Advanced Machine Intelligence
种子期物理 AI 实验室,团队和资本都异常强,但公开产品和客户证据仍有限。
AMI 是资本充足度异常高的实体 AI 实验室,领导层顶级;但公开证据支持继续研究而不是买入,因为产品、客户、经济性与治理证明仍落后于种子轮估值。
封面要素
公司概况
Advanced Machine Intelligence(公开品牌为 AMI Labs)是一家总部在巴黎的前沿 AI 实验室,面向医疗、工业自动化、机器人、可穿戴设备及相邻领域的安全关键物理世界流程开发世界模型。公开记录能支撑几件事:2025 年 12 月在法国注册、2026 年 3 月公开融资启动、围绕 Yann LeCun 和 Alexandre LeBrun 的深厚创始领导层,以及 Nabla 首个优先接入合作;但尚未证明 AMI 已交付产品、形成收入基础或拥有付费客户名单。
- 成立时间
- 2025-12-15
- 创始人
- Yann LeCun, Alexandre LeBrun, Laurent Solly, Saining Xie, Pascale Fung, Michael Rabbat
- 创立地点
- Paris, France
- 总部
- 10 rue de Penthièvre, 75008 Paris, France
- 产品
- AMI 尚未披露可销售 SKU;公司在开发世界模型系统,用于学习真实世界传感器表征、预测后果,并在护栏下支持受控规划。
- 客户
- 医疗、工业过程控制、自动化、可穿戴设备和机器人等受监管、安全关键运营方;Nabla 是首个具名的医疗合作伙伴渠道。
- 商业模式
- 尚未披露;公开证据指向未来可能的模型授权、API 访问、垂直合作伙伴部署、联合开发,以及通过合作伙伴推进医疗商业化。
- 阶段
- Seed; pre-product and pre-revenue in public evidence
- 融资情况
- 2026-03-10 公布 $1.03B 种子轮,据报投前估值 $3.5B;在未知结构前,隐含投后估值约 $4.53B。
执行摘要
主要优势
- Yann LeCun、Alexandre LeBrun 以及资深世界模型研究团队,给了公司创始人与研究履历背书。
- 种子轮资本规模异常充足,财务和战略投资人组合质量很高。
- 逆向的世界模型逻辑,契合医疗、机器人、工业自动化与可穿戴设备里的实体 AI 需求。
- Nabla 提供了早期战略合作伙伴通道,并能优先接入新兴 AMI 技术。
主要风险
- 没有公开证据显示公司已经交付产品、发布 AMI 专属 benchmark、拿下直接付费客户、形成 ARR 或签署商业合同。
- 算力、数据整理与资本强度可能在收入证据出现前吃掉种子轮资金。
- 围绕 LeCun、Meta/NYU、LeBrun 与 Nabla 的关键人、治理、IP 和关联方依赖仍未解决。
- 医疗和其他安全关键目标领域带来重大的监管、可审计性与责任负担。
- 对一个种子期、产品前实验室来说,估值偏贵;如果里程碑落后,估值可能压缩。
未决问题
- 融资条款、清算优先权、分批提款、期权池处理、投资人权利与有效进入价格。
- AMI 专属产品 benchmark、路线图里程碑、SKU/API 定义、部署接口与可靠性证据。
- Nabla 经济性、数据权利、授权条款、最低承诺、排他性与转让定价安排。
- 算力采购承诺、供应商集中度、月度烧钱速度、现金跑道,以及对 GPU 价格或容量冲击的敏感性。
- 直接客户证据、已签试点或合同、客户数、ARR、收入质量与定价模型。
- AMI、Meta、NYU、Nabla、开源承诺与监管领域安全控制之间的治理和 IP 边界。
目录
01公司概览
1.1 身份、阶段与产品论点
AMI 更应被视为一家新成立的法国前沿 AI 实验室,而不是已有成熟商业指标的软件运营商。法律锚点是 ADVANCED MACHINE INTELLIGENCE:这是一家巴黎 SASU,SIREN 994675254,2025-12-15 完成 RCS Paris 注册,现总部位于 10 rue de Penthièvre, 75008 Paris。公司的公开入口是 amilabs.xyz;其使命声明称,AMI 正在构建世界模型,用于学习真实世界传感器数据的抽象表征,并在安全护栏下规划。这个身份与融资公告一致:AMI 还没有销售打包产品,而是用异常庞大的种子轮押注长周期研究、算力和人才。因此,核心尽调框架是阶段错配。资本规模像后期公司,业务仍处在产品前、收入前阶段,并依赖一套非共识技术架构落地到可靠性要求很高的应用。[CO001, CO002, CO003, CO004, CO005, CO016]
| 指标 | 数值 / 状态 | 日期 / 版本 | 置信度 | 缺口 / 尽调备注 |
|---|---|---|---|---|
| 法律实体 | ADVANCED MACHINE INTELLIGENCE;SASU 法人;SIREN 994675254 | 2026-07-03 注册记录核查 | 高 | 从官方文件核验完整章程、受益所有权以及种子轮后的任何资本变更。 |
| 总部 | 10 rue de Penthièvre, 75008 Paris 注册地址 | 主要机构创建于 2026-02-12 | 中 | 注册地址清楚,但运营足迹和租约细节未公开。 |
| 官方网站域名 | amilabs.xyz | 2026-07-03 抓取 | 高 | Pappers 未列出网站;域名根据官网和相互链接的报道推断。 |
| 最新轮次 | $1.03B / ~€890M 种子轮 | 2026-03-10 宣布 | 高 | 轮次结构、清算优先权和二级交易部分未披露。 |
| 估值 | $3.5B 投前;~$4.5B 隐含投后 | 2026-03-10 | 高 | 投后估值是算术推算,不是发行方披露数。 |
| 收入 / ARR | 2026-07-03 | 低 | 未找到公开产品收入、ARR 或商业定价证据。 | |
| 具名客户 / 合作伙伴 | Nabla 是首个具名战略合作伙伴 | 2025-12-18 / 2026-03-10 | 中 | STAT 称尚未披露正式股权或授权协议。 |
| 员工人数 | 据报道近期计划招聘 20-30 人;当前总人数未披露 | 2026-03-10 | 中 | Ashby 页面确认招聘漏斗,但不确认员工人数或已填补岗位。 |
| 地点 | Paris、New York、Montreal 与 Singapore | 2026-03-10 | 高 | 需要办公室级别人员配置、法律分支和当地雇佣实体。 |
空值单元格表示公开指标没有支撑,不代表为零。投后估值由已披露融资金额加上所报道投前估值估算得出。
[CO001, CO002, CO003, CO004, CO008, CO016]AMI 的可投资性取决于:可信研究团队能否把世界模型科学和战略伙伴转化为经过验证的安全关键部署。
流程是尽调逻辑图,不是运营流程图。
[CO004, CO005, CO010, CO016, CO021, CO024]轮次和估值很具体,商业牵引与运营指标大多仍未披露。
隐含投后估值为已披露轮次金额加上已报道投前估值。空 KPI 值代表公开指标缺乏支撑,不是零。
[CO016, CO017, CO024, CO032, CO033, CO045]1.2 领导层、治理与关键人物风险
以公司的年龄看,领导班底罕见地可信,但公开证据也把故事压缩在少数个人身上。AMI 称 Yann LeCun 担任董事长,Alexandre LeBrun 担任 CEO,Laurent Solly 担任 COO,Saining Xie 是首席科学官,Pascale Fung 是首席研究与创新官,Michael Rabbat 负责世界模型。这套组合把 LeCun 的 JEPA / 世界模型论点、LeBrun 通过 Wit.ai 和 Nabla 积累的公司建设经验,以及多位 Meta/FAIR 相关的研究和运营履历放在一起。人才确定性强于治理确定性:不同来源对 LeCun 的表述在执行董事长与非执行董事长之间摇摆,公开资料中也没有出现董事会、投票控制或投资人权利安排。后续章节因此应把本章角色表当作具名领导者地图,而不是完整治理深度的证明。最高优先级尽调问题是:公司能否把决策机制制度化,不只依赖 LeCun 的科学权威和 LeBrun 的运营角色。[CO010, CO011, CO012, CO013, CO014, CO015]
| 人物 | 公开职务 | 相关背景 | 职能覆盖 | 关键人依赖 |
|---|---|---|---|---|
| Yann LeCun | 主席 / 创始人;不同来源头衔不一 | Turing Award 得主,前 Meta 首席 AI 科学家,NYU 教授,JEPA 倡导者 | 科学主线、投资人吸引力、开放研究可信度 | 极高——AMI 的叙事和估值高度依赖他的可信度。 |
| Alexandre LeBrun | CEO;转任后担任 Nabla 首席 AI 科学家兼董事长 | 连续 AI 创始人;创办 Wit.ai 和 Nabla;曾在 FAIR 与 LeCun 共事 | 公司搭建、医疗入口、合作伙伴落地 | 极高——负责把研究转成公司执行。 |
| Laurent Solly | COO | 前 Meta 欧洲副总裁 | 全球运营、公司规模化、欧洲生态关系 | 中——对研究色彩很重的创始团队是重要运营补位。 |
| Saining Xie | 首席科学官 | 视觉表征学习研究者,关联 NYU / Google DeepMind / Meta | 核心感知与表征学习研究 | 中高——对技术项目深度至关重要。 |
| Pascale Fung | 首席研究与创新官 | 以人为本 AI 教授,曾任高级 AI 研究负责人 | 研究议程、以人为本系统、外部可信度 | 中——把研究范围扩展到 LeCun 个人主线之外。 |
| Michael Rabbat | 世界模型副总裁 | 前 Meta/FAIR 研究负责人,常驻 Montreal | 世界模型研究领导与 Montreal 人才节点 | 中——直接对应 AMI 具名技术类别。 |
列举范围覆盖公开具名的创始领导班底;不代表完整董事会、投资人权利安排或完整员工名册。
[CO010, CO011, CO012, CO013, CO014, CO015]1.3 资本结构、投资人与战略期权价值
AMI 的 $1.03B 种子轮,是公司可投也高风险的同一个事实。公司、投资人和独立来源都指向约 €890M / $1.03B 融资,投前估值 $3.5B;扣除费用或任何未披露结构前,隐含投后约 $4.5B。本轮由 Cathay Innovation、Greycroft、Hiro Capital、HV Capital 和 Bezos Expeditions 共同领投,战略和长期支持方包括 NVIDIA、Samsung、Temasek、Toyota Ventures、Sea、SBVA 和 Alpha Intelligence Capital。这个财团不只是钱:硬件、工业、亚洲和医疗相邻资源,在 AMI 证明科学路径后可能转成部署通道。反向解读是,估值已经把多年研究突破提前计价,而 AMI 仍无收入、审计报表,也未披露 Nabla 之外的客户。Sequoia 对 AI 资本开支的批评、Reuters 来源关于 AI 估值泡沫的评论,并不是 AMI 专属失败证据,但足以成为承销十亿美元种子轮的护栏。[CO016, CO017, CO018, CO019, CO020, CO021]
| 利益相关方 | 角色 | 控制 / 经济重要性 | 已有证据的立场 | 尽调请求 |
|---|---|---|---|---|
| Cathay Innovation | 共同领投方 | 与 Nabla 有历史关系的财务投资方 | 共同领投本轮,并发布详细投资论点 | 确认持股比例、董事席位和后续跟投储备。 |
| Greycroft | 共同领投方 | 美国风险投资方 | 公司和新闻材料中具名共同领投 | 确认董事 / 观察员权利以及美国商业化角色。 |
| Hiro Capital | 共同领投方 | 欧洲风险投资方;资料报道提到 LeCun 顾问关联 | 本轮具名共同领投 | 检查任何顾问冲突和治理权利。 |
| HV Capital | 共同领投方 | 欧洲成长资本,并与既有 Nabla 投资人生态有交集 | 公司材料中具名共同领投 | 确认能否借力医疗 / 欧洲投资组合。 |
| Bezos Expeditions | 共同领投方 | 高知名度私人资本和信号价值 | 具名共同领投;Jeff Bezos 参与被重点提及 | 澄清除品牌信号外是否存在战略帮助。 |
| NVIDIA | 战略支持方 | 算力生态相关性和物理 AI 平台邻近性 | 具名长期战略支持方 | 确定算力承诺、云积分或商业义务。 |
| Samsung | 战略支持方 | 硬件 / 设备生态和亚洲分发相关性 | 具名参与方 | 确定是否有设备、传感器或端侧 AI 协作权利。 |
| Temasek | 战略 / 主权关联投资人 | 亚洲资本和 Singapore 运营相关性 | 具名参与方 | 确认持股、亚洲扩张支持和治理角色。 |
| Toyota Ventures | 战略支持方 | 机器人、出行和工业 AI 相关性 | 具名长期支持方 | 核实是否考虑汽车 / 机器人试点。 |
| Nabla | 首个具名战略合作伙伴 | 医疗验证路径,而非已披露收入 | 已宣布优先 / 首次访问;STAT 称未披露正式股权或授权协议 | 获取已签署协议、数据权利、定价、FDA / 监管计划和试点里程碑。 |
| 法国生态 / 公共信号 | 政治和生态支持 | 法国 AI 主权叙事,并强化 Paris 总部定位 | Macron 公开称赞 AMI 发布 | 区分非稀释性支持和象征性背书。 |
由于 AMI 未披露持股比例、董事席位、清算优先权或商业合同经济性,利益相关方重要性只能定性判断。
[CO018, CO019, CO024, CO025, CO043, CO044]1.4 合作伙伴、用例与商业化路径
Nabla 是第一个具体外部验证点,但应归为战略伙伴,而不是经常性收入证据。Nabla 宣布可优先或首批接入 AMI 正在形成的世界模型,并把合作放在临床流程上:LLM 在幻觉、非确定性、连续信号和可审计行动上容易失手。STAT 补上一条有用边界:两家公司合作紧密,但尚未披露正式股权或授权协议关系。医疗之外,AMI 和独立报道反复点名工业过程控制、自动化、可穿戴设备、机器人、自动驾驶、喷气发动机、发电厂和患者器官等目标领域。这些例子说明世界模型为何有价值,也说明验证会很难:最有价值的应用往往安全关键、受监管、数据密集。因此,商业化路径应先是伙伴驱动的实验,而不是立即推出 SaaS。[CO024, CO025, CO026, CO027, CO028, CO029]
AMI 约四个月内从法人成立推进到十亿美元级种子轮,但产品和收入证据仍是未来里程碑。
时间线只纳入公开里程碑;私下设立步骤、董事会批准和融资交割机制不可见。
[CO001, CO002, CO003, CO006, CO007, CO016]1.5 里程碑、反向检查与未解缺口
里程碑显示公司创建节奏很压缩:2025 年 12 月完成法律设立,Nabla 在 12 月宣布合作与 LeBrun 职位转换,1 月获得公开画像,2 月迁至当前巴黎地址,2026 年 3 月完成 $1.03B 融资。Pappers 上未出现法国登记制裁、诉讼或集体程序,但公开运营记录仍太短,这份干净登记快照的预测力有限。更重要的反向证据是战略层面而非法律层面:Le Monde Informatique 称 AMI 尚无可运行系统,Forbes 强调视频 / 传感器世界模型的成本和监管负担,更广泛的 AI 市场评论警示早期估值可能脱离收入。这些并不否定 AMI 的上行空间;它们定义下一步尽调路径。公司必须把资本和人才转化为公开研究、伙伴试点、安全证据,并最终转化为商业模式。[CO006, CO007, CO031, CO032, CO033, CO034]
| 日期 | 事件 | 类型 | 金额 / 估值 / 状态 | 参与方 | 含义 |
|---|---|---|---|---|---|
| 2025-12-15 | ADVANCED MACHINE INTELLIGENCE 在 RCS Paris 注册 | 创立 | SIREN 994675254 | 法国注册登记 | 公开融资叙事前,法律实体已经存在。 |
| 2025-12-18 | Nabla 宣布独家战略合作伙伴关系及 LeBrun 领导层过渡 | 合作 | 首次 / 优先访问;无公开定价 | Nabla、AMI、LeCun、LeBrun | 首个具名合作伙伴,也是最清晰的医疗验证路径。 |
| 2026-01-22 | MIT Technology Review 发布 LeCun 访谈,阐述 AMI 主线 | 治理 | 执行董事长措辞;Paris 总部,并计划布局北美 / 亚洲 | LeCun、MIT Technology Review | 公开把 AMI 定位为对 LLM 的逆向押注。 |
| 2026-01-23 | TechCrunch 报道 AMI 领导层和 Nabla 关联 | 治理 | CEO LeBrun、执行董事长表述不明、办公室 | TechCrunch、AMI 领导层 | 确认关键人叙事和早期合作伙伴结构。 |
| 2026-02-12 | Pappers 显示当前 Paris 机构在 10 rue de Penthièvre 创建 | 创立 | 主要机构处于存续 | 法国注册登记 | 锚定后续章节使用的总部地址。 |
| 2026-03-10 | AMI 宣布 $1.03B / ~€890M 种子轮融资 | 融资 | $3.5B 投前;~$4.5B 隐含投后 | 共同领投方和战略支持方 | 创造异常长的跑道,但估值负担很高。 |
| 2026-03-10 | AFP 联稿报道称近期计划招聘 20-30 人 | 扩张 | 近期招聘;当前员工数未披露 | AMI / AFP | 招聘已有计划,但实际员工数仍是缺口。 |
| 2026-03-10 | Macron 称赞 AMI 是法国 AI 里程碑 | 治理 | 公开政治背书 | 法国总统 Emmanuel Macron | 支撑法国主权叙事,但不代表商业牵引力。 |
| 2026-03-10 | 独立报道称该轮为欧洲最大种子轮 | 融资 | 留存报道称为欧洲最大种子轮 | TechCrunch、Crunchbase、TNW | 参照支持类别重要性,也提高估值审查压力。 |
| 2026-03-10 | 报道指出 AMI 尚无运营系统 / 尚无产品收入 | 反向 | 产品前、收入前状态 | Le Monde Informatique、TNW 与 TechCrunch | 这是承销种子轮估值的核心反向里程碑。 |
| 2026-07-03 | 注册记录核查在 Pappers 上未发现账目、制裁、诉讼或集体程序 | 监管 | 未列示账目;0 项程序 / 制裁 / 诉讼 | Pappers | 记录干净但法律历史很短;没有历史不等于执行质量有证据。 |
里程碑合并法律备案、合作伙伴公告、公司 / 投资人发布和独立报道。反向行被有意保留,因为运营准备度是当前主要风险。
[CO001, CO002, CO003, CO006, CO007, CO016]1.6 展项
02市场分析
2.1 市场边界与现状替代方案
AMI 的市场规模应放在物理 AI 和具身 AI 栈里衡量:模型从传感器流学习,模拟后果,规划行动,并支持安全部署到真实世界系统。这个边界包括机器人基础模型、自动驾驶和机器人仿真、工业数字孪生 / 模型层、医疗相邻的多模态流程智能,以及受监管部署所需的安全工具。它不包括通用聊天助手、横向企业 AI 助手、传统机器人硬件收入,也不包括完整自动驾驶车队经济,除非 AMI 因模型层获得付费。现状替代方案也很实在:制造商可以继续使用 PLC / MES / SCADA 自动化、经典计算机视觉、系统集成商、任务专用机器人、内部数据科学团队,或供应商绑定的仿真栈。因此,本章把广义具身 AI 市场数字当作上限背景,再收窄到 AMI 可能合理销售 API、模型、授权或联合开发权的软件 / 模型层机会。[CM001, CM002, CM003, CM004, CM010, CM011]
| 细分 / 类别 | AMI 论点纳入支出 | 排除支出 | 主要买方 / 付费方 | 相关性 |
|---|---|---|---|---|
| 机器人基础模型层 | 世界模型、VLA 策略、合成数据、评估、微调、安全护栏 | 机器人硬件、执行器、电池、安装服务 | 机器人 CTO、自主系统负责人、机器人工程 VP | 如果 AMI 销售模型授权或联合开发,这是最接近的类别。 |
| 工业自动化智能 | 传感器融合、异常预测、数字孪生推理、面向工厂和厂区的安全规划 | PLC/MES/SCADA 许可收入和完整系统集成人力 | COO、厂长、自动化总监 | 预算很大,但必须证明集成和运行时间。 |
| 自动驾驶出行 / AV 工具链 | 用于场景、仿真、评估和车队学习辅助的世界模型 | 整车制造、网约车车队收入、保险、地图硬件 | 自主系统平台负责人、OEM 软件买方 | 相关性来自仿真和规划,而不是完整 AV TAM。 |
| 医疗邻近多模态工作流 | 面向临床工作流上下文、监测、文档和未来具身照护支持的可审计模型 | 医疗服务报销、医院硬件、无关临床 SaaS | 临床平台负责人、合规官、服务提供方运营 | Nabla 展示了一条路径,但监管和责任会拖慢采用。 |
| 战略研究与创新 | 付费试点、联合研究、战略授权、投资人支持的概念验证 | 未披露且没有付费权利的内部 R&D | 企业风险投资、创新、AI 实验室领导层 | 是有用的切入楔子,但不足以证明收入可重复。 |
| 通用 AI 软件 / copilots | 除非与物理感知、仿真或安全动作绑定,否则不纳入 | 横向聊天机器人、办公 copilots、内容生成、CRM copilots | CIO / 业务应用负责人 | 排除在 AMI 市场之外,因为买方任务和验证负担不同。 |
边界表把 AMI 可能变现的软件 / 模型层,与更广义物理系统和通用 AI 软件分开;排除支出不计入收窄后的 SAM。
[CM001, CM002, CM003, CM010, CM032, CM033]广义具身 AI 预测只是上限背景;AMI 可承销机会是更窄的软件 / 模型层。
所有数值均为十亿美元。USD 23.06B 的广义层来自来源报道;USD 6.9B 和 USD 2.3B 是 TM002 所示 30% 和 10% 的四舍五入转换,不是独立分析师预测。
[CM009, CM033, CM034, CM035]2.2 规模测算视角与口径调和
证据支持多个测算视角,而不是单一 TAM。MarketsandMarkets 估计具身 AI 2025 年市场为 $4.44B,2030 年为 $23.06B,但它的产品边界包括机器人、外骨骼、自主系统和智能家电,远宽于 AMI 可能变现的层。IFR 的 2025 World Robotics 页面显示物理自动化需求按台数看很强——2024 年工业机器人安装量 542,000 台,专业服务机器人销量接近 200,000 台——但这不是软件收入池。Deloitte 的智能制造调查显示买方已准备投入 AI、数据、传感器和自动化,但也意味着采用会嵌在更大的转型项目里。调和后的市场判断是:物理系统大背景很大且在增长,具身 AI 市场估算中等规模,而软件 / 世界模型 SAM 要小得多,必须靠试点和定价证明。[CM009, CM012, CM013, CM014, CM015, CM016]
| 视角 / 发布方 | 年份或期限 | 地域 / 范围 | 价值或采用信号 | 方法 / CAGR | 置信度 | 局限 |
|---|---|---|---|---|---|---|
| MarketsandMarkets 具身 AI | 2025 | 全球具身 AI 产品 | USD 4.44B 市场规模 | 预测基准年;产品范围包括机器人、自主系统、外骨骼、家电 | 中 | 对 AMI 来说过宽,因为包含硬件和完整系统。 |
| MarketsandMarkets 具身 AI | 2030 | 全球具身 AI 产品 | USD 23.06B 市场预测 | 2025 至 2030 年 CAGR 39.0% | 中 | 上限语境,不是 AMI SAM。 |
| 狭义世界模型 / 模型层 SAM 估算 | 2030 | 具身 AI 软件 / 模型层子集 | USD 2.3B 至 USD 6.9B | MarketsandMarkets 2030 年具身 AI 预测值的 10% 至 30% | 低 | 分析师来源未拆分模型层收入;尽调必须核验定价。 |
| IFR World Robotics 2025 | 2024 | 全球工业机器人 | 安装 542,000 台工业机器人 | 年度安装量连续第四年超过 500,000 台 | 中 | 部署台数代理指标,不代表软件收入。 |
| IFR World Robotics 2025 | 2024 | 专业服务机器人 | 售出近 200,000 台;+9% | 服务机器人供应商样本 | 中 | 样本构成会变化,不能外推到整个行业。 |
| Deloitte 智能制造调查 | 2025 年报告 | 收入超过 $500M、员工超过 1,000 人的美国制造商 | 29% 在工厂或网络层面使用 AI/ML,24% 使用生成式 AI | 2024 年 8–9 月对 600 名高管的调查 | 中 | 衡量采用准备度,不是 AMI 可捕获支出。 |
| Bain 人形机器人 | 2024 年资本背景 | 人形机器人风险融资 | 约 USD 2.5B VC 投资 | Bain Technology Report 2025 | 中 | 资本形成信号;部署仍处早期。 |
| Wayve 具身 AI | 2024 | 自动驾驶基础模型 | USD 1.05B Series C 轮 | 公司新闻稿 | 中 | 单家公司融资显示资本兴趣,不代表市场收入。 |
主要图表数值与本表勾稽:FM001 和 FM002 使用 MarketsandMarkets 2030 年 USD 23.06B 上限, 以及推导出的 USD 2.3B-6.9B 收窄 SAM 区间;机器人台数行仅作为采用代理指标。
[CM009, CM012, CM014, CM017, CM018, CM028]可用市场区间从狭义软件层估算延伸到完整广义具身 AI 预测。
区间使用十亿美元,并与 TM002 对齐:低位=USD 23.06B 的 10%,中位=30%,高位=100% 广义具身 AI 背景。
[CM009, CM033, CM034]2.3 买方、用户与付款方分层
AMI 尚未打包成应用,买方地图因此很碎片化。制造和物流场景里,经济买方通常是负责产出、稼动率、安全和劳动生产率目标的运营、制造或自动化负责人;机器人工程师和系统集成商则是技术用户。机器人和 AV 公司里,买方可能是机器人工程、仿真或自动驾驶平台组织,需要更好的数据整理、合成场景、模型评估和跨形态泛化。医疗相邻流程里,临床平台负责人和风险 / 合规团队很重要,因为价值命题取决于可审计性和安全行动,不只是模型准确率。战略创新团队和企业风投部门可以资助早期试点,但除非试点转成运营、临床或工程预算,否则还不足以证明可重复的预算归属。[CM002, CM005, CM007, CM008, CM028, CM029]
| 细分市场 | 经济买方 | 技术用户 | 付款方 / 预算所有者 | 工作流 | 采用触发因素 |
|---|---|---|---|---|---|
| 工业制造与物流 | COO、工厂经理、自动化负责人 | 机器人工程师、控制工程师、集成商 | 运营卓越、资本开支、自动化预算 | 产能、质量、维护、排程、物料搬运 | 劳动力约束、回流生产、产出 / 产能 ROI、安全论证。 |
| 机器人 OEM 与机器人软件供应商 | 机器人 CTO 或产品 GM | 模型训练、仿真、控制、评估团队 | R&D 与平台工程预算 | 跨机器人本体的策略开发与验证 | 需要减少特定任务编程和数据采集。 |
| 自动驾驶车辆 / 出行 | 自动驾驶平台负责人、OEM 软件高管 | 仿真、感知、规划、验证团队 | OEM 软件、AV R&D、车队学习预算 | 场景生成、边缘案例搜索、辅助驾驶升级路径 | 需要扩展安全验证和无地图 / 基础模型自动驾驶。 |
| 医疗周边工作流平台 | 临床平台 CEO/CTO、首席医疗信息官 | 临床 AI、产品、合规、工作流团队 | 服务提供方运营、平台 R&D、合规预算 | 可审计的多模态工作流推理,以及未来行动支持 | 幻觉、非确定性、监控、受监管变更等顾虑。 |
| 战略创新与企业风投 | 首席创新官、企业 VC、AI 实验室负责人 | 应用 AI 研究员与试点团队 | 创新、战略合作或风投预算 | 探索性试点和数据共享合作 | 业务线 ROI 得到证明之前,为物理 AI 保留期权价值。 |
AMI 的可能市场进入路径因垂直行业而异,且可能先从试点开始,再由运营预算承接续约。因此本表区分经济买方、技术用户和付款方。
[CM002, CM005, CM007, CM028, CM036, CM037]各细分的差异不在于是否看好 AI,而在于哪道证明门槛控制生产预算。
矩阵是基于来源用例和调研证据推导的定性买方地图,不估算细分收入。
[CM002, CM028, CM036, CM037, CM038, CM039]2.4 增长驱动因素
最强市场驱动不是泛泛的 AI 热情,而是物理经济约束让更好的感知、仿真和行动模型变得有价值。工业机器人安装量已连续四年高于 500,000 台,中国继续提前拉动物理自动化需求,物流和医疗机器人等专业服务机器人品类也在扩大。制造商报告智能制造在产出、生产率和产能上带来收益,同时把流程自动化、传感器、视觉系统、数据分析和 AI 列为优先事项。基础模型进展是另一条驱动:NVIDIA Cosmos、Gemini Robotics、Genie 2、π0、Helix,以及 LeRobot 等开放工具,都指向一个从手工编程机器人走向数据驱动策略、合成环境和多模态行动模型的市场。只要 AMI 能把研究产出转成企业级集成点,它的世界模型论点就踩在时间窗口上。[CM004, CM005, CM006, CM007, CM008, CM012]
| 驱动因素 / 约束 | 方向 | 时间 | 含义 | 尽调追问 |
|---|---|---|---|---|
| 工业机器人部署基础 | 驱动因素 | 当前至 2028 年 | 已安装自动化设备为更好的模型创造数据和集成界面。 | 核实 AMI 可以接入哪些机器人 / OEM 技术栈。 |
| 劳动力短缺与生产率压力 | 驱动因素 | 当前 | 制造商和服务运营商需要自动化,但只有 ROI 可衡量的场景才会采用。 | 按工作流量化客户回本门槛。 |
| 仿真与合成数据 | 驱动因素 | 当前 | 世界模型可能降低数据整理、边缘案例生成和评估成本。 | 获取 AMI 试点证据,对比真实数据与合成数据表现。 |
| 多模态基础模型进展 | 驱动因素 | 当前 | VLA / 世界模型系统正从实验室验证走向合作伙伴预览和开放工具。 | 用 Cosmos、Gemini Robotics、π0、Helix 和开源基线对标 AMI。 |
| 监管与安全治理 | 约束因素 | 当前并在增强 | 安全关键部署需要风险管理、监控、透明度和变更控制。 | 将 AMI 目标用途映射到 EU AI Act、FDA 和买方治理义务。 |
| 算力与数据强度 | 约束因素 | 当前 | 训练和定制成本可能迫使 AMI 采用高定价或战略合作。 | 索取算力预算、数据权利、模型效率和毛利率计划。 |
| 与现有自动化技术栈集成 | 约束因素 | 当前 | 没有适配器和支持,买方不会为了研究模型拆掉可信 OT 系统。 | 验证集成、正常运行时间 SLA、网络安全模型和系统集成商渠道。 |
| 基准测试与 ROI 证明尚不成熟 | 约束因素 | 短期 | 通用演示不能证明在受监管或高差异运营中的价值。 | 要求围绕具体任务做试点,并交代基线、安全、成本和部署指标。 |
方向是作者基于引用证据作出的分类。时间判断取决于该因素已在公开来源中显现,还是仍有赖未来部署浪潮。
[CM004, CM005, CM006, CM007, CM012, CM015]物理 AI 部署从研究证明到规模化生产会急剧收窄,因为安全、集成和 ROI 门槛叠加。
漏斗值只是用于渲染的序数阶段权重,不是实测转化率;公开 AMI 转化数据不可得。
[CM004, CM018, CM020, CM021, CM022, CM023]2.5 商业化约束与尽调缺口
市场有吸引力,但门槛很高。Bain 和 MIT Technology Review 都提醒,类人和物理智能体采用仍处早期、分阶段推进,并受自主性、灵巧度、电力、安全和信任限制。受监管或安全关键流程还叠加一层要求:EU AI Act 义务、FDA 对自适应医疗软件的生命周期监管、NIST 风险管理实践,以及前沿 AI 安全承诺,都会把买方推向验证、监测、治理和事件响应。算力强度同样重要,因为 Epoch 的训练成本研究和 Sequoia 的 $600B 批评显示,前沿模型经济性可能跑在终端用户收入前面。因此,关键尽调问题首先是商业问题,不只是技术问题:AMI 到底销售什么价值单元、谁拥有预算、试点如何证明 ROI、买方会接受哪些基准,以及能否在竞争对手或既有厂商商品化模型层之前拿出安全证据。[CM018, CM019, CM020, CM021, CM022, CM023]
| 缺口 | 重要性 | 当前证据 | 估值风险 | 下一步尽调 |
|---|---|---|---|---|
| AMI 定价单位未披露 | 无法把技术承诺换算为 ARR、毛利率或市场份额。 | 公开材料描述使命和合作关系,但没有说明打包方式。 | 高 | 索取模型访问、许可、服务和共同开发的定价假设。 |
| 狭义模型层 SAM 没有直接公开数据 | 广义具身 AI 估算包含硬件和整套系统。 | 推导出的 10%-30% SAM 视角是分析假设。 | 高 | 访谈分析师来源和买方,估算各工作流的软件 / 模型占比。 |
| 试点转生产尚未验证 | 企业买方需要正常运行时间、安全、集成和 ROI 证据。 | Nabla 是战略验证点,不是已披露的收入证明。 | 高 | 审阅试点合同、里程碑、成功指标和扩展权利。 |
| 基准测试接受度尚不成熟 | 世界模型和 VLA 论文显示进展,但还不是买方级标准基准。 | 竞争对手发布演示、论文或预览,指标各不相同。 | 中 | 定义覆盖工业、机器人和医疗周边任务的基准套件。 |
| 按用例划分的监管分类尚未解决 | 医疗、安全和前沿 AI 义务可能改变上市节奏。 | EU AI Act、FDA、NIST 和英国安全承诺带来治理预期。 | 中 | 获取法律备忘录,将目标工作流映射到义务和验证成本。 |
本表有意保留实质缺口,而不是强行给出单一 TAM/SAM/SOM。AMI 披露定价、试点和目标垂直行业之前,这些未知项应保持开放。
[CM021, CM022, CM023, CM024, CM033, CM034]2.6 展项
03竞争对手
3.1 格局与替代地图
AMI 的竞争场景是拥挤的物理 AI 栈,而不是边界清晰的单一软件品类。最接近的直接同行包括 World Labs、Google DeepMind Genie 2 等世界模型和空间智能实验室,也包括 Physical Intelligence、Skild AI、Figure、Covariant 等机器人策略实验室;它们都在尝试把多模态感知转成行动。赋能平台同样重要:NVIDIA Cosmos 和 Hugging Face LeRobot 向开发者提供模型、数据集、仿真和策略工具,可能降低使用 AMI 专有层的必要性。当买方更偏好全栈自主能力或内部自建,而不是授权模型供应商时,Wayve 和 Tesla 等垂直具身 AI 公司就是替代方案。因此,实际格局由直接竞争者、赋能平台、替代 / 内部自建、现状自动化,以及可能凭算力、数据或流程控制切入的新进入者构成。[CP001, CP002, CP041, CP045, CP047, CP048]
| 竞争对手 / 替代方案 | 类别 | 规模 / 融资信号 | 目标细分市场 | 差异化 | AMI 对比限制 |
|---|---|---|---|---|---|
| AMI | 本次评估公司 | 已融资 $1.03B;无公开产品指标 | 工业、机器人、医疗、自动化、可穿戴设备 | 面向传感器数据、规划和安全护栏的世界模型 | 公开来源显示尚未推出产品;定价和数据集未披露 |
| World Labs / Marble | 直接世界模型同业 | 官方 $1B 新融资;此前 $230M 和 Autodesk 投资已有报道 | 游戏、VFX、VR、设计、机器人仿真邻近领域 | 通用可用的 3D 世界模型,支持导出和编辑 | 与 AMI 宽泛的物理控制论点相比,更偏创意 / 空间产品 |
| Physical Intelligence | 直接机器人策略同业 | CNBC 报道 $400M、估值 $2.4B;2026 年谈判据报约 ~$1B、估值 >$11B | 机器人 OEM 与通用机器人控制 | 迭代 VLA / 机器人策略栈,并开源 π0 | TechCrunch 报道没有商业化时间表 |
| Skild AI | 直接机器人基础模型同业 | 官方 $300M Series A 轮,估值 $1.5B | 工业、家庭、危险环境、低成本机器人 | 跨操作、运动、导航的通用大脑 | 商业部署证据不如融资 / 模型叙事具体 |
| Figure / Helix | 直接人形机器人同业 | 官方 >$1B Series C 轮,投后估值 $39B | 人形机器人家庭和商业运营 | 具身人形硬件加 Helix VLA,号称已可商用 | 垂直整合硬件路径可能无法映射到 AMI 许可模式 |
| Covariant / RFM-1 | 直接机器人基础模型同业 | 生产仓储客户和车队数据,而非来源中的近期融资 | 仓储拣选、组套、拆垛、物流 | 商业机器人数据和面向仓储的 RFM-1 世界模型推理 | 起点比 AMI 跨行业论点更窄,聚焦物流 |
| Google DeepMind | 在位直接与研究同业 | Alphabet 级资源;未披露独立价格 | 研究合作伙伴、机器人、具身智能体 | Gemini Robotics 私有预览和 Genie 2 世界模型研究 | 访问受限,产品打包未披露 |
| NVIDIA Cosmos | 赋能平台 / 潜在进入者 | 开放平台,初始采用者广泛 | 机器人和 AV 开发者 | WFMs、分词器、护栏、合成数据、文档、计算栈 | 可能赋能 AMI 而非替代 AMI,但会让工具商品化 |
| Wayve | 具身 AI 欧洲同业 / 垂直替代 | SoftBank 领投的 $1.05B Series C 轮 | 汽车 OEM 和车队所有者 | 面向驾驶的硬件无关、无地图具身 AI | 聚焦 AV,而非通用工业世界模型 |
| Tesla Optimus | 垂直替代 / 内部自建信号 | Q1 2026 材料显示 Optimus 产线和 AI 算力爬坡 | Tesla 工厂和未来人形机器人应用 | 硬件、制造、机器人数据、AI 算力整合 | 封闭内部栈;短期内不是外部模型供应商 |
画像结合官方和独立公开证据;未披露的定价、ARR 和私有客户指标不作推断。
[CP003, CP009, CP012, CP018, CP020, CP024]AMI 的竞争评分卡里,资本规模和研究主线最强,产品验证和数据闭环可见度最弱。
1-10 序数评分是基于留存证据的尽调判断,不是公司披露指标。
[CP003, CP005, CP006, CP029, CP033, CP043]3.2 直接同行能力对比
公开商业化证据上,直接同行对比不利于 AMI;但在野心上,AMI 占优。来源支撑的 AMI 论点是面向真实传感器数据和规划的抽象世界模型,但公司尚未披露已交付产品、定价页、基准、客户案例或通用 API。World Labs 已经推出 Marble,作为付费和免费 3D 世界产品;Physical Intelligence 展示了围绕 π0 及 2026 年后续版本迭代的模型线;Figure 声称 Helix 已可在嵌入式 GPU 上商业就绪;Covariant 的 RFM-1 背后有仓储部署数据;Google DeepMind 则有私有预览版 Gemini Robotics 和研究级 Genie 2。整个行业的能力宽度真实存在,但矩阵中缺乏支撑的单元格仍标为 unknown,因为演示、预览和投资人说法无法证明 AMI 的生产可靠性。[CP007, CP009, CP010, CP014, CP015, CP016]
| 购买标准 | AMI | World Labs | Physical Intelligence / Skild | Figure / Covariant | DeepMind / NVIDIA |
|---|---|---|---|---|---|
| 世界模型范围 | 声称可做传感器抽象和动作条件规划 | 3D 空间世界;持久导出 | 机器人策略和物理智能 | 人形机器人和仓储动作模型 | Genie 2 虚拟世界;Cosmos WFMs |
| 具身动作 | 通过世界模型规划;无公开演示 | 具备机器人仿真潜力,但不是机器人控制产品 | 核心聚焦机器人动作策略 | 核心聚焦机器人动作 / 硬件部署 | Gemini Robotics 和 Cosmos 瞄准物理动作栈 |
| 商业可用性 | 未发现公开 SKU 或 API | Marble 已通用可用,并设分层 | 据报道 PI 无时间表;Skild 产品条款未知 | Figure 声称已可商用;Covariant 有仓储产品 | Gemini 私有预览;Cosmos 开放 / 开发平台 |
| 定价证据 | 未披露 | 据报道有免费、$20、$35、$95 月费档 | 未披露 | 未披露 | 多数未披露,或为平台 / 许可访问 |
| 分发切入点 | Nabla 首发访问医疗合作伙伴 | Autodesk 设计 / 媒体合作 | 投资人和合作伙伴信号,细节有限 | 硬件 / 物流客户环境 | DeepMind 合作伙伴;NVIDIA 广泛采用者生态 |
| 数据优势 | 未披露合作伙伴数据权利 | 生成式 3D 世界和设计工作流 | 跨机器人和灵巧任务数据集 | 人形机器人数据采集和仓储轨迹 | 视频 / 仿真 / 模型基础设施,以及 AV / 机器人生态 |
| 信任 / 安全姿态 | 已表述安全性和可控性 | 商业资产生成风险仍然存在 | 机器人可靠性仍偏研究 | 物理安全和正常运行时间必须证明 | NVIDIA 护栏;受监管机器人 / AV 验证仍然困难 |
不支持或未披露的单元格写为未知 / 未披露,而不是估算。
[CP001, CP009, CP010, CP011, CP014, CP016]AMI 目标和资本都很大,但在已发布产品、私有预览或生产数据闭环方面落后于同行。
评分是有来源支撑的序数判断,不是报告 KPI;x 轴强调公开可用性 / 部署证明。
[CP014, CP019, CP023, CP026, CP029, CP032]竞争对手按其是否掌握空间生成、机器人动作、基础设施或部署渠道形成集群。
单元格为定性判断,只使用留存来源支持的表述;未知价格有意保留。
[CP009, CP016, CP021, CP025, CP027, CP029]3.3 商业化、定价与分销
定价多数未披露,所以本章不编造价格。Marble 是审阅范围内唯一有公开订阅层级的直接世界模型产品。LeRobot 是开放工具,不是 AMI 式付费供应商合同。AMI、Physical Intelligence、Skild AI、Figure Helix、Covariant RFM-1、Gemini Robotics,以及多数 Cosmos 企业条款,要么未披露,要么绑定伙伴关系、预览、部署或平台访问,而不是公开标价。分销差异比定价更大:World Labs 有 Autodesk 作为战略设计工作流渠道,NVIDIA 有广泛的首批采用者生态,Wayve 有 OEM 和云关系,Tesla 垂直整合制造;AMI 目前有 Nabla 这个具名医疗首批接入伙伴,但尚未披露可比的机器人或工业渠道。[CP003, CP004, CP005, CP006, CP011, CP012]
| 替代方案 | 公开价格 / 打包 | 包含能力 | 未知项 | 竞争含义 |
|---|---|---|---|---|
| AMI | 未披露 | 世界模型研究论点和 Nabla 首发访问合作关系 | SKU、API、许可、试点费用、企业条款 | 目前无法判断价格竞争力 |
| World Labs Marble | 据报道有免费档;Standard $20/mo;Pro $35/mo;Max $95/mo | 3D 世界生成、编辑、导出,付费档包含商业权利 | 企业 / 模型许可经济性未披露 | 为空间世界模型设定了可见的低摩擦基准 |
| Physical Intelligence | 未披露 | 机器人基础模型、开源 π0,官网提到合作伙伴 | 商业化时间表和付费打包 | 融资额高但没有定价,尽调负担加重 |
| Skild AI | 未披露 | 通用机器人“大脑”主张 | 客户合同、部署费用、模型访问 | 竞争靠能力叙事,而不是价格透明度 |
| Figure Helix | 未披露 | 人形机器人和 Helix VLA,与制造计划整合 | 机器人租赁 / 销售 / 服务模式,以及 Helix 单独访问 | 垂直硬件经济性可能绕开 AMI 式许可 |
| Covariant RFM-1 | 未披露 | 仓储机器人拣选与 RFM-1 推理能力 | 按机器人、SaaS、服务或部署计价 | 商业数据证明比标价更关键 |
| NVIDIA Cosmos / LeRobot | 开放模型许可 / 开源工具;已审阅文件未公开企业支持条款 | WFMs、文档、数据工具、机器人数据集、策略 | NVIDIA 企业支持与云消耗经济性 | 支持内部自建,也压低专有模型定价 |
| Wayve / Tesla | 外部模型价格未披露 | 全栈汽车或人形机器人项目 | 模型组件能否单独授权 | 它们是替代品,不是简单的软件可比公司 |
在已审阅来源中,只有 Marble 公开了美元订阅档位;其他定价单元格均有意保留“未披露”状态。
[CP005, CP011, CP019, CP023, CP024, CP029]| 参与方 | 分销 / 合作伙伴触达 | 数据优势 | 准备度信号 | AMI 含义 |
|---|---|---|---|---|
| AMI | Nabla 首发接入;战略投资方包括 NVIDIA、Toyota、Samsung 等 | 未公开披露合作伙伴数据权利或基准数据集 | 融资和团队已有证明,但产品尚无证明 | 必须把投资方 / 合作伙伴网络转成付费试点和数据接入 |
| World Labs | 与 Autodesk 开展研究 / 模型合作,并切入设计工作流界面 | 生成式 3D 世界,以及创意 / 设计反馈循环 | Marble 产品已正式开放 | AMI 需要更强的工业 / 机器人切入点,避免交付速度被拉开 |
| Physical Intelligence | CNBC 称支持方包括 OpenAI/Bezos;官网提到合作伙伴应用 | 跨机器人与灵巧任务数据集 | 模型线活跃;未报道商业化时间表 | 技术上接近 AMI,机器人控制指向更明确 |
| Figure | 人形机器人硬件与 BotQ 制造路径 | 计划采集人类视频和多模态传感数据 | Helix 称已具备商用条件,Series C 用于规模化 | 垂直数据闭环可能比 AMI 单纯研究更快复利 |
| Covariant | 仓储自动化客户覆盖多个国家和行业 | 来自生产机器人数千万条轨迹 | 商用仓储机器人与 RFM-1 演示 | 生产轨迹数据是一道 AMI 尚未展示的护城河 |
| NVIDIA | Cosmos 初始采用者广泛,另有计算 / 软件生态 | 合成数据、分词器、Omniverse、Blackwell/NGC/Hugging Face 分发 | 开放模型和开发者文档已可用 | 可能成为 AMI 必须基于其构建、或与之竞争的默认平台 |
| Wayve | 有 SoftBank、NVIDIA、Microsoft 支持的 OEM / 车队路径 | 驾驶试验、车队学习、仿真与验证平台 | 融资投向量产车产品 | 通用 AI 之外,欧洲具身 AI 的强证明点 |
| Tesla | 内部工厂、AI 算力、车辆、机器人制造 | 专有车辆 / 机器人 / 工厂数据闭环 | Q1 2026 材料显示 Optimus 在准备产线 | 对大型工业买家构成内部自建威胁 |
本表强调有来源支撑的分销和数据信号,而非私有客户数或收入。
[CP004, CP005, CP008, CP013, CP026, CP030]3.4 数据优势与护城河耐久性
物理 AI 里最持久的竞争优势不太可能只来自模型架构。真正的优势是具身交互数据、仿真闭环、客户流程、评估基础设施和值得信任的部署渠道。Covariant 的仓储车队和轨迹说法、Figure 的类人数据采集、Physical Intelligence 的跨形态训练、Wayve 的道路试验和 OEM 路径、Tesla 的 AI 算力和 Optimus 量产准备、NVIDIA 的平台控制,以及 LeRobot 的开放数据集 / 工具体系,都会挤压 AMI。AMI 的资本、人才和战略支持方有价值,但在公司展示专有伙伴数据权利、安全基准、付费试点或买方无法通过内部自建和开放 / 赋能平台复制的集成界面之前,护城河还没有被来源证明。[CP008, CP026, CP029, CP030, CP033, CP034]
| 护城河主张 | 威胁 | 严重性 | 缓解措施 / 尽调问题 |
|---|---|---|---|
| 科学领军团队与世界模型论点 | 同行可以招到类似人才,并发布可比的机器人 / 世界模型演示 | 高 | 要求 AMI 研究团队提供基准证据和留任数据 |
| $1.03B 种子轮级资本 | World Labs、Physical Intelligence、Figure、Wayve 及基础设施巨头同样资本充裕 | 高 | 对比烧钱速度、算力分配和基于里程碑的融资纪律 |
| 战略投资方网络 | 投资方不等于客户分销或数据权利 | 中 | 索取已签商业试点、数据使用权和共同开发条款 |
| Nabla 首发接入合作 | 医疗切入口较窄,未必能迁移到机器人或工业自动化 | 中 | 核验交付物、排他性、里程碑、监管路径和付费经济性 |
| 世界模型 IP | 开放的 Cosmos、LeRobot 和 PI 权重可能降低内部自建的切换成本 | 高 | 识别开放栈难以复制的专有数据集、安全评测或部署工具 |
| 物理 AI 市场拉力 | 人形机器人和机器人部署仍早,自治、灵巧性、电池和信任都是关口 | 高 | 在给出广义 TAM 信用前,要求受控环境中的分阶段试点 |
| 算力规模 | GPU 算力可能商品化,模型层定价权可能被侵蚀 | 中 | 看到定价、COGS 和差异化结果后再承保毛利率 |
| 欧洲主权定位 | World Labs、Wayve、DeepMind 和全球巨头也能提供非美国或战略替代方案 | 中 | 测试买家选择 AMI 究竟看重主权、性能还是合作伙伴触达 |
风险登记表把反向分析师证据与竞争对手特定来源证据合并;严重性是序数型尽调判断。
[CP003, CP005, CP012, CP018, CP024, CP029]3.5 反向证据、商品化风险与潜在进入者
反向证据很重要。Bain 认为类人机器人部署仍处早期且需要结构化推进,自主性、灵巧度、电池、认证、劳动力接受度和信任都会卡住规模化。LeCun 本人告诉 MIT Technology Review,目前没人知道如何做出广泛有用的机器人,重大概念突破仍然必要。Sequoia 对 AI 基础设施的批评增加第二层风险:如果算力变得像商品,而终端用户收入滞后于资本开支,没有流程锁定的模型供应商会承受价格压力。潜在进入者站位很强,因为 NVIDIA 控制物理 AI 基础设施,Google DeepMind 控制前沿机器人研究,Tesla 控制机器人硬件 / 制造和专有数据,Autodesk 等 CAD / 仿真既有厂商拥有流程界面。AMI 的尽调负担,是证明其世界模型不只是商品化栈里又一条昂贵研究路径。[CP035, CP036, CP037, CP038, CP044, CP048]
3.6 展项
04财务
4.1 收入模型与披露缺口
AMI 尚未公开披露收入、ARR、客户数量、标价或通用产品。最清楚的事实是使命、$1.03B 种子轮和伙伴驱动的商业化路径:AMI 称自己在为安全关键的真实世界领域构建世界模型,Nabla 称自己可首批接入 AMI 面向医疗的新兴技术。这些证据支撑收入模型假设,不支撑收入确认。未来可能的收入流包括企业模型授权、API 访问、联合开发费用、垂直部署、医疗合作伙伴经济分成,以及可能的战略研究合同。没有合同、用量、定价、开票和续约数据,任何一项都不能当作已入账收入。尽调姿态因此必须明确:当前收入指标用 null,Nabla 归为伙伴信号,并在承销可重复性前要求管理层提供财务包。[CI001, CI006, CI007, CI008, CI009, CI014]
| 潜在收入流 | 机制 | 计量单位 / 确认问题 | 当前公开价值 / 状态 | 收入质量 | 尽调问题 |
|---|---|---|---|---|---|
| 世界模型企业许可 | 向工业、机器人、医疗或自动化合作伙伴授权 AMI 模型或权重 | 年度或多年许可;收入确认取决于访问权、更新和支持义务 | 未披露;未发现公开许可合同 | 仅为假设 | 索取已签合同、价格手册、条款清单和收入确认备忘录。 |
| API / 用量制访问 | 向开发者或合作伙伴开放推理、仿真、规划或评估端点 | 按用量、席位、token、环境或计算小时计量;COGS 绑定推理和验证 | 未找到公开的 AMI API 或价格表 | 仅为假设 | 检查产品路线图、API 遥测、单位成本、可用性 SLA 和计划定价。 |
| Nabla / 医疗合作伙伴经济性 | 借 Nabla 的临床 AI 渠道获得医疗商业化首发接入 | 可能是许可费、版税、转让价或嵌入式合作伙伴收入 | 已披露 Nabla 接入;经济条款未披露 | 证明战略价值,不证明收入 | 审阅 Nabla 协议、数据权利清单、最低承诺和收入分成条款。 |
| 垂直共同开发 / 试点 | 与工业、机器人、自动驾驶或医疗合作伙伴开展付费试点或共同开发 | 里程碑费用或服务收入;若高度定制,毛利率可能偏低 | 有报道称在与企业洽谈,但未披露试点经济性 | 可能是通向收入的桥 | 索取管线、SOW、发票、试点转化历史和客户 ROI 证据。 |
| 开放论文 / 开源 | 带来招聘、生态、基准或采用飞轮,而非直接变现 | 通常是间接收益;变现需要托管服务、支持或企业权利 | AMI 提到开放论文和开源,但没有付费支持产品 | 分发选项,不是收入 | 区分开放研究策略和可变现的企业级封装。 |
各行区分已披露事实与假设;null 或未披露表示没有公开价值,不等于收入为零。
[CI001, CI006, CI007, CI014, CI015, CI035]| 定价项 | 公开价值 | 有来源支撑的状态 | 重要性 | 尽调路径 |
|---|---|---|---|---|
| AMI 标价 | null | AMI 官方页面未发现公开定价页、API 费率或 SKU | 估算 ACV、折扣或毛利率前,必须先有标价 | 索取当前和计划中的价格手册。 |
| Nabla 许可或收入分成 | null | 已披露首发接入;费用、最低承诺、版税和转让定价未披露 | 只有具名合作伙伴信号,缺少经济条款就无法转成 ARR | 审阅已签署的 AMI-Nabla 协议。 |
| API 用量指标 | null | 未披露公开端点、token、仿真小时、环境或席位单位 | 用量单位决定 COGS 转嫁和毛利波动 | 获取 API 设计、单位成本模型和价格测试证据。 |
| 企业模型许可 | null | 收入模式只是从合作伙伴 / 产品表述推导出的假设 | 许可结构影响收入确认、支持负担和续约质量 | 索取 MSA 模板、许可范围、支持 SLA 和续约假设。 |
| 试点 / 共同开发费用 | null | 企业合作洽谈可能已启动,但缺少试点费用证据 | 类服务试点可能看起来像收入,却掩盖重复性差 | 索取管线、SOW、交付人员配置和转化指标。 |
| 结果或工作流定价 | null | 医疗工作流价值主张存在;定价公式未公开 | 结果定价需要临床风险分配和验证证据 | 审查支付方 / 服务方经济性和监管主张佐证。 |
所有定价值均有意保留为 null,因为没有来源披露 AMI 标价或实际成交价。
[CI008, CI014, CI015, CI016, CI022, CI024]AMI 当前公开路径从研究资产走向合作伙伴验证;收入机制还没有披露。
桥图为定性判断;AMI 未公开收入、价格或利润率数字。
[CI015, CI035, CI039, CI042, CI049]4.2 GTM 与销售效率代理指标
公开 go-to-market 证据很薄,但方向一致。AMI 和独立报道都指向先做研究、未来六到十二个月推进伙伴讨论,以及 Nabla 作为首个医疗首批接入伙伴。这不足以计算 CAC、回本期、配额产出、销售周期、管道转化、渠道毛利或客户集中度。最好的代理指标是定性判断:AMI 可能从战略伙伴和受监管或工业试点切入,先完成技术验证,再进入业务线预算归属。如果能转成高价值授权,这条路有价值;但它也慢,而且服务含量重。在 AMI 披露已签试点条款、转化率和经济买方归属之前,销售效率应标为 unavailable,而不是拿 SaaS 常模做基准。[CI005, CI006, CI012, CI013, CI016, CI038]
| 指标 | 数值 / 状态 | 置信度 | 重要性 | 尽调问题 |
|---|---|---|---|---|
| 收入 / ARR | null | 中 | CAC 回收、NRR 或毛利测算都没有公开收入基线 | 索取 ARR、订阅额、已确认收入、递延收入和发票。 |
| 付费客户 | null | 中 | Nabla 是战略合作伙伴信号,不是已披露付费客户数 | 索取客户名单、已签合同和客户集中度明细。 |
| CAC / 回收期 | null | 低 | 未披露销售动作、管线转化、配额产能或渠道经济性 | 索取 CRM 导出和销售效率仪表盘。 |
| 毛利率 | null | 低 | 算力、验证、支持和定价均未披露 | 用工作负载遥测和合同定价搭建毛利率桥。 |
| 训练算力 COGS / 研发 | 未披露;外部前沿成本在上升 | 中 | 收入出现前,前沿训练可能主导烧钱 | 索取训练路线图、实验预算和模型会计政策。 |
| 推理算力 COGS | 未披露;GPU 云租 / 购买价格区间可从外部观察 | 中 | 单位毛利取决于利用率和成本转嫁 | 索取推理遥测、GPU 小时成本、预留条款和利用率。 |
| 数据整理 / 测试 | 未披露;物理 AI 需要大量数据和测试 | 中 | 传感器 / 视频工作流会在 GPU 之外带来人力和基础设施成本 | 索取数据集权利、标注预算、仿真支出和测试计划。 |
| 安全 / 验证成本 | 未披露;FDA/NIST 风险工作流相关 | 中 | 医疗和关键基础设施会推迟收入并抬高服务成本 | 索取验证预算、质量管理计划和监管映射。 |
| 债务 / 项目融资 | 无公开披露 | 中 | 即便种子轮金额很大,已承诺义务也可能缩短跑道 | 索取债务明细、云承诺和资本开支授权。 |
代理表有意将私有指标保留为 null,仅用外部成本背景识别尽调驱动因素。
[CI005, CI009, CI016, CI017, CI019, CI021]4.3 成本结构与毛利驱动因素
AMI 的成本结构很可能由算力、顶尖研究人才、数据整理、仿真、安全验证和伙伴集成主导。这是基于 AMI 的前沿世界模型野心和外部成本基准作出的推断,不是公司披露的烧钱模型。Epoch 显示前沿训练成本快速上升;NVIDIA 称物理 AI 系统可能需要 PB 级视频数据和数万小时算力;CloudZero 报告 H100 购买 / 租赁区间很宽;FDA / NIST 材料则显示,医疗和关键基础设施流程会增加风险管理和验证工作。Blackwell 式效率可以降低单位推理成本,但在 AMI 证明工作负载组合、利用率、定价权和合同收入之前,毛利问题不会被解决。因此,毛利率应继续保持 null,成本驱动地图仅作为尽调指南。[CI017, CI018, CI019, CI020, CI021, CI022]
| 成本驱动项 | 证据基础 | 可能对应的损益科目 | 对利润率的影响 | 尽调指标 |
|---|---|---|---|---|
| 训练算力 | Epoch 关于前沿成本增长的证据;AMI 的世界模型目标 | 研发,也可能是资本化模型开发 | 如果大规模训练先于试点启动,可能在收入出现前吃掉种子轮资金 | 单次实验成本、最终训练预算和训练路线图。 |
| 推理算力 | NVIDIA 效率说法;CloudZero 的 H100 购买 / 租赁区间 | 客户开始使用后计入 COGS | 如果定价不能转嫁 GPU-hours 成本,毛利率可能被压缩 | 每个工作流的 GPU-hour、利用率、预留折扣和成本转嫁条款。 |
| 传感器 / 视频数据与整理 | NVIDIA Cosmos 强调物理 AI 的数据和测试强度 | 研发、COGS 或合作伙伴数据费用 | 可能产生 GPU 成本估算看不到的人力和数据权利成本 | 数据集规模、权利、标注成本、仿真成本和刷新节奏。 |
| 顶尖 AI 人才 | 投资人资金用途提到全球招聘 | 研发和 G&A | 在收入前推高固定烧钱 | 已到岗人数、薪酬结构和按季度招聘计划。 |
| 安全与验证 | Nabla 的 FDA-certifiable 表述;FDA/NIST 风险管理语境 | 研发、质量、法务和实施服务 | 可能推迟收入确认,并要求成本高的客户专属验证 | 验证预算、监管路径、QMS 配置和责任分配。 |
| 合作伙伴集成与支持 | Nabla 优先接入路径和企业合作伙伴节奏 | 服务、客户成功和解决方案工程 | 标准化产品包装成型前,早期试点毛利率可能较低 | 实施工时、SOW 毛利率、支持 SLA 和续约转化。 |
成本驱动项只能定性,因为 AMI 尚未披露工作负载遥测、人员规模或合同。
[CI004, CI017, CI019, CI020, CI021, CI023]未来单位经济性取决于使用量、GPU 成本、数据工作、安全验证和定价权。
仅为定性桥图;AMI 尚未披露工作负载组合、利用率、价格或毛利率。
[CI017, CI019, CI020, CI021, CI023, CI040]唯一站得住脚的数字区间是资本化和外部 GPU 成本背景;AMI 收入、烧钱速度、现金跑道和利润率仍为空。
图中有意排除收入、烧钱速度、现金跑道和利润率区间,因为 AMI 的公开数据无法支撑。
[CI002, CI003, CI021, CI047, CI048]4.4 资本充足性与融资依赖
大额种子轮买来时间和战略期权价值,但它本身不能证明资本充足。来源支撑的事实是 $1.03B 融资、据报 $3.5B 投前估值,以及投资人称这笔钱支持长期研究、全球招聘和可靠智能系统开发。真正判断现金续航所需的信息并未被支撑:扣除费用后的到账现金、已花现金、月度 burn、已承诺算力、债务、GPU 采购、办公室面积、员工数和下一轮里程碑。公开来源显示 AMI 计划做高算力 R&D,且当前没有收入指标;因此,在伙伴试点转成合同经济性,或公司证明种子轮资金足以支撑研究路线图之前,AMI 仍依赖后续融资。[CI002, CI003, CI004, CI025, CI026, CI031]
| 资本项 | 公开价值 / 状态 | 置信度 | 解读 | 尽调问题 |
|---|---|---|---|---|
| 最新一轮 | USD 1.03B / 约 EUR 890M 种子轮 | 高 | 如果烧钱受控,足以支撑大量研究、算力和招聘跑道 | 核验总募集额与净到款、交割机制和到账现金。 |
| 报道估值 | 投前 USD 3.5B;不考虑结构,隐含投后约 USD 4.53B | 中 | 估值先把未来突破计入价格,而公开收入指标尚未出现 | 审阅股权结构表、清算优先权、期权池和老股部分。 |
| 在手现金 | null | 低 | 已知轮次规模,但支出后的当前现金未知 | 索取资金报告和交割后现金调节表。 |
| 月度烧钱 | null | 低 | 算力和顶尖人才可能让烧钱显著高于普通种子期 SaaS | 索取月度 P&L、现金流量表、工资单和云账单。 |
| 跑道月数 | null | 低 | 没有当前现金和烧钱就无法计算 | 核验现金和烧钱后再计算。 |
| 计划资金用途 | 长期研究、全球招聘、可靠系统、重算力开发 | 中 | 资金用途偏研发,不是近期销售扩张计划 | 按工作流和里程碑索取董事会预算。 |
| 下一轮触发条件 | null | 低 | 未公开里程碑计划,将未来融资与收入或技术关口挂钩 | 索取里程碑模型和融资敏感性方案。 |
| 债务 / 项目融资义务 | 未发现公开披露 | 中 | 没有债务证据,不等于没有承诺义务 | 索取债务、云、GPU、租赁和数据中心义务明细。 |
| 算力成本敞口 | 重大但未量化 | 中 | 外部来源显示训练、数据、GPU 和 TCO 成本可能很高 | 索取已承诺 GPU / 云支出和模型训练路线图。 |
跑道和烧钱保持 null,因为用种子轮金额估算会编造私有数据。
[CI002, CI003, CI004, CI017, CI018, CI021]4.5 财务结论与尽调阻塞项
财务结论是 research-more,不是因为 AMI 缺钱,而是承销所需的财务报表基本缺位。正向案例很强:异常充足的种子轮融资、顶级投资人,以及伙伴驱动的医疗切口。反向案例同样具体:Sequoia 质疑 AI 基础设施支出是否有终端用户收入匹配,Le Monde Informatique 称 AMI 缺少可运行系统,Silicon Republic 提醒一家今年成立的公司已有数十亿美元估值。这些批评重要,因为 AMI 未来成本基础会重算力、重验证。阻塞性尽调路径很具体:拿到合同、价格手册、ARR / bookings、现金和 burn、算力承诺、验证预算、数据权利和董事会里程碑。在此之前,所有收入和现金续航字段都应保持未披露,而不是估算。[CI027, CI028, CI029, CI030, CI044, CI049]
| 缺口 | 对承保的影响 | 严重性 | 具体尽调路径 |
|---|---|---|---|
| ARR / 已确认收入 | 无法评估收入质量、增长、留存或估值倍数 | 阻断性 | 索取 ARR、订阅额、发票、递延收入和收入确认政策。 |
| 定价 / 已成交合同条款 | 无法估算 ACV、折扣、毛利率或客户支付意愿 | 阻断性 | 索取价格手册、已签合同、试点 SOW 和折扣审批。 |
| 现金、烧钱和跑道 | 无法判断资本充足性或下一轮融资依赖 | 阻断性 | 索取现金台账、月度 P&L、预测和董事会批准预算。 |
| 算力承诺 | 无法评估成本底线、capex/opex 组合或下行烧钱情形 | 重大 | 索取云 MSA、GPU 预留、硬件 capex、利用率和抵扣额度。 |
| Nabla 经济条款和数据权利 | 无法把合作伙伴信号转成 AMI 收入或护城河证据 | 阻断性 | 审阅协议、排他性、数据权利、最低承诺和收入分成条款。 |
| 客户数量 / 管线 | 无法区分市场兴趣和可重复需求 | 重大 | 索取 CRM 管线、客户推荐、转化漏斗和流失预期。 |
| 安全 / 监管验证预算 | 无法为医疗或关键基础设施上线要求定价 | 重大 | 索取监管地图、质量体系、验证测试计划和责任分配。 |
| 员工数和薪酬 | 无法区分研发野心和薪酬烧钱 | 重要 | 索取组织架构图、已到岗 / 计划人数、薪酬带和招聘承诺。 |
| 下一轮融资里程碑触发点 | 无法判断种子轮是否覆盖下一个价值拐点 | 阻塞 | 索取与技术、安全、合作伙伴和商业里程碑挂钩的融资计划。 |
本表是尽调工作计划,用可审计的私有指标替换 null 占位符。
[CI009, CI014, CI025, CI026, CI029, CI030]4.6 展项
05产品与技术
5.1 产品状态与客户流程
AMI 的产品外露面更适合描述为世界模型的研究平台和伙伴集成路线图,而不是已交付 SKU。官网称,世界模型在表征空间中预测,并在安全护栏下规划行动序列;更新页面称,AMI 正在构建能够理解世界、保持持久记忆、推理、规划,并保持可控和安全的系统。这映射到客户流程,而不是可定价产品:医疗伙伴可以用仿真和确定性规划扩展临床助手;工业运营方可以把传感器状态喂给模型,预测过程后果;机器人团队可以用视频条件规划选择安全动作;安全关键团队可以要求执行前先由人审核计划。关键尽调边界很严格:Nabla 有首批接入权,但公开来源没有显示 AMI 已有通用模型、API、SLA、基准或发布日期。[CE001, CE002, CE003, CE004, CE005, CE006]
| 用户任务 | 现有工作流痛点 | 与 AMI 相关的解决方案概念 | 需要证明的可衡量收益 | 当前限制 |
|---|---|---|---|---|
| 医疗临床人员 / 运营团队 | LLM 助手能帮写文档,但确定性多模态规划仍吃力 | 世界模型先模拟临床工作流后果,再建议动作 | 降低认知负荷、减少不安全自主动作、验证任务完成 | 目前只有 Nabla 优先接入;未展示获 FDA 批准的 AMI 产品。 |
| 工业过程控制操作员 | 规则和控制回路可能漏掉罕见传感器状态组合 | 预测未来过程状态,并提出受约束的动作序列 | 降低停机时间、减少过程偏离、让干预更安全 | AMI 未披露工业试点、传感器接口或可靠性指标。 |
| 机器人工程师 | 机器人策略往往需要特定环境数据或大量校准 | 视频条件世界模型围绕当前状态和目标状态做规划 | 泛化到新物体、降低数据采集负担、让操作更安全 | 证据来自 Meta V-JEPA 2,而非 AMI 部署。 |
| 自动化 / 物流负责人 | 物理环境变化时,手工处理异常会限制自主性 | 抽象状态表征过滤不可预测细节,并规划下一步动作 | 异常解决率和人工干预减少幅度 | 未公布客户部署、支持模式或集成架构。 |
| 可穿戴 / 个人设备厂商 | 连续多模态信号噪声大,且依赖上下文 | 持久记忆和传感器状态模型识别上下文,并推荐下一步 | 误报减少、延迟、电池、隐私和用户信任 | AMI 提到可穿戴,但未披露设备、传感器栈或隐私设计。 |
| 安全关键基础设施团队 | 在高风险环境中,自主系统行动前必须可审计、受约束 | 带人工闭环、已验证风险控制和监控的规划器 | 已验证安全论证、事故率、可审计性和操作员接受度 | AMI 尚未拿出监管和风险管理控制的证据。 |
收益项是需要索取的尽调指标,不是 AMI 已报告结果;当前限制保留路线图层级的 证据强度。
[CE002, CE005, CE006, CE025, CE026, CE027]产品工作流由合作伙伴牵引:观察状态、预测后果、约束动作,再接入人工复核的工作流。
流程根据来源描述可能的客户工作流;AMI 未发布集成文档。
[CE002, CE005, CE025, CE026, CE027, CE029]5.2 模块、资产与架构地图
模块地图应严格受证据约束。AMI 尚未发布产品包,因此资产地图只是把公开研究主张翻译成可部署层:传感器接入、表征编码器、JEPA 预测器、动作条件规划、评估 / 安全护栏,以及伙伴流程适配器。LeCun 的架构论文、基于能量模型笔记、I-JEPA、V-JEPA 和 V-JEPA 2 支撑技术谱系,但它们不是 AMI 产品基准。Meta 的 V-JEPA 2 是最具体的公开类比,因为它把视频训练、动作条件预测和机器人任务的模型预测控制连在一起。AMI 尽调要证明架构,需要说明哪些部分是专有的,哪些借用了开放文献,哪些已开源,哪些已经接入伙伴数据系统。[CE010, CE011, CE012, CE013, CE014, CE015]
| 模块 / 资产 | 主要用户 | 公开成熟度 | 差异化主张 | 尽调缺口 |
|---|---|---|---|---|
| 世界模型研究平台 | AMI 研究人员和战略合作伙伴 | 公开论点;没有 SKU | 面向真实世界传感器数据的表征空间预测 | 索取内部路线图、模型卡和发布标准。 |
| 传感器数据表征编码器 | 机器人、工业、可穿戴、医疗团队 | I-JEPA/V-JEPA 文献里可见研究类比 | 不是像素 / token 重建,而是非生成式语义嵌入 | 展示 AMI 自有训练数据、模态和评测结果。 |
| JEPA 预测器 / 世界模型 | 安全关键工作流开发者 | 研究谱系已公开;AMI 实现未披露 | 预测抽象未来状态和后果 | 提供 AMI 基准、失效模式和模型治理。 |
| 动作条件规划器 | 机器人和自动化工程师 | Meta V-JEPA 2 类比;AMI 产品未披露 | 基于候选动作做模型预测规划 | 在合作伙伴环境中演示 AMI 规划器,并给出安全边界。 |
| 安全与评测护栏 | 临床、工业和合规负责人 | 原则有主张;控制措施未公开 | 在高风险场景执行前约束规划 | 提供安全论证、红队结果、监控和事故流程。 |
| Nabla 医疗适配器 | Nabla 产品和临床工作流团队 | 优先接入合作;没有 AMI SKU | 以仿真和确定性推理切入临床工作流 | 审阅数据权利、产品范围、验证计划和经济性。 |
| 开放发表 / 代码 | 研究社区和开发者 | 意向已表明;尚未找到 AMI 专属产物 | 招聘和生态飞轮 | 区分开放研究、企业支持和专有护城河。 |
各行是受公开证据约束的模块假设;并非 AMI 已宣布的 SKU 或已发布产品。
[CE001, CE003, CE005, CE009, CE010, CE012]| 层级 / 组件 | 作用 | 依赖 | 风险 |
|---|---|---|---|
| 真实世界传感器和工作流数据 | 为表征学习和规划提供观测 | 合作伙伴数据权利、模态覆盖、隐私批准 | 未披露 AMI 数据语料、权利清单或模态列表。 |
| 表征编码器 | 将噪声观测转成语义嵌入 | 自监督学习和整理后的视频 / 图像 / 传感器数据 | 嵌入质量未必能跨医疗、工业和机器人领域迁移。 |
| JEPA 预测器 / 能量模型 | 预测目标表征和未来状态 | H-JEPA / 基于能量建模的研究谱系 | AMI 发布实现证据前,架构证明仍停留在文献层面。 |
| 动作条件规划器 | 评估候选动作,并选择安全下一步 | 机器人或流程动作数据、目标、模型预测控制回路 | 超出短周期或受控环境后,规划可能失效。 |
| 评测和基准套件 | 衡量物理推理、因果性和安全性 | 基准、红队、验收测试、人工基线 | 没有公开的 AMI 专属基准或可靠性指标。 |
| 合作伙伴应用适配器 | 将模型输出嵌入 Nabla 或工业工作流 | API、schema、审计日志、支持、安全控制 | 没有公开 API、部署拓扑或 SLA。 |
架构表把 AMI 主张与 JEPA/V-JEPA 研究类比并列;不主张 AMI 已发布技术栈。
[CE010, CE011, CE012, CE013, CE015, CE016]AMI 的公开主线可映射为分层世界模型平台,但目前外部只能看到研究脉络。
层级图是受证据约束的综合,不是 AMI 发布的架构图。
[CE001, CE010, CE011, CE015, CE018, CE019]5.3 相对 LLM 与开放物理 AI 平台的差异化
AMI 最强的差异化主张是概念层面的:token 预测型 LLM 针对语言续写优化,而 JEPA 式世界模型学习抽象状态表征,并预测世界可能如何演化,尤其是在候选动作之下。这在医疗和机器人里相关,因为幻觉、缺失因果性和弱物理扎根都可能不可接受。风险在于,架构本身不是护城河。Meta、NVIDIA、Hugging Face 和 DeepMind 已经在相邻的世界模型和机器人控制领域公开论文、代码、模型集合、仿真工具或私有预览。因此,AMI 需要专有伙伴数据权利、安全评估、集成界面和部署证据来证明差异化。没有这些,买方可能把 AMI 视为开放平台、内部自建项目和既有机器人栈之间的一条研究路径而已。[CE007, CE008, CE017, CE020, CE021, CE022]
AMI 在产品化前依赖数据权利、算力、合作伙伴工作流、开放研究姿态和受监管验证。
依赖图是定性判断,并有意标出公开产品验证缺口。
[CE021, CE022, CE023, CE024, CE029, CE030]5.4 部署、信任、安全、隐私与合规缺口
部署路径由伙伴驱动,而且高度依赖受监管流程。AMI 的公开外露面不包括 API 文档、客户 onboarding、监控、支持、数据处理条款、隐私控制、安全认证或上市后证据。在医疗场景里,Nabla 的 FDA-certifiable 表述应先按愿景处理,直到设备范围、predicate 或 De Novo 策略、验证计划和变更控制计划可见。FDA、NIST 和 EU AI Act 材料把控制负担讲得很具体:风险管理、可审计性、验证、人工监督、修改控制,以及保护健康、安全和权利,在安全关键用例里都不是可选项。立即部署建议应限制在小范围试点:保留 human-in-the-loop 审核,锁定运行域,明确事件流程,并在自主行动前收集证据。[CE029, CE030, CE031, CE032, CE033, CE034]
| 控制 / 认证领域 | 公开状态 | 范围 | 缺口 / 尽调要求 |
|---|---|---|---|
| 可获 FDA 认证的医疗 AI | 来自 Nabla 的路线图表述,不是获批证据 | 潜在临床智能体 AI 功能 | 获取监管策略、器械范围、验证和 FDA 往来函件。 |
| 预定变更控制计划 | 外部 FDA 指引已存在 | AI 驱动器械变更 | 将 AMI 模型更新映射到计划变更、方法论和影响评估。 |
| NIST AI 风险管理 | 外部框架已存在 | 关键基础设施和值得信赖 AI 实践 | 提供风险台账、控制措施、监控和治理责任归属。 |
| EU AI Act 合规 | 外部法规已存在 | 高风险 AI、健康、安全和基本权利保护 | 给用例分类,并记录人工监督、数据治理和合规路径。 |
| 可靠性 / 正常运行时间 / 事故响应 | AMI 未公开披露 | 产品运营和支持 | 索取 SLA、状态页、事故日志和升级模型。 |
| 隐私和安全控制 | AMI 未公开披露 | 临床、工业、可穿戴和合作伙伴数据 | 索取 DPA、加密、访问控制、审计日志、留存和认证。 |
| 人工监督和安全护栏 | 原则有主张;实现未公开 | 行动前自主规划 | 要求限定自主性、操作员批准、验证闸门和回滚设计。 |
表格区分外部监管要求与 AMI/Nabla 主张;认证单元格不视为 已达成。
[CE031, CE032, CE033, CE034, CE035, CE036]5.5 路线图、开源姿态与产品成熟度
路线图很长。AMI 已筹集种子轮级别资本,并称将与行业伙伴、产品开发者、学术界、公开发表和开源协作。TechCrunch 访谈是最清楚的商业化时点护栏:管理层把 AMI 定位为基础研究,可能需要数年才能变成商业应用。Nabla 合作提供可信的医疗切口,但仍是首批接入,不是 AMI 软件已交付的证据。外部证据也给机器人假设降温:类人部署面临电池、制造、安全、认证和人类接受度障碍,而 Meta 自己的 V-JEPA 2 基准相对人类物理推理仍有差距。成熟度结论是 research-more:技术论点自洽,但在 AMI 发布模型工件、伙伴试点、安全案例和部署指标前,公开证据还停在路线图层面。[CE009, CE039, CE040, CE041, CE042, CE044]
| 日期 / 阶段 | 功能或里程碑 | 状态 | 含义 | 来源依据 |
|---|---|---|---|---|
| 2022 | LeCun 关于自主机器智能的架构愿景 | 已发表研究愿景 | 提供概念架构,但不是 AMI 产品证据 | OpenReview 论文。 |
| 2023 | I-JEPA 和能量模型笔记 | 已发表研究谱系 | 支撑表征空间和 H-JEPA 框架 | arXiv 论文。 |
| 2024–2025 | V-JEPA 和 V-JEPA 2 产物 | 来自 Meta 生态的公开论文、代码和模型 | 可作为规划和开发者信号的有用类比;不是 AMI 基准 | arXiv、Meta、GitHub 与 Hugging Face。 |
| December 2025 | Nabla 独家 AMI 合作伙伴关系 | 已宣布优先接入合作伙伴 | 医疗切入口存在,但产品范围、经济性和审批状态未披露 | Nabla 新闻稿。 |
| March 2026 | AMI $1.03B 融资和团队建设更新 | 获融资支持的研究扩张 | 资本和人才足以支撑研究;没有发布日期或 SKU | AMI 更新和 TechCrunch。 |
| 未来几年 | 商业世界模型应用 | 仅路线图层级 | 管理层暗示,基础研究可能需要数年才能商业化 | TechCrunch 访谈。 |
| 运行日期 2026-07-03 | AMI 产品可用性 | 未找到公开 GA SKU、API、SLA、基准或支持路径 | 在发布证据出现前,按研究平台评估 | 已审阅官方和新闻来源。 |
开发阶段使用公开日期和来源表述;未发布被记录为尽调发现, 不推断为失败。
[CE004, CE009, CE014, CE017, CE039, CE040]公开研究脉络成熟度最高,AMI 特定产品、可靠性和监管验证最低。
定性矩阵;不推断 AMI 数字化基准或产品发布评分。
[CE003, CE015, CE025, CE026, CE027, CE035]5.6 展项
06客户
6.1 AMI 直接客户证据仍处商业化前
公开客户尽调应从一个负面发现开始:AMI 尚未披露付费客户基础、客户数量、产生 ARR 的试点、NRR 或合同条款。最强直接证据是 AMI 自己的定位——一家为可靠性关键领域构建世界模型的研究实验室——以及 Nabla 的独家首批接入合作。这让 Nabla 具有战略重要性,但不会把 Nabla 的医疗系统客户转成 AMI 客户。AMI 可能的买方是需要可靠预测、规划和仿真的企业或受监管运营方;在医疗切口里,经济买方很可能是 Nabla 或医疗系统渠道,而不是单个临床医生。因此,尽调姿态是商业化前:先拆分可能的买方、用户和付款方群体,再要求 AMI 特定生产部署证据,才可承销可重复性。实际操作上,本章每个尽调展项都把客户证明按层级处理:AMI 官方定位在底部,其上是 Nabla 战略接入,Nabla 客户部署证明放在独立的渠道验证通道里,AMI 直接生产证明仍为空。这个层级防止报告把伙伴势能误转成尚未赚到的客户指标。[CU001, CU002, CU003, CU004, CU005, CU007]
| 客群 | 买方 / 付款方 | 主要用户 | 用例 | 当前证据 | 战略价值 | 缺口 |
|---|---|---|---|---|---|---|
| AMI 直接医疗合作伙伴 | Nabla 或医疗系统合作伙伴渠道 | 通过 Nabla 触达临床医生和护理团队 | 超出文档记录的智能体临床工作流 | Nabla 拥有独家优先接入;未披露 AMI 部署 | 第一个受监管垂直切入口 | 缺少 AMI 专属试点、定价和部署证据 |
| AMI 直接工业 / 自动化买方 | 工业运营商或自动化厂商 | 工厂操作员、控制工程师、机器人团队 | 安全关键系统中的预测、规划和控制 | AMI 官网将工业过程控制和自动化列为目标领域 | 如果可靠性得到证明,可进入大型安全关键市场 | 未披露具名工业客户或 PoC |
| AMI 直接机器人 / 物理 AI 买方 | 机器人 OEM、仓储运营商、出行公司 | 机器人工程师和现场操作员 | 用动作条件世界模型在约束下规划 | AMI 官网和 TechCrunch 描述真实世界应用和未来客户 | 潜在高价值横向智能层 | 未具名现场部署、基准或付费客户 |
| Nabla 间接医疗系统用户 | 购买 Nabla 的医疗系统 | 医生、APPs、护士、编码团队 | 环境式文档记录、编码、EHR 指令、未来智能体工作流 | 多个 Nabla 案例研究和 Series C 材料 | 为 AMI 医疗切入口提供渠道学习和分发 | 除非 AMI 技术签约或部署,否则 Nabla 客户不是 AMI 客户 |
| 战略投资人 / 生态伙伴 | 战略支持方或被投组合渠道 | 产品团队和领域专家 | 数据访问、技术验证、未来分发 | 投资人名单包含与工业相关的支持方 | 潜在设计伙伴 | 投资人参与不等于客户收入 |
客户细分区分 AMI 直接商业化假设与 Nabla 渠道验证;没有任何一行意味着 AMI 披露了付费客户。
[CU001, CU002, CU004, CU012, CU013, CU014]路径从 AMI 研究开始;只有在 AMI 赋能的工作流签约、部署、衡量并续约之后,才会变成客户验证。
阶段是基于公开证据搭出的尽调模型;AMI 尚未披露直接试点或续约阶段。
[CU001, CU003, CU004, CU012, CU013, CU031]6.2 Nabla 提供具名客户证明,但只是间接渠道验证
Nabla 是唯一具名且可优先接入 AMI 技术的战略伙伴,也带来有意义的间接证明:医疗系统已在真实临床流程中部署 Nabla 的环境式助手,报告使用率,并从试点扩展到更大范围部署。这些证据重要,因为 AMI 第一条可信医疗路径很可能通过 Nabla 的装机基础和反馈闭环推进。但它必须放在单独的尽调桶里。Denver Health、Carle、McFarland、Tia、CHLA、UToledo 和 Aultman 验证的是 Nabla 销售、集成和支持临床 AI 的能力;它们不能证明 AMI 已交付自己的世界模型,也不能证明 AMI 已获得收入。因此,具名证明表把 AMI 关系与下游 Nabla 客户结果分开标注。最强间接信号不是单个 logo,而是 Nabla 账户中反复出现的模式:明确的流程痛点、试点或评估、EHR 集成、临床医生层面采用,以及量化的文档或职业倦怠结果。缺失的桥,是这个模式里披露的 AMI 衍生功能。[CU015, CU016, CU017, CU018, CU019, CU020]
| 指标 | 数值 | 日期 / 资料期 | 来源 | 置信度 | 含义 | 缺失分母 |
|---|---|---|---|---|---|---|
| AMI 直接付费客户 | 未公开披露 | 截至 2026-07-03 | 已审阅 AMI、Nabla、TechCrunch、STAT 和追踪器来源 | 中 | 将 AMI 视为商业化前公司 | 客户数、客户标识名单、试点数、ACV |
| AMI 收入时间 | 未报告当前收入计划 | TechCrunch,March 2026 | TechCrunch | 中 | 商业证明不是近期事件 | ARR、预订额、已签试点 |
| Nabla 医疗系统覆盖 | 因来源不同,为 130+ 至超过 150 家机构 | 2025-2026 材料 | Nabla / Highland / The Healthcare Technology Report 等来源 | 中 | 存在大型间接渠道 | 准确活跃客户数和流失率 |
| Nabla 支持的临床人员 | 85,000 名临床医生 | 2025 年 6 月 Series C 材料 | Nabla / Highland / The Healthcare Technology Report 等来源 | 高 | 庞大用户基数可支撑反馈循环 | 月活用户数和使用结构 |
| Nabla 年度就诊量 | 2,000 万次年度就诊 | 2025 年 6 月 Series C 材料 | Nabla / Highland | 中 | 可能触达临床数据和工作流 | 可用于 AMI 衍生功能的就诊占比 |
| Denver Health 部署 | 300,000+ 次就诊;试点后一周内 400 名临床医生采用 | 抓取时案例研究为最新 | Nabla Denver Health 案例 | 中 | 证明试点后的规模化能力 | 分队列留存和合同期限 |
| UToledo 评估 | 病历关闭速度提升 29%;积压从 >400 降至 <30 | 2026 年 4 月 | PR Newswire / HIT Consultant | 高 | 证明 Nabla 有可量化的运营成效 | 长期续约和净扩张 |
| Aultman 推广 | 每天节省 30-60 分钟;单名患者文档时间减少 20-40% | 2026 年 1 月 | TMCnet / PRNewswire 联合稿 | 中 | 证明采购不只来自 Epic,也进入 Oracle Cerner | 全系统使用率和续约经济性 |
采用指标均为 Nabla 指标,除非明确标注为 AMI;AMI 这一行保持 null,因为公司没有披露直接客户数量。
[CU008, CU009, CU012, CU015, CU016, CU020]| 客户 / 证据对象 | 与 AMI 的关系 | 细分领域 | 部署或用例 | 生产环境还是试点 | 成效证据 | 局限 |
|---|---|---|---|---|---|---|
| Nabla | 拥有首批访问权的战略伙伴 | 临床 AI 平台 | 未来为智能体式医疗 AI 接入 AMI 世界模型 | 战略访问;AMI 生产状态未披露 | 独家合作和首批访问公告 | 未披露 AMI 收入、许可或生产部署条款 |
| Denver Health | 通过 Nabla 间接相关 | 安全网医疗系统 | 在多类护理场景中接入 Epic 的环境式文档记录 | 试点扩展为广泛采用 | 文档时间减少 40%;职业倦怠降低 30%;300,000+ 次就诊 | 验证的是 Nabla,不是 AMI 世界模型 |
| Carle Health | 通过 Nabla 间接相关 | 综合医疗系统 | 集成 Epic 的环境式助手 | 评估后推广 | 55% 至少节省一小时;89% 愿意推荐;1,500 名服务提供者 | Nabla 案例研究没有显示 AMI 功能使用 |
| McFarland Clinic | 通过 Nabla 间接相关 | 医生所有的多专科集团 | Epic 文档支持 | 试点并披露留存 | 每月 10,000 次就诊;试点留存率 80%;100+ 名服务提供者 | 留存指标仅对应 Nabla |
| Tia Health | 通过 Nabla 间接相关 | 女性健康服务提供者 | 混合和虚拟护理文档 | 生产环境案例研究 | 笔记提交时间减少 50%;50,000+ 条笔记;90+ 名服务提供者 | 未识别出 AMI 技术 |
| Children’s Hospital Los Angeles 医院 | 通过 Nabla 间接相关 | 儿科医院 | 儿科文档和职业倦怠降低 | 生产环境案例研究 | 文档时间减少 50%;职业倦怠降低 47%;89% 当日完成笔记 | 未识别出 AMI 部署 |
| University of Toledo Health 医疗系统 | 通过 Nabla 间接相关 | 学术医疗系统 | 跨专科 Epic 文档 | 评估后转向更广泛部署 | 病历关闭速度提升 29%;积压从 >400 降至 <30 | 未披露 AMI 衍生模块 |
| Aultman Health System | 通过 Nabla 间接相关 | 综合医疗系统 | Oracle Cerner 环境式 AI 部署 | 初步成效后全系统扩张 | 每天节省 30-60 分钟;文档时间减少 20-40% | 未披露 AMI 衍生模块 |
枚举只是公开证据的部分样本:包括唯一具名的 AMI 合作伙伴,以及仅用于间接验证的若干具名 Nabla 客户。
[CU003, CU004, CU007, CU010, CU011, CU018]公开证据在合作伙伴访问权和 Nabla 部署环节较强,但到 AMI 直接收入和留存时降为零。
漏斗为定性判断,因为 AMI 未披露试点、生产或续约阶段的数量。
[CU002, CU004, CU012, CU018, CU031, CU035]Nabla 客户在部署证据上得分较好;AMI 直接客户得分较低,因为公开记录止步于战略访问权。
各行评估公开证据可见度,而不是产品质量;AMI 相关性有意与 Nabla 部署质量分开。
[CU012, CU015, CU019, CU020, CU021, CU024]6.3 留存、扩张和集中度都应按缺口承销
Nabla 采用耐久性有一些有用证据,例如 McFarland 的试点留存率和医疗系统扩张故事,但 AMI 特定留存完全未披露。买方不应从 Nabla 的环境式助手指标推断 AMI 续约行为,因为 AMI 技术仍被描述为新兴世界模型架构。早期集中度也异常二元:Nabla 是唯一具名首批接入伙伴,所以战略验证和渠道学习都集中在这一段关系里,而正式授权经济性尚未公开。上行空间来自 Nabla 从记录文书延伸到编码、EHR 命令和智能体流程,但证据包仍需要 AMI 特定试点、部署里程碑、定价、续约条款和扩张经济性。承销时,尽调负担应从“产品有多粘?”改成“什么产品、谁在用、基于什么协议、续约时钟如何启动?”在这些答案出现前,即使 Nabla 有有用的满意度或试点留存指标,留存表里 AMI 也应显示 null。[CU006, CU010, CU011, CU024, CU031, CU032]
| 指标 | 数值 / null | 细分 | 置信度 | 解读 | 尽调事项 |
|---|---|---|---|---|---|
| AMI NRR / GRR / 流失 | null | AMI 直接客户 | 中 | 公开资料没有留存经济性 | 向管理层索取分队列留存、续约日期和总 / 净留存 |
| AMI 客户满意度 | null | AMI 直接客户 | 中 | 没有直接 AMI 客户背书 | 要求与任何 AMI 试点客户或设计伙伴做推荐访谈 |
| McFarland 试点留存 | 试点后 80% | Nabla 间接客户 | 中 | Nabla 可在一个诊所案例中证明试点耐久性 | 弄清分母、时间跨度和续约状态 |
| Carle 推荐意愿 | 89% 非常可能推荐 | Nabla 间接客户 | 中 | 环境式助手释放出正向用户满意度信号 | 将临床医生 NPS 与买方续约和扩张拆开 |
| Denver 职业倦怠降低的持续性 | 30 天和 90 天均持续下降 30% | Nabla 间接客户 | 中 | 运营成效在短期随访中延续 | 索取 6 个月和 12 个月使用率及续约数据 |
| UToledo 积压和病历关闭 | 关闭速度提升 29%;积压从 >400 降至 <30 | Nabla 间接客户 | 高 | 运营证据支撑采购 ROI 叙事 | 确认推广后的持续性和合同经济性 |
null 值是 AMI 的刻意证据缺口;非 null 指标均为 Nabla 的间接信号,不能当作 AMI 留存。
[CU012, CU024, CU021, CU020, CU027, CU028]| 扩张驱动因素 | 证据 | 集中度风险 | 影响 | 尽调路径 |
|---|---|---|---|---|
| Nabla 首批访问渠道 | 独家合作并首批访问 AMI 世界模型 | 唯一具名 AMI 战略伙伴 | 首批客户学习高度依赖 Nabla | 审查合作协议、数据权利、独家性和终止权 |
| Nabla 已安装基础 | 85,000 名临床医生;130+ 家机构;2,000 万次就诊 | 全部为间接客户,不是 AMI 客户 | 可能加快分发,但若混同会夸大 AMI 证据 | 梳理哪些 Nabla 账户会测试 AMI 衍生功能 |
| 智能体式工作流扩张 | 编码、EHR 命令、住院和护理路线图 | 在单一医疗垂直内执行 | 若安全案例跑通,文档之外还有上行空间 | 索取 AMI 启用功能的产品里程碑和监管计划 |
| 工业 / 机器人扩张 | AMI 网站列出工业控制、自动化、机器人 | 没有具名设计伙伴 | TAM 很大,但没有客户背书 | 识别已签约设计伙伴和现场试点 |
| 战略投资者网络 | 战略支持者包括与工业相关的名字 | 投资者不是买方 | 可能带来准入,但证据偏弱 | 询问哪些投资者拥有商业评估权 |
| 正式 Nabla 经济条款 | STAT 称尚无正式股权或许可协议 | 变现路径不清晰 | 客户集中不一定等于收入集中 | 获取确定性的许可、收入分成和 IP 条款 |
风险评级为定性判断,因为 AMI 尚未披露客户合同、客户收入或伙伴经济性。
[CU006, CU010, CU015, CU016, CU032, CU033]6.4 在 AMI 证明生产就绪前,采购摩擦可能很高
反向信号不是 AMI 技术论点不可信,而是如果公司停留在基础研究阶段,客户证明会滞后;与此同时,竞争对手和垂直 AI 供应商会积累生产数据、信任和采购参考案例。医疗买方从战略接入走向生产前,会要求安全性、可审计性、隐私姿态、EHR 集成、监管路径、可衡量 ROI 和支持义务。Sacra 明确警告 AMI 可能很慢才能证明生产优势,Forbes 追问世界模型能否越过炒作,Sequoia 对更广泛 AI 支出的批评也强化了客户正在审视 ROI 的事实。这些来源使客户章节在 AMI 具名直接试点或商业客户前成为尽调阻塞项。采购门槛尤其高,因为 AMI 的价值主张触及安全关键决策支持,而不是低风险后台自动化。买方可以认可研究项目,同时仍会推迟采用,直到责任分配、监控、兜底行为和可衡量流程经济性在具名账户中被记录。[CU034, CU036, CU037, CU038, CU039, CU040]
| 阻力点 | 为什么买方在意 | 证据 | 严重度 | 缓释 / 尽调事项 |
|---|---|---|---|---|
| 生产准备度 | 买方依赖世界模型前,需要真实环境中的证据 | TechCrunch 称 AMI 从基础研究起步,可能需要数年 | 高 | 要求具名试点、成功标准和部署日期 |
| 监管和安全路径 | 医疗智能体式 AI 需要可审计且确定性的行为 | Nabla 将世界模型定位为可获 FDA 认证的智能体系统 | 高 | 索取监管策略、验证协议和人工监督控制 |
| EHR 与工作流集成 | 医疗系统采购能嵌入现有 Epic / Cerner 工作流的工具 | Nabla 客户案例强调 Epic 和 Oracle Cerner 集成 | 高 | 在真实 EHR 工作流中展示 AMI 启用功能 |
| ROI 审查 | AI 买方和投资者在审视收入和 AI 支出回报 | Sequoia 强调更广泛的 AI 支出转收入问题 | 中 | 证明可衡量的劳动力、质量或收入周期影响 |
| 竞品交付风险 | 能部署的对手会更快收集失败数据并建立客户信任 | Sacra 警告,聚焦部署的竞争者可能比 AMI 学得更快 | 高 | 优先打穿一个垂直生产切口,而不是铺开宽泛研究叙事 |
| 炒作 / 类别稀释 | 世界模型可能先变成标签,客户价值还没证明 | Forbes 追问 AMI 世界模型能否走出炒作 | 中 | 发布 AMI 专属基准和可被引用的部署案例 |
本表有意纳入反向和怀疑性来源,避免把合作伙伴公告当作客户证据。
[CU005, CU008, CU034, CU036, CU037, CU038]6.5 展项
07风险
7.1 主导风险是产品前、高估值的研究押注
AMI 的风险栈应自上而下读:估值和声誉先于产品证据到来,而最难的主张依赖世界模型系统,这些系统必须在受监管、安全关键领域足够可靠。公司拥有异常强的创始人可信度和很大的资产负债表期权,但这些优势也抬高门槛。严重风险不是孤立的:算力强度拖慢产品化,产品化延迟让客户证明稀缺,证明稀缺使 $3.5 billion 投前估值更难辩护,融资门槛又会增加公司在治理、安全和医疗控制充分举证前启动产品的压力。反向来源在这里很重要。Forbes 追问世界模型能否越过炒作,Sequoia 质疑基础设施支出背后的广义 AI 收入缺口,Epoch 显示前沿训练经济性可能只对资金最充足的实验室可承受。投资含义因此是 track 或 research-more,而不是按已规模化公司承销:在把本轮规模视为验证之前,要求 AMI 特定基准、受治理试点和合同证据。[CR001, CR002, CR004, CR005, CR006, CR008]
| 排名 | 风险 | 可能性 | 影响 / 严重度 | 缓释成熟度 | 剩余暴露 | 投资含义 |
|---|---|---|---|---|---|---|
| 1 | 在产品前商业化阶段就达到 $3.5B 投前估值 | 高 | 严重 | 早期:已披露融资和研究计划;没有产品证据 | 高 | 具名试点和产品包出现前,不承保规模化收入 |
| 2 | 算力强度和数据整理成本 | 高 | 高 | 早期:大额融资到位;没有公开算力计划 | 高 | 要求跑道、云 / GPU 承诺、训练预算和里程碑映射 |
| 3 | 通过 Nabla 进入医疗监管 / 责任路径 | 中高 | 高 | 部分:Nabla 表述了可获 FDA 认证的目标;FDA 路径未披露 | 高 | 在分类、PCCP、验证和责任计划形成文件前,阻断临床自主性 |
| 4 | 来自 World Labs、DeepMind、NVIDIA 和其他物理 AI 平台的竞争 | 高 | 高 | 部分:AMI 有人才和资本;产品差异化未经验证 | 中高 | 要求经基准检验的用例证据和伙伴 / 客户切口 |
| 5 | 对 LeCun 和 LeBrun 的关键人依赖 | 中高 | 高 | 部分:创始人知名度高;接班安排和团队纵深未披露 | 中高 | 审查留任、接班、技术领导层纵深和董事会监督 |
| 6 | Nabla 伙伴集中度和未披露条款 | 高 | 中高 | 部分:战略访问已公开;经济条款未披露 | 中高 | 要求协议条款、数据权利、独家期限和终止权 |
| 7 | Meta / NYU / Nabla IP 和冲突边界 | 中 | 高 | 未知:公开重叠已有记录;豁免未公开 | 中高 | 要求 IP 转让、发明披露、冲突豁免和发表流程 |
| 8 | 客户证据缺口和采购阻力 | 高 | 高 | 早期:未披露直接 AMI 客户 | 高 | 在客户证据转化前,按研究阶段处理 |
| 9 | 安全、隐私和质量控制未披露 | 中高 | 高 | 未知:外部框架存在;AMI 控制未披露 | 高 | 要求 SOC / 安全态势、临床数据治理、模型审计日志和事件响应 |
| 10 | 估值驱动的未来融资预期 | 中高 | 高 | 早期:大额融资买来时间;下一阶段证明门槛已抬高 | 中高 | 将下一轮融资绑定到技术和商业风险降低 |
定性严重度排序基于截至 2026-07-03 保留的来源;可能性和剩余暴露是分析师判断,不是公司披露评分。
[CR001, CR006, CR017, CR018, CR020, CR032]最高残余敞口出现在高可能性与估值、算力、监管、客户验证影响叠加处。
定性矩阵;未取得公司提供的风险概率。
[CR032, CR033, CR035, CR036, CR039, CR040]7.2 监管、法律和治理边界未解
法律风险比单一法规更宽。如果 AMI 技术成为临床流程组件,FDA AI-enabled-device 和 predetermined-change-control 材料就会变得相关,因为模型行为、更新、验证证据和上市后监测都是核心尽调项。在欧洲,AI Act 把高风险部署变成提供方责任问题,而不只是研究问题。UK 前沿承诺和 NIST RMF 又提供了实务上的安全测试和治理词汇。行业监管之外,IP 和冲突边界需要特别关注,因为公开记录把 LeCun 与 Meta、NYU 相连,把 LeBrun 与 Nabla、Meta/FAIR 相连,也把 AI 训练生态与针对 Meta 的活跃版权诉讼相连。这些事实都不能证明 AMI 有不当行为;它们说明,投资人在接受公司治理姿态前,应要求提供转让、数据权利、发表审查、冲突豁免和伙伴排他性文件。[CR010, CR011, CR012, CR013, CR014, CR015]
| 规则 / 案例 / 边界 | 司法辖区 | 状态 | 可能性 | 严重度 | 缓释 | 剩余暴露 | 尽调路径 |
|---|---|---|---|---|---|---|---|
| EU AI Act 高风险 / GPAI 义务 | 欧盟 | 法规已生效;适用性取决于 AMI 的产品角色 | 中 | 高 | 明确产品角色、提供者身份、GPAI / 高风险义务和合规路径 | 中高 | 将预期用例映射到 AI Act 条款和义务 |
| FDA AI 辅助医疗软件与 PCCP 预期 | 美国 | Nabla / AMI 产品若进入临床决策或器械工作流,即相关 | 中高 | 高 | 预提交计划、PCCP、验证证据、上市后监测 | 高 | 获取监管律师备忘录和 FDA 路径时间表 |
| 前沿 AI 安全测试承诺 | 英国 / 全球政策 | 自愿性质,但会为前沿模型尽调设基准 | 中 | 中高 | 安全评估计划、红队方案、发布闸门、事件报告 | 中 | 对照英国安全测试声明审查 AMI 测试方案 |
| AI 训练版权 / IP 诉讼背景 | 美国 | Kadrey v. Meta 是进行中的诉讼背景,不是针对 AMI 的主张 | 中 | 高 | 数据来源、训练许可审查、IP 赔偿保障、发表审查 | 中高 | 审查数据集、许可、Meta 来源资产、发明转让 |
| Meta / NYU / Nabla 利益冲突与权利转让边界 | 法国 / 美国 / 合作伙伴合同 | 公开资料已记录重叠;正式边界未公开 | 中高 | 高 | 利益冲突豁免、董事会批准、数据权利清单、雇佣 / IP 转让 | 中高 | 索取已签政策和交易对手同意书 |
该清单覆盖已留存来源识别出的重大法律 / 监管风险类别,不覆盖所有可能辖区。
[CR011, CR012, CR013, CR014, CR016, CR024]7.3 运营风险集中在可靠性、算力和未披露控制上
AMI 的技术承诺也是它的运营负担。面向医疗、机器人或工业控制的世界模型,不能只生成看似合理的输出;它必须支撑测试、监控、可审计性、变更控制、兜底行为和事件响应。外部材料说明这会很贵:Epoch 估计前沿训练成本升级,NVIDIA 描述物理 AI 工作负载需要 PB 级视频和大规模算力小时预算,Sequoia 警告基础设施支出必须由真实收入证明。安全和隐私叙事尚未公开。AMI 未来或许能建立强控制,但留存来源没有披露临床数据治理、模型日志、红队测试结果、访问控制或泄露响应姿态。因此,运营质量是门控风险,而不是投后增强项;尤其是受监管客户会在生产自主性前索要证据。[CR015, CR016, CR017, CR018, CR019, CR020]
| 失效模式 | 发生概率 | 严重性 | 缓释成熟度 | 剩余风险敞口 | 未解决缺口 |
|---|---|---|---|---|---|
| 世界模型在动态临床 / 物理场景中达不到可靠性门槛 | 中高 | 高 | 概念层面:外部框架存在;AMI 指标未披露 | 高 | AMI 专属基准、安全论证、兜底行为 |
| 训练和推理成本超过里程碑融资可承受范围 | 高 | 高 | 早期:大额融资已披露;算力预算未披露 | 高 | 云 / GPU 合同、烧钱速度、现金跑道、模型规模路线图 |
| 数据整理和来源证明过慢,或受法律约束 | 中高 | 高 | 未知:未公开数据集或许可计划 | 中高 | 数据集清单、合作伙伴数据权利、留存 / 删除控制 |
| 安全 / 隐私控制不足以支撑医疗或企业数据 | 中 | 高 | 未知:AMI 未披露控制措施 | 高 | 访问治理、日志、加密、事件响应、隐私评估 |
| 评估流程达不到监管方或医院风控委员会预期 | 中 | 高 | 部分:披露了 Nabla 与可按 FDA 路径认证的意图;无申报材料包 | 高 | PCCP、验证方案、监测计划、独立审计 |
| 发表 / 开放姿态泄露自研或受限的合作伙伴专有技术 | 中 | 中高 | 未知:AMI 称会发表;边界未公开 | 中 | 发表审查、IP 过滤、合作伙伴审批流程 |
各行把直接来自来源的外部要求与推断的 AMI 风险敞口合并;没有任何一行声称 AMI 已披露事故。
[CR015, CR016, CR017, CR019, CR020, CR033]算力、治理和产品验证一旦失手,风险会传导到客户采用、融资预期和估值。
DAG 是尽调监控使用的定性因果图。
[CR017, CR018, CR020, CR033, CR034, CR036]7.4 合作伙伴、人员与竞争者依赖压窄执行路径
AMI 的集中风险分三层。第一,公开具名的早期垂直路径只有 Nabla;它有战略价值,但也让 AMI 依赖一个医疗合作伙伴、一组数据权利谈判,以及一条临床验证路径。第二,AMI 与关键人物绑定很深:LeCun 提供研究可信度,LeBrun 搭起 Nabla 和运营之间的桥。公开叙事因此暴露在关键人可用性、继任和治理风险之下。第三,世界模型赛道已经不再空白。World Labs、Google DeepMind 和 NVIDIA 都在发布相邻的空间、机器人和物理 AI 工作;在 AMI 拿出公开产品证据之前,既有平台就能先把算力、工具、分发和客户关系合在一起。缓释手段不是泛泛招聘,而是让外界看见足够深的梯队、Nabla 之外的设计合作伙伴、有文件记录的合作条款,以及 AMI 能在某个具体工作流里打赢资本更充足平台的证据。[CR009, CR021, CR022, CR023, CR024, CR025]
| 依赖项 | 交易对手 | 角色 | 集中度 | 失效情景 | 严重性 | 缓释措施 | 剩余风险敞口 |
|---|---|---|---|---|---|---|---|
| 医疗首发切入点 | Nabla | 首个具名合作伙伴和垂直场景验证场 | 高 | Nabla 条款、数据权利或产品节奏未能产出 AMI 证明 | 高 | 披露合同经济性、数据权利、路线图里程碑 | 中高 |
| 临床监管路径 | FDA / 医疗系统风控委员会 | 自主临床工作流的守门人 | 医疗切入点高度集中 | 产品需要类器械控制,或卡在合规审查 | 高 | 监管策略、PCCP、验证和上市后计划 | 高 |
| 算力和物理 AI 基础设施 | 云 / GPU 供应商和数据整理栈 | 训练、仿真、推理能力 | 高 | 成本上涨或容量受限拖慢路线图 | 高 | 容量合同、模型扩展计划、支出治理 | 高 |
| 创始人研究信誉 | Yann LeCun / AMI 技术领导层 | 投资逻辑、招聘、投资人信心 | 高 | 可投入度、继任或研究死胡同削弱信心 | 高 | 梯队深度、独立技术评审、继任计划 | 中高 |
| 运营者 / 医疗桥梁 | Alexandre LeBrun / Nabla 网络 | CEO、合作伙伴桥梁、医疗语境 | 高 | 双重履历复杂性或离任削弱 GTM 路径 | 高 | 角色清晰、治理、替补梯队、合作伙伴升级预案 | 中高 |
| 学术 / 前雇主边界 | Meta、NYU、Nabla | 潜在 IP / 利益冲突相对方 | 中 | 围绕资产、发表、人员或数据权利的争议 | 高 | 权利转让、豁免、发表批准、律师审查 | 中高 |
| 世界模型生态节奏 | World Labs、DeepMind、NVIDIA | 竞争证明和平台压力 | 中高 | AMI 产品化之前,既有玩家先定义标准 | 高 | 聚焦用例、合作伙伴排他、以基准验证差异化 | 中高 |
依赖集中度是定性判断;正式合同金额、排他期限和服务水平未公开。
[CR009, CR021, CR023, CR024, CR025, CR026]| 角色 / 职能 | 依赖或缺口 | 发生概率 | 严重性 | 缓释措施 | 尽调路径 |
|---|---|---|---|---|---|
| 执行主席 / 研究方向 | LeCun 是投资逻辑可信度和人才吸引力的核心 | 中高 | 高 | 继任计划和独立技术委员会 | 访谈创始人以下的技术负责人 |
| CEO / 医疗桥梁 | LeBrun 串起 AMI、Nabla、Meta / FAIR 履历和运营者叙事 | 中高 | 高 | 角色清晰、董事会监督、副手运营者 | 审查治理和运营节奏 |
| 研究梯队 | 需要把 JEPA / 世界模型研究转成 AMI 专属系统 | 中 | 高 | 招聘里程碑和同行评审 | 索取组织架构图、论文、基准负责人 |
| 监管 / 质量负责人 | AMI 未公开监管、临床安全或质量体系负责人 | 中高 | 高 | 临床自主化之前先招聘受监管产品负责人 | 审查 FDA / EU 法律顾问和质量体系计划 |
| 安全 / 隐私负责人 | 临床 / 企业数据控制未公开 | 中 | 高 | 明确安全负责人、审计路线图、事件流程 | 索取安全计划文档 |
| GTM / 企业销售 | AMI 未披露定价、试点或客户成功团队 | 高 | 中高 | 设计伙伴打法和企业支持模型 | 审查销售管线、定价、支持人员配置 |
人才风险基于公开角色集中度和缺位证据;如能用文档证明已有非公开招聘,风险敞口可能降低。
[CR024, CR025, CR026, CR028, CR033, CR037]客户验证出现前,AMI 依赖少数关键人物、合作伙伴、基础设施和治理对手方。
依赖关系图仅反映公开证据;若能核验非公开合同,集中度可能降低。
[CR009, CR024, CR025, CR026, CR031, CR035]7.5 缓释措施必须落到可监测触发项
风险清单只有在缓释措施可衡量时才值得投资。第一道证据门槛是 AMI 专属产品产物:演示、基准测试或试点,且评估协议有记录、可复现。第二道是治理证据:IP 转让、数据权利安排、利益冲突豁免、监管分类分析和安全控制。第三道是商业证据:Nabla 之外的具名设计合作伙伴、定价或试点经济性、客户 ROI 和部署支持义务。最后一道是资本纪律:算力承诺、跑道、烧钱速度和融资预期,应当对应技术里程碑,而不是对应头条估值。单个门槛失败未必致命;但产品验证、治理和融资同时失守,就应推翻投资论点,因为这意味着本轮买到的是时间,而不是风险下降。今天的剩余风险仍然很高:公开记录里有很强的野心和融资,但 AMI 专属控制证据还不够。[CR028, CR032, CR036, CR037, CR038, CR040]
| 风险 | 可监测触发项 | 门槛 / 事件 | 行动含义 |
|---|---|---|---|
| 产品前商业化 | 出现 AMI 专属基准或试点 | 到下一轮融资流程时,仍无具名试点、基准或产品包 | 不要上调估值 |
| 算力强度 | 算力预算映射到里程碑 | 尽调期间未披露烧钱速度、现金跑道或云承诺 | 把轮次规模视为买时间,而不是降风险 |
| 医疗监管路径 | FDA / EU 监管备忘录和 PCCP 式计划 | 没有分类、验证、上市后监测或责任分配 | 阻断临床自主化投资论证 |
| Nabla 集中度 | 审阅已签合作伙伴条款 | 没有经济性、排他期限、数据权利或终止权 | 大幅折价渠道证明 |
| 客户证明缺口 | 直接客户或设计伙伴证据 | 公开证据仍只有 Nabla 相关证明 | 建议维持在继续研究 / 观察 |
| IP / 利益冲突边界 | 出具权利转让和豁免文件 | 没有 Meta / NYU / Nabla 利益冲突与 IP 材料包 | 升级为法律阻断项 |
| 运营质量 | 独立安全 / 评估报告 | 没有可复现评估方案或红队结果 | 阻断生产级风险投资论证 |
| 安全 / 隐私 | 安全和临床数据控制材料包 | 没有日志、访问控制、隐私或事件响应计划 | 阻断医疗部署投资逻辑 |
| 竞争压力 | 相对 World Labs / DeepMind / NVIDIA 同类方案的差异化基准 | 目标工作流中没有经基准验证的差异化 | 假设护城河受挤压 |
| 估值预期 | 基于里程碑的融资计划 | 下一轮主要依赖创始人声望或品类热度 | 将估值立场标为昂贵 / 偏高 |
否决标准是根据当前证据缺口推导出的投资人监测门槛;不是公司提供的契约。
[CR028, CR032, CR036, CR037, CR038, CR040]7.6 展项
08估值
8.1 当前价格与建议
AMI 应被视为一笔定价极高的特殊研究型公司融资,而不是常规种子期软件投资。公开记录支持本轮规模、投资人质量和 $3.5 billion 投前估值,但不支持收入倍数测算,因为 AMI 尚未披露收入、ARR、定价、毛利率或付费客户。这个缺口比通常更关键:隐含投后价格已经约为 $4.5 billion,而商业化可能还要数年。因此立场应是跟踪或继续研究:若条款受到保护,可以继续尽调、保留选择权;但不要仅凭公开证据按当前报道价格承销买入。IC 真正要问的不是 AMI 是否令人印象深刻,而是私下条款和里程碑能否补偿过早买入的风险。[CV001, CV002, CV003, CV005, CV006, CV008]
| 决策项 | 章节立场 | 证据基础 | 决策含义 |
|---|---|---|---|
| 建议 | 观察 / 继续研究;不要仅凭公开证据按据报道的种子轮价格给出买入判断 | 创始人与投资人质量强,但产品收入、定价和客户证明未披露 | 只推进到尽调或受保护的内部人条款,不接受无保护加价 |
| 置信度 | 中低 | 融资事实和合作伙伴说法公开;经济性和合同事实仍是私有信息 | 如果 AMI 披露产品牵引或条款,建议可迅速上调 |
| 风险评级 | 高 | 产品前、算力重、受监管、合作伙伴集中且竞争拥挤 | 配置资本前需要明确否决标准 |
| 估值立场 | 昂贵 / 偏高 | 披露收入或 SKU 之前,隐含投后估值约 $4.5B | 价格需要里程碑保护,或更低有效入场价 |
| 入场纪律 | 基于里程碑分期,或等待 | 股权结构表条款、优先权和按比例认购权未公开 | 任何价格判断落地前,先索取条款 |
| 退出姿态 | 尚未具备退出条件 | IPO / M&A 逻辑需要商业证明,目前还没有 | 把它视作长期研究期权,而不是近期流动性 |
这些立场是从公开证据推导出的定性投资判断;不主张目标价格或回报。
[CV001, CV002, CV008, CV033, CV038, CV039]决策链把前沿实验室的强成色和公开经济性的弱证据,落到「跟踪 / 继续研究」立场。
定性决策流;边的方向体现投资逻辑,而非因果证明。
[CV033, CV034, CV038, CV039, CV047, CV049]8.2 投资论点与反论点
可投资的论点是:世界模型成为新的物理 AI 前沿栈,AMI 赢下足够的人才、算力、合作伙伴数据和安全可信度,从而定义这一层。这个判断并不空想:AMI 有战略支持方,World Labs 和 Physical Intelligence 证明投资人愿意押注相邻的世界模型和机器人基础模型,NVIDIA 和 DeepMind 也把物理 AI 验证为一条重要技术方向。反论点同样重要。同样的关注度可能在买方看到持久产品价值之前,把世界模型变成融资标签。AMI 首个具名合作伙伴 Nabla 提供了有用验证,但公开证据尚未披露合同经济性、数据权利、FDA 路径归属或客户转化。因此,投资论点可信,但在所报道价格下还不足以承销,尤其外部报道仍把本轮描述为在收入出现前给顶尖研究员支付的溢价。[CV007, CV010, CV011, CV012, CV013, CV020]
| 论点 | 支持证据 | 改变判断的因素 | 当前权重 |
|---|---|---|---|
| 投资逻辑:世界模型成为新的前沿技术栈 | AMI、World Labs、NVIDIA、DeepMind 和机器人同行都指向物理 AI / 世界模型动能 | AMI 发布 AMI 专属基准、产品架构和合作伙伴部署 | 正面但未证实 |
| 投资逻辑:创始人与投资人质量能赢得人才 / 算力 | Yann LeCun、Alexandre LeBrun 和战略投资人释放强信号 | 确认高级招聘、算力供应和数据合作 | 正面 |
| 投资逻辑:医疗合作伙伴给出可信切入点 | Nabla 的首发接入关系给安全关键学习提供真实场域 | 已签经济条款、数据权利范围和对齐 FDA 的验证计划 | 正面但集中 |
| 投资逻辑:市场能支撑超大赢家 | 具身 AI 预测和同行融资显示资本胃口很大 | AMI 能捕获软件经济性,而不只是研究声望的证据 | 有条件 |
| 反向逻辑:估值先于产品证明 | 据报道约 $4.5B 投后估值出现在 ARR、SKU、定价或客户披露之前 | 公开试点结果、收入合同或受保护定价条款 | 高权重 |
| 反向逻辑:护城河可见前,算力先烧掉资本 | Epoch 和 Sequoia 将前沿 AI 定义为资本密集,收入缺口仍不确定 | 预算纪律、模型效率证明和供应商承诺 | 高权重 |
| 反向逻辑:品类可能沦为炒作标签 | TechCrunch 援引 AMI 领导层警告:世界模型可能变成融资热词 | 独立基准把 AMI 和泛泛世界模型主张区分开 | 中高 |
| 反向逻辑:公司平台和开放平台把这一层商品化 | NVIDIA Cosmos 和 DeepMind 机器人研究降低世界模型工具的稀缺性 | 专有数据权利、垂直工作流集成和客户切换成本 | 中高 |
表格衡量证据质量,而非预期回报;每一行在资本配置前都需要后续尽调。
[CV010, CV020, CV021, CV028, CV034, CV035]AMI 的野心和资本可得性得分高,公开经济性验证和价格支撑得分低。
评分是定性的投委会就绪度指标;除上述风险强度外,分数越高越好。
[CV006, CV007, CV008, CV028, CV031, CV032]8.3 情景与进场纪律
公开证据缺少 AMI 财务数据,情景分析不应编造退出价值或目标回报。真正有用的是看里程碑能否支撑价格。牛市情景里,AMI 发布产品基准,Nabla 拿出受监管工作流验证,合作伙伴数据或算力形成稀缺平台,当前价格才可能合理。基准情景里,AMI 仍是一家高质量实验室;在产品和客户证据出现前,估值应保持不变。熊市情景里,算力成本、监管摩擦、估值周期压缩,或开放 / 企业平台跑在 AMI 证据之前,迫使公司平轮或下轮融资。因此,进场纪律应是结构性的:分期、按比例跟投权、治理报告和更低的有效进入价格,比无保护加价更站得住。[CV024, CV025, CV026, CV027, CV035, CV036]
| 情景 | 假设 | 估值 / 回报逻辑 | 概率信号 | 下行触发因素 |
|---|---|---|---|---|
| 乐观 | AMI 拿出差异化世界模型基准,Nabla 跑通受监管工作流证明,合作伙伴提供数据 / 算力 | 如果 AMI 成为稀缺的物理 AI 平台,当前约 $4.5B 的投后估值可以自洽;回报仍取决于情景 | World Labs、Mistral、Figure、Physical Intelligence 和 Wayve 表明,投资者愿意以数十亿美元估值押注前沿物理 AI | 下一轮融资前仍没有 AMI 专属产品或合作伙伴经济性 |
| 基准 | AMI 仍是高质量研究实验室,投资人阵容强,但公开客户证据有限 | 应以观望 / 跟踪为主;估值应按里程碑校准,而不是套收入倍数 | TechCrunch 称商业化可能需要数年,AMI 目前暂无收入计划 | 在没有产品或收入证明的情况下,下一轮定价高于当前隐含投后估值 |
| 悲观 | 算力成本、监管摩擦、合作伙伴集中度或平台商品化压力超过证明进展 | 即便团队质量强,当前价格也容易承受平轮 / 下轮压力或重度稀释 | Sequoia 与 Reuters 的泡沫警示,加上 Epoch 的算力成本压力,都是反向信号 | 开放平台或直接同行在 AMI 之前展现更强产品牵引 |
由于公开证据不足以支撑虚构 AMI 收入预测或目标回报,情景采用定性里程碑逻辑。
[CV002, CV008, CV024, CV025, CV027, CV035]估值判断最受产品验证、客户经济性、算力计划和融资条款影响。
评分是 1-10 的定性敏感度,不是财务预测。
[CV024, CV025, CV027, CV035, CV036, CV040]公开证据只能支撑入场估值防御性的情景区间,不能支撑虚构回报目标。
指数表达当前估值在不同证据情景下的防御性,不是回报预测。
[CV002, CV035, CV036, CV037, CV044, CV046]8.4 可比公司与市场背景
可比组同时支持两个结论。第一,前沿 AI、具身 AI 和世界模型公司可以在成熟利润出现前拿到数十亿美元估值:World Labs、Mistral、Physical Intelligence、Figure、Wayve 和 Skild 都显示,相邻主题有大量资本需求。第二,AMI 的商业证据弱于其中几个参照。Mistral 有战略产业投资人和已部署的模型业务;World Labs 有 Marble;Physical Intelligence 和 Figure 展示了机器人演示或产品叙事;CoreWeave 和 NVIDIA 则是带有基础设施经济性的公开市场或申报参照。因此,AMI 的估值在过热的前沿 AI 市场里并不荒唐,但相对 AMI 专属公开证据仍然偏高。[CV012, CV013, CV014, CV015, CV016, CV017]
| 可比对象 | 融资 / 估值信号 | AMI 参考意义 | 局限 |
|---|---|---|---|
| AMI | 种子轮 $1.03B,投前估值 $3.5B,隐含投后约 $4.5B | 本估值章节的直接入场价格 | 没有公开收入、优先权、产品或客户经济性 |
| World Labs | 2026 年融资 $1B;此前以 $1B 估值融资 $230M;据报洽谈估值约 $5B,且有 Marble 产品证据 | 最接近的世界模型 / 空间 AI 私营同行 | 估值报道不如官方融资确定;产品偏创意 / 3D,不是 AMI 的医疗 / 工业方向 |
| Physical Intelligence | 2024 年以 $2.4B 投后估值融资 $400M;据报 2026 年洽谈估值超过 $11B | 机器人基础模型同行;当机器人演示存在时,物理 AI 溢价成立 | 后续估值来自报道;商业模式和收入仍未公开 |
| Mistral AI | C 轮 €1.7B,投后估值 €11.7B;ASML 投资 €1.3B,取得完全稀释后约 11% | 欧洲前沿 AI 实验室,拥有战略工业投资者 | Mistral 的模型 / 产品分发已超出 AMI 当前公开状态 |
| Figure AI | 超过 $1B 的 C 轮承诺投资,投后估值 $39B | 体现具身 AI / 人形机器人叙事的极端溢价 | 人形硬件 / 制造牵引不是 AMI 当前证据集 |
| Wayve | $1.05B C 轮,由 SoftBank 领投,用于具身 AI 自动驾驶 | 显示物理世界 AI 部署能吸引战略资本 | 汽车 AV 产品化不同于 AMI 研究阶段的世界模型 |
| Skild AI | A 轮 $300M,估值 $1.5B | 机器人基础模型早期估值基准 | 轮次更小,机器人定位更直接 |
| CoreWeave | S-1/A 披露 IPO 发行价区间为每股 $47-$55,面向 AI 云基础设施 | AI 基础设施需求和算力稀缺的公开市场路径 | 已有收入的云基础设施,不可与 AMI 这种收入前研究机构直接相比 |
| NVIDIA | 2025 财年文件披露非关联方持有市值约 $2.7T | 显示 AI 基础设施供应商捕获的公开市场价值 | 即便模型实验室回报滞后,供应商经济性也能受益于创业公司支出 |
| AI 巨额融资市场 | SiliconAngle/PitchBook 报道,2026 年 Q1 美国风投交易额为 $267.2B,其中 5 笔交易占 73% | 设定 AI 资本集中和 xAI 级融资胃口的宏观背景 | 巨额融资会抬高可比估值,但不能证明 AMI 自身价值 |
列举样本来自为价格背景审阅的前沿 AI、物理 AI 和基础设施参照,并非完整私募市场可比公司集。
[CV001, CV002, CV012, CV013, CV014, CV015]8.5 尽调请求、退出准备度与论点失效触发项
最后的尽调清单应聚焦能改变建议的证据,而不是泛泛评估公司质量。必须拿到的请求包括融资条款、产品验证、Nabla 经济性、算力承诺、监管准备度、专有数据权利和治理可见度。今天的退出准备度很低,因为可信 IPO 或大型战略收购叙事,需要产品、收入、监管或基础设施规模证据,而这些证据对 AMI 并不公开。若 AMI 在没有产品或客户验证的情况下,以高于当前隐含投后估值再融资;若 Nabla 仍只是营销关系;若算力烧钱不透明;或若 NVIDIA、DeepMind、World Labs、Physical Intelligence 或开源系统在 AMI 建成专有数据和分发前商品化这一层,投资论点就会失效。[CV031, CV032, CV043, CV044, CV045, CV046]
| 触发因素 | 阈值 | 对投资论点的传导 | 行动含义 |
|---|---|---|---|
| 无保护加价 | 下一轮定价高于当前隐含投后估值,且未披露 SKU、收入或具名客户 | 把质量故事变成估值风险 | 没有重置或保护就不参与 |
| 无产品证明 | 下一轮融资窗口前没有 AMI 专属基准、演示、API、试点或产品规格 | 世界模型论点仍停留在未转化的研究 | 从跟踪下调为回避 |
| 缺少 Nabla 经济性 | 首个准入合作关系没有合同经济性、数据权利、监管责任方或使用指标 | 合作伙伴证明停留在营销,而非商业化 | 投资前要求客户尽调 |
| 算力计划不透明 | 没有预算、供应协议、模型效率路径或烧钱敏感性 | 无法承销资金续航和稀释风险 | 任何投资都以运营计划为条件 |
| 监管阻滞 | 医疗使用场景无法映射到 FDA/EU 义务或安全验证 | 限制最有价值的早期切入口 | 转向非医疗用例证明,或暂停 |
| 平台商品化 | NVIDIA、DeepMind 或开放平台在 AMI 建立数据护城河前就追平世界模型能力 | 压缩定价权和退出可选性 | 要求专有数据或垂直分发证明 |
| 治理缺口 | 在高估值下,条款仍缺少信息权、里程碑报告或投资人保护 | 投资人无法监测投资论点何时失效 | 拒绝或重新谈判条款 |
否决触发因素是尽调阈值;没有任何一项假设公开目标价或保证回报。
[CV031, CV032, CV035, CV036, CV040, CV041]| 主题 | 缺失证据 | 重要性 | 尽调路径 |
|---|---|---|---|
| 轮次条款 | 股权结构表、清算优先权、分批交割安排、期权池、按比例跟投权和信息权 | 决定有效入场价格和下行保护 | 索取融资文件和投资人附函 |
| 产品证明 | AMI 专属演示、技术规格、基准和产品路线图 | 把研究声誉转化为可承销的里程碑证据 | 审阅基准包、路线图和独立评估 |
| Nabla 经济性 | 合同价值、数据权利、责任分配、监管计划和试点里程碑 | 唯一具名合作伙伴证明贴近医疗且集中 | 访谈 Nabla 管理层并审查协议 |
| 算力计划 | 预算、供应商承诺、训练运行、推理成本目标和模型效率 | 算力强度会吃掉资金续航并迫使稀释 | 审阅董事会预算、云 / GPU 合同和敏感性模型 |
| 人才与治理 | 关键人物依赖、招聘计划、留任方案和董事会控制 | 创始人主导的实验室一旦招聘或治理滑坡就会脆弱 | 访谈领导层并审阅组织计划 |
| 监管准备度 | FDA/EU AI Act 分类、安全论证、审计日志和质量管理计划 | 安全关键用例商业化前需要合规证据 | 开展监管律师审查和红队计划 |
| 竞争护城河 | 数据权利、合作伙伴排他性、模型评估差异化和开源姿态 | NVIDIA 和 DeepMind 能将通用世界模型工具商品化 | 将 AMI 基准与 Cosmos、Gemini Robotics、World Labs 和 PI 对比 |
| 退出路径 | 战略收购方地图、收入里程碑和 IPO 准备前置条件 | 没有商业证明,退出可选性仍是推测 | 把里程碑映射到潜在战略买家和公开市场可比对象 |
尽调问题聚焦那些足以改变建议、置信度或估值立场的证据。
[CV040, CV041, CV042, CV043, CV045, CV046]8.6 展项
免责声明
本报告基于章节证据中收集的公开来源,不构成投资、法律、税务或会计建议。
证据索引
| 编号 | 陈述 | 可信度 | 来源 |
|---|---|---|---|
| CO001 | ADVANCED MACHINE INTELLIGENCE is a French SASU legal entity tied to SIREN 994675254. | 高 | SO005, SO006 |
| CO002 | Pappers lists the company’s RCS status as registered in Paris on 2025-12-15. | 高 | SO005, SO006 |
| CO003 | The current registered headquarters is 10 rue de Penthièvre, 75008 Paris, and Pappers shows the main Paris establishment was created on 2026-02-12. | 中 | SO005 |
| CO004 | The official public domain retained for the company overview is amilabs.xyz. | 高 | SO001, SO002 |
| CO005 | AMI describes its product direction as world models that learn abstract representations of real-world sensor data rather than token-only language models. | 高 | SO001, SO002 |
| CO006 | AMI publicly launched its partnership narrative in December 2025 and was profiled as a newly announced venture in January 2026. | 高 | SO008, SO009 |
| CO007 | The company announced its $1.03B financing on 2026-03-10 after press coverage and partner materials had surfaced the venture in late 2025 and January 2026. | 高 | SO002, SO003, SO011 |
| CO008 | AMI says it is operating from Paris, New York, Montreal, and Singapore from day one. | 高 | SO002, SO003, SO011 |
| CO009 | The company’s public hiring funnel points applicants to an AMI Ashby jobs page. | 中 | SO002, SO022 |
| CO010 | AMI’s announced leadership includes Yann LeCun as chair, Alexandre LeBrun as CEO, Laurent Solly as COO, Saining Xie as chief science officer, Pascale Fung as chief research and innovation officer, and Michael Rabbat as VP of world models. | 高 | SO002, SO003, SO017 |
| CO011 | MIT Technology Review described LeCun as executive chairman, while AFP-syndicated coverage described him as non-executive chairman. | 中 | SO007, SO011, SO012 |
| CO012 | Alexandre LeBrun moved from Nabla’s CEO role into AMI’s CEO role while becoming Nabla’s chief AI scientist and chairman. | 高 | SO008, SO009 |
| CO013 | Laurent Solly’s COO role adds former Meta Europe operating leadership to a company otherwise anchored by research scientists. | 中 | SO002, SO013, SO014 |
| CO014 | Saining Xie, Pascale Fung, and Michael Rabbat give the founding bench coverage in visual representation learning, human-centered AI, and world-model research. | 高 | SO002, SO003, SO017 |
| CO015 | AMI remains highly key-person-dependent because the public narrative is concentrated around LeCun’s scientific thesis and LeBrun’s operator role. | 中 | SO003, SO007, SO018 |
| CO016 | AMI raised $1.03B, approximately €890M, in its first disclosed financing round. | 高 | SO002, SO003, SO004 |
| CO017 | The financing was priced at a $3.5B pre-money valuation, implying roughly a $4.5B post-money valuation if the full $1.03B round is added to the pre-money figure. | 高 | SO003, SO004, SO018 |
| CO018 | Cathay Innovation, Greycroft, Hiro Capital, HV Capital, and Bezos Expeditions co-led the round. | 高 | SO002, SO004, SO018 |
| CO019 | Strategic and long-term backers publicly named for the round include NVIDIA, Samsung, Temasek, Toyota Ventures, Sea, SBVA, and Alpha Intelligence Capital. | 高 | SO002, SO003, SO004 |
| CO020 | Independent coverage described the financing as Europe’s largest seed round. | 高 | SO003, SO016, SO018 |
| CO021 | The round’s stated use of proceeds is long-term research, global hiring, compute, and development of reliable intelligent systems. | 高 | SO003, SO004 |
| CO022 | TechCrunch reported LeBrun saying AMI has no near-term revenue plan and could take years before world models become commercial applications. | 中 | SO003 |
| CO023 | The Next Web summarized the same risk as no product, no revenue, and no near-term prospect of either while AMI focuses on R&D. | 中 | SO018 |
| CO024 | Nabla is the first named strategic partner expected to receive early or first access to AMI world-model technologies. | 高 | SO003, SO009, SO010 |
| CO025 | STAT reported that there was not yet a formal equity or licensing agreement relationship between AMI and Nabla even though the companies were working closely together. | 中 | SO020 |
| CO026 | Nabla frames the AMI relationship around healthcare workflows where LLMs face hallucination, non-determinism, and multimodal-data limits. | 中 | SO009, SO010 |
| CO027 | Nabla says world models could support deterministic, auditable decision-making, simulation-based reasoning, multimodal medical signals, and a regulatory path for agentic clinical AI. | 中 | SO009, SO010 |
| CO028 | AMI’s official mission names industrial process control, automation, wearable devices, robotics, healthcare, and beyond as target areas where reliability and safety matter. | 高 | SO001, SO008, SO015 |
| CO029 | AFP-syndicated coverage reported LeCun saying AMI would focus on R&D in its first year and could hold partner discussions within six to twelve months. | 高 | SO011, SO012 |
| CO030 | France 24 reported LeCun’s goal of fairly universal intelligent systems within three to five years. | 高 | SO011, SO012 |
| CO031 | AFP-syndicated coverage reported LeCun saying AMI would hire 20 to 30 people in the very short term after the round. | 中 | SO011, SO012, SO017 |
| CO032 | Forbes flagged video-model training cost, long task inference cost, healthcare regulation, audit requirements, and incident reporting as execution risks for AMI-style world models. | 中 | SO019 |
| CO033 | Sequoia warned that AI infrastructure spend implies a large revenue gap, and Reuters-sourced U.S. News coverage warned that early-stage AI valuations can look frothy. | 高 | SO023, SO024 |
| CO034 | Le Monde Informatique’s March 2026 report noted that AMI had attracted international investors without yet having operational systems. | 中 | SO014 |
| CO035 | Pappers listed no available annual accounts for ADVANCED MACHINE INTELLIGENCE as of the reviewed registry page. | 中 | SO005 |
| CO036 | Pappers listed zero collective proceedings, zero litigation entries, and zero sanctions on the reviewed company page. | 中 | SO005 |
| CO037 | World Labs and NVIDIA Cosmos show that world models and physical AI are already an active competitive category rather than an AMI-only concept. | 高 | SO016, SO027, SO028 |
| CO038 | Crunchbase and TechCrunch both framed AMI as part of a smaller but increasingly funded world-model category alongside World Labs. | 高 | SO003, SO016 |
| CO039 | LeCun’s JEPA thesis is a stated technical foundation for AMI, and arXiv/Meta materials show the architecture lineage predates the startup. | 高 | SO007, SO025, SO026 |
| CO040 | AMI and TechCrunch state that the company intends to publish papers and open-source significant code as it builds the research ecosystem. | 高 | SO001, SO003 |
| CO041 | LeCun told MIT Technology Review that AMI systems would train on video, audio, and sensor data, including robot-arm position, lidar, and audio. | 中 | SO007 |
| CO042 | AFP coverage described prospective world-model applications such as jet engines, power plants, and organs of a human patient. | 高 | SO011, SO012 |
| CO043 | French President Emmanuel Macron publicly praised LeCun’s AMI launch as a new page in artificial intelligence for France. | 中 | SO011, SO013 |
| CO044 | Investor demand was reportedly high enough that AMI could choose backers for alignment and value-add rather than only capital availability. | 中 | SO003, SO018 |
| CO045 | The company’s public filings and coverage do not disclose ARR, revenue run-rate, customer count beyond Nabla, or current total headcount. | 中 | SO003, SO005, SO020, SO022 |
| CM001 | AMI frames its mission around world models that learn from real-world sensor data and plan safely rather than around generic enterprise chat software. | 高 | SM001, SM003 |
| CM002 | Nabla positions AMI’s first disclosed external use case around agentic healthcare workflows that need lower hallucination risk and auditable action. | 高 | SM004, SM002 |
| CM003 | NVIDIA defines physical AI development around world foundation models for robots and autonomous vehicles. | 高 | SM005, SM006 |
| CM004 | NVIDIA Cosmos is designed to generate photoreal, physics-based synthetic data and support evaluation for robots and autonomous vehicles. | 高 | SM005, SM006 |
| CM005 | Google DeepMind describes Gemini Robotics as a vision-language-action model that turns visual information and instructions into robot motor commands. | 高 | SM007, SM008 |
| CM006 | DeepMind’s Genie 2 is a foundation world model for generating action-controllable 3D environments for embodied-agent training and evaluation. | 中 | SM009 |
| CM007 | Physical Intelligence’s π0 work argues that robot foundation models require broad multi-robot data because robotics lacks a web-scale equivalent of language data. | 高 | SM010, SM011 |
| CM008 | Hugging Face’s LeRobot standardizes robotics models, datasets, and hardware interfaces, lowering experimentation barriers for physical AI developers. | 中 | SM012 |
| CM009 | MarketsandMarkets forecasts embodied AI rising from USD 4.44B in 2025 to USD 23.06B in 2030 at a 39.0% CAGR. | 中 | SM017 |
| CM010 | The MarketsandMarkets embodied-AI definition includes robots, exoskeletons, autonomous systems, and smart appliances, making it broader than AMI’s likely software/model revenue pool. | 中 | SM017 |
| CM011 | ABI Research explicitly tracks physical-AI robotics foundation models as an ecosystem category, supporting the narrower market label even without public revenue details. | 中 | SM018 |
| CM012 | IFR reported 542,000 global industrial robot installations in 2024, more than double the level from ten years earlier. | 中 | SM019 |
| CM013 | IFR reported Asia represented 74% of 2024 industrial robot deployments, while China alone installed 295,000 industrial robots. | 中 | SM019 |
| CM014 | IFR reported almost 200,000 professional service robots sold in 2024 and around 16,700 medical robots sold, with medical robots up 91%. | 中 | SM019 |
| CM015 | Deloitte found that 92% of surveyed manufacturers believe smart manufacturing will be the main driver of competitiveness over the next three years. | 中 | SM020 |
| CM016 | Deloitte reported average net impacts from smart-manufacturing initiatives of 10%-20% production-output improvement, 7%-20% employee-productivity improvement, and 10%-15% unlocked capacity. | 中 | SM020 |
| CM017 | Deloitte reported that 29% of surveyed manufacturers use AI/ML and 24% use generative AI at facility or network level, while many remain in pilots. | 中 | SM020 |
| CM018 | Bain says humanoid deployments remain mostly early-stage and highly structured despite about USD 2.5B of venture investment in 2024. | 中 | SM021 |
| CM019 | Bain identifies battery life, dexterity, ecosystem readiness, safety certification, and public trust as constraints on humanoid robot deployment. | 中 | SM021 |
| CM020 | MIT Technology Review reported roboticist skepticism that humanoid fleets are already useful at scale and argued adoption is likely slow, industry-specific, and drawn out. | 中 | SM022 |
| CM021 | The EU AI Act creates a uniform legal framework for AI systems in the Union, raising compliance relevance for safety-critical physical-AI deployments. | 中 | SM023 |
| CM022 | FDA’s AI/ML SaMD plan emphasizes lifecycle oversight, real-world performance monitoring, and predetermined change-control concepts for adaptive medical software. | 中 | SM024 |
| CM023 | NIST AI RMF 1.0 organizes AI risk management around Govern, Map, Measure, and Manage functions. | 中 | SM026 |
| CM024 | The UK Frontier AI Safety Commitments call for risk assessments, thresholds for intolerable severe risk, red-teaming, cybersecurity safeguards, and public reporting of limitations. | 中 | SM025 |
| CM025 | Epoch AI estimates frontier-model training costs have grown about 2.4x per year since 2016 and could exceed USD 1B for the largest training runs by 2027. | 中 | SM027 |
| CM026 | Sequoia’s AI’s $600B Question argues that the revenue gap implied by AI infrastructure spending had expanded to roughly USD 500B. | 中 | SM028 |
| CM027 | Reuters-reported investor commentary warned that early-stage AI valuations were becoming frothy as AI startups captured a majority of global venture funding in early 2025. | 中 | SM029 |
| CM028 | Wayve raised USD 1.05B to develop embodied-AI products for automated driving and says OEMs can upgrade vehicles from L2+ to L4 automation as models advance. | 中 | SM013 |
| CM029 | Skild AI raised USD 300M at a USD 1.5B valuation and positions its model as a general-purpose brain for robots across embodiments and industries. | 中 | SM014 |
| CM030 | CNBC reported Physical Intelligence raised USD 400M at a USD 2.4B post-money valuation while pursuing general-purpose AI for robots. | 中 | SM015 |
| CM031 | Figure says Helix is a generalist VLA that controls humanoid upper-body actions, runs onboard embedded GPUs, and handles novel household objects from natural-language prompts. | 中 | SM016 |
| CM032 | AMI’s relevant market excludes generic horizontal AI software unless the product is tied to physical sensing, simulation, planning, or safe action. | 中 | SM001, SM003, SM005, SM017 |
| CM033 | Published market estimates are contradictory for AMI underwriting because some count hardware and complete robotics systems while AMI would likely monetize only a model, software, or co-development layer. | 中 | SM017, SM019, SM020 |
| CM034 | A constrained 2030 software/model-layer SAM of USD 2.3B to USD 6.9B equals 10%-30% of the MarketsandMarkets USD 23.06B broad embodied-AI forecast. | 中 | SM017 |
| CM035 | AMI’s near-term SOM is more likely to come from paid pilots and co-development in industrial, robotics, mobility, or healthcare-adjacent workflows than from broad self-serve SaaS. | 中 | SM004, SM013, SM020 |
| CM036 | Operations and automation leaders are the likely economic buyers in manufacturing because they own throughput, uptime, labor, and ROI budgets. | 中 | SM020, SM019 |
| CM037 | Robotics engineering teams are likely users for model-training, simulation, evaluation, data, and cross-embodiment tooling. | 中 | SM005, SM007, SM010, SM012 |
| CM038 | Clinical workflow and platform leaders are plausible healthcare-adjacent buyers only if AMI can meet auditability, monitoring, and regulatory expectations. | 中 | SM004, SM024 |
| CM039 | Strategic innovation teams can fund AMI pilots, but production adoption requires transfer to operating, clinical, or engineering budget owners. | 中 | SM013, SM014, SM020 |
| CM040 | Industrial robot deployment, labor shortages, reshoring, simulation, synthetic data, and multimodal foundation-model progress jointly support physical-AI demand. | 中 | SM005, SM019, SM020, SM021 |
| CM041 | Commercialization is constrained by reliability, ROI proof, integration with existing automation stacks, regulation, safety, compute intensity, and immature benchmarks. | 中 | SM020, SM021, SM022, SM023, SM027 |
| CM042 | Simulation and synthetic-data capabilities are especially important because real-world robotics and AV data collection is expensive, scarce, and hard to cover for edge cases. | 中 | SM005, SM006, SM009, SM010 |
| CM043 | Buyer-grade benchmarks remain immature because public sources emphasize demos, papers, and previews rather than standardized ROI, safety, and uptime metrics across workflows. | 中 | SM016, SM021, SM022 |
| CM044 | Open tooling and datasets such as LeRobot increase experimentation but also raise commoditization pressure on any closed model layer that lacks proprietary data or validation advantages. | 中 | SM012, SM010, SM011 |
| CM045 | Nabla is evidence of a strategic healthcare entry path for AMI, but not yet proof of repeatable revenue or broad buyer willingness to pay. | 中 | SM004, SM024 |
| CM046 | The market is simultaneously attractive and risky: capital and adoption signals are strong, while robotics constraints and AI-capex economics warn against over-scaling the TAM. | 中 | SM018, SM021, SM022, SM027, SM028, SM029 |
| CP001 | AMI is developing world models that learn abstract representations of real-world sensor data and predict in representation space. | 高 | SP001, SP003 |
| CP002 | AMI lists industrial process control, automation, wearable devices, robotics, healthcare, and beyond as application areas where reliability, controllability, and safety matter. | 中 | SP001 |
| CP003 | AMI disclosed a $1.03 billion round co-led by Cathay Innovation, Greycroft, Hiro Capital, HV Capital, and Bezos Expeditions. | 高 | SP002, SP003 |
| CP004 | AMI’s disclosed strategic backers include Toyota Ventures, NVIDIA, Samsung, Publicis Groupe, Bpifrance Digital Venture, Temasek, and Sea. | 高 | SP002, SP003 |
| CP005 | Nabla announced an exclusive strategic partnership under which it would gain first access to AMI world-model technologies. | 中 | SP004 |
| CP006 | The Nabla partnership is healthcare-oriented and does not by itself prove AMI distribution into robotics, manufacturing, or autonomous-vehicle buyers. | 中 | SP004, SP001 |
| CP007 | MIT Technology Review reported LeCun’s view that AMI’s focus on physical-world world models is different from Meta’s generative-AI and LLM focus. | 中 | SP005 |
| CP008 | MIT Technology Review reported AMI recruitment from OpenAI, Google DeepMind, and xAI, supporting talent access but not commercial traction. | 中 | SP005 |
| CP009 | World Labs made Marble generally available as a multimodal world model for creating 3D worlds from text, images, video, or coarse 3D layouts. | 高 | SP006, SP008 |
| CP010 | Marble can export generated worlds as Gaussian splats, meshes, or videos and includes interactive editing and expansion workflows. | 高 | SP006, SP008 |
| CP011 | TechCrunch reported Marble subscription tiers of Free, Standard at $20 per month, Pro at $35 per month, and Max at $95 per month. | 中 | SP008 |
| CP012 | World Labs announced $1 billion in new funding, and TechCrunch reported Autodesk invested $200 million as part of that larger round. | 高 | SP007, SP009 |
| CP013 | World Labs and Autodesk planned research- and model-level collaboration, but TechCrunch reported the partnership was early and did not include data sharing. | 中 | SP009 |
| CP014 | World Labs’ near-term Marble use cases are gaming, visual effects, virtual reality, and design, with robotics framed as potential simulation use rather than deployed robot control. | 高 | SP006, SP008 |
| CP015 | Physical Intelligence describes its goal as developing learning algorithms to create a model that can control any robot to do any task. | 中 | SP010 |
| CP016 | Physical Intelligence’s π0 model spans images, text, and actions, can control multiple robot types, and can be prompted or fine-tuned for tasks. | 中 | SP011 |
| CP017 | Physical Intelligence’s 2026 homepage listed π0.7, embodied memory, online reinforcement learning, partner applications, and open-sourcing π0 as active model-line developments. | 中 | SP010 |
| CP018 | CNBC reported Physical Intelligence raised $400 million at a $2.4 billion post-money valuation, and TechCrunch later reported discussions for about $1 billion at more than $11 billion valuation. | 高 | SP012, SP013 |
| CP019 | TechCrunch reported Physical Intelligence had no timeline for commercialization despite substantial funding and compute appetite. | 中 | SP013 |
| CP020 | Skild AI announced $300 million in Series A funding at a $1.5 billion valuation to scale its robotics foundation model and team. | 中 | SP015 |
| CP021 | Skild AI claims its general-purpose robot brain generalizes across manipulation, locomotion, navigation, scenarios, tasks, and robot embodiments. | 高 | SP015, SP014 |
| CP022 | Figure says Helix is a generalist VLA model that unifies perception, language understanding, and learned control for humanoid upper-body action. | 中 | SP017 |
| CP023 | Figure says Helix runs onboard embedded low-power GPUs and is immediately ready for commercial deployment. | 中 | SP017 |
| CP024 | Figure announced more than $1 billion of Series C committed capital at a $39 billion post-money valuation to scale Helix and BotQ manufacturing. | 中 | SP018 |
| CP025 | Covariant’s RFM-1 is an 8 billion parameter multimodal robotics foundation model trained on text, images, video, robot actions, and physical measurements. | 高 | SP019, SP020 |
| CP026 | Covariant’s RFM-1 benefits from warehouse automation deployments, including a large fleet, tens of millions of trajectories, and customers across 15 countries. | 高 | SP020, SP019 |
| CP027 | Google DeepMind lists Gemini Robotics 1.5 as a private-preview VLA model that turns visual information and instructions into motor commands across multiple embodiments. | 中 | SP021 |
| CP028 | Google DeepMind’s Genie 2 is a foundation world model that generates action-controllable 3D environments for training and evaluating embodied agents. | 中 | SP022 |
| CP029 | NVIDIA Cosmos provides open world foundation models, tokenizers, guardrails, and data pipelines for physical-AI development in robots and autonomous vehicles. | 高 | SP023, SP024 |
| CP030 | NVIDIA named 1X, Agility, Figure, Skild AI, Waabi, XPENG, Uber, and other robotics or AV companies among initial Cosmos adopters. | 中 | SP023 |
| CP031 | Wayve raised $1.05 billion in Series C funding to develop embodied-AI products for automated driving and production vehicles. | 中 | SP026 |
| CP032 | Wayve positions its hardware-agnostic mapless foundation models for OEMs and fleet owners moving from assisted driving toward automated driving. | 中 | SP026 |
| CP033 | Hugging Face LeRobot provides open models, datasets, policies, and tools for real-world robotics, lowering barriers to internal prototyping. | 中 | SP027, SP028 |
| CP034 | Tesla’s Q1 2026 update says Optimus production lines are being installed and first large-scale Optimus factory preparations were to begin shortly in Q2. | 中 | SP031 |
| CP035 | Bain cautions that humanoid robot deployments are mostly early-stage and limited to structured environments with substantial human supervision. | 中 | SP029 |
| CP036 | Bain identifies dexterity, handling, battery life, safety, certification, workforce acceptance, and public trust as gating factors for humanoid deployment. | 中 | SP029 |
| CP037 | Sequoia argues AI infrastructure economics face a large revenue gap and that GPU computing can commoditize as prices are competed down. | 中 | SP030 |
| CP038 | LeCun told MIT Technology Review that useful general robots still require major conceptual breakthroughs and will not arrive in the next year or two. | 中 | SP005 |
| CP039 | AMI is less commercially ready than World Labs, Figure, Covariant, and Wayve because those peers disclose available products, deployment claims, production plans, or customer environments while AMI discloses no public product. | 中 | SP001, SP006, SP017, SP020, SP026 |
| CP040 | Public pricing evidence is asymmetric: Marble discloses subscription tiers, LeRobot is open tooling, and AMI, Physical Intelligence, Skild AI, Figure Helix, Covariant RFM-1, Gemini Robotics, and most Cosmos enterprise terms remain undisclosed in reviewed sources. | 中 | SP001, SP008, SP010, SP015, SP017, SP020, SP021, SP023, SP027 |
| CP041 | The competitive landscape separates into direct world-model peers, robot-policy peers, infrastructure enablers, vertical embodied-AI companies, open-source substitutes, internal-build paths, and likely incumbent entrants. | 中 | SP006, SP010, SP017, SP023, SP026, SP027, SP031 |
| CP042 | World Labs has Autodesk as a strategic design-channel partner, NVIDIA has a broad Cosmos adopter list, Wayve has OEM and cloud partners, and AMI’s only named application partner in reviewed sources is Nabla. | 中 | SP004, SP009, SP023, SP026 |
| CP043 | Covariant, Figure, Physical Intelligence, Wayve, and Tesla disclose or imply richer real-world robot or vehicle data loops than AMI has publicly documented. | 中 | SP011, SP017, SP020, SP026, SP031 |
| CP044 | Open and semi-open tooling from NVIDIA Cosmos, LeRobot, and Physical Intelligence can commoditize some model-layer capabilities by enabling buyers to build prototypes without AMI. | 中 | SP010, SP023, SP024, SP027, SP028, SP030 |
| CP045 | Direct robotics competitors create more immediate buyer substitution risk for AMI than horizontal LLM vendors because they already frame APIs, policies, datasets, or robots around physical action. | 中 | SP010, SP015, SP017, SP020, SP021 |
| CP046 | AMI’s most visible moat today is a combination of scientific leadership, capital, global hiring, and strategic investors rather than disclosed proprietary datasets or deployed customer outcomes. | 中 | SP001, SP002, SP003, SP005 |
| CP047 | Internal build is a credible substitute because developers can combine open robotics tooling, Cosmos synthetic data, and proprietary operational data before purchasing a specialist AMI model. | 中 | SP023, SP024, SP027, SP028 |
| CP048 | Likely entrants include NVIDIA, Google DeepMind, Tesla, large robotics OEMs, AV platforms, and CAD/simulation incumbents because they control compute, datasets, workflows, or distribution surfaces adjacent to world models. | 中 | SP009, SP021, SP022, SP023, SP026, SP031 |
| CP049 | The main adverse competitive evidence is that physical-AI buyers may delay broad deployment while model infrastructure simultaneously becomes cheaper and less differentiated. | 中 | SP029, SP030, SP005 |
| CP050 | AMI needs partner data rights, benchmark wins, safety evidence, and paid pilot terms to convert its research thesis into a durable competitive moat. | 中 | SP004, SP005, SP029, SP030 |
| CI001 | AMI’s public product thesis is to build world models for real-world sensor data, planning under safety guardrails, and applications in industrial process control, automation, wearables, robotics, healthcare, and beyond. | 高 | SI001, SI003 |
| CI002 | AMI says it raised $1.03B, approximately €890M, in a March 2026 round. | 高 | SI002, SI003 |
| CI003 | Cathay Innovation and TechCrunch report AMI’s $1.03B round at a $3.5B pre-money valuation. | 高 | SI003, SI004 |
| CI004 | Investor materials describe the round as supporting long-term research, global hiring, and development of reliable intelligent systems rather than near-term commercial scaling. | 高 | SI003, SI004 |
| CI005 | TechCrunch quotes Alexandre LeBrun saying AMI is not a typical applied-AI startup that can release a product in three months, have revenue in six months, and reach $10M ARR in 12 months. | 中 | SI004 |
| CI006 | AMI’s first named partner is Nabla, the healthcare AI company connected to AMI CEO Alexandre LeBrun. | 高 | SI004, SI007, SI008 |
| CI007 | Nabla says its exclusive strategic partnership gives it first access to AMI’s emerging world-model technologies for healthcare. | 中 | SI007, SI008 |
| CI008 | AMI’s reviewed official homepage and updates page disclose mission and funding but do not disclose public list pricing, a commercial SKU, or a self-serve API price. | 中 | SI001, SI002 |
| CI009 | Reviewed official, investor, and independent sources did not disclose AMI revenue, ARR, paying-customer count, gross margin, or burn rate as of the run date. | 中 | SI001, SI002, SI003, SI004, SI006 |
| CI010 | Pappers describes AMI’s corporate purpose around software, digital programs, AI model development, and AI-model training, but the reviewed registry page does not provide underwritable operating metrics for AMI. | 中 | SI010 |
| CI011 | Pappers and Annuaire des Entreprises provide registry anchors for ADVANCED MACHINE INTELLIGENCE but not private financial statements, audited accounts, or investor-rights terms. | 高 | SI010, SI011 |
| CI012 | France 24 and RFI report that LeCun said AMI would focus on research and development in its first year. | 高 | SI015, SI016 |
| CI013 | France 24 and RFI report that corporate-partner discussions could happen within six to twelve months. | 高 | SI015, SI016 |
| CI014 | The Nabla materials and STAT coverage evidence privileged access to AMI technology but do not disclose AMI-Nabla licensing fees, revenue-share economics, minimum commitments, or transfer pricing. | 中 | SI007, SI008, SI009 |
| CI015 | AMI’s supportable revenue-model hypotheses are future model licensing, API access, vertical partner deployments, co-development, and healthcare commercialization through partners rather than currently disclosed recurring revenue. | 中 | SI001, SI007, SI008, SI031 |
| CI016 | AMI’s public go-to-market evidence is partner-led and research-led, not a self-serve SaaS funnel with disclosed CAC, sales cycle, conversion, or payback. | 中 | SI001, SI004, SI007, SI013 |
| CI017 | Epoch AI estimates that final-run frontier-model training costs have grown at roughly 2.4x per year since 2016, making compute a central cost driver for frontier labs. | 高 | SI018, SI019 |
| CI018 | Epoch AI says the largest training runs would cost more than $1B by 2027 if historical training-cost growth continues. | 中 | SI018 |
| CI019 | NVIDIA says physical-AI model development requires vast real-world data and testing, including petabytes of video data and tens of thousands of compute hours for processing, curation, and labeling. | 中 | SI021 |
| CI020 | NVIDIA says its Blackwell platform can reduce LLM inference operating cost and energy by up to 25x versus its predecessor for relevant workloads. | 中 | SI020 |
| CI021 | CloudZero’s 2026 H100 cost guide reports new H100 purchase prices of $25,000-$40,000 and on-demand rental rates of $1.38-$8.00+ per GPU-hour. | 中 | SI029 |
| CI022 | Lambda and Modal show cloud GPU and container-concurrency access is commercially available, implying AMI can rent compute but must still manage variable usage costs and capacity planning. | 中 | SI026, SI027 |
| CI023 | FDA and NIST materials show healthcare and critical-infrastructure AI require explicit regulatory and risk-management workstreams that can add validation cost before production deployment. | 高 | SI023, SI024 |
| CI024 | Nabla frames the AMI collaboration around FDA-certifiable, safe, auditable agentic healthcare systems, which implies validation is part of the product path rather than a post-sale afterthought. | 中 | SI007, SI008 |
| CI025 | The $1.03B seed materially improves capital adequacy relative to most seed-stage labs, but public sources do not disclose cash on hand, monthly burn, runway, or committed compute obligations. | 中 | SI002, SI003, SI004, SI018, SI021 |
| CI026 | Because AMI has no disclosed revenue metrics and is pursuing compute-heavy frontier R&D, financing dependency remains high until partner pilots convert into contracted economics. | 中 | SI003, SI004, SI017, SI018, SI021 |
| CI027 | Sequoia’s AI-infrastructure critique asks “Where is all the revenue?” and highlights a gap between AI infrastructure build-out and actual AI ecosystem revenue growth. | 中 | SI017 |
| CI028 | Sequoia argues that GPUs are only about half of AI data-center total cost of ownership, with energy, buildings, backup generators, and other infrastructure forming the other half. | 中 | SI017 |
| CI029 | Le Monde Informatique characterized AMI as attracting investors while not yet having an operational system, making it an adverse commercialization source for this financial chapter. | 中 | SI014 |
| CI030 | Silicon Republic noted that AMI reached a $3.5B valuation despite having only been established in 2026, reinforcing stage-versus-valuation risk. | 中 | SI030 |
| CI031 | Crunchbase News describes AMI’s financing as Europe’s largest seed round. | 高 | SI004, SI006 |
| CI032 | Independent coverage reports strategic investor participation from large groups including Toyota, NVIDIA, and Samsung. | 高 | SI015, SI016, SI003 |
| CI033 | AMI’s official target domains include industrial process control, automation, wearable devices, robotics, healthcare, and beyond. | 高 | SI001, SI003 |
| CI034 | Nabla describes healthcare as operationally complex and high-consequence, with world models aimed at safer, more auditable workflows beyond documentation. | 中 | SI007, SI008 |
| CI035 | AMI says it wants to build with industry partners, product developers, the academic research community, open publications, and open source, which widens distribution options but does not define who pays. | 中 | SI001 |
| CI036 | No reviewed official, filing, or investor source disclosed debt, project-finance obligations, GPU purchase commitments, or credit facilities for AMI. | 中 | SI002, SI003, SI010, SI011 |
| CI037 | Global hiring is a disclosed use of funds, but public sources reviewed for this chapter do not disclose filled headcount, payroll run-rate, or compensation mix. | 中 | SI003, SI004 |
| CI038 | A first-year R&D focus plus six-to-twelve-month corporate-partner discussion window means AMI has no public near-term sales-efficiency proxy beyond partner access and investor confidence. | 中 | SI004, SI013, SI015, SI016 |
| CI039 | Current revenue quality is not underwritable because the evidence supports strategic access and research intent, not repeatable revenue, recognized bookings, renewals, or customer concentration. | 中 | SI004, SI007, SI008, SI009, SI014 |
| CI040 | AMI’s eventual gross margin will depend on model size, utilization, GPU pricing, inference efficiency, safety-validation labor, partner data rights, and pricing power. | 中 | SI017, SI018, SI020, SI021, SI023, SI029 |
| CI041 | Blackwell efficiency and falling H100 cost can improve compute economics, but those inputs do not prove AMI margin because AMI has not disclosed workload mix, pricing, utilization, or customer contracts. | 中 | SI020, SI021, SI029 |
| CI042 | If Nabla becomes a monetization path, AMI’s economics could take the form of licensing, royalty, transfer-price, or embedded partner revenue, but none of those terms are public. | 中 | SI007, SI008, SI009, SI031 |
| CI043 | Physical-AI cost is not only model training: data collection, video processing, scenario curation, labeling, simulation, and evaluation also drive cost of goods and R&D burn. | 中 | SI001, SI021, SI024 |
| CI044 | The registry evidence supports AMI’s legal existence and broad AI/software purpose but does not solve the core financial gaps of ARR, burn, runway, margin, or customer economics. | 中 | SI010, SI011 |
| CI045 | The disclosed use-of-funds profile points to a long-horizon research-and-compute budget, not a conventional seed-stage plan to scale a validated sales motion. | 中 | SI003, SI004, SI015, SI018 |
| CI046 | No reviewed source disclosed a paid AMI customer separate from Nabla’s strategic access relationship. | 中 | SI001, SI002, SI004, SI007, SI008, SI009 |
| CI047 | Adding the $1.03B round to the reported $3.5B pre-money valuation implies an approximate $4.53B post-money valuation before fees, option-pool effects, or undisclosed structure. | 中 | SI002, SI003, SI004 |
| CI048 | A revenue forecast, burn forecast, runway calculation, or gross-margin estimate would be speculative without AMI’s cash balance, monthly burn, pricing, usage, and contracted revenue. | 中 | SI009, SI017, SI018, SI021, SI029 |
| CI049 | The financial underwriting blocker is not lack of funding; it is the absence of disclosed revenue, ARR, pricing, customer conversion, burn, runway, gross margin, and partner-contract economics. | 中 | SI003, SI004, SI007, SI017, SI018, SI029 |
| CI050 | Sequoia’s revenue-gap critique and NVIDIA’s physical-AI cost context imply AMI must prove workflow lock-in and pricing power before compute-heavy models can become attractive unit economics. | 中 | SI017, SI021, SI029 |
| CE001 | AMI says it is developing world models that learn abstract representations of real-world sensor data, make predictions in representation space, and support action-conditioned planning under safety guardrails. | 中 | SE001, SE002 |
| CE002 | AMI names industrial process control, automation, wearable devices, robotics, healthcare, and other safety-sensitive domains as target application areas. | 中 | SE001 |
| CE003 | Reviewed AMI official pages and contemporaneous coverage do not disclose a shipped SKU, public API, price list, uptime SLA, or AMI-specific product benchmark. | 中 | SE001, SE002, SE005 |
| CE004 | TechCrunch quoted AMI CEO Alexandre LeBrun saying AMI starts with fundamental research and is not the kind of startup that can release a product in three months or reach revenue in six months. | 中 | SE005 |
| CE005 | Nabla has first access to AMI emerging world-model technologies for healthcare, but the public partnership materials do not disclose a standalone AMI product SKU or economics. | 中 | SE003, SE004 |
| CE006 | Nabla frames AMI world models as a path toward safe, auditable agentic healthcare systems that complement today’s LLM-based clinical assistants. | 中 | SE003, SE004 |
| CE007 | MIT Technology Review characterizes LeCun’s AMI venture as a contrarian bet against large language models and in favor of world models that reflect real-world dynamics. | 中 | SE006 |
| CE008 | LeCun told MIT Technology Review that broadly useful domestic robots require good world models and planning. | 中 | SE006 |
| CE009 | AMI and its CEO publicly state an intent to work with industry partners, product developers, academia, open publications, and substantial open-source code. | 中 | SE001, SE005 |
| CE010 | LeCun’s 2022 architecture paper frames autonomous machine intelligence around predictive world models, hierarchical representations, planning, and action modules rather than next-token generation. | 中 | SE007 |
| CE011 | The latent-variable energy-based-model notes explain H-JEPA as a building block for reliable world models that can reason and plan complex action sequences. | 中 | SE008 |
| CE012 | I-JEPA is a non-generative self-supervised image approach that predicts target-block representations from context blocks. | 中 | SE009 |
| CE013 | V-JEPA extends feature prediction to video and trains without text, negative examples, reconstruction, or other sources of supervision. | 中 | SE010 |
| CE014 | Meta’s V-JEPA repository provides official PyTorch code and models for video joint-embedding predictive architecture research. | 中 | SE011 |
| CE015 | V-JEPA 2 is presented by Meta and arXiv as a video-trained world model for visual understanding, prediction, planning, and zero-shot robot control. | 中 | SE012, SE013 |
| CE016 | V-JEPA 2 reporting says it uses internet-scale video and images plus less than 62 hours of robot data for action-conditioned planning. | 中 | SE012, SE013 |
| CE017 | The V-JEPA 2 code repository and Hugging Face collection provide developer-visible model and checkpoint artifacts outside AMI itself. | 中 | SE014, SE015 |
| CE018 | The JEPA lineage differs from token-prediction LLMs by predicting in embedding or representation space and by avoiding direct pixel or token reconstruction as the core objective. | 中 | SE001, SE009, SE010 |
| CE019 | Meta’s V-JEPA 2 description uses model-predictive control: encode current and goal states, imagine candidate actions, score them, and re-plan the next action. | 中 | SE013 |
| CE020 | LeRobot’s public repository shows a practitioner ecosystem for real-world robotics models, datasets, policies, and world-model approaches such as VLA-JEPA. | 中 | SE022 |
| CE021 | NVIDIA states that physical-AI world-model development can require petabytes of video data and tens of thousands of compute hours, which makes data and compute pipeline assets central to AMI diligence. | 中 | SE020, SE021 |
| CE022 | Google DeepMind’s Genie 2 demonstrates a competitor world-model direction: generating action-controllable 3D environments for embodied-agent training and evaluation. | 中 | SE023 |
| CE023 | Google DeepMind’s Gemini Robotics 1.5 is described as a vision-language-action model that turns visual inputs and instructions into robot motor commands in private preview. | 中 | SE024 |
| CE024 | Because Meta, NVIDIA, Hugging Face, and DeepMind expose world-model or robotics tooling, AMI’s durable differentiation must come from proprietary partner data, evaluation, integration, or safety evidence rather than the JEPA label alone. | 中 | SE013, SE014, SE020, SE022, SE023, SE024 |
| CE025 | In healthcare workflow terms, the AMI-Nabla opportunity is to move from documentation and retrieval toward simulation, persistent memory, deterministic reasoning, and planned clinical actions, but this remains roadmap-level. | 中 | SE003, SE004, SE026 |
| CE026 | In industrial automation workflow terms, AMI’s public thesis maps sensor observations to predicted consequences and guarded control plans, but no industrial deployment or integration interface is disclosed. | 中 | SE001, SE020 |
| CE027 | In robotics workflow terms, Meta’s V-JEPA 2 evidence supports the plausibility of video-conditioned planning, but it is not evidence of an AMI robot product or AMI deployment. | 中 | SE013, SE014 |
| CE028 | In wearables and safety-critical settings, AMI’s named opportunity is multimodal state understanding and safe planning; the public evidence does not identify sensors, form factors, or regulated product requirements. | 中 | SE001 |
| CE029 | The public deployment path looks partner-led rather than self-serve because Nabla is the named first-access partner and AMI does not publish API documentation, onboarding docs, or support terms. | 中 | SE001, SE003, SE005 |
| CE030 | Data rights, event schemas, model endpoints, deployment topology, monitoring, support SLAs, and partner acceptance tests remain undisclosed integration dependencies. | 中 | SE001, SE002, SE003 |
| CE031 | FDA explains that AI software may be subject to medical-device review, clearance, approval, or modification review depending on risk. | 中 | SE016 |
| CE032 | FDA PCCP guidance recommends documenting planned AI-device modifications, validation and implementation methodology, and impact assessment. | 中 | SE017 |
| CE033 | NIST’s AI RMF materials emphasize risk-management practices for trustworthy AI, including critical-infrastructure profiles. | 中 | SE018 |
| CE034 | The EU AI Act sets a framework for trustworthy AI while protecting health, safety, and fundamental rights. | 中 | SE019 |
| CE035 | Nabla’s FDA-certifiable language is a roadmap claim because reviewed sources do not show FDA clearance, certification, or a submitted AMI-Nabla medical-device product. | 中 | SE003, SE016, SE017 |
| CE036 | AMI public pages do not disclose clinical data privacy controls, security certifications, deployment isolation, audit-log design, or data-retention commitments. | 中 | SE001, SE003 |
| CE037 | AMI public pages do not disclose uptime, incident history, red-team results, model cards, safety-case artifacts, post-market monitoring, or support SLAs for a productized system. | 中 | SE001, SE002, SE003 |
| CE038 | Before autonomous healthcare or industrial action, AMI would need human oversight, auditability, validation, monitoring, change-control, and risk-management controls aligned with FDA, NIST, and EU AI Act expectations. | 中 | SE003, SE016, SE017, SE018, SE019 |
| CE039 | AMI’s March 2026 update says it raised a $1.03B round and is building a distributed team across Paris, New York, Montreal, and Singapore. | 中 | SE002, SE005 |
| CE040 | Nabla’s December 2025 announcement says it will invest over coming months in multimodality, simulation, deterministic reasoning, and product quality while it expands its assistant. | 中 | SE003 |
| CE041 | TechCrunch reports that AMI’s world models could take years to move from theory to commercial applications. | 中 | SE005 |
| CE042 | No public AMI product launch date, release notes, generally available terms, customer support path, or enterprise documentation surfaced in reviewed AMI pages. | 中 | SE001, SE002 |
| CE043 | Open publications and code could help AMI recruit and seed an ecosystem, but they also reduce architecture-only moat if open platforms can replicate comparable workflows. | 中 | SE005, SE011, SE014, SE022 |
| CE044 | MIT Technology Review reports that humanoid demos do not overcome power, battery, manufacturing, and safety constraints, which limits near-term robot workflow assumptions. | 中 | SE027 |
| CE045 | Bain says humanoid deployment expansion depends on certification, safety, human acceptance, and battery or charging constraints. | 中 | SE028 |
| CE046 | Meta says human performance remains meaningfully above top models including V-JEPA 2 on several new physical-reasoning benchmarks, so world-model capability remains incomplete. | 中 | SE013 |
| CE047 | The fetched source set does not show AMI-specific benchmarks, IP filings, released model weights, or released products; those items should remain diligence gaps rather than assumed assets. | 中 | SE001, SE002, SE005, SE006 |
| CE048 | AMI’s strongest technical differentiation versus token-prediction LLMs is the world-model premise of representation-space prediction and action-conditioned planning, but public evidence does not yet prove commercial deployment. | 中 | SE001, SE003, SE004, SE009, SE013 |
| CE049 | Safety-critical deployments should start with human-in-the-loop, constrained-scope pilots until AMI can show validation evidence, regulatory path, monitoring, and incident response. | 中 | SE016, SE017, SE018, SE019, SE028 |
| CE050 | AMI’s current product should be underwritten as a research platform and partner-integration roadmap for world models, not as a shipped commercial SKU. | 中 | SE001, SE003, SE005, SE006 |
| CU001 | AMI’s own site positions the company as a frontier research lab building world models for safety-critical applications rather than a customer-deployment business. | 高 | SU001, SU008 |
| CU002 | AMI names healthcare, industrial process control, automation, wearables, and robotics as target application domains, but the official site does not name paying customers in those domains. | 中 | SU001, SU002 |
| CU003 | Nabla announced an exclusive strategic partnership with AMI in December 2025. | 高 | SU003, SU004 |
| CU004 | Nabla is the only named party with first access to AMI’s emerging world-model technology in the reviewed public evidence. | 高 | SU003, SU004, SU005 |
| CU005 | Nabla frames AMI access as a route toward FDA-certifiable, auditable, agentic healthcare AI, not as proof that AMI already sells a deployed product. | 高 | SU003, SU004, SU024 |
| CU006 | Alex LeBrun’s move to AMI CEO while retaining Nabla chair/chief-scientist roles creates a deep strategic link between the companies. | 高 | SU003, SU006 |
| CU007 | TechCrunch reported that Nabla is AMI’s first disclosed partner for early models. | 中 | SU005 |
| CU008 | AMI’s CEO told TechCrunch the company is not an applied AI startup expected to ship a product in three months, generate revenue in six months, or reach $10 million ARR in twelve months. | 中 | SU005 |
| CU009 | TechCrunch reported that AMI had no plans to generate revenue for the time being while still planning early engagement with prospective customers. | 中 | SU005 |
| CU010 | STAT reported there was no formal equity or licensing agreement relationship yet between AMI and Nabla, despite close collaboration. | 中 | SU006 |
| CU011 | HIT Consultant and HLTH corroborate that Nabla’s first-access role is the key public commercialization bridge for AMI technology. | 中 | SU007, SU024 |
| CU012 | Across the reviewed sources, no public evidence names a paying AMI customer, AMI customer count, AMI NRR, AMI pilot count, or AMI contract term. | 中 | SU001, SU002, SU003, SU005, SU006, SU023 |
| CU013 | The likely direct AMI buyer is an enterprise or regulated operator needing reliable world models, while the initial end users are most visible through Nabla’s clinician workflows. | 中 | SU001, SU003, SU004, SU005 |
| CU014 | AMI’s likely payer for the healthcare wedge would be a partner or health-system channel buyer rather than an individual clinician, but public sources do not disclose pricing or contract structure. | 中 | SU003, SU004, SU025 |
| CU015 | Nabla public materials cite a large but inconsistently stated installed base, ranging from 130-plus healthcare organizations to over 150 health systems and provider groups. | 中 | SU003, SU025, SU026, SU028 |
| CU016 | Nabla’s 2025 Series C materials report 85,000 clinicians, 20 million annual encounters, and more than 130 healthcare organizations. | 高 | SU025, SU026, SU027, SU028 |
| CU017 | Nabla states that its assistant integrates with major EHRs and supports more than 35 languages, which matters for healthcare deployment friction. | 高 | SU003, SU025, SU028 |
| CU018 | Nabla’s case-study hub lists multiple health-system customer stories, but those stories validate Nabla deployment rather than AMI deployment. | 中 | SU009 |
| CU019 | Denver Health began with an eight-week Nabla pilot across 12 specialties, 50 clinicians, and more than 6,000 visits before broader system adoption. | 中 | SU010 |
| CU020 | Denver Health reported 40% lower post-visit documentation time, 30% lower burnout scores, more than 300,000 encounters captured, and 400 clinicians adopting within one week after pilot success. | 中 | SU010 |
| CU021 | Carle Health’s Nabla case study reports 1,500 providers, 55% of clinicians saving at least one documentation hour, 89% willingness to recommend, and 78% likelihood of replacing the prior documentation method. | 中 | SU011 |
| CU022 | Fierce Healthcare independently reported Carle’s rollout and said the partnership expanded Nabla’s reach by 1,500 providers. | 中 | SU017 |
| CU023 | Healthcare IT Today reported Carle chose Nabla after testing many ambient voice products and emphasized workflow fit and EHR integration. | 中 | SU018 |
| CU024 | McFarland Clinic’s Nabla case study reports 10,000 average monthly encounters, 100-plus providers, and an 80% retention rate following the pilot. | 中 | SU012 |
| CU025 | Tia Health’s Nabla case study reports a 50% reduction in clinical note submission time, more than 50,000 clinical notes generated, and more than 90 providers. | 中 | SU013 |
| CU026 | CHLA’s Nabla case study reports a 50% drop in documentation time, 47% decrease in physician burnout, and 89% same-day note completion after deployment. | 中 | SU014 |
| CU027 | UToledo Health sources report deployment to hundreds of clinicians after an eight-week evaluation in which Nabla users reduced time to chart closure by 29%. | 中 | SU015, SU016 |
| CU028 | UToledo Health sources report documentation backlogs in several departments dropped from more than 400 open charts to fewer than 30 during evaluation. | 中 | SU015, SU016 |
| CU029 | Aultman Health System implemented Nabla within Oracle Cerner for hundreds of clinicians, completing native integration in less than 60 days. | 中 | SU019 |
| CU030 | Aultman reported clinicians saved 30 to 60 minutes per day, documentation time per patient often fell 20% to 40%, and some clinicians could see three to five more patients per day. | 中 | SU019 |
| CU031 | Nabla’s customer proof shows a credible healthcare channel for eventual AMI technology, but it is indirect because the case studies concern Nabla’s ambient assistant before public AMI product deployment. | 中 | SU009, SU010, SU011, SU012, SU013, SU014, SU024 |
| CU032 | AMI’s customer concentration risk is currently extreme because Nabla is the only named strategic partner with first access to the technology. | 中 | SU003, SU004, SU005 |
| CU033 | AMI’s expansion path through Nabla is plausible because Nabla is already extending from documentation into coding, agentic EHR commands, inpatient, nursing, and multimodal workflows. | 中 | SU004, SU024, SU025, SU028 |
| CU034 | Procurement friction for AMI’s healthcare wedge will center on safety, auditability, EHR integration, privacy, and regulatory posture rather than model quality alone. | 中 | SU003, SU010, SU017, SU019, SU025 |
| CU035 | Public sources do not disclose AMI direct retention, gross retention, net retention, renewal rate, contract length, pilot-to-production conversion, or customer satisfaction metrics. | 中 | SU005, SU006, SU012 |
| CU036 | Sacra’s adverse analysis argues that AMI could be too slow to prove production advantage while deploy-focused rivals build data loops, customer trust, and shipping products. | 中 | SU022 |
| CU037 | Forbes framed the key customer-readiness question for AMI as whether world models can move beyond hype into real-world benefits. | 中 | SU021 |
| CU038 | Sequoia’s broader AI analysis underscores buyer and investor scrutiny over whether AI infrastructure spending converts into durable revenue and ROI. | 中 | SU020 |
| CU039 | AMI’s strategic and industrial backers may become customer-development channels, but public investor lists are not customer contracts. | 中 | SU005, SU008 |
| CU040 | The French Tech Journal tracker highlights AMI funding, hiring, product-brand, and team milestones but not named customer deployments. | 中 | SU023 |
| CU041 | The Healthcare Technology Report independently repeats Nabla’s 85,000-clinician and 130-plus-organization scale claims from its Series C announcement. | 中 | SU027 |
| CU042 | Highland Europe repeats Nabla’s 85,000-clinician, 20-million-encounter, and 130-plus-health-system disclosures while adding investor corroboration. | 中 | SU028 |
| CU043 | Fierce Healthcare reported that Nabla had contracts with The Permanente Medical Group, CHLA, Stratum Med, and Mankato Clinic before the later Series C scale-up. | 中 | SU017 |
| CR001 | AMI raised $1.03 billion at a $3.5 billion pre-money valuation in March 2026. | 高 | SR003, SR020, SR021 |
| CR002 | TechCrunch reported in December 2025 that AMI was associated with a prospective $5 billion valuation discussion before the March 2026 round. | 中 | SR004 |
| CR003 | AMI is building world models intended to learn from reality rather than language alone. | 高 | SR001, SR003, SR005 |
| CR004 | AMI’s public materials and investor coverage frame the company as a long-term research and hiring program, not a near-term packaged SaaS product. | 中 | SR002, SR003, SR020 |
| CR005 | TechCrunch reported that AMI investments may take a while to turn into commercial applications. | 中 | SR003 |
| CR006 | LeBrun told TechCrunch that world models may become the next fundraising buzzword, creating adverse hype and category-noise risk. | 高 | SR003, SR021 |
| CR007 | MIT Technology Review described LeCun’s AMI thesis as a contrarian bet against the prevailing large-language-model direction. | 中 | SR005 |
| CR008 | Forbes framed AMI around the question of whether world models can move beyond hype. | 中 | SR006 |
| CR009 | Nabla is AMI’s first named partner and has exclusive first access to AMI’s emerging world-model technology. | 高 | SR003, SR016, SR017 |
| CR010 | Nabla says the AMI partnership is intended to help bring safe, auditable agentic AI systems into healthcare. | 中 | SR016, SR017 |
| CR011 | Nabla’s AMI announcement explicitly positions the partnership around FDA-certifiable agentic AI systems for healthcare. | 高 | SR016, SR044 |
| CR012 | FDA maintains an inventory of AI-enabled medical devices, indicating that healthcare AI deployment already sits inside active device-regulatory workflows. | 高 | SR013, SR044 |
| CR013 | The EU AI Act subjects high-risk AI systems to strict obligations before they can be put on the market. | 高 | SR039, SR040 |
| CR014 | The EU AI Act assigns responsibility to providers placing high-risk AI systems on the market or putting them into service. | 高 | SR039, SR040 |
| CR015 | NIST’s AI RMF provides a governance frame for mapping, measuring, managing, and monitoring AI risks but does not itself prove AMI has implemented those controls. | 中 | SR011 |
| CR016 | UK frontier-AI summit materials emphasize safety testing and research as a recognized frontier-model obligation. | 高 | SR037, SR038 |
| CR017 | Epoch AI estimates frontier model training costs have grown around 2.4x annually and could exceed $1 billion for the largest runs by 2027 if trends continue. | 高 | SR008, SR009 |
| CR018 | Sequoia argued that the AI infrastructure revenue gap had expanded to a $600 billion annual question. | 中 | SR042 |
| CR019 | NVIDIA says physical-AI model development can require petabytes of video data and tens of thousands of compute hours. | 中 | SR027 |
| CR020 | AMI’s world-model roadmap therefore carries material compute-capacity and data-curation execution risk. | 中 | SR008, SR027, SR042 |
| CR021 | World Labs, DeepMind, and NVIDIA all publish world-model, robotics, or physical-AI initiatives that pressure AMI’s differentiation window. | 中 | SR024, SR025, SR026, SR027 |
| CR022 | World Labs’ Marble materials describe reconstructing, generating, and simulating 3D worlds, overlapping the spatial/world-model narrative AMI uses. | 中 | SR024, SR025 |
| CR023 | Google DeepMind’s Gemini Robotics and NVIDIA Cosmos show that incumbents can pair foundation models with robotics or simulation tooling before AMI ships. | 中 | SR026, SR027 |
| CR024 | LeCun is AMI’s executive chair and is publicly tied to Meta AI and NYU roles in the public record. | 高 | SR004, SR028, SR029 |
| CR025 | LeBrun is AMI’s CEO, was Nabla’s CEO/co-founder, and previously worked with LeCun at Meta/FAIR. | 高 | SR004, SR016, SR022 |
| CR026 | AMI’s public profile depends unusually heavily on LeCun’s research reputation and LeBrun’s Nabla/operator bridge. | 中 | SR003, SR004, SR005, SR016 |
| CR027 | CourtListener’s Kadrey v. Meta docket demonstrates active copyright litigation risk around AI training and Meta-related model-development practices. | 中 | SR045 |
| CR028 | AMI has no publicly disclosed product SKU, direct paying customer list, ARR, pilot count, or pricing in the reviewed source set. | 中 | SR001, SR002, SR003, SR016, SR017 |
| CR029 | AMI is registered as a French legal entity, but public corporate registries do not resolve its product, IP, or conflict-boundary questions. | 中 | SR030, SR031 |
| CR030 | LeCun’s open-source advocacy and AMI’s publication posture create a governance tension between transparency, IP protection, and regulated deployment. | 中 | SR003, SR005, SR045 |
| CR031 | AMI’s Meta, NYU, and Nabla overlap creates diligence questions about IP assignment, data rights, board approvals, conflict policies, and publication boundaries. | 中 | SR016, SR028, SR029, SR045 |
| CR032 | The top residual risk is pre-product commercialization: a large financing round precedes disclosed customer proof and product readiness. | 中 | SR001, SR003, SR020, SR042 |
| CR033 | Healthcare use cases intensify liability, auditability, and model-change-control diligence because AMI and Nabla explicitly point toward agentic clinical workflows. | 中 | SR016, SR017, SR013, SR044 |
| CR034 | Operational quality risk is material because the promised system must be reliable in dynamic, safety-critical environments rather than just plausible in demos. | 中 | SR011, SR016, SR027, SR038 |
| CR035 | Partner concentration is high because Nabla is the only named first-access partner and the only disclosed path to an early regulated vertical. | 中 | SR003, SR016, SR017 |
| CR036 | Financing risk is valuation-driven because AMI must convert a $3.5 billion pre-money round into proof that can justify future compute, hiring, and product capital. | 中 | SR001, SR003, SR020, SR042 |
| CR037 | A credible mitigation plan would require AMI-specific safety cases, evaluation results, data-rights documentation, regulatory pathway mapping, and named design partners. | 中 | SR011, SR013, SR016, SR038, SR044 |
| CR038 | The principal thesis-break trigger is failure to disclose a governed pilot, benchmark, or FDA/EU-ready control plan within the next financing window. | 中 | SR003, SR013, SR039, SR042, SR044 |
| CR039 | Competitor pressure can compress AMI’s time-to-proof because better-capitalized incumbents already own compute, model platforms, and customer distribution. | 中 | SR024, SR025, SR026, SR027, SR042 |
| CR040 | Security and privacy diligence remains open because no source discloses AMI-specific controls for clinical data, model logging, incident response, or access governance. | 低 | |
| CR041 | No retained source discloses AMI’s burn rate, runway, cloud commitments, or compute procurement terms. | 中 | SR002, SR003, SR020, SR042 |
| CR042 | No retained source discloses the AMI-Nabla commercial agreement economics, exclusivity duration, data rights, or termination rights. | 中 | SR016, SR017, SR018 |
| CR043 | No retained source discloses whether Meta, NYU, or Nabla have formal conflict waivers or IP boundary documents tied to AMI. | 中 | SR016, SR028, SR029, SR045 |
| CR044 | Residual exposure remains high for the highest-ranked risks because most mitigations are plans or external frameworks rather than AMI-specific public controls. | 中 | SR011, SR013, SR016, SR039, SR044 |
| CR045 | AMI’s adverse-source profile is meaningful because independent skeptics and market analysts question hype, compute economics, and revenue realization in the AI stack. | 中 | SR006, SR008, SR042 |
| CV001 | AMI announced a $1.03 billion seed round, approximately €890 million, to fund development of world models. | 高 | SV001, SV002, SV004 |
| CV002 | AMI’s round was reported at a $3.5 billion pre-money valuation, implying roughly $4.5 billion post-money before unknown structure. | 高 | SV002, SV004, SV012 |
| CV003 | Crunchbase and Observer described AMI’s financing as Europe’s largest seed round and a major world-model funding milestone. | 中 | SV011, SV012 |
| CV004 | TechCrunch reported that AMI had initially been seeking about €500 million at about a €3 billion valuation before closing the larger financing. | 中 | SV003, SV002 |
| CV005 | TechCrunch reported that AMI intends to prioritize compute and talent and may take years to reach commercial applications. | 中 | SV002 |
| CV006 | AMI’s disclosed backers include strategic or high-profile names such as NVIDIA, Samsung, Toyota Ventures, Temasek, Mark Cuban, and others. | 高 | SV001, SV002, SV004 |
| CV007 | Nabla receives first access to AMI world-model technologies, but STAT reported no formal equity or licensing relationship had yet been established. | 中 | SV005, SV013 |
| CV008 | Public sources reviewed for this chapter do not disclose AMI revenue, ARR, pricing, gross margin, paying customers, or signed commercial contracts. | 中 | SV002, SV005, SV006, SV013 |
| CV009 | AMI’s CEO told TechCrunch that world models could require years before theory becomes commercial application. | 中 | SV002 |
| CV010 | Cathay frames world models as systems that learn real-world representations, predict consequences, and plan actions under constraints. | 中 | SV004 |
| CV011 | TechCrunch quoted AMI’s CEO predicting that world models could become a fundraising buzzword within six months. | 中 | SV002, SV011 |
| CV012 | World Labs announced $1 billion in new funding from investors including AMD, Autodesk, Emerson Collective, Fidelity, NVIDIA, and Sea. | 中 | SV014, SV015 |
| CV013 | TechCrunch reported that World Labs emerged from stealth with $230 million at a $1 billion valuation and later sought funding around a $5 billion valuation. | 中 | SV015, SV016 |
| CV014 | CNBC reported that Physical Intelligence raised $400 million at a $2.4 billion post-money valuation in 2024. | 中 | SV018, SV017 |
| CV015 | TechCrunch reported that Physical Intelligence was discussing a roughly $1 billion raise at a valuation above $11 billion in 2026. | 中 | SV019, SV018 |
| CV016 | Mistral announced a €1.7 billion Series C at an €11.7 billion post-money valuation, with ASML investing €1.3 billion for about an 11% fully diluted stake. | 高 | SV020, SV021 |
| CV017 | Figure announced more than $1 billion of committed Series C capital at a $39 billion post-money valuation for humanoid robots. | 中 | SV036, SV031 |
| CV018 | Wayve announced a $1.05 billion Series C led by SoftBank to develop embodied-AI products for automated driving. | 中 | SV037, SV022 |
| CV019 | Skild AI announced a $300 million Series A that valued the company at $1.5 billion for a robotics foundation model strategy. | 中 | SV038, SV017 |
| CV020 | NVIDIA’s Cosmos platform targets world foundation models, video tokenizers, guardrails, and data pipelines for robots and autonomous vehicles. | 中 | SV022 |
| CV021 | Google DeepMind presents Gemini Robotics as adaptable across diverse robot forms, adding corporate-lab competition in physical AI. | 中 | SV035 |
| CV022 | CoreWeave’s S-1/A showed AI infrastructure companies can reach public-market financing, but through revenue-bearing infrastructure rather than pre-product research alone. | 中 | SV024, SV027 |
| CV023 | NVIDIA’s fiscal 2025 Form 10-K reported about $2.7 trillion of non-affiliate market value, underscoring public-market scale behind AI infrastructure comparables. | 中 | SV023 |
| CV024 | Sequoia argued that AI infrastructure economics had become a $600 billion annual revenue question with a roughly $500 billion remaining gap. | 中 | SV009 |
| CV025 | Reuters coverage quoted large investors warning that early-stage AI valuations were frothy and driven by hype, with PitchBook citing $73.1 billion of Q1 2025 AI startup funding. | 高 | SV010, SV029 |
| CV026 | Reuters quoted TPG’s Todd Sisitsky saying some early-stage AI ventures were valued at $400 million to $1.2 billion per employee, which he called breathtaking. | 中 | SV010, SV029 |
| CV027 | Epoch AI warned that the largest training runs could cost more than $1 billion by 2027 if scaling trends continue. | 中 | SV030 |
| CV028 | MarketsandMarkets projects embodied AI market growth from $4.44 billion in 2025 to $23.06 billion in 2030. | 中 | SV039 |
| CV029 | Bain says humanoid robot deployments remain early-stage and heavily reliant on human supervision despite billion-dollar valuations. | 中 | SV031 |
| CV030 | MIT Technology Review described humanoid robotics as a hype cycle in which slick demos can raise investor expectations before impact is verified. | 中 | SV032 |
| CV031 | The EU AI Act establishes harmonized rules intended to protect health, safety, and fundamental rights for AI systems in the Union. | 中 | SV033 |
| CV032 | FDA guidance pages state that many changes to AI/ML-driven medical device software may need premarket review. | 中 | SV034 |
| CV033 | AMI’s current public valuation is expensive because the approximately $4.5 billion post-money price precedes disclosed product revenue, ARR, pricing, or paying customers. | 中 | SV002, SV004, SV008, SV010 |
| CV034 | AMI has upside if world models become a new frontier stack, because investors, World Labs, Mistral, NVIDIA, and physical-AI peers show multi-billion-dollar willingness to fund the category. | 中 | SV004, SV014, SV020, SV022, SV036 |
| CV035 | The base case should require AMI to show a product milestone, AMI-specific benchmark, and Nabla pilot economics before a higher price is underwritten. | 中 | SV002, SV005, SV013, SV034 |
| CV036 | The bear case is that compute intensity, slow commercialization, and AI-valuation compression create flat-round or down-round risk before AMI reaches revenue proof. | 中 | SV009, SV010, SV027, SV030 |
| CV037 | The bull case requires AMI to convert its founder/investor quality into defensible product traction in a market that reaches meaningful embodied-AI spend. | 中 | SV004, SV015, SV019, SV028, SV039 |
| CV038 | The appropriate public-evidence recommendation is research-more or track, not buy at the current valuation, unless investors receive milestone protection or a lower effective entry. | 中 | SV002, SV008, SV009, SV010, SV029 |
| CV039 | The chapter’s confidence is medium-low because financing and strategic-partner evidence are public, while cap-table terms, product metrics, ARR, and customer proof remain private. | 中 | SV001, SV002, SV005, SV013 |
| CV040 | Diligence should request cap table, liquidation preferences, tranche terms, pro-rata rights, and option-pool/dilution details before underwriting the entry price. | 中 | SV001, SV002, SV004 |
| CV041 | Diligence should request AMI’s compute budget, training roadmap, utilization plan, supplier commitments, and sensitivity to GPU price declines or shortages. | 中 | SV002, SV023, SV030 |
| CV042 | Diligence should request Nabla contract economics, data rights, safety responsibilities, and FDA pathway ownership for healthcare deployments. | 中 | SV005, SV013, SV034 |
| CV043 | Diligence should require evidence of AMI-specific product milestones, benchmark results, and customer pipeline beyond the first-access Nabla relationship. | 中 | SV002, SV005, SV016, SV017 |
| CV044 | A thesis-break trigger is any new financing above the current implied post-money valuation without public product, revenue, or named-customer proof. | 中 | SV002, SV010, SV029 |
| CV045 | A thesis-break trigger is evidence that open or corporate world-model platforms commoditize AMI’s layer before it controls proprietary data or distribution. | 中 | SV022, SV035, SV009 |
| CV046 | Exit readiness is low today because likely IPO or large M&A paths require product, revenue, regulatory, or infrastructure-scale proof not disclosed for AMI. | 中 | SV002, SV021, SV024, SV027 |
| CV047 | Entry discipline should emphasize milestone-based tranches, pro-rata rights, governance information rights, or a lower effective price instead of a blind seed-stage markup. | 中 | SV002, SV009, SV010, SV029 |
| CV048 | Preference and dilution overhang remain unknown because public round disclosures do not publish liquidation preferences, tranche conditions, option-pool treatment, or investor rights. | 低 | SV001, SV002, SV004 |
| CV049 | Public evidence supports AMI’s quality as a founder-led frontier lab, but not the full valuation because economics, product readiness, and customer conversion remain undisclosed. | 中 | SV004, SV006, SV007, SV002 |
| CV050 | The anti-thesis includes the risk that world models become a fundraising label before AMI proves differentiated product performance. | 中 | SV002, SV008, SV011 |
| CV051 | Dataconomy characterized AMI’s financing as one of the largest pre-revenue AI raises in history and a premium on elite researchers despite the company’s young age. | 中 | SV040, SV002 |
| 编号 | 出版方 | 标题 | 引文 |
|---|---|---|---|
| SO001 | AMI | AMI homepage | AMI is developing world models that learn abstract representations of real-world sensor data. |
| SO002 | AMI | AMI raises $1.03B to build world models | We’ve raised a $1.03B USD (~€890M) round from global investors. |
| SO003 | TechCrunch | Yann LeCun’s AMI Labs raises $1.03 billion to build world models | AMI Labs has no plans to generate revenue for the time being. |
| SO004 | Cathay Innovation | Advanced Machine Intelligence (AMI) is enabling the next AI revolution built on foundational world models | The company has raised $1.03B USD (~€890M) – based on a $3.50B (~€3B) pre-money valuation. |
| SO005 | Pappers | Informations juridiques de ADVANCED MACHINE INTELLIGENCE | SIREN : 994 675 254 |
| SO006 | Annuaire des Entreprises | Advanced Machine Intelligence - 994675254 | |
| SO007 | MIT Technology Review | Yann LeCun’s new venture is a contrarian bet against large language models | I am going to be the executive chairman of the company, and Alex LeBrun will be the CEO. |
| SO008 | TechCrunch | Who’s behind AMI Labs, Yann LeCun’s world model startup? | Nabla’s board supported LeBrun’s shift from CEO to chief AI scientist and chairman, clearing the way for his new role. |
| SO009 | Nabla | Nabla announces exclusive partnership with Advanced Machine Intelligence to pioneer the next era of agentic healthcare AI | Through this partnership, Nabla will gain first access to Advanced Machine Intelligence’s emerging world model technologies. |
| SO010 | Nabla | AMI raises $1.03B to build world models — powering the next generation of healthcare AI with Nabla | AMI’s research focuses on developing world models that can maintain persistent memory, reason about evolving situations, and plan actions under real-world constraints. |
| SO011 | France 24 | French AI startup AMI raises $1B to develop universal intelligent systems | Based in Paris with offices in New York, Singapore and Montreal, AMI was valued at around $3.5 billion before this funding round. |
| SO012 | RFI | French startup raises $1 billion to shift AI research into high gear | LeCun told French news agency AFP that, with the funding round complete, AMI would bring aboard 20-30 people very shortly. |
| SO013 | Euronews | AMI, une startup française de l’IA, annonce une levée de fonds d’un milliard de dollars | Ce premier tour de table ... valorise l'entreprise française à 3,5 milliards de dollars. |
| SO014 | Le Monde Informatique | La start-up AMI Labs de Yann LeCun lève 890 millions d’euros | Elle a séduit des investisseurs internationaux et français sans pour l'instant avoir de systèmes opérationnels. |
| SO015 | EU-Startups | Beyond LLMs: AI pioneer Yann LeCun’s new venture AMI raises €890 million to build world model AI systems | This is among the largest Seed rounds ever raised by a European company. |
| SO016 | Crunchbase News | World Model AI Lab AMI Raises Europe’s Largest Seed Round | The funding for Paris-based AMI represents the largest seed round ever for a European startup. |
| SO017 | Observer | Yann LeCun’s AMI Startup Funding Round | The company is also hiring across offices in Paris, New York, Montreal and Singapore. |
| SO018 | The Next Web | Yann LeCun just raised $1bn to prove the AI industry has got it wrong | AMI has no product, no revenue, and no near-term prospect of either. |
| SO019 | Forbes | Yann LeCun’s New Startup AMI Labs: Can World Models Move Beyond Hype? | Training models on video is expensive. Running them for long, complex tasks pushes costs up fast. |
| SO020 | STAT | World model developer tied to Nabla gets $1 billion | There’s no formal equity or licensing agreement relationship yet between AMI and Nabla. |
| SO021 | Offcall | AI world models in medicine with Yann LeCun and Alex LeBrun | AI Legends Yann LeCun and Alex LeBrun Debut AMI Labs' Bold Ambitions for World Models in Healthcare |
| SO022 | Ashby | AMI Jobs | AMI Jobs |
| SO023 | Sequoia Capital | AI’s $600B Question | AI’s $200B question is now AI’s $600B question. |
| SO024 | U.S. News / Reuters | AI startup valuations raise bubble fears as funding surges | There's a little bit of a hype bubble going on in the early-stage venture space. |
| SO025 | arXiv | A Path Towards Autonomous Machine Intelligence | |
| SO026 | Meta | Our new model helps AI think before it acts | |
| SO027 | NVIDIA Newsroom | NVIDIA launches Cosmos world foundation model platform to accelerate physical AI development | NVIDIA launches Cosmos world foundation model platform to accelerate physical AI development |
| SO028 | World Labs | Introducing Marble, a world model for 3D worlds | Marble is our first step toward spatial intelligence and world models. |
| SM001 | AMI | Advanced Machine Intelligence homepage | We are building systems that understand the world and can plan safely. |
| SM002 | AMI | AMI Updates | |
| SM003 | Cathay Innovation | Advanced Machine Intelligence (AMI) is enabling the next AI revolution built on foundational world models | |
| SM004 | Nabla | Nabla announces exclusive partnership with Advanced Machine Intelligence to pioneer the next era of agentic healthcare AI | |
| SM005 | NVIDIA Newsroom | NVIDIA launches Cosmos world foundation model platform to accelerate physical AI development | |
| SM006 | arXiv | Cosmos World Foundation Model Platform for Physical AI | |
| SM007 | Google DeepMind | Gemini Robotics 1.5 | |
| SM008 | arXiv | Gemini Robotics: Bringing AI into the Physical World | |
| SM009 | Google DeepMind | Genie 2: A large-scale foundation world model | |
| SM010 | Physical Intelligence | π0: A vision-language-action flow model for general robot control | |
| SM011 | arXiv | π0: A Vision-Language-Action Flow Model for General Robot Control | |
| SM012 | Hugging Face | LeRobot | |
| SM013 | Wayve | Wayve raises over $1 billion led by SoftBank to develop embodied AI products for automated driving | |
| SM014 | Skild AI | Announcing our $300M Series A | |
| SM015 | CNBC | Jeff Bezos and OpenAI invest in robot startup Physical Intelligence at $2.4 billion valuation | |
| SM016 | Figure AI | Introducing Helix | |
| SM017 | MarketsandMarkets | Embodied AI Market Size, Share and Trends - Global Forecast to 2030 | |
| SM018 | ABI Research | Physical AI Robotics Foundation Models | |
| SM019 | International Federation of Robotics | World Robotics Report 2025 | |
| SM020 | Deloitte | 2025 Smart Manufacturing and Operations Survey | |
| SM021 | Bain & Company | Humanoid Robots: From Demos to Deployment | |
| SM022 | MIT Technology Review | Why the humanoid workforce is running late | |
| SM023 | European Union | Regulation (EU) 2024/1689 laying down harmonised rules on artificial intelligence | |
| SM024 | U.S. Food and Drug Administration | Artificial Intelligence/Machine Learning-Based Software as a Medical Device Action Plan | |
| SM025 | GOV.UK | Frontier AI Safety Commitments, AI Seoul Summit 2024 | |
| SM026 | NIST | Artificial Intelligence Risk Management Framework (AI RMF 1.0) | |
| SM027 | Epoch AI | How much does it cost to train frontier AI models? | |
| SM028 | Sequoia Capital | AI’s $600B Question | |
| SM029 | U.S. News / Reuters | AI startup valuations raise bubble fears as funding surges | |
| SP001 | AMI | AMI homepage | AMI is developing world models that learn abstract representations of real-world sensor data. |
| SP002 | AMI | AMI raises $1.03B to build world models | We’ve raised a $1.03B USD (~€890M) round from global investors who believe in our vision. |
| SP003 | Cathay Innovation | Advanced Machine Intelligence is enabling the next AI revolution built on foundational world models | The company has raised $1.03B USD (~€890M) – based on a $3.50B (~€3B) pre-money valuation. |
| SP004 | Nabla | Nabla announces exclusive partnership with Advanced Machine Intelligence | Nabla will gain first access to Advanced Machine Intelligence’s emerging world model technologies. |
| SP005 | MIT Technology Review | Yann LeCun’s new venture is a contrarian bet against large language models | Nobody—absolutely nobody—knows how to make those robots smart enough to be useful. |
| SP006 | World Labs | Marble: A Multimodal World Model | Today we are making Marble, a first-in-class generative multimodal world model, generally available for anyone to use. |
| SP007 | World Labs | World Labs Announces New Funding | World Labs has raised $1 billion in new funding. |
| SP008 | TechCrunch | World Labs speeds up the world model race with Marble | Marble is now available via freemium and paid tiers. |
| SP009 | TechCrunch | World Labs lands $1B, with $200M from Autodesk | Autodesk will serve as an adviser to World Labs, and the two will collaborate at the research and model level. |
| SP010 | Physical Intelligence | Physical Intelligence homepage | Physical Intelligence is bringing general-purpose AI into the physical world. |
| SP011 | Physical Intelligence | Our First Generalist Policy | π0 uses Internet-scale vision-language pre-pretraining, open-source robot manipulation datasets, and our own datasets. |
| SP012 | CNBC | Jeff Bezos and OpenAI invest in robot startup Physical Intelligence | Physical Intelligence ... has raised $400 million at a $2.4 billion post-money valuation. |
| SP013 | TechCrunch | Physical Intelligence is reportedly in talks to raise $1B, again | Co-founder Lachy Groom told TechCrunch the company has no timeline for commercialization. |
| SP014 | Skild AI | Skild AI homepage | Skild AI |
| SP015 | Skild AI | Announcing our $300M Series A Funding | We’ve raised $300 million in Series A funding, which values our company at $1.5 billion. |
| SP016 | Figure AI | Figure homepage | Figure |
| SP017 | Figure AI | Helix: A Vision-Language-Action Model for Generalist Humanoid Control | Helix is the first VLA that runs entirely onboard embedded low-power-consumption GPUs, making it immediately ready for commercial deployment. |
| SP018 | Figure AI | Figure Exceeds $1B in Series C Funding at $39B Post-Money Valuation | Today we’re announcing that we have exceeded more than $1 billion in committed capital through our Series C financing round, at a post-money valuation of $39 billion. |
| SP019 | IEEE Spectrum | Covariant Announces a Universal AI Platform for Robots | The robotics foundation model is built on tons of warehouse data. |
| SP020 | RoboticsTomorrow | Covariant Introduces RFM-1 to Give Robots the Human-like Ability to Reason | Covariant currently offers the broadest portfolio of AI-powered robotic picking applications for warehouse environments. |
| SP021 | Google DeepMind | Gemini Robotics 1.5 | Name Gemini Robotics 1.5 Status Private preview. |
| SP022 | Google DeepMind | Genie 2: A large-scale foundation world model | Genie 2 is a world model, meaning it can simulate virtual worlds, including the consequences of taking any action. |
| SP023 | NVIDIA Newsroom | NVIDIA launches Cosmos world foundation model platform | Cosmos models will be available under an open model license to accelerate the work of the robotics and AV community. |
| SP024 | NVIDIA Docs | NVIDIA Cosmos | NVIDIA Cosmos is a platform purpose-built for physical AI. |
| SP025 | GitHub | NVIDIA Cosmos organization | NVIDIA Cosmos |
| SP026 | Wayve | Wayve raises over $1 billion led by SoftBank | Wayve announces a $1.05 billion Series C investment round led by SoftBank Group. |
| SP027 | GitHub | Hugging Face LeRobot repository | LeRobot aims to provide models, datasets, and tools for real-world robotics in PyTorch. |
| SP028 | Hugging Face | LeRobot | LeRobot already provides a set of pretrained models, datasets with human collected demonstrations, and simulated environments. |
| SP029 | Bain & Company | Humanoid Robots: From Demos to Deployment | Most humanoid robots today remain in pilot phases, heavily dependent on human input for navigation, dexterity, or task switching. |
| SP030 | Sequoia Capital | AI’s $600B Question | GPU computing is increasingly turning into a commodity, metered per hour. |
| SP031 | Tesla Investor Relations | Q1 2026 Update | First-generation production lines for Optimus are being installed in anticipation of volume production. |
| SI001 | AMI | AMI homepage | AMI will advance AI research and develop applications where reliability, controllability, and safety really matter. |
| SI002 | AMI | AMI Labs - Updates | We’ve raised a $1.03B USD (~€890M) round from global investors. |
| SI003 | Cathay Innovation | Advanced Machine Intelligence is enabling the next AI revolution built on foundational world models | The company has raised $1.03B USD (~€890M) – based on a $3.50B (~€3B) pre-money valuation. |
| SI004 | TechCrunch | Yann LeCun's AMI Labs raises $1.03B to build world models | It’s not your typical applied AI startup that can release a product in three months, have revenue in six months, and make $10 million in annual recurring revenue in 12 months. |
| SI006 | Crunchbase News | Turing Winner LeCun’s New World Model AI Lab Raises $1B In Europe’s Largest Seed Round Ever | The funding for Paris-based AMI represents the largest seed round ever for a European startup. |
| SI007 | Nabla | Nabla Announces Exclusive Partnership With Advanced Machine Intelligence to Pioneer the Next Era of Agentic Healthcare AI | Through this partnership, Nabla will gain first access to Advanced Machine Intelligence’s emerging world model technologies. |
| SI008 | Nabla | AMI Raises $1.03B to Build World Models — Powering the Next Generation of Healthcare AI with Nabla | Nabla will gain first access to these emerging world model technologies, positioning us to help bring the next generation of safe, auditable agentic AI systems into healthcare. |
| SI009 | STAT | Health AI startup to benefit from $1 billion funding round for Yann LeCun’s AMI | AI documentation company Nabla will have early access to this new technology. |
| SI010 | Pappers | Société ADVANCED MACHINE INTELLIGENCE : Chiffre d’affaires, statuts, extrait d’immatriculation | La conception, le développement et l’entraînement de tous modèles d’intelligence artificielle. |
| SI011 | Annuaire des Entreprises | Advanced Machine Intelligence - 994675254 | Advanced Machine Intelligence - 994675254 |
| SI013 | MIT Technology Review | Yann LeCun’s new venture is a contrarian bet against large language models | The truly difficult part is understanding the real world. |
| SI014 | Le Monde Informatique | La start-up AMI Labs de Yann Lecun lève 890 millions d’euros | Elle a séduit des investisseurs internationaux et français sans pour l’instant avoir de système opérationnel. |
| SI015 | France 24 | French AI startup AMI raises $1B to develop universal intelligent systems | AMI would focus on research and development in its first year. |
| SI016 | RFI | French start-up raises $1 billion to shift AI research into high gear | Discussions with corporate partners could be held within six to 12 months. |
| SI017 | Sequoia Capital | AI’s $600B Question | Where is all the revenue? |
| SI018 | Epoch AI | How much does it cost to train frontier AI models? | The amortized hardware and energy cost for the final training run of frontier models has grown rapidly, at a rate of 2.4x per year since 2016. |
| SI019 | Epoch AI | Data on AI Models | Frontier models are models that were in the top 10 by training compute at the time of their release. |
| SI020 | NVIDIA Newsroom | NVIDIA Blackwell Platform Arrives to Power a New Era of Computing | New Tensor Cores and TensorRT-LLM Compiler reduce LLM inference operating cost and energy by up to 25x. |
| SI021 | NVIDIA Newsroom | NVIDIA Launches Cosmos World Foundation Model Platform to Accelerate Physical AI Development | Building physical AI models requires petabytes of video data and tens of thousands of compute hours to process, curate and label that data. |
| SI023 | U.S. Food and Drug Administration | Artificial Intelligence in Software | Medical device manufacturers are using these technologies to innovate their products to better assist health care providers. |
| SI024 | National Institute of Standards and Technology | AI Risk Management Framework | The profile will guide critical infrastructure operators towards specific risk management practices to consider when engaging AI-enabled capabilities. |
| SI026 | Lambda | Instances | Launch NVIDIA HGX B200, H100, A100, or GH200 instances in minutes with self-serve, first-come access. |
| SI027 | Modal | Plan Pricing | compute / month; 100 containers + 10 GPU concurrency; higher GPU concurrency. |
| SI029 | CloudZero | H100 GPU Cost In 2026: Buy, Rent, And Cloud Pricing Compared | A new NVIDIA H100 GPU costs $25,000–$40,000; cloud rental runs $1.38–$8.00+ per GPU-hour on-demand. |
| SI030 | Silicon Republic | Yann LeCun’s AI start-up AMI raises $1.03bn in seed funding | The seed funding round values the start-up at $3.5bn, despite having only been established this year. |
| SI031 | HIT Consultant | AMI Raises $1.03B to Build World Models: How Nabla is Powering the Next Generation of Healthcare AI | Clinical AI company Nabla holds an exclusive strategic partnership with AMI to get first access to these new AI models. |
| SE001 | AMI | AMI Labs: Real World. Real Intelligence. | AMI is developing world models that learn abstract representations of real-world sensor data, ignoring unpredictable details, and that make predictions in representation space. |
| SE002 | AMI | AMI Labs - Updates | Advanced Machine Intelligence (AMI) is building a new breed of AI systems that understand the world, have persistent memory, can reason and plan, and are controllable and safe. |
| SE003 | Nabla | Nabla Announces Exclusive Partnership With Advanced Machine Intelligence to Pioneer the Next Era of Agentic Healthcare AI | Nabla will gain first access to Advanced Machine Intelligence’s emerging world model technologies, positioning the company to become the first to bring FDA-certifiable agentic AI systems to healthcare. |
| SE004 | Nabla | AMI Raises $1.03B to Build World Models — Powering the Next Generation of Healthcare AI with Nabla | Nabla will gain first access to these emerging world model technologies, positioning us to help bring the next generation of safe, auditable agentic AI systems into healthcare. |
| SE005 | TechCrunch | Yann LeCun's AMI Labs raises $1.03B to build world models | AMI Labs is a very ambitious project, because it starts with fundamental research. It’s not your typical applied AI startup that can release a product in three months. |
| SE006 | MIT Technology Review | Yann LeCun’s new venture is a contrarian bet against large language models | Instead, he thinks we should be betting on world models—a different type of AI that accurately reflects the dynamics of the real world. |
| SE007 | OpenReview | A Path Towards Autonomous Machine Intelligence | How could machines learn to reason and plan? How could machines learn representations of percepts and action plans at multiple levels of abstraction, enabling them to reason, predict, and plan? |
| SE008 | arXiv | Introduction to Latent Variable Energy-Based Models: A Path Towards Autonomous Machine Intelligence | We introduce energy-based and latent variable models and combine their advantages in the building block of LeCun's proposal, that is, in the hierarchical joint embedding predictive architecture (H-JEPA). |
| SE009 | arXiv | Self-Supervised Learning from Images with a Joint-Embedding Predictive Architecture | We introduce the Image-based Joint-Embedding Predictive Architecture (I-JEPA), a non-generative approach for self-supervised learning from images. |
| SE010 | arXiv | Revisiting Feature Prediction for Learning Visual Representations from Video | This paper explores feature prediction as a stand-alone objective for unsupervised learning from video and introduces V-JEPA. |
| SE011 | GitHub | facebookresearch/jepa: PyTorch code and models for V-JEPA | Official PyTorch codebase for the video joint-embedding predictive architecture, V-JEPA, a method for self-supervised learning of visual representations from video. |
| SE012 | arXiv | V-JEPA 2: Self-Supervised Video Models Enable Understanding, Prediction and Planning | V-JEPA 2 achieves strong performance on motion understanding and state-of-the-art performance on human action anticipation and video question-answering benchmarks. |
| SE013 | Meta AI | Introducing the V-JEPA 2 world model and new benchmarks for physical reasoning | V-JEPA 2 is a world model that achieves state-of-the-art performance on visual understanding and prediction in the physical world. Our model can also be used for zero-shot robot planning. |
| SE014 | GitHub | facebookresearch/vjepa2: PyTorch code and models for VJEPA2 | V-JEPA 2-AC is a latent action-conditioned world model post-trained from V-JEPA 2 using a small amount of robot trajectory interaction data. |
| SE015 | Hugging Face | V-JEPA 2 - a facebook Collection | A frontier video understanding model developed by FAIR, Meta, which extends the pretraining objectives of V-JEPA. |
| SE016 | U.S. Food and Drug Administration | Artificial Intelligence in Software | The FDA may also review and clear modifications to medical devices, including software as a medical device, depending on the significance or risk posed to patients of that modification. |
| SE017 | U.S. Food and Drug Administration | Predetermined Change Control Plan for Artificial Intelligence-Enabled Device Software Functions | This guidance recommends that a PCCP describe the planned device modifications, the associated methodology to develop, validate, and implement those modifications, and an assessment of the impact. |
| SE018 | National Institute of Standards and Technology | AI Risk Management Framework | NIST released a concept note for an AI RMF Profile on Trustworthy AI in Critical Infrastructure. |
| SE019 | European Union | Regulation (EU) 2024/1689 Artificial Intelligence Act | The purpose of this Regulation is to improve the functioning of the internal market and promote the uptake of human centric and trustworthy artificial intelligence while ensuring a high level of protection of health, safety, fundamental rights. |
| SE020 | NVIDIA Newsroom | NVIDIA Launches Cosmos World Foundation Model Platform to Accelerate Physical AI Development | Building physical AI models requires petabytes of video data and tens of thousands of compute hours to process, curate and label that data. |
| SE021 | arXiv | Cosmos World Foundation Model Platform for Physical AI | Physical AI needs to be trained digitally first. It needs a digital twin of itself, the policy model, and a digital twin of the world, the world model. |
| SE022 | GitHub | huggingface/lerobot: Making AI for Robotics more accessible | LeRobot aims to provide models, datasets, and tools for real-world robotics in PyTorch. |
| SE023 | Google DeepMind | Genie 2: A large-scale foundation world model | Genie 2 is a foundation world model capable of generating an endless variety of action-controllable, playable 3D environments for training and evaluating embodied agents. |
| SE024 | Google DeepMind | Gemini Robotics | Gemini Robotics 1.5 is a vision-language-action model that turns visual information and instructions into motor commands to perform a task. |
| SE025 | STAT | Health AI startup to benefit from $1 billion funding round for Yann LeCun's AMI | Advanced Machine Intelligence (AMI), the new company from former Meta chief AI scientist Yann LeCun, announced it had raised $1 billion for its quest to develop world models. |
| SE026 | HIT Consultant | AMI Raises $1.03B to Build World Models: How Nabla is Powering the Next Generation of Healthcare AI | Instead of just predicting the next word, AMI is building world models that learn abstract representations of reality, allowing them to simulate environments, anticipate consequences, and plan sequential actions. |
| SE027 | MIT Technology Review | Why the humanoid workforce is running late | Some impressive humanoid demos don’t overcome core constraints as much as they display other impressive features. |
| SE028 | Bain & Company | Humanoid Robots: From Demos to Deployment | Safety will remain paramount, and use cases will expand into open, guest-facing areas only as certification and human-acceptance thresholds are met. |
| SU001 | Advanced Machine Intelligence | AMI Labs: Real World. Real Intelligence. | AMI will advance AI research and develop applications where reliability, controllability, and safety really matter, especially for industrial process control, automation, wearable devices, robotics, healthcare, and beyond. |
| SU002 | Advanced Machine Intelligence | AMI Labs - Updates | AMI Labs - Updates |
| SU003 | Nabla | Nabla Announces Exclusive Partnership With Advanced Machine Intelligence to Pioneer the Next Era of Agentic Healthcare AI | Through this partnership, Nabla will gain first access to Advanced Machine Intelligence’s emerging world model technologies. |
| SU004 | Nabla | AMI Raises $1.03B to Build World Models — Powering the Next Generation of Healthcare AI with Nabla | Through our exclusive strategic partnership with AMI announced at the end of 2025, Nabla will gain first access to these emerging world model technologies. |
| SU005 | TechCrunch | Yann LeCun’s AMI Labs raises $1.03B to build world models | AMI Labs is a very ambitious project, because it starts with fundamental research. It’s not your typical applied AI startup that can release a product in three months, have revenue in six months, and make $10 million in annual recurring revenue in 12 months. |
| SU006 | STAT | Health AI startup to benefit from $1 billion funding round for Yann LeCun’s AMI | There’s no formal equity or licensing agreement relationship yet between AMI and Nabla, but the companies are already working closely together, Nabla COO Delphine Groll told me. |
| SU007 | HIT Consultant | AMI Raises $1.03B to Build World Models: How Nabla is Powering the Next Generation of Healthcare AI | AMI Raises $1.03B to Build World Models: How Nabla is Powering the Next Generation of Healthcare AI |
| SU008 | Cathay Innovation | Advanced Machine Intelligence (AMI) is Enabling the Next AI Revolution — Built on Foundational World Models | The round brings together leading global investment firms and strategic partners, aligned with AMI’s long-term scientific and industrial ambition. |
| SU009 | Nabla | Case Studies · Nabla | Discover how leading health systems have successfully implemented Nabla in their organizations to improve clinical workflows. |
| SU010 | Nabla | Denver Health Deploys Nabla’s Ambient AI to Expand Care by Reducing Administrative Burden Across Its Health System | After beginning with an eight-week pilot deployment of Nabla’s ambient AI assistant, Denver Health saw meaningful reductions in documentation burden and measurable improvements in patient experience and clinician workflow. |
| SU011 | Nabla | Nabla Rolls Out at Carle Health Through Epic Integration | 55% of clinicians saved at least 1 hour in documentation time with Nabla. |
| SU012 | Nabla | McFarland Clinic Taps Nabla to Support Clinicians and Streamline Documentation in Epic | 80% retention rate following the pilot. |
| SU013 | Nabla | Tia Health Selects Nabla’s AI Assistant to Enhance Patient-Centered Care | 50% reduction in clinical note submission time. |
| SU014 | Nabla | Beating Burnout: How CHLA Used Nabla to Support Physician Wellbeing | 47% decrease in physician burnout. |
| SU015 | HIT Consultant | University of Toledo Health to Deploy Nabla’s Ambient AI Documentation in Epic EHR | Clinicians reduced their time to chart closure by 29%. |
| SU016 | PR Newswire | University of Toledo Health Scales Nabla’s Ambient AI After Unlocking Better Documentation Precision and Revenue Cycle Performance | During the 8-week initial evaluation phase, clinicians using Nabla reduced time to chart closure by 29%. |
| SU017 | Fierce Healthcare | Carle Health teams up with Nabla for AI scribe assistant | Nabla now has contracts with The Permanente Medical Group, Children’s Hospital of Los Angeles, Stratum Med and the Mankato Clinic and is deployed in more than 40 health organizations. |
| SU018 | Healthcare IT Today | Carle Health Chooses Nabla for Its Workflow Focus | Carle Health chose Nabla after testing many ambient voice products; Dr. Lovinger says they are not all created equal. |
| SU019 | TMCnet | Aultman Health System Scales Nabla’s Ambient AI Through Oracle Cerner Integration | Early results from Aultman’s deployment show clinicians saving between 30 and 60 minutes a day on documentation. |
| SU020 | Sequoia Capital | AI’s $600B Question | AI’s $600B Question |
| SU021 | Forbes | Yann LeCun’s New Startup AMI Labs: Can World Models Move Beyond Hype? | Can World Models Move Beyond Hype? |
| SU022 | Sacra | AMI Labs commercialization risk from deploy-focused rivals | The key risk is not that AMI is wrong in theory, it is that it may be too slow to prove a clear production advantage while rivals build data loops, customer trust, and shipping products. |
| SU023 | The French Tech Journal | AMI Labs Tracker | AMI Labs begins scaling its team across Paris, New York, Montreal, and Singapore. |
| SU024 | HLTH | AMI and Nabla Advance World Models to Power Agentic Healthcare AI | Nabla plans to integrate AMI’s technology into its next phase of product development, moving beyond ambient documentation toward Agentic AI. |
| SU025 | Nabla | Nabla Raises $70M Series C to Deliver Agentic AI to the Heart of Clinical Workflows, Bringing Total Funding to $120M | Trusted by 130+ healthcare organizations and 85,000 clinicians, Nabla is expanding its AI assistant to support coding, agentic EHR commands, and a wider range of clinical roles. |
| SU026 | Nabla | $70M Series C: Just Getting Started | Nabla’s Ambient AI is used by 85,000 clinicians across more than 130 healthcare organizations from rural hospitals to academic medical centers, FQHCs, and national providers. |
| SU027 | The Healthcare Technology Report | Nabla Secures $70M Series C to Advance Clinical AI Platform | Nabla’s AI assistant is currently used by over 85,000 clinicians across more than 130 U.S. healthcare organizations. |
| SU028 | Highland Europe | Nabla Raises $70M Series C to Deliver Agentic AI to the Heart of Clinical Workflows, Bringing Total Funding to $120M | The company has multiplied its revenue by five over the past 6 months and now supports more than 85,000 clinicians and 20 million annual encounters. |
| SR001 | Advanced Machine Intelligence | AMI Labs: Real World. Real Intelligence. | AMI presents itself as building world models for real-world intelligence. |
| SR002 | Advanced Machine Intelligence | AMI Labs - Updates | AMI announced a $1.03B funding round to build world models. |
| SR003 | TechCrunch | Yann LeCun’s AMI Labs raises $1.03B to build world models | AMI Labs has raised $1.03 billion at a $3.5 billion pre-money valuation. |
| SR004 | TechCrunch | Yann LeCun confirms his new world model startup reportedly seeks $5B valuation | AMI hired Alex LeBrun, co-founder and CEO of Nabla, as CEO. |
| SR005 | MIT Technology Review | Yann LeCun’s new venture is a contrarian bet against large language models | LeCun thinks the industry should be betting on world models rather than large language models. |
| SR006 | Forbes | Yann LeCun’s New Startup AMI Labs: Can World Models Move Beyond Hype? | The article frames AMI around whether world models can move beyond hype. |
| SR008 | Epoch AI | How much does it cost to train frontier AI models? | If trends continue, the largest training runs will cost more than a billion dollars by 2027. |
| SR009 | arXiv | The rising costs of training frontier AI models | |
| SR011 | NIST | Artificial Intelligence Risk Management Framework (AI RMF 1.0) | The AI RMF is intended to be a living document for AI risk management. |
| SR013 | U.S. Food and Drug Administration | AI-Enabled Medical Devices | FDA maintains a public list of AI-enabled medical devices. |
| SR016 | Nabla | Nabla announces exclusive partnership with Advanced Machine Intelligence | Nabla will gain first access to AMI’s emerging world model technologies. |
| SR017 | Nabla | AMI Raises $1.03B to Build World Models — Powering the Next Generation of Healthcare AI with Nabla | Nabla says the AMI partnership positions it to bring safe, auditable agentic AI systems into healthcare. |
| SR018 | STAT | AI startup to benefit from $1 billion funding round for Yann LeCun’s AMI | STAT describes the world-model developer tied to Nabla receiving $1 billion. |
| SR019 | HIT Consultant | AMI Raises $1.03B to Build World Models: How Nabla is Powering the Next Generation of Healthcare AI | |
| SR020 | Cathay Innovation | Advanced Machine Intelligence is enabling the next AI revolution built on foundational world models | The round supports long-term research, global hiring, and reliable intelligent systems. |
| SR021 | Crunchbase News | Turing Winner LeCun’s New World Model AI Lab Raises $1B In Europe’s Largest Seed Round Ever | Crunchbase characterizes the raise as Europe’s largest seed round ever. |
| SR022 | Observer | Yann LeCun’s Paris A.I. Startup AMI Labs Raises Record $1B Seed Round | AMI is hiring across New York, Montreal, Paris and Singapore. |
| SR023 | EU-Startups | Beyond LLMs: Yann LeCun’s new venture AMI raises €890 million | |
| SR024 | World Labs | Marble: A Multimodal World Model | World Labs describes Marble as reconstructing, generating, and simulating 3D worlds. |
| SR025 | World Labs | World Labs Announces New Funding | World Labs says it is focused on building world models for spatial intelligence. |
| SR026 | Google DeepMind | Gemini Robotics 1.5 | |
| SR027 | NVIDIA Newsroom | NVIDIA launches Cosmos world foundation model platform | Building physical AI models requires petabytes of video data and tens of thousands of compute hours. |
| SR028 | Meta AI | Yann LeCun | Yann is Chief AI Scientist for Facebook AI Research and a part-time NYU professor. |
| SR029 | NYU Center for Data Science | Yann LeCun | NYU lists LeCun as Professor of Computer Science, Neural Science, Data Science, and Electrical and Computer Engineering. |
| SR030 | Pappers | Informations juridiques de Advanced Machine Intelligence | Pappers lists the French legal entity Advanced Machine Intelligence. |
| SR031 | Annuaire des Entreprises | Advanced Machine Intelligence - 994675254 | |
| SR037 | GOV.UK | AI Safety Summit 2023: The Bletchley Declaration | |
| SR038 | GOV.UK | AI Safety Summit 2023: Chair’s statement – safety testing | Frontier AI companies recognised increased emphasis on AI safety testing and research. |
| SR039 | EUR-Lex | Regulation (EU) 2024/1689 - Artificial Intelligence Act | High-risk AI systems should be subject to conformity assessment before being placed on the market. |
| SR040 | European Commission | AI Act | High-risk AI systems are subject to strict obligations before they can be put on the market. |
| SR042 | Sequoia Capital | AI’s $600B Question | AI’s $200B question is now AI’s $600B question. |
| SR044 | HHS / FDA | Marketing Submission Recommendations for a Predetermined Change Control Plan for AI-Enabled Device Software Functions | |
| SR045 | CourtListener | Kadrey v. Meta Platforms, Inc., 3:23-cv-03417 | The docket includes a class action complaint alleging copyright infringement against Meta Platforms. |
| SV001 | Advanced Machine Intelligence | AMI Labs - Updates | We’ve raised a $1.03B USD (~€890M) round from global investors who believe in our vision of universally intelligent systems centered on world models. |
| SV002 | TechCrunch | Yann LeCun’s AMI Labs raises $1.03 billion to build world models | AMI Labs has raised $1.03 billion at a $3.5 billion pre-money valuation. |
| SV003 | TechCrunch | Yann LeCun confirms his new world model startup reportedly seeks $5B valuation | AMI Labs is also reportedly seeking to raise €500 million ... at a €3 billion valuation ... before even launching. |
| SV004 | Cathay Innovation | Advanced Machine Intelligence is enabling the next AI revolution built on foundational world models | The company has raised $1.03B USD (~€890M) – based on a $3.50B (~€3B) pre-money valuation. |
| SV005 | Nabla | Nabla announces exclusive partnership with Advanced Machine Intelligence | Nabla will gain first access to Advanced Machine Intelligence’s emerging world model technologies. |
| SV006 | Nabla | AMI raises $1.03B to build world models — powering the next generation of healthcare AI with Nabla | AMI’s $1.03B funding round reflects growing recognition that the next phase of AI will require new foundational architectures. |
| SV007 | MIT Technology Review | Yann LeCun’s new venture is a contrarian bet against large language models | Meta might be our first client! We’ll see. The work we are doing is not in direct competition. |
| SV008 | Forbes | Yann LeCun’s New Startup AMI Labs: Can World Models Move Beyond Hype? | Can World Models Move Beyond Hype? |
| SV009 | Sequoia Capital | AI’s $600B Question | The AI bubble is reaching a tipping point. Navigating what comes next will be essential. |
| SV010 | U.S. News / Reuters | AI startup valuations raise bubble fears as funding surges | Any company startup with an AI label will be valued right up there at huge multiples of whatever the small revenue is. |
| SV011 | Crunchbase News | World Model AI Lab AMI Raises Europe’s Largest Seed Round | The funding for Paris-based AMI represents the largest seed round ever for a European startup. |
| SV012 | Observer | Yann LeCun’s Paris A.I. Startup AMI Labs Raises Record $1B Seed Round | The funding values AMI at $3.5 billion pre-money. |
| SV013 | STAT | World model developer tied to Nabla gets $1 billion | There’s no formal equity or licensing agreement relationship yet between AMI and Nabla, but the companies are already working closely together. |
| SV014 | World Labs | World Labs announces new funding | World Labs has raised $1 billion in new funding. |
| SV015 | TechCrunch | World Labs lands $1B, with $200M from Autodesk | World Labs ... emerged from stealth in 2024 with $230 million at a $1 billion valuation. |
| SV016 | World Labs | Introducing Marble, a world model for 3D worlds | Marble is the first of its kind - a next-generation world model making strides toward this vision. |
| SV017 | Physical Intelligence | π0: A vision-language-action flow model for general robot control | Over the past eight months, we’ve developed a general-purpose robot foundation model that we call π0. |
| SV018 | CNBC | Jeff Bezos and OpenAI invest in robot startup Physical Intelligence at $2.4 billion valuation | Physical Intelligence ... has raised $400 million at a $2.4 billion post-money valuation. |
| SV019 | TechCrunch | Physical Intelligence is reportedly in talks to raise $1B, again | Physical Intelligence ... is in discussions to raise about $1 billion in new funding at a valuation exceeding $11 billion. |
| SV020 | Mistral AI | Mistral AI raises 1.7B€ to accelerate technological progress with AI | We are announcing a Series C funding round of 1.7B€ at a 11.7B€ post-money valuation. |
| SV021 | ASML | ASML and Mistral AI enter strategic partnership | ASML is investing 1.3 billion EUR in Mistral AI’s Series C funding round as lead investor. |
| SV022 | NVIDIA Newsroom | NVIDIA launches Cosmos world foundation model platform to accelerate physical AI development | NVIDIA today announced NVIDIA Cosmos, a platform comprising ... generative world foundation models. |
| SV023 | U.S. Securities and Exchange Commission | NVIDIA Corporation Form 10-K for fiscal year 2025 | The aggregate market value of the voting stock held by non-affiliates ... was approximately $2.7 trillion. |
| SV024 | U.S. Securities and Exchange Commission | CoreWeave Inc. S-1/A registration statement | We expect that the initial public offering price per share of our Class A common stock will be between $47.00 and $55.00. |
| SV027 | SiliconANGLE | PitchBook: US venture funding surges to record $267B as OpenAI, Anthropic and xAI dominate AI deals | Databricks Inc. also raised $7 billion, with the five deals representing 73% of total U.S. venture deal value during the quarter. |
| SV028 | National Venture Capital Association | PitchBook-NVCA Venture Monitor | Q1 2026 PitchBook-NVCA Venture Monitor ... offers an in-depth view of the US venture capital. |
| SV029 | The Hindu BusinessLine / Reuters | AI start-up valuations raise bubble fears as funding surges | Artificial intelligence start-ups are attracting record sums of venture capital, but some of the world's largest investors warned that early-stage valuations are starting to look frothy. |
| SV030 | Epoch AI | How much does it cost to train frontier AI models? | If the trend of growing training costs continues, the largest training runs will cost more than a billion dollars by 2027. |
| SV031 | Bain & Company | Humanoid Robots: From Demos to Deployment | While demonstrations dazzle, most deployments remain early-stage, with heavy reliance on human supervision. |
| SV032 | MIT Technology Review | Why the humanoid workforce is running late | A frenzied venture capital market ... is betting that humanoids will create the largest market for robotics the field has ever seen. |
| SV033 | European Union | Regulation (EU) 2024/1689 laying down harmonised rules on artificial intelligence | A Union legal framework laying down harmonised rules on AI is therefore needed. |
| SV034 | U.S. Food and Drug Administration | Artificial Intelligence in Software | Many changes to artificial intelligence and machine learning-driven devices may need a premarket review. |
| SV035 | Google DeepMind | Gemini Robotics 1.5 | Adapts to a diverse array of robot forms. |
| SV036 | Figure AI | Figure exceeds $1B in Series C funding at $39B post-money valuation | We have exceeded more than $1 billion in committed capital ... at a post-money valuation of $39 billion. |
| SV037 | Wayve | Wayve raises over $1 billion led by SoftBank to develop embodied AI products for automated driving | Wayve announces a $1.05 billion Series C investment round led by SoftBank Group. |
| SV038 | Skild AI | Announcing our $300M Series A | We’ve raised $300 million in Series A funding, which values our company at $1.5 billion. |
| SV039 | MarketsandMarkets | Embodied AI Market Size, Share and Trends - Global Forecast to 2030 | The embodied AI market size is projected to reach USD 23.06 billion in 2030 from USD 4.44 billion in 2025. |
| SV040 | Dataconomy | Yann LeCun’s AMI Labs hits $3.5 billion pre-money valuation | The raise is one of the largest pre-revenue AI raises in history. |