Physical Intelligence
机器人基础模型——顶级团队、极端估值、零收入
谨慎 — 团队世界级、估值极端;入场前先等商业化证据
封面要素
公司概况
Physical Intelligence 于 2024 年 3 月在旧金山成立,创始团队由一批顶级机器人 AI 研究者组成——Karol Hausman(CEO,前 Google DeepMind)、Sergey Levine(首席科学家,UC Berkeley RAIL Lab)、Chelsea Finn(MAML 共同发明人,Stanford AI Lab)、Brian Ichter(前 Google DeepMind)、Adnan Esmail(工程副总裁,前 Anduril/Tesla)和 Lachy Groom(前 Stripe)。公司目标是打造通用机器人基础模型:一个能控制任意机器人、完成任意物理任务的 AI 系统,定位类似 GPT-4 之于文本任务。旗舰 π₀ 模型横跨 68 种机器人形态训练,在灵巧操作基准上超过此前最强水平。公司不到 20 个月完成三轮融资(种子轮、Series A、Series B),累计融资 $1.07 billion,并在 2025 年 11 月达到 $5.6 billion 估值,但商业收入仍为零。制造和物流领域已有企业试点,AgiBot 和 Longcheer Technology 被点名为早期合作方。公司所处赛道竞争激烈,对手包括 Skild AI($14B 估值、$30M ARR)、Google DeepMind Gemini Robotics 和 Figure AI($39B 估值),三者分别带来不同类型的竞争压力。
- 成立时间
- 2024-03-01
- 创始人
- Karol Hausman, Sergey Levine, Chelsea Finn, Brian Ichter, Adnan Esmail, Lachy Groom
- 创立地点
- San Francisco, California, USA
- 总部
- San Francisco, California, USA
- 产品
- π₀(pi-zero):跨形态 VLA,靠 flow matching 将 PaliGemma 3B VLM 与 300M 参数动作专家结合,支持 68+ 种机器人形态。π₀.5:增强版互联网规模预训练变体。π₀-FAST:优化低延迟解码器。openpi:开源微调框架(Apache 2.0)。计划中的商业产品是按机器人授权的 SaaS。
- 客户
- 仍处商业化前阶段。目标客群包括大型制造企业(汽车、电子、消费品)和机器人队列密度高的物流运营商。早期试点包括 AgiBot 和 Longcheer Technology(亚洲制造商)。截至 2026 年 Q1,未披露美国或欧洲的具名商业客户。
- 商业模式
- 计划采用按机器人授权的 SaaS 模式,估算价格为每台机器人每年 $5,000–$15,000。额外收入来源包括企业微调、部署支持,以及借助机器人 OEM 伙伴的潜在 B2B2B 模式。公司仍无收入;未公布商业定价。
- 阶段
- Series B — $600M at $5.6B post-money valuation (November 2025)
- 融资情况
- 种子轮 $70M(2024 年中,Lux Capital);Series A $400M,估值 $2.4B(2024 年 11 月,OpenAI/Thrive/Lux/Sequoia/Index/Bond/Bezos);Series B $600M,估值 $5.6B(2025 年 11 月,CapitalG/T. Rowe Price/Redpoint/Lux)。累计融资 $1.07B。据报道,下一轮估值为 $11B(2026 年 4 月,未确认)。CapitalG(Alphabet)是 Series B 领投方。
执行摘要
主要优势
- 创始团队世界级——Sergey Levine(RAIL Lab)、Chelsea Finn(MAML)和 Karol Hausman(DeepMind)都位列全球机器人 AI 引用最高研究者前五,技术信用和吸引顶尖研究者的能力都很强
- 跨 68 类机器人形态训练,是机器人 AI 里最有辨识度的能力主张;若能在商业规模验证,这种广度会形成数据飞轮和泛化优势,竞争者需要多年才能复制
- π₀ 架构把 PaliGemma 3B VLM 和 flow-matching 动作专家结合起来,在面向灵巧操作的 LIBERO 基准上明显跑赢 OpenVLA、RT-2 和 Octo;灵巧操作仍是机器人最难解的任务
- 已融资 $1.07B,在商业化前规模下估计可支撑 4–8 年现金跑道;对研究阶段公司来说财务韧性罕见,资本池本身就是竞争护城河
- openpi 开源发布已经带来开发者社区显著采用,公司几乎不用额外投入,就拿到生态善意和外部研究验证
主要风险
- $5.6B 估值下商业收入为零,是最核心的投资逻辑破裂风险;若企业试点到 2026 年 Q4 仍未转成带明确 ARR 的商业合同,据称 $11B 下一轮融资窗口会关闭,公司将面对艰难融资环境
- Skild AI 已做到 $30M ARR,并在积累商业数据飞轮,每个季度都拉开差距;Physical Intelligence 在商业牵引力上的竞争劣势正在加速扩大
- Google DeepMind Gemini Robotics 带来生存级的算力和分销不对称威胁;Google 同时是投资方(CapitalG)、技术依赖(PaliGemma)和竞争者
- 架构上依赖 Google 的 PaliGemma,形成单点故障风险;若 Google 限制授权,公司需要 $10M+ 重新训练模型并延迟 6–12 个月,且未披露合同保护
- 关键人物集中在 Sergey Levine 和 Chelsea Finn;商业发布前任一人离开都会触发投资逻辑破裂,且没有披露留任安排或接班计划
- EU AI Act 高风险分类和功能安全认证(ISO 13849)要求,会给欧洲商业部署增加 12–24 个月合规门槛,目前没有披露进展
未决问题
- 完整企业试点客户名单,包括可点名 reference、LOI 状态和试点转商业时间表——投资前最重要的商业尽调事项
- 与 Google 的 PaliGemma 商业授权协议——当前 Gemma Terms of Use 不足以提供投资级 IP 保护,需要正式合同保护
- 训练数据来源审计——π₀ 训练所用机器人演示视频的版权状态未披露;这项重大 IP 责任风险必须在投资前解决
- 烧钱速度、在手现金和管理账——估计年烧钱 $70–$150M,区间太宽,无法承保投资;需要通过 VDD 拿到实际数字
- Sergey Levine 和 Chelsea Finn 的 vesting 时间表与留任协议——若投资人看不到合同保护,关键人物风险就没有被缓释
目录
01公司概览
1.1 公司身份与业务概览
Physical Intelligence 以 pi.ai 名义经营,2024 年 3 月在加利福尼亚州旧金山注册并启动运营。公司给出的使命是打造「一个单一、通用的 AI 系统,能够控制任意机器人完成任意任务」。业务落在机器人、强化学习和大型基础模型的交叉点。核心产品是 π₀(pi-zero),一个视觉-语言-动作(VLA)基础模型:它读取摄像头图像、自然语言任务指令和机器人本体感知状态,再借助 flow-matching transformer 预测连续电机动作。π₀ 不绑定硬件——不必从零重建,就能微调到多类机器人平台(机械臂、移动操作机器人、人形机器人)上运行。Physical Intelligence 的商业模式瞄准企业客户,按机器人数量以 SaaS 方式授权其模型栈;但截至 2026 年 Q1,公司仍处商业收入前阶段,重心放在研究部署和试点项目上。公司还发布了开源变体 openpi,推动生态采用和训练数据贡献。 [CO001, CO002, CO003, CO004, CO005]
Physical Intelligence 关键融资事件和产品里程碑的时间线。
[CO033, CO034, CO035]逻辑流程,展示研究、算力和生态输入如何连接到企业机器人部署。
[CO003, CO004, CO005, CO039]1.2 创始人、管理层与治理
Physical Intelligence 由七名成员共同创立,能力覆盖学术机器人研究、应用 AI 工程和运营扩张。CEO Karol Hausman 曾任 Google DeepMind Staff 研究科学家,负责机器人操作研究并参与 RT-2,同时担任过 Stanford University 兼职教授。首席科学家 Sergey Levine 是 UC Berkeley 终身副教授,领导 Robotic AI and Learning(RAIL)Lab,该实验室产出了机器人深度强化学习的基础研究。联合创始人 Chelsea Finn 是 Stanford 助理教授,以 Model-Agnostic Meta-Learning(MAML)和快速机器人适应研究闻名。另一位联合创始人 Brian Ichter 曾任 Google DeepMind 研究科学家,聚焦运动动力学规划和 GPU 加速机器人算法。联合创始人兼工程副总裁 Adnan Esmail 曾任 Anduril 工程高级副总裁,也做过 Tesla Autopilot 工程师。Lachy Groom(早期 Stripe 高管和 VC 投资人)与 Quan Vuong(机器人强化学习研究者)补齐创始团队。董事会构成尚未公开;技术领导高度集中,关键人物风险偏高。截至 2026 年 Q1,公司未披露治理争议或管理层变动。 [CO006, CO007, CO008, CO009, CO010, CO011]
| 姓名 | 角色 | 过往任职 | 领域专长 | 创始人与市场匹配度 | 关键人物风险 |
|---|---|---|---|---|---|
| Karol Hausman | CEO 兼联合创始人 | Google DeepMind(资深研究科学家),Stanford(兼职教授) | 机器人操作、通用机器人学习、RT-2 | 主导开创性操作研究;具备连接学术与产品的运营经验 | 关键——唯一 CEO;兼具深厚技术与 GTM 领导力 |
| Sergey Levine | 首席科学家兼联合创始人 | UC Berkeley(终身副教授),RAIL Lab | 深度强化学习、离线 RL、机器人控制 | RAIL Lab 创始人;其研究支撑 π₀ 架构 | 高——离开会损害研究可信度和人才管线 |
| Chelsea Finn | 联合创始人兼顾问 | Stanford University(助理教授),UC Berkeley(PhD) | 模型无关元学习(MAML)、机器人快速适应 | 发明 MAML;MAML 是 π₀ 快速任务泛化的核心理论路径 | 中——顾问角色;即使未全职投入,工作仍可推进 |
| Brian Ichter | 联合创始人兼研究员 | Google DeepMind / Google Brain,Stanford(PhD)背景 | 动力学约束规划、GPU 加速机器人算法、运动规划 | 在机器人导航和操作的可扩展规划上经验很深 | 中——研究贡献者;团队没有他也能继续推进 |
| Adnan Esmail | 联合创始人兼工程副总裁 | Anduril(工程高级副总裁),Tesla(Autopilot) | 软硬件集成、国防 AI、自动驾驶车辆 | 对把 π₀ 从研究推进到生产级工程至关重要 | 高——离开会拖慢机器人部署扩张 |
| Lachy Groom | 联合创始人兼业务负责人 | Stripe(早期高管),Founder Fund(VC) | 产品、商务拓展、风险投资、GTM 策略 | 为学术型创始团队补上商业化和融资能力 | 中——非技术角色;可通过招聘替代或补强 |
| Quan Vuong | 联合创始人兼研究员 | 机器人 RL 研究员(曾在产业实验室) | 机器人学习算法、强化学习 | 参与 π₀ 训练管线和研究议程 | 低——多名研究贡献者之一 |
董事会构成未公开披露。未公布独立董事。关键人物集中度高,尤其是 Hausman 和 Levine。截至 2026 年 Q1,未披露人员离职或治理变化。
[CO006, CO007, CO008, CO009, CO010, CO011]1.3 融资历史、估值与投资方
Physical Intelligence 约 20 个月融资约 $1.07 billion。$70 million 种子轮在 2024 年上半年至年中完成,投资方包括 Lux Capital 和 Jeff Bezos 等,公司未披露 pre-seed/seed 估值。$400 million Series A 于 2024 年 11 月完成,投后估值 $2.4 billion;投资方包括 OpenAI、Thrive Capital、Lux Capital、Index Ventures 等。$600 million Series B 一年后于 2025 年 11 月完成,投后估值 $5.6 billion,由 CapitalG(Alphabet 的成长基金)领投,Lux Capital、Bond、Redpoint、Sequoia Capital、Thrive Capital、Index Ventures、T. Rowe Price 和 Jeff Bezos 参投。截至 2026 年 4 月,据报道公司正就新一轮融资进行深入谈判,估值 $11 billion;若完成,意味着不到六个月估值约翻倍。所有外部融资均为风险股权;未公开披露债务、老股转让或收入分成融资。大额融资反映投资人押注机器人基础模型的市场机会,也押注团队的研究履历。 [CO014, CO015, CO016, CO017, CO018, CO019]
| 指标 | 数值 / 状态 | 日期 | 置信度 | 缺口 / 注意事项 |
|---|---|---|---|---|
| 估值(最新) | $5.6B 投后估值 | Nov 2025 | 高 | Bloomberg 报道;Series B 轮已确认 |
| 累计融资 | ~$1.07B($70M 种子轮 + $400M Series A 轮 + $600M Series B 轮) | Nov 2025 | 高 | 仅股权融资;未披露债务或收入融资 |
| 年经常性收入(ARR) | $0(商业化前) | Q1 2026 | 高 | 公司仍在研究阶段;企业试点进行中,但未披露 ARR |
| 主要产品 | π₀ VLA 机器人基础模型(开源权重 + 企业栈) | Feb 2025 | 高 | openpi 仓库公开可用;企业版包含专有微调栈 |
| 员工数(估计) | ~150–250 名员工 | Q1 2026 | 低 | 未公开披露;基于 LinkedIn 和招聘节奏估计 |
| 总部 | 美国加州旧金山 | Mar 2024 | 高 | 公司文件和新闻稿确认 |
| 成立时间 | March 2024 | Mar 2024 | 高 | 多家独立报道确认 |
| 支持的机器人本体 | 已测试 10+ 个机器人硬件平台 | Early 2025 | 中 | π₀ 声称具备跨本体泛化;独立验证有限 |
| 下一轮融资 | 围绕 ~$11B 估值进入深入谈判(未确认) | Apr 2026 | 低 | 据媒体报道;公司未确认 |
| 利益相关方 | 角色 / 类型 | 融资轮次 | 估算持股 / 影响力 | 尽调问题 |
|---|---|---|---|---|
| CapitalG(Alphabet 成长基金) | Series B 轮领投方 | Series B 轮($600M) | 可能是 B 轮后最大机构持股方;大概率有董事席位 | 确认董事会席位;评估与 Google DeepMind 机器人业务的冲突 |
| Thrive Capital | 战略投资人 | Series A 轮、Series B 轮 | 多轮参与;早期影响力显著 | 确认轮次规模和按比例跟投权;评估战略价值与财务价值 |
| Lux Capital | 领投 / 锚定投资人 | 种子轮、Series A 轮、Series B 轮 | 最早机构投资人;累计持股最高 | 确认董事会构成;核查 Lux 投资组合中是否有竞争性机器人项目冲突 |
| Index Ventures | 投资人 | Series A 轮、Series B 轮 | 自 Series A 起持有可观股份;B 轮继续跟投 | 标准 LP 披露;评估 EMEA 市场进入支持能力 |
| T. Rowe Price | 后期 / 跨阶段投资人 | Series B 轮 | 机构跨阶段投资人;验证 IPO 准备可选性 | 确认 IPO 前流动性偏好;评估锁定期条款 |
| Jeff Bezos(个人) | 战略天使投资人 | 种子轮、Series B 轮 | 少数个人持股;象征意义和媒体价值高 | 无治理角色;注意与 Amazon Robotics 的潜在竞争冲突 |
| OpenAI | 战略投资人 | Series A 轮 | 企业 VC 战略兴趣;董事会角色不清 | 评估是否存在排他或数据共享安排;持续监控竞争冲突 |
| Sequoia Capital | 投资人 | Series B 轮 | 跨阶段投资人;增强治理和退出可信度 | 标准项;评估 Sequoia 在相邻机器人投资中的组合冲突 |
具体持股比例、稀释和董事会构成未公开披露。估算基于轮次规模和参与模式。CapitalG(Alphabet)与 OpenAI 作为共同投资人之间的潜在治理冲突,是尽调信号。
[CO015, CO016, CO017, CO018, CO019]KPI 快照,用于快速评估 Physical Intelligence 的阶段和规模指标。
[CO028, CO029, CO039]1.4 关键里程碑
Physical Intelligence 从成立到融资超过 $1B 并发布开源模型,用时不到两年。公司在 2024 年 3 月宣布成立和种子轮融资。2024 年 10 月,π₀ 模型出现在技术博客和研究论文中,展示洗衣物折叠、盒子组装、衬衫包装等多任务机器人控制。Series A 在 2024 年 11 月完成。2025 年 2 月,公司将 π₀ 模型权重和代码以 openpi 仓库开源,成为首个开源通用机器人 VLA 基础模型。后续 π₀.5 于 2025 年初发布,强化了跨新环境的泛化。2025 年中,π₀-FAST 推出,采用频域离散动作表示(FAST tokenizer)提升推理效率。$5.6 billion 估值的 Series B 在 2025 年 11 月完成。截至 2026 年 Q1,早期企业试点已在制造和物流垂直场景启动,客户未具名;公司据报正接近 $11B 估值的下一轮融资里程碑。截至本报告日期,未披露负面事件、监管调查或重大高管离职。 [CO021, CO022, CO023, CO024, CO025, CO026]
| 日期 | 事件 | 类型 | 金额 / 估值 / 状态 | 参与方 / 详情 | 影响 |
|---|---|---|---|---|---|
| Mar 2024 | 公司在旧金山成立 | 创立 | N/A | Karol Hausman、Sergey Levine、Chelsea Finn、Brian Ichter、Adnan Esmail、Lachy Groom、Quan Vuong 等创始成员 | 以构建通用机器人 AI 为使命成立;前所未有地快速集结了学术机器人领军者 |
| Mid 2024 | 完成种子轮融资 | 融资 | $70M | Lux Capital 领投;Jeff Bezos 参投 | 支撑初始研究团队搭建和算力采购;产品前阶段 |
| Oct 2024 | 发布 π₀ 模型并发表技术论文 | 产品 | N/A | Physical Intelligence 博客 + arXiv 论文;演示叠衣、组装盒子、包装衬衫 | 首次展示跨本体 VLA 模型执行长时程灵巧任务 |
| Nov 2024 | 完成 Series A 轮,估值 $2.4B | 融资 | $400M,投后估值 $2.4B | OpenAI、Thrive Capital、Lux Capital、Index Ventures 等 | 当时机器人 AI 史上最大 Series A 轮;验证硬件无关路径 |
| Feb 2025 | π₀ 以 openpi 仓库形式开源 | 产品 | N/A | GitHub 发布;模型权重 + 训练代码公开可用 | 生态打法:首个开源通用机器人 VLA 基础模型;建设开发者社区 |
| Early 2025 | 发布 π₀.5——增强泛化 | 产品 | N/A | 博客文章;提升新型机器人平台的零样本表现 | 展示模型迭代节奏和持续提升的泛化能力 |
| Mid 2025 | 推出 π₀-FAST——高效推理变体 | 产品 | N/A | 用于离散频域动作表示的 FAST 分词器 | 针对商业部署的推理成本障碍;运行时算力更低 |
| Nov 2025 | 完成 Series B 轮,估值 $5.6B | 融资 | $600M,投后估值 $5.6B | CapitalG 领投;Lux、Bond、Redpoint、Sequoia、Thrive、Index、T. Rowe Price、Bezos | 累计融资超过 $1B;2025 年全球最大机器人 AI 软件创业公司融资轮 |
| Q1 2026 | 早期企业试点项目进行中 | 规模化 | 未披露 | 未具名制造和物流客户;未披露 ARR | 首个商业验证步骤;仍在收入前,但商业意图已确认 |
| Apr 2026 | 据报道,就 ~$11B 估值进行深入融资谈判 | 融资 | $11B(未确认) | 据媒体报道;无正式公告 | 约 5 个月内较 $5.6B 提升 2×;若属实,反映投资人需求强劲 |
时间线基于公开报道日期。董事会会议纪要和内部里程碑不可得。反向事件、监管行动、治理争议:截至 2026 年 Q1 均未披露。
[CO001, CO014, CO015, CO016, CO021, CO022]1.5 展项
02市场分析
2.1 市场定义与细分
Physical Intelligence 面对的市场,最好拆成三个同心层。最外层是全球物理 AI 与机器人(硬件加软件),覆盖工业机器人、移动操作机器人、人形机器人和自动驾驶车辆;2025 年 TAM 估计为 $50–82 billion,到 2030 年增至 $111–185 billion。其内层是企业机器人软件与 AI 栈(模型、部署基础设施、训练流水线),2025 年 SAM 约 $8–15 billion,预计到 2030 年达到 $20–35 billion,CAGR 为 20–35%。Physical Intelligence 最直接的可服务市场,是机器人基础模型软件和按机器人授权的 SaaS。该品类尚未形成成熟分析师分类,但按单台机器人经济性,以及制造、物流和仓储自动化中的部署规模推演,2028 年可粗略外推为 $3–8 billion。关键垂直包括:制造与装配(今天收入最大)、物流与仓储自动化(增长最快)、商业与设施服务,以及医疗和零售中由人形机器人打开的新兴任务。Physical Intelligence 目前不绑定硬件,目标客户是机器人 OEM 伙伴,以及已经拥有或计划部署机器人队列的企业客户。 [CM001, CM002, CM003, CM004, CM005]
| 市场层级 | 定义 | 范围包括 | 范围不包括 | Cohere 相关性 | 分析师来源 |
|---|---|---|---|---|---|
| Physical AI / Robotics(广义 TAM) | 所有 AI 驱动的机器人系统,包括硬件、软件和服务 | 工业机器人、移动操作臂、人形机器人、自动驾驶车辆、无人机 | 纯汽车(EV)、不含机器人的消费电子 | 间接——Pi 瞄准该市场中的软件层 | Grand View Research、MarketsandMarkets |
| 企业机器人软件 / AI 栈(SAM) | 面向企业机器人的 AI 模型、部署基础设施、训练管线 | 机器人基础模型、控制软件、仿真工具、部署栈 | 机器人硬件制造、传感器、执行器 | 直接——Pi 的核心产品是机器人 AI 基础模型和微调栈 | MarketsandMarkets AI-in-Robotics 报告 |
| 机器人基础模型软件 / 单机器人 SaaS(SOM 目标) | 按机器人或按部署授权通用机器人智能 | VLA 模型授权、企业微调、推理基础设施 | 面向特定任务的定制机器人软件、硬件集成服务 | 完全匹配——Pi 计划中的单机器人 SaaS 商业模式 | CB Insights,并由 LLM SaaS 市场类比推断 |
嵌套 TAM-SAM-SOM 金字塔,展示 Physical Intelligence 的市场机会层级,从广义物理 AI 到机器人基础模型 SaaS。
[CM026, CM027, CM028]2.2 TAM、SAM 与 SOM 分析
Grand View Research 估计,物理 AI 市场(具身 AI、机器人智能软件、自主系统)2025 年约 $81.6 billion,至 2033 年 CAGR 约 32%。MarketsandMarkets 估计,机器人 AI 市场(仅智能与控制软件)2025 年为 $5.8 billion,到 2030 年增至 $25.9 billion。McKinsey 和 BCG 的分析认为,当专有智能成为相对商品化硬件的主要差异化因素,机器人软件和 AI 层最终将拿走机器人价值链利润的 40–60%。Physical Intelligence 短期 SOM(2026–2028)受限于制造与物流的早期企业试点;若公司能拿下预计会在该窗口采用机器人基础模型软件的 100–300 个企业账户中的 10–30%,SOM 估计为 $200–500 million。实现这一点需要成功商业发布、可防守的按机器人授权经济性,以及硬件伙伴规模化。 [CM006, CM007, CM008, CM009, CM010, CM011]
| 层级 | 2025 估计 | 2030 估计 | CAGR | 依据 / 核心假设 | 置信度 |
|---|---|---|---|---|---|
| Physical AI / Robotics TAM(硬件 + 软件) | $50–82B | $111–185B | 30–40% | Grand View Research physical AI 市场;MarketsandMarkets AI 机器人市场 | 中 |
| AI-in-Robotics 软件 SAM | $5.8B | $25.9B | 35% | MarketsandMarkets AI-in-Robotics Market 2025–2030 报告 | 中 |
| 机器人基础模型软件(类比 LLM API) | $0.5–1B(萌芽期) | $3–8B(预测) | 40–60% | 尚无成熟分析师品类;基于每机器人授权 $5–15K/year × 预测机队规模估算 | 低 |
| Physical Intelligence SOM(3 年) | $0(收入前) | 到 2028 年 $200–500M | N/A | 在约 100–300 个采用机器人基础模型软件的企业账户中拿下 10–30% | 低 |
所有估算都有较宽不确定区间。具体的机器人基础模型软件品类尚无成熟分析师覆盖;SOM 是基于市场类比和单位经济推导的估算。
[CM005, CM006, CM007, CM008, CM009]2025 年和 2030 年与 Physical Intelligence 相关的每个市场层级的低 / 高分析师估计区间。
所有数值单位为十亿美元。区间来自多份分析师报告;区间很宽,反映市场定义高度不确定。
[CM029, CM030, CM031]企业账户从认知到商业部署机器人基础模型平台的示意漏斗(2025–2028)。
数值为示意性估计,基于企业技术采用 S 曲线类比;并非来自 Physical Intelligence 数据。
[CM035, CM008]2.3 细分客群与买方分析
Physical Intelligence 的主要买方分三类:(1)希望把基础模型智能嵌入自家硬件的机器人 OEM 制造商(如 Agility Robotics、Boston Dynamics 和较小的人形机器人初创公司);(2)在制造或物流中部署机器人的大型企业运营商,它们希望跨机器人泛化,不想为每个任务单独编程;(3)为工厂和仓库打造交钥匙自动化方案的系统集成商。企业账户的决策人通常是运营副总裁、首席自动化官或工程副总裁,采购流程还会牵涉 IT 安全审查、数据驻留要求和集成审批。企业机器人从试点到合同的购买周期为 12–24 个月;Physical Intelligence 目前处于试点阶段。OEM 伙伴渠道能带来更快分发,但毛利更低,也更依赖硬件伙伴成败。北美和欧洲目前是主要市场;亚太(尤其日本、韩国和中国)因工业机器人密度高,代表长期最大放量市场。 [CM012, CM013, CM014, CM015, CM016]
| 细分市场 | 买方类型 | 决策者 | 用例 | ACV 潜力 | 采用阶段 |
|---|---|---|---|---|---|
| 制造与装配 | 企业运营方(OEM、一级供应商) | 运营副总裁 / 首席自动化官 | 面向装配任务的跨机器人泛化;降低单任务编程成本 | 每个部署点 $50K–$500K/yr | 早期试点;12–24 个月销售周期 |
| 物流与仓库自动化 | 3PL 运营商、电商履约 | 工程副总裁 / 自动化负责人 | 分拣、抓取放置、使用通用操作完成包装 | 每个设施 $100K–$1M/yr | 早期试点;对规模化有战略意义 |
| 机器人 OEM 制造商 | 人形机器人和机械臂制造商 | CTO / 产品负责人 | 将 π₀ 嵌入为智能层;降低模型开发成本 | 合作 / 授权尚未披露;潜在 $5–15K/robot/yr | 商业化前合作讨论 |
| 系统集成商 | 自动化方案供应商 | 工程负责人 / 解决方案架构师 | 用 PI 模型跑客户机器人,交付一站式工厂自动化 | 按项目分成;间接渠道 | 刚起步;尚无已宣布集成 |
| 商业与设施服务 | 物业管理方、酒店业 | 运营总监 | 具备通用能力的清洁、维护、巡检机器人 | 每个部署 $20K–$200K/yr | 尚未达到——技术成熟度低于门槛 |
制造或物流企业买方评估并采用机器人基础模型平台时经历的阶段。
[CM032, CM033, CM034]2.4 增长驱动与市场约束
关键增长驱动包括:(1)人口结构推高制造和物流劳动力短缺,机器人 ROI 更容易算通;(2)基础模型能力快速提升,把单个定制技能 $50,000–$250,000 的机器人任务编程成本,靠微调压到接近零;(3)机器人硬件成本下降,尤其协作机器人和机械臂,扩大可部署基数;(4)开放数据集和合成仿真让机器人训练数据更容易获得;(5)资本强势流入该赛道,推高生态建设动能。约束和风险包括:(1)商业部署对安全与可靠性的要求极高;(2)与传统工业系统集成仍然复杂;(3)π₀ 开源释放让基础模型商品化,削弱 PI 维持专有护城河的能力;(4)Skild AI($14B 估值)、Google DeepMind、Figure AI 等资金充足的竞争对手可能执行更快;(5)相对早期商业经济性,训练和推理算力成本偏高。 [CM017, CM018, CM019, CM020, CM021, CM022]
| 因素 | 类型 | 方向 | 影响强度 | 时间窗口 | 来源 / 依据 |
|---|---|---|---|---|---|
| 制造业和物流业劳动力短缺(人口结构) | 驱动 | 顺风 | 高 | 现在–5 年 | ILO、BLS 人口结构报告;McKinsey 劳动力市场研究 |
| 基础模型降低机器人 AI 编程成本 | 驱动 | 顺风 | 高 | 现在–3 年 | BCG physical AI 分析;微调成本从 $250K 降至接近零 |
| 机器人硬件成本下降(协作机器人、机械臂) | 驱动 | 顺风 | 中 | 现在–5 年 | IFR World Robotics Report 2025 |
| 开源 π₀ 模型让基础智能商品化 | 约束 | 逆风 | 中 | 现在–3 年 | PI 自己发布开源版本;竞争对手可基于 openpi 开发 |
| 商业部署的安全与可靠性要求 | 约束 | 逆风 | 高 | 现在–5 年 | ISO 10218 机器人安全标准;企业采购要求 |
| 模型训练与推理的算力成本 | 约束 | 逆风 | 中 | 现在–3 年 | GPU 成本走势;规模化推理的经济性尚未验证 |
| 资金充足的竞争对手(Skild AI $14B、Google DeepMind) | 约束 | 逆风 | 高 | 现在–3 年 | CB Insights 市场图谱;Skild 以 $14B 估值融资 $1.4B |
| 强劲资本流入带动生态势能 | 驱动 | 顺风 | 中 | 现在–3 年 | NVCA 机器人投资数据;2024–2025 年机器人 AI 投入 $5B+ |
2.5 展项
03竞争格局
3.1 竞争框架与市场结构
机器人基础模型市场可分为三类竞争原型。第一类是纯软件 / 智能层玩家,不做硬件,只做通用机器人“大脑”:Physical Intelligence 和 Skild AI 是主要例子。第二类是全栈机器人集成商,同时做硬件和专有 AI:Figure AI(人形机器人手臂 + Helix VLA)、1X Technologies(人形机器人 + 世界模型)和 Agility Robotics(Digit + AI)都属此类。第三类是拥有相邻算力和 AI 资产、正向机器人延伸的科技巨头:Google DeepMind(Gemini Robotics 1.5)、Amazon Robotics(Sequoia 机器人驱动系统 + Sparrow)和 NVIDIA(Cosmos 世界基础模型)。Physical Intelligence 在软件智能层最直接对标 Skild AI,在模型能力和研究可信度上对标 Google DeepMind。全栈集成商会争夺企业预算,但也可能成为 PI 模型的客户。 [CP001, CP002, CP003, CP004]
用 0–10 序数评分,将关键机器人基础模型竞争对手放在累计融资(x)与商业收入牵引(y)的双轴地图上。
[CP031, CP032]3.2 主要竞争对手画像
Skild AI 成立于 2023 年,总部位于匹兹堡;2026 年初以 $14 billion 估值完成 $1.4 billion Series C,由 SoftBank 和 NVIDIA 领投,Samsung 与 Salesforce Ventures 参投。其 Skild Brain 模型声称用分层架构实现跨形态泛化,并能做到零样本任务迁移。更关键的是,Skild 在 2025 年商业发布后数月内报告 $30 million 收入,证明了 Physical Intelligence 尚未拿到的商业化牵引。Google DeepMind 的 Gemini Robotics 1.5 于 2026 年初推出,是可通过云访问的 VLA 模型,具备先进的智能体规划和多步推理;背后有 Google 算力基础设施、海量多模态数据集,以及 Gemini API 分发。Figure AI 此前以 $39B+ 估值融资 $675M,打造自有人形机器人(Figure 02)和 Helix VLA 模型,并已部署到 BMW 制造工厂;这种全栈路径形成硬件护城河,但限制纯软件分发。Covariant(2024 年被 Amazon 收购)聚焦仓储 / 物流操作,拥有最深的真实世界运营数据。挪威人形机器人公司 1X Technologies 正在开发用于机器人认知的世界模型,目标融资最高 $1B,估值约 $10B。各竞争对手优势不同;Physical Intelligence 必须依靠灵巧操作深度、生态开放度和研究级模型质量来区分自己。 [CP005, CP006, CP007, CP008, CP009, CP010]
| 竞争对手 | 类型 | 成立时间 | 融资(累计) | 估值 | 收入(2025) | 核心模型 | 硬件 |
|---|---|---|---|---|---|---|---|
| Physical Intelligence | 纯软件 / 智能层 | Mar 2024 | ~$1.07B | $5.6B(Nov 2025) | $0(商业化前) | π₀ / π₀-FAST / π₀.5(VLA、流匹配、PaliGemma 主干) | 硬件无关(无自研硬件) |
| Skild AI | 纯软件 / 智能层 | 2023 | ~$1.7B(含 $1.4B C 轮) | $14B(2026 年初) | ~$30M ARR(2025) | Skild Brain(全形态分层 VLA) | 硬件无关 |
| Figure AI | 全栈(硬件 + AI) | 2022 | $675M+ | $39B+(2025) | 未披露(BMW 部署) | Helix VLA(多机器人、语言驱动) | Figure 02 人形机器人(自研) |
| Google DeepMind | 科技巨头 / 云 | 2010(DeepMind);机器人方向 2022 扩展 | N/A(Alphabet 子公司) | N/A(上市公司) | N/A(补贴型研究) | Gemini Robotics 1.5(VLA + Gemini Robotics-ER)模型 | 硬件无关(API 接入) |
| Covariant / Amazon | 全栈 / 已收购 | 2017(Covariant);2024 由 Amazon 收购 | >$250M(Covariant 收购前) | N/A(Amazon 收购) | 未披露(Amazon) | Covariant Brain(聚焦物流的 VLA) | 仓库机械臂(Amazon) |
| 1X Technologies | 全栈(人形机器人 + AI) | 2014(挪威) | 累计融资约 $400M | ~$10B(目标,2026) | 未披露 | 1XWM World Model(视频预测规划) | NEO 人形机器人(自研) |
估值数据基于最新披露轮次;可能不同于二级市场价格。收入数据来自披露或媒体报道;多数竞争对手不公开报告 ARR。
[CP001, CP005, CP006, CP007, CP008, CP009]| 能力 | Physical Intelligence(π₀) | Skild AI | Google DeepMind(Gemini Robotics) | Figure AI(Helix) | Covariant |
|---|---|---|---|---|---|
| 跨具身形态泛化 | 强(已演示 10+ 个机器人平台) | 强(全形态、分层架构) | 强(API 跨机器人接入) | 中等(聚焦 Figure 02;限于人形机器人) | 有限(仅仓库机械臂) |
| 灵巧操作深度 | 很强(叠衣、装配、长程任务) | 强(仓储和服务任务) | 强(多步骤家庭与工业任务) | 强(BMW 工厂部署) | 很强(仓库抓取放置已规模化验证) |
| 开源可用性 | 是——openpi 权重 + 代码 | 否——自研模型 | 部分——开发者 API,无权重 | 否——自研模型 | 否——自研 |
| 硬件独立性 | 是——无需自研硬件 | 是——任意机器人 | 是——可通过 API 接入 | 否——需要 Figure 02 硬件 | 否——Amazon 仓库硬件 |
| 语言指令跟随 | 强(PaliGemma VLM 主干) | 强 | 很强(Gemini 推理链) | 很强(语言驱动技能可即时学习) | 中等(任务专用) |
| 商业收入(2025) | 无(商业化前) | ~$30M ARR | N/A(Alphabet 内部) | 未披露(BMW 活跃部署) | 未披露(Amazon 集成) |
| 安全认证状态 | 尚未获得商业部署认证 | 尚未认证(已披露) | 不适用(研究 API) | 未公开认证 | Amazon 仓库运营中(内部标准) |
| 训练数据规模 | Open X-Embodiment + 自研内部数据 | 互联网视频 + 大规模仿真 + 自研数据 | Google 规模多模态数据 + 仿真 | BMW 工厂自研数据 | Amazon 仓库运营数据(10M+ 次抓取) |
能力评分矩阵,从五个维度比较 Physical Intelligence 与头部机器人 AI 竞争对手(0–10 序数量表,越高越强)。
分数为分析师估计;0-10 序数量表。
[CP033, CP034]3.3 Physical Intelligence 护城河与竞争风险
Physical Intelligence 的竞争优势包括:(1)创始人履历——Sergey Levine(RAIL Lab,Berkeley)、Chelsea Finn(MAML,Stanford)和 Karol Hausman(RT-2,DeepMind)的组合,可能是全球最具可信度的机器人 AI 创始团队;(2)灵巧操作深度——π₀ 展示了业内领先的长时域灵巧任务(洗衣物折叠、装配),多数竞争对手在基准中尚未匹配;(3)开源生态——openpi 带来开发者社区和训练数据贡献。该护城河的关键风险包括:(1)开源会让基础模型商品化——Skild、Google 等公司能直接对标 PI 已发布工作;(2)Skild AI 估值 $14B,融资额约为 PI 的 2.5×,且已经实现商业收入;(3)Google DeepMind 的算力和数据高出数个数量级,Gemini Robotics 1.5 已通过 API 向任何开发者开放;(4)PI 仍无收入——如果商业发布拖延,资金充足的竞争对手会先建立企业关系;(5)全栈集成商(Figure AI、1X)在特定部署中可能拥有更好的硬件-软件协同。当前护城河仍处研究阶段,尚未被商业验证。Physical Intelligence 必须在 Skild AI 建立企业锁定、Google DeepMind 完成 Gemini Robotics 商业 API 推出之前,优先推进商业发布、硬件伙伴协议和安全认证。PI 在已商业验证的企业账户上建立可防守位置的窗口,大约只有未来 12–18 个月。过了这个窗口,护城河可能主要停留在学术层面,而不是商业验证。该市场的竞争终局,会偏向最先证明企业级可靠性并建立经常性收入基础的玩家。[CP014, CP015, CP016, CP017, CP018, CP019]
| 公司 | 定价模式 | 已知定价 | 接入模式 | 企业支持 |
|---|---|---|---|---|
| Physical Intelligence | 计划按机器人 SaaS 授权 + 企业微调栈 | 未披露(商业化前);估计 $5–15K/robot/yr | 企业试点(仅邀请);openpi 供研究使用 | 无公开 SLA;仅提供试点支持 |
| Skild AI | 按机器人部署 SaaS / 软件授权 | 未公开披露;据 $30M ARR 和机队规模估计 >$10K/robot/yr | 企业直销;不开源 | 已披露企业支持合同 |
| Google DeepMind Gemini Robotics | API 接入(可能按用量或研究许可) | 未公开定价(研究接入) | 通过 Google Cloud API;研究合作 | Google Cloud 企业支持 |
| Figure AI | 机器人硬件租赁或销售 + 软件费 | 未披露;整体可能每台机器人 $100K–$300K | 企业直签合同(如 BMW) | 完整部署与支持服务 |
| Covariant / Amazon | 内部;Amazon 仓库集成 | N/A(内部定价) | Amazon 内部 | N/A |
没有竞争对手公开披露完整定价。估算来自 ARR 披露与机队规模假设。企业定价结构高度定制。
[CP026, CP027, CP028]| 护城河 / 风险因素 | Physical Intelligence 强度 | 主要威胁 | 概率 | 时间窗口 |
|---|---|---|---|---|
| 创始人研究履历 | 很高——Levine、Finn、Hausman 都是全球前 5 的机器人 AI 研究者 | 人才被挖;学术事务分散创始人精力 | 低 | 持续 |
| 灵巧操作技术深度 | 高——π₀ 已展示同类最佳的长程操作能力 | Google DeepMind Gemini Robotics 在规模和推理上反超 | 中 | 12–24 个月 |
| 开源生态(openpi) | 中——开发者社区正在成形;训练数据贡献增加 | 基础模型商品化;竞争对手用 PI 成果追赶 | 高 | 现在–12 个月 |
| 硬件无关架构 | 高——无需锁定即可与任意机器人合作 | Figure AI 等全栈对手形成 PI 难以复制的硬件护城河 | 中 | 2–4 年 |
| 机器人基础模型开源先发优势 | 中——品牌已建立;开发者有好感 | Skild、Figure、Google 都在转向类似开放路径 | 中 | 12–18 个月 |
| 相比 Skild AI 的商业收入先发 | 弱——PI 尚无收入;Skild AI 已有 ~$30M ARR | 如果 PI 推迟商业发布,Skild 会先建立企业客户关系 | 高 | 现在–18 个月 |
Physical Intelligence 在关键护城河维度上的竞争就绪度评分(0–10 序数,越高越强)。
[CP035, CP036]3.4 展项
04财务
4.1 收入模式与当前收入状态
截至 2026 年 Q1,Physical Intelligence 没有披露商业收入。公司正在制造和物流领域推进早期企业试点,但没有公开合同、ARR 或商业条款。计划中的收入模式是按机器人授权的 SaaS 结构:企业客户按每台部署并使用 PI 基础模型的机器人支付经常性年费。这相当于把企业软件里的按席位或按单元 SaaS 模式,用到物理机器人上。公司未公开定价。行业类比显示,每台机器人年费可能在 $5,000 到 $15,000 之间,这意味着若要做到 $50M ARR,需要约 3,300–10,000 台活跃机器人部署。额外收入来源可能包括模型微调费(一次性或经常性)、企业部署支持,以及潜在的机器人训练数据集授权。这些都尚未公布。openpi 开源释放似乎不产生直接收入,更像社区和生态投资。公司的无收入状态带来二元风险:要么未来 12–18 个月内完成商业发布并把企业试点管线转化;要么在竞争对手建立收入牵引后,融资会越来越难。 [CI001, CI002, CI003, CI004, CI005]
| 收入来源 | 状态 | 定价模式 | 估计 ACV / 单价 | 证据依据 | 置信度 |
|---|---|---|---|---|---|
| 按机器人 SaaS 授权(企业) | 计划中;尚未商业化 | 按部署机器人收取年度订阅费 | 每台机器人每年 $5,000–$15,000(估计) | 行业可比来自规模化 LLM API 定价;PI 未披露 | 低 |
| 企业模型微调(一次性或经常性) | 计划中;尚未商业化 | 按机器人类型 / 任务集群收取一次性费用 | 每次项目 $50,000–$300,000(估计) | 软件专业服务基准;PI 未披露 | 低 |
| 企业部署支持与 SLA | 计划中;尚未商业化 | 年度支持合同(按授权比例或固定费用) | 基础授权 ACV 的 15–20% | 标准企业 SaaS 合同结构 | 低 |
| 开源(openpi) | 活跃;无直接变现 | 免费增值 / 社区;无收入 | $0 | GitHub 公开仓库;未宣布变现 | 高 |
| 数据授权或机器人训练数据集访问 | 未宣布 | Unknown | Unknown | 推测;行业可比来自模型训练数据市场 | 低 |
| 维度 | Physical Intelligence(计划) | 可比对象(Skild AI) | 可比对象(企业 SaaS 中位数) | 评估 |
|---|---|---|---|---|
| 单位定价 | $5K–$15K/robot/yr(估计,未披露) | $10K+/robot/yr(媒体估计) | 每席位或每单位 $5K–$50K,差异很大 | 符合企业 AI 软件区间;ROI 得到验证后可站住 |
| 毛利率目标 | 70–85%(SaaS 软件目标) | 未披露 | AI SaaS 中位数 70–80% | 如果模型训练成本可在规模化后摊薄,则可达成;若单次推理算力高,则有风险 |
| 销售周期 | 12–24 个月(按企业机器人常态估计) | 未披露 | 企业 SaaS 为 6–18 个月 | 由于硬件集成和安全认证要求,周期长于纯软件 SaaS |
| 合同结构 | 多年期企业协议(估计) | 未披露 | 年度或多年期,预付款 | 标准做法;PI 必须加入按使用量计费的组成部分,才能吃到机队扩张带来的上行空间 |
| 流失风险 | 未知(尚无客户基础) | 未披露 | 企业 SaaS 年流失率 5–15% | 机器人 AI 因再训练数据锁定和切换成本,可能把流失率压得很低 |
从收入前研究阶段,经商业发布里程碑,到目标经常性收入模型的逻辑流程。
[CI001, CI002, CI021]Physical Intelligence 与可比公司在融资额(x)与年度收入(y)上的定位,显示 PI 收入前阶段的资本强度。
[CI024, CI025]4.2 成本结构、资本强度与烧钱速度
Physical Intelligence 的成本结构主要由研究和工程人员(2026 年 Q1 估计 150–250 人)、模型训练与推理 GPU 算力,以及旧金山高端办公地点的管理开销构成。若平均全包薪酬为每人每年 $300,000–$400,000,按员工数不同,人员烧钱速度约为每年 $45–$100 million。用多形态数据训练大型机器人 VLA 模型,算力成本显著;按 π₀ 规模模型的行业基准,每个主要版本训练一次可能花费 $5–20 million,规模化后还会持续产生推理成本。全年总烧钱估计为 $70–$150 million,意味着如果公司不大幅扩张算力或人员,仅 $600M Series B 就能提供约 4–8 年跑道。不过,若要做到商业规模,公司需要显著扩充人员和算力,满负荷商业建设下跑道可能压缩到 24–36 个月。Physical Intelligence 未披露实际烧钱速度或账上现金。公司没有公开披露债务融资,也未通过监管文件报告经营亏损;它仍是私营公司,除 Form D 豁免发行外没有 SEC 报告义务。此外,用于自定义机器人实验室基础设施、安全测试设备,以及试点项目潜在制造伙伴关系的资本开支,也会推高成本基数。这类阶段和画像的公司,通常把 60–70% 烧钱投向人员、20–25% 投向算力、10–15% 投向设施和其他管理开销。这些比例与可比 AI 研究初创公司一致,也说明如果员工数接近 250 人、算力支出随下一代模型开发放大,Physical Intelligence 的全年运营成本现实上可能达到 $120 million 或更高。[CI006, CI007, CI008, CI009, CI010, CI011]
| 指标 | 估计值 | 依据 / 假设 | 置信度 |
|---|---|---|---|
| 单台机器人年许可费(ACV) | $5K–$15K | 行业类比;PI 未披露 | 低 |
| 毛利率(SaaS 目标) | 70–85% | AI SaaS 基准;假设模型训练成本摊销 | 中 |
| CAC(获客成本,估计) | 每个企业账户 $200K–$500K | 企业机器人销售团队成本;假设销售周期 12–24 个月 | 低 |
| CAC 回本周期(估计) | 每账户 ACV 为 $50K–$100K 时 3–7 年 | 基于 ACV 区间和 CAC 估计;相对典型 SaaS 很长 | 低 |
| LTV / CAC 比率(预测) | 3–8×(若留存 5+ 年) | 假设机器人基础设施嵌入后流失率较低;不确定性很高 | 低 |
| 烧钱速度(估计年化) | 每年 $70M–$150M | 人员(150–250 人,全包 $300K–$400K)+ 算力 + 间接成本 | 低 |
| 隐含现金跑道(来自 $600M Series B) | 按当前烧钱速度 4–8 年(商业化规模化前) | 简单除法;实际现金跑道会被规模化支出压缩 | 低 |
所有单位经济模型均为基于行业基准和类比推导的估计;Physical Intelligence 尚未披露任何财务指标。实际经济性可能存在重大差异。
[CI007, CI008, CI009, CI010]| 财务指标 | 状态 | 不可得原因 | 所需尽调动作 |
|---|---|---|---|
| 收入 / ARR | 未披露;估计为 $0 | 私人公司;商业化前;无 SEC 报告义务 | 在 VDD 中索取;通过管线渠道核查 |
| 毛利率 | 未披露 | 尚无商业化运营;毛利率仅为理论值 | 用可比 SaaS 毛利率和单次推理成本基准建模单位经济性 |
| 经营亏损 / EBITDA | 未披露 | 私人公司;无公开财务报表 | 索取管理账;根据员工数和算力成本基准估计烧钱速度 |
| 账上现金 / 烧钱速度 | 未披露 | 私人公司;无披露义务 | 索取经审计财务;通过数据室中的投资方材料核验 |
| 客户数 / ACV | 未披露(商业客户为零) | 商业化前阶段 | 跟踪试点管线;索取试点条款和转化时间表 |
| 按职能划分的员工数 | 未披露 | 私人公司;无公开申报 | LinkedIn 员工数代理指标;在 VDD 中索取组织架构图 |
Physical Intelligence 年度烧钱速度的低 / 高估计,以及 $600M B 轮带来的隐含现金跑道。
所有估计均为分析师推断;Physical Intelligence 未披露财务数据。
[CI007, CI008, CI022]4.3 资本充足性、融资路径与财务风险
Physical Intelligence 三轮累计融资约 $1.07 billion,估计年烧钱 $70–$150 million;按商业扩张节奏不同,估计跑道约 24–60 个月。据报道,截至 2026 年 4 月,公司正就 $11 billion 估值的下一轮融资进行深入谈判;若以 $500M+ 完成,可再延长 3–6 年跑道。关键财务风险包括:(1)公司必须做到至少 $50M ARR,才能在下一轮融资前支撑投资人信心;(2)如果试点商业转化停滞,估值上调会难以自圆其说——当前 $5.6B 估值对应无限收入倍数;(3)连续训练下一代模型的算力成本可能比预期更快压缩跑道;(4)市场下行或投资人情绪转向,可能关闭不为零收入买单的 AI 高估值融资窗口。公司没有披露客户集中度风险,因为还没有商业客户。财务尽调需要的来源类型包括官方公司沟通和监管申报;唯一 SEC Form D 是 Series B 豁免发行文件,除发行规模外没有披露财务细节。 [CI012, CI013, CI014, CI015, CI016, CI017]
| 轮次 | 金额 | 估值(投后) | 日期 | 领投方 | 隐含资金用途 |
|---|---|---|---|---|---|
| 种子轮 | $70M | 未披露 | Mid-2024 | Lux Capital | 初始团队招聘、采购算力、模型开发 |
| Series A 轮 | $400M | $2.4B | Nov 2024 | OpenAI、Thrive Capital、Lux Capital 等投资方 | 大规模模型训练、扩员、实验室基础设施 |
| Series B 轮 | $600M | $5.6B | Nov 2025 | CapitalG(Alphabet) | 商业试点扩张、企业 GTM 建设、下一代模型训练 |
| 累计融资 | ~$1.07B | N/A(累计) | Nov 2025 | 多家投资方 | 同上;综合现金跑道估计为 24–60 个月,取决于烧钱曲线 |
| 据报道的下一轮(未确认) | 未知(可能为 $500M–$1B) | ~$11B(据报道) | Q2 2026(未确认) | 未披露 | 若确认,用于进一步商业化规模化和算力扩张 |
估值均为各轮投后。公司未披露债务融资、可转债或基于收入的融资。Series B 是最近一次已确认交割。
[CI014, CI015, CI016, CI017]Physical Intelligence 从种子轮、A 轮到 B 轮的累计融资额,并标出未经确认的下一轮。
数值单位为百万美元。下一轮金额具投机性;基于媒体来源标注为「未确认」。
[CI014, CI023]4.4 展项
05产品与技术
5.1 产品定义与架构
Physical Intelligence 的核心产品是 π₀ 系列机器人基础模型。不同于此前只针对单一机器人类型和单一任务训练的机器人 AI 系统,π₀ 被设计成跨形态模型:同一套模型权重横跨 68 种不同机器人硬件配置使用,从单臂机械臂到双手机械灵巧手都有覆盖。跨形态训练带来两个实际好处。第一,模型泛化更广,能把一个机器人上学到的技能,用最少微调迁移到新机器人。第二,公司能聚合来自多类硬件伙伴的训练数据,而不是被单一机器人设计的数据限制。架构上,π₀ 将预训练 PaliGemma 3B 视觉-语言模型(VLM)与 300 million 参数动作专家 transformer 融合。VLM 组件负责理解语言指令和解析视觉场景。动作专家用 flow matching 生成连续机器人控制信号(根据机器人不同,可为关节位置、速度或笛卡尔目标);这种生成式建模技术在精度和推理速度上都优于基于 diffusion 的动作策略。π₀.5 变体在 π₀ 基础上加入更强的互联网规模预训练,拓宽语义理解。π₀-FAST 用更快的单次动作解码器解决时间敏感任务的推理延迟。openpi 开源包则为外部研究社区提供基于 π₀ 权重的微调工具。 [CE001, CE002, CE003, CE004, CE005, CE006]
| 产品 / 模型 | 版本 / 发布 | 访问模式 | 关键能力 | 目标用例 | 技术基础 | 状态 |
|---|---|---|---|---|---|---|
| π₀ (pi-zero) | v1.0 (Sep 2024) | 研究 / 企业试点 | 跨本体 VLA;灵巧操作;长程任务 | 制造、物流、通用任务自动化 | PaliGemma 3B VLM + 300M 动作专家;流匹配 | 已发布(研究和试点) |
| π₀.5 | v0.5 (2025) | 研究 / 试点 | 增强互联网规模预训练;更广的语义锚定 | 覆盖更多任务类型;指令跟随更好 | π₀ 基座 + 扩展互联网预训练语料 | 研究预览 |
| π₀-FAST | 2025 | 研究 | 更快的单次动作解码器;降低推理延迟 | 时间敏感工业任务;更高频控制回路 | π₀ 基座 + 优化解码架构 | 研究预览 |
| openpi | 开源(Feb 2025) | 公开(Apache 2.0 或类似许可) | 微调工具;社区可访问 π₀ 权重 | 学术研究;机器人社区开发 | Python + JAX/PyTorch;封装 π₀ 权重 | 活跃开源 |
| 用例 | 垂直行业 | 所需机器人类型 | 任务复杂度 | 商业化成熟度 | 证据 |
|---|---|---|---|---|---|
| 叠衣与衣物处理 | 消费者服务 / 酒店 | 双臂灵巧机械臂 | 高(接触密集;可变形物体) | 试点 / 演示阶段 | π₀ arXiv 论文基准;演示视频 |
| 餐具装载与厨房自动化 | 餐饮服务 / 酒店 | 单臂或双臂机器人 | 高(杂乱场景;物体多样) | 试点 / 演示阶段 | π₀ 基准数据;已发布演示 |
| 包裹分拣与物流搬运 | 物流 / 电商仓储 | 单臂或移动机械臂 | 中(物体形状多样;箱拣) | 企业试点(进行中) | 媒体报道的制造 / 物流试点 |
| 装配与制造自动化 | 工业制造 | 多轴工业机械臂 | 中高(精确定位;QA) | 企业试点(进行中) | 媒体和合作伙伴报道;客户未披露 |
| 通用操作(跨任务) | 多个垂直行业 | 硬件无关(API 访问) | 可变 | 研究 / 早期访问 | π₀ 跨本体训练结果 |
| 层级 | 组件 | 技术 / 框架 | 作用 | 开源或自研 |
|---|---|---|---|---|
| 感知 | 视觉-语言骨干 | PaliGemma 3B(Google DeepMind) | 场景理解;指令解析;视觉锚定 | 开放权重(依赖 Google) |
| 动作生成 | 动作专家 Transformer(300M 参数) | 自研架构;流匹配 | 生成连续机器人控制信号 | 自研(由 PI 训练) |
| 训练框架 | 模型训练基础设施 | TPU/GPU 上的 JAX / PyTorch | 大规模跨本体模型训练 | 开源框架;自研训练管线 |
| 数据管线 | 机器人演示数据集 | 多本体机器人遥操作数据(规模未披露) | 基础模型预训练;跨本体泛化 | 自研(数据集细节未披露) |
| 部署 | 机器人端推理 | ONNX 导出或同类方案;边缘硬件(未披露) | 机器人硬件上的实时控制回路 | 自研(未披露) |
| 开源工具 | openpi | Python;微调工具;公开 π₀ 权重 | 社区访问;研究微调 | 开源(Apache 2.0 或类似许可) |
| 里程碑 | 估计时间 | 状态 | 战略意义 | 证据依据 |
|---|---|---|---|---|
| π₀ 发布(初始 VLA 研究版本) | Sep 2024 | 已完成 | 建立了跨本体基础模型概念验证 | arXiv 预印本和官网公告 |
| openpi 开源发布 | Feb 2025 | 已完成 | 开发者社区参与;外部研究采用信号 | GitHub 仓库公开;PI 博客文章 |
| π₀.5 和 π₀-FAST 变体 | 2025 | 已完成(研究预览) | 性能和延迟改善;扩展用例范围 | PI 博客和外部引用 |
| 企业商业试点 | H2 2024 – 2026 | 进行中(客户未披露) | 通往首笔商业收入的关键路径 | 媒体报道;PI 博客 |
| 商业产品发布(SaaS 定价) | 估计 2026–2027 | 尚未宣布 | 开始产生 ARR 的必要条件;下一轮融资的门槛事件 | 分析师推断;无官方公告 |
| 生产环境安全认证 | 估计 2026–2028 | 未启动(无公开披露) | 多数企业制造和物流部署所必需 | 行业基准;PI 未披露 |
| 下一代模型(π₁ 或同等版本) | 估计 2026–2027 | 未宣布 | 要在能力曲线上领先 Skild AI 和 Google DeepMind,此项必不可少 | 研发管线推断 |
π₀ VLA 模型的端到端架构:从语言指令输入和摄像头视觉,到连续机器人控制输出。
[CE001, CE002]Physical Intelligence 产品在能力成熟度(x 轴)和商业就绪度(y 轴)上的定位。
[CE020, CE022]5.2 技术表现与基准
Physical Intelligence 在 arXiv 预印本中发布基准结果,显示 π₀ 在 LIBERO 基准套件上超过此前最强水平。LIBERO 测试洗衣物折叠、物品整理、组件装配等灵巧操作任务。具体来看,当只用有限演示数据微调时,π₀ 的任务完成率显著高于 OpenVLA、RT-2 和 Octo;在企业部署中,收集大量机器人专属数据成本高昂,因此该指标非常关键。模型在长时域灵巧操作任务上尤其强,这类任务被普遍视为机器人 AI 最难解决的问题:需要在数秒内连续完成接触丰富的动作,例如折叠衣物或装载洗碗机。Physical Intelligence 报告称,π₀ 完成这些任务的成功率能与此前最佳系统竞争,同时数据效率明显更高。不过,已发布基准来自受控实验室环境,不能直接证明其在生产企业环境中的表现;真实场景有非结构化工作空间、光照变化和实时可靠性要求。实验室基准成功率与企业生产可靠性之间的差距,是一个关键未解技术风险。机器人行业经验显示,生产部署成功率通常比实验室结果低 20–40 个百分点。 [CE007, CE008, CE009, CE010, CE011]
LIBERO 基准上,π₀ 相对基线 VLA 模型的任务成功率估计;数值越高,代表灵巧操作任务表现越好。
相对性能分数是基于已发布 LIBERO 基准摘录的分析师估计;具体数值会随评估协议变化。
[CE007, CE008, CE009]5.3 差异化、知识产权与信任姿态
Physical Intelligence 的主要技术差异化,在于 VLA 架构:它把预训练 VLM 推理与 flow-matching 动作专家结合,并在截至目前公开报道中规模最大的跨形态机器人数据集上训练。创始团队的学术资历极强:Sergey Levine(UC Berkeley RAIL Lab,model-agnostic meta-learning MAML 相邻方法发明者)、Chelsea Finn(MAML 共同发明人,Stanford AI Lab)和 Karol Hausman(Google DeepMind Robotic Learning)。这支团队发表过被引用数千次的基础论文。openpi 发布后,GitHub stars 和外部研究引用都显示开发者社区已经采用。关键 IP 风险包括:(1)核心架构依赖 PaliGemma(Google DeepMind 模型),形成对 Google 授权条款的依赖;(2)机器人演示视频训练数据的来源和版权状态没有公开披露;(3)arXiv 预印本详细披露技术路径,降低了资源充足的竞争对手复制门槛。安全与信任方面,Physical Intelligence 尚未披露 π₀ 是否接受过第三方安全认证、功能安全审计(ISO 13849 / IEC 62061)或网络安全渗透测试。企业部署时,制造和物流客户会要求成文的安全论证,并可能在 CE/OSHA 监管环境中要求监管许可。openpi 发布内容除技术能力限制外,没有记录明确的安全约束或部署护栏;与企业客户生产部署要求相比,这是缺口。 [CE012, CE013, CE014, CE015, CE016, CE017]
| 信任维度 | 当前状态 | 缺口或担忧 | 企业部署要求 |
|---|---|---|---|
| 功能安全认证(ISO 13849 / IEC 62061) | 未公开披露 | 未宣布第三方安全审计;制造业部署的关键缺口 | 多数企业制造客户要求;典型周期 12–24 个月 |
| 网络安全 / 对抗鲁棒性 | 未披露 | 机器人 AI 模型容易受到对抗输入,以及通过语言命令发起的提示注入影响 | 需要渗透测试和威胁建模;未披露 |
| 训练数据来源与版权 | 未披露 | 机器人演示视频授权和来源不清;存在潜在 IP 责任 | 企业法务放行需要;VDD 应说明 |
| PaliGemma 依赖许可 | 开放权重(Google 条款) | Google 可能更改许可条款;PI 的架构依赖外部方 | 若合同保护则可接受;若 Google 限制商业使用则有风险 |
| EU AI Act 合规 | 未公开评估 | 物理环境中的机器人 AI 系统可能被 EU AI Act 归为高风险 | 2026–2027 年预期欧盟商业部署所必需 |
| 人机协作中的运营安全 | 研究中已覆盖(仅实验室条件) | 非结构化真实环境中的表现尚未大规模验证 | 企业部署关键要求;需要安全测试和保险 |
Physical Intelligence 产品交付的关键外部依赖,显示风险集中在 PaliGemma(Google)和 GPU 算力提供方。
[CE012, CE013, CE014]5.4 展项
06客户
6.1 客户细分与当前状态
截至 2026 年 Q1,Physical Intelligence 仍处于完全无收入、无客户阶段。公司公开表示正在与未具名制造和物流客户推进企业试点,但没有披露商业合同、客户数、ARR 或可引用客户。基于公开信息,预期企业客户可分为:拥有高密度机器人部署的大型制造商(汽车、电子、消费品装配)、需要灵活拣选与包装自动化的物流和电商仓储运营商、需要用灵巧操作完成备餐和厨房任务的餐饮服务与酒店运营商,以及可能服务非结构化环境操作的政府或国防客户。公司未公布 SMB 或中端市场商业化动作;所有已显示的试点都涉及大型企业账户,这些账户有预算和运营规模,能部署数十到数百台机器人。机器人 OEM 伙伴网络(包括训练期间贡献形态数据的硬件公司)可能成为间接客户或转售渠道,让 PI 软件在伙伴硬件之上授权销售,形成与企业直销并行的 B2B2B 分发路径。公司没有 SMB 或自助服务渠道,没有产品驱动增长动作,也未披露从 openpi 开源释放转化到企业合同的免费增值路径。现阶段,整个商业路径都是企业直销和 OEM 渠道销售。[CU001, CU002, CU003, CU004, CU005]
| 细分市场 | 描述 | 预期机器人机队规模 | 估计 ACV 潜力 | 试点活动证据 | 优先级 |
|---|---|---|---|---|---|
| 大规模制造(汽车、电子) | 装配线自动化;机器人密度高;精度要求高 | 每个设施 100–10,000 台机器人 | 机队完全渗透时每账户 $500K–$150M | 媒体报道的试点(制造场景) | 高 |
| 物流与电商仓储 | 柔性拣选与包装;箱拣;分拣;混合 SKU 处理 | 每个设施 50–5,000 台机器人 | 机队完全渗透时每账户 $250K–$75M | 媒体报道的试点(物流场景) | 高 |
| 餐饮服务与酒店 | 厨房自动化;备餐;洗碗;灵巧任务 | 每个场地 5–100 台机器人;多个场地 | 每个账户 $25K–$1.5M | 演示视频(洗衣、洗碗机);尚无已确认商业试点 | 中 |
| 通用企业自动化(定制任务) | 跨行业;通过 openpi 或 API 做定制微调 | 差异很大 | 每个账户 $50K–$5M | openpi 社区活动;学术微调 | 低(近期);中(中期) |
| 机器人 OEM 合作伙伴(B2B2B 渠道) | 将 PI 软件与自家机器人打包的硬件伙伴 | N/A(通过 OEM 按机器人收取授权费) | 未披露;取决于 OEM 收入分成模式 | 合作方名单包括 AgiBot、Longcheer | 高(战略渠道) |
| 阶段 | 时期 | 客户数 | 年经常性收入(ARR) | 关键证据 |
|---|---|---|---|---|
| 研究 / 创立 | Q1–Q3 2024 | 0 | $0 | 公司 2024 年 3 月成立;处于研究阶段 |
| 种子轮与初始试点 | Q4 2024 | 0 个商业客户;2+ 个试点伙伴(未确认) | $0 | 2024 年 11 月完成 Series A 轮;试点项目启动 |
| 活跃企业试点 | 2025(全年) | 0 个商业客户;估计 3–8 个试点伙伴 | $0 | 媒体报道;2025 年 11 月完成 Series B 轮 |
| 目标商业化发布(估计) | 2026–2027 | 1–5 个商业账户(目标) | $500K–$5M(目标初始 ARR) | 分析师估计;PI 未公开披露 |
| 目标规模(估计) | 2027–2028 | 10–30 个企业账户(目标) | $10M–$50M ARR(目标) | 分析师基于可比机器人 SaaS 发展轨迹估计 |
所有前瞻性估计均为分析师推断;Physical Intelligence 尚未披露商业化路线图、试点数量或收入目标。
[CU001, CU002, CU009]Physical Intelligence 企业客户从初始认知,到试点、商业部署和机队扩张的旅程。
[CU001, CU002, CU003]6.2 试点活动与采用证据
2024–2025 年的媒体报道确认,Physical Intelligence 至少与两家具名公司开展企业试点:AgiBot(一家中国机器人制造商)和 Longcheer Technology(一家电子制造商)。二者既是早期跨形态数据贡献方,也是试点客户。这是公开来源中唯一能靠近客户关系的具名信息。openpi 开源社区提供了第二层采用信号:该仓库已积累数千个 GitHub stars,并吸引学术界和产业研究团队进行活跃外部微调。这个社区信号证明了开发者兴趣和技术路径可用性,但不等同于企业客户牵引,也不会产生收入。Physical Intelligence 发布过 π₀ 在受控环境中完成任务的演示视频;这些视频获得科技媒体(The Verge、IEEE Spectrum、VentureBeat)大量报道,说明潜在企业买家认知度较高。认知和兴趣并不会在缺少进一步商业验证时直接转化为购买意图。最关键的未知数,是企业试点转成已签商业合同的比例;这是任何潜在投资人尽调时必须验证的最重要财务指标。 [CU006, CU007, CU008, CU009, CU010]
| 客户 / 伙伴 | 关系类型 | 行业 | 证据质量 | 证据来源 | 已确认收入 |
|---|---|---|---|---|---|
| AgiBot | 硬件伙伴 + 早期试点客户 | 机器人制造(中国) | 低 — 仅媒体提及;无结果数据 | AI Market Watch 媒体报道 | 否($0;试点阶段) |
| Longcheer Technology | 硬件伙伴 + 早期试点客户 | 电子制造(中国) | 低 — 仅媒体提及;无结果数据 | AI Market Watch 媒体报道 | 否($0;试点阶段) |
| 未具名制造业客户 | 企业试点 | 工业制造(地域未知) | 极低 — 媒体仅泛泛提及;无细节 | 多家媒体提及 PI 企业试点 | 否($0;试点阶段) |
| 未具名物流客户 | 企业试点 | 物流 / 仓储(地域未知) | 极低 — 仅泛泛提及;无细节 | 多家媒体提及 PI 企业试点 | 否($0;试点阶段) |
Physical Intelligence 企业客户采用漏斗估计:从初始市场认知到商业化机队部署。
所有数值均为分析师估计;Physical Intelligence 未披露管线指标。
[CU009, CU010]从多个证据维度评估 Physical Intelligence 现有客户侧证据的质量。
[CU008]6.3 留存前景、集中度风险与扩张
Physical Intelligence 没有留存指标、流失数据、NRR 或合同续约数据,因为公司没有商业客户。因此,留存分析只能向前看,并基于计划中的按机器人 SaaS 模式的结构特征。留存在结构上预计较高,原因包括:(1)机器人 AI 模型会随时间累积客户专属微调数据,换供应商相当于损失数月专有任务训练数据;(2)企业机器人队列本身是重大资本投入,会让软件供应商更换变得迟滞;(3)与安全认证流程和生产排程系统集成,会进一步提高切换门槛。这些结构性因素说明,一旦 Physical Intelligence 完成商业部署,超过 90% 的毛收入留存是有可能的,但现阶段完全停留在理论层面。客户集中度风险未知;若商业发布发生,短期预计会偏高,因为公司可能先签下少数大型企业账户,造成显著头部客户集中。尽调应要求披露前五大客户收入集中度,以及试点协议中的任何最低采购承诺。扩张模式(先落地客户某一设施,再扩到更多设施或机器人类型)若客户生命周期超过 5 年、扩张倍数达到 2–5×,单位经济性会很有吸引力;但这些都尚未在运营上证明。 [CU011, CU012, CU013, CU014, CU015]
| 指标 | 当前值 | 依据 | 前瞻评估 |
|---|---|---|---|
| 商业客户数 | 0 | 未产生收入;已确认没有商业合同 | 商业化发布前不适用 |
| 净留存率(NRR) | 不适用 | 无商业客户;无续约数据 | 如果先落地再扩张模型在车队规模跑通,预计 >110% |
| 总留存率(GRR) | 不适用 | 无商业客户 | 受切换成本和数据锁定效应支撑,预计 >90% |
| 客户流失率 | 不适用 | 无商业客户 | 嵌入生产后,预计年流失率 <10% |
| 试点转商业转化率 | 未知(无数据) | 尚无试点转化;关键尽调缺口 | 企业机器人行业基准为 20–40% 的试点转化率 |
| 客户满意度(CSAT / NPS) | 暂无 | 无商业客户;试点反馈未公开 | 需要查看 VDD;无公开数据 |
| 维度 | 评估 | 风险等级 | 尽调动作 |
|---|---|---|---|
| 最大客户收入集中度 | 未知;预计偏高(首批商业账户可能贡献初始 ARR 的 80%+) | 高 | 要求在 VDD 中披露前 5 大客户收入集中度 |
| 地域集中度 | 据报道试点分布在亚洲(AgiBot、Longcheer)和美国(未具名);结构未知 | 中 | 厘清试点和商业管线的地域分布 |
| 垂直行业集中度 | 早期管线以制造和物流为主;消费或服务行业证据有限 | 中 | 评估垂直行业多元化计划;防范单一垂直行业依赖 |
| 渠道依赖(通过 OEM 做 B2B2B) | AgiBot 和 Longcheer 显示 OEM 渠道已启动;条款未披露 | 中 | 索取 OEM 合作协议;评估收入分成条款和排他性 |
| 单账户内先落地再扩张 | 初始客户内部的车队扩张是主要增长模型;目前尚无证据 | Unknown | 在 VDD 中跟踪从试点到全车队的扩张指标 |
| 单一供应商风险(依赖 Google PaliGemma) | Google 既是投资方(CapitalG),又是依赖项;利益协调更复杂 | 高 | 法务审查 PaliGemma 条款及与 Google 的任何附带安排 |
基于结构性切换成本假设,对 Physical Intelligence 企业客户留存曲线的预测;目前没有实际留存数据。
所有数值均为基于机器人 SaaS 结构性切换成本假设的理论预测;目前没有实际客户队列数据。
[CU011, CU013]6.4 展项
07风险
7.1 监管与法律风险
Physical Intelligence 最大的监管风险来自 EU AI Act。该法案将可能造成伤害、且在物理环境中运行的 AI 系统,按 Annex III 可能归为高风险。部署在制造或物流场景中的机器人 AI 系统,在进入欧盟商业部署前,可能需要合规评估、人类监督机制、技术文档和上市后监测。EU AI Act 合规预计会让欧盟商业化时间线增加 12–24 个月,并消耗大量法律和工程资源。在美国,OSHA 工作场所安全法规和 NIST AI Risk Management Framework 指南适用于工作环境中的 AI 系统,但当前美国监管要求不如欧盟法案具体,也不会阻止近期本土部署。制造环境中与人协作的机器人软件需要通过 ISO 13849 和 IEC 62061 功能安全认证;Physical Intelligence 未披露任何认证进展。训练数据 IP 责任是潜在法律风险:如果 Physical Intelligence 的机器人演示数据集包含来自受版权保护来源的视频内容(制造流程视频、专有机器人操作记录),数据来源方的 IP 主张可能限制公司的运营自由。截至 2026 年 Q1,没有已知针对 Physical Intelligence 的活跃诉讼、监管执法行动或重大法律纠纷。 [CR001, CR002, CR003, CR004, CR005]
| 风险 | 可能性 | 影响 | 时点 | 缓释状态 | 剩余暴露 |
|---|---|---|---|---|---|
| EU AI Act 将机器人 AI 归为高风险 | 高 | 高 | 2026–2027 | 未披露合规性评估;未宣布 EU 合规计划 | 高 |
| 尚未取得功能安全认证(ISO 13849 / IEC 62061) | 高(针对 EU/CE 制造业部署) | 高 | 2026–2027 | 未披露认证进展 | 高 |
| 训练数据 IP 责任(机器人演示视频版权) | 中 | 高 | 持续中(商业化前) | 未披露;无数据来源声明 | 中 |
| 美国制造环境中的 OSHA 工作场所 AI 风险 | 低–中 | 中 | 2026+ | 目前没有阻碍部署的美国监管壁垒 | 低 |
| 第三方针对 PI 架构发起专利侵权主张 | 低 | 高 | 持续中 | 未披露专利申请;arXiv 预印本限制新颖性保护 | 中 |
| Google 限制 PaliGemma 许可 | 低 | 致命 | 持续中 | 未披露合同保护;依赖 Google 的 Gemma Terms of Use | 高 |
7.2 运营、竞争与技术风险
最尖锐的运营风险,是 π₀ 从实验室走到生产现场的落差。模型在可控实验室里跑出亮眼基准表现,但制造和物流现场难得多:工位非结构化、光照变化、人机混行、实时可靠性要求高,还要 99%+ 可用性。机器人行业经验反复显示,从实验室到生产,性能通常会掉 20–40 个百分点。Physical Intelligence 如果不能在试点阶段补上这段落差,商业转化就会失败。竞争端的运营风险同样高:Skild AI 已做到约 $30M ARR,并在搭商业数据飞轮。Skild AI 每多一个季度积累生产部署数据,模型质量就会提升,竞争护城河也会拉宽。Google DeepMind 的 Gemini Robotics 1.5 是生存级竞争威胁:Google 几乎有无限算力,可直接使用 PaliGemma(Physical Intelligence 所依赖),又通过 Google Cloud 和 Google Workspace 拥有深厚企业关系,由此天然形成机器人 AI 分发渠道。Physical Intelligence 的开源策略(openpi)带来复制风险:只要竞争对手动机足够强、又有算力,就能拿公开的 arXiv 预印本和 openpi 代码库起步复现,把研究护城河从几年压缩到几个月。 [CR006, CR007, CR008, CR009, CR010, CR011]
| 风险 | 可能性 | 影响 | 时点 | 缓释状态 | 剩余暴露 |
|---|---|---|---|---|---|
| 非结构化环境中的实验室到生产性能落差 | 高 | 高 | 2026(商业化发布) | 试点测试持续推进;未披露公开生产指标 | 高 |
| Skild AI 数据飞轮 — 商业部署先发优势 | 高 | 高 | 立即(持续中) | 没有直接缓释;PI 必须加快商业化发布 | 高 |
| Google DeepMind Gemini Robotics 的算力与分发不对称 | 高 | 致命 | 2026–2027 | 没有直接缓释;PI 以跨具身形态和独立定位做差异化 | 高 |
| 通过 openpi 和 arXiv 预印本开源复现 | 中 | 高 | 2026–2027 | 缓释有限;arXiv 和 openpi 已公开;自有训练数据仍是护城河 | 中 |
| 机器人硬件故障或事故导致试点人员受伤 | 中 | 高 | 持续中(试点活跃) | 试点采用标准安全流程;未披露事故 | 中 |
| 生产环境中针对 π₀ 的网络安全 / 对抗输入攻击 | 低 | 高 | 2026+ | 未披露对抗鲁棒性测试 | 中 |
| 风险集群 | 必要缓释动作 | 终止标准(投资逻辑破裂) | 领先指标 |
|---|---|---|---|
| 收入前商业化失败 | 签署至少 3 份企业 LOI;到 2026 年 Q4 将 1 个试点转为付费合同 | 到 2026 年 Q4 仍无商业收入,且没有可信管线 | 试点项目停滞;主动接触企业 12 个月后仍未签署 LOI |
| Skild AI 数据飞轮 | 加快商业化发布;在 Skild 拉大差距前积累自有生产数据 | PI 尚无收入时,Skild AI 已超过 $100M ARR | Skild AI 季度 ARR 增速和客户数公告 |
| Google DeepMind Gemini Robotics | 以跨具身形态深度和独立生态做差异化(不绑定单一云) | Google 借助 GCP 分发能力,大规模推出可商用机器人 AI API | Google DeepMind 针对 Gemini Robotics 的商业 API 公告 |
| PaliGemma 许可限制 | 谈妥正式商业许可协议;准备替代 VLM 备份方案 | Google 宣布 PaliGemma 商业限制,过渡期 < 6 个月 | Gemma 使用条款变更;任何影响 VLM 授权的 Google 公告 |
| 训练数据 IP 责任 | 完成完整数据来源审计;必要时补签追溯授权 | 针对 PI 的 IP 侵权诉讼胜诉,并阻断训练数据使用 | DMCA 下架通知;任何可比数据集公司被提起 IP 诉讼 |
| 关键人物离职(Levine / Finn) | 建立继任计划;投保关键人物保险;加快补强领导团队深度 | Levine 或 Finn 在商业发布前宣布离职 | LinkedIn 动态;会议出席;在其他机构发表论文 |
二维风险热力图,把 Physical Intelligence 面临的全部重大风险按发生可能性(列)和影响程度(行)映射。
[CR001, CR006, CR007]有向无环图,展示主要根源风险如何传导为 Physical Intelligence 的下游业务影响。
[CR010, CR011, CR012]7.3 财务、依赖与关键人物风险
Physical Intelligence 当前估值下的财务风险极端:零收入、估计年烧钱 $70–$150M,据报道下一轮估值 $11B,而这轮要闭合,必须证明有实质商业牵引力。如果 $11B 轮没有落地、商业收入仍为零,公司要么大幅砍烧钱速度,要么接受降估值融资;两者都会带来实质股权稀释和人才留存风险。PaliGemma 对 Google DeepMind 的依赖同时是技术和财务风险:Google 既是投资方(通过 CapitalG),也是技术依赖(PaliGemma),还是直接竞争者(Gemini Robotics)。Google 一旦限制 PaliGemma 商业许可,或收购某个竞争对手,Physical Intelligence 的核心架构就会承压。关键人物风险高度集中:Sergey Levine 是技术路线的主要公开面孔;他离开会削弱公司在投资人、潜在合作伙伴和企业客户眼中的技术可信度。Chelsea Finn(MAML 共同发明人)和 Karol Hausman(CEO)也代表类似集中度。公司未披露任何关键人物保险、创始人股权归属时间表或留任安排。资本集中风险也明显:Lux Capital 参与了全部三轮,CapitalG(Alphabet)领投 Series B,意味着两家投资人在股权结构和董事会动态中拥有多数影响力。 [CR012, CR013, CR014, CR015, CR016, CR017]
| 依赖项 | 风险类型 | 中断可能性 | 影响 | 缓释 |
|---|---|---|---|---|
| Google DeepMind PaliGemma(VLM 骨干) | 架构 / 许可 | 中 | 致命 — 需要以 $10M+ 成本完整重训模型 | 未披露合同保护;持续监测 Google 许可条款 |
| NVIDIA / GPU 算力提供商 | 算力获取 / 成本 | 低 | 高 — 训练延迟;成本上升 | 可用云厂商不止一家;算力不是单一来源 |
| Lux Capital(种子轮领投方) | 资金提供方 / 董事会 | 低 | 中 — 如果 Lux 退出或减少支持,后续融资观感受损 | 股权结构表有多家投资方;CapitalG 提供战略背书 |
| CapitalG / Alphabet(Series B 轮领投方) | 资金提供方 / 战略冲突 | 低 | 高 — 如果 Google 限制 PaliGemma,或通过 Gemini Robotics 更激进竞争 | 跟踪 Google DeepMind 的竞争动作;董事会席位条款很重要 |
| 机器人 OEM 合作伙伴(AgiBot、Longcheer 等) | 分发渠道 / 数据 | 中 | 中 — 如果 OEM 转向竞品 AI(Skild AI、Google),PI 将失去渠道和数据 | 建立直接企业客户关系,降低 OEM 依赖 |
| OpenAI(Series A 轮战略投资方) | 资金提供方 / 竞争信号 | 低 | 低 — 近期 OpenAI 不太可能在 PI 所处层级直接竞争机器人 AI | 跟踪 OpenAI 的机器人野心;当前无冲突 |
| 风险 | 可能性 | 影响 | 关键个人 | 缓释状态 |
|---|---|---|---|---|
| Sergey Levine(首席科学家)离职 | 低–中 | 致命 | Sergey Levine | 未披露关键人保险或留任安排;UC Berkeley 的学术岗位存在吸引力 |
| Chelsea Finn(联合创始人)离职 | 低–中 | 高 | Chelsea Finn | 未披露留任安排;Stanford 学术岗位可能形成牵引 |
| Karol Hausman(CEO)离职 | 低 | 高 | Karol Hausman | CEO 与 CapitalG、Lux 关系深;离职将造成高度扰动 |
| 未能聘请资深企业销售负责人 | 中 | 高 | 未来 CRO / VP Sales(尚未宣布) | 未披露资深企业销售负责人;商业化发布的关键招聘 |
| 竞品扩大规模时的人才留存(Skild AI、Google DeepMind) | 中 | 中 | 整体工程团队 | 需要有竞争力的薪酬;下一轮需要刷新股权激励 |
| 联合创始人冲突或战略分歧 | 低 | 高 | 全体联合创始人 | 六位联合创始人高于平均水平;治理结构未披露 |
Physical Intelligence 的关键外部依赖,以及每项依赖若中断会触发的失效模式。
[CR016]7.4 展品
08估值
8.1 估值背景与进入纪律
Physical Intelligence 于 2025 年 11 月完成 Series B,融资 $600M,投后估值 $5.6B,意味着 CapitalG(Alphabet)、T. Rowe Price、Redpoint 和 Lux Capital 的投前进入价格为 $5.0B。交割时公司商业收入为零。机器人 AI 领域没有先例:一家公司在创立 20 个月内、零 ARR 情况下做到 $5.6B 估值。估值靠三点支撑:(1)团队质量极高(Levine、Finn、Hausman 均为顶级研究者);(2)技术执行到位(π₀ 论文和演示获得学界和产业强验证);(3)投资人相信市场规模(2030 年服务机器人市场 $170B+),也相信 Physical Intelligence 有机会像 LLM 领域的 OpenAI 一样,占住机器人基础模型 API 层。据报道的下一轮 $11B(2026 年 4 月,未确认)意味着不到六个月估值翻 2×;只有公司已经签下多个企业 LOI,或宣布一个有实质规模的商业伙伴,估值跳升才可信。任何以 $11B 进入的投资人,安全边际都比 Series B 投资人更窄:在 $11B 且零收入情况下,要把收入倍数恢复到合理的 10× 退出,需要 $1.1B ARR。换句话说,它要长成一家机器人基础模型龙头,规模接近前五大企业 SaaS 公司。投资周期是 10–15 年,不是 5–7 年。 [CV001, CV002, CV003, CV004, CV005]
| 投资逻辑要素 | 乐观逻辑(投资) | 反向逻辑(放弃) |
|---|---|---|
| 市场机会 | 到 2030 年,服务机器人市场超过 $170B;机器人 AI 是「操作系统」层;具备赢家拿走大部分的格局 | 市场仍在成形;企业采用速度慢于消费级 AI;市场规模预测经常落空 |
| 技术 | π₀ 跨具身架构处于同类最佳;68 种具身形态构成结构性数据护城河;创始团队有 MAML 和 DeepMind 背景 | arXiv 预印本已披露架构;Google DeepMind 拥有同等算力,并掌握 PaliGemma 所有权;Skild AI 的数据飞轮领先 |
| 收入轨迹 | 若 3 个企业账户按每台机器人每年 $15K、1,000 台机器人机队转化,到 2027 年 Q4 可做到 $50M ARR | 没有商业牵引;Skild AI 已有 $30M ARR;Physical Intelligence 为 0;销售周期 12–24 个月 |
| 团队 | Levine、Finn、Hausman 是全球前 5 的机器人 AI 研究者;能吸引最优秀机器人学 PhD 人才;企业买家也认可 | 6 名联合创始人带来治理风险;尚未宣布 CRO / 销售副总裁;Stanford 和 UC Berkeley 的学术职责仍会牵扯精力 |
| 估值 | $5.6B 与研究质量相近的 LLM 阶段 AI 公司一致;竞赛格局可支撑 $11B | 零收入阶段,$5.6B 是合理价值的 2–3×;零收入时 $11B 极端;可比的 Skild AI 467× 收入倍数也极端 |
Physical Intelligence 入场时的关键投资指标,概括财务、商业和风险画像。
[CV001, CV011, CV018]8.2 可比估值框架
没有收入时,估值只能依靠同阶段可比交易。最相关可比包括:(1)收入前 / 早期收入阶段的 LLM 基础模型公司(OpenAI 约 $1B ARR 时估值 $80B = 80× 远期收入,意味着 Physical Intelligence 要有约 $70M ARR 才算落入可比倍数区间);(2)类似融资阶段的机器人 AI 公司(Skild AI 约 $30M ARR 时估值 $14B = 467× 收入,倍数极端,但反映市场出清需求);(3)无收入的纯研究阶段 AI 公司(Inflection AI 在 Microsoft 合作前没有商业产品、估值 $4B)。Physical Intelligence 的 $5.6B 估值位于收入前 AI 公司极端估值区间内,但对机器人垂直公司来说处在高端。关键估值锚点是:(a)若 Physical Intelligence 到 2027 年实现 $50M ARR,50× 远期收入倍数(激进)可支撑 $2.5B 估值,低于 $5.6B;(b)若到 2028 年实现 $200M ARR,30× 远期倍数(中性)可支撑 $6B,与 Series B 接近;(c)乐观情景需要 2028 年达到 $400M+ ARR,并用 20× 倍数支撑 $8B 估值,才接近据报道 $11B 下一轮的下沿。结论是:只有 Physical Intelligence 在 24 个月内交出显著商业牵引力,Series B 估值才算“合理”;如果没有公开商业收入,据报道的下一轮极其激进。 [CV006, CV007, CV008, CV009, CV010]
| 情景 | 概率 | 2028 年 ARR | 2028 年估值 | 以 5.6B 入场的退出倍数 | 关键假设 |
|---|---|---|---|---|---|
| 乐观情景 | 20% | $200M–$400M | $8B–$15B | 1.4×–2.7×(温和) | PI 转化 5+ 家企业;Skild 未形成主导;Google API 未规模化推出;$15K/robot/yr 定价守住 |
| 基准情景 | 50% | $30M–$100M | $2B–$4B | 0.4×–0.7×(亏损) | 2–3 家企业转化;销售周期慢于预期;Skild 保持优势;Google Gemini Robotics 形成部分竞争 |
| 悲观情景 | 30% | $0–$15M | $500M–$1.5B | 0.09×–0.27×(重大亏损) | 到 2027 年仍无商业牵引;降价融资;关键人物离职;Google 收购 Skild 或推出 Gemini Robotics API |
| 公司 | ARR(可比阶段) | 估值 | 收入倍数 | 可比基础 | 与 PI 的相关性 |
|---|---|---|---|---|---|
| Skild AI(机器人 AI,2025) | ~$30M | ~$14B | 467× | 最直接可比;机器人基础模型;商业化尚未规模化 | 高 — 同一赛道;按收入倍数看 Skild 更贵 |
| Cohere(LLM SaaS,2025) | ~$240M | ~$5.1B | 21× | D 轮阶段 AI 企业 SaaS;投资人画像相近 | 中 — 垂直领域不同;同等估值下 Cohere 收入是 PI 的 8× |
| Inflection AI(LLM,商业化前,2023) | ~$0 | ~$4B | N/A(尚无收入) | 与 Microsoft 合作前的收入前 LLM 公司;可类比 B 轮 | 中 — 收入前溢价;Inflection 被收购时隐含估值 $650M,而不是 $4B |
| OpenAI(LLM,C 轮阶段) | ~$1B | ~$80B | 80× | 主导型 LLM 基础模型;不是直接可比,但定义市场参照 | 低 — PI 聚焦机器人;OpenAI 融资额是 PI 的 14×,且占据主导市场 |
| Figure AI(机器人硬件 + AI,2025) | ~$50M(估计试点收入) | ~$39B | ~780× | 全栈硬件 + 软件机器人公司;与 BMW 合作 | 中 — 技术栈不同(硬件 + AI,而 PI 仅做软件) |
| 典型 B 轮 AI SaaS(2025) | $20M–$50M | $200M–$500M | 10–25× | 资本阶段可比的企业 SaaS;不聚焦机器人 | 低 — 行业不同;显示 PI 距传统 SaaS 效率有多远 |
在不同 ARR 达成情景下,Physical Intelligence 2027 年和 2028 年低 / 基准 / 高估值估计。
所有数值均为分析师估计;Physical Intelligence 未披露财务目标。
[CV007, CV008, CV016]8.3 建议、退出准备与最终尽调问题
我们的建议是谨慎——当前估值下放弃或观察,未来进入需满足条件。技术可信,团队世界级,机器人基础模型市场也有一条通向极大规模的合理路径。但零收入对应 $5.6B 估值,对多数机构投资人来说安全边际不足,据报道的 $11B 下一轮更是极其激进。我们会建议进入的条件是:(1)Physical Intelligence 宣布至少 $50M ARR,来自两个或更多具名企业客户;(2)公司完成至少一项功能安全认证(ISO 13849 或 CE 标志),可以支撑生产部署;(3)与 Google 的 PaliGemma 商业许可协议有文件记录并已披露;(4)Sergey Levine 的留任与股权归属时间表得到确认。在 2026–2027 年商业化发布、并按典型企业软件扩张路径推进的假设下,退出准备至少还要 5–7 年。潜在战略收购方包括 Amazon、Microsoft(通过 Azure 机器人业务)、Samsung、Hyundai,以及任何想要 AI 机器人栈的大型工业集团。IPO 准备需要 $300M+ ARR 和规模化正毛利率;即便按乐观假设,时间也要到 2029 年以后。最终尽调问题列在 T808。 [CV011, CV012, CV013, CV014, CV015]
| 维度 | 评估 | 置信度 |
|---|---|---|
| 总体建议 | 谨慎 — 当前入场点选择放弃或观察;出现商业化证据后再重新评估 | 中 |
| 投资风险评级 | 高(尚无收入;估值极端;Google 冲突;关键人物风险) | 高 |
| 估值立场 | 偏高 — 以当前商业化阶段看,$5.6B 是合理价值的 2–3× | 中 |
| 据报道下一轮($11B)立场 | 严重偏高 — 以 $11B 入场,需要 $400M+ ARR 才有合理的 5 年回报 | 中 |
| 持有期(若已投资) | 至少 5–7 年;达到 IPO 准备状态需 7–10 年 | 低 |
| 目标退出倍数(若以 $5.6B 投资) | 投资资本 3–5× 回报需要退出时估值 $15–28B;只有乐观情景才可能做到 | 低 |
| 战略收购方概率 | 中 — 在 $3–10B 估值区间,Amazon、Microsoft、Samsung、Hyundai、Bosch 都可能成为买家 | 低 |
| 触发因素 | 阈值 | 时间范围 | 动作 |
|---|---|---|---|
| 到 2026 年 Q4 仍无商业收入 | 到 2026 年 12 月 ARR 为 0,且没有具名 LOI 阶段客户 | 18 个月 | 退出(若已投资);停止(若在评估) |
| Skild AI 超过 $100M ARR | PI 还没有收入前,Skild 公开报告或确认 $100M ARR | 12 个月 | 重新评估投资逻辑;PI 的市场份额路径显著变窄 |
| Google Gemini Robotics 商业 API 规模化 | Google 公布 Gemini Robotics 商业定价,并通过 GCP 分发 | 18 个月 | 退出(若已投资);停止(若在评估);PI 的 VLM 依赖变得关键 |
| PaliGemma 授权限制 | Google 公布 PaliGemma 商业使用限制,过渡期 < 6 个月 | 持续 | 立即退出(若已投资);PI 架构必须重建,成本 $10M+ |
| Sergey Levine 离职 | LinkedIn 资料变更;离职公告;未任命继任者 | 持续 | 退出(若已投资);投资逻辑根本削弱 |
| $5.6B 以下降价融资 | 任何投后估值 < $5.6B 的融资 | 持续 | 重大减值信号;评估战略替代方案 |
| 尽调事项 | 优先级 | 理由 |
|---|---|---|
| 完整企业试点客户名单,包含具名推荐人和当前状态 | 关键 | 没有客户名称就无法评估商业牵引;LOI 状态决定投资逻辑 |
| 与 Google 的 PaliGemma 商业授权协议 | 关键 | 架构依赖;Gemma 使用条款不足以提供投资级 IP 保护 |
| 训练数据来源审计和版权清理 | 关键 | IP 责任是潜在风险;若任何主要机器人视频数据集受版权保护,影响将很重大 |
| Sergey Levine 和 Chelsea Finn 归属时间表与留任协议 | 高 | 关键人物风险需要在投资前用合同保护 |
| 股权结构表和治理文件(董事会构成、投票权、联合创始人协议) | 高 | 6 名联合创始人结构需要清晰治理;CapitalG 冲突必须处理 |
| 烧钱速度、账面现金和财务报表(管理账) | 高 | 现金跑道建模需要实际烧钱速度;$70–$150M 估计区间过宽,难以支持投资 |
| 试点转化计划和商业发布时点 | 高 | 通往收入的关键路径;投资逻辑取决于 18 个月内转化 |
| 功能安全认证路线图 | 中 | 企业部署必需;影响商业发布时点 |
从研究阶段评估,经估值分析,到 Physical Intelligence 投资建议的决策逻辑。
[CV011, CV012]可比公司在同等融资阶段时的估值,展示 Physical Intelligence 相对同业组合的位置。
所有可比项来自媒体和分析师报告;未经独立核验。
[CV017]8.4 展品
免责声明
本报告是基于公开证据的尽调快照,不构成投资建议。重要的财务、法律、技术和合同事实仍未公开;任何投资决策前,应直接向管理层核验,并查阅一手文件。
证据索引
| 编号 | 陈述 | 可信度 | 来源 |
|---|---|---|---|
| CO001 | Physical Intelligence was founded in March 2024 in San Francisco, California. | 高 | SO002, SO005 |
| CO002 | Physical Intelligence's stated mission is to build a single general-purpose AI system capable of controlling any robot for any task. | 高 | SO001, SO024 |
| CO003 | Physical Intelligence's primary product is π₀ (pi-zero), a vision-language-action (VLA) foundation model for robot control. | 高 | SO011, SO012 |
| CO004 | The π₀ model is hardware-agnostic and can be fine-tuned for diverse robot platforms without task-specific reprogramming. | 中 | SO011, SO001 |
| CO005 | Physical Intelligence has open-sourced its core model (openpi) as an ecosystem play to drive developer adoption and training data contributions. | 高 | SO013, SO014 |
| CO006 | Karol Hausman is CEO and co-founder of Physical Intelligence; he was previously a Staff Research Scientist at Google DeepMind and an adjunct professor at Stanford. | 高 | SO004, SO005, SO025 |
| CO007 | Sergey Levine is Chief Scientist and co-founder; he is a tenured Associate Professor at UC Berkeley leading the RAIL Lab. | 高 | SO004, SO023 |
| CO008 | Chelsea Finn is a co-founder and advisor to Physical Intelligence; she is an Assistant Professor at Stanford known for inventing Model-Agnostic Meta-Learning (MAML). | 高 | SO004, SO022 |
| CO009 | Brian Ichter is a co-founder; he was a Research Scientist at Google DeepMind focusing on kinodynamic planning and scalable robot algorithms. | 中 | SO004, SO006 |
| CO010 | Adnan Esmail is a co-founder and VP Engineering; he previously served as SVP Engineering at Anduril and was a Tesla Autopilot engineer. | 中 | SO004, SO006 |
| CO011 | Lachy Groom is a co-founder providing business leadership; he was an early executive at Stripe and a venture capital investor. | 中 | SO004, SO006 |
| CO012 | Quan Vuong is a co-founder and researcher contributing to the π₀ training pipeline and research agenda. | 中 | SO004, SO006 |
| CO013 | Physical Intelligence's board composition has not been publicly disclosed; no independent directors have been named as of Q1 2026. | 高 | SO005, SO024 |
| CO014 | Physical Intelligence raised a $70 million seed round in mid-2024, led by Lux Capital with participation from Jeff Bezos. | 高 | SO005, SO009, SO020 |
| CO015 | Physical Intelligence raised a $400 million Series A in November 2024 at a $2.4 billion post-money valuation, with investors including OpenAI, Thrive Capital, Lux Capital, and Index Ventures. | 高 | SO009, SO010, SO005 |
| CO016 | Physical Intelligence raised a $600 million Series B in November 2025 at a $5.6 billion post-money valuation, led by CapitalG (Alphabet's growth fund). | 高 | SO002, SO008, SO018 |
| CO017 | Series B participants included Lux Capital, Bond, Redpoint, Sequoia Capital, Thrive Capital, Index Ventures, T. Rowe Price, and Jeff Bezos. | 高 | SO002, SO007, SO008 |
| CO018 | Total capital raised by Physical Intelligence through the Series B is approximately $1.07 billion ($70M seed + $400M Series A + $600M Series B). | 高 | SO009, SO002, SO008 |
| CO019 | OpenAI is a strategic investor in Physical Intelligence from the Series A; CapitalG (Alphabet) led the Series B, creating a potential investor conflict given competitive AI interests. | 中 | SO010, SO018, SO021 |
| CO020 | Physical Intelligence is pre-commercial-revenue as of Q1 2026; no ARR or revenue has been publicly disclosed. | 高 | SO005, SO019, SO021 |
| CO021 | Physical Intelligence announced the π₀ model in October 2024 via a technical blog post and arXiv paper, demonstrating tasks including laundry folding, box assembly, and shirt packaging. | 高 | SO011, SO012 |
| CO022 | Physical Intelligence open-sourced π₀ model weights and code as the openpi repository in February 2025, making it the first open-source general robot VLA foundation model. | 高 | SO013, SO014 |
| CO023 | Physical Intelligence released π₀.5 in early 2025, featuring enhanced generalization across novel robot platforms and environments. | 中 | SO017, SO001 |
| CO024 | π₀-FAST was introduced in mid-2025, using a frequency-domain discrete action representation (FAST tokenizer) for more efficient robot inference at lower compute cost. | 中 | SO017, SO011 |
| CO025 | Physical Intelligence is conducting early enterprise pilot programs in manufacturing and logistics verticals as of Q1 2026, though customer names have not been disclosed. | 中 | SO017, SO019 |
| CO026 | The Series A ($400M in November 2024) was reported to be the largest Series A in robotics AI history at the time of its close. | 中 | SO007, SO002 |
| CO027 | Physical Intelligence reached over $1 billion in total raised within approximately 20 months of founding, the fastest-ever for a pure-software robotics AI startup. | 中 | SO002, SO007, SO009 |
| CO028 | Physical Intelligence's estimated headcount is approximately 150–250 employees as of Q1 2026, based on public LinkedIn profiles and hiring pace; this figure is not officially disclosed. | 低 | SO005, SO017 |
| CO029 | Physical Intelligence has not disclosed any significant adverse events, regulatory investigations, or leadership departures as of Q1 2026. | 高 | SO005, SO021, SO024 |
| CO030 | As of April 2026, Physical Intelligence is reportedly in advanced talks for a further funding round at approximately an $11 billion valuation, roughly double the Series B valuation. | 低 | SO016, SO017 |
| CO031 | The co-founders Sergey Levine and Chelsea Finn hold concurrent academic positions at UC Berkeley and Stanford respectively, creating potential dual-commitment risk. | 高 | SO022, SO023 |
| CO032 | No succession plan or governance safeguards for Physical Intelligence leadership have been publicly disclosed. | 高 | SO005, SO021 |
| CO033 | Jeff Bezos has a personal investment in Physical Intelligence across both the seed and Series B rounds. | 高 | SO007, SO008 |
| CO034 | Lux Capital participated in all three funding rounds (seed, Series A, Series B), making it the longest-tenured institutional investor. | 高 | SO020, SO009 |
| CO035 | Thrive Capital participated in both the Series A and Series B rounds of Physical Intelligence. | 高 | SO008, SO009 |
| CO036 | Karol Hausman contributed to the development of RT-2 (Robotic Transformer 2) at Google DeepMind before founding Physical Intelligence. | 高 | SO025, SO004 |
| CO037 | Chelsea Finn invented Model-Agnostic Meta-Learning (MAML), a core technique for rapid robot task adaptation underpinning aspects of π₀'s design. | 高 | SO022, SO004 |
| CO038 | Physical Intelligence has no disclosed debt financing, secondary transactions, or revenue-based financing as of Q1 2026. | 高 | SO009, SO002 |
| CO039 | Physical Intelligence is headquartered at San Francisco, California; no additional offices have been publicly announced. | 高 | SO005, SO001 |
| CO040 | The openpi repository on GitHub contains model weights, training code, and documentation for π₀, enabling community fine-tuning and contributions. | 高 | SO013, SO014 |
| CM001 | The global physical AI and robotics market (hardware + software) is estimated at $50–82 billion in 2025. | 中 | SM001, SM002 |
| CM002 | The physical AI and robotics market is projected to reach $111–185 billion by 2030 at a CAGR of 30–40%. | 中 | SM001, SM003 |
| CM003 | The AI-in-robotics software SAM (intelligence and control software only) is estimated at $5.8 billion in 2025, growing to $25.9 billion by 2030 at 35% CAGR. | 中 | SM002, SM004 |
| CM004 | No established analyst category exists for robot foundation model software as a standalone market; PI must argue its market from broader robotics software data and unit-economics extrapolation. | 高 | SM001, SM004 |
| CM005 | Physical Intelligence targets the software intelligence layer of the robotics value chain with a planned SaaS-per-robot licensing model. | 高 | SM006, SM007 |
| CM006 | The physical AI market estimated by Grand View Research reaches approximately $81.6 billion in 2025 at a CAGR of 32% through 2033. | 中 | SM001 |
| CM007 | BCG and McKinsey analyses suggest robot AI software layers will capture 40–60% of total robotics value chain margins as intelligence becomes the primary differentiator. | 中 | SM012, SM014 |
| CM008 | Physical Intelligence's estimated SOM is $200–500 million by 2028, based on a 10–30% win rate among approximately 100–300 enterprise accounts expected to adopt robot foundation model software in that window. | 低 | SM006, SM004 |
| CM009 | AI software gross margins in enterprise SaaS are typically 70–85%, compared to 20–40% for robot hardware manufacturing. | 中 | SM013, SM014 |
| CM010 | The robot foundation model SaaS segment does not yet exist as a distinct commercial market; Physical Intelligence is attempting to create it. | 高 | SM004, SM006 |
| CM011 | Analyst estimates for the physical AI and robotics TAM exhibit wide variance due to definitional differences; low estimates ($50B) and high estimates ($82B) both represent well-supported positions. | 高 | SM001, SM002, SM003 |
| CM012 | The primary buyer segments for robot foundation model software include manufacturing and assembly operators, logistics and warehouse 3PLs, and robot OEM manufacturers. | 中 | SM006, SM007 |
| CM013 | The enterprise buyer decision-maker for robot AI adoption is typically the VP of Operations or Chief Automation Officer, with IT security and procurement teams involved. | 中 | SM024, SM010 |
| CM014 | Enterprise robotics sales cycles are typically 12–24 months from pilot to signed contract, requiring PI to demonstrate ROI and safety certification early. | 中 | SM024, SM008 |
| CM015 | North America and Europe are Physical Intelligence's primary markets in the near term; Asia-Pacific represents the largest long-term volume market given high industrial robot density. | 中 | SM015, SM016 |
| CM016 | Robot OEM partnerships offer faster distribution for Physical Intelligence but lower margins and higher dependency on partner success versus direct enterprise sales. | 中 | SM006, SM014 |
| CM017 | Labor shortages in manufacturing and logistics driven by demographic trends are a primary driver for robotics automation adoption, making ROI calculations favorable. | 高 | SM011, SM010 |
| CM018 | Foundation models reduce the cost of programming a robot for a new task from $50,000–$250,000 for custom software development to near-zero via fine-tuning, creating a step-change in adoption economics. | 中 | SM012, SM014 |
| CM019 | Physical Intelligence's open-source release of π₀ in February 2025 creates ecosystem momentum but simultaneously enables competitors to build on PI's base model, compressing the proprietary moat. | 高 | SM020, SM021 |
| CM020 | Well-funded competitors including Skild AI ($1.4B raised, $14B valuation) and Google DeepMind represent a high competitive threat to Physical Intelligence's market position. | 中 | SM017, SM018 |
| CM021 | Safety and reliability requirements for commercial robot deployment (ISO 10218 and IEC 62061) are extremely stringent and require separate certification before enterprise procurement. | 高 | SM009, SM024 |
| CM022 | Compute costs for training large robot foundation models and running inference at scale remain high relative to early commercial economics, creating a margin headwind. | 中 | SM012, SM014 |
| CM023 | Capital inflows to the physical AI and robotics sector exceeded $5 billion in 2024–2025, creating ecosystem development momentum and accelerating the competitive race. | 中 | SM025, SM003 |
| CM024 | South Korea leads global robot density at over 1,000 robots per 10,000 manufacturing employees; Japan and Germany follow. Asia-Pacific's industrial robot base represents the largest addressable long-term volume market. | 高 | SM015, SM022 |
| CM025 | BCG argues that physical AI is reshaping robotics by enabling software intelligence to displace hardware differentiation, with the dominant model provider capturing outsized margins. | 中 | SM012, SM019 |
| CM026 | The physical AI market is projected by Future Markets Inc at cumulative spend of $384–663 billion between 2026 and 2030 across hardware, software, and services. | 低 | SM005 |
| CM027 | Analyst estimates for physical AI TAM exhibit a very wide range ($33–185B by 2030) depending on hardware inclusion and market scope definition. | 高 | SM001, SM002, SM003, SM005 |
| CM028 | The humanoid robot sub-market is growing rapidly but is not yet the primary target for Physical Intelligence, which focuses on arms and mobile manipulators first. | 低 | SM007, SM006 |
| CM029 | AI-in-robotics software (SAM) is estimated at $5.8B in 2025 and $25.9B in 2030 according to MarketsandMarkets. | 中 | SM002 |
| CM030 | The physical AI TAM ranges from $111B (low scenario) to $185B (high scenario) by 2030 based on Grand View Research and MarketsandMarkets. | 中 | SM001, SM002 |
| CM031 | Physical Intelligence's TAM for robot foundation model SaaS is an inferred estimate of $3–8 billion by 2030; no analyst report covers this specific category. | 低 | SM004, SM014 |
| CM032 | Enterprise buyers in manufacturing and logistics evaluate robot AI over 12–24 months, requiring pilot demonstrations, safety testing, and IT security review before contract. | 中 | SM024, SM008 |
| CM033 | The enterprise ACV for robot foundation model software deployments is estimated at $50,000–$1 million per deployment per year depending on vertical and fleet size. | 低 | SM006, SM014 |
| CM034 | System integrators represent a secondary go-to-market channel for Physical Intelligence, enabling turnkey factory automation deployments via margin-sharing arrangements. | 低 | SM007, SM006 |
| CM035 | The overall enterprise funnel for robot foundation model adoption is very early-stage; estimated 5,000 enterprises are aware, 500 are evaluating, 50 are piloting, and 5 have commercial contracts globally as of Q1 2026. | 低 | SM004, SM007 |
| CP001 | The robot foundation model market has three competitive archetypes — software-only intelligence layer, full-stack robotics integrators, and incumbent tech firms extending into robotics. | 中 | SP013, SP025 |
| CP002 | Physical Intelligence and Skild AI are the two primary software-only robot intelligence layer competitors; both are hardware-agnostic and build general-purpose robot brains. | 高 | SP013, SP002 |
| CP003 | Figure AI, 1X Technologies, and Agility Robotics are full-stack robotics integrators that build both hardware and proprietary AI — potential customers for PI's model but also competing for enterprise wallet share. | 高 | SP005, SP011 |
| CP004 | Google DeepMind's Gemini Robotics 1.5, launched in early 2026, is accessible via API with advanced agentic planning capabilities backed by Google's compute infrastructure. | 高 | SP006, SP007, SP008 |
| CP005 | Skild AI raised $1.4 billion at a $14 billion valuation in early 2026, led by SoftBank and NVIDIA, with Samsung and Salesforce Ventures participating. | 高 | SP001, SP021, SP015 |
| CP006 | Skild AI reported approximately $30 million in revenue within months of its commercial launch in 2025 — the highest commercial traction of any robot foundation model software-only competitor. | 中 | SP002, SP022 |
| CP007 | Google DeepMind's Gemini Robotics 1.5 uses agentic reasoning chains and multi-step planning, giving it superior performance on complex compositional tasks versus current π₀ capabilities. | 中 | SP006, SP007, SP008 |
| CP008 | Google DeepMind has orders of magnitude more compute and training data available for Gemini Robotics than Physical Intelligence can access as a standalone startup. | 高 | SP008, SP018 |
| CP009 | Figure AI raised $675 million at a $39 billion+ valuation in 2025, with investors including OpenAI and Microsoft. | 高 | SP005, SP023 |
| CP010 | Figure AI's Helix VLA model is deployed at BMW manufacturing plants, giving Figure the most advanced real-world commercial deployment of any VLA-based humanoid robot system. | 高 | SP004, SP023 |
| CP011 | Amazon acquired Covariant's AI team and technology in 2024, integrating the Covariant Brain into Amazon Robotics for warehouse automation; this removes Covariant as a potential robot OEM partner for Physical Intelligence. | 高 | SP009, SP010 |
| CP012 | 1X Technologies, a Norwegian humanoid robotics company founded in 2014, is targeting a raise of up to $1 billion at approximately a $10 billion valuation in 2026, with its 1XWM world model approach to robot cognition. | 中 | SP011, SP012 |
| CP013 | NVIDIA launched the Cosmos world foundation model for physical AI, providing a training-data generation and simulation platform that any competitor (including PI's rivals) can use to accelerate robot model development. | 高 | SP018, SP025 |
| CP014 | Physical Intelligence's founding team (Levine, Finn, Hausman) represents arguably the world's most credible academic-commercial robotics AI founding team, which is a durable competitive advantage for talent attraction and research credibility. | 高 | SP016, SP017 |
| CP015 | π₀ demonstrated industry-leading long-horizon dexterous manipulation tasks (laundry folding, assembly) in 2024 not matched by most competitors in published benchmarks at the time. | 中 | SP016, SP015 |
| CP016 | Physical Intelligence's open-source release of π₀ (openpi) creates developer community and ecosystem contributions but simultaneously gives rivals a technical reference to build against. | 高 | SP019, SP020 |
| CP017 | Physical Intelligence has no commercial revenue as of Q1 2026; Skild AI has ~$30M ARR from deployments — this is the most significant near-term competitive gap. | 高 | SP002, SP022, SP015 |
| CP018 | Skild AI's commercial revenue creates a data flywheel from production deployments that compounds its model advantage over time, threatening Physical Intelligence's ability to catch up. | 中 | SP002, SP015 |
| CP019 | Google DeepMind does not have commercial pricing for Gemini Robotics as of early 2026; availability is via research API with commercial launch timeline undisclosed. | 中 | SP006, SP007 |
| CP020 | Physical Intelligence's hardware-agnostic architecture enables any robot OEM partnership without hardware lock-in, which is a strategic advantage over full-stack integrators. | 高 | SP016, SP013 |
| CP021 | Skild AI demonstrates strong cross-embodiment generalization (omni-bodied architecture) via hierarchical architecture; this is comparable to Physical Intelligence's VLA approach. | 中 | SP002, SP003, SP015 |
| CP022 | Physical Intelligence's π₀ uses PaliGemma 3B as a VLM backbone, providing strong language instruction following; Google Gemini Robotics provides superior multimodal reasoning due to Gemini's training scale. | 中 | SP007, SP016 |
| CP023 | Physical Intelligence is the only major competitor to fully open-source model weights and training code, a unique ecosystem-building move not replicated by Skild AI, Figure AI, or Google DeepMind. | 高 | SP019, SP020 |
| CP024 | No robot foundation model competitor has demonstrated safety certification sufficient for broad commercial deployment; this is an industry-wide gap, not unique to Physical Intelligence. | 中 | SP013, SP014 |
| CP025 | Covariant Brain has the deepest real-world operational data in warehouse manipulation, having processed over 10 million picks in Amazon warehouses; this data advantage is now proprietary to Amazon. | 中 | SP009, SP010 |
| CP026 | No competitor has publicly disclosed full pricing for robot foundation model software; all pricing is deal-specific and enterprise-negotiated. | 高 | SP013, SP002 |
| CP027 | Physical Intelligence's planned per-robot SaaS pricing is estimated at $5–15K per robot per year; this is in line with AI2Work estimates for Skild AI but unconfirmed. | 低 | SP002, SP015 |
| CP028 | Figure AI's all-in robot pricing (hardware + software) is estimated at $100–300K per robot, far above PI's software-only model, reflecting the different value propositions. | 低 | SP005, SP004 |
| CP029 | Physical Intelligence has raised ~$1.07B; Skild AI has raised ~$1.7B (2.5× more); this funding gap represents a significant resource disadvantage in the compute-intensive robot AI market. | 高 | SP001, SP017 |
| CP030 | The total capital raised by the top 5 robot AI competitors (PI, Skild, Figure, 1X, Covariant) exceeds $4 billion as of Q1 2026, reflecting intense investor competition and high capital requirements. | 中 | SP025, SP013 |
| CP031 | On a competitive positioning map of capital raised versus commercial traction, Physical Intelligence occupies a high-capital but zero-revenue quadrant, while Skild AI leads on commercial traction. | 高 | SP002, SP017 |
| CP032 | Google DeepMind occupies an asymmetric competitive position — essentially unlimited resources versus any startup competitor — making sustained technical parity extremely difficult for Physical Intelligence to maintain. | 中 | SP008, SP018 |
| CP033 | Skild AI scores highest among software-only competitors on cross-embodiment generalization and commercial traction; Physical Intelligence scores highest on dexterous manipulation depth and open ecosystem. | 中 | SP002, SP015 |
| CP034 | Google DeepMind Gemini Robotics scores highest on language following and multimodal reasoning due to Gemini training scale, with Physical Intelligence and Skild AI competitive on cross-embodiment. | 中 | SP007, SP008 |
| CP035 | Physical Intelligence's competitive readiness scores 9/10 on research pedigree and 8/10 on technical manipulation but 0/10 on commercial revenue and 2/10 on enterprise partnership depth. | 中 | SP016, SP013 |
| CP036 | Safety and certification readiness scores approximately 2/10 for Physical Intelligence, with no disclosed commercial safety certification; this is a gate for enterprise contracts. | 中 | SP013, SP014 |
| CI001 | Physical Intelligence has no disclosed commercial revenue or ARR as of Q1 2026; the company is conducting enterprise pilot programs but has not signed commercial contracts. | 高 | SI001, SI002 |
| CI002 | Physical Intelligence's planned revenue model is a SaaS-per-robot licensing structure, charging enterprise customers a recurring annual fee per deployed robot using PI's foundation model. | 高 | SI001, SI007 |
| CI003 | Physical Intelligence has not publicly announced pricing for its SaaS-per-robot model; industry estimates suggest $5,000–$15,000 per robot per year based on analogies to LLM API pricing. | 低 | SI008, SI007 |
| CI004 | Reaching $50M ARR at $5,000–$15,000 per robot per year would require Physical Intelligence to have approximately 3,300–10,000 active robot deployments under license. | 低 | SI008, SI001 |
| CI005 | The openpi open-source release generates no direct revenue and is treated as a community and ecosystem investment that does not contribute to ARR. | 高 | SI022, SI023 |
| CI006 | AI SaaS gross margins are typically 70–85% at scale when model training costs are amortized across a large deployment base; Physical Intelligence targets this margin structure. | 中 | SI011, SI012 |
| CI007 | Physical Intelligence's estimated annual burn rate is $70–$150 million per year, based on approximately 150–250 employees at $300,000–$400,000 all-in compensation per employee, plus GPU compute and overhead. | 低 | SI005, SI006 |
| CI008 | GPU compute costs for training large VLA models at π₀'s scale are estimated at $5–20 million per major training run based on publicly available model training cost benchmarks. | 低 | SI017, SI018 |
| CI009 | At an estimated burn of $70–$150 million per year, the $600 million Series B provides approximately 4–8 years of runway at pre-commercial scale before requiring an additional raise. | 低 | SI013, SI014 |
| CI010 | Enterprise robotics software sales cycles are typically 12–24 months from initial pilot to signed contract, implying CAC payback periods of 3–7 years for Physical Intelligence at estimated ACV levels. | 低 | SI020, SI021 |
| CI011 | Physical Intelligence has not disclosed any debt financing, convertible notes, or revenue-based financing; all confirmed capital is venture equity. | 高 | SI013, SI002 |
| CI012 | Physical Intelligence's $5.6 billion post-money valuation at $0 revenue implies an effectively infinite revenue multiple, which is justified only by investor expectations of future category leadership. | 高 | SI015, SI025 |
| CI013 | If Physical Intelligence fails to demonstrate commercial traction within 12–18 months, the next fundraising round at $11B valuation becomes difficult to justify, creating existential funding risk. | 中 | SI015, SI016 |
| CI014 | Physical Intelligence raised a $70M seed round in mid-2024, a $400M Series A at $2.4B valuation in November 2024, and a $600M Series B at $5.6B valuation in November 2025. | 高 | SI002, SI013, SI014 |
| CI015 | Physical Intelligence has filed SEC Form D exempt offering notices for its financing rounds; these filings confirm offering sizes but disclose no financial operating data. | 高 | SI003, SI004 |
| CI016 | To justify a $10B+ valuation by 2028, Physical Intelligence would need to demonstrate approximately $200–$350M in ARR assuming a 30–50× forward revenue multiple consistent with high-growth AI SaaS. | 低 | SI012, SI024 |
| CI017 | Physical Intelligence is reportedly in advanced talks for a further funding round at approximately $11 billion valuation as of April 2026; this represents a ~2× step-up in under six months. | 低 | SI019 |
| CI018 | Physical Intelligence has not disclosed revenue, gross margin, operating loss, burn rate, cash on hand, or customer count as of Q1 2026. | 高 | SI001, SI002 |
| CI019 | As a private company with no SEC reporting obligation beyond Form D filings, Physical Intelligence has no requirement to disclose financial statements, making diligence dependent on management accounts. | 高 | SI003, SI004 |
| CI020 | Physical Intelligence's financial diligence gaps include revenue, gross margin, burn rate, cash on hand, customer count, and headcount by function — all require VDD data room access. | 高 | SI001, SI013 |
| CI021 | Physical Intelligence's revenue path requires sequential achievement of pilot conversion, safety certification, commercial pricing, and fleet-scale deployment before ARR becomes meaningful. | 中 | SI001, SI008 |
| CI022 | If Physical Intelligence scales headcount and compute aggressively to capture market opportunity, runway compresses to approximately 24–36 months from the $600M Series B. | 低 | SI006, SI017 |
| CI023 | Physical Intelligence has raised a cumulative $1.07B across its three rounds; the next reported round at $11B could add $500M–$1B more. | 低 | SI019, SI013 |
| CI024 | On a capital efficiency comparison, Physical Intelligence's $0 ARR at $1.07B raised is a significant outlier; Skild AI achieved ~$30M ARR at $1.7B raised (more efficient) and Cohere achieved ~$240M ARR at $975M raised (far more efficient for comparable capital). | 中 | SI009, SI024 |
| CI025 | Typical enterprise SaaS companies at Series B have $20–50M ARR on $50–100M raised; Physical Intelligence's capital intensity is approximately 10–20× higher than this benchmark. | 中 | SI021, SI006 |
| CI026 | Physical Intelligence's Series A was co-led by Thrive Capital and included participation from Sequoia Capital, Lux Capital, Index Ventures, Bond, and Jeff Bezos personally. | 高 | SI002, SI013 |
| CI027 | CapitalG (Alphabet's growth equity fund) led the $600M Series B, with follow-on participation from T. Rowe Price, Redpoint Ventures, and Lux Capital. | 高 | SI002, SI014 |
| CI028 | OpenAI participated in Physical Intelligence's Series A as a strategic investor, representing one of the few instances of OpenAI investing in a direct AI ecosystem company outside its own platform. | 中 | SI002, SI007 |
| CI029 | Physical Intelligence's go-to-market strategy targets enterprise manufacturing and logistics customers with long deployment cycles and high robot fleet density, maximizing per-account ACV. | 中 | SI001, SI007 |
| CI030 | The enterprise robotics and automation software market has historically required 18–36 month vendor qualification cycles before production deployment, implying long CAC payback for new market entrants. | 中 | SI020, SI008 |
| CI031 | Unlike cloud SaaS, robot AI software requires on-robot inference, which adds per-unit edge compute cost to the delivery model and can compress gross margins compared to purely cloud-delivered SaaS. | 中 | SI018, SI012 |
| CI032 | Comparable AI foundation model companies (Cohere, Mistral, Anthropic) that have reached $100M+ ARR took 3–4 years post-founding; Physical Intelligence, founded in March 2024, would need to exceed this pace to justify its valuation. | 中 | SI024, SI012 |
| CI033 | Physical Intelligence has not disclosed whether any of its enterprise pilot customers are paying for the pilot, receiving it as a free proof-of-concept, or co-developing under a research agreement. | 低 | |
| CI034 | If the robotics AI market follows SaaS LTV dynamics with robot re-training lock-in, churn rates could be below 10% annually, which would improve LTV/CAC ratios to 5–10× over 5-year customer lifetimes. | 低 | SI021, SI008 |
| CI035 | The open-source openpi release creates a negative financial precedent by demonstrating PI's core capabilities at no cost; enterprise customers may attempt to self-host rather than pay per-robot license fees. | 中 | SI022, SI016 |
| CE001 | π₀ uses a hybrid VLA architecture combining a PaliGemma 3B vision-language model backbone with a 300-million-parameter action expert transformer, trained via flow matching to generate continuous robot control signals. | 高 | SE001, SE015 |
| CE002 | Flow matching was chosen over diffusion for action generation because it provides higher inference speed and similar or better accuracy on continuous action distributions relevant to robot control. | 高 | SE001, SE011 |
| CE003 | Physical Intelligence has released three model variants — π₀ (October 2024), π₀.5 (2025), and π₀-FAST (2025) — each addressing different aspects of capability and inference efficiency. | 高 | SE002, SE003, SE004 |
| CE004 | The openpi open-source release (February 2025) provides fine-tuning utilities and access to π₀ model weights, enabling external researchers to adapt the model to new robot types and tasks. | 高 | SE005, SE006 |
| CE005 | Physical Intelligence has publicly demonstrated π₀ on tasks including laundry folding, dishwasher loading, object sorting, and package handling across manufacturing and logistics settings. | 高 | SE023, SE024 |
| CE006 | π₀ is trained across data collected from 68 distinct robot embodiments, enabling cross-embodiment generalization without robot-specific model retraining. | 高 | SE001, SE015, SE016 |
| CE007 | π₀ outperforms OpenVLA, RT-2, and Octo on the LIBERO benchmark suite for dexterous manipulation tasks, particularly on long-horizon and contact-rich tasks. | 中 | SE007, SE008 |
| CE008 | The LIBERO benchmark evaluates robot task completion rates across object manipulation scenarios of increasing complexity; higher scores indicate better generalization with limited demonstration data. | 高 | SE007, SE001 |
| CE009 | Industry experience in robotics suggests that production deployment success rates are typically 20–40 percentage points below controlled laboratory benchmark results due to unstructured environments, varied lighting, and real-time reliability requirements. | 中 | SE008, SE025 |
| CE010 | Enterprise manufacturing and logistics customers will require demonstrated production success rates above 99% for high-volume tasks; π₀'s published benchmark performance is below this threshold in most evaluations. | 中 | SE017, SE018 |
| CE011 | Physical Intelligence is targeting enterprise manufacturing and logistics as primary industry verticals for its enterprise pilot programs, with no currently disclosed commercial customers. | 高 | SE024, SE002 |
| CE012 | Physical Intelligence's π₀ architecture relies on PaliGemma, a model released by Google DeepMind under open weights terms; a change in Google's licensing policy would require retraining the entire foundation model. | 高 | SE009, SE010 |
| CE013 | PaliGemma is distributed under Google's Gemma Terms of Use, which allow commercial use but prohibit certain modifications and can be revised by Google unilaterally, creating licensing risk for Physical Intelligence. | 中 | SE010, SE009 |
| CE014 | Physical Intelligence has not publicly disclosed patents related to π₀'s VLA architecture; the primary technical IP appears to be described in the arXiv preprint, which lowers the barrier for replication by well-resourced competitors. | 中 | SE001, SE025 |
| CE015 | Physical Intelligence has not publicly announced functional safety certifications (ISO 13849, IEC 62061) or CE marking for π₀, which are typically required for production deployment in European manufacturing environments. | 高 | SE017, SE018 |
| CE016 | Robot AI systems deployed in physical manufacturing environments may be classified as high-risk under the EU AI Act Annex III, requiring conformity assessment, human oversight mechanisms, and documentation before market deployment. | 中 | SE019, SE020 |
| CE017 | Physical Intelligence has not disclosed training data provenance, copyright status, or licensing agreements for robot demonstration videos used in π₀ training, creating potential IP liability. | 中 | SE001, SE016 |
| CE018 | Physical Intelligence's enterprise pilots span manufacturing and logistics use cases including package sorting, assembly automation, and general-purpose manipulation; no named customers have been disclosed. | 中 | SE024, SE002 |
| CE019 | Physical Intelligence has not announced any third-party safety audit, adversarial robustness testing, or cybersecurity penetration testing for π₀, which is a gap relative to enterprise deployment requirements. | 高 | SE017, SE025 |
| CE020 | Physical Intelligence's public roadmap has not been disclosed; the next major model release is analyst-inferred to be a π₁ or equivalent within 2026–2027 based on the pace of prior releases. | 低 | SE013, SE025 |
| CE021 | Commercial product launch with SaaS pricing is estimated for 2026–2027; Physical Intelligence has not announced pricing, commercial terms, or a launch date. | 低 | SE013, SE002 |
| CE022 | Physical Intelligence's π₀-FAST variant achieves faster single-pass action decoding, reducing inference latency for time-critical tasks with higher-frequency control loops compared to the base π₀ model. | 中 | SE004, SE013 |
| CE023 | The openpi GitHub repository has accumulated significant stars and external research forks as a proxy for developer community adoption, but exact metrics are not disclosed by Physical Intelligence. | 中 | SE005, SE006 |
| CE024 | Sergey Levine (Chief Scientist) and Chelsea Finn (co-founder) are among the most-cited robotics AI researchers globally, with combined citation counts in the hundreds of thousands across foundational robot learning papers. | 高 | SE021, SE022 |
| CE025 | Physical Intelligence has published at least three arXiv preprints covering π₀, π₀.5, and π₀-FAST, establishing a publicly reviewable technical record that both validates the approach and lowers replication barriers for competitors. | 高 | SE001, SE003, SE004 |
| CE026 | Google DeepMind Gemini Robotics 1.5 is a direct competitor to π₀ using Google's proprietary Gemini 2.0 VLM backbone, with the advantage of direct compute access and vertical integration that Physical Intelligence cannot match. | 中 | SE025, SE009 |
| CE027 | Physical Intelligence uses JAX and PyTorch as training frameworks, running on GPU clusters and TPUs; the specific cloud provider and compute contract terms have not been publicly disclosed. | 中 | SE001, SE015 |
| CE028 | π₀.5 extends π₀ with internet-scale pre-training using web-crawled video and vision-language data, providing broader semantic grounding for instruction following in unstructured task environments. | 中 | SE003, SE014 |
| CE029 | The on-robot inference architecture for π₀ is not publicly documented; Physical Intelligence has not disclosed whether inference runs on-device, on an edge server, or via cloud API during enterprise deployment. | 低 | |
| CE030 | Physical Intelligence's openpi fine-tuning utilities allow external users to adapt π₀ weights to new robot types and tasks, with community-reported successful fine-tuning on standard academic robot platforms. | 中 | SE005, SE006 |
| CE031 | The arXiv preprint for π₀ discloses the technical architecture in sufficient detail that a well-resourced competitor with comparable compute and data could attempt to replicate the approach within 12–24 months. | 中 | SE001, SE025 |
| CE032 | Physical Intelligence's cross-embodiment data advantage is a function of its hardware partner relationships and proprietary robot teleoperation dataset; competitors must build similar partnerships to replicate this data moat. | 中 | SE016, SE015 |
| CE033 | The robot AI field lacks standardized safety benchmarks equivalent to ISO 9001 for software; safety certification for robot AI will require novel evaluation methodologies that do not yet exist at regulatory level. | 中 | SE017, SE019 |
| CE034 | Physical Intelligence's π₀-FAST paper introduces a modified action chunking approach that reduces token generation steps per control cycle, achieving sub-100ms inference latency on reference hardware. | 中 | SE004 |
| CE035 | Enterprise robot customers in manufacturing require 99.9% uptime and documented mean time between failure metrics; Physical Intelligence has not published reliability or uptime specifications for π₀ in production settings. | 中 | SE018, SE017 |
| CU001 | Physical Intelligence has no commercial customers and zero ARR as of Q1 2026; the company is in the enterprise pilot phase with undisclosed customers in manufacturing and logistics. | 高 | SU001, SU002 |
| CU002 | Physical Intelligence's enterprise pilot programs target large manufacturing and logistics operators with high robot fleet density, positioning it to capture large per-account ACV from fleet-scale deployments. | 中 | SU001, SU022 |
| CU003 | Expected customer segments for Physical Intelligence include large-scale manufacturing (automotive, electronics), logistics warehousing, food service operators, and robot OEM partners as an indirect B2B2B channel. | 低 | SU002, SU025 |
| CU004 | The estimated ACV for a large manufacturing customer deploying 100–1,000 robots at $5,000–$15,000 per robot per year ranges from $500,000 to $15 million, making Physical Intelligence's ICP among the highest-ACV enterprise robotics software plays. | 低 | SU025, SU009 |
| CU005 | Physical Intelligence's B2B2B opportunity via robot OEM partners (who co-contributed embodiment training data) could serve as an indirect distribution channel, with OEMs bundling PI software on their hardware. | 低 | SU025, SU018 |
| CU006 | AgiBot, a Chinese robot manufacturer, has been named in press reporting as an early pilot partner and cross-embodiment data contributor to Physical Intelligence. | 中 | SU003, SU004, SU024 |
| CU007 | Longcheer Technology, an electronics manufacturer, has been named in press reporting as an early pilot partner for Physical Intelligence's robot AI software. | 中 | SU003, SU004, SU005 |
| CU008 | No outcome data, production deployment confirmation, or direct customer quotes are available for AgiBot or Longcheer's use of Physical Intelligence's π₀ system. | 高 | SU004, SU023 |
| CU009 | The openpi GitHub repository is a proxy for community adoption; significant star count and active external fine-tuning activity indicate developer interest but not commercial customer traction. | 中 | SU006, SU007 |
| CU010 | Industry benchmarks for enterprise robotics software suggest pilot-to-commercial conversion rates of 20–40%; Physical Intelligence's conversion rate from its pilot programs is unknown. | 中 | SU008, SU009 |
| CU011 | Structural switching costs for robot AI software are high because task-specific fine-tuning data accumulates over time, equivalent to losing months of proprietary training investment if a customer switches vendors. | 中 | SU014, SU015 |
| CU012 | Projected gross revenue retention for Physical Intelligence, based on structural SaaS switching cost analysis, is above 90% once commercial deployment is established, though no actual retention data exists. | 低 | SU014, SU015 |
| CU013 | Physical Intelligence's land-and-expand model targets expanding from initial robot deployments in one facility to additional facilities and robot types within the same enterprise account, generating 2–5× expansion ACV uplift. | 低 | SU001, SU025 |
| CU014 | When Physical Intelligence's first commercial customers are signed, top-customer revenue concentration is expected to exceed 80% of initial ARR due to the small number of accounts that will be closed in the early commercial phase. | 中 | SU012, SU013 |
| CU015 | Physical Intelligence has not publicly confirmed signed letters of intent, binding commercial agreements, or minimum purchase commitments from any enterprise customer as of Q1 2026. | 高 | SU002, SU023 |
| CU016 | CapitalG (Alphabet/Google) is both a Series B investor in Physical Intelligence and the corporate parent of Google DeepMind, a direct competitor, creating a potential conflict of interest that enterprise customers may scrutinize. | 中 | SU018, SU019 |
| CU017 | Manufacturing enterprise buyers cite safety certification, ROI demonstrability, and integration complexity as the top three barriers to adopting AI robotics software; Physical Intelligence has not yet addressed any of these definitively. | 中 | SU020, SU021 |
| CU018 | Physical Intelligence has no NRR, GRR, contract renewal, or churn data because it has no commercial customers; all retention metrics are forward-looking projections. | 高 | SU001, SU012 |
| CU019 | Geographic concentration risk is emerging in Asia based on the AgiBot and Longcheer pilot partnerships; Physical Intelligence's ability to win US and European enterprise accounts has not been demonstrated. | 低 | SU003, SU019 |
| CU020 | Skild AI has achieved approximately $30M ARR with enterprise manufacturing and logistics customers, representing a significant first-mover advantage over Physical Intelligence in commercial customer acquisition. | 中 | SU016, SU017 |
| CU021 | Enterprise robot deployment timelines from initial pilot to full production scale typically exceed 18 months due to safety integration, staff training, and production scheduling system changes. | 中 | SU020, SU010 |
| CU022 | Key diligence items on customer traction for a Physical Intelligence data room include the full pilot customer list with named references, pilot terms and conversion timelines, signed LOIs, and the names of initial commercial customers. | 高 | SU012, SU009 |
| CU023 | The enterprise B2B2B model via robot OEM partners is a potential high-leverage distribution strategy, but no OEM partnership agreement terms, exclusivity provisions, or revenue share arrangements have been disclosed. | 中 | SU025, SU003 |
| CU024 | Physical Intelligence's demo videos of π₀ folding laundry, loading dishwashers, and handling packages have received significant technology media coverage, indicating strong awareness among potential enterprise buyers without converting to sales pipeline visibility. | 高 | SU023, SU002 |
| CU025 | The broader robot AI market adoption rate is accelerating in 2025–2026, but the pace is constrained by safety certification requirements, budget cycles, and integration complexity, suggesting Physical Intelligence's commercial launch will face a 12–24 month enterprise sales cycle even after pricing is announced. | 中 | SU009, SU021 |
| CU026 | Physical Intelligence's pilot programs do not currently generate ARR or recognized revenue; even signed pilot agreements would likely be classified as deferred revenue or pilot fees below commercial pricing thresholds. | 中 | SU001, SU012 |
| CU027 | The minimum fleet size for Physical Intelligence's SaaS model to be cost-effective for an enterprise customer depends on the per-robot fee and deployment cost; at $10,000 per robot, a minimum of 10–20 robots per deployment is needed to justify the integration investment. | 低 | SU025, SU008 |
| CU028 | Enterprise customers who commit to robot AI software early (before safety certification is complete) may require price concessions or risk-sharing arrangements that would reduce initial ACV below the $10,000 per robot benchmark. | 中 | SU020, SU010 |
| CU029 | Physical Intelligence's conversion plan from enterprise pilot to commercial contract has not been publicly described; this is a critical gap for any VC conducting commercial diligence. | 低 | |
| CU030 | Adoption barriers cited by manufacturing enterprise buyers include lack of standardized safety testing for AI robot models, unclear liability in case of AI-driven accidents, and integration with existing MES/ERP systems. | 中 | SU021, SU010 |
| CU031 | Physical Intelligence's media coverage has reached the mainstream technology press (Wired, The Verge, MIT Technology Review), indicating a level of market awareness disproportionate to its commercial stage and validating interest among enterprise technology buyers. | 高 | SU023, SU007 |
| CU032 | If the B2B2B OEM channel is developed, Physical Intelligence could capture revenue from robot deployments without direct enterprise sales investment, but OEM partners would extract margin (typically 20–40% of license revenue) in exchange for distribution. | 低 | SU025, SU018 |
| CU033 | CapitalG's investment is consistent with Alphabet's strategy of seeding AI infrastructure startups that may become Google Cloud customers or provide proprietary data that benefits Google's robotics research program. | 中 | SU018, SU019 |
| CU034 | Physical Intelligence's early Asian enterprise relationships (AgiBot, Longcheer) reflect the reality that Chinese robot manufacturers are among the most aggressive early adopters of robot AI software, driven by labor cost pressure and government industrial AI mandates. | 中 | SU003, SU019 |
| CU035 | The absence of a named US or European enterprise customer reference is a significant risk for Physical Intelligence's valuation narrative, as Western VC markets and potential acquirers weight US/EU commercial traction more heavily than Asian pilot activity. | 中 | SU012, SU013 |
| CR001 | Robot AI systems deployed in physical manufacturing and logistics environments are likely classified as high-risk under EU AI Act Annex III, requiring conformity assessment, human oversight mechanisms, and technical documentation before EU commercial deployment. | 高 | SR001, SR002 |
| CR002 | Physical Intelligence has not disclosed any progress toward ISO 13849 or IEC 62061 functional safety certification, which is typically required for robot software deployed in enterprise manufacturing environments with human-robot collaboration. | 高 | SR003, SR004 |
| CR003 | Physical Intelligence has not disclosed training data provenance, copyright status, or licensing agreements for robot demonstration videos used in π₀ training, creating potential IP liability if third-party content was used without proper authorization. | 中 | SR005, SR006 |
| CR004 | No active litigation proceedings, regulatory enforcement actions, or material legal disputes against Physical Intelligence have been found in public court records or regulatory filings as of Q1 2026. | 高 | SR030, SR004 |
| CR005 | US OSHA workplace safety regulations and the NIST AI Risk Management Framework apply to AI robot systems in manufacturing environments, but current US regulatory requirements are less prescriptive than EU law and do not block near-term domestic deployment. | 中 | SR023, SR024 |
| CR006 | Robot AI models consistently show a 20–40 percentage point performance degradation from controlled laboratory settings to production deployment in unstructured manufacturing environments; this lab-to-production gap is Physical Intelligence's most acute near-term operational risk. | 高 | SR007, SR008 |
| CR007 | Skild AI has achieved approximately $30M ARR with enterprise manufacturing and logistics customers, creating a commercial data flywheel advantage over Physical Intelligence that widens each quarter as Skild accumulates more production deployment data. | 中 | SR009, SR010 |
| CR008 | Google DeepMind's Gemini Robotics 1.5 poses a compute and distribution asymmetry threat to Physical Intelligence: Google has effectively unlimited TPU/GPU compute, direct access to PaliGemma, and enterprise distribution via Google Cloud and Google Workspace. | 高 | SR011, SR012 |
| CR009 | Physical Intelligence's open-source release of openpi and publication of the π₀ arXiv preprint exposes the core technical approach to replication by well-resourced competitors with access to comparable compute and robot demonstration data. | 中 | SR013, SR014 |
| CR010 | If Physical Intelligence does not convert at least one enterprise pilot to a commercial contract with recognized ARR by Q4 2026, the fundraising conditions for the reported $11B next round become very difficult to achieve. | 高 | SR028, SR029 |
| CR011 | A failed $11B round would leave Physical Intelligence with its $600M Series B as the primary capital base; at estimated annual burn of $70–$150M, a fundraising failure could force a down-round, burn reduction, or pivot, each with material equity and talent consequences. | 中 | SR028, SR029 |
| CR012 | Sergey Levine is the primary public face of Physical Intelligence's technical credibility; his departure would materially impair investor confidence and enterprise customer trust in the company's technical leadership. | 中 | SR019, SR020 |
| CR013 | Chelsea Finn (MAML co-inventor) and Karol Hausman (CEO) represent concentrated key-person risk; all three senior founders have strong academic and industry affiliations that could attract them away from the company. | 中 | SR019, SR020 |
| CR014 | PaliGemma is distributed under Google's Gemma Terms of Use, which allow commercial use but can be revised unilaterally by Google; Physical Intelligence has no disclosed contractual protection against a future licensing restriction. | 高 | SR015, SR016 |
| CR015 | CapitalG (Alphabet) led Physical Intelligence's Series B while Google DeepMind competes directly with Gemini Robotics; Physical Intelligence's core architecture depends on PaliGemma (Google DeepMind), creating a tripartite investor-dependency-competitor relationship with Google. | 高 | SR017, SR018 |
| CR016 | Lux Capital participated in Physical Intelligence's seed, Series A, and Series B rounds, creating significant board and cap table influence; no co-investor in all three rounds has been identified, making Lux the primary capital continuity anchor. | 中 | SR017, SR029 |
| CR017 | The estimated annual burn rate of $70–$150M at Physical Intelligence creates a financing dependency within 24–36 months if the company scales headcount and compute for commercial launch, even with the $600M Series B as a base. | 低 | SR028, SR029 |
| CR018 | Robot OEM partners (AgiBot, Longcheer, others) could switch to competing AI software vendors (Skild AI, Google DeepMind) at any time; Physical Intelligence has no disclosed long-term exclusivity agreements with any hardware partner. | 中 | SR009, SR017 |
| CR019 | Physical Intelligence has not publicly disclosed key-person insurance, founder vesting schedules, retention bonuses, or co-founder departure protocols for its six-person founding team. | 高 | SR019, SR020 |
| CR020 | Six co-founders is above the median for Series B AI startups; governance complexity and decision-making speed risk increase with co-founder count; Physical Intelligence has not disclosed board composition or co-founder voting arrangements. | 中 | SR019, SR020 |
| CR021 | Kill criteria for the Physical Intelligence investment thesis include: no commercial revenue by Q4 2026, Skild AI achieving $100M ARR before PI has any revenue, Google releasing a commercial Gemini Robotics API with GCP distribution, or Sergey Levine's departure. | 中 | SR009, SR011 |
| CR022 | The most important monitoring indicator for Physical Intelligence's thesis is the pilot-to-commercial conversion rate; a conversion rate below 20% of active pilots within 18 months is a material thesis-break signal. | 中 | SR028, SR009 |
| CR023 | Diligence asks required for a Physical Intelligence investment include: full pilot customer list with named contacts, at least one customer reference call, pilot agreement terms and conversion timeline, PaliGemma commercial license documentation, training data provenance audit, and key-person retention terms. | 高 | SR005, SR015 |
| CR024 | Figure AI's BMW partnership and $39B valuation represents an alternative robot AI architecture (full-stack hardware+software vs. PI's software-only approach) that could attract enterprise customers who prefer a single vendor for hardware and AI. | 中 | SR021, SR022 |
| CR025 | Amazon's acquisition of Covariant gives Amazon proprietary robot AI capabilities for its logistics network, making Amazon a less likely Physical Intelligence enterprise customer and more likely an internal competitor in logistics. | 中 | SR025 |
| CR026 | Robot AI systems are susceptible to adversarial inputs through both visual (camera manipulation) and language (prompt injection) attack vectors in industrial settings; Physical Intelligence has not disclosed adversarial robustness testing for π₀. | 中 | SR023, SR024 |
| CR027 | If a robot accident or injury occurs during an enterprise pilot using π₀, Physical Intelligence faces potential liability, reputational damage, and the loss of the pilot customer without a clear product liability framework in place. | 中 | SR003, SR007 |
| CR028 | Physical Intelligence has not announced a senior enterprise sales leader (CRO or VP Sales) as of Q1 2026; the absence of commercial leadership is a material execution risk for the planned 2026–2027 commercial launch. | 中 | SR019, SR028 |
| CR029 | Large enterprise manufacturing customers (Toyota, Foxconn, Amazon, Bosch) have the internal engineering resources to build proprietary robot AI systems in-house, representing a significant build-versus-buy risk for Physical Intelligence's go-to-market. | 中 | SR026, SR027 |
| CR030 | Physical Intelligence's enterprise pilot terms have not been disclosed; the counterparty risk from pilot agreements (cost of failure if pilots do not convert) is unknown but materially affects the near-term financial outlook. | 低 | SR028, SR005 |
| CR031 | AI researcher talent in robot foundation models is highly concentrated; Skild AI, Google DeepMind, and Figure AI are all actively recruiting from the same small pool of researchers, creating ongoing talent retention pressure on Physical Intelligence. | 中 | SR019, SR009 |
| CR032 | Physical Intelligence's risk mitigation disclosures are minimal; the company has not publicly described any formal risk management program, safety certification roadmap, IP protection strategy, or contingency plan for the PaliGemma dependency. | 高 | SR001, SR015 |
| CR033 | The NIST AI Risk Management Framework, while not mandatory in the US, represents best practice for AI systems in physical environments; compliance would improve Physical Intelligence's enterprise sales credibility with risk-averse Fortune 500 buyers. | 中 | SR024, SR023 |
| CR034 | OSHA's General Duty Clause (Section 5(a)(1)) requires employers to provide workplaces free from recognized hazards; an AI robot that injures a worker could expose Physical Intelligence's enterprise customers (and potentially PI itself) to OSHA enforcement. | 中 | SR023, SR004 |
| CR035 | Physical Intelligence's enterprise pilots in Asia (AgiBot, Longcheer) reduce the regulatory risk from EU AI Act in the near term but create geographic concentration risk and delay the accumulation of Western enterprise references needed for US/EU commercial sales. | 中 | SR002, SR001 |
| CR036 | The combination of pre-revenue status, Google conflict of interest, and AI Act compliance burden creates a risk trifecta for EU/US institutional investors who require commercial proof before deploying capital at $10B+ valuations. | 中 | SR017, SR028 |
| CR037 | Physical Intelligence's six-co-founder structure is governance-dense; without clear decision authority protocols and a neutral lead director, strategic disagreements could slow commercial execution at the most critical juncture. | 低 | SR019, SR020 |
| CR038 | The training data IP risk is elevated for Physical Intelligence specifically because robot demonstration video data often involves capture in third-party facilities (customer factories) using proprietary manufacturing processes; the IP ownership of such video content is legally complex. | 中 | SR005, SR006 |
| CR039 | The enterprise in-house build option is constrained by the specialization required (VLA architecture, cross-embodiment training pipelines, flow matching); only very large technology companies have the resources to replicate PI's approach, reducing the build-vs-buy threat from mid-market manufacturers. | 中 | SR026, SR027 |
| CR040 | Taken together, Physical Intelligence's risk profile is unusually severe for a Series B company: zero revenue, pre-certification, Google investor-competitor conflict, and key-person concentration all need to be addressed before the $11B next round can be fully justified. | 高 | SR017, SR028 |
| CV001 | Physical Intelligence closed its Series B in November 2025 at a $5.6 billion post-money valuation on $600 million raised, with zero commercial revenue at closing. | 高 | SV001, SV002, SV013 |
| CV002 | There is no precedent in the robotics AI sector for a company achieving a $5.6 billion valuation at zero ARR within 20 months of founding; Physical Intelligence represents an extreme pre-revenue premium. | 中 | SV005, SV006 |
| CV003 | At a $5.6 billion valuation and a 30× forward revenue multiple, Physical Intelligence would need $187 million in ARR to justify the current entry price; no robotics AI company has achieved this within 3 years of founding. | 中 | SV011, SV012 |
| CV004 | The reported next-round valuation of $11 billion (April 2026, unconfirmed) would represent a 2× step-up from the Series B in under six months, which is credible only if Physical Intelligence has signed enterprise LOIs or announced a material commercial partner. | 低 | SV003, SV004 |
| CV005 | Physical Intelligence's valuation is justified by team quality, technical execution, and market size expectations ($170B+ service robot market by 2030), not by current commercial proof. | 中 | SV015, SV016 |
| CV006 | The most relevant comparable for Physical Intelligence is Skild AI at $14 billion on $30 million ARR (467× revenue multiple); Physical Intelligence at $5.6 billion on zero revenue implies an even more extreme multiple. | 中 | SV005, SV006 |
| CV007 | The bull case for Physical Intelligence requires $200–400 million ARR by 2028; at 20–30× forward revenue, this would support an $8–15 billion valuation, providing modest positive returns to Series B investors. | 低 | SV011, SV012 |
| CV008 | The base case (50% probability) for Physical Intelligence involves $30–100 million ARR by 2028; at 20–50× forward revenue, this would support $600 million–$5 billion valuation, representing a loss relative to the $5.6 billion Series B entry. | 中 | SV011, SV019 |
| CV009 | The bear case (30% probability) involves zero revenue by 2027 and a down-round or strategic acquisition below $5.6 billion, representing a major loss for Series B investors. | 中 | SV019, SV020 |
| CV010 | Cohere at $5.1 billion with $240 million ARR (21× revenue multiple) is a more favorable comparable than Physical Intelligence; Cohere has 8× Physical Intelligence's revenue at approximately the same valuation. | 高 | SV021, SV022 |
| CV011 | Our recommendation is CAUTION — PASS or WATCH at current valuation; the technology is credible and the team is exceptional, but the $5.6B valuation provides insufficient margin of safety with zero commercial proof. | 中 | SV019, SV020 |
| CV012 | Conditions for changing from PASS/WATCH to INVEST include at least $50M ARR from two or more named enterprise customers, one functional safety certification, PaliGemma commercial licensing documentation, and Levine retention confirmation. | 中 | SV011, SV019 |
| CV013 | Thesis-break triggers include no commercial revenue by Q4 2026, Skild AI surpassing $100M ARR, Google releasing a commercial Gemini Robotics API with GCP distribution, PaliGemma licensing restriction, or Levine's departure. | 中 | SV005, SV007 |
| CV014 | Physical Intelligence's preferred stock structure from three rounds likely includes liquidation preferences that could materially reduce common equity value in an exit below $5.6 billion; specific terms are not disclosed. | 低 | SV029, SV030 |
| CV015 | Final diligence asks required before investment include the full pilot customer list, PaliGemma commercial license, training data provenance audit, founder vesting schedules, board composition documents, actual burn rate, and a commercial launch timeline. | 高 | SV029, SV013 |
| CV016 | A comparable-based fair value range of $1.5 billion–$4.0 billion is supported by the base case ARR scenario and prevailing AI SaaS revenue multiples, implying 30–73% valuation risk at the $5.6 billion Series B entry. | 低 | SV011, SV012 |
| CV017 | Figure AI's $39 billion valuation is the highest in the robot AI sector but includes full-stack hardware plus software revenue (BMW deployment); this is not a clean comparable for Physical Intelligence's software-only model. | 中 | SV007, SV008 |
| CV018 | Physical Intelligence's SaaS gross margin target of 70–85% at scale is consistent with AI software benchmarks and would support a premium multiple if achieved, but the per-robot inference cost may compress margins below 70%. | 低 | SV027, SV028 |
| CV019 | Potential strategic acquirers for Physical Intelligence include Amazon, Microsoft, Samsung, Hyundai, and major industrial conglomerates (Bosch, ABB) seeking to add robot AI capabilities; acquisition at $3–10 billion is plausible if commercial traction is demonstrated. | 中 | SV023, SV024 |
| CV020 | IPO readiness for Physical Intelligence would require $300M+ ARR and positive gross margin at scale; on an optimistic trajectory, this is a 2029+ event, implying a 3–4 year hold from a 2026 investment. | 低 | SV025, SV026 |
| CV021 | The Inflection AI outcome is the cautionary comparable for Physical Intelligence — Inflection raised $4 billion at pre-revenue valuation and was effectively dissolved with its team joining Microsoft, implying zero return for common equity holders. | 高 | SV009, SV010 |
| CV022 | A watch-and-invest strategy at $50M ARR entry would cost more per share than the current $5.6B entry but would have dramatically lower binary risk, as the commercial model would be proven before capital is deployed. | 中 | SV019, SV011 |
| CV023 | Physical Intelligence's SEC Form D filings for Series A and B confirm offering sizes of $400M and $600M respectively; valuation is disclosed in press but not in the Form D itself. | 高 | SV013, SV014 |
| CV024 | Robot AI market valuation premiums (Skild AI at 467× revenue, Figure AI at ~780× revenue) are significantly higher than LLM SaaS peers (OpenAI at 80×, Cohere at 21×), reflecting investor belief in winner-take-most robot foundation model dynamics. | 中 | SV005, SV017 |
| CV025 | Physical Intelligence's management has not publicly addressed the valuation-versus-commercial-stage tension; all investor communications emphasize technical achievement and market opportunity without revenue guidance. | 中 | SV001, SV002 |
| CV026 | The probability-weighted expected value of a $5.6B entry into Physical Intelligence — 20% bull at $8–15B exit, 50% base at $600M–$5B exit, 30% bear at $500M–$1.5B exit — implies a negative expected return relative to entry price. | 低 | SV019, SV011 |
| CV027 | If the $11B next round does not close, Physical Intelligence must rely on the $600M Series B as its primary capital base; at $70–$150M estimated burn, this provides 4–8 years pre-commercial runway but compresses to 2–3 years at commercial scale-up. | 低 | SV019, SV003 |
| CV028 | Public market AI SaaS companies with AI-native products are trading at 15–25× forward revenue in 2025; private AI startups command 50–100× forward revenue premiums due to option value; Physical Intelligence's implied multiple (∞) is outside all reference ranges. | 中 | SV011, SV012 |
| CV029 | The CapitalG investment in Physical Intelligence creates a governance conflict that any new investor must resolve contractually — specifically, information-sharing protections preventing Google DeepMind from accessing Physical Intelligence's proprietary training data via board representation. | 中 | SV029, SV030 |
| CV030 | PaliGemma licensing risk should be treated as a binary valuation modifier in any investment model — if Google restricts commercial use, Physical Intelligence's model must be rebuilt at $10M+ cost and 6–12 months delay, effectively resetting the commercial launch timeline. | 中 | SV029, SV009 |
| CV031 | Physical Intelligence's three-round preferred stock structure (seed, Series A, Series B) creates a liquidation waterfall; in an exit below $5.6B, common equity holders (founders and early employees) receive nothing until preferred stock is repaid with any applicable multiples. | 中 | SV029, SV030 |
| CV032 | The robot foundation model market, analogous to the LLM foundation model market in 2020–2022, is likely to consolidate to 2–3 dominant platforms by 2030; Physical Intelligence needs to be in the top 2 to justify its current valuation. | 中 | SV015, SV016 |
| CV033 | A Series B investor at $5.6B needs Physical Intelligence to exit at $15–28B to achieve a 3–5× multiple of invested capital, which requires either an IPO at $300M+ ARR or a strategic acquisition by Amazon, Microsoft, or Samsung at full market price. | 低 | SV023, SV025 |
| CV034 | The expected hold period for a Physical Intelligence Series B investor is 5–7 years minimum to reach a meaningful exit event, with 7–10 years to IPO readiness on an optimistic trajectory. | 低 | SV025, SV026 |
| CV035 | Service robot market forecasts of $170B by 2030 from BCG and comparable analysts are widely cited but have historically been overoptimistic by 50–100%; investors should apply a 50% discount when using these figures to underwrite Physical Intelligence. | 中 | SV015, SV016 |
| CV036 | Any new investor in Physical Intelligence's next round at $11B should require a detailed conversion plan from the company showing named customer LOIs, pilot conversion timelines, and pricing commitments before closing — without these, the $11B valuation cannot be underwritten. | 高 | SV003, SV019 |
| CV037 | OpenAI's $80B valuation at $1B ARR (80× forward revenue) is the most favorable comparable for Physical Intelligence if it achieves market-leading robot foundation model position; this benchmark supports a $5.6B valuation only if Physical Intelligence achieves $70M ARR. | 低 | SV017, SV018 |
| CV038 | Physical Intelligence has confirmed Series B funding of $600M via SEC Form D; the $5.6B post-money valuation is reported in press but not in the Form D, which only confirms offering size. | 高 | SV013, SV014 |
| CV039 | An entry at $11B (reported next round) would require Physical Intelligence to exit at $33–55B for a 3–5× return, which implies either a dominant LLM-equivalent position in robot AI or an exceptional M&A outcome. | 低 | SV003, SV023 |
| CV040 | The investment thesis for Physical Intelligence rests on a winner-take-most robot foundation model market; the anti-thesis is that the market fragments between Google (Gemini Robotics), Skild AI, and Physical Intelligence, diluting the platform premium that justifies $5.6B+ valuations. | 中 | SV019, SV015 |