO trajectory. SV034
CV048 Independent analysts note that U.S. venture investors systematically reward long-term potential and global reach, often funding companies without demonstrated commercial revenue, whereas investors in other geographies anchor valuations to verifiable deployments. This behavior explains the premium commanded by U.S.-based pre-revenue AI robotics companies — including Generalist AI's $2 billion mark — and represents a structural risk if U.S. investor sentiment toward physical AI shifts to requiring revenue proof before writing Series C checks. SV035
来源
编号出版方标题引文
SO001 Generalist AI Generalist AI — Homepage Generalist is a frontier AI research and product driven company building general intelligence for the physical world.
SO002 Generalist AI About — Generalist AI We are building general intelligence for the physical world. At Generalist, we are on a mission to make general-purpose robots a reality.
SO003 Generalist AI GEN-0: Embodied Foundation Models That Scale with Physical Interaction GEN-0 marks the beginning of a new era: embodied foundation models whose capabilities predictably scale with physical interaction data.
SO004 Generalist AI GEN-1: Scaling Embodied Foundation Models to Mastery We believe GEN-1 to be the first general-purpose AI model that crosses a new performance threshold: mastery of simple physical tasks.
SO005 Generalist AI Accelerating the next phase of physical AI Today, we're announcing $400 million in new funding, bringing our total raised to more than half a billion dollars.
SO006 Generalist AI Going Beyond World Models and VLAs In GEN-1, approximately 99% of the parameters are trained from scratch.
SO007 Generalist AI Careers — Generalist AI
SO008 Generalist AI Blog Index — Generalist AI
SO009 The Robot Report Generalist raises $400M to scale its general-purpose AI models Generalist AI Inc., a company creating AI for a range of robot form factors, today said it has raised $400 million in new funding.
SO010 Robotics and Automation News Generalist AI unveils GEN-1 model, claiming breakthrough in real-world robotic task performance GEN-1 achieves '99 percent success rates' on certain tasks, compared with around 64 percent for its previous-generation system.
SO011 Robotics and Automation News Generalist AI raises $400 million to scale robot intelligence platform
SO012 Forbes Generalist Is Betting Its Robot-Training Gloves Will Usher In Robotics' ChatGPT Moment Just brute forcing a huge amount of data against a not-perfect architecture is really expensive and not necessarily going to get you the result you want.
SO013 Humanoids Daily Physical AI Arms Race Accelerates: Generalist AI Secures $400M to Scale Robot Learning
SO014 Humanoids Daily Generalist AI Unveils GEN-0, Claims Scaling Laws for Robotics Backed by 270,000 Hours of Real-World Data
SO015 Humanoids Daily Generalist AI Unveils GEN-1: The Quest for Robot Mastery and Intelligent Improvisation
SO016 SiliconANGLE Generalist AI raises $400M at $2B valuation to build general intelligence for robotics Artificial intelligence startup Generalist AI Inc., a startup building embodied robotics intelligence, said today it has raised $400 million in new funding, bringing the company's valuation to $2 billion.
SO017 Silicon Republic Nvidia, Fei-Fei Li back Generalist's $400m round to scale AI robotics
SO018 Ashby (Generalist AI job board) Generalist Jobs — Open Positions
SO019 8VC Announcing our Investment in Generalist Pete Florence, CEO, was previously a senior scientist on DeepMind's robotics team and a senior author on PaLM-E and RT-2, two of the most influential papers in embodied AI.
SO020 CryptoBriefing Generalist AI raises $400M in funding round led by Radical Ventures, hitting $2B valuation
SO021 RoboHorizon Generalist's GEN-1 Brain Hits 99% Success, 3x Speed
SO022 EmbodiedGlobal Generalist AI Raises $400M to Build Foundation Models for Any Robot
SO023 Fundraise Insider Generalist AI Raises $400M Series B at $2B Valuation
SO024 FinSMEs Generalist AI Raises $400M in Funding at $2 Billion Valuation
SO025 Generalist AI Contact — Generalist AI
SM001 International Federation of Robotics (IFR) Top 5 Global Robotics Trends 2026
SM002 International Federation of Robotics (IFR) IFR World Robotics Data — Installation Statistics 2026
SM003 Mordor Intelligence Industrial Robotics Market Size, Analysis, Share & Growth Trends 2031
SM004 Grand View Research Warehouse Automation Market Size And Share Report, 2030
SM005 Fortune Business Insights Warehouse Robotics Market Size, Share Report 2026–2034
SM006 Grand View Research Artificial Intelligence In Robotics Market Size Report, 2033
SM007 Intel Market Research (IMR) AI Robotics Software Market Outlook 2026–2034
SM008 Research and Markets Humanoid Robot Market Report 2026
SM009 Technavio Humanoid Robot Market Growth Analysis — Size and Forecast 2026–2030
SM010 RoboticsTomorrow Top 5 Global Robotics Trends 2026 — International Federation of Robotics Reports
SM011 StartUs Insights Global Robotics Report 2026
SM012 MarketsandMarkets Artificial Intelligence in Manufacturing Market Size, Share & Trends 2025–2030
SM013 Mordor Intelligence Laboratory Robotics Market Size, Growth, Share Analysis 2026–2031
SM014 NeuroForge GTM Embodied AI Commercialization: 2026 Challenges & Trends
SM015 arXiv Embodied AI in Action: Insights from SAE World Congress 2026 on Safety, Trust, Robotics, and Real-World Deployment
SM016 Grand View Research Embodied AI Market Size & Share | Industry Report, 2033
SM017 Morgan Stanley Investment Management EDGE: Embodied AI and the Rise of Humanoid Robots
SM018 Sacra Physical Intelligence — Company Overview and Funding Analysis
SM019 Physical Intelligence Physical Intelligence — Official Website
SM020 The Mimic Physical Intelligence: $1B Funding and the Humanoid Robot Race in 2026
SM021 Sacra Figure AI — Company Overview and Market Analysis
SM022 QubitTool Embodied AI 2026: From Robot Foundation Models to Industrial Deployment
SM023 Failory Top Robotics Unicorns and Startups 2026
SM024 NVIDIA Investor Relations NVIDIA Releases New Physical AI Models as Global Partners Unveil Next-Generation Robots
SM025 Research and Markets Embodied AI Market Size, Share & Trends Analysis Report 2033
SP001 Generalist AI GEN-1: Scaling Embodied Foundation Models to Mastery
SP002 Generalist AI Going Beyond World Models and VLAs
SP003 The Robot Report Generalist introduces GEN-1 general-purpose model for physical AI
SP004 Robotics and Automation News Generalist AI unveils GEN-1 model, claiming breakthrough in real-world robotic task performance
SP005 Figure AI Helix: A Vision-Language-Action Model for Generalist Humanoid Control
SP006 Sacra Figure AI valuation, funding and news
SP007 Google DeepMind Gemini Robotics 1.5 brings AI agents into the physical world
SP008 Google AI for Developers Gemini Robotics-ER 1.6 — API Overview
SP009 NVIDIA Investor Relations NVIDIA Releases New Physical AI Models as Global Partners Unveil Next-Generation Robots
SP010 NVIDIA News NVIDIA and Global Robotics Leaders Take Physical AI to the Real World
SP011 Boston Dynamics Boston Dynamics Unveils New Atlas Robot to Revolutionize Industry
SP012 Skild AI Skild AI Acquires Zebra Technologies' Robotics Arm to Bring Omni-Bodied Intelligence to Warehouses
SP013 Business Wire Skild AI Acquires Zebra Technologies' Robotics Automation Business
SP014 Robotics and Automation News Skild AI acquires Zebra Technologies' robotics automation business
SP015 AI2Work Skild AI's $1.4B Raise: Why Robotics Foundation Models Are 2026's Mega-Bet
SP016 CB Insights The physical AI models market map: Behind the arms race to control robot intelligence
SP017 EVST International Top Robotics Foundation Model and Embodied AI Companies 2026
SP018 QubitTool Embodied AI 2026: From Robot Foundation Models to Industrial Deployment
SP019 MarkTechPost Top 10 Physical AI Models Powering Real-World Robots in 2026
SP020 TechFunding News Physical Intelligence eyes $1B raise at $11B valuation, Founders Fund and Lightspeed in talks
SP021 Humanoids Daily Physical AI Arms Race Accelerates: Generalist AI Secures $400M to Scale Robot Learning
SP022 Humanoids Daily Generalist AI Unveils GEN-0, Claims Scaling Laws for Robotics Backed by 270,000 Hours of Real-World Data
SP023 8VC Announcing our Investment in Generalist
SP024 EVST International Embodied AI This Week — Five Storylines: NVIDIA Cosmos 3, Unitree IPO, BMW x Hexagon AEON (Jun 1–7, 2026)
SP025 Physical Intelligence Physical Intelligence (π) — Official Company Site
SI001 Sacra Physical Intelligence valuation, funding & news Physical Intelligence runs a B2B software-as-a-service model for robotics companies, manufacturers, and automation integrators. Pricing is a $300 monthly subscription per connected robot, yielding recurring revenue that scales with fleet deployments.
SI002 ahr.so Physical Intelligence: How π0.7 Created a GPT-3 Moment Skild AI Commercial deployments; Skild Brain ~$30M (2025) Real-world data flywheel; industrial validation.
SI003 The Mimic Physical Intelligence Is Raising Another $1B — Here's Why Investors Keep Betting on Humanoid Robots Training generalist robot policies requires massive amounts of diverse physical interaction data. More capital means more robots, more environments, more variation — all of which translates to better models.
SI004 RoboticsTomorrow Cobot Announces Second-Generation Proxie, Bringing Production-Tested Physical AI to Real Operations Brad Porter, founder and CEO of Cobot: "For decades, deploying robots has meant choosing between mobility and dexterity, and always required custom software integration." Cobot's approach uses on-robot edge AI rather than cloud-dependent large-scale data pretraining.
SI005 Outset Capital Physical Intelligence Raises $600M at $5.6B Valuation Physical Intelligence has raised $600 million at a $5.6 billion valuation. The round was led by CapitalG, with participation from Lux Capital, Thrive Capital, Jeff Bezos, Index Ventures, and T. Rowe Price.
SI006 The Robot Report Generalist introduces GEN-1 general-purpose model for physical AI "Building GEN-1 was not easy — we redesigned our distributed training infrastructure to support petabytes of physical interaction data as a first-class citizen," said Generalist AI. The company said that early-access partners can now gain access to the model.
SI007 udit.co Physical Intelligence Raises $1B at $11B Valuation
SI008 HumansAreObsolete Physical Intelligence Raises $600M at $5.6B Valuation: General-Purpose Robot Startup Becomes Unicorn Physical Intelligence has secured pilot programs with major logistics, manufacturing, and service companies, demonstrating 70% faster deployment times compared to traditional automation solutions.
SI009 Generalist AI Accelerating the Next Phase of Physical AI — Funding Announcement We are beginning to see a flywheel take shape: scaling robot learning creates better models, better models can do more useful physical work, and data from real businesses drives the next generation of more capable models.
SI010 Generalist AI GEN-1 Model Launch Blog Post We believe GEN-1 to be the first general physical AI model to cross a key threshold: unlocking commercial viability across a broad range of tasks.
SI011 Generalist AI Generalist AI Careers Page
SI012 Ashby (Generalist AI) Generalist AI Jobs — Applied & Partnerships Open Role Applied AI & Partnerships: Applied & Partnerships — San Francisco Bay Area (San Mateo) or Boston (Somerville) — Full time — On-site. Only one outward-facing commercial role listed.
SI013 SiliconANGLE Generalist AI raises $400M at $2B valuation to build general intelligence for robotics
SI014 The Robot Report Generalist raises $400M to scale its general-purpose AI models
SI015 8VC Announcing Our Investment in Generalist We followed Generalist's progress closely across successive generations of its models, and the signal was unmistakable: rapid adaptation to new robots and tasks, strong sample efficiency, and remarkable early commercial traction.
SI016 Forbes Generalist Is Betting Its Robot Training Gloves Will Usher In Robotics' ChatGPT Moment Brad Porter, CEO of Cobot, argues: "just brute forcing a huge amount of data against a not-perfect architecture is really expensive and not necessarily going to get you the result you want."
SI017 Humanoids Daily Generalist AI Unveils GEN-1: The Quest for Robot Mastery and Intelligent Improvisation Critics like Brad Porter, CEO of Cobot, argue that brute-forcing data against imperfect architectures is "really expensive and not necessarily going to get you the result you want." This echoes skepticism from Yann LeCun, who maintains that world models must learn through observation rather than just action-token prediction.
SI018 Humanoids Daily Physical AI Arms Race Accelerates: Generalist AI Secures $400M to Scale Robot Learning
SI019 Robotics and Automation News Generalist AI raises $400 million to scale robot intelligence platform
SI020 Fundraise Insider Generalist AI Raises $400M Series B at $2B Valuation
SI021 Generalist AI GEN-0 Blog Post — Scaling Laws in Robotics
SI022 Generalist AI Beyond World Models — Technical Blog
SI023 Generalist AI The Dark Matter of Robotics: Physical Commonsense Physical commonsense emerges from the sensorimotor loop. And in the process of interacting with the world, action produces information. Generalist AI's core thesis that physical intelligence requires real physical interaction data — not internet text — underpins the data engine strategy.
SI024 Robohorizon Generalist's GEN-1 Brain Hits 99% Success and 3x Speed
SI025 Silicon Republic Nvidia and Fei-Fei Li back Generalist's $400M round to scale AI robotics
SI026 Embodied Global Generalist AI $400M Funding 2026
SI027 CryptoBriefing Generalist AI Raises $400M Led by Radical Ventures
SI028 U.S. Securities and Exchange Commission Form D — Notice of Exempt Offering of Securities, Generalist AI, Inc. (CIK 0002137708) Generalist AI, Inc. (CIK 0002137708) filed Form D (Notice of Exempt Offering of Securities, Item 06b, Reg D Rule 506(b)) on June 5, 2026. The filing lists the business address as San Francisco, CA and state of incorporation as Delaware; accession number 0002137708-26-000003.
SI029 U.S. Securities and Exchange Commission Form D Primary Document — Generalist AI, Inc. (CIK 0002137708), Accession 0002137708-26-000003 The primary Form D XML for Generalist AI's June 2026 offering discloses a total offering amount of $399,998,660, with $363,810,091 already sold to 39 investors as of the filing date (June 5, 2026). The filing also identifies the company's former legal name as "Artificial General Dexterity, Inc." and confirms the Rule 506(b) exempt offering signed by CEO Peter Florence.
SE001 Generalist AI Generalist AI — Homepage Generalist is a frontier AI research and product driven company building general intelligence for the physical world.
SE002 Generalist AI About — Generalist AI
SE003 Generalist AI GEN-0: Embodied Foundation Models That Scale with Physical Interaction GEN-0 marks the beginning of a new era: embodied foundation models whose capabilities predictably scale with physical interaction data.
SE004 Generalist AI GEN-1: Scaling Embodied Foundation Models to Mastery We believe GEN-1 to be the first general-purpose AI model that crosses a new performance threshold: mastery of simple physical tasks.
SE005 Generalist AI Accelerating the next phase of physical AI Today, we're announcing $400 million in new funding, bringing our total raised to more than half a billion dollars.
SE006 Generalist AI Going Beyond World Models & VLAs In GEN-1, approximately 99% of the parameters are trained from scratch.
SE007 Generalist AI Careers — Generalist AI
SE008 Generalist AI Blog Index — Generalist AI
SE009 Generalist AI The Dark Matter of Robotics: Physical Commonsense Physical commonsense is the reactive, closed-loop intelligence behind acting in the real world: an intuition for forces, friction, compliance, and uncertainty, learned through a lifetime of sensorimotor experience.
SE010 Generalist AI The Real Breakthrough Behind Our GTC Demo We took a robot platform that didn't exist and had a live, public demo of GEN-0 running on the system within a handful of days.
SE011 Generalist AI The Robots Build Now, Too As far as we know this is the world's first robot to assemble Legos with end-to-end visuomotor control.
SE012 The Robot Report Generalist raises $400M to scale its general-purpose AI models
SE013 Robotics and Automation News Generalist AI unveils GEN-1 model, claiming breakthrough in real-world robotic task performance GEN-1 achieves '99 percent success rates' on certain tasks, compared with around 64 percent for its previous-generation system.
SE014 Robotics and Automation News Generalist AI raises $400 million to scale robot intelligence platform
SE015 Forbes Generalist Is Betting Its Robot-Training Gloves Will Usher In Robotics' ChatGPT Moment Just brute forcing a huge amount of data against a not-perfect architecture is really expensive and not necessarily going to get you the result you want.
SE016 Humanoids Daily Generalist AI Unveils GEN-0, Claims Scaling Laws for Robotics Backed by 270,000 Hours of Real-World Data
SE017 Humanoids Daily Generalist AI Unveils GEN-1: The Quest for Robot Mastery and Intelligent Improvisation
SE018 Humanoids Daily Physical AI Arms Race Accelerates: Generalist AI Secures $400M to Scale Robot Learning
SE019 SiliconANGLE Generalist AI raises $400M at $2B valuation to build general intelligence for robotics
SE020 Silicon Republic Nvidia, Fei-Fei Li back Generalist's $400m round to scale AI robotics
SE021 The Daily Upside $2 Billion Startup Generalist AI Wants to Solve the Robot Training Conundrum The company is doing so by way of what it calls 'grippers,' a sort of dummy-prototype version of its flagship robotic hands that real-life humans can wear like gloves and puppeteer.
SE022 8VC Announcing our Investment in Generalist The crux of the robot intelligence problem lies in dexterity. Progress at the cognitive frontier was limited by the scarcity of intelligent experts; dexterity poses the opposite problem.
SE023 Ashby (Generalist AI job board) Generalist Jobs — Open Positions
SE024 AI Daily Post Generalist launches physical robotics AI with production-level success rates Generalist now claims it has collected over half a million hours and 'petabytes of physical interaction data' to help train its physical model.
SE025 Humphrey Theodore Generalist AI Raised $400M to Put an AI Foundation Model Inside Robots GEN-1 runs roughly three times faster than comparable state-of-the-art models, holds about 99% reliability across a diverse spread of physical tasks, and — per the reporting — outperforms Physical Intelligence's pi-0.
SE026 Bloomberg Nvidia-Backed Robotics Startup Generalist AI Valued at $2 Billion
SE027 Quartz Generalist AI raises $400M at $2B valuation, backed by Nvidia
SE028 TechCrunch A key DeepMind robotics researcher left Google, and Nvidia has already backed his stealth startup
SU001 Generalist AI GEN-1: Scaling Embodied Foundation Models to Mastery Early access to GEN-1 is now available to selected partners.
SU002 Generalist AI Accelerating the Next Phase of Physical AI "Only two months after GEN-1, we are beginning to see a flywheel take shape: scaling robot learning creates better models, better models can do more useful physical work, and data from real businesses drives the next generation of more capable models."
SU003 Generalist AI The Real Breakthrough Behind Our GTC Demo "We were fortunate to be asked by Universal Robots, the world's #1 cobot manufacturer by volume, to join them in their booth to live demo our GEN-0 model on their new mobile manipulation platform."
SU004 8VC Announcing Our Investment in Generalist "We followed Generalist's progress closely across successive generations of its models, and the signal was unmistakable: rapid adaptation to new robots and tasks, strong sample efficiency, and remarkable early commercial traction."
SU005 The Robot Report Generalist introduces GEN-1 general-purpose model for physical AI
SU006 The Robot Report Generalist raises $400M to scale its general-purpose AI models
SU007 SiliconANGLE Generalist AI raises $400M at $2B valuation to build general intelligence for robotics
SU008 NVIDIA Corporation NVIDIA and Global Robotics Leaders Take Physical AI to the Real World "Generalist AI is using Cosmos to explore generating synthetic data."
SU009 Robotics and Automation News Generalist AI unveils GEN-1 model claiming breakthrough in real-world robotic task performance
SU010 Robotics and Automation News Generalist AI raises $400 million to scale robot intelligence platform
SU011 Forbes Generalist Is Betting Its Robot-Training Gloves Will Usher In Robotics' ChatGPT Moment "Just brute forcing a huge amount of data against a not-perfect architecture is really expensive and not necessarily going to get you the result you want." — Brad Porter, CEO of Cobot
SU012 robotics.press Generalist AI Inc. "Zero verified customer deployments, paid pilots, case studies, or named partners disclosed — commercial readiness is entirely unproven."
SU013 Sacra Physical Intelligence — Company Research Report Physical Intelligence runs a B2B SaaS model priced at $300 per connected robot per month.
SU014 Silicon Republic Nvidia and Fei-Fei Li back Generalist's $400M round to scale AI robotics
SU015 CB Insights The physical AI models market map: Behind the arms race to control robot intelligence
SU016 MarkTechPost Top 10 Physical AI Models Powering Real-World Robots in 2026
SU017 The Daily Upside Generalist AI Is Betting an Army of Humans Will Create Its Robot Army
SU018 Robo Horizon Generalist's GEN-1 Brain Hits 99% Success, 3x Speed
SU019 Humphrey Theodore Generalist AI Raised $400M to Put an AI Foundation Model Inside Robots
SU020 Generalist AI Generalist AI Careers
SU021 Generalist AI About — Generalist AI
SU022 Cryptobriefing Generalist AI Raises $400M at $2B Valuation — Radical Ventures Leads
SU023 Embodied Global Generalist AI Raises $400M: Building One Brain for All Robots
SU024 AI2.work Skild AI's $1.4B Raise: Why Robotics Foundation Models Are 2026's Mega-Bet
SU025 Robotics Tomorrow Cobot Announces Second-Generation Proxie, Bringing Production-Tested Physical AI to Real Operations
SU026 Generalist AI GEN-0: Embodied Foundation Models That Scale with Data
SU027 AI Robotic Daily Embodied AI and Robotics Financing Boom — A Health Check on Commercial Readiness "Truly productive revenue from industrial scenarios accounts for merely three to five percent of total sales. The remaining ninety five percent comes from universities, corporate showrooms, and consumers buying expensive gadgets."
SU028 Acgram Robotics and Liability — The Legal Frameworks You Cannot Ignore
SU029 Finerva Robotics and AI — 2026 Valuation Multiples Report
SU030 Robotics Meta Robot Law and Ethics — Who Is Liable When an Autonomous System Fails?
SU031 Agent Market Cap Physical Intelligence — Fastest Robotics Valuation, Embodied AI Commercialization Context
SU032 Finance and Money AI Startups With No Product, No Revenue Drawing Massive Investor Bets
SR001 European Union AI Act Service Desk AI Act Service Desk – Article 6: Classification rules for high-risk AI systems An AI system shall be considered to be high-risk where it is intended to be used as a safety component of a product covered by Union harmonisation legislation listed in Annex I and that product is required to undergo a third-party conformity assessment.
SR002 NIST (National Institute of Standards and Technology) AI Risk Management Framework | NIST NIST AI RMF 1.0 is structured around four core risk management functions: Govern, Map, Measure, and Manage. Seven trustworthy AI characteristics underpin the framework: valid/reliable, safe, secure/resilient, accountable/transparent, explainable/interpretable, privacy-enhanced, and fair with bias managed.
SR003 Legiscope EU AI Act: Practical Compliance Guide for 2026 Penalties for non-compliance with the EU AI Act can reach up to €35 million or 7% of global annual revenue for the most severe infractions, including non-adherence to prohibited or high-risk use requirements.
SR004 GDPR Register EU AI Act Compliance 2026 | Timeline, High-Risk AI Guide August 2, 2026: High-risk AI obligations become fully enforceable and compliance is required. August 2, 2028: Special rules for some embodied/robotic systems integrated into certain products take effect.
SR005 European Commission European approach to artificial intelligence The European Commission will finalise interpretative guidance and example use-cases by February 2026, with further consultations ongoing in 2026 regarding high-risk AI system classification for embodied and robotics applications.
SR006 RAND Corporation Liability for Harms from Artificial Intelligence Systems: The Application of U.S. Tort Law The fragmentation of AI development chains makes causation attribution complex and leaves model providers with unpredictable exposure under existing US strict product liability doctrines, as courts continue to work through how liability attaches when the AI software provider is separate from the hardware manufacturer.
SR007 Robotics Meta Robot Law and Ethics: Who is Liable When an Autonomous System Fails? Strict product liability means that if an autonomous robot malfunctions or is defective, manufacturers can be held liable regardless of intent or negligence. The claimant only needs to prove the product was defective and caused harm.
SR008 ACGRAM Robotics & Liability — Legal Frameworks for AI Robotics The EU AI Liability Directive is beginning to address gaps in strict liability for AI and robotics risks, clarifying insurance requirements for operators and deployers and moving toward shared liability models where product designers, software developers, hardware manufacturers, deployers, and users may all share liability.
SR009 International AI Safety Report (UK AI Security Institute) International AI Safety Report 2026 – Extended Summary for Policymakers General-purpose AI capabilities have continued to improve rapidly, but performance remains jagged, with systems still failing at some seemingly simple tasks. Some models are now capable of distinguishing between evaluation and deployment contexts and can alter their behaviour accordingly, creating new challenges around evaluation and safety testing.
SR010 PR Newswire 2026 International AI Safety Report Charts Rapid Changes and Emerging Risks Chaired by Turing Award-winner Yoshua Bengio, the 2026 International AI Safety Report brings together over 100 international experts backed by an Expert Advisory Panel with nominees from more than 30 countries and international organisations, including the EU, OECD and UN.
SR011 Inside Privacy (Covington) International AI Safety Report 2026 Examines AI Capabilities, Risks, and Safeguards The International AI Safety Report 2026 identifies accountability gaps in embodied AI as one of the field's most urgent governance challenges, emphasising that globally consistent standards for lifecycle governance and safety assurance have not yet been established.
SR012 Spheron Network GPU Shortage 2026: How to Secure AI Compute When GPUs Are Sold Out Lead times for H100 and even successor chips from major resellers remain 36–52 weeks, reflecting a structural supply bottleneck driven by High Bandwidth Memory (HBM3) supply constraints and TSMC CoWoS advanced packaging oversubscription.
SR013 Vamsi Talks Tech The GPU Supply Chain Crisis: What Every Enterprise CIO Must Know in 2026 NVIDIA is shifting capacity from mature H100 lines to higher-margin Blackwell production, meaning fewer new H100s, longer lead times, and allocation priority to cloud providers and hyperscalers, with the rest crowded out. Embodied AI teams must diversify sourcing and plan 12–18 months in advance.
SR014 8VC Announcing Our Investment in Generalist Pete Florence is as much a builder as a researcher — magnetic and unusually commercial for someone of his research caliber. We wanted to back him from the first conversation. Generalist is earliest on that curve and best positioned to lead.
SR015 Generalist AI GEN-1: Towards Robot Mastery GEN-1's improvisational intelligence is a strength—and a potential liability—that is central to its commercial proposition and its safety profile in real-world deployment environments.
SR016 SEC EDGAR Generalist AI Form D filing (CIK 0002137708) Generalist AI, Inc. (CIK 0002137708) filed Form D on 2026-06-05 for a securities offering under Rule 506(b); business address listed as San Francisco, CA; incorporated in Delaware.
SR017 Generalist AI Physical Commonsense At Generalist, we built lightweight handheld, ergonomic devices that let people manipulate objects almost as they would with their own hands. After a few minutes of doing a task, operators stop thinking and start reacting.
SR018 NVIDIA NVIDIA Releases New Physical AI Models as Global Partners Unveil Next-Generation Robots NVIDIA Isaac GR00T N1.6 is an open reasoning vision language action model, purpose-built for humanoid robots, that unlocks full body control and uses NVIDIA Cosmos Reason for better reasoning and contextual understanding. Open models are available on Hugging Face.
SR019 QubittTool Embodied AI 2026: The Year of Robot Foundation Models 2026 marks the inflection point where Embodied AI transitions from proof-of-concept to industrial-scale deployment. Industrial Deployment Analysis confirms that logistics warehousing achieves commercial scale while manufacturing flexible assembly enters batch pilots.
SR020 The Robot Report Generalist raises $400M to scale its general-purpose AI models Generalist AI has raised $400 million to scale its general-purpose AI platform for robotics, with stated plans to build next-generation models, scale the physical data engine, and expand the team.
SR021 Sacra Physical Intelligence – valuation, funding & news Physical Intelligence runs a B2B SaaS model for robotics companies, manufacturers, and automation integrators. Pricing is a $300 monthly subscription per connected robot, yielding recurring revenue that scales with fleet deployments.
SR022 Morgan Stanley Investment Management Embodied AI and the Rise of Humanoid Robots Broad adoption of humanoid robotics is still years away; early deployments may catalyse powerful learning flywheels, but the economic forces driving adoption must be weighed against the technical challenges that remain and the capital intensity required for broad deployment.
SR023 NeuroForge GTM Embodied AI Commercialization Challenges 2026 Physical AI startups have raised over $3.4 billion globally. Developing a humanoid platform like Tesla Optimus required $3–4 billion in R&D. Figure AI burns $200–300 million annually to sustain its development cycles. Long ROI horizons remain a moving target.
SR024 arXiv / SAE World Congress 2026 Embodied AI in Action: Insights from SAE World Congress 2026 on Safety, Trust, Robotics, and Real-World Deployment The panel reached broad agreement that long-term success will depend not only on advances in AI capability, but equally on safe and trustworthy deployment, requiring engineering rigor, lifecycle governance, human-centered design, and evolving safety standards.
SR025 Generalist AI Beyond World Models Generalist's architecture is explicitly goal-driven rather than method-driven; the company selects methods for scaling at capability and data scale rather than for architectural elegance.
SR026 Robotics Tomorrow Cobot Announces Second-Generation Proxie — Production-Tested Physical AI Cobot CEO Brad Porter, former VP of robotics at Amazon, cited the gap between lab-demonstrated generalisation and production-tested reliability as the central challenge facing embodied AI deployments in 2026.
SR027 Silicon Angle Generalist AI Raises $400M at $2B Valuation to Build General Intelligence for the Real World Generalist AI raised $400 million at a $2 billion post-money valuation led by Spark Capital, with participation from NVIDIA NVentures, 8VC, Union Square Ventures, Radical Ventures, and Hanabi Capital; the round includes angel investors Eric Yuan, Bin Lin, Fei-Fei Li, and Naval Ravikant.
SR028 The Mimic Physical Intelligence Is Raising Another $1B — Here's Why Investors Keep Betting on Humanoid Robots Training generalist robot policies requires massive amounts of diverse physical interaction data. More capital means more robots, more environments, more variation — all of which translates to better models. Physical Intelligence raised $1 billion at an $11 billion valuation.
SR029 Generalist AI Accelerating the Next Phase of Physical AI The data flywheel is beginning to take shape: real businesses are generating task data that feeds successive model generations. Planned use of the $400M: building next-generation models, scaling the physical data engine, and expanding the team.
SR030 The Robot Report Generalist introduces GEN-1 general-purpose model for physical AI Generalist AI unveiled GEN-1, a general-purpose model for physical AI, available via an early-access program to selected industry partners starting April 2, 2026; the company has not named any commercial customers or disclosed pricing.
SV001 VentureBeat Generalist AI raises $400M at $2B valuation to build general intelligence for robotics Generalist AI just closed a $400 million funding round that values the robotics startup at $2 billion post-money. Radical Ventures led the round.
SV002 The Robot Report Generalist raises $400M to scale its general-purpose AI models Radical Ventures led Generalist AI's latest funding. New investors included 8VC, Union Square Ventures, Hanabi Capital, and Norwest.
SV003 SiliconANGLE Generalist AI raises $400M at $2B valuation to build general intelligence for robotics Chief executive Pete Florence is a former DeepMind senior scientist who helped build RT-2 and PaLM-E.
SV004 8VC Announcing our investment in Generalist AI Pete Florence is as much a builder as a researcher — magnetic and unusually commercial for someone of his research caliber. We saw remarkable early commercial traction.
SV005 Humphrey Theodore Generalist AI Raised $400M to Put an AI Foundation Model Inside Robots GEN-1 runs roughly three times faster than comparable state-of-the-art models, holds about 99% reliability across a diverse spread of physical tasks.
SV006 Sacra Physical Intelligence company profile — valuation, funding, business model Physical Intelligence closed a $600 million Series B in November 2025 led by CapitalG at a $5.6 billion post-money valuation. Pricing is a $300 monthly subscription per connected robot.
SV007 Sacra Figure AI company profile — valuation, funding, business model Figure AI reached a $39 billion post-money valuation in September 2025 following a Series C funding round; Figure 02 ran daily 10-hour shifts at BMW Spartanburg, loading over 90,000 parts.
SV008 TechCrunch Robotics software maker Skild AI hits $14B valuation Skild AI has raised a $1.4 billion Series C round that values it at more than $14 billion. The round was led by SoftBank, and Nvidia, Macquarie Group, and others also invested.
SV009 AgentMarketCap Physical Intelligence Hits $11B in 4 Months: The Fastest Valuation Markup in Robotics VC History From a $400 million seed valuation in March 2024 to an $11 billion raise just two years later, Physical Intelligence has become the fastest-appreciating startup in robotics history.
SV010 TechFundingNews Physical Intelligence eyes $1B raise at $11B valuation, Founders Fund and Lightspeed in talks Physical Intelligence is reportedly closing a $1 billion round that would double its $5.6 billion valuation from just four months ago.
SV011 Finerva Robotics & AI: 2026 Valuation Multiples After bottoming out at 2.5x in Q1 2025, the median revenue multiple rose to 3.4x by Q4 2025. EV/EBITDA multiples reached 16.8x in Q4 2025, ranging from 0.9x to 78.2x.
SV012 Finro AI Valuation Multiples (Q1 2026) | 575 Company Dataset LLM Vendors: 575 companies covered; 27 in LLM Vendors segment; avg EV/Rev 73.5x, median EV/Rev 39.5x; 25th–75th percentile 16.7x–75.9x.
SV013 AIRoboticDaily Embodied AI Bubble: Humanoid Robot Market Valuation & Trends Truly productive revenue from industrial scenarios accounts for merely three to five percent of total sales. Investors are paying massive premiums for a universal productivity tool, but the current financials reflect a high tech toy.
SV014 SEC EDGAR Generalist AI, Inc. — Form D (Filing date 2026-06-05) CIK 0002137708; Generalist AI, Inc.; incorporated Delaware; 455 Market Street, Suite 1940, San Francisco CA 94105; Pete Florence listed as Executive Officer, Director, Promoter; filing date 2026-06-05.
SV015 CB Insights The physical AI models market map: Behind the arms race to control robot intelligence The robotics sector raised a record $40.7B in 2025 — up 74% YoY and 9% of all venture funding — making it a funding leader alongside AI software.
SV016 HumanoidsDaily Physical AI arms race accelerates — Generalist AI secures $400M to scale robot learning Rather than focusing on a single embodiment, Generalist is positioning its software as a cross-form-factor intelligence layer for humanoids, industrial arms, and mobile platforms.
SV017 Generalist AI GEN-1: General-purpose model for physical AI — official blog post GEN-1 improves average success rates to 99% on tasks where previous models achieved 64%, completes tasks roughly 3x faster, and adapts to any robot in one hour of task-specific data.
SV018 Generalist AI Accelerating the next phase of physical AI — Series B announcement blog A data flywheel has begun to take shape: real businesses are generating task data that feeds the next generation of more capable models.
SV019 The Robot Report Generalist introduces GEN-1: general-purpose model for physical AI Generalist said its GEN-1 unlocks commercial viability across a broad range of applications and launched an early-access program with selected industry partners.
SV020 Robotics and Automation News Generalist AI raises $400 million to scale its robot intelligence platform The round was led by Radical Ventures, with participation from NVIDIA NVentures, Bezos Expeditions, 8VC, Union Square Ventures, Hanabi Capital, and Norwest.
SV021 EmbodiedGlobal Generalist AI Raises $400M to Build Foundation Models for Any Robot Generalist AI has raised $400M at a $2B valuation; the round brings total funding to more than $500M since founding.
SV022 FinsmeS Generalist AI Raises $400M in Funding at $2 Billion Valuation Generalist AI raises $400M in funding at a $2 billion valuation led by Radical Ventures.
SV023 Silicon Republic NVIDIA and Fei-Fei Li back Generalist's $400M round to scale AI robotics US AI robotics company Generalist has raised $400m at a reported $2bn valuation to accelerate plans toward achieving physical AGI.
SV024 CryptoBriefing Generalist AI raises $400M led by Radical Ventures New participants include 8VC, Union Square Ventures, Hanabi Capital, and Norwest. Angel investors included Zoom founder Eric Yuan and renowned AI researcher Fei-Fei Li.
SV025 Daily Upside $2 Billion Start-Up Generalist AI Wants to Solve the Robot Data Bottleneck Nvidia- and Jeff Bezos-backed robotics software startup Generalist AI raised $400 million at a $2 billion valuation to help build its AI-for-robots system.
SV026 Fundraise Insider Generalist AI Raises $400M Series B at $2B Valuation Generalist AI, a robotics AI company, has raised $400 million in a Series B round at a post-money valuation of $2 billion, led by Radical Ventures. The company has raised more than $500 million since its founding.
SV027 Morgan Stanley Embodied AI and the Rise of Humanoid Robots Advances in AI are accelerating the transition of humanoid robots from long-term ambition to early industrial deployment; broad adoption is still years away.
SV028 Grand View Research / Research and Markets Embodied AI Market Size, Share & Trends Analysis Report 2026–2033 Embodied AI market was valued at $6.5 billion in 2026 and is projected to reach $67.6 billion by 2033 at a 39.7% CAGR; North America held 35.6% of the 2025 market.
SV029 Robotics Tomorrow Cobot Announces Second-Generation Proxie, Bringing Production-Tested Physical AI to Real Operations First generation Proxie has logged 12,627 operating hours in production environments. Brad Porter: "For decades, deploying robots has meant choosing between mobility and dexterity."
SV030 Generalist AI GEN-0: Bringing robots into the pretraining era GEN-0 trained on 270,000 hours of real-world physical interaction data; demonstrated scaling laws in robotics — more data and larger models predictably produce more capable systems.
SV031 Generalist AI About Generalist AI — company mission and overview Generalist AI is a frontier AI research and product-driven company building general intelligence for the physical world.
SV032 TechCrunch Many AI startups are raising at astronomical valuations AI startups are raising at multiples that would be hard to justify even with optimistic revenue projections, with many at Series B still pre-revenue and valued above $1 billion.
SV033 AI Funding Tracker Top Humanoid Robotics Startups Funded in 2026 Goldman Sachs projects the humanoid robotics market to reach $38 billion by 2035; robotics startups raised $13.8 billion globally in 2025, up from $7.8 billion in 2024. The top 10 humanoid robotics companies have captured nearly 80% of all capital raised in the category since 2022. Companies without strong corporate partnerships, real deployment data, or a defensible software moat will find it much harder to raise in 2026 than in 2024.
SV034 TechCrunch Anthropic raises $65 billion, nears $1T valuation ahead of IPO Anthropic has snagged $65 billion in funding at a $965 billion post-money valuation, marking what could be the AI startup's last private fundraising before debuting on the public markets. The company said its run rate revenue crossed $47 billion earlier this month. OpenAI last raised a whopping $122 billion round in March at an $852 billion post-money valuation.
SV035 RobotToday China's Robot Boom: $1.3 Billion in Orders in 3 Months as Unicorn Valuations Climb but Still Trail the U.S. Investors note that if Chinese robotics firms were based in the U.S., their valuations could easily be five to ten times higher. U.S. investors reward long-term potential and global reach, often funding companies without immediate commercial revenue. Chinese funds tend to be more pragmatic, anchoring valuations to demonstrable product deployments.