AMI Labs
尚未产生收入的前沿 AI 研究实验室——以 JEPA 架构构建世界模型
AMI Labs 是科学上可信、商业上未验证的世界模型研究押注;在 JEPA 基准证据和 Nabla 之外的商业锚点出现前,应按 $3.5B 种子轮估值继续观察。
封面要素
公司概况
AMI Labs(Advanced Machine Intelligence Labs)是一家总部位于巴黎的前沿 AI 研究公司,由图灵奖得主 Yann LeCun(执行董事长)和连续创业者 Alexandre LeBrun(CEO)于 2025 年底共同创立。公司正在基于 LeCun 的 Joint Embedding Predictive Architecture(JEPA)开发世界模型——这是一种非生成式架构,不在像素或 token 空间预测未来世界状态,而是在抽象表征空间中预测,目标是带来持久记忆、规划能力和可控性。目标应用领域包括工业流程控制、医疗诊断、临床工作流、机器人和可穿戴设备。AMI 在四个枢纽运营:巴黎、纽约、蒙特利尔和新加坡。公司于 2026 年 3 月以 $3.5B 投前估值完成 $1.03B 种子轮,并将资金投向算力和人才,而非短期收入。截至运行日,公司尚未发布产品,也未产生商业收入。
- 成立时间
- 2025-12-15
- 创始人
- Yann LeCun, Alexandre LeBrun, Saining Xie
- 创立地点
- Paris, France
- 总部
- Paris, France
- 产品
- 基于 JEPA 构建的行动条件世界模型,可在抽象表征空间预测环境转移;面向工业流程控制、医疗诊断、机器人和可穿戴设备设计。截至 2026 年 6 月尚未发布产品;公司仍处于延长研究阶段。计划商业化路径是向行业伙伴授权技术;开源论文和代码发布并行推进。
- 客户
- 制造、医疗和机器人领域的工业企业买家,需要比 LLM 更可靠、可控的 AI 替代方案;欧洲和亚太的主权买家及受监管市场买家。
- 商业模式
- 向行业伙伴授权世界模型基础设施技术;当前无收入。开源研究发布作为并行定位路径。公司明确不计划短期变现。
- 阶段
- Seed (post-$1.03B round)
- 融资情况
- 2026 年 3 月 10 日完成 $1.03B 种子轮;投前估值 $3.5B;由 Cathay Innovation、Greycroft、Hiro Capital、HV Capital 和 Bezos Expeditions 共同领投。战略参与方包括 NVIDIA、Temasek、Samsung 和 Toyota Ventures。法国机构支持者包括 Bpifrance Digital Venture、Association Familiale Mulliez 和 Publicis Groupe。资金指定用于算力和人才;近期不计划产生收入。
执行摘要
主要优势
- Yann LeCun 的 2018 年图灵奖、12 年 FAIR 领导经历和 JEPA 作者身份,让 AMI 在世界模型 AI 上拥有思想权威。
- $1.03B 种子资金估计能为重算力研究提供 2–3 年现金跑道,短期没有融资压力。
- 战略投资人组合(NVIDIA、Temasek、Samsung、Toyota Ventures)释放出工业部署意图,也可能带来 Nabla 之外的合作伙伴管线。
- 主权 AI 与巴黎定位,让公司更容易触达欧盟受监管行业采购、Bpifrance 公共资金,以及偏好非美国超大规模云厂商的买家。
- CEO LeBrun 曾创办 Nabla,使团队与 AMI 的主要商业垂直有直接领域重叠,也拥有一个运行中的临床 AI 试验场。
主要风险
- 对 Yann LeCun 的关键人物依赖是最大生存风险;他是执行董事长而非全职 CEO,一旦离开,会显著打击投资人信心和招聘。
- JEPA 的架构押注未必能跑赢 Meta、Google DeepMind 等多模态 LLM 的进展;这些对手的资源量级远高于 AMI。
- 年度算力成本升高 4–5×,意味着商业收入验证投资逻辑之前,公司需要连续融资。
- 多年研究周期带来融资缺口风险:若 2028 年前无法以不低于 $3.5B 的投前估值完成 Series A,就会进入承压情景。
- Goldman Sachs / Sequoia 所称 $600B 资本开支与收入缺口,是宏观层面的反向风险,可能削弱企业买家为未产生收入的 AI 研究大规模付费的意愿。
未决问题
- 独立 JEPA 基准结果,需要证明其在物理世界任务上优于多模态 LLM;这是最主要的投资逻辑破裂触发点。
- Series A 时间表和定价条件(预计 2027–2028;目标投前估值 $5B+);若无法从 $3.5B 抬升,将释放估值压缩信号。
- 当前员工数,以及 2026 年 3 月种子轮完成时估计约 12 人之外的团队建设情况。
- Nabla 合作的商业条款、排他条件和收入分成结构。
- 董事会构成、治理节奏和完整顾问名单;这些尚未公开。
- 当 AMI 模型接近发布或部署门槛时,对 EU AI Act GPAI 系统性风险要求的合规姿态。
目录
01公司概览
1.1 身份、总部与商业模式
Advanced Machine Intelligence Labs——以 AMI Labs 名义运营——是一家在法国巴黎注册并设总部的前沿 AI 研究公司。自 2025 年底成立以来,公司就在四个国际枢纽运营:巴黎(总部)、纽约、蒙特利尔和新加坡。巴黎地址把 AMI 锚定在法国日渐成形的 AI 生态中,与 Mistral AI 和 Meta 的 FAIR 实验室并列;LeCun 表示,这一选择反映了他的判断:前沿 AI 不应只是美国或中国的事业。 AMI 的全名 Advanced Machine Intelligence 直接写出了野心:构建超越以语言为中心的模型、能直接理解物理世界的 AI 系统。公司当前处于收入前研究阶段;没有商业产品,也没有近期产生收入的计划。未来启动变现时,AMI 打算向行业伙伴授权世界模型技术,而不是直接销售消费者或企业软件。公司公开承诺发布开源代码和开放学术论文,把自己定位为服务更广泛社区的研究平台。法国总统 Emmanuel Macron 公开欢迎 AMI 将总部设在巴黎,并称其为法国 AI 的里程碑。 [CO001, CO002, CO003, CO004, CO005, CO006]
| 指标 | 数值 / 状态 | 日期 / 版本 | 信心 | 缺口 / 尽调路径 |
|---|---|---|---|---|
| 累计融资额 | $1.03B USD (~€890M) | March 2026 | 高 | 无;已由一手来源交叉确认 |
| 投前估值 | $3.5B USD (~€3B) | March 2026 | 高 | 无;官方新闻稿及多家一线新闻源确认 |
| 成立日期 | Late 2025 (Nov–Dec) | December 2025 | 高 | 精确法律注册日期未披露 |
| 总部 | 法国巴黎 | January 2026 | 高 | 公司和新闻源确认 |
| 运营枢纽 | 巴黎、纽约、蒙特利尔、新加坡 | Jan 2026 | 高 | 新加坡办公室人员规模未知 |
| 员工数 | 种子轮发布时 ~12 人(估计) | March 2026 | 低 | 当前员工数未披露;招聘页显示仍在招聘 |
| 年收入 | 收入前 / $0 | June 2026 | 高 | 公司确认近期无收入计划;授权模式尚未部署 |
| 产品 | 无 / 研究阶段 | June 2026 | 高 | 截至运行日未发布原型、benchmark 或论文 |
员工数估计(~12)来自 Futurum Group 在种子轮发布时的分析;此后可能已大幅增长。收入和产品行反映公司自己的披露。所有 USD 数字均基于发布时披露的 USD/EUR 汇率。
[CO003, CO004, CO006, CO007, CO029, CO030]AMI Labs 关键状态指标快照,涵盖资本、团队规模、产品成熟度和执行风险信号。
员工数(约 12 人)是 Futurum Group 在 2026 年 3 月种子轮公告时的估计;此后可能已有增长。产品数为零、论文数为零反映运行日期状态;AMI 已承诺未来开放发表。
[CO046, CO047, CO048, CO049]1.2 技术愿景——世界模型与 JEPA
AMI Labs 的整个技术论点建立在 LeCun 的 Joint Embedding Predictive Architecture(JEPA)之上。JEPA 最早出现在 2022 年一篇立场论文中,随后通过包括 I-JEPA(2023)在内的论文继续完善。JEPA 的核心洞察是:生成式方法在高维像素或 token 空间中预测世界未来状态,但物理世界 AI 必须处理连续、嘈杂、高维的传感器数据,这种路径从根上不适配。JEPA 转而训练模型在抽象表征空间中预测:学习世界如何变化的关键因素,而不是试图重建表面外观。 AMI 正在开发行动条件世界模型:系统预测智能体行动的后果,并在真实世界约束和安全护栏下规划多步行动序列。AMI 认为,这类世界模型会表现出持久记忆、推理和规划能力,以及可控性——这些特性是当前 LLM 在结构上无法提供的。目标应用领域包括工业流程控制、工厂自动化、医疗诊断和临床工作流、机器人以及可穿戴设备。这些场景里,LLM 幻觉会带来真实成本,可靠性和可控性不能让步。Nabla 合作是 AMI 的第一个商业联盟,专门用来在高风险临床环境中测试世界模型的可靠性。 AMI 的论点还带有结构性经济含义:更小、特定领域的世界模型如果用相关传感器数据训练,应当比万亿参数 LLM 需要少得多的算力,从而有可能以当前基础设施成本的一小部分,把前沿 AI 部署到设备端或工业现场。 [CO009, CO010, CO011, CO012, CO013, CO014]
AMI 的资本、人才和 JEPA 研究如何连接到世界模型授权收入和伙伴价值。
[CO006, CO009, CO011, CO014, CO029, CO043]1.3 创始团队与领导层
AMI Labs 围绕 Yann LeCun 的科学声誉和 Alexandre LeBrun 的运营履历搭建,支撑高管大多来自 Meta 的 AI 研究组织。 Yann LeCun 获得 2018 年 ACM A.M. 图灵奖——计算领域的诺贝尔奖——并与 Yoshua Bengio、Geoffrey Hinton 共同因深度神经网络奠基性工作获奖。在 2025 年 11 月离开前,他在 Meta 的 FAIR 实验室担任 VP 和首席 AI 科学家超过十年;据报道,离职源于与 Mark Zuckerberg 在 AI 战略上的分歧。LeCun 保留 NYU 教职,并担任 AMI 执行董事长;他明确不是 CEO。他把 AMI 总部设在巴黎,同时保留纽约基地,凸显公司横跨多洲的全球结构。 Alexandre LeBrun 以 CEO 身份带来运营深度。他共同创立 Wit.ai(2015 年出售给 Facebook),随后共同创立总部位于巴黎的医疗 AI 公司 Nabla;在转入 AMI 前,Nabla 已融资 $120M,ARR 增长三倍。LeBrun 也曾在 Meta FAIR 为 LeCun 工作。首席科学官 Saining Xie 来自 Google DeepMind,拥有超过 90,000 次研究引用,并共同创造 Diffusion Transformers(DiT)——许多领先生成系统背后的架构。首席研究与创新官 Pascale Fung 是 HKUST 教授,也是以人为中心 AI 的先驱。Michael Rabbat(世界模型 VP)和 Laurent Solly(COO,前 Meta 欧洲 VP)补齐高管团队。对 LeCun 的关键人物依赖是最重大的单人风险:他的可信度支撑投资者信心和招聘。 [CO016, CO017, CO018, CO019, CO020, CO021]
| 人物 | 角色 | 关键背景 | 创始人—市场匹配 | 关键人物依赖 |
|---|---|---|---|---|
| Yann LeCun | 执行董事长(联合创始人) | 2018 ACM Turing Award;在 Meta FAIR 担任 VP 兼首席 AI 科学家 10+ 年;NYU 教授;JEPA 提出者 | 发明现代深度学习骨干;创造 AMI 正在执行的技术论点 | 关键——科学可信度、投资人吸引力和 JEPA IP 都集中在 LeCun 身上 |
| Alexandre LeBrun | CEO(联合创始人) | Wit.ai 联合创始人(2015 年出售给 Facebook);Nabla 联合创始人及前 CEO;Meta FAIR 老兵 | 巴黎 AI 生态操盘手;临床 AI 经验直接贴合首批用例 | 高——唯一已识别 CEO 人选;未披露继任计划 |
| Saining Xie | 首席科学官(联合创始人) | Google DeepMind;Meta FAIR;NYU 教职;DiT 共同创建者;90K+ 研究引用 | 顶级 ML 架构能力;把理论 JEPA 接到工程执行 | 高——把 LeCun 研究愿景转化为可训练系统的关键人物 |
| Pascale Fung | 首席研究与创新官 | HKUST 教授;以人为中心 AI 先驱;NLP 和多模态专长 | 全球研究网络;新加坡枢纽入口;覆盖 HRI 领域 | 中——重要,但有团队深度兜底 |
| Michael Rabbat | 世界模型 VP | 前 Meta 研究员;世界模型领域专家 | 掌握 AMI 主要研究项目的核心技术 | 中——领域关键,但外部可见度较低 |
| Laurent Solly | COO | 前 Meta 欧洲 VP;大规模运营和政府 / 企业关系 | 欧洲监管和企业市场准入;运营扩张能力 | 中——组织执行赋能者;可替代但有成本 |
枚举覆盖截至 2026 年 3 月公开确认的执行层领导。董事会构成和顾问名单未公开披露;尽调应要求提供完整治理清单。
[CO016, CO017, CO018, CO019, CO020, CO021]1.4 融资、估值与投资方联盟
2026 年 3 月 10 日,AMI Labs 宣布完成 $1.03B 美元(约 €890M)种子轮,投前估值 $3.5B。多家媒体引用的 PitchBook 数据确认,这是欧洲有记录以来最大种子轮。该轮融资显著超过 Financial Times 在 2025 年 12 月报道的 €500M 初始目标。 本轮由五家投资方共同领投:Cathay Innovation(曾在 Nabla 支持 LeBrun)、Greycroft、Hiro Capital、HV Capital,以及 Jeff Bezos 的 Bezos Expeditions。战略投资者包括 NVIDIA(显示 GPU 基础设施协同)、Toyota Ventures(工业 / 汽车应用)、Temasek(新加坡枢纽和主权财富)、Samsung(设备 / 消费电子)和 Sea。法国机构支持者包括 Bpifrance Digital Venture、Association Familiale Mulliez、Groupe Industriel Marcel Dassault 和 Publicis Groupe。知名天使投资人包括 Eric Schmidt、Mark Cuban、Jim Breyer、Tim 和 Rosemary Berners-Lee、Xavier Niel 以及 Mark Leslie。 LeBrun 明确表示,这笔资金投向两个成本中心:算力和人才,而不是收入生成。公司将在四个地点招聘时优先质量,而非数量。战略投资方构成(NVIDIA、Toyota、Samsung、Temasek)反映了公司有意押注工业和主权 AI 市场;这些市场里的政府和企业买家正在寻找非美国 hyperscaler 的 AI 替代方案。 [CO029, CO030, CO031, CO032, CO033, CO034]
| 利益相关方 | 角色 / 类别 | 经济或战略重要性 | 关键尽调问题 |
|---|---|---|---|
| Cathay Innovation | 联合领投方(VC) | 曾投资 LeBrun 的 Nabla;巴黎 / EU / 亚洲 VC,AUM 超 €2.5B;管理联合领投关系 | 出资规模、持股比例和董事席位条款 |
| Greycroft | 联合领投方(VC) | 美国风险投资机构,AMI 为其组合公司;提供美国市场入口 | 组合构成,以及除 AMI 承诺外的论点匹配 |
| Hiro Capital | 联合领投方(VC) | 欧洲 AI / 前沿模型论点投资人;明确引用世界模型契合 | 既有基金规模和未来轮次跟投能力 |
| HV Capital | 联合领投方(VC) | 德国欧洲风险投资机构;2000 年以来投资 200+ 组合公司 | 与 AMI 赛道重叠度和医疗 AI 敞口 |
| Bezos Expeditions | 联合领投方(个人 VC) | Jeff Bezos 个人资本;可信度信号;通常不占董事席位 | 是否存在任何优先权或 side-letter 条款 |
| NVIDIA | 战略投资人 | GPU 算力供应链对齐;潜在优先硬件获取 | 商业条款、算力 credits 或转售商关系 |
| Toyota Ventures | 战略投资人 | 工业过程控制和汽车机器人应用管线 | 是否计划在工厂或车辆 AI 上开展 proof-of-concept 合作 |
| Temasek | 主权财富基金(新加坡) | 新加坡枢纽入口;东南亚市场和主权 AI 政策对齐 | Temasek 是否有任何治理或报告契约 |
| Samsung | 战略投资人 | 可穿戴和消费设备应用;硬件集成潜力 | 是否有端侧世界模型部署路线图 |
| Bpifrance Digital Venture | 法国政府 VC | 主权 AI 背书;法国政治支持;未来 grant 的非稀释路径 | 任何共同投资条件、报告契约或 EU AI Act 合规承诺 |
| Meta(前雇主) | 潜在未来客户 / 竞争者 | LeCun 明确称 Meta 可能成为智能眼镜的首批客户;竞争性的 FAIR 实验室也在做相邻 AI | LeCun / LeBrun 与 Meta 之间的 IP 转让协议;如有,竞业限制条款 |
投资财团来自 AMI Labs 官方公告、Cathay Innovation 新闻稿和新闻报道。出资规模和股权比例未公开披露。Meta 行不是投资人; 但鉴于 LeCun 公开表示 Meta 可能成为 AMI 潜在客户,该行作为重要未来利益相关方纳入。
[CO032, CO033, CO034, CO035, CO043]1.5 里程碑、合作与轨迹
AMI Labs 的创立弧线异常压缩:从 LeCun 2025 年 11 月离开 Meta,到 2026 年 3 月完成 $1.03B 种子轮,只用了四个月。Nabla 合作与融资公告同步披露,是公司唯一公开确认的商业联盟。Nabla 将获得 AMI 世界模型的特权早期访问权,用于在高风险临床环境中测试。LeBrun 保留 Nabla 董事长兼首席 AI 科学家职务,维持两家公司之间的战略桥梁。 截至 2026 年 6 月运行日,AMI 仍处于延长研究阶段。公司尚未公开发布论文、基准测试或原型演示。LeBrun 在融资公告期间确认,有意义的商业产品可能需要数年才能出现。公司眼下的优先事项是研究基础设施、算力采购,以及在巴黎、纽约、蒙特利尔和新加坡组建团队。世界模型赛道已经吸引竞争者:World Labs(Fei-Fei Li)在 2026 年 2 月完成 $1B 融资,General Intuition 在 2025 年 10 月融资 $134M。LeBrun 自己预测,“world model”会成为下一个 AI 热词,许多公司会改名贴上这个标签。 [CO039, CO040, CO041, CO042, CO043, CO044]
| 日期 | 事件 | 类型 | 金额 / 估值 / 状态 | 参与方 | 含义 |
|---|---|---|---|---|---|
| Nov 2025 | Yann LeCun 在担任首席 AI 科学家 12 年后离开 Meta | 创立 | N/A | LeCun、Meta / Zuckerberg | 关键人物得以创立 AMI;JEPA 研究与 Meta FAIR 之间的 IP 边界必须厘清 |
| Dec 2025 | AMI Labs 通过 Nabla 新闻稿和 LeCun LinkedIn 发帖公开确认 | 创立 | N/A | LeCun(董事长)、LeBrun(CEO)、Nabla | 正式创立;LeBrun 同时从 Nabla CEO 转任 AMI CEO |
| Dec 2025 | FT 报道 AMI 寻求以 €3B 投前估值融资 €500M | 融资 | €500M / €3B 投前 | 未披露 VC 方、Cathay Innovation、Greycroft、Hiro Capital 被报道 | 投资人兴趣得到确认;融资目标在产品或团队完全组建前即已设定 |
| Jan 22–23, 2026 | AMI 网站上线,发布 JEPA 世界模型使命宣言 | 产品 | N/A | AMI Labs 团队 | 确立相对 LLM 的公开定位;网站成为主要官方来源 |
| Jan 2026 | 完整执行团队公开点名 | 治理 | N/A | LeBrun、Xie、Fung、Rabbat、Solly 核心团队 | C-suite 完整;确认 LeCun 和 Xie 的关键人物集中 |
| Feb–Mar 2026 | 投资财团横跨美国、EU、亚洲组建 | 融资 | $1.03B / $3.5B 投前 | Cathay、Greycroft、Hiro、HV、Bezos、NVIDIA、Toyota、Temasek、Samsung、15+ 其他方 | 多市场战略投资人基础释放主权 AI、工业和设备定位信号 |
| Mar 10, 2026 | $1.03B 种子轮宣布 | 融资 | $1.03B USD / $3.5B 投前 | 所有已披露投资人;LeBrun 和 LeCun 公开宣布 | 欧洲最大种子轮;资本市场验证世界模型论点 |
| Mar 10, 2026 | Nabla 被披露为首个商业合作伙伴 | 合作 | 商业条款未披露 | AMI Labs、Nabla | 医疗 AI 用例确认;Nabla 以真实世界评估数据换取特权模型访问 |
| Jun 2026(报告运行日) | 延长研究阶段;未发布论文或原型 | 产品 | N/A | AMI Labs 团队 | 产品前阶段按预期持续;按 CEO 表述,商业时间线仍需数年 |
时间线基于公开新闻报道和公司公告。精确法律成立日期和系列 / 股份类别细节未公开。未列监管、负面和规模事件,因为尚无报告; 这与公司产品前、收入前状态一致。
[CO002, CO029, CO030, CO031, CO036, CO037]从 LeCun 离开 Meta,到 2026 年 3 月种子轮及后续阶段的关键带日期里程碑。
时间线日期来自新闻报道;部分事件(2025 年 12 月 FT 报道、2026 年 1 月网站上线)可能有几天误差。
[CO002, CO029, CO038, CO039, CO040, CO043]1.6 展示材料
02市场分析
2.1 市场边界——物理世界 AI 基础设施
AMI Labs 位于一个尚无标准分析师定义市场的基础层:物理世界 AI 基础设施,也被称作具身智能、物理 AI,或基于世界模型的 AI。这个市场的核心特征,是要求 AI 系统能在连续、嘈杂、高维的真实环境中理解、预测并行动——工业厂房、临床场景、机器人系统和可穿戴设备——而不是处理离散文本或语言 token。 LeCun 明确用 Moravec Paradox 来界定市场边界:对人类来说轻而易举的事情(感知、导航、物理操控),对当前 AI 仍然难以计算。大型语言模型在架构上受限于离散 token 序列,无法可靠预测物理行动后果,也无法在多步操作序列中维持一致的世界状态表征。在 2026 年 1 月接受 MIT Technology Review 采访时,LeCun 点出了四个具体应用领域:工业流程控制(喷气发动机、钢厂、化工厂)、工厂自动化和机器人、医疗诊断与临床工作流,以及智能或可穿戴设备。 AMI 公司网站、Cathay Innovation 的投资者沟通,以及媒体报道,都持续确认这四个细分市场就是公司声明的市场边界。NVIDIA 也把 “Physical AI” 并行定位成一个独立基础设施类别——投资 Isaac 仿真、Cosmos 世界模型训练和机器人推理基础设施——这从外部佐证了该市场边界的连贯性,也说明资本密集型玩家认可这一类别。 AMI 核心市场不包括:纯文本和以语言为中心的应用(LLM、聊天机器人、文档摘要)、生成式图像和视频创作(扩散模型),以及基于结构化表格数据的传统 ML 分类。边界不是地理性的:AMI 的巴黎总部和多枢纽结构(纽约、蒙特利尔、新加坡),叠加欧洲和亚洲投资方联盟,传递的是全球市场野心,并对欧洲 AI 主权要求特别敏感。 AMI 目标技术的现状替代品包括传统 SCADA 和 PLC 工业控制系统、基于规则的临床决策支持软件、基于遥操作的机器人系统、专用计算机视觉模型,以及把行动规划外挂到语言模型上的 LLM agent 框架。每类替代品在可泛化物理推理上都有记录在案的限制——世界模型正是为弥合这一缺口而设计。Moravec Paradox 框架是 LeCun 自己的解释镜头,代表公司声称的市场逻辑,而不是独立验证过的市场边界研究。 [CM001, CM002, CM003, CM004, CM015, CM016]
| 细分 / 类别 | 纳入支出 | 排除支出 | 主要买方 / 付款方 | 与 AMI 的相关性 |
|---|---|---|---|---|
| 工业过程控制 | 用于工厂监控、预测性维护、连续流程(化工、能源、航空航天)自主控制的 AI | 非 AI 自动化、传统 SCADA/PLC、旧式 ERP/MES 软件 | 制造业 CTO / CapEx 预算(3–5 年周期) | 核心目标;LeCun 点名提到喷气发动机、钢厂、化工厂 |
| 工厂自动化与工业机器人 | AI 赋能工业机器人、自主移动机器人、AI 增强操控和装配 | 非智能机器人、没有 AI 泛化能力的遥操作系统 | 运营总监 / CapEx 预算 | 核心目标;Toyota Ventures 投资释放汽车和工业机器人兴趣 |
| 医疗诊断与临床 AI | 临床决策支持、医学影像分析、AI 辅助病历记录、agentic 临床工作流 AI | 收入周期管理、行政 SaaS、纯 EHR 软件 | 医院 CIO / OpEx + 保险报销;FDA/EMA 作为监管闸门 | 通过 Nabla 合作形成近期主要锚点;LeBrun 具备领域经验 |
| 可穿戴 / 智能设备 | 嵌入智能眼镜、AR/VR 头显、具备实时环境推理的可穿戴健康监测设备的 AI | 没有主动 AI 推理的消费 IoT、传统纯传感器应用 | 消费电子 R&D 预算 / 首席产品官 | LeCun 描述的视觉用例;Samsung 投资释放设备雄心 |
| 自动驾驶汽车(相邻) | 需要理解物理世界的自动驾驶感知和规划栈 | ADAS 规则系统、非自动巡航控制、传统地图 | OEM AI R&D 部门 | 相邻;LeCun 将 Level 5 自动驾驶列为 JEPA 用例,但这不是 AMI 宣布的主要细分市场 |
| 文本 / 语言 / 生成式 AI | N/A — 明确排除 | LLM、聊天机器人、代码生成、图像生成、文档 AI、NLP 应用 | N/A | 明确排除;LeCun 称其为竞争者领域,也是理解物理世界的死胡同 |
细分定义基于 LeCun 访谈(MIT Technology Review 2026 年 1 月)和 AMI 公开沟通。投资人构成提供了相互印证的行业信号。 支出边界为分析师推断,并非公司披露。
[CM001, CM002, CM003, CM004]2.2 市场规模——TAM、SAM 与证据受限视角
还没有分析师把 “world models” 或 “physical-world AI infrastructure” 定义为独立商业类别,因此 TAM 测算只能框定相邻市场,再施加相关性筛选。 广义 TAM——物理世界 AI 可触达市场(约 $258–610B,2026 年估计):三个独立分析师锚点勾勒上限。Grand View Research 将 2025 年总 AI 市场估为 $390.9B,预计 2033 年达到 $3,497.3B,CAGR 为 30.6%。MarketsandMarkets 将 2026 年市场估为 $601.9B,预计 2033 年达到 $3,638.1B,CAGR 为 29.3%。两组数字都涵盖包括文本和生成式 AI 在内的所有 AI 垂直领域;AMI 可触达部分只是其中子集。与 AMI 声明领域最相关的三个子细分是:(1)医疗 AI——2026 年 $36.7B,预计 2031 年达到 $194.8B,CAGR 39.7%(MarketsandMarkets);(2)工业自动化——2024 年 $206.3B,预计 2030 年达到 $378.6B,CAGR 10.8%(Grand View Research);以及(3)机器人 AI 软件——2025 年 $21.0B,预计 2032 年达到 $70.8B,CAGR 19.0%(MarketsandMarkets)。 SAM——世界模型基础设施授权(无独立规模测算):AMI 声明的收入模式是向行业伙伴授权技术,而不是直接做垂直部署。基础设施层授权业务会从上述垂直市场中抽取一部分价值——历史上,基础设施层通常捕获其赋能应用层价值的约 3–15%。对 $258B 的合并物理 AI 子细分(医疗 AI + 机器人 AI 软件,2026 年中估计)施加 5–10% 捕获率,得到 2031 年视角下粗略 SAM 为 $13–26B。鉴于 AMI 研究优先的时间线且没有任何可部署产品,不确定性很宽。这是分析师推导;没有公开来源独立验证捕获率或子细分构成。 SOM——2026–2028 可服务可获得市场(实际为零):AMI 是研究优先公司,当前没有产品。CEO LeBrun 在 2026 年 1 月表示,商业推出可能 “约一年后” 开始,乐观情况下意味着最早商业收入在 2027 年。短期 SOM 受 Nabla 医疗合作约束,只是单一渠道客户,而非大众市场。$3.5B 投前种子轮估值说明,投资者主要是在为科学可信度和物理世界 AI 类别的长期期权价值付费,而不是近期收入。 一个重要的相反规模信号:工业自动化——按价值计最大的物理世界细分——CAGR 为 10.8%,显著低于 AI 特定市场细分 29–40% 的 CAGR 区间。这种分化反映了工业部署资本密集、采纳保守的性质。投资者如果把头部 AI 增长率外推到 AMI 的物理世界市场,很可能高估近期可触达需求。 [CM005, CM006, CM007, CM008, CM009, CM010]
| 发布方 | 发布年份 | 地域 | 类别 | 基准值 | 预测值 | CAGR | 方法论 | 信心 | 对 AMI 的限制 |
|---|---|---|---|---|---|---|---|---|---|
| Grand View Research | 2026 | 全球 | 整体 AI 市场 | $390.9B (2025) | $3,497.3B (2033) | 30.6% | 自下而上调查 + 二手研究 | 中 | 覆盖包括文本和语言在内的所有 AI;大幅高估 AMI 可服务市场 |
| MarketsandMarkets | 2026 | 全球 | 整体 AI 市场 | $601.9B (2026) | $3,638.1B (2033) | 29.3% | 一手和二手研究;硬件、软件、服务分层 | 中 | 同样受广义市场口径限制;Mistral AI 与 NVIDIA 和 Microsoft 一同列为关键玩家 |
| MarketsandMarkets | 2026 | 全球 | 医疗 AI | $36.7B (2026) | $194.8B (2031) | 39.7% | 按功能拆分:影像、机器人、AI scribe、CDS、精准医疗 | 高 | 包含基于 LLM 的医疗 AI;世界模型份额是未定义子细分 |
| MarketsandMarkets | 2026 | 全球 | 机器人 AI 软件 | $21.0B (2025) | $70.8B (2032) | 19.0% | 机器人平台、中间件、分析和 AI 框架 | 中 | 最接近世界模型基础设施市场的 proxy;但仍包含非 JEPA 和非世界模型路径 |
| Grand View Research | 2026 | 全球 | 工业自动化 | $206.3B (2024) | $378.6B (2030) | 10.8% | 按收入口径;包括 DCS、SCADA、PLC、智能工厂集成 | 高 | 覆盖非 AI 自动化;未单列 AI 子板块;较慢 CAGR 显示企业仍偏保守 |
| IFR | 2026 | 美国 | 工业机器人安装量 | 38,000 units (2025) | N/A | +11% YoY | 机器人出货量年度普查;行业协会方法论 | 高 | 这是数量指标,不是价值指标;未捕捉 AI 软件层价值或世界模型授权机会 |
| IFR | 2026 | 中国 | 工业机器人安装量 | 295,000 台(2024) | N/A | ~54% 全球份额 | 年度普查;2025 年初步估计尚未发布 | 高 | 中国机器人规模约为美国 10 倍,说明物理 AI 市场足够大,但也带来竞争性的市场格局 |
| 分析师推断(本报告) | 2026 | 全球 | 世界模型基础设施 SAM(估计) | N/A | $13–26B(2031 年估计) | N/A | 对医疗 AI 与机器人 AI 软件子板块应用 5–10% 基础设施捕获率 | 低 | 没有已发布的分析师报告把世界模型授权作为独立类别测算;基础设施捕获假设高度推测 |
已发布估计覆盖的是广义 AI 市场或细分垂直领域,并非专门覆盖世界模型。SAM 行由本报告分析师推算,采用基础设施层捕获率;不存在独立市场规模测算。所有金额均为 USD。信心评级反映来源可靠性,而非与 AMI 具体市场的相关度。
[CM005, CM006, CM007, CM008, CM009, CM010]AMI 物理世界 AI 基础设施业务的三层市场规模测算,从宽口径 AI TAM,到推测性世界模型 SAM,再到当前接近零的 SOM。
SAM 和 SOM 数值由分析师基于基础设施捕获率假设推断,并非已发布市场研究。宽口径 TAM 反映总体 AI 市场(包含与 AMI 无关的 LLM 和生成式 AI 细分)。物理 AI 子 TAM 排除工业自动化(2024 年 $206B),原因是 CAGR 较低且采用周期慢。
[CM005, CM006, CM007, CM038, CM039]与 AMI 相关的三个物理世界 AI 子领域分析师预测;低 / 基准 / 高边界反映市场定义和采用速度的不确定性。
所有数值单位均为十亿美元。医疗 AI 基准来自 MarketsandMarkets(到 2031 年 $194.8B,CAGR 39.7%);不确定性边界 ±33%。机器人 AI 软件基准来自 MarketsandMarkets(到 2032 年 $70.8B,CAGR 19.0%);边界 ±30%。工业自动化 AI 子领域由分析师估算,约占 $379B 工业自动化总市场(Grand View Research 2030)的 15–25%;不确定性高。世界模型 SAM 基于医疗 AI + 机器人子市场预测应用 5–10% 基础设施捕获率计算;没有独立来源。
[CM006, CM007, CM009, CM038]2.3 买家、用户与付款方分层
AMI 的授权模式意味着,公司的直接买家是构建或部署 AI 赋能系统的企业,而不是这些系统的终端用户。这种 B2B 基础设施动态形成两层市场结构:AMI 必须同时赢得系统构建者的技术信任,以及运营导向企业买家的商业采纳。 工业自动化(到 2030 年最大的可触达细分):买家是化工、汽车、能源和航空航天等大型工业企业的制造 CTO、工厂运营总监和自动化工程师。付款方是资本开支预算,周期为每 3–5 年。采纳由劳动力成本上升、工厂现代化计划,以及中国制造商机器人密度优势加速带来的竞争压力共同触发。IFR 数据确认,2025 年美国机器人安装量增长 11%,达到 38,000 台,其中食品行业采纳激增 30%——说明自动化需求正在从传统汽车行业向外扩散。2024 年中国年度安装量约 295,000 台(全球份额 54%),进一步加大西方制造商加速 AI 赋能自动化的竞争压力。 医疗(考虑 Nabla 合作,近期信号最强):买家是医疗系统 CIO 和临床信息学团队。付款方是医院运营预算,并最终取决于医保报销框架。Forbes 报道称,LeBrun 特别提到当前 AI 在临床场景的限制,将其作为 AMI 的切入点——世界模型能够提供比 LLM 更安全、更可靠的临床决策支持。FDA 和 EMA 审批限制面向患者的应用;Nabla 合作看起来被定位为工作流工具(AI 病历记录),而不是 FDA 监管的临床设备,这可能加快初始商业时间线。 机器人 OEM(平台 / 渠道买家):AMI 可以把世界模型 API 授权给机器人硬件制造商。Toyota Ventures 参与种子轮,传递了汽车和工业机器人兴趣。LeCun 的明确表述是,当前机器人只适用于特定任务,无法泛化——这正是世界模型要填补的缺口。该细分的战略买家是机器人 OEM 的 AI / 软件 VP,拥有产品开发预算和多年硬件周期。LeCun 承认,当前人形机器人通过遥操作训练,无法跨环境泛化——这一说法既描述了市场问题,也抬高了商业就绪速度的不确定性。 可穿戴 / 设备 OEM(最长前置期):Samsung 的战略投资显示,可穿戴设备和智能眼镜是潜在应用领域。LeCun 曾把能够 “观察你在做什么、识别你的动作,并预测你接下来要做什么” 的智能眼镜描述为世界模型用例。预算归属消费电子 OEM 的 R&D 部门,硬件设计周期为 2–3 年。 投资者构成提供了最清晰的分层信号:NVIDIA(物理 AI 基础设施)、Toyota Ventures(汽车和工业机器人)、Samsung(设备和可穿戴)、Temasek(亚太主权与企业资本)都是战略投资者,而不是财务投资者;每家都指向一个世界模型技术可能嵌入的细分市场。 [CM019, CM020, CM021, CM022, CM023, CM024]
| 细分市场 | 买方 | 用户 | 付费方 | 受影响工作流 | 预算负责人 | 采用触发因素 |
|---|---|---|---|---|---|---|
| 工业过程控制 | 制造业 CTO / 数字化转型负责人 | 工厂运营工程师、过程控制团队 | CapEx 预算(3–5 年周期) | 预测性维护、过程异常检测、自主过程优化 | 工厂 VP / 运营总监 | 劳动力成本上升、工厂现代化、中国机器人竞争压力 |
| 医疗健康——临床 AI | 医疗系统 CIO / 临床信息学负责人 | 临床医生、护士、医务人员 | 医院 OpEx + 保险报销(AI 临床工具的付费覆盖正在出现) | 临床文档、诊断辅助、照护路径优化、AI 病历记录 | CFO + CMO / 医院管理层 | 临床劳动力短缺、监管质量要求、已证明的病历记录 ROI |
| 工业机器人 / 自动化 OEM | 机器人 OEM(工程 VP 或 AI/软件负责人) | 工厂一线操作员、机器人集成工程师、终端制造商 | OEM R&D 预算(嵌入式产品成本)+ 终端用户 CapEx | 机器人感知、多任务泛化、自适应操作、自主导航 | 机器人 OEM 的工程 VP / CTO | 对灵活多任务自动化的需求;中国人形机器人竞争;客户要求任务通用性 |
| 可穿戴 / 智能设备 OEM | 消费电子 OEM(首席产品官或 R&D VP) | 终端消费者、企业现场作业人员、临床人员 | OEM 产品开发预算;现场作业应用的企业部署预算 | 情境辅助、活动预测、AR 叠加、可穿戴健康监测 | 首席产品官 / R&D VP | 硬件设计周期拐点、消费 AI 差异化需求、企业 AR 用例发展 |
| 自动驾驶汽车(相邻) | Tier-1 汽车供应商或 OEM AI 团队负责人 | 安全工程师、测试驾驶员、车队运营商 | OEM AI R&D 预算 + 软件授权 | 面向 L4/L5 自动驾驶的感知、世界状态预测、路径规划 | 自动驾驶 VP / CTO | 监管推动 L4/L5 自动驾驶认证;中国 OEM 带来的 AV 市场竞争格局 |
买方和付费方结构基于 LeCun 访谈(MIT Technology Review,2026 年)、Forbes 医疗健康分析、Nabla 合作结构和投资方构成信号。预算周期估计反映典型企业技术采购模式,不是 AMI 的具体披露。FDA/EMA 批准要求未纳入本表,但它是医疗健康细分市场收入的前提。
[CM019, CM021, CM022, CM023, CM024, CM025]映射 AMI 声明的四个市场细分中的买方类型、用户、付款方和主要采用触发因素;也纳入横跨所有细分的机器人能力矛盾。
矩阵基于 LeCun 访谈、Forbes 医疗分析、Nabla 新闻稿和战略投资人构成。收入兑现时间是分析师根据典型企业技术采用周期推断,并非 AMI 披露的预测。
[CM019, CM021, CM022, CM023, CM024, CM025]2.4 增长驱动因素与采纳约束
增长驱动因素: 劳动力套利与自动化需求:制造、医疗和物流的全球劳动力短缺构成结构性驱动。IFR 确认,2025 年美国机器人安装量增长 11%,食品行业采纳激增 30%。中国第十五个五年规划(2026–2030)把机器人置于现代工业体系核心,明确将 AI 研究聚焦在物理应用,并把机器人指定为经济增长的主要驱动力。这迫使西方制造商和政府投资物理 AI 基础设施,以匹配中国的部署规模。 物理 AI 基础设施投资:NVIDIA 已在 GTC 2026 验证 “Physical AI” 是一个独立基础设施类别。NVIDIA 的 Isaac 仿真平台、Cosmos 世界模型训练基础设施和机器人推理栈,代表着 hyperscaler 级别的投入,建设 AMI 目标市场所需的赋能基础设施。Futurum Group 分析师指出,NVIDIA 在 AMI 的投资方身份传递的是协同,而非竞争——AMI 的基础模型层可能补充 NVIDIA 的算力基础设施,而不是与之竞争。 欧洲 AI 主权需求:欧洲政府、主权财富基金和企业买家正在积极寻找不经由美国 hyperscaler 供应链、也不把敏感数据暴露在美国云管辖下的 AI 基础设施。Bpifrance 参与 AMI 种子轮,反映了法国国家资本对本土 AI 能力的投入。LeCun 明确把 AMI 描述为 “既不是美国,也不是中国”,直接回应主权 AI 需求,并创造出超越纯技术优劣的结构性市场偏好。法国国家 AI 生态和 EU AI Act 合规要求,也为一家总部位于欧洲的前沿 AI 实验室提供进一步顺风。 LLM 失败模式意识:在高风险物理环境中,LLM 的失败模式已有记录——幻觉、无法建模行动后果、时间推理不可靠——这些问题正在催生企业对更可靠 AI 架构的需求。SiliconAngle 报道和 Futurum Group 分析都指出,工业和医疗买家非常清楚 LLM 可靠性边界。 采纳约束: 世界模型技术尚未商业验证:作为商业产品类别,世界模型还不存在。Futurum Group 明确指出 “研究优先使命与按十亿美元融资校准的投资者预期之间存在结构性张力”。AMI 押注基于 JEPA 的架构会比替代路径更快在物理任务上达到商业级性能,但没有外部验证支持这一时间线。 监管路径:面向患者的临床 AI 需要 FDA 和 EMA 审批,上市前审查以年计。自主系统的工业安全认证(IEC 61511、ISO 13849)需要大量验证。医疗和工业控制高风险应用的 EU AI Act 合规要求,在 AMI 主要欧洲目标市场进一步增加监管复杂度。 企业采纳保守:工业和医疗企业买家资本密集,规划周期多年,从既有平台切换成本高(工业领域 Siemens、GE、Honeywell;医疗领域 Epic、Cerner)。Grand View Research 的工业自动化数据确认,即便纳入 AI 集成,该行业 CAGR 仍为 10.8%——远低于头部 AI 增速——反映了这种保守性。资本开支周期平均为 3–5 年。 竞争性物理 AI 架构:Google DeepMind、Meta Superintelligence Labs、Physical Intelligence 和 1X 都在开发相邻物理 AI 能力。LLM 提供商正在投入推理和行动能力,可能向世界模型用例收敛。LeCun 承认,如果主要 LLM 提供商加快世界模型和空间学习工作,AMI 的架构差异窗口会变窄。没有公开基准测试独立比较基于 JEPA 的世界模型与竞争路径在物理世界任务上的表现。 机器人泛化缺口:LeCun 在 2026 年 1 月公开表示,“没人,绝对没人,知道怎样让这些机器人聪明到真正有用”,并把原因归结为依赖遥操作的训练数据无法跨环境泛化。这个坦率承认一方面勾勒了世界模型的市场机会,另一方面也显示距离商业部署仍有很长技术距离。 [CM026, CM027, CM028, CM029, CM031, CM032]
| 因素 | 类型 | 方向 | 时间 | 对 AMI 的含义 | 尽调问题 |
|---|---|---|---|---|---|
| 劳动力短缺推动自动化投资 | 驱动因素 | 正向 | 当前 / 2–5 年 | 为通用 AI 自动化形成结构性拉力;利好工业和机器人论点 | 量化 AMI 具体目标行业的劳动力成本与自动化 ROI;与既有自动化成本比较 |
| 中国“十五五”规划——机器人优先 | 驱动因素 | 正向(对西方形成竞争压力) | 2026–2030 | 加速全球机器人部署,也倒逼西方加大物理 AI 投资 | 判断竞争压力究竟转化为 AMI 合作需求,还是只会加快中国侧竞争者的物理 AI 开发 |
| 欧洲 AI 主权需求 | 驱动因素 | 正向 | 当前 / 3–7 年 | AMI 的欧洲总部和投资方组合,使其在符合欧盟要求、非美国 / 非中国 AI 基础设施上具备结构性偏好 | 验证买方意向:有多少欧洲企业潜在客户明确偏好非美国 AI 基础设施? |
| NVIDIA 物理 AI 基础设施投资 | 驱动因素 | 正向 | 当前 / 2–4 年 | 在超大规模云厂商量级验证物理 AI 类别;NVIDIA 的 Cosmos 平台可能与 AMI 互补 | 厘清 AMI 与 NVIDIA 自身世界模型工作的关系;确认合作条款与潜在竞争重叠 |
| 企业对 LLM 失效模式的认知 | 驱动因素 | 正向 | 当前 | 企业越来越清楚 LLM 会幻觉、在物理世界不可靠,从而拉动对更可靠架构的需求 | 调研工业和医疗 CIO 对 LLM 可靠性的看法,以及尝试替代 AI 路径的意愿 |
| 世界模型技术商业化尚未验证 | 约束 | 负向 | 近期(2–4 年) | 尚无可部署产品;多年研究时间线让 $1.03B 融资对应的执行风险更高 | 跟踪 JEPA 基准论文;确认首批商业试点时间线和试点转化率 |
| FDA 与 EMA 监管路径 | 约束 | 负向 | 中期(3–7 年) | 面向患者的临床 AI 需要以年计的审批流程,限制近期医疗收入 | 确认 Nabla 合作范围:AI 工作流工具还是受 FDA 监管的临床设备;厘清监管策略 |
| EU AI Act 高风险分类 | 约束 | 负向 | 中期(2–5 年) | EU AI Act 将工业控制和医疗 AI 归入高风险类别;合规成本不可忽视 | 评估 AMI 的监管事务能力;确认主要欧洲市场的合规路线图 |
| 工业 CapEx 保守(3–5 年周期) | 约束 | 负向 | 持续 | 规划周期长,销售难以快速推进;AMI 早期收入不能依赖企业快速采用 | 跟踪工业客户中的概念验证管线规模,以及企业转试点的时间线 |
| 竞争性物理 AI 架构 | 约束 | 负向 | 近期至中期 | Google DeepMind、Physical Intelligence、1X 以及 NVIDIA 自身世界模型,可能收敛到 AMI 的目标用例 | 监控 LLM 供应商的物理 AI 路线图;获取 JEPA 与竞争架构的独立基准 |
时间估计基于行业周期和监管基准,属指示性判断,不是 AMI 的具体预测。尽调问题是下一步研究问题,不是已验证发现。EU AI Act 高风险分类适用于关键基础设施、医疗健康和自主设备中的 AI 系统。
[CM026, CM027, CM028, CM029, CM030, CM031]从 AMI 当前研究阶段到商业授权和企业部署的分阶段采用路径,并列出每一阶段的关口条件。
漏斗数值代表分析师估算的初始研究池转化率,并非 AMI 披露的管线数据。AMI 仍处于产品前阶段,没有商业管线数据。漏斗用于说明结构性采用壁垒,不是量化收入预测。
[CM032, CM033, CM034, CM036]2.5 规模矛盾、证据缺口与尽调优先级
本章保留三个仍未解决的重大矛盾。 矛盾 1——AI 增速与物理世界部署速度:头部 AI 市场 CAGR 为 29–31%(MarketsandMarkets、Grand View Research),显著高于工业自动化市场 10.8% 的 CAGR;后者是 AMI 物理世界范围内最大的单一细分。差异源于文本、语言和生成式 AI 在整体 AI 市场规模测算中占主导。投资者在研究第一天就给 AMI $3.5B 投前估值,可能是在外推头部 AI 增速;但这些增速并不代表物理世界 AI 采纳速度,尤其考虑工业和医疗买家的资本开支保守性。 矛盾 2——机器人商业就绪说法:LeCun 2026 年 1 月称 “没人,绝对没人,知道怎样让这些机器人聪明到真正有用”,这与 Agility Robotics、Figure AI、1X 和 Boston Dynamics 的商业部署说法直接冲突;这些公司都描述了近期通用人形机器人商业化。关于当前机器人泛化能力,这两种判断不可能同时准确。矛盾仍未解决,也给任何物理 AI 基础设施提供商带来显著市场时机不确定性。 矛盾 3——世界模型差异性与 LLM 收敛:AMI 的论点要求 LLM 在结构上无法拿下物理世界 AI 市场。然而,Google DeepMind 的 Gemini Robotics、NVIDIA 的 Project GR00T,以及 OpenAI 的物理智能体工作,都试图用多模态基础模型处理类似物理世界任务。基于 JEPA 的世界模型是否在物理世界任务上架构更优,还是各路径最终收敛,是一个关键经验问题,目前没有公开基准测试给出答案。 关键证据缺口——没有世界模型授权的分析师 SAM:没有公开分析师报告把世界模型基础设施授权市场作为独立类别测算规模。所有市场估计都针对更宽类别,里面包含非世界模型应用。本章引用的 $13–26B SAM 是基于基础设施捕获率推导,并非来自任何公开研究。这是投资尽调中最重大的规模缺口。 [CM038, CM039, CM040, CM041]
2.6 展示材料
03竞争对手
3.1 竞争格局概览
AMI Labs 的竞争格局横跨四个不同竞争簇:(1)瞄准 3D 空间和仿真输出的直接世界模型创业公司;(2)在大型企业平台中嵌入世界模型项目的前沿研究实验室;(3)把物理世界 AI 商业化到机器人硬件上的具身 AI 和机器人创业公司;以及(4)把世界基础模型作为平台层提供的基础设施玩家。 AMI 最清晰的直接竞争对手是 World Labs(Fei-Fei Li 的空间智能创业公司)、Odyssey(通用世界模型实验室)和 SpAItial(3D Gaussian Splat 世界模型)。三家公司都是私营融资创业公司,追求用于生成式 3D 环境的世界模型架构。不同于 AMI,三者都采用生成式架构——预测像素级或体素级内容,而不是抽象表征——且已经部署公开产品或 API。 前沿实验室簇包括 Google DeepMind(Genie 1/2/3、Gemini Robotics)、Meta FAIR(V-JEPA,LeCun 离开后)和 OpenAI(Sora,已于 2026 年 4 月停用)。NVIDIA 位置特殊:其 Cosmos 世界基础模型平台采用开放授权,已通过机器人和 AV 伙伴嵌入物理 AI 开发者生态。Physical Intelligence 和 Wayve 代表具身 AI 商业化簇,把世界模型和基础策略直接部署到机器人硬件和自动驾驶车辆中。 AMI 基于 JEPA 的路径占据一条技术上不同的赛道——非生成式、可控、可靠性优先——当前没有已商业部署的竞争对手直接瞄准其声明的工业流程控制和医疗垂直领域。战略风险在执行速度:AMI 承诺多年产品前时间线,而竞争者已经在交付。 [CP001, CP005, CP006, CP009, CP010, CP012]
| 竞争者 | 类别 | 估计融资 | 目标细分市场 | 核心技术 / 产品 | 关键差异化 | 相对 AMI 的限制 |
|---|---|---|---|---|---|---|
| World Labs | 直接竞争初创公司 | ~$230M+(估计) | 3D 空间 / 创意 AI | Marble 生成式 3D 世界模型;World API(2026 年 1 月) | 已部署商业产品;Fei-Fei Li 背书;空间分类法领先 | 生成式(非 JEPA);偏创意 / 空间,不是工业 / 医疗 |
| Odyssey | 直接竞争初创公司 | <$50M(估计) | 研究 / 游戏模拟 | Odyssey-2(世界模型)、Starchild-1、Agora-1(多智能体)、PROWL(RL) | 多智能体世界;用 RL 对抗框架提升准确性 | 阶段更早;商业足迹更窄;生成式架构 |
| SpAItial | 直接竞争初创公司 | 未披露 | 开发者 / 3D 应用 | Echo(3D Gaussian Splat 世界模型);Echo-2;开发者 API | 开发者优先的 3D API;明确支持 3D 导出格式;Gaussian Splat 差异化 | 2025 年 5 月成立;非常早期;领域限于 3D 内容生成 |
| Google DeepMind | 前沿实验室 | Google 支持 | 研究 + 具身 AI | Genie 1/2/3(世界模型);Gemini Robotics;SIMA 2 智能体 | 算力不受限;支持多形态机器人;SIMA 智能体研究 | 研究目标(非商业);GTM 不同;不瞄准工业 / 医疗 |
| Meta FAIR | 前沿实验室 | Meta 支持 | 研究 | V-JEPA(开放 CC-NC);I-JEPA 开源;Muse Spark(新) | 原始 JEPA 架构发源方;开源代码可用 | LeCun 已离开;战略重心转离 JEPA;没有工业 / 医疗垂直 |
| Physical Intelligence | 具身 AI 初创公司 | ~$470M+(估计) | 机器人 / 硬件 OEM | π0.7 VLA 策略;跨形态;组合式泛化;伙伴计划 | 已与商业伙伴部署;π0.7 带来泛化能力跃迁(2026 年 4 月) | VLA 本身不算世界模型;仅限机器人;不瞄准医疗或工业过程控制 |
| OpenAI | 前沿实验室 | 融资 >$6B | 消费 / 企业 LLM | Sora(2026 年 4 月 26 日停用) | 庞大用户基础;LLM 集成;算力规模 | Sora 已停用;截至 2026 年中没有活跃世界模型产品 |
| NVIDIA | 平台 / 基础设施 | 上市公司 | 物理 AI / AV / 机器人 | Cosmos WFMs(开放许可证);Isaac 平台;GEAR 和 Spatial Intelligence 实验室 | 开放商业许可证;9,000T token 训练;庞大伙伴生态(1X、Agility、Waabi、Uber) | 基础设施层,不是世界模型研究;GPU 中心激励;不瞄准医疗 |
| Wayve | 具身 AI(AV) | ~$1B+(估计) | 自动驾驶 OEM | AV2.0 端到端驾驶 AI;无地图;车型无关;自监督 | 强 OEM 合作;已验证真实世界 AV 部署;自监督规模 | 领域专用(仅驾驶);不在工业或医疗垂直 |
融资数字来自公开报道和公司声明的估计;未披露轮次和估值可能存在重大差异。Physical Intelligence 估计合计了公开二手报道中的 $70M 种子轮(2024 年 5 月)和 $400M Series B(2024 年 10 月)。World Labs 估计来自二手媒体报道,未由一手文件确认。限制列相对于 AMI 所述使命(工业 / 医疗世界模型),不是一般性批评。
[CP001, CP005, CP006, CP009, CP010, CP012]AMI Labs 独处非生成式、研究阶段象限;所有已商业部署的竞争者都使用生成式或 VLA 架构。OpenAI Sora(已停止)退出活跃空间。 Physical Intelligence 和 NVIDIA Cosmos 在部署成熟度上领先。
X 轴:1 = 完全生成式(像素 / 体素重建),10 = 非生成式 / 预测式(抽象表征)。Y 轴:0 = 已停止 / 无产品,10 = 已商业部署并产生收入。所有位置均为有证据支撑的序数估计;没有数值来源。
[CP005, CP026, CP037, CP040]3.2 直接世界模型创业同业:World Labs、Odyssey、SpAItial
Fei-Fei Li 创立的 World Labs,是 AMI 最突出的直接创业竞争对手。它的论点——“空间智能” 是生成式 AI 的下一前沿——与 AMI 的物理世界 AI 野心相互映照,但在架构和应用上明显分化。World Labs 的 Marble 产品可从文本、图像、视频和 360 全景生成空间一致、持久的 3D 世界。2026 年 1 月推出的 World API 为开发者构建仿真、创意工具和空间应用提供程序化访问。World Labs 于 2026 年 6 月发布世界模型功能分类,把模型分为 Renderers、Simulators 和 Planners——这个框架隐含地把 AMI 定位为 Simulator / Planner 类型,而 World Labs 领先 Renderer 细分。World Labs 已有商业产品和开发者采纳,截至 2026 年中拥有 AMI 尚不具备的先发优势。 Odyssey 是规模较小的世界模型实验室,但野心更宽:Odyssey-2 模型瞄准通用物理准确性,Starchild-1 把从视觉观察中学习扩展到更丰富的多模态交互,Agora-1 支持多智能体共享世界,PROWL 框架用对抗式 RL 改善世界模型表现。Odyssey 的产品套件比 World Labs 面向消费者的 Marble 更贴近研究,但其以 RL 驱动提升物理准确性的路径,与 AMI 的准确性优先定位高度重叠。 SpAItial 成立于 2025 年 5 月,是直接同业中最年轻、规模最小的一家。其 Echo 模型家族通过开发者 API 从图像、文本或全景生成持久 3D Gaussian Splat 世界。Echo-2 已公布。SpAItial 的开发者优先定位(API + SPZ、PLY、SOG 等 3D 格式导出)以及对 3D 世界生成的窄聚焦,代表了不同于 AMI 计划工业授权模式的商业化策略。 三家同业与 AMI 之间的关键架构分野在于:World Labs、Odyssey 和 SpAItial 都是生成式系统——它们重建像素级或体素级内容,用算力效率换取照片级真实输出。AMI 的 JEPA 路径在抽象表征空间中预测,创始人认为这能带来可控性和可靠性,而这些特性是生成式系统在安全关键环境中结构上无法匹配的。 [CP001, CP002, CP003, CP004, CP005, CP006]
| 能力维度 | AMI Labs | World Labs | Physical Intelligence | Google DeepMind | NVIDIA Cosmos | Meta FAIR |
|---|---|---|---|---|---|---|
| 已部署商业产品(2026 年中) | 否(研究) | 是(Marble + API) | 是(π0.7 + 伙伴) | 仅研究 | 是(开放模型) | 仅开源 |
| 非生成式 JEPA 架构 | 是(核心) | 否(生成式 3D) | 部分(VLA 混合) | 否(自回归) | 否(扩散 + AR) | 是(V-JEPA,开放) |
| 面向工业过程控制 | 计划中 | 否 | 否 | 否 | 部分(Cosmos) | 否 |
| 医疗 / 临床应用 | 计划中(Nabla) | 否 | 否 | 否 | 否 | 否 |
| 开源模型 / 代码 | 否(研究优先) | 部分(API) | 是(π0 于 2025 年 2 月开放) | 部分(Genie 代码) | 是(开放许可证) | 是(CC-NC) |
| 多形态机器人控制 | 计划中 | 否 | 是(8+ 平台) | 是(Gemini Robotics) | 是(Isaac + Cosmos) | 否 |
| 语言条件动作输出 | 计划中 | 否(仅内容) | 是 | 是(Gemini Robotics) | 是 | 部分(V-JEPA) |
| 开发者 API / SDK | 否 | 是(World API) | 部分 | 否 | 是(Isaac SDK + API) | 否(研究) |
单元格反映截至 2026 年 6 月的公开信息;“计划中”表示公司公开路线图,尚未确认部署。“未知”用于无公开信息处。Physical Intelligence 的 VLA 架构在 π0.7 中包含用于子目标生成的世界模型组件,但其主要价值是直接机器人控制,而不是作为基础设施层的世界建模。Meta FAIR 的 V-JEPA 是非生成式,但以 CC-NC(非商业)发布;其商业适用性不同于 AMI 授权的世界模型。
[CP001, CP002, CP003, CP005, CP016, CP017]AMI Labs 在非生成式 JEPA 以及医疗 / 工业目标上位置独特,但截至 2026 年中,在已部署产品、开源可用性和开发者 API 上落后于所有活跃竞争者。
“部分”表示有限或 beta 可用。“计划”反映公司公开路线图,截至 2026 年 6 月尚无确认部署。单元格仅反映公开可得信息。
[CP002, CP005, CP019, CP021, CP023, CP028]3.3 前沿实验室与 Hyperscaler:DeepMind、Meta FAIR、OpenAI 和 NVIDIA
在前沿实验室中,Google DeepMind 的世界模型组合最完整。Genie 1(2024 年 2 月)是第一个基于互联网视频训练的无监督基础世界模型,参数量 110 亿。Genie 2(2024 年 12 月)把能力扩展到可由行动控制的 3D 环境,用于训练和评估具身智能体;它能从单张提示图生成最多持续一分钟的一致世界。Genie 3 被列为 DeepMind 当前世界模型,但完整技术披露尚未公开。另一路线中,Gemini Robotics 把语言条件推理应用到多种机器人硬件形态,包括 ALOHA、Bi-arm Franka 和 Apptronik Apollo。DeepMind 的规模——算力、数据、研究人才和 Google 生态整合——使其成为长期竞争威胁,尽管其目标(推进 AI 研究)不同于 AMI 的商业路线图。 Meta FAIR 的 V-JEPA(2024),即 Video Joint Embedding Predictive Architecture,是与 AMI 路径最相似的已发表系统:非生成式、自监督,在抽象表征空间中预测。LeCun 在 Meta 共同设计 JEPA,如今把该架构带到 AMI。随着 LeCun 于 2025 年底离开,Meta 截至 2026 年 6 月的 AI 研究头条转向 “Muse Spark”,显示战略重心离开以 JEPA 为中心的世界模型。Meta FAIR 的 I-JEPA 和 V-JEPA 仍以 Creative Commons NonCommercial 许可证在 GitHub 开源——这构成任何实验室都可继续利用的公开先例,也给 AMI 带来商品化问题。 OpenAI 的 Sora 在技术论文中被明确定位为 “能理解并模拟真实世界的模型基础”,因此是直接世界模型竞争者。但这项战略押注被逆转:OpenAI 于 2026 年 4 月 26 日停用 Sora 网页和应用体验,API 将于 2026 年 9 月 24 日随后停止。此次退却大幅削弱了世界仿真器空间的一大竞争者,也验证了 LeCun 的论点:基于扩散的生成式系统不适合可靠的物理世界应用。 NVIDIA 与众不同,它是基础设施层玩家,而不是终端市场竞争者。Cosmos 于 CES 2025 发布,目前为第 3 版,是一组世界基础模型,基于来自真实世界数据的 20,000,000 小时、9,000 万亿 token 训练,并以开放商业许可证提供。Cosmos 在单一平台中支持视觉推理、机器人策略训练和基于物理的仿真。NVIDIA 的开发者生态庞大:Cosmos 伙伴包括 1X、Agility Robotics、XPENG、Uber 和 Waabi。其 Isaac 平台提供完整机器人软件栈。NVIDIA 的世界模型野心和 GPU 市场主导地位意味着,它将设定每一家世界模型公司用来对标的算力基础设施标准。 [CP012, CP013, CP014, CP015, CP016, CP017]
3.4 具身 AI 与物理机器人:Physical Intelligence 和 Wayve
尽管架构策略不同,Physical Intelligence(π)是商业野心上最接近 AMI 的物理世界 AI 类比对象。AMI 计划把世界模型作为基础设施授权给行业伙伴,Physical Intelligence 则构建并部署直接控制机器人硬件的 vision-language-action(VLA)模型。π0(2024 年 10 月)是第一个使用八种不同机器人形态的跨形态数据训练的通用机器人策略,结合了互联网规模视觉语言预训练和机器人传感运动数据。π0.7(2026 年 4 月)在泛化上实现跃迁:它通过组合多样训练数据中的技能,完成训练中从未见过的任务,实现组合式任务泛化,并能在差异很大的机器人形态间迁移而无需额外数据。Physical Intelligence 已有活跃伙伴计划,并与机器人 OEM 建立具名商业关系——这是 AMI 尚未达到的商业牵引力基准。 值得注意的是,π0.7 使用世界模型生成视觉子目标:给定语言命令后,一个轻量世界模型会生成下一子步骤应呈现的图像,再把它提供给策略模型。这种集成确认,即便机器人基础模型公司也开始把世界模型组件——AMI 的核心产品方向——纳入流水线。它说明 AMI 既可能与 Physical Intelligence 竞争,也可能成为其下方的基础设施层。 Wayve 位于自动驾驶垂直领域,采用它称为 AV2.0 的端到端具身 AI 路径:一个神经网络把原始传感器数据转化为驾驶命令,不依赖 HD 地图,也不依赖特定领域标注数据。Wayve 的路径使用大规模自监督学习——与 AMI 的无监督世界模型论点有哲学血缘——但产品应用专注驾驶,Wayve 并不争夺 AMI 瞄准的工业或医疗市场。Wayve 总部在英国,投资方包括主要汽车 OEM 和科技公司。 HuggingFace 的 LeRobot 开源框架值得一提,它是一股结构性力量:该框架提供硬件无关、Python 原生的机器人控制,配套 Hugging Face Hub 上的数据集和预训练策略,正在主动把物理 AI 普及给开发者和研究者。随着时间推移,LeRobot 会削弱任何单家公司机器人基础模型的专有优势。 [CP020, CP021, CP022, CP023, CP024, CP033]
| 实体 | 定价模式 | 已发布定价 | 主要包含能力 | 商业状态(2026 年中) |
|---|---|---|---|---|
| World Labs (Marble) | API + 订阅(估计) | 未公开披露 | 从文本 / 图像 / 视频生成 3D 世界;World API 访问;Marble Labs 教程 | 已商业部署;World API 于 2026 年 1 月推出 |
| SpAItial (Echo) | 按 API 使用量计费(估计) | 未公开披露 | Echo 世界生成;SPZ/PLY/SOG 导出;浏览器查看器;图像 / 文本 / 全景输入 | API 可用;Echo-2 已发布 |
| Odyssey | 尚未商业化 | N/A | Odyssey-2 世界模型;PROWL RL 框架(研究) | 研究 / 早期访问;无公开定价 |
| Physical Intelligence(π0.7) | 企业授权(估算) | 未公开披露 | 面向多类机器人平台的 VLA 策略;合作伙伴计划;跨具身迁移 | 合作伙伴计划已启动;商业部署已确认 |
| NVIDIA Cosmos | 开放模型许可(免费);DGX Cloud 用于训练 | 开放模型(免费下载);云算力另行计费 | Cosmos WFMs;Isaac 平台;NeMo 微调;Omniverse 集成;分词器 | 已商业可用;可选择 DGX Cloud 部署 |
| Google DeepMind(Gemini Robotics) | 未公开授权 | N/A | 多具身机器人推理;Apptronik Apollo 集成;工具使用 | 研究 / 合作伙伴试点;无商业许可 |
| Meta FAIR(V-JEPA) | 开放 CC-NC(非商业) | 免费(仅限非商业) | V-JEPA 预训练模型;I-JEPA 代码库;GitHub 访问 | 非商业开源;无商业授权 |
| AMI Labs | 计划技术授权 | 尚不可用 | 面向工业 / 医疗应用的世界模型技术(计划) | 收入前研究阶段;尚无产品部署 |
截至 2026 年 6 月,所有非 NVIDIA 定价要么为估算,要么未公开披露;数字来自公开定价页面和媒体报道。NVIDIA Cosmos 可免费下载,但企业级大规模训练需要付费使用 NVIDIA DGX Cloud 算力。AMI Labs 计划中的授权模式尚未在公开文件中说明;公司表述为技术授权。定价中的 “N/A” 表示截至 runDate,该主体没有公开商业产品。
[CP003, CP011, CP024, CP028, CP029]3.5 护城河持久性、差异化与竞争风险
AMI Labs 最核心的架构差异,在于它基于 JEPA、走非生成式路线:它不重建像素或体素,而是在抽象表征空间里预测。因此,AMI 的世界模型理论上更省算力、更可控,也更适合那些幻觉会带来物理成本的场景。当前已商业部署的竞争对手里,还没有哪家公司同时宣称具备这种非生成式架构,并瞄准安全关键的物理世界应用。AMI 的创始团队(LeCun、Xie、Rabbat、Fung)以及明确聚焦医疗和工业领域的选择——Nabla 临床合作已经验证这一点——构成了可信的早期护城河。 但需要评估四类竞争风险。第一,开源商品化:Meta FAIR 的 I-JEPA 和 V-JEPA 已在 GitHub 公开,HuggingFace LeRobot 生态也在持续降低具身 AI 策略的构建成本。JEPA 架构本身属于既有技术,资金充足的竞争者完全可以绕开 AMI,训练一个基于 JEPA 的工业模型。第二,部署速度:World Labs、Physical Intelligence 和 NVIDIA Cosmos 都已有活跃产品和开发者社区;AMI 多年研究承诺意味着竞争者会不断积累分发、数据和客户反馈,并随时间复利。第三,被超大云厂吸收:Google DeepMind 和 NVIDIA 算力几乎不受限,也有路径把任何 AMI 式世界模型进展纳入自家平台。第四,架构套利:世界模型领域变化很快——如果生成式或混合式架构以更低成本达到相近可靠性,AMI 的非生成式差异会收窄。 AMI 最耐久的优势,是 LeCun 无法复制的信誉,与面向受监管行业、专注可控性和安全属性的具体方向叠加。医疗和工业企业面对监管、责任和安全要求,消费级或开发者向世界模型产品并不是为这些要求设计的。如果 AMI 跑通 Nabla 用例并发表结果,它可以在一个尚无竞争者可信宣称占位的垂直领域建立概念验证。 [CP035, CP036, CP037, CP038, CP039, CP040]
| 护城河主张 | 竞争威胁 | 严重性 | 缓释措施 / 尽调问题 |
|---|---|---|---|
| JEPA 非生成式架构优势(可靠性、效率) | Meta FAIR 按 CC-NC 开源 I-JEPA 和 V-JEPA;任何实验室都能训练 JEPA 模型 | 高 | AMI 是否拥有超出 JEPA 架构本身的 IP?训练数据筛选、领域适配或安全评估里是否有商业秘密? |
| LeCun 品牌与人才吸引力 | 关键人物风险:单一创始人的信誉支撑投资人与合作伙伴信心 | 高 | 独立治理核查;如果 LeCun 参与度下降,AMI 是否有足够接班梯队? |
| 医疗 / 工业垂直聚焦(当前无竞争者) | Physical Intelligence 和 NVIDIA Cosmos 正在扩展到工业垂直领域;Gemini Robotics 有企业试点 | 中 | Nabla 合作是否排他,是否能产出可复现证据?更多医疗或工业部署的时间表是什么? |
| 算力效率(用于端侧部署的小模型) | NVIDIA 开放 Cosmos 生态让大规模生成式模型几乎以零边际成本可用 | 中 | 是否有已确认的基准,在同等准确率下比较 AMI 模型算力与生成式竞争对手?边缘部署规格是否已定义? |
| 研究发表速度与人才留存 | 前沿实验室(DeepMind、Meta FAIR)争夺同一批研究人才;Meta 和 HuggingFace 的开源文化抬高外部机会成本 | 中 | AMI 的薪酬和股权结构竞争力如何?是否有已确认离职? |
| 与竞争对手相比的商业部署缺口 | World Labs、Physical Intelligence 和 NVIDIA 都已有部署产品或开放模型,带来开发者采用和数据飞轮;AMI 尚无产品,并承诺多年研究 | 高 | AMI 的首个外部部署里程碑是什么?除 Nabla 外,首批商业被许可方是谁? |
| HuggingFace LeRobot 生态带来的开源商品化 | LeRobot 提供硬件无关策略和开放数据集,持续降低具身 AI 门槛;AMI 必须靠专有训练数据和工业领域专长做出差异 | 中 | AMI 是否拥有无法从开放数据集复制的特定领域专有数据资产? |
严重性评级(高 / 中 / 低)是基于截至 2026 年 6 月收集到的竞争证据作出的定性判断;它们反映风险落地后的商业影响幅度,而非发生概率。“尽调问题” 是投资人尽调中的开放问题,并非指控当前失败。AMI 尚未公开披露 IP 申请、专有数据集范围或详细路线图里程碑。
[CP035, CP036, CP037, CP038, CP039]AMI Labs 在世界模型初创公司中资金基础最大,但距离商业部署的缺口最长。已有部署产品的竞争者数量是 AMI 的 5 倍。
“活跃商业产品或开放模型”包括 World Labs(Marble)、Physical Intelligence(π0.7 合作伙伴项目)、 NVIDIA(Cosmos 开放模型)、SpAItial(API)和 Wayve(AV 产品)。AMI 产品时间线基于其公开声明的多年期、 研究优先承诺估算;公司尚未披露官方时间表。融资对比来自二级媒体报道估算;未披露金额可能不同。
[CP003, CP028, CP037, CP038]3.6 展品
04财务
4.1 资本结构与轮次机制
AMI Labs 在 2026 年 3 月 10 日宣布完成 $1.03 billion USD(约 €890 million)种子轮融资。该轮按 $3.5 billion 投前估值定价,隐含投后估值约 $4.53 billion。PitchBook 确认这是欧洲有记录以来最大的一笔种子轮。AMI 直到 2025 年 12 月前后最初还在寻求约 €500 million;最终交割金额接近目标两倍,说明投资团获得超额认购。 本轮由五家 venture investors 共同领投:Cathay Innovation、Greycroft、Hiro Capital、HV Capital 和 Bezos Expeditions。战略企业投资者包括 NVIDIA、Samsung、Temasek、Toyota Ventures 和 Sea。法国机构和产业支持方包括 Association Familiale Mulliez、Groupe Industriel Marcel Dassault、Publicis Groupe 和 Bpifrance Digital Venture。知名天使包括 Eric Schmidt、Mark Cuban、Jim Breyer、Tim and Rosemary Berners-Lee、Xavier Niel 和 Mark Leslie。 2026 年 4 月 3 日提交给 US Securities and Exchange Commission 的 Form D 显示,美国投资载体 “Advanced Machine Intelligence I, L.P.” 和 “Advanced Machine Intelligence II, L.P.”——由旧金山的 Model I, L.P. 管理——合计募集约 $16.9 million,用于投资 AMI。这种结构符合把小额美国 LP 份额汇入更大国际投资团的安排。AMI 本身是一家法国私营公司,没有公开财务申报或资产负债表披露。 战略投资者的逻辑看起来分成两条线:财务 VC(Cathay、Greycroft、Hiro、HV)接受了多年研究时间线,并以高于同阶段创业公司的溢价定价;企业战略方(NVIDIA、Samsung、Toyota、Temasek)则可能获得世界模型研究的早期接触权,这些研究与其算力、半导体、汽车和主权 AI 优先事项相关。超额认购说明两类投资者都认为,考虑到 LeCun 的技术信誉和世界模型叙事,$3.5B 投前估值是合理价格。 [CI001, CI002, CI003, CI004, CI005, CI006]
| 参数 | 数值 / 估算 | 来源 / 依据 | 置信度 | 备注 |
|---|---|---|---|---|
| 种子轮总融资 | $1.03B USD (~€890M) | AMI Labs 官方;TechCrunch | 高 | 2026 年 3 月 10 日宣布 |
| 投前估值 | $3.5B USD | AMI Labs;Dataconomy | 高 | 2026 年 3 月种子轮交割时确定 |
| 投后估值(隐含) | ~$4.53B USD | 推断:投前估值 + 轮次规模 | 高 | 推导计算 |
| 在手现金(估算) | ~$1.03B | 根据轮次交割推断;无提款数据 | 低 | 假设交割前支出很少 |
| 月度烧钱率(估算,第 1 年) | $12M–$25M | 参照早期前沿实验室;仅为估算 | 低 | AMI 未披露烧钱数据 |
| 年度烧钱率(估算,第 1 年) | $150M–$300M | 算力 + 人才参照;见成本表 | 低 | 区间很宽;未验证 |
| 低烧钱情形下的估算资金续航期 | ~6.9 年 | 推断:$1.03B ÷ $150M/yr | 低 | 对烧钱加速敏感 |
| 高烧钱情形下的估算资金续航期 | ~3.4 年 | 推断:$1.03B ÷ $300M/yr | 低 | 对算力扩张敏感 |
| 计划资金用途(已披露) | 算力和人才 | LeBrun 明确表述 | 高 | 未披露预算拆分 |
| 下一轮融资触发因素(估算) | 研究里程碑或合作伙伴试点(估算为 2028–2030 年) | Futurum Group;推断 | 低 | 无官方 Series A 时间表 |
| 债务 / 项目融资义务 | 未披露 | 无公开记录 | 低 | 私营公司;未知 |
在手现金和烧钱数字均为参照前沿 AI 实验室案例得出的估算;AMI 未披露任何专属财务数据。投前和投后估值直接按已宣布轮次计算。资金续航期数字是线性外推,未计入后续研究阶段的烧钱加速。
[CI001, CI002, CI003, CI004, CI009, CI025]4.2 收入模型与 GTM 时间线
AMI Labs 没有收入,没有商业产品,并已明确表示近期不计划创收。CEO Alexandre LeBrun 把公司描述为“一个非常有雄心的项目,因为它从基础研究开始。它不是那种典型的应用 AI 创业公司:三个月发产品、六个月有收入、十二个月做到 $10 million ARR。” 他表示,世界模型从理论走向商业应用可能需要“数年”。 预期的变现路径——一旦到来——是技术许可:向工业伙伴出售特定领域的世界模型能力,而不是销售消费级或企业软件。Yann LeCun 预计,2026 年 3 月种子轮宣布后 6–12 个月内,可能开始与企业伙伴讨论;而通向“准通用智能系统”的时间线是 3–5 年。Nabla 是 LeBrun 之前创办的医疗 AI 公司,也是 AMI 首个且唯一公开披露的商业伙伴;它在非收入条款下获得早期研究产出的访问权。 因此,AMI 的 GTM 动作是从研究走向伙伴,而不是从产品走向客户。没有标价、没有交易历史、没有客户 pipeline,也没有可评估的 ARR。销售效率指标(CAC、回本期、NRR)在这个阶段结构性不适用。Nabla 合作是医疗场景中的试点环境和早期概念验证,不是创收型商业协议。并行的开源和学术发表承诺,把 AMI 定位成研究平台而不是产品公司;这可能吸引后续合作,但会推迟支撑高价许可所需的传统 IP 护城河形成。 [CI010, CI011, CI012, CI013, CI014, CI015]
| 收入流 | 机制 | 当前状态 | 收入质量(激活后) | 关键尽调问题 |
|---|---|---|---|---|
| 特定领域世界模型授权 | 工业 / 医疗合作伙伴为领域调优世界模型支付授权费 | 未激活 — 产品前阶段 | 若可复制,潜在毛利率高;若逐项安装,质量较低 | 定价模型、最低承诺、排他条款 |
| 研究合作伙伴早期访问 | 合作伙伴(如 Nabla)获得模型早期访问;条款未披露 | 已激活(仅 Nabla)— 非收入 | N/A — 当前无收入 | 是否涉及任何费用、数据或收入? |
| 开源生态杠杆 | AMI 发布代码 / 论文;为未来商业化建设社区 | 已激活 — 计划发布学术 / 代码成果 | 直接收入为零;通过人才 / 合作伙伴间接产生价值 | 开源许可是否允许合作伙伴商业使用? |
| 战略投资人的算力实物支持 | NVIDIA、Samsung 可能把 GPU / 硬件访问作为战略投资的一部分提供 | 可能 — 条款未披露 | 非现金;抵消烧钱,而不是产生收入 | NVIDIA 提供的确切算力配额和定价 |
| 政府 / 主权 AI 补助 | Bpifrance Digital Ventures 共同投资;可能获得欧盟 / 法国公共补助 | 可能 — 尚未宣布正式补助 | 非稀释性资本;不是经营收入 | 是否已有公共 R&D 补助或主权 AI 资金承诺? |
AMI 没有经营收入。所有商业路径描述都基于 LeBrun 表述的意图,以及可比研究实验室的案例。收入质量评估是假设性判断,无法用实际交易打分。
[CI010, CI011, CI012, CI013, CI014, CI015]展示 AMI Labs 的研究管线如何计划在多年维度转化为特定领域授权收入,同时把开源发布作为并行路径。
收入路径来自公司声明的意图;截至运行日期,公司尚无收入合同。开源路径是并行承诺。LeCun/LeBrun 表示,走到收入需要多年。
[CI012, CI013, CI015, CI027]4.3 成本结构、burn 对标与 runway 估算
LeBrun 明确把算力和人才列为 AMI 两大成本中心。本章所有 burn 数字都是参照 frontier AI lab 对标推导的估算;AMI 未披露任何公司自身财务数据。 算力方面:Epoch AI 数据显示,自 2020 年以来,前沿模型训练算力每年增长 4–5 倍;已知最大的 AI data center 需要约 $24 billion 资本成本,才能运行 800,000 张 H100-equivalent 芯片。不过,AMI 的架构论点——JEPA 世界模型需要数亿而不是数千亿参数——意味着单次训练所需算力应显著低于 frontier LLM。AI 芯片单位美元性能约每年提升 37%(Epoch AI),也就是说算力成本按实际口径在下降。Sequoia Capital 的分析估计,GPU 约占 AI data center 总持有成本的 50%,其余为能源、建筑和网络。 人才方面:AMI 启动时团队估计约 12 人(Futurum Group analysis,2026 年 3 月),并正在巴黎、纽约、蒙特利尔和新加坡招聘。LeBrun 表示 AMI 会“重质量胜过重数量”。可比实验室(OpenAI、Anthropic、DeepMind)的 frontier AI research 薪酬公开记录显示,高级研究人员的年度总薪酬从数十万美元到超过 $1 million 不等。如果 AMI 扩张到 100 名研究人员和员工,平均 fully-loaded cost 为 $400,000–600,000,则年度人才支出会达到 $40–60 million——还未计入创始团队股权稀释。 把保守算力支出(早期研究规模下 $50–150M/year)与人才成本(80–150 人规模下 $40–80M/year)合并,再加上基础设施和 overhead,未来 1–3 年每年 $150–300 million 的 burn rate 属于合理区间。Goldman Sachs 质疑,全球估计 $1 trillion 的 AI 基础设施支出最终能否产生相称收入回报。Sequoia 警示,投机性 AI 基础设施投资狂热在历史上会导致资本焚烧、GPU 算力定价能力侵蚀,以及快于预期的硬件折旧。这些宏观风险不直接作用于 AMI 近期运营,但会让 Series A 融资环境更困难。 按 $150M/year 的 burn,AMI 的 $1.03B 约可支撑 6.9 年资金续航。按 $300M/year,资金续航降至约 3.4 年。这些只是指示性区间;实际 burn 轨迹高度取决于 AMI 是否会在第 2–3 年扩大到大算力训练。 [CI009, CI017, CI018, CI019, CI020, CI021]
| 成本类别 | 低位估算(USD M/yr) | 高位估算(USD M/yr) | 估算依据 | 置信度 |
|---|---|---|---|---|
| 算力 — 本地或云 GPU 集群 | 40 | 150 | 按 Epoch AI 数据缩放;前沿 JEPA 训练规模小于 LLM | 低 |
| 人才 — 研究科学家和工程师 | 30 | 70 | 80–150 名 FTE,完全负担平均成本 $400K–$600K | 低 |
| 人才 — 运营、法务、财务、行政 | 10 | 25 | 支持人员估为研究人员人数的 20–30% | 低 |
| 云基础设施、数据和工具 | 5 | 20 | 标准 AI 实验室基础设施栈;NVIDIA 战略投资可能抵消部分成本 | 低 |
| 设施和办公室成本(4 个城市) | 5 | 15 | 巴黎总部 + 纽约、蒙特利尔、新加坡办公室 | 低 |
| 年度总烧钱估算 | 150 | 300 | 上述类别合计 | 低 |
所有数字都是基于可比前沿 AI 实验室公开数据、Epoch AI 算力成本数据和典型前沿 AI 人才市场价格作出的外部推断。AMI 未披露专属成本数据。NVIDIA 的战略投资可能通过优惠 GPU 访问抵消部分算力支出;这些估算没有计入该收益。
[CI009, CI018, CI019, CI022, CI023, CI025]展示 AMI 年度烧钱率、隐含跑道和轮次估值指标的低、中、高估算;所有烧钱 / 跑道数字均为分析师推断。
烧钱率估算来自外部对前沿 AI 实验室类比和 Epoch AI 数据的推断;AMI 没有可用财务披露。跑道按 $1.03B 除以相应烧钱率估算。估值数字直接来自 2026 年 3 月公告。
[CI001, CI025, CI026, CI040]示意第 1 年从 $1.03B 种子资本流向估算算力、人才和运营支出类别后的烧钱瀑布,并给出第一年末预计剩余跑道。除种子轮融资外, 下列所有数字均为分析师估算。
除种子轮融资外,所有项目均为外部估算,并以可比前沿 AI 实验室为基准。若 NVIDIA 提供战略性算力抵扣, 将降低算力支出项。$870M 剩余数字采用低端烧钱情景的中点;实际资金消耗取决于 AMI 的招聘和算力决策, 这些决策尚未披露。
[CI009, CI022, CI025, CI034]4.4 公开 traction 与财务数据缺口
从公开来源看,AMI 的财务画像几乎不透明。它是一家法国私营公司(société par actions simplifiée),无义务公布账目,没有 SEC reporting,没有公开 ARR,也没有披露收入。外部来源能观察到的唯一财务事实,是种子轮规模和估值;多家独立媒体和公司自有公告都确认了这些信息。美国投资载体 2026 年 4 月的 SEC Form D filing 显示,其合计募集约 $16.9 million,为美国投资者 tranche 结构打开了一扇小窗口,但不涉及 AMI 内部财务。 员工数数据稀少:Futurum Group 估计种子轮宣布时约 12 名员工。AMI 的 Ashby job board(2026 年 5 月归档)可访问,但只提供有限结构化数据。可信第三方没有发布 LinkedIn 员工数、Glassdoor 薪资数据,或 AMI 员工的独立统计。考虑到 2026 年 3 月种子轮交割和 LeBrun 明确的招聘指令,当前员工数几乎肯定高于 12 人,但没有精确数字。 与 NVIDIA(战略投资者)的算力访问条款未披露。NVIDIA 对 AI lab 的战略投资通常包括算力分配协议,这可能实质性抵消 AMI 的硬件支出。Samsung 和 Toyota 合作也可能提供数据或硬件访问,而这些不反映在基于纯财务投资者得出的 burn-rate analogues 中。如果存在这些 in-kind 安排,AMI 的有效 runway 会长于仅按现金估算的结果。 Nabla 合作条款未披露:AMI 是否从 Nabla 获得任何收入、in-kind 数据访问或医疗领域 expertise,以及时间线如何,都不得而知。LeBrun 将其描述为“对 AMI 研究的早期访问”,不是商业许可,说明这段关系主要是 AMI 的数据和验证 pipeline,而非收入来源。 [CI029, CI032, CI034, CI036]
| 指标 | 数值 | 置信度 | 重要性 | 尽调闭环路径 |
|---|---|---|---|---|
| 年度经常性收入(ARR) | N/A | SaaS / 授权业务的核心健康指标 | 需要商业产品和付费客户 | |
| 收入运行率 | N/A | 体现商业牵引力规模 | 授权启动前不可得 | |
| 毛利率 | N/A | 决定规模化后的长期单位经济 | 需要模型交付成本和授权合同结构 | |
| 获客成本(CAC) | N/A | 销售效率基准 | 尚未开始获客活动 | |
| LTV / CAC 比率 | N/A | SaaS 或授权可行性的核心指标 | 需要商业部署中的 CAC 和流失数据 | |
| 净收入留存(NRR) | N/A | 授权客户的扩张效率 | 上线前不适用;Series B 尽调时询问 | |
| 员工人数 | 种子轮时约 12 人(估算);仍在增长 | 低 | 烧钱驱动项;团队密度反映研究产出速度 | LinkedIn 人数统计;内部 HR 数据 |
| 人均收入 | N/A | 收入规模化后的团队效率 | 需要收入和已确认员工数 | |
| 年度烧钱率 | $150M–$300M 估算 | 低 | 衡量 runway 和融资依赖 | 经审计现金流量表;提款计划 |
| 总烧钱(每月现金流出) | $12M–$25M 估算 | 低 | 运营节奏和投资人风险 | 尽调中通过银行流水验证 |
所有为 null 的单位经济字段都反映截至 2026 年 6 月 AMI Labs 仍处于收入前、产品前阶段。烧钱率估算是分析师参照可比前沿 AI 研究实验室得出的推断;没有公开可得的 AMI 财务数据。Series A 尽调拿到实际运营数据后,必须重新评估这些字段。
[CI010, CI022, CI023, CI025, CI026]| 缺失数据项 | 对分析的影响 | 尽调路径 |
|---|---|---|
| 实际现金烧钱率和月度提款计划 | 无法验证资金续航期或烧钱轨迹 | 经审计财务;Series A 尽调中访谈 CFO |
| 按角色、级别和地点拆分的员工人数 | 无法建模人才成本或研究团队密度 | LinkedIn 人数;数据室中的内部 HR 披露 |
| NVIDIA 算力访问条款(战略投资中的实物支持) | 可能显著降低硬件烧钱;是否存在未知 | NVIDIA 和 AMI 在投资材料中直接披露 |
| Nabla 合作商业条款(费用、数据权、期限) | 收入、数据访问和纯非商业合作之间界限不清 | 审阅合作协议;访谈 Nabla CFO |
| Series A 时间表、目标规模和估值预期 | 对资金续航期规划和稀释模型至关重要 | 访谈 CEO;数据室中的投资人沟通 |
| AMI 算力架构 — 每次训练运行的确切参数量和 GPU 需求 | 决定 JEPA 规模算力究竟比 LLM 规模低很多,还是只低一些 | 内部模型架构披露;技术尽调 |
| 法国 / 欧盟公共项目中的任何政府补助、补贴或非稀释性资本 | 可能显著延长 runway;是否申请或获批未知 | Bpifrance、EU Horizon 或 PIIA 补助记录;公司披露 |
所有缺口都源于 AMI Labs 是收入前私营公司,且没有公开财务申报义务。要做有意义的承销,必须在正式 Series A 尽调中补齐这些缺口。
[CI029, CI034, CI036]映射 AMI Labs 标准 SaaS 与授权单位经济链条,标出当前因尚未产生收入而全部为 null 的节点,以及填充每个节点所需的证据。
截至 2026 年 6 月,AMI 没有收入、客户或定价,因此所有节点均为 null。本图展示的是证据缺口,而非实际数值。
[CI010, CI011, CI029]4.5 财务结论与尽调阻断项
AMI 的财务位置结构简单,但执行不透明。公司持有约 $1.03 billion 种子资本,全部用于算力和人才;收入为零,且处于多年 pre-revenue 时间线。按保守 burn rate,资本充足性似乎可以支撑其声明的研究任务;若每年 burn 为 $150–200M,$1.03B 种子资金大约能让 AMI 在需要 Series A 之前支撑五到七年。但如果第 2–3 年算力需求像可比实验室从小规模研究扩张到大训练一样加速,资金续航可能迅速压缩。 没有收入,因此 revenue quality 无法评级。利润率路径完全停留在理论上(许可模型意味着高毛利,但还没有可定价产品)。按研究实验室标准,资本强度很高:算力和人才两个成本中心规模相近,研究阶段没有收入抵消。融资依赖是完全的:未来至少 3–5 年,AMI 离不开外部资本。 两个最重大的财务风险是:(1)如果训练需要扩张到高于初始预测的规模,burn 会加速,可能缩短 runway,并迫使公司在不如超额认购种子轮有利的条件下融资 Series A;(2)time-to-revenue 风险,即世界模型研究周期长于预期,迫使公司以可能更低估值追加融资。第三个风险,是 AMI 的开源承诺与高价许可模型所需 IP 保护之间的张力:开源服务研究社区,也可能吸引人才和伙伴,但会削弱高价许可的可执行性。Goldman Sachs 和 Sequoia 都系统性追问过,AI 基础设施投资能否在 venture timeline 内产生回本。 Futurum Group 指出,AMI 的 Series A 将是研究产出能否转化为商业可信度的 “the first real market test”。尚未完成的关键财务尽调事项包括:(a)实际 cash burn rate 和 draw-down schedule;(b)与 NVIDIA 的任何 in-kind compute 安排;(c)Nabla 合作的商业条款;(d)headcount ramp plan 和总薪酬预算;(e)Series A 时间线和估值目标;(f)AMI 的具体算力架构要求——frontier scale 的 JEPA 训练成本远低于 frontier LLM,但精确数量级仍是私密信息。 [CI028, CI029, CI039, CI040]
4.6 展品
05产品与技术
5.1 技术架构与 JEPA 脉络
AMI Labs 的整个技术身份,都可追溯到 Yann LeCun 在 2022 年 6 月立场论文 “A Path Towards Autonomous Machine Intelligence” 中提出的 Joint Embedding Predictive Architecture(JEPA)。核心洞见是:世界模型应该在抽象表征空间,而不是原始像素或 token 空间里预测。这条非生成式路线让系统可以有意忽略不可预测的细节——树叶移动、表面噪声、随机传感器伪影——同时保留高层语义结构。LeCun 架构提出五个相互作用的模块:perceptor(多模态感知编码器)、world model(JEPA 预测器)、actor(行动规划器)、configurator(目标和策略设定器),以及双记忆系统(短期和长期 episodic)。没有任何单一模块构成 AMI 的产品;完整认知架构才是公司声明的终点。JEPA 脉络已有两个公开实现。I-JEPA(CVPR 2023)只用未标注数据,训练 Vision Transformer 在潜空间预测被遮蔽的图像区域;一个 ViT-Huge/14 在 16 张 A100 GPU 上不到 72 小时完成 ImageNet 训练。V-JEPA(Meta,2024)把方法扩展到视频:遮蔽大块时空区域,要求预测器在表征空间而非像素中补全。相对既有视频表征学习 baseline,V-JEPA 训练和样本效率提升 1.5× 到 6×,并擅长“冻结评估”——编码器预训练一次后保持冻结,轻量 probes 处理下游任务。V-JEPA 以 Creative Commons NonCommercial licence 发布。AMI 计划的下一步,是 action-conditioned world models:能够模拟拟议动作的后果,并在安全 guardrails 约束下选择行动序列——这种能力尚未公开展示。 [CE001, CE002, CE003, CE004, CE005, CE006]
| 层 / 组件 | 角色 | 依赖项 | 风险 |
|---|---|---|---|
| 情境编码器(ViT 主干) | 将部分观测编码为潜在表示 | 大规模无标签预训练语料;GPU 算力 | 需要海量数据;传感器输入偏离分布时质量会下降 |
| 目标编码器(EMA) | 借助指数移动平均提供稳定的目标嵌入 | 情境编码器参数 | EMA 超参数调错会导致训练不稳定 |
| JEPA 预测器(世界模型核心) | 基于情境编码预测目标区域表示 | 情境编码器;目标编码器 | 当前 V-JEPA 只覆盖约 10 秒时间跨度;长时域问题未解决 |
| 动作条件规划器 | 基于世界模型预测选择并模拟动作序列 | JEPA 预测器;安全护栏模块 | 尚未构建;设计只存在于 LeCun 2022 年立场论文中 |
| 多模态融合层 | 将非视觉传感器流(音频、生命体征、lidar)与视觉编码器整合 | 配对的多传感器数据集;模态对齐训练 | 架构未披露;尚无已发表的多模态 JEPA 论文 |
| 安全护栏模块 | 将规划动作限制在安全运行边界内 | 领域安全规范;监管输入 | 尚未演示;FDA 或安全标准合规路径未定义 |
架构层来自 LeCun 2022 年立场论文、I-JEPA 和 V-JEPA 论文,以及 AMI 官方沟通材料的重构。动作条件规划器和安全模块是设计目标,不是已实现组件。公开证据不足,无法独立验证该架构。
[CE001, CE002, CE003, CE004, CE005, CE027]AMI 从原始传感器(底层)到行动条件规划(顶层)的 JEPA 认知架构,展示非生成式预测流。
架构根据 LeCun 2022 年立场论文、I-JEPA 和 V-JEPA 论文重建。未构建层标注为(计划中)。AMI 实际实现细节尚未公开。
[CE001, CE002, CE003, CE027]5.2 产品定义、用例与许可模型
AMI Labs 短期内明确不是一家产品公司。公司官网和 CEO 表态确认,它是一家 pre-revenue 研究实验室,追求技术许可模型:开发世界模型,并提供给构建垂直应用的行业伙伴。Nabla 是首个披露的伙伴;这家医疗 AI 公司由 Alexandre LeBrun 共同创办,如今 LeBrun 领导 AMI。2025 年 12 月宣布的 Nabla 合作,给予 Nabla 对 AMI 新兴世界模型技术的优先且特权访问权,目标是为临床护理构建可获 FDA 认证的 agentic AI systems——具体是对音频、生命体征、影像等连续医疗信号进行确定性、可审计决策。医疗之外,AMI 官网列出五个目标垂直领域:工业流程控制、自动化、可穿戴设备、机器人,以及兜底的“其他”。这些领域的共同点是安全关键:LLM 幻觉不可接受,可靠性、可控性和安全是一级要求。许可模型尚未披露类似 SaaS 或按 token API 计费的机制;实际商业安排预计要等世界模型成熟到足以进入伙伴试点之后才会出现。CEO LeBrun 明确表示,公司可能需要“数年”才能让世界模型从理论进入商业应用;早期伙伴参与会围绕真实世界数据和联合评估,而不是离散产品发布。 [CE011, CE014, CE015, CE016, CE021, CE022]
| 模块 / 资产 | 目标用户 / 买方 | 状态 / 成熟度 | 差异化 | 尽调缺口 |
|---|---|---|---|---|
| I-JEPA 图像世界模型 | AI 研究人员和开发者 | 已发布 + 开源(CVPR 2023) | 非生成式语义图像表征;不需要手工增强 | 无企业部署;CC BY-NC 排除商业使用 |
| V-JEPA 视频世界模型 | AI 研究人员、计算机视觉从业者 | 研究发布(Meta 2024,CC BY-NC) | 训练效率较此前高 1.5–6×;支持冻结评估 | 仅限约 10 秒片段;无商业许可 |
| 动作条件世界模型 | 行业合作伙伴(机器人、工业) | 预发布 / R&D 概念 | 围绕预测动作后果做规划,并加入安全护栏 | 尚未公开演示;架构只存在于设计论文 |
| 多模态传感器融合模块 | 工业、医疗、可穿戴合作伙伴 | 研究概念 / 路线图 | 在视觉之外处理连续非视觉传感器模态 | 无原型或论文;架构细节未披露 |
| AMI 世界模型平台(授权) | Nabla(医疗)、战略工业合作伙伴 | 试点合作阶段(Nabla 于 2025 年 12 月宣布) | 临床护理中可获 FDA 认证的智能体 AI 先发者 | 未交付产品;无 API 可用;无监管申报 |
成熟度评级基于 AMI 官方表述,以及 LeCun 在 Meta FAIR 时期发表的 JEPA 论文。I-JEPA 和 V-JEPA 是 Meta 时期的研究发布,不是 AMI 品牌产品。尽调缺口反映缺少独立技术验证。
[CE001, CE004, CE005, CE006, CE013, CE014]| 用户任务 | 当前工作流 | AMI 方案(拟议) | 可衡量收益 | 限制 |
|---|---|---|---|---|
| 临床文档与决策支持 | 基于 LLM 的环境记录;容易幻觉 | 世界模型围绕音频 / 生命体征进行确定性、可审计推理 | 降低幻觉风险;有可信的 FDA 认证路径 | 无可用产品;Nabla 试点时间表未披露 |
| 工业流程监控与控制 | 基于规则的 SCADA 系统;反应式异常告警 | 用连续传感器世界模型做预测控制和异常仿真 | 提前预测故障;基于仿真的情景假设分析 | 概念阶段;未宣布试点;需要特定领域训练数据 |
| 机器人任务规划 | 手工策略;窄域 RL 智能体;昂贵遥操作 | 用动作条件世界模型做零样本目标条件规划 | 更快泛化到新任务;降低遥操作成本 | 规划模块尚未构建;未展示机器人集成 |
| 可穿戴情境智能 | 简单手势 / 计步;孤立的应用级模型 | 带情景记忆的持久多模态世界模型 | 持续的情境感知;环境式健康监测 | 尚未做出原型;硬件无关的说法尚未在边缘侧验证 |
所有拟议中的 AMI 方案都来自已发表的研究意图和合作伙伴公告,而不是已部署产品。可衡量收益是架构设计目标,不是已验证的性能指标。
[CE010, CE014, CE015, CE023, CE026, CE032]行业合作伙伴(以 Nabla 医疗场景示意)如何从研究合作开始,获取并部署 AMI 的世界模型,最终走向监管提交。
流程根据 AMI 和 Nabla 公开声明推断。除最初的 Nabla 合作公告外,任何步骤都没有披露时间线。
[CE014, CE015, CE016, CE021, CE022]5.3 开放研究策略、开发者社区与 IP 差异化
AMI Labs 已公开承诺开放论文和开源代码,把自己定位为研究社区建设者,而非封闭 incumbent。CEO LeBrun 明确表示“我们也会开源大量代码”,并称“开放时事情推进得更快”。I-JEPA 代码库已经在 GitHub 的 facebookresearch(Meta repository)下公开,V-JEPA 则以 Creative Commons NonCommercial licence 发布。AMI 维护一个 Medium blog,但渠道还处于早期。Yann LeCun 保留 NYU 教职,并继续指导博士和博士后研究,形成直接的学术到商业 pipeline。公司的差异化 IP 论点有三根支柱:JEPA 架构族(表征空间中的非生成式世界建模)、带安全 guardrails 的 action-conditioned planning,以及超越纯视觉的多模态传感器数据处理。这不同于 NVIDIA Cosmos——后者是生成式(diffusion 和 autoregressive),面向机器人和 AV 开发者生成模拟数据——也不同于 DeepMind Genie 2,后者同样是生成式,面向 embodied agents 的训练环境。两者都没有直接处理 AMI 正在构建的 action-conditioned non-generative architecture。竞争者 SpAItial(面向 3D Gaussian Splatting 的 Echo world model)、World Labs(spatial 3D world generation)和 Physical Intelligence(robot foundation models)都走生成式或窄机器人路线。AMI 的 JEPA 脉络由 LeCun 的 Turing Award 声望和已发表的 CVPR / 开源记录锚定,提供了竞争者不易复制的可信 prior art 和学术合法性。 [CE012, CE013, CE017, CE018, CE019, CE024]
| 日期 / 阶段 | 功能 / 里程碑 | 状态 | 含义 | 来源 |
|---|---|---|---|---|
| 2022 年 6 月 | LeCun 发表「A Path Towards Autonomous Machine Intelligence」——包含 JEPA 的完整认知架构 | 已发表(OpenReview) | 奠定理论框架;AMI 的整个产品愿景都源自该文件 | SE003 |
| 2023 年 4 月 | I-JEPA 发表在 CVPR 2023;代码在 GitHub 开源 | 已发表 + 已开源 | 首个得到验证的 JEPA 实现;证明非生成式表示学习可规模化 | SE004 |
| 2024 年 2 月 | Meta FAIR 以 CC BY-NC 许可证发布 V-JEPA(Video JEPA) | 已发布(研究) | 视频世界模型能力得到验证;相对既有方法效率提升 1.5–6× | SE005 |
| 2024 年 12 月 | DeepMind 推出 Genie 2——可由动作控制的生成式 3D 世界模型 | 竞争对手里程碑——已发布 | 生成式世界模型快速推进;AMI 必须靠非生成式安全论点做出差异化 | SE013 |
| 2025 年 12 月 | AMI Labs 成立;宣布与 Nabla 独家合作;$1.03 B 种子轮启动 | 已宣布 | 医疗合作落定;在临床智能体 AI 中获得先发定位 | SE007 |
| 2026 年 4 月 | OpenAI Sora 停止运营;Physical Intelligence 发布 π0.7 可操控机器人模型 | 竞争对手事件 | 生成式视频世界模型面临商业阻力;机器人专用模型正在实现商业部署 | SE020 |
| 2026 年(进行中) | AMI 动作条件世界模型和多传感器整合仍在开发中 | 发布前 / 研发 | 首个合作伙伴部署时间表未确认;没有公开 alpha 或 beta | SE001 |
AMI 内部里程碑仅来自官方沟通材料。竞争对手里程碑(Genie 2、Sora、π0.7)是公开记录事件,用于放置 AMI 在世界模型格局中的位置。
[CE001, CE004, CE005, CE019, CE020, CE021]AMI 要把世界模型平台推向商业部署,必须解决的上游关键依赖——研究传承、算力、人才、数据和监管。
依赖图根据公开证据重建。截至运行日期,安全模块和监管批准节点尚未构建 / 尚未达成。
[CE001, CE013, CE015, CE023]5.4 安全、可控性与监管路径
AMI Labs 把安全和可控性放在架构层,而不是事后 guardrail。非生成式 JEPA 方法旨在从结构上消除幻觉:因为预测器在抽象表征空间工作,并丢弃不可预测的像素级细节,它不能像 LLM token sampler 那样编造具体事实内容。AMI 官网写道:“Action-conditioned world models 让 agentic systems 能预测自身行动的后果,并在 safety guardrails 约束下规划完成任务的行动序列。” Nabla 合作公告强化了这一框架:世界模型支持“确定性、可审计决策”和“基于模拟的推理与‘假设’分析”——这些属性是任何自主临床 AI 提交 FDA 监管申请的核心。Nabla 的既定目标,是成为首个用 AMI 技术把可获 FDA 认证的 agentic AI systems 带入医疗的公司。但截至 2026 年 6 月,还没有监管申报,没有获得安全认证,也没有公开呈现针对世界模型系统的正式安全评估。架构安全论证——非生成式等于无幻觉——尚未被独立验证,也未在临床场景中与 LLM 错误率做 benchmark。AI Snake Oil / AI as Normal Technology 的批评者质疑,学术 JEPA 演示与生产级临床 AI 之间的缺口,能否在没有大量领域专用训练数据、监管工程和临床验证的情况下跨越;这些条件无法从当前公开证据中推导出来。 [CE010, CE015, CE020, CE023, CE025, CE031]
| 控制项 / 指标 | 状态 | 范围 | 缺口 |
|---|---|---|---|
| 幻觉规避(架构层面) | 设计原则——在表示空间中做非生成式预测 | 所有基于 JEPA 的模型 | 尚未与 LLM 做正式基准对比;没有独立临床安全研究 |
| 开源代码可用性 | 活跃:I-JEPA(GitHub/facebookresearch);V-JEPA(CC BY-NC) | 仅限研究模型;许可证限制商业使用 | 没有 AMI 品牌开源仓库;来自 Meta FAIR 传承,不是 AMI IP |
| 可获 FDA 认证的 AI 路径 | Nabla 已表述合作目标;仅为设计意图 | 医疗垂直领域(Nabla 合作) | 没有监管申报;没有预提交会议证据;时间表未披露 |
| 确定性 / 可审计推理 | 按 AMI 和 Nabla 公告,为架构设计意图 | 计划用于动作条件智能体系统 | 没有已部署系统可审计;生产环境中尚未验证 |
| 安全认证 / ISO 合规 | 未披露 | AMI 未说明 | 未提及认证;没有第三方审计证据 |
所有状态条目都基于 AMI 和 Nabla 官方公开沟通材料。尚未确认任何独立第三方审计、认证或监管提交。缺口反映公开证据不存在,并不确认内部工作不存在。
[CE010, CE015, CE023, CE034, CE035]AMI 世界模型能力在研究、演示、合作伙伴集成和商业化维度上的成熟度。
成熟度评估仅基于公开证据。所有 I-JEPA 和 V-JEPA 研究与代码均属于 Meta FAIR 传承;截至 2026 年 6 月, AMI 尚未以自己名义发布独立论文或代码库。
[CE004, CE005, CE006, CE008, CE021, CE040]5.5 竞争格局与执行缺口
世界模型领域竞争激烈、资金充足,并在快速分化。NVIDIA Cosmos(CES 2025 发布)是已完全部署的生成式世界基础模型平台,基于 20 million 小时真实世界数据和 9 trillion tokens 训练,已被 1X、Agility Robotics、Waabi 等机器人和 AV 公司采用。DeepMind Genie 2(2024 年 12 月)是 action-controllable generative world model,可生成多样 3D 环境,用于训练 embodied agents。Physical Intelligence 在 2026 年 4 月推出 π0.7 steerable robot foundation model。OpenAI 的 Sora 视频生成模型在 2026 年 4 月停用,说明生成式视频世界模型面对商业逆风,验证了 AMI 的非生成式押注——但也说明,即便资金充足的生成式路线,也未能获得广泛商业采用。AMI 的具体执行缺口很实质:(1)AMI Labs 自身尚未公开发布任何产品或模型;(2)V-JEPA 仅限约 10 秒视频片段,更长时域推理仍是开放研究问题;(3)action-conditioned planning module 只存在于 LeCun 2022 年论文的设计概念中;(4)截至 run date,尚无 AMI Labs 署名论文发表,仍依赖此前 Meta FAIR 论文;(5)没有公开分享任何把 AMI 方法与 NVIDIA Cosmos、Genie 2 或 Physical Intelligence 在任一任务上比较的 benchmarks;(6)视觉之外的多模态传感器整合仍是愿景,未被展示;(7)技术许可模型需要大型工业伙伴愿意试点未经验证的研究阶段 AI,这可能限制早期签约速度。Epoch AI 数据显示,frontier AI 训练算力每年增长 5×——意味着 AMI 必须持续扩张,才能与资源充足的 incumbents 竞争。公司 $1.03 billion 融资提供了可观 runway,但商业可行性仍完全未经证明。 [CE017, CE018, CE019, CE020, CE021, CE022]
5.6 展品
06客户
6.1 商业化前阶段与客户格局
截至 2026 年 6 月,AMI Labs 没有付费客户,没有披露收入,也没有生产部署。公司在 2025 年底启动,完成 $1.03 billion 种子轮;CEO Alexandre LeBrun 在公开表态中说得很清楚:“AMI Labs 不是典型的应用 AI 创业公司,不是三个月发产品、六个月有收入、十二个月做到 $10 million ARR。” 这是有意且明示的战略选择,不是疏忽。AMI 在商业许可前,先资助一项多年基础研究议程。 公司的客户策略遵循业内所说的 design-partner model——让早期用户或伙伴组织参与共同开发,在正式商业发布前验证技术。AMI 的使命陈述把“工业流程控制、自动化、可穿戴设备、机器人、医疗及更多领域”列为最终垂直领域;Yann LeCun 在 2026 年 3 月接受 AFP 采访时表示,融资公告后“六到十二个月内可能展开”与企业伙伴的讨论。这个窗口大约覆盖 2026 年 9 月到 2027 年 3 月,但不意味着已经签署商业合同。 投资者基础本身就是早期商业兴趣的 proxy。Toyota Ventures(robotics/automotive)、Groupe Industriel Marcel Dassault(aerospace/industrial)、NVIDIA(hardware and inference infrastructure)、Samsung(devices)、ZEBOX Ventures(CMA CGM logistics)和 Publicis Groupe(creative/media)等战略投资者,各自指向 AMI 可能首先获得 design-partner traction 的行业。但没有公开披露显示 AMI 与其中任何一方签有合同性 design-partner agreement。 [CU001, CU002, CU003, CU004, CU005, CU006]
| 分群 | 买方 / 用户 / 付款方角色 | 主要用例 | 规模 / 体量 | 对 AMI 的战略价值 | 当前状态 | 证据缺口 |
|---|---|---|---|---|---|---|
| 医疗(临床 AI) | 医院系统 CTO/CMO,临床 AI 平台(例如 Nabla) | 可获 FDA 认证的自主文档生成、智能体式临床工作流 | 通过 Nabla 触达 190+ 医疗机构;Nabla 服务 150+ 医疗系统 | 旗舰概念验证;FDA 认证锚点;LeBrun 的领域专长 | 仅有设计合作伙伴访问;没有商业许可证;技术尚未部署 | AMI 技术整合时间表未披露;FDA 路径未定义 |
| 工业过程控制 / 制造 | 制造商的工程或运营 VP、自动化负责人 | 传感器密集环境建模(喷气发动机、化工厂、钢厂) | 大型工业企业;可能以每家公司 $10B+ 收入目标为主 | 首个商业许可证目标;用专有传感器数据验证世界模型 | 潜在机会;没有确认的接触;仅有投资人信号(Dassault、Toyota) | 未披露设计合作伙伴协议;数据共享条款未定 |
| 机器人 | 机器人工程负责人、自动化研发经理 | 面向实体机器人的动作条件规划;可靠任务执行 | 机器人 OEM、物流自动化公司、人形机器人开发商 | 验证 JEPA 在具身 AI 中的适用性;区别于 NVIDIA Isaac | 潜在机会;没有确认的接触;ZEBOX/Toyota Ventures 释放兴趣信号 | 没有与既有方案(Isaac、Physical Intelligence)的基准对比 |
| 可穿戴设备 | 设备 OEM(Samsung、Garmin)产品和 AI 团队 | 持久记忆环境智能;设备端边缘推理 | 消费电子 OEM;Samsung 是战略投资人 | 设备端世界模型证明算力效率;与医疗形成互补 | 潜在机会;仅有 Samsung 投资人信号;没有披露产品路线图 | 未宣布设备整合时间表或芯片合作伙伴 |
| 主权 AI / 政府相关企业 | 国家 AI 计划办公室、非美国辖区的企业 IT 决策者 | 欧洲和亚洲 AI 主权——获取非美国、非中国前沿模型 | 欧盟、新加坡、日本的政府机构和大型国有关联企业 | 验证 AMI 的战略定位;可能成为公共部门锚点 | 潜在机会;Temasek、Bpifrance、法国机构投资人释放兴趣信号 | 未披露政府合同或招标;主权 AI 叙事仍属推测 |
状态分类只反映公开披露证据。「潜在机会」表示投资人匹配信号显示可能接触,但不存在合同关系。规模引用来自第三方估计或 Nabla 自身发布指标,不是 AMI 商业数据。所有缺口反映截至 2026-06-22 尚不可得的信息。
[CU005, CU006, CU008, CU009, CU014, CU021]| 指标 | 数值 | 日期 / 期间 | 来源 | 置信度 | 含义 | 缺失分母 / 缺口 |
|---|---|---|---|---|---|---|
| 付费客户数量 | 0(收入前) | 截至 2026-06-22 | AMI 管理层公开声明 | 高 | CEO 确认仍处商业化前阶段 | 未披露管线规模或合格线索数量 |
| 设计合作伙伴协议(已披露) | 1(Nabla) | 2025 年 12 月至今 | Nabla 新闻稿;TechCrunch | 高 | 唯一真实世界反馈回路;没有多元化 | 可能存在未披露合作伙伴,但尚未验证 |
| 年度经常性收入(ARR) | 截至 2026-06-22 | 未披露 | 高 | 无法评估 SaaS 指标;收入模型尚未形成 | 未披露定价模型、合同结构或费率 | |
| 收入运行率 | 截至 2026-06-22 | 未披露 | 高 | AMI 已明确推迟收入 | 最早商业许可证可能也要到 2027 年中期 | |
| 医疗合作伙伴部署覆盖(Nabla,代理指标) | 190+ 医疗机构;100,000+ 临床医生 | 截至 2026 年 6 月 | Nabla 主页 | 中 | 若合作成熟,显示 AMI 最终渠道触达 | AMI 目前没有从 Nabla 部署中获得收入 |
| 预期首次企业合作伙伴讨论(管理层预测) | 2026 年 3 月融资完成后 6–12 个月 | 2026 年 3 月 | LeCun 对 AFP / France24 的表述 | 中 | 仅为指示性时间线;讨论 ≠ 已签协议 | 初始公告后没有更新或确认 |
截至运行日期,AMI 所有商业指标为 null 或 0,反映其收入前状态。Nabla 代理指标是 Nabla 自身披露数字,只作为渠道背景;AMI 不从 Nabla 现有客户群获得收入。null 字段置信度为高,因为无收入来自管理层声明确认,而非仅仅无法验证。
[CU001, CU003, CU041, CU044]从早期研究伙伴合作到商业授权的示意路径,展示各阶段及其证据质量。截至 2026 年 6 月,除设计伙伴访问外,后续阶段尚未达成。
旅程阶段状态仅反映已披露证据。设计伙伴与试点之间的边界为示意;AMI 或 Nabla 均未披露里程碑时间表。
[CU003, CU005, CU010, CU011, CU037]6.2 Nabla 作为首个设计伙伴:医疗入口
Nabla 是 AMI 唯一公开点名的伙伴,也是公司最早的真实世界导向证据;但它不是商业收入证据。2025 年 12 月的新闻稿把该安排描述为“排他性战略合作”,Nabla 获得 AMI 新兴世界模型技术的“首批访问权”。Nabla 的既定目标,是成为“第一家把可获 FDA 认证的 agentic AI systems 带入医疗的公司”——这一定位完全依赖 AMI 成功构建可认证的世界模型技术,而结果还在多年之后。 Nabla 的规模为 AMI 技术未来触达医疗市场提供了重要背景。截至 2026 年 6 月,Nabla 服务超过 190 家医疗组织和 150+ 个 health systems and provider groups,覆盖 35+ languages,超过 100,000 名临床医生使用其 ambient documentation platform。这些客户属于 Nabla,不属于 AMI。Nabla 的临床部署没有收入流向 AMI;当前 AMI-Nabla 安排的性质是共同开发访问权,而非许可费。 Nabla 合作带有显著结构性纽带:LeBrun 在成为 AMI CEO 前,是 Nabla 的联合创始人和 CEO;LeCun 自 Nabla 创立以来就是投资人和顾问。LeBrun 仍担任 Nabla 的 Chairman and Chief AI Scientist。这种深层关系让 AMI 真正获得早期医疗买方需求和临床场景 LLM 局限的情报,但也意味着关系不是 arm's-length,不能作为独立商业 proof point。 关键是,Nabla 当前的 AI 系统完全基于 LLM。AMI 的世界模型技术仍处于研究阶段,没有 production-ready API 或产品。AMI 世界模型实际整合进 Nabla 临床平台的时间线未披露;而医疗中自主 agentic AI 的 FDA 监管路径,远比现有 narrow diagnostic AI tools 的 510(k) 路径复杂且缓慢。 [CU010, CU011, CU012, CU013, CU014, CU015]
| 名称 | 关系类型 | 分群 / 用例 | 部署状态 | 结果证据 | 证据局限 |
|---|---|---|---|---|---|
| Nabla | 独家设计合作伙伴(非付费客户) | 医疗——可获 FDA 认证的智能体式临床 AI | 设计合作伙伴访问;技术尚未在生产中部署 | 190+ 医疗机构、100K+ 临床医生(Nabla 自有基础);均非来自 AMI 模型 | AMI 世界模型尚未整合;合作关系是结构性的(LeBrun 是 Nabla 董事长) |
| Meta(LeCun 表示有兴趣) | 潜在未来客户(公司表示有兴趣,无合同) | 面向消费设备的物理世界 AI(Ray-Ban Meta 智能眼镜) | 无协议;仅为潜在机会 | LeCun 表示「Meta 可能是我们的第一个客户」 | 推测性;表述兴趣不等于谈判或承诺 |
| Groupe Industriel Marcel Dassault(投资人) | 战略投资人;潜在工业设计合作伙伴 | 航空航天 / 制造——传感器密集工业过程建模 | 仅为投资人关系;未披露设计合作伙伴协议 | 参与种子轮显示商业匹配 | 股权投资不产生成为客户的义务 |
| Toyota Ventures(投资人) | 战略投资人;潜在机器人 / 汽车设计合作伙伴 | 汽车 / 机器人——面向自主系统的动作条件世界模型 | 仅为投资人关系;未披露设计合作伙伴协议 | 参与种子轮显示汽车 AI 兴趣 | Toyota Motor(客户身份)不同于 Toyota Ventures(投资人身份) |
| ZEBOX Ventures(投资人,CMA CGM 物流基金) | 战略投资人;潜在物流 / 运输设计合作伙伴 | 物流 / 工业——供应链 AI、港口自动化、航运协调 | 仅为投资人关系;未披露设计合作伙伴协议 | CMA CGM 关联关系显示物流用例兴趣 | AMI 未披露航运或物流产品路线图 |
本表仅列出已披露关系。「关系类型」区分设计合作伙伴(Nabla:正式协议)与潜在 / 投资人信号驱动的机会(其他所有主体)。任何一行都不对应商业收入。Nabla 关系是唯一有公开合同基础的关系;其他关系都由投资人参与推断。
[CU005, CU010, CU011, CU012, CU013, CU016]截至 2026 年 6 月,展示已知 / 估算意向方、已披露设计伙伴、已确认试点和商业授权的漏斗。漏斗严重集中在顶部的意向方; 除设计伙伴外,后续阶段尚未进入。
“战略投资者”数量根据已披露融资轮财团估算。并非所有投资者都是设计伙伴候选;该数量用作温热商业兴趣的上限代理。 截至 2026 年 6 月运行日期,所有下游漏斗阶段均确认为零。
[CU001, CU003, CU008, CU009, CU041, CU042]6.3 目标买方群体与采购路径
AMI 的目标客户画像,是处在数据密集、安全关键环境中的企业买方;在这些环境里,LLM 幻觉会带来实质运营或监管风险。LeCun 描述的典型用例,是一家飞机发动机制造商希望建立覆盖数千个传感器的整体模型,以优化效率并预测故障——这是工业制造垂直领域的买方,其设备生命周期长达数十年,供应商评估流程也常跨多年。医疗买方(医院系统、医疗器械制造商、Nabla 这类 health AI platforms)同样具备安全关键特征,但采购治理更严格,受 FDA clearance 要求、临床工作流整合和责任风险驱动。 机器人领域中,相关买方角色是制造或物流公司的自动化或工程负责人,他们需要可靠的实体机器人 action-prediction 能力——目前这类买方由 NVIDIA Isaac robotics platform,或 Physical Intelligence、1X Technologies 等公司服务。可穿戴设备领域中,买方很可能是 device OEM(Samsung、Garmin、Oura),希望获得设备硬件层可高效运行的 persistent-memory ambient intelligence。 无论哪种情况,AMI 世界模型的采购路径都需要:(1)design partner agreement,让客户以提供领域数据和评估反馈为交换,访问早期 model checkpoints;(2)pilot phase,可能持续 6–18 个月,在客户传感器环境中测试世界模型;(3)基于试点结果进行商业许可谈判。这个多步骤流程意味着,即便今天最早介入的伙伴(Nabla,2025 年 12 月参与),距离签署商业许可也还有数年。 除 Nabla 外,没有公开证据显示正式 design-partner agreements 已披露。Toyota Ventures、Dassault、ZEBOX、Samsung 和 SEA 出现在投资 syndicate 中,与商业协同一致,但股权投资并不迫使任何一方成为客户或设计伙伴。 [CU021, CU022, CU023, CU024, CU025, CU026]
| 指标 | 数值 | 分群 | 置信度 | 尽调问题 |
|---|---|---|---|---|
| 净收入留存(NRR) | 所有分群 | 高(缺失已确认) | 在未来投资人更新中索取;目前没有客户可留存 | |
| 毛收入留存(GRR) | 所有分群 | 高(缺失已确认) | 同 NRR——收入前 | |
| 流失率 | 所有分群 | 高(缺失已确认) | 未定义;没有活跃订阅 | |
| 合同期限 / 续约条款 | 所有分群 | 高(缺失已确认) | 未来尽调;商业模式尚未成形 | |
| Nabla 设计合作伙伴延续信号 | 持续(合作伙伴公开声明,2026 年 3 月) | 医疗 | 中 | 向 Nabla 确认共同开发里程碑是否已达成 |
| 客户满意度 / NPS | 所有分群 | 高(缺失已确认) | 不适用;没有可调研的付费客户 |
AMI 没有付费客户,因此所有留存和满意度指标都是 null。基于管理层声明,null 值是已确认的不存在(不是数据缺失)。Nabla 设计合作伙伴延续信号是唯一可用的关系稳定性代理,但它基于公开声明,不是正式续约或里程碑文件。
[CU041, CU042, CU044]对每个已披露或投资者释放信号的关系,在六个尽调维度上给证据质量打分。评分从 0(无证据)到 3(确认 / 强)。 所有关系在生产部署、结果证据和收入上均为 0。
质量评估基于已发表证据,采用分类口径(无 / 低 / 中等 / 正式 / 零)。所有行均确认不存在生产部署和收入。 独立性评级反映 AMI 领导层与伙伴 / 投资方实体之间的结构性关系。
[CU005, CU011, CU012, CU016, CU025, CU026]6.4 采购摩擦、监管壁垒与采用时间线
多重结构性力量会把 AMI 的 go-to-market 时间线拉长到标准 frontier AI 商业化周期之外。 医疗领域中,根据 Menlo Ventures 2025 State of AI in Healthcare survey,即便最激进的买方,现在处理 AI 采购决策也平均需要 6.6 个月(低于传统 IT 的 8 个月)。但这种加速适用于 production-ready ambient scribing 和 billing automation tools,不适用于需要 FDA clearance 的研究阶段世界模型。FDA 已 cleared 的 AI/ML medical devices(截至 2026 年初超过 900 个)主要是 narrow diagnostic detection tools,通过 510(k) 或 De Novo pathways 获批。Nabla 设想的自主 agentic clinical AI,很可能需要复杂得多的监管接触——最可能是 PMA(pre-market approval)或新的 De Novo pathway,从实质提交起,时间线通常跨 2–5 年。 Menlo Ventures 还记录到,医疗买方明确优先考虑“技术成熟度”和“能在规模化环境中可靠运行、可直接投产的解决方案”——AMI 目前达不到这个标准。上一阶段的“试点致死”动态虽然对已验证 AI 工具有所缓解,但会继续困住 world models 这类未经验证的架构。Goldman Sachs research 指出,累计 AI capex 已达 $1 trillion,但部署型 AI 收入有限;Sequoia Capital 更新版 “AI's $600B question” 分析量化了截至 2024 年 GPU 基础设施支出与实际终端用户价值之间的缺口。这些结构性观察适用于更广泛的 AI 市场,但也直接影响客户采用未经验证架构的意愿。 工业 / 制造市场中,复杂 process control systems 的企业软件销售周期,即便是成熟供应商也常常需要 12–24 个月。AMI 还面对额外障碍:它需要客户共享专有运营数据(喷气发动机遥测、工厂传感器流)来训练世界模型,由此引发数据主权和 IP ownership 担忧;Futurum Group 和独立分析师都把这些问题标记为“AI 驱动交易中最先被审查的问题之一”。这种数据共享前提制造了鸡生蛋动态:AMI 需要领域数据来构建有用模型,但客户没有经验证的 ROI 就不会共享数据。 BMJ 的 TRIPOD+AI 临床 AI 报告框架强调,任何临床预测模型在进入 health system 部署前,都必须满足严格的外部验证、预先指定 calibration 和 bias assessment 要求——AMI 尚未发表的世界模型还无法回应这些标准。 [CU031, CU032, CU033, CU034, CU035, CU036]
根据管理层表述、设计伙伴模式类比以及医疗 / 工业采购周期,估算 AMI 关键商业里程碑区间。所有区间都是推断估算, 不是 AMI 披露的目标。
所有数值均为分析师估算,来自管理层表述,以及可比前沿 AI 和医疗 AI 商业化先例。低端估算假设最佳执行; 高端估算假设安全关键应用的典型企业采购和监管周期。这些并非 AMI 披露的预测。
[CU003, CU032, CU034, CU037, CU038]6.5 扩张、集中度风险与证据缺口
AMI 已披露的全部客户接触面,是单一非付费设计伙伴(Nabla),在当前阶段形成极端集中度风险。由于没有商业收入,传统集中度指标(top-customer share of ARR、NRR、GRR)均未定义。真正有意义的集中度风险是结构性的:如果 Nabla 合作未能产出可运行的世界模型整合,AMI 就失去唯一真实世界医疗反馈回路。如果 Nabla 的 FDA 认证之路停滞——考虑到 agentic clinical AI 的新颖性,这在结构上很可能发生——医疗垂直入口会无限期推迟。 融资本身也制造扩张压力:投资者以 $3.5 billion 投前估值为一家研究阶段实验室付费,下一轮融资时就会期待商业 traction 证据。LeBrun “years, not quarters” 的表述诚实,但会在 Series A 时制造困难动态;Series A 很可能在种子轮交割后 18–24 个月到来(约 2027 年中到 2027 年底)。到那时,AMI 至少需要证明:(a)可运行的世界模型架构和已发表 benchmarks;(b)至少一个活跃 design-partner pilot,能生成真实领域数据;(c)围绕 Nabla 或另一家医疗伙伴的可信 regulatory filing strategy。 AMI 设想的 land-and-expand model——把世界模型技术许可给一个垂直领域或伙伴,再扩展到相邻买方——需要一个成功的首个锚点。如果任何 AMI 世界模型技术整合进 Nabla,Nabla 的 190+ health organization 部署足迹会提供真实广度,但目前这仍是假设。工业和机器人垂直领域则没有已点名锚点。 无法报告 customer count、ARR、NRR 或 GRR;由于 pre-revenue 状态,这些指标都未定义。本章无法区分 “zero customers” 与 “undisclosed early customers”,因为 AMI 没有提交需要披露收入的 Form D 或 SEC disclosure,全部商业状态证据都来自管理层公开表态。 [CU041, CU042, CU043, CU044, CU045, CU046]
| 扩张驱动因素 / 集中度风险 | 当前状态 | 潜在影响 | 严重程度 | 尽调路径 |
|---|---|---|---|---|
| 单一已披露合作伙伴(Nabla)集中 | 100% 设计合作伙伴关系都来自 Nabla;没有多元化 | 如果 Nabla 整合失败或延迟,AMI 会失去唯一医疗反馈回路 | 高 | 确认是否有其他设计合作伙伴讨论正在进行;在 Series A 时向 CEO 索取管线更新 |
| Nabla 的 FDA 路径依赖 | Nabla 目标是可获 FDA 认证的智能体式 AI;需要整合 AMI 世界模型 | 自主临床 AI 的 FDA 路径可能需要 3–7 年;会阻断近期收入 | 高 | 跟踪 FDA 关于自适应 AI 的新兴预提交指引;索取 Nabla 的 510(k) vs. PMA 策略 |
| 投资人与合作伙伴重叠(非公平交易) | LeBrun 同时是 AMI CEO 和 Nabla 董事长;LeCun 是 Nabla 投资人 | 关联方安排可能削弱合作伙伴里程碑评估的客观性 | 中 | 要求独立技术顾问委员会验证 Nabla 整合里程碑 |
| 没有具名锚点进入工业垂直领域 | 尽管有 Dassault / Toyota 投资人信号,仍没有确认的工业设计合作伙伴 | 延迟收入多元化;如果 Nabla 是唯一路径,会形成医疗单一路径依赖 | 高 | 确认是否有工业试点讨论正在推进;跟踪 Dassault 或 Toyota 公开公告 |
| 研究阶段的 land-and-expand 模式未验证 | AMI 的扩张模式依赖先把首个设计合作伙伴转为商业许可证,再扩大规模 | 没有可运行的商业模板,扩张到第二个垂直领域就没有先例 | 中 | 索取预期 Series A 里程碑,以及 AMI 从设计合作伙伴模式转向商业模式的门槛 |
| LLM 既有厂商挤出设计合作伙伴 | OpenAI、Microsoft、Google 补贴 AI 采用,以在医疗和工业领域抢份额 | 在 AMI 世界模型可用于生产之前,大型既有厂商可能先向 Nabla 客户提供相邻方案 | 中 | 监控 Nabla 的竞争挤出风险;跟踪 Epic 和 Microsoft Nuance AI 与 Nabla 现有医疗系统客户的合作 |
所有严重性判断均基于公开证据和行业惯例。「高」严重性表示存在单点故障或结构性障碍,下一次融资前需要管理层直接处理。AMI 目前没有客户多元化选项,因为公司尚无付费客户。
[CU037, CU038, CU039, CU040, CU041, CU042]6.6 展品
07风险
7.1 执行与商业化风险
AMI Labs 在 2026 年 3 月按 $3.5 billion 投前估值完成 $1.03 billion 种子轮;当时没有收入、没有已部署产品,并自称采用多年研究优先路线图。LeCun 本人在 2026 年 1 月 MIT Technology Review 采访中表示,实现人类水平 AI 需要“重大概念突破”,而且“明年或后年不会发生”。公司智识旗手的这句坦白,把执行风险框定得很清楚:全部价值主张押在一次科学范式转移上,而按创始人自己的说法,时间点无法确定。即便条件有利,从世界模型研究到可商业部署的许可产品之间的产品化缺口,行业分析师和历史先例估计也需要三到七年。 AMI 的开源发表承诺——写在使命陈述中,也被 LeCun 反复强调——与其未来许可模型形成额外张力。提前发表核心架构洞见,意味着资源更充足的竞争者(Meta FAIR、Google DeepMind、NVIDIA)可以比 AMI 更快吸收并实现 AMI 的研究。Gary Marcus 在 Substack 分析中认为,alternative-to-LLM AI 路线缺乏具体商业证据是一种模式,不是异常;这类投机性估值带有实质 dud risk。 [CR001, CR002, CR017, CR018, CR019, CR021]
| 风险 | 可监控触发信号 | 阈值 / 事件 | 行动含义 |
|---|---|---|---|
| 研究到产品缺口 | 没有任何获许可的商业产品 | 到 2028 年底仍无创收部署 | 论点破裂:AMI 无法把研究商业化;考虑减仓 |
| 关键人物离职(LeCun) | LeCun 降低参与度或公开批评 AMI 方向 | LeCun 退出执行董事长角色或降为顾问 | 立即重新评估估值支撑;通过公开声明监控 |
| 种子轮后资本失败 | AMI 未能在种子轮后 24 个月内完成 Series A | 到 2028 年 Q1 未宣布新一轮融资 | 跑道风险升高;要求直接披露 burn rate |
| EU AI Act GPAI 执法行动 | EU AI Office 对 AMI 的 GPAI 合规展开调查 | 正式通知或罚款 | 声誉和运营风险重大;解决前暂停投资 |
| 被 hyperscaler 世界模型竞争替代 | Google DeepMind 或 NVIDIA 先于 AMI 发布生产级世界模型 | 已商业部署、能力与 AMI 路线图相当的世界模型 | 重新评估差异化论点;AMI 护城河收窄,只剩主权叙事 |
| 开源 IP 侵蚀 | AMI 在许可收入建立前发布核心 JEPA 模型权重 | 向社区发布开放权重模型及训练数据 | 许可收入模型被结构性削弱;收入路径不清楚 |
触发信号和阈值是定性尽调框架;均非 AMI 的正式契约。行动含义是投资分析指引,不构成法律义务。
[CR002, CR009, CR003, CR025, CR034, CR023]7.2 技术、研究与竞争风险
AMI 押注 JEPA(Joint Embedding Predictive Architecture)会接替 LLM,这在科学上站得住脚——LeCun 的立场论文以及 I-JEPA/V-JEPA 论文都说明其研究根基很强——但商业化路径明显比基于 LLM 的产品更长。训练算力成本每年增长 4–5×(Epoch AI,2024),前沿世界模型训练预计需要超过 10^25 浮点运算的 exascale 级算力运行,而这正是 EU AI Act 推定系统性风险的门槛。NVIDIA 在 AI 训练芯片上的近乎垄断,意味着 AMI 在运营上依赖 NVIDIA 的定价和配额决策; NVIDIA 参与 AMI 种子轮投资就是证据——这层关系带来供给通道,也带来利益绑定风险。 竞争压力极端:Google DeepMind Genie 2、NVIDIA Cosmos、Meta FAIR 后续项目、Physical Intelligence(π)和 World Labs 都在争夺重叠技术地带,而且资产负债表更厚、迭代循环更快。DeepSeek 等中国开源模型说明,接近世界模型的能力可能很快商品化。 Sequoia 的「$600B question」分析直接适用于 AMI:hyperscaler 的 GPU 资本开支越来越商品化,任何依赖算力的 AI 产品定价权都会被侵蚀。 Goldman Sachs 也指出,预计 $1 trillion 的 AI 资本开支迄今几乎没有产生可衡量 ROI。 [CR007, CR008, CR024, CR034, CR035, CR036]
| 失效模式 | 可能性 | 严重性 | 缓释成熟度 | 剩余敞口 | 未解决缺口 |
|---|---|---|---|---|---|
| NVIDIA 算力供应中断 — GPU 配额削减或价格飙升 | 中 | 严重 | 低 — 单一来源依赖;NVIDIA 是共同投资方(利益一致),但没有供应保证 | 高 | 没有公开协议保证按合同价格获得算力 |
| 前沿训练运行失败或成本超支 — 世界模型预训练超过预算 | 中 | 高 | 低 — AMI 未公开披露算力成本治理框架 | 高 | 未披露初始模型的 capex 计划或算力成本上限 |
| 人才流失 — 关键研究科学家离职(Meta 出走之后) | 中 | 高 | 中 — 假定有股权激励;hyperscaler 可能给出竞争性反报价 | 中 | 未披露留任合同;股权归属时间表未知 |
| 数据管线失效 — 世界模型训练缺少足够的未标注传感器 / 视频数据 | 中 | 高 | 低 — 除 Nabla 外,未公开披露合作伙伴数据协议 | 高 | 数据合作覆盖范围和质量未知 |
| 网络安全 / IP 外流 — 专有模型权重或研究成果被盗 | 低 | 严重 | 未知 — 未披露安全认证或审计 | 中 | 研究 IP 基础设施未披露 SOC 2 / ISO 27001 认证 |
可能性和严重性是定性估算,来自分析师评论、同业公司披露以及缺少公开缓释证据。所有单元格均反映外部分析;AMI 未披露运营风险控制。
[CR007, CR008, CR013, CR024]截至 2026 年 6 月,对 AMI Labs 主要风险的发生可能性与严重性进行评估。风险为作者基于监管来源、分析师报告和上市公司声明收集证据后的定性估算。
所有位置均为作者基于本章证据作出的定性估算。未进行定量风险建模。单元格表示该格最具代表性的单一风险; 多个风险可能落在同一“可能性—严重性”象限。
[CR002, CR007, CR009, CR013, CR023, CR025]7.3 监管与法律风险
AMI Labs 面对一套分层、跨司法辖区的监管栈。作为在法国(EU 成员国)注册并设总部的通用 AI(GPAI)模型开发商, AMI 明确落在 EU AI Act(Regulation 2024/1689)范围内;该法案于 2024 年 3 月 13 日由欧洲议会以 523 票赞成通过。 GPAI 模型提供商必须履行透明度义务,包括发布详细训练数据摘要、遵守 EU 版权法、落实模型文档要求;这些义务在该法案生效后 12 个月(2025 年 8 月)开始适用。根据第 51 条,若 GPAI 模型训练累计计算量超过 10^25 FLOPs,即被推定存在系统性风险, 进而触发模型评估、系统性风险缓释计划和事件报告等额外义务。 AMI 规划中的医疗和机器人应用属于 EU AI Act 的高风险 AI 类别,需要合格评定、质量管理体系和上市后监测。在美国, AI 赋能医疗器械需要 FDA 510(k) 许可或上市前批准(PMA);FDA 于 2021 年 1 月发布 AI/ML Action Plan,并持续搭建自适应 AI 医疗软件监管框架。美国 Bureau of Industry and Security(BIS)负责的出口管制限制向特定国家转让先进 AI 芯片和相关技术; 若地缘政治恶化,AMI 的欧洲算力基础设施会承受供应链风险。截至 2026 年 6 月,AMI 未披露未决诉讼或 IP 纠纷,但 LeCun 与 Meta 之间没有公开的竞业限制或 IP 转让协议,这让 Meta FAIR 资助下发展出的 JEPA 架构存在潜在 IP 归属模糊。 [CR025, CR026, CR027, CR028, CR029, CR030]
| 规则 / 许可 / 案件 | 司法辖区 | 状态 | 可能性 | 严重性 | 缓释措施 | 剩余敞口 | 尽调路径 |
|---|---|---|---|---|---|---|---|
| EU AI Act 第 51 条 — GPAI 系统性风险义务(>10^25 FLOPs 门槛) | EU | GPAI 透明度规则 2025 年 8 月生效;系统性风险规则 2026 年 8 月生效 | 高 | 严重 | 建立合规计划;对接 EU AI Office;发布训练数据摘要 | 高 — AMI 的前沿模型训练很可能突破 10^25 FLOP 门槛 | 要求说明 GPAI 合规状态;审计算力运行估算 |
| EU AI Act — 高风险 AI 合格评定(医疗、机器人) | EU | 新高风险系统的义务自 2026 年 8 月起适用 | 高 | 高 | 与公告机构合作取得 CE 标志;建立质量管理体系 | 中 — 合格评定成本和周期会把产品发布推迟 1–2 年 | 要求提供 Nabla 合作产品的合格评定路线图草案 |
| FDA 510(k) / PMA 对 AI 医疗器械的许可 | US | 监管框架已生效;FDA AI/ML 行动计划 2021 仍在推进 | 高 | 高 | 及早对接 FDA;将 Nabla 临床试点设计为研究性器械研究 | 中 — 许可周期 12–36 个月,且需要临床证据 | 核实任何临床 AI 功能的 FDA 预提交会议状态 |
| US Export Administration Regulations — 先进 AI 芯片转让管制 | US/Global | BIS EAR 已生效;限制特定国家和终端用户 | 中 | 高 | 通过美国云服务商(AWS、Azure、GCP)采购算力;避开受限国家的数据中心 | 中 — 若美欧贸易条件恶化,风险可能升级 | 确认 AMI 的 GPU 采购渠道;核实 BIS 许可证豁免是否适用 |
| GDPR — 为临床 AI 处理特殊类别健康数据 | EU | GDPR 第 9 条对健康数据的限制已生效 | 高 | 中 | 数据处理协议;本地部署或 EU 主权云;为临床产品开展 DPIA | 低-中 — 标准医疗数据治理控制已经存在 | 审查 Nabla DPA 条款及 AMI 的数据处理责任 |
| 潜在 IP / 竞业风险 — LeCun JEPA IP 与 Meta 雇佣协议 | France/US | 未知 — 条款无公开披露 | 中 | 高 | 获取 IP 归属法律意见;寻求 LeCun/Meta 的公开声明或确认 | 高 — JEPA 架构在 Meta FAIR 资助下开发;IP 转让不清楚 | 尽调 JEPA 的 IP 权属链;取得 LeCun 竞业限制陈述 |
行按严重性排序(严重优先)。GPAI 门槛和合规日期来自欧洲议会 2024 年 3 月通过的 EU AI Act(Regulation 2024/1689)。FDA 许可周期为行业估算。IP 风险行基于缺少公开披露,而非已确认争议。
[CR025, CR026, CR027, CR028, CR029, CR030]AMI Labs 所依赖的关键外部要素(算力、资本、合作伙伴、监管机构、人才),以及这些要素传导至公司运营和商业化能力的路径。
依赖图仅代表公开已知关系。未纳入私下协议、数据合同和未披露合作。
[CR008, CR013, CR030, CR031, CR032, CR041]7.4 关键人物、治理与融资风险
AMI Labs 的治理结构形成两条不同的关键人物风险。LeCun 担任执行董事长,而非 CEO;他明确把自己的角色定义为战略性而非运营性, 保留 NYU 教职并继续常驻纽约。公司总部在巴黎,CEO Alex LeBrun 同时担任 Nabla(AMI 锚定临床合作伙伴)的董事长兼首席 AI 科学家。 这种双重角色安排意味着,最具知名度的科学权威是兼职,运营负责人又同时向另一家实体承担职责。LeCun 2025 年 11 月在 Meta 任职 12 年后离开,原因是与 Zuckerberg 在 AI 战略上存在分歧;这说明高知名度 AI 科学家一旦出现理念错位,可能脱离机构承诺, AMI 面临不小的复发风险。 融资风险很实质。$1.03B 种子轮、$3.5B 投前估值建立在一条短期没有收入里程碑的研究时间线上。Anthropic、Mistral、Cohere 等前沿 AI 实验室为维持类似的重算力研究项目,通常每 18–24 个月再融资一轮。AMI 在 2 年研究期内几乎肯定需要种子轮后的资本。 Sequoia 的分析把算力定义为定价权坍塌的商品,这意味着 AMI 后续每一次训练运行都会同时面对绝对成本上升(每年增长 4–5×) 和回报压缩。Goldman Sachs 也独立指出前沿 AI 支出存在 ROI 缺口。 [CR009, CR010, CR011, CR012, CR013, CR014]
| 角色 / 职能 | 依赖或缺口 | 可能性 | 严重性 | 缓释措施 | 尽调路径 |
|---|---|---|---|---|---|
| 执行董事长(Yann LeCun) | 兼职;保留 NYU 教授职位;常驻纽约;非 CEO | 高 | 严重 | 将投入时间正式写入协议;确认 IP 归属 | 谈判明确时间分配;确认竞业限制和 IP 转让 |
| CEO(Alex LeBrun) | 双重角色:AMI CEO + Nabla 董事长 / 首席 AI 科学家 | 中 | 高 | 正式划清 AMI 与 Nabla 之间的时间分配 | 审查 LeBrun 的雇佣协议;评估利益冲突治理 |
| 首席科学官(Saining Xie) | NYU 教授;兼职;关键技术领导均为学术 / 兼职身份 | 中 | 高 | 转为全职,或任命专职全职研究负责人 | 明确 Xie 在 AMI 的雇佣状态;确认研究团队深度 |
| 首席研究与创新官(Pascale Fung) | 常驻香港;远离巴黎总部和纽约运营 | 低 | 中 | 远程协作基础设施;访问安排 | 确认 Fung 对 AMI 的全职等价投入 |
| 世界模型副总裁(Michael Rabbat) | 前 Meta;对世界模型研究领导的连续性至关重要 | 中 | 高 | 留任方案;股权激励 | 核实雇佣合同和股权归属条款 |
角色来自 Observer 和 TechCrunch 报道(2026 年 3 月)。雇佣条款和时间分配承诺未公开披露。严重性反映对各职能单点故障风险的评估。
[CR009, CR010, CR011, CR013, CR014, CR016]有向无环图展示 AMI 主要风险因素如何传导至收入、运营、融资和估值结果。
DAG 结构代表作者对风险传导的因果模型;不代表量化概率模型。边权重相等且为名义设定。
[CR002, CR007, CR009, CR023, CR025, CR034]7.5 地缘政治与主权风险
AMI Labs 明确围绕地缘政治叙事搭建:成为有别于美国主导和中国主导模型的「第三条路径」欧洲主权 AI 替代方案。LeCun 在 MIT Technology Review 2026 年 1 月访谈中阐述了这一愿景,称许多国家希望掌握 AI 主权控制权,而一家开源欧洲冠军会回应这种需求。 法国总统 Macron 也公开背书 AMI 的巴黎总部。这一叙事带来特定风险:公司的战略定位依赖 EU 政治意愿、资本和监管友好度持续存在, 而这些条件都可能变化。 美国对先进 AI 芯片的出口管制(由 BIS 依据 Export Administration Regulations 执行)形成结构性供应链风险:如果对 EU 实体的芯片出口限制收紧, 或 EU-US 贸易关系恶化,AMI 的训练算力供给可能中断。多司法辖区运营(巴黎、纽约、蒙特利尔、新加坡)带来复杂监管敞口, GDPR 义务、美国数据本地化偏好和新加坡数据治理规则之间可能冲突。OECD AI Principles 与 EU AI Act 叠加,形成一张合规义务拼图, 一家尚未有收入的公司也必须穿过去。主权 AI 叙事还带来集中风险:如果 AMI 无法证明技术优势,或欧洲 AI 资金枯竭, 叙事就会坍塌,公司也会失去主要战略差异化。 [CR041, CR044, CR045, CR046, CR043, CR031]
| 依赖项 | 交易对手 | 角色 | 集中度 | 失效情景 | 严重性 | 缓释措施 | 剩余敞口 |
|---|---|---|---|---|---|---|---|
| AI 训练算力 | NVIDIA | 主要 GPU 供应商和共同投资方 | 严重 | 供应削减、价格飙升、出口限制 | 严重 | 多云备援;AMD ROCm;EU 算力合作 | 高 |
| 锚定临床合作 | Nabla | 独家世界模型访问合作伙伴和首个客户 | 高 | Nabla 转向、放弃 AMI 合作;LeBrun 利益冲突 | 高 | 正式合作协议;独立的 AMI 商业管线 | 中 |
| 资本供给 | Cathay Innovation / Greycroft / HV Capital / NVIDIA / Bezos / Schmidt / Cuban 等投资方 | 种子轮投资方;后续融资把关方 | 高 | 若里程碑未达成,后续融资被拒;下调估值融资风险 | 高 | 投资方基础多元;围绕里程碑沟通 | 中 |
| 研究基础设施和人才管线 | FAIR (Meta) 校友网络 | 创始团队和研究文化 | 高 | Meta 反向挖人;LeCun 声誉变化;人才战升级 | 高 | 股权留任;围绕 AMI 专属研究声望定位 | 中 |
| EU 监管和政治支持 | 法国政府 / EU Commission | 政治背书;潜在公共资金 | 中 | 政府优先级变化;EU 监管负担增加 | 中 | 对接 EU AI Office;申请 Horizon Europe 资金 | 低 |
交易对手角色和失效情景依据公开公告及可比前沿 AI 实验室结构推断。集中度评级为定性判断。除 Nabla 合作外,没有正式依赖协议公开披露。
[CR008, CR013, CR034, CR038, CR041, CR044]7.6 附录
08估值
8.1 估值背景与可比公司组
AMI Labs 2026 年 3 月种子轮按 $3.5 billion 投前估值定价,融资 $1.03 billion 后投后估值为 $4.53 billion。PitchBook 确认这是欧洲有记录以来最大种子轮。交割时公司约有十几名员工,没有产品,且自己给出的时间线按年而非季度计算——这让估值几乎纯粹押注于科学可信度和战略期权价值, 而不是收入或近期现金流。 最接近的结构性可比,是商业规模形成前就被定价的精英创始人前沿 AI 实验室。Fei-Fei Li 的 World Labs 于 2024 年 8 月以 $1 billion 估值融资 $230 million;这在当时已被视为产品前融资中的大额交易;到 2026 年初,随着 Marble 3D 世界模型产品发布, World Labs 据称正讨论按 $5 billion 估值融资。Andreessen Horowitz 领投 World Labs Series A,并用一份宣言把空间智能定义为下一代 AI 前沿,这与 AMI 的 JEPA 定位形成直接智识平行。Mira Murati 的 Thinking Machines Lab 种子轮估值约 $12 billion, 体现市场给予前 OpenAI CTO 履历的溢价。DeepSeek 据称在 2026 年 6 月以 $50 billion-plus 估值融资 $7.4 billion, 但 DeepSeek 已发布生产模型并产生可衡量使用量,与种子阶段的 AMI 截然不同。世界模型初创公司 Odyssey 在 2026 年 6 月 18 日当周完成 $310 million 融资,Crunchbase 周度交易摘要将其列为当周领跑交易,说明到 2026 年中,资本仍在大规模流向世界模型子赛道。 作为公开市场锚点,C3.ai 报告 2025 财年(截至 2025 年 4 月 30 日)总收入 $389.1 million,同比增长 25.3%,其中订阅收入 $327.6 million;这为判断 AI 软件平台达到商业成熟时的样子提供了有用的 SaaS 可比,也提醒我们 AMI 当前研究阶段距离这种变现水平还有很长距离。 PitchBook 引述 NVIDIA 对 AI 初创公司的累计投资约 $53 billion、覆盖 170 笔交易,这凸显战略资本涌入该领域的规模——也包括 NVIDIA 自己参与 AMI 这一轮。 投资人给 AMI 定价的逻辑,与「主权 AI」溢价和人才稀缺溢价一致。AMI 的投资财团由 Cathay Innovation、Greycroft、Hiro Capital、 HV Capital 和 Bezos Expeditions 联合领投,NVIDIA、Samsung、Temasek、Toyota Ventures 等战略方参与,横跨美国、欧洲、亚洲主权资本和战略资本。 $3.5B 投前估值高于 World Labs 初始 $1B 估值,但明显低于 Thinking Machines Lab 的 $12B 种子轮;这说明市场对 AMI 更长技术周期、 以及押注 JEPA 架构而非 LLM 邻近路径的特定风险,给出了校准后的折价。[CV001, CV002, CV003, CV004, CV005, CV006]
| 公司 | 阶段 / 状态 | 估值或上一轮融资 | 收入 / 产品状态 | 与 AMI 的相关性 | 可比对象的关键局限 |
|---|---|---|---|---|---|
| World Labs (Fei-Fei Li) | 私有;2024 年 8 月 Series A → 据称 2026 年初正按 $5B 洽谈 | Series A 估值 $1B;2026 年初估算约 $5B | 2025 年 11 月推出 Marble 3D 世界模型 | 最接近的架构同类——空间世界模型,高端创始人溢价 | 已有上线产品;AMI 还没有 |
| Thinking Machines Lab(Mira Murati)实验室 | 私营;2024–2025 年种子轮 | 种子轮估值约 $12B | 截至 2026 年中无公开产品 | 产品前 AI 实验室的明星创始人溢价基准 | 邻近 LLM,但不是世界模型;技术假设不同 |
| Odyssey | 私营;2026 年 6 月 Series B | $310M Series B(2026 年 Q2 完成) | Odyssey-2、Starchild-1 等世界模型产品已在部署 | 世界模型同类;与 AMI 有直接战略重叠 | 收入未知;总部在美国,不同于 AMI 的欧洲定位 |
| DeepSeek | 私营;2026 年 6 月融资 | 按 $50B+ 估值融资 $7.4B | R1 和 V4-Pro 已投产;有可衡量使用量 | 有主权雄心的前沿研究实验室 | 已有生产模型和收入信号;AMI 仍是产品前阶段 |
| AMI Labs(标的) | 私营;2026 年 3 月种子轮 | 投前估值 $3.5B | 无产品、无收入、无商业客户 | — | — |
| Mistral AI | 私营;欧洲前沿 LLM 实验室 | 2024 年 Series B 约 $6B;2025 年后轮次估值估计更高 | 商业 API 和产品;收入为正 | 欧洲主权 AI 定位相似 | 基于 LLM;架构不同于 JEPA 世界模型 |
| SpAItial | 私营;种子轮 | €13M 种子轮(按欧洲种子轮标准异常大) | 公布时处于产品前阶段 | 欧洲世界模型相邻初创公司的空间 AI 种子轮基准 | 规模小得多;产品重点不同 |
| C3.ai(NYSE: AI)上市公司 | 上市公司 | 市值波动;FY2025 收入 $389M | FY2025 总收入 $389.1M(同比增长 25.3%);上市 AI 软件公司 | 大规模 AI 软件变现的公开市场锚点 | 已有收入;AMI 处于收入前阶段——比较凸显差距,而不是估值同等 |
私营公司估值来自新闻和分析师报告中的已披露轮次或分析师估计。收入数字如有列示,来自公开申报文件或新闻稿。C3.ai 收入来自其 10-K(截至 2025 年 4 月 30 日期间)。
[CV001, CV002, CV003, CV004, CV005, CV006]8.2 投资论点与反方论点
正向论点靠四根柱子支撑:科学可信度、架构差异化、战略可选性和主权溢价。LeCun 的 Turing Award 履历和十年执掌 Meta FAIR 的经历, 构成一个罕见、已大幅去风险的科学领导画像。JEPA 在抽象表示空间做预测,而不是生成像素或 token 层级预测;这是一条可测试的差异化假设, 且已有公开研究记录。投资财团的战略构成——尤其是 NVIDIA、Temasek、Samsung、Toyota Ventures——表明具备长期规划周期的产业买家相信, 世界模型 AI 会在 AMI 的运营窗口内进入其供应链和产品。主权 AI 维度(欧洲总部、跨大陆投资人组合、Bpifrance 参与)让 AMI 能接触公共资金流和受监管行业采购管道, 这些是仅在美国运营的同行拿不到的。 反方论点同样有力。AMI CEO LeBrun 在发布时说得很清楚:这不是一家 6 个月内有收入的典型应用 AI 初创公司。公司会先做多年研究, 再进入商业化。Goldman Sachs 对 AI 资本开支的分析估计,约 $1 trillion 的计划中 AI 基础设施投资迄今在收入或可衡量生产率收益上「little to show」。 Sequoia 更新后的分析把「$200B question」扩大为「$600B question」——也就是 AI 基础设施 CapEx 与支撑这些支出所需收入之间的缺口——但尚未给出答案。 如果这种资本开支到收入的结构性缺口是 LLM 邻近 AI 的风险,那么世界模型的风险更重,因为它还多了一层架构赌注:JEPA 路线必须在 AMI 目标应用垂直领域跑赢多模态 LLM,而且要赶在那些资源多 AMI 数个数量级的 Big Labs 收敛到类似架构之前。 CB Insights 2026 年 AI 100 显示,能建立持久业务的垂直 AI 公司,护城河来自数据——要么是非文本数据、稀缺数据,要么深度嵌入受监管工作流的数据。 AMI 设想的垂直领域(制造、医疗、机器人)符合这一论点,但 AMI 还没有这些数据、工作流或客户。从研究实验室走到拥有数据护城河的企业级在位者, 路很长,历史上通常需要 hyperscaler 赞助者或非常长的 runway。[CV011, CV012, CV013, CV014, CV015, CV016]
| 维度 | 评估 | 依据 |
|---|---|---|
| 建议 | 跟踪 | 没有产品或收入;估值偏高,但技术假设在科学上可信 |
| 信心 | 低 | 商业证据接近于零;研究优先的公司,公开时间线为多年 |
| 风险评级 | 高 | 关键人物、技术、竞争、算力成本和资本风险相互交织 |
| 估值立场 | 偏高 | $3.5B pre-money 已计入大量期权价值,但当前没有收入锚 |
| 决策含义 | 监控;没有 JEPA benchmark 结果和 >1 个商业锚定客户前,不在 Series A 承诺投资 |
评估反映截至 2026 年 6 月的公开证据。除 Nabla 合作外,公司没有收入、没有产品,也没有披露商业合同。
[CV025, CV026, CV027, CV028, CV029]| 维度 | 看多论点 | 看空反驳 | 什么会改变判断 |
|---|---|---|---|
| 科学领导力 | LeCun 图灵奖 + FAIR 履历;团队可信度独特 | 品牌不保证商业化;DeepMind 先例显示,研究到收入需要 10 年以上 | JEPA 在公开 benchmark 上跑赢多模态 LLM |
| 架构 | JEPA 通过在表征空间预测,避开生成式幻觉 | Big Labs 正在向类似方法收敛;差距可能在 AMI 商业化前更快收窄 | 同行论文无法以同等保真度复现 JEPA 结果 |
| 主权溢价 | 欧洲主权 AI 定位能吸引受监管买方和公共资助方需求 | 算力供应链(NVIDIA 芯片)削弱真正的技术独立性 | EU 国家 AI 采购合同授予 AMI |
| 投资方构成 | NVIDIA、Samsung、Temasek、Toyota 释放工业部署意图 | 若 Big Lab 替代方案出现,战略投资方可能转向,削弱后续融资信号 | 领投战略投资方在 Series A 增持 |
| 估值 vs 可比公司 | 相比 Thinking Machines Lab 的 $12B、World Labs 接近 $5B,$3.5B 不算高 | World Labs 已有部署产品(Marble);AMI 没有——比较对 AMI 有利 | 没有产品却以 $8B+ 定价完成 Series A,说明泡沫加重 |
| 收入时间线 | 一旦收入启动,工业 B2B 定价逻辑可支撑高倍数 | CEO 称商业化需要「数年」;收入前 burn rate 会要求多轮融资 | 首个商业客户以重大 ACV($5M+)签约 |
论点来自公开分析师报告(Futurum、Goldman Sachs、Sequoia)和公司声明。截至 2026 年 6 月,所有看空反驳均无法从公开来源直接评估。
[CV011, CV012, CV013, CV014, CV015, CV016]决策链从 AMI 的证据画像出发,穿过风险和估值因素,落到最终的“Track”建议。
逻辑链是解释性综合;边权重为定性。
[CV025, CV026, CV027]8.3 情景分析:牛市、基准与熊市
三种情景主要区别在三个变量:JEPA 商业可用性的推进速度、Big Labs 的竞争反应、以及能否以不低于 $3.5B 投前底价拿到后续资本。 三种情景都假设 AMI 需要 $1.03B 种子轮之外的追加资本;原因是同类实验室在前沿规模下每年算力和人才 burn 估计为 $300–500M, 而 Epoch AI 记录的算力每年 4–5× 增长,会让任何固定算力预算在多年研究周期内不够用。 牛市情景下,JEPA 到 2028 年实现可展示的真实世界 benchmark,例如在工业过程控制或自主机器人中,相比多模态 LLM 形成可复现优势; AMI 随后按 $10–15 billion 估值完成 Series A。与两家或更多产业投资人(Toyota Ventures、Samsung、NVIDIA)的战略合作在 2029–2030 年转化为已部署用例。IPO 在 2031–2033 年具备可行性,前提是收入达到 $200M+ ARR;这意味着 20–25× 收入倍数, 与 2025–2026 年前沿 AI 公司群体的高端溢价一致。 基准情景下,AMI 用 2–3 年产出高质量开放研究,以基本持平到小幅上修的估值($5–8B)融资 $500M– $1B Series A, 并把应用焦点收窄到一两个能证明早期客户防御性的垂直领域。商业产品在 2029–2031 年到来。退出最可能通过 hyperscaler 以 $8–15 billion 战略收购实现——若届时公司拥有独特架构、主权叙事一致的投资人基础,并以 50–100 名员工做到 $40–60M ARR, 这一结果具备可信度。 熊市情景下,JEPA 架构赌注在实践中没有形成差异化:Big Labs 发布有竞争力的世界模型系统,开源替代品扩散,企业买家看不到购买理由。 到 2028 年,资本市场对收入前 AI 研究收紧。AMI 难以按种子轮估值融资,被迫大幅收窄范围、接受会稀释主权叙事架构的战略投资, 或以困境条件出售。$1–2 billion 困境退出或关闭,是不利结果。[CV019, CV020, CV021, CV022, CV023, CV024]
| 情景 | 关键假设 | 估值 / 回报逻辑 | 退出或里程碑时间线 | 关键风险 | 概率信号 |
|---|---|---|---|---|---|
| 牛市 | 2028 年前 JEPA benchmark 突破;两个工业部署;Series A 估值 $12–15B | 2031–2033 年 IPO 或战略出售,估值 $30–50B;按 $3.5B 进入价计算回报 8–14× | 2031–2033 年 IPO,ARR $200M+;收入倍数 20–25× | JEPA 结果无法在生产规模复现 | 低至中;需要达成具体 benchmark 里程碑 |
| 基准 | 研究继续 2–3 年;Series A 估值 $5–8B;聚焦单一垂直商业化;被 hyperscaler 收购 | 2029–2032 年以 $8–15B M&A 退出;按 $3.5B 进入价计算回报 2–4×;回报温和 | Series A 2027–2028;2030–2032 年被收购 | 资本市场收紧;Series A 估值持平或下降 | 中;符合前沿 AI 实验室先例 |
| 熊市 | JEPA 相比 LLM 改进没有差异化;Big Labs 推出竞争性世界模型;2028 年资金缺口 | 困境出售或清算,估值 $1–2B;进入价亏损 | 2028–2030 年被迫退出,估值低于种子轮 | 需要多个相互交织的失败同时发生 | 低至中;需要宏观恶化和 JEPA 失败 |
情景来自分析师估算和同业实验室先例,并非专有财务模型。收入数字仅为示意;AMI 未披露 ARR。回报倍数假设进入估值为 $3.5B。
[CV019, CV020, CV021, CV022, CV023, CV024]示意关键估值驱动因素对 AMI 隐含企业价值的影响;每个驱动因素在不同假设组合下给出低到高区间。
数值为不同假设下隐含退出企业价值的示意性 USD 百万美元,不是财务模型输出。基准情景隐含价值约为 $8–15B(收购); 牛市情景约为 $30–50B(IPO)。敏感性条表示每个维度低假设与高假设之间的估算差距。
[CV019, CV020, CV021]以 $3.5B 投前种子轮入场估值为锚,给出三种情景下的退出企业价值区间。熊市假设困境结果;基准假设被超大规模云厂商收购; 牛市假设规模化后 IPO。
所有数值均为 USD 百万美元。熊市:$1–2B 区间;基准:$8–15B 区间;牛市:$30–50B 区间。入场 $3.5B 是 2026 年 3 月投前估值。 这些是分析师构建的情景边界,不是财务模型输出。
[CV022, CV023, CV024]8.4 建议与估值立场
当前价格下建议立场是 TRACK;在上调前,应对具体过关里程碑继续 RESEARCH-MORE。估值立场是 STRETCHED:$3.5B 投前估值可以被解释为对 LeCun 科学项目和欧洲主权 AI 定位的一份期权,但相对于公开可核验信息(无产品、无收入、除 Nabla 外无商业合作、无 JEPA benchmark 结果), 其中计入了大量乐观预期。 由于商业证据几乎完全缺席,且 JEPA 技术假设高度不确定,本建议的信心为 LOW。风险评级为 HIGH:研究优先的时间线、LeCun 关键人物集中(执行董事长而非全职 CEO)、 算力成本升级、Big Lab 竞争压力交织在一起,形成难以对冲的风险。 下方 KPI 评分框架概括了投委会在七个维度上的定位:市场、技术证明、竞争护城河、单位经济、风险组合、估值和证据质量。 在现有证据下,除市场机会外,没有任何维度得分高于「中等」。 破题触发项(详见附录)包括:任何证据显示 JEPA 在 AMI 目标应用类型上无法跑赢多模态 LLM;Yann LeCun 不再担任执行董事长; 种子轮交割后 24 个月内未能以不低于 $5B 投前估值完成 Series A;或 Nabla 合作关系解散且没有替代方。[CV025, CV026, CV027, CV028, CV029]
投资委员会综合评分覆盖七个维度,采用 0–10 分制。基于当前公开证据,没有任何维度超过 6/10。
评分为 0–10 分制的定性判断(10 = 最佳)。市场机会因可触达 TAM 和政策顺风而得分较高。技术证明、单位经济和证据质量得分较低, 原因是公司仍处于产品前阶段,且缺少已发布的 JEPA 基准。
[CV025, CV026, CV028, CV029]8.5 退出成熟度、尽调问题与融资可选性
截至 2026 年 6 月,AMI 尚未具备退出条件。IPO 至少需要已验证收入和可复制 go-to-market;两者都不存在。结合 Futurum 分析师评估、 公司自己的公开表述,以及同类实验室先例(DeepMind 2014 年被 Google 收购,而非 IPO),近期到中期最可信的流动性路径是被 hyperscaler 战略收购, 最可能发生在 2029–2033 年窗口。 潜在收购方因战略契合而区分:Samsung 和 Toyota Ventures(现有投资人,设备 / 工业契合)、Apple(空间智能是 Vision Pro 和机器人战略核心)、 Microsoft(主权 AI 计算基础设施),或由法国 / 欧洲国家冠军出手收购以保留主权维度。Meta 仍是自然的技术合作方,但鉴于 LeCun 离开时的摩擦, 作为收购方存在冲突。 $1.03B 种子轮按前沿实验室 burn 率($300–500M/年)估计可支撑 2–3 年 runway,这意味着 AMI 需要在 2027 年末或 2028 年回到市场融资 Series A。 这一轮会成为真正的市场出清估值测试:研究产出能否支撑从 $3.5B 上调,还是倍数压缩。种子轮进入的投资人应把研究发表速度和 JEPA benchmark 结果, 作为 Series A 定价的主要领先指标。 下方附录列出的尽调问题,是建议参与 Series A 共同投资前的最低证据标准。最关键的问题,是独立验证 JEPA 在标准化物理世界 benchmark 上相对 state-of-the-art 多模态 LLM 的性能优势;截至 2026 年 6 月,公开资料无法回答这个问题。[CV030, CV031, CV032, CV033, CV034]
| 触发条件 | 阈值 | 对投资假设的传导 | 行动含义 |
|---|---|---|---|
| JEPA 基准失败 | 多模态 LLM 在目标应用基准(机器人 / 工业控制)上追平或超过 JEPA | 架构差异化消失;AMI 只能靠团队竞争 | 下调信心;暂停 Series A 跟投 |
| Yann LeCun 离职 | LeCun 退出执行董事长角色,或实质性减少参与 | 关键人物风险兑现;品牌溢价坍塌 | Series A 硬性止步;按替代团队质量重新评估 |
| Series A 定价低于 $3.5B | 下一轮机构融资低于当前投后估值($4.53B) | 市场给出信号:AMI 未能证明进展;新投资人出现逆向选择 | 按可得条款退出;不要继续持有并承受后续稀释 |
| 2028 年前除 Nabla 外无商业伙伴 | 到 2028 年底,AMI 未能宣布第二个付费或正式研究合作伙伴 | 表明工业买家尚未被说服;AMI 声称的垂直场景没有转化 | 降低敞口;重新评估基准情景 |
| 主权 AI 资金收缩 | 法国政府或欧盟 AI 资助项目大幅削减;Bpifrance 退出;欧盟 AI 战略转向 | AMI 的主权 AI 溢价缩水;监管顺风转为逆风 | 监测政策信号;调整退出时间表 |
触发条件基于分析师风险分析(Futurum、Goldman Sachs、Sequoia)和前文风险评估。AMI 披露有限,目前没有公开的正式监测触发条件。
[CV030, CV031, CV032, CV033]| 主题 | 缺失证据 | 为何重要 | 负责人 / 尽调路径 |
|---|---|---|---|
| JEPA 基准结果 | 没有公开基准比较 JEPA 与多模态 LLM 在工业任务(机器人、过程控制、可穿戴设备)上的表现 | 决定架构押注是已有实证验证,还是仍属投机 | 申请查看技术演示;跟踪论文发表;对比 World Labs / Odyssey 发布节奏 |
| 算力支出和烧钱拆分 | 除“算力和人才”外,没有公开募资用途细节;成本结构未披露 | 跑道建模的关键;每年 $300–500M 的同类估计是外推值,并未确认 | 向 CFO 索取;对照本轮 NVIDIA 算力承诺验证 |
| Nabla 合作商业条款 | AMI-Nabla 合作结构(股权、IP、排他性、费用)未公开披露 | 决定首个商业锚点是真实收入期权,还是关联方安排 | 索取条款清单或意向书;核验关系独立性 |
| Series A 投资人沟通 | 尚无信号显示谁预计领投 Series A,或预期估值是多少 | 决定基准情景估值底线($5–8B)是否可达 | 跟踪投资人情绪;对照 Thinking Machines Lab 的 Series A 时间线 |
| 人员计划和人才管线 | 种子轮完成时仅公开确认“约十几名”员工;未披露人员目标 | 没有临界规模的世界模型研究员,研究实验室质量会下滑 | 索取招聘计划;验证 LinkedIn 增长;对标 World Labs 人员扩张轨迹 |
尽调问题是给出中等信心的 Series A 共同投资建议前所需的最低证据集。所有项目均来自分析师评论和公开报道缺口。
[CV034, CV035, CV036]8.6 附录
免责声明
本报告基于截至运行日期(2026-06-22)的公开来源生成。不构成投资建议。AMI Labs 是私营公司;财务预测和估值估算为分析性推断, 并非经审计披露。所有 USD/EUR 换算均采用底层来源在公告时引用的汇率。
证据索引
| 编号 | 陈述 | 可信度 | 来源 |
|---|---|---|---|
| CO001 | AMI Labs stands for Advanced Machine Intelligence Labs; the company's full legal name is Advanced Machine Intelligence Labs. | 高 | SO001, SO004 |
| CO002 | AMI Labs was co-founded in late 2025 by Yann LeCun and Alexandre LeBrun following LeCun's departure from Meta. | 高 | SO003, SO004 |
| CO003 | AMI Labs is headquartered in Paris, France. | 高 | SO001, SO005 |
| CO004 | The company operates across four international locations from its founding: Paris, New York, Montreal, and Singapore. | 高 | SO001, SO002 |
| CO005 | AMI Labs' mission is to build intelligent systems that understand the real world; it believes real intelligence starts in the world, not in language. | 高 | SO001, SO004 |
| CO006 | AMI Labs plans to license its world model technology to industry partners rather than building consumer or enterprise software directly. | 中 | SO007, SO005 |
| CO007 | AMI Labs is a pre-revenue research lab with no plans to generate revenue in the near term; the company is explicitly not a typical applied AI startup. | 高 | SO005, SO010 |
| CO008 | AMI's name is intentionally pronounced 'a-mee' — the French word for 'friend' — as noted by LeCun. | 中 | SO004 |
| CO009 | AMI Labs is building world models based on LeCun's Joint Embedding Predictive Architecture (JEPA), which learns abstract representations of real-world sensor data rather than predicting in output space. | 高 | SO001, SO025 |
| CO010 | World models based on JEPA make predictions in abstract representation space rather than pixel or token space, learning what matters about how the world changes. | 高 | SO025, SO018 |
| CO011 | Action-conditioned world models allow agentic systems to predict the consequences of their actions and plan multi-step action sequences subject to safety guardrails. | 高 | SO001, SO025 |
| CO012 | AMI's core thesis is that generative architectures trained by self-supervised learning are unsuitable for unpredictable, continuous, high-dimensional real-world sensor data. | 高 | SO001, SO009 |
| CO013 | AMI Labs' primary target application domains include industrial process control, factory automation, healthcare, robotics, and wearable devices. | 高 | SO001, SO005 |
| CO014 | AMI Labs has committed to publishing research openly and releasing code as open source, planning to build a community and research ecosystem. | 中 | SO001, SO005 |
| CO015 | The JEPA concept was proposed by LeCun in a 2022 position paper titled 'A Path Towards Autonomous Machine Intelligence'; I-JEPA was published in 2023 at ICCV. | 高 | SO025, SO018 |
| CO016 | Yann LeCun is AMI Labs' Executive Chairman and co-founder; he explicitly is not the CEO. | 高 | SO004, SO005 |
| CO017 | LeCun shared the 2018 ACM A.M. Turing Award with Yoshua Bengio and Geoffrey Hinton for conceptual and engineering breakthroughs in deep neural networks. | 高 | SO019, SO022 |
| CO018 | LeCun was VP and Chief AI Scientist at Meta (formerly Facebook) for more than 12 years, leading the FAIR research organization, before departing in November 2025. | 高 | SO007, SO014 |
| CO019 | LeCun retains his NYU professorship as Jacob T. Schwartz Professor of Computer Science, Data Science, Neural Science, and Electrical and Computer Engineering at the Courant Institute. | 高 | SO022, SO007 |
| CO020 | Alexandre LeBrun is AMI Labs' CEO and co-founder. | 高 | SO003, SO004 |
| CO021 | LeBrun previously co-founded Nabla (healthcare AI for clinical documentation) and Wit.ai (a natural-language startup sold to Facebook in 2015). | 高 | SO003, SO008 |
| CO022 | LeBrun previously worked at Meta's FAIR division under LeCun's leadership, creating a direct professional link between the two co-founders. | 高 | SO004, SO014 |
| CO023 | Saining Xie is AMI Labs' Chief Science Officer (CSO) and co-founder. | 高 | SO023, SO005 |
| CO024 | Saining Xie previously served as a research scientist at Google DeepMind and before that at Meta's FAIR; he remains a faculty member at NYU. | 高 | SO023, SO010 |
| CO025 | Saining Xie co-created Diffusion Transformers (DiT), a generative framework that powers many leading generative AI systems including Sora; his research has been cited more than 90,000 times. | 中 | SO023 |
| CO026 | Pascale Fung is AMI Labs' Chief Research and Innovation Officer (CRIO) and a professor at the Hong Kong University of Science and Technology. | 高 | SO005, SO011 |
| CO027 | Michael Rabbat is AMI Labs' VP of World Models; he is a leading researcher in world models who joined from Meta. | 高 | SO005, SO010 |
| CO028 | Laurent Solly is AMI Labs' COO; he previously served as Meta's VP for Europe, departing Meta in December 2025. | 高 | SO005, SO004 |
| CO029 | AMI Labs announced a seed round of $1.03 billion USD (approximately €890 million) on March 10, 2026. | 高 | SO005, SO002 |
| CO030 | The seed round values AMI Labs at $3.5 billion pre-money (approximately €3 billion). | 高 | SO005, SO009 |
| CO031 | The AMI Labs seed round is Europe's largest seed round on record per PitchBook data cited by multiple independent sources. | 高 | SO005, SO014 |
| CO032 | The seed round was co-led by Cathay Innovation, Greycroft, Hiro Capital, HV Capital, and Bezos Expeditions. | 高 | SO002, SO005 |
| CO033 | Strategic investors in the round include NVIDIA, Samsung, Sea, Temasek, and Toyota Ventures. | 高 | SO002, SO011 |
| CO034 | Individual angel investors include Jeff Bezos, Mark Cuban, Eric Schmidt, Tim and Rosemary Berners-Lee, Jim Breyer, Xavier Niel, and Mark Leslie. | 高 | SO002, SO005 |
| CO035 | French institutional investors include Bpifrance Digital Venture, Association Familiale Mulliez, Groupe Industriel Marcel Dassault, Publicis Groupe, Aglaé Lab, Artémis, and ZEBOX Ventures. | 高 | SO002, SO015 |
| CO036 | The Financial Times reported in December 2025 that AMI Labs was seeking to raise €500 million at a €3 billion pre-money valuation before its website had even launched. | 中 | SO003, SO004 |
| CO037 | The final AMI Labs raise of approximately €890M significantly exceeded the initially reported €500M target, likely reflecting strong investor demand. | 中 | SO005 |
| CO038 | LeCun departed Meta in November 2025 after more than 12 years; the Observer reported it followed disagreements with Mark Zuckerberg over the future of AI. | 高 | SO014, SO007 |
| CO039 | AMI Labs was publicly confirmed in December 2025 via a Nabla press release disclosing LeBrun's CEO role and LeCun's LinkedIn post confirming his Executive Chairman role. | 高 | SO006, SO003 |
| CO040 | AMI Labs' public website (amilabs.xyz) launched in January 2026, publishing the company's JEPA world model mission statement. | 中 | SO004, SO007 |
| CO041 | French President Emmanuel Macron publicly welcomed AMI's decision to headquarter in Paris, pledging government support. | 高 | SO004, SO012 |
| CO042 | LeBrun has explicitly stated that AMI Labs could take years to produce commercial applications; it is not a typical startup with six-month product and twelve-month ARR timelines. | 高 | SO005, SO010 |
| CO043 | Nabla is AMI Labs' only publicly disclosed commercial partner as of the run date; it will receive privileged access to AMI's world models as they are developed. | 高 | SO006, SO005 |
| CO044 | The Nabla-AMI partnership arose from LeBrun's transition: Nabla's board approved his CEO-to-AMI move in exchange for privileged world model access. | 高 | SO004, SO006 |
| CO045 | LeBrun indicated that the AMI seed round attracted strong interest from industrial players and potential partners, hinting at further commercial alliances beyond Nabla. | 中 | SO005 |
| CO046 | Futurum Group's independent analysis identifies structural tension between AMI's research-first mandate and investor expectations set by a billion-dollar raise, creating difficult follow-on financing dynamics. | 中 | SO010 |
| CO047 | At the time of the seed round announcement, AMI Labs employed approximately 12 people and had no product; the company had been operating for only a few months. | 中 | SO010 |
| CO048 | The claim that world models categorically solve LLM hallucination problems is contested by independent analysts; world models face their own generalization challenges in novel environments. | 中 | SO010 |
| CO049 | AMI Labs faces competitive threats from Google DeepMind, Meta FAIR (now under Meta Superintelligence Labs), Physical Intelligence, and other labs working on world models and embodied AI. | 中 | SO010, SO017 |
| CO050 | LeCun publicly criticized Meta's restructuring around Alexandr Wang, calling him 'inexperienced,' indicating continued tension between LeCun and his former employer. | 中 | SO014 |
| CM001 | AMI Labs declares its market boundary as physical-world AI infrastructure — AI systems that understand, predict, and act within continuous, noisy, high-dimensional real-world environments rather than manipulating discrete text tokens. | 高 | SM023, SM009 |
| CM002 | AMI explicitly excludes pure text, language, and generative AI applications from its market; LeCun describes LLMs as competitors' domain and a structural dead-end for physical-world understanding. | 高 | SM009, SM015 |
| CM003 | AMI's four declared application domains are: industrial process control, factory automation and robotics, healthcare diagnostics and clinical workflows, and smart wearable devices. | 高 | SM015, SM019, SM009 |
| CM004 | Status-quo substitutes for world model technology include conventional SCADA and PLC industrial control systems, rule-based clinical decision support software, teleoperation-based robotic systems, specialized computer vision models, and LLM-based agentic frameworks. | 中 | SM009, SM010 |
| CM005 | The global AI market was estimated at USD 390.9 billion in 2025 and projected to grow to USD 3,497.3 billion by 2033 at a CAGR of 30.6% (Grand View Research 2026). | 中 | SM001 |
| CM006 | The global AI in healthcare market is projected to reach USD 194.79 billion by 2031 from USD 36.67 billion in 2026, at a CAGR of 39.7% (MarketsandMarkets 2026). | 中 | SM002 |
| CM007 | The global Robotics AI Software market is estimated at USD 21.0 billion in 2025 and projected to reach USD 70.8 billion by 2032 at a CAGR of 19.0%, driven by convergence of AI, industrial automation, simulation platforms, and physical AI systems (MarketsandMarkets July 2026). | 中 | SM003 |
| CM008 | The global AI market was estimated at USD 601.93 billion in 2026 and projected to reach USD 3,638.08 billion by 2033 at a CAGR of 29.3% (MarketsandMarkets June 2026). | 中 | SM004 |
| CM009 | The global industrial automation and control systems market was estimated at USD 206.33 billion in 2024 and projected to reach USD 378.57 billion by 2030 at a CAGR of 10.8% (Grand View Research 2026). | 中 | SM006 |
| CM010 | Industrial robot installations in the United States rose 11% year-on-year to reach 38,000 units in 2025, with the food industry surging 30%; the automotive industry remains the largest adopter at 13,500 units (IFR, June 2026). | 高 | SM005, SM006 |
| CM011 | China's annual industrial robot installations reached approximately 295,000 units in 2024, representing a 54% global market share; China's 15th Five-Year Plan (2026–2030) places robotics at the heart of its modern industrial system with physical AI as a core focus (IFR, June 2026). | 中 | SM005 |
| CM012 | The healthcare sector led the global AI market in 2025 with the highest revenue share among all industries, driven by AI adoption in diagnostics, patient care, and operational efficiency (Grand View Research 2026). | 中 | SM001 |
| CM013 | Automotive and transportation AI is expected to grow at the fastest CAGR of 33.2% from 2026 to 2033 among all AI market end-use segments, driven by autonomous vehicles and advanced driver-assistance systems (Grand View Research). | 中 | SM001 |
| CM014 | NVIDIA has positioned Physical AI as a distinct infrastructure category, investing in Isaac simulation, Cosmos world model training, and robotics inference infrastructure, and validated the category at GTC 2026 as the next major AI wave. | 中 | SM007, SM010 |
| CM015 | LeCun frames the market opportunity using the Moravec Paradox: what is easy for humans — perception, navigation, physical manipulation — remains computationally intractable for AI; LLMs are limited to discrete text and cannot truly reason or plan because they lack a world model. | 高 | SM009, SM024 |
| CM016 | The JEPA (Joint Embedding Predictive Architecture) developed by LeCun at Meta FAIR trains AI models to predict in abstract representation space rather than generating pixel-level or token-level predictions; the I-JEPA paper was published in January 2023 and accepted at CVPR. | 高 | SM020, SM009 |
| CM017 | LeCun identifies industrial processes with thousands of sensors — jet engines, steel mills, chemical factories — as a primary world model application because no existing technique builds a complete holistic model of these systems from sensor data. | 中 | SM009 |
| CM018 | AMI claims world models will exhibit persistent memory, reasoning and planning capability, and controllability — properties that LLMs structurally cannot provide due to their token-prediction architecture. | 中 | SM023, SM015 |
| CM019 | AMI's declared revenue model is technology licensing to industry partners rather than direct B2C or enterprise SaaS; the company is positioning as a foundational physical-world AI infrastructure provider. | 中 | SM009, SM010 |
| CM020 | LeCun stated in the January 2026 MIT Technology Review interview that Meta might be AMI's first client, indicating incumbent technology companies are potential buyers for world model APIs. | 中 | SM009 |
| CM021 | Strategic investors in AMI's seed round signal target market segments: NVIDIA (physical AI infrastructure), Toyota Ventures (automotive and industrial robotics), Samsung (devices and wearables), Temasek (Asia-Pacific sovereign and enterprise capital), and Bpifrance (European sovereign AI investment). | 中 | SM013, SM017, SM010 |
| CM022 | Healthcare AI budget ownership lies with hospital CIOs, clinical informatics teams, and healthcare system executives; patient-facing AI requires FDA and EMA regulatory clearance as a prerequisite; North American healthcare AI represents 42.4% of global market share (MarketsandMarkets). | 中 | SM002, SM012 |
| CM023 | Industrial automation enterprise buyers have capital expenditure cycles averaging 3–5 years and high switching costs from incumbent automation platforms including Siemens, GE, and Honeywell; technology adoption in manufacturing is structurally slower than in software-first industries. | 中 | SM006, SM010 |
| CM024 | Robotics OEMs — including industrial robot suppliers and humanoid robot startups — are a potential channel buyer for world model APIs, embedding the technology in robots as a software layer; Toyota Ventures' participation signals automotive and industrial OEM interest. | 中 | SM017, SM009, SM010 |
| CM025 | The Nabla healthcare partnership demonstrates AMI's B2B infrastructure licensing model: AMI provides world model infrastructure while Nabla builds and distributes clinical AI applications to healthcare providers; the partnership is exclusive per the Nabla press release. | 中 | SM014, SM012 |
| CM026 | Global labor shortages across manufacturing, healthcare, and logistics are a structural driver of automation investment; IFR confirms US robot installations rose 11% in 2025 and food industry robot adoption surged 30%, indicating broadening automation demand beyond traditional automotive. | 中 | SM005, SM006 |
| CM027 | China's 15th Five-Year Plan (2026–2030) places robotics at the heart of its modern industrial system and explicitly focuses AI research on physical applications with robots as the main driver of economic growth, creating competitive pressure for Western physical AI investment. | 中 | SM005 |
| CM028 | European AI sovereignty demand creates structural preference for AI infrastructure that does not route through US hyperscaler supply chains or expose data to US cloud jurisdiction; LeCun explicitly frames AMI as 'neither American nor Chinese' in positioning to these buyers. | 中 | SM009, SM021, SM022 |
| CM029 | Bpifrance (the French state investment bank) participated in AMI's $1.03B seed round, reflecting direct French government investment in domestic frontier AI capability as part of a national AI sovereignty strategy. | 中 | SM013, SM021 |
| CM030 | Documented LLM failure modes — hallucination, inability to model physical action consequences, and unreliable temporal reasoning — create growing enterprise demand for more reliable AI architectures in industrial and clinical settings. | 中 | SM010, SM009 |
| CM031 | North America's A3 (Association for Advancing Automation) trade association submitted a 'Vision for a National Robotics Strategy' to US lawmakers at the 2026 Automate Show, advocating for a Federal Robotics Office, national policy coordination, and public-private partnerships to accelerate commercial deployment. | 中 | SM005 |
| CM032 | World models as a commercial product category do not yet exist; the Futurum Group explicitly identifies 'the structural tension between a research-first mandate and investor expectations calibrated to a billion-dollar raise' as the primary commercial risk for AMI. | 中 | SM010 |
| CM033 | Healthcare AI regulatory review, including FDA pre-market submissions for AI/ML medical devices, typically requires 2–7 years from prototype to clearance; the UK's MHRA allocated USD 4.1M to its AI Airlock regulatory sandbox in April 2026 to accelerate AI medical device review. | 中 | SM002 |
| CM034 | Industrial enterprise buyers have multi-year CapEx planning cycles and high switching costs from incumbent automation platforms; Grand View Research's industrial automation CAGR of 10.8% is substantially below the 29–40% CAGR range for pure AI software markets, reflecting this conservatism. | 中 | SM006, SM001 |
| CM035 | Competing physical AI architectures — Google DeepMind Gemini Robotics, NVIDIA Project GR00T, Physical Intelligence, and 1X — are advancing multimodal and embodied AI capabilities that could converge on AMI's target use cases and narrow its architectural window of distinctiveness. | 中 | SM010, SM016 |
| CM036 | LeCun stated publicly in January 2026 that 'nobody, absolutely nobody, knows how to make robots smart enough to be useful,' attributing this to teleoperation-dependent training data that fails to generalize when the environment changes — a statement that simultaneously describes the market opportunity and the technical distance from commercial deployment. | 高 | SM009, SM010 |
| CM037 | The Futurum Group analyst notes that AMI's initial valuation of $3.5B pre-money implies investors are paying primarily for scientific credibility and long-term option value, creating real execution pressure when the next financing round requires evidence of progress. | 中 | SM010 |
| CM038 | No published analyst report independently sizes the world model infrastructure licensing market as a distinct commercial category; all available market estimates encompass generative AI, LLMs, and diverse applications, substantially overstating the market addressable by AMI's specific technology. | 中 | SM001, SM002, SM003, SM004 |
| CM039 | The divergence between headline AI market CAGR (29–31%) and industrial automation market CAGR (10.8%) reflects the dominance of text and generative AI in aggregate AI estimates; physical AI markets where AMI operates grow substantially slower than headline AI market forecasts suggest. | 中 | SM001, SM006 |
| CM040 | LeCun's January 2026 statement that current robots cannot generalize directly contradicts commercial deployment claims from Agility Robotics, Figure AI, 1X, and Boston Dynamics; both positions cannot simultaneously be accurate about current-state robot generalization capability. | 中 | SM009, SM010 |
| CM041 | It is an open question whether AMI's JEPA-based world models will achieve commercial-grade performance on specific vertical tasks before competing architectures from NVIDIA, Physical Intelligence, or LLM providers close the gap in physical-world AI capabilities. | 低 | |
| CP001 | World Labs is a spatial intelligence company founded by Fei-Fei Li to build frontier models that can perceive, generate, reason, and interact with the 3D world. | 高 | SP001, SP002 |
| CP002 | World Labs' Marble product generates spatially consistent, high-fidelity, persistent 3D worlds from multimodal inputs including text, images, video, and 360 panoramas. | 高 | SP001, SP002 |
| CP003 | World Labs launched a public World API in January 2026, providing programmatic access to Marble's 3D world generation capabilities for developers and applications. | 中 | SP001, SP002 |
| CP004 | World Labs published a functional taxonomy of world models in June 2026 classifying models as Renderers, Simulators, and Planners, establishing a competitive framing for the world-model market. | 中 | SP002 |
| CP005 | World Labs' generative 3D approach constructs pixel- or voxel-level content, contrasting with AMI Labs' non-generative JEPA architecture which predicts in abstract representation space, a fundamental architectural divergence. | 高 | SP001, SP010, SP027 |
| CP006 | Odyssey is an AI lab pioneering general world models, with its flagship Odyssey-2 model targeting general-purpose physical accuracy of world modeling. | 中 | SP003, SP004 |
| CP007 | Odyssey's product portfolio includes Odyssey-2 (general world model), Starchild-1 (multimodal interaction), Agora-1 (multi-agent world model), and PROWL (RL adversarial framework for world model improvement). | 中 | SP004 |
| CP008 | Odyssey's Starchild-1 model moves beyond world models that learn only from visual observation toward systems that learn from richer multimodal interaction with the world. | 中 | SP004 |
| CP009 | SpAItial is building physically-grounded world models using its Echo model family, which outputs persistent 3D Gaussian Splatting worlds explorable in real time through a developer API. | 中 | SP005 |
| CP010 | SpAItial was founded in May 2025, making it the most recently founded competitor in the direct world-model startup cluster and significantly earlier stage than World Labs. | 中 | SP006 |
| CP011 | SpAItial's API provides direct access to Echo for agents, tools, simulations, and products requiring generated Gaussian Splat worlds, with Echo-2 announced as the current flagship version. | 中 | SP005 |
| CP012 | Google DeepMind's Genie 1 (February 2024) was the first generative interactive environment trained without supervision on internet videos, a foundation world model at 11 billion parameters. | 高 | SP007, SP009 |
| CP013 | DeepMind's Genie 2 (December 2024) is a foundation world model that generates diverse action-controllable 3D environments for training and evaluating embodied agents, from a single prompt image, capable of sustaining consistent worlds for up to a minute. | 高 | SP007, SP009 |
| CP014 | DeepMind's Gemini Robotics model enables robots of any shape and size to perceive, reason, use tools, and interact with humans across multiple embodiments including ALOHA, Bi-arm Franka, and Apptronik Apollo. | 高 | SP007, SP008 |
| CP015 | Google DeepMind lists Genie 3 as its current world model on its public models page as of June 2026, indicating continued rapid iteration in the world-model research space. | 中 | SP008 |
| CP016 | Meta FAIR's V-JEPA (Video Joint Embedding Predictive Architecture) is a non-generative model that predicts in abstract representation space, the same foundational architectural principle on which AMI Labs is built. | 高 | SP010, SP027 |
| CP017 | V-JEPA was released under a Creative Commons NonCommercial license in 2024 while Yann LeCun was still at Meta, directly establishing open-source JEPA prior art before his departure to found AMI Labs. | 高 | SP010, SP025 |
| CP018 | Following LeCun's departure, Meta AI's headline research as of June 2026 focuses on Muse Spark — a new foundation model — indicating a strategic redirect away from JEPA-centric world models. | 中 | SP011 |
| CP019 | Meta FAIR's I-JEPA codebase remains publicly available on GitHub with over 90,000+ associated research citations, establishing JEPA as open-source prior art accessible to any research institution. | 高 | SP025, SP027 |
| CP020 | Physical Intelligence is a generalist robotics company building vision-language-action (VLA) models that directly control diverse robot hardware through embodied sensorimotor learning. | 高 | SP012, SP013 |
| CP021 | Physical Intelligence launched π0 in October 2024 — the first generalist robot policy trained on cross-embodiment data from 8 distinct robot types — combining internet-scale vision-language pretraining with robot sensorimotor data. | 高 | SP012, SP013 |
| CP022 | Physical Intelligence's π0.7 (April 2026) exhibits compositional task generalization, following language coaching to solve tasks never seen in training and demonstrating emergent cross-embodiment transfer across substantially different robot platforms. | 高 | SP014, SP015 |
| CP023 | Physical Intelligence's π0.7 incorporates a lightweight world model to generate visual subgoal images for each language-defined sub-task, indicating that world models are now embedded in leading robotics foundation model pipelines. | 高 | SP014, SP015 |
| CP024 | Physical Intelligence has an active partner program with real-world robotics companies deploying its models as described in a February 2026 blog post about 'The Physical Intelligence Layer,' establishing commercial traction AMI Labs does not yet have. | 中 | SP015 |
| CP025 | OpenAI's Sora was explicitly positioned as 'a foundation for models that can understand and simulate the real world,' directly making it a world-model competitor before its discontinuation. | 高 | SP016, SP017 |
| CP026 | OpenAI discontinued the Sora web and app experience on April 26, 2026, with the Sora API to follow on September 24, 2026, representing a complete retreat from the video-based world-simulator product space. | 高 | SP016, SP018 |
| CP027 | Sora used a diffusion-based transformer architecture that generates video by removing noise from a noisy baseline — a generative approach structurally different from AMI's non-generative JEPA prediction in abstract representation space. | 高 | SP016, SP017 |
| CP028 | NVIDIA Cosmos is a family of world foundation models trained on 9,000 trillion tokens from 20 million hours of real-world interaction data, available under a permissive open commercial model license. | 高 | SP020, SP022 |
| CP029 | NVIDIA Cosmos 3 supports three physical-AI use cases: vision-language reasoning over real-world scenarios, robotic policy training as a World Action Model backbone, and physics-grounded world simulation for closed-loop evaluation. | 高 | SP020, SP022 |
| CP030 | NVIDIA's Isaac platform provides a full robotics software stack including CUDA-accelerated motion planning (cuMotion), pose estimation (FoundationPose), stereo depth (FoundationStereo), teleoperation, and mobility foundation models (COMPASS). | 高 | SP019, SP021 |
| CP031 | NVIDIA's robotics research labs include GEAR (Generalist Embodied Agent Research, focused on world models and large action models) and the NVIDIA Spatial Intelligence Lab (focused on 3D understanding and spatial AI), directly overlapping with AMI Labs' research mission. | 高 | SP019, SP020 |
| CP032 | NVIDIA's Cosmos partner ecosystem includes 1X, Agility Robotics, XPENG, Uber, and Waabi, giving NVIDIA deep distribution into the physical-AI stack through relationships AMI Labs does not currently have. | 高 | SP019, SP022 |
| CP033 | Wayve is building a general-purpose driving intelligence using an AV2.0 end-to-end embodied AI approach, replacing modular sense-plan-act architecture with a single neural network converting raw sensor data into driving commands without HD maps. | 中 | SP023, SP024 |
| CP034 | Wayve's AV2.0 technology uses self-supervised learning at scale without labeled data and is vehicle-agnostic, mapless, and sensor-agnostic — sharing philosophical DNA with AMI's unsupervised world model approach but targeting automotive exclusively. | 中 | SP024 |
| CP035 | AMI Labs' JEPA architecture avoids generative reconstruction, making it more compute-efficient and potentially more reliable in safety-critical environments compared to diffusion-based or autoregressive generative competitors. | 中 | SP010, SP027 |
| CP036 | AMI Labs' primary differentiation is its non-generative, controllable world model approach targeting industrial and healthcare verticals where LLM hallucinations carry unacceptable physical costs — a segment with no current commercially deployed competitor. | 中 | SP001, SP005, SP012, SP020 |
| CP037 | AMI's primary near-term competitive risk is execution velocity: World Labs, Physical Intelligence, and NVIDIA Cosmos all have deployed products and active developer communities, while AMI has committed to a multi-year research-first timeline with no commercial product. | 中 | SP002, SP015, SP020 |
| CP038 | The open-sourcing of JEPA architectures by Meta FAIR (I-JEPA, V-JEPA) under open licenses creates commoditization pressure on AMI Labs' architectural advantage, as any well-resourced lab can now train a JEPA-based world model from public code. | 中 | SP010, SP025, SP027 |
| CP039 | HuggingFace's LeRobot provides a hardware-agnostic, open-source robotics policy framework with standardized datasets on the Hugging Face Hub, continuously lowering barriers to embodied AI development and eroding proprietary moat. | 中 | SP026 |
| CP040 | World Labs, Odyssey, SpAItial, and DeepMind's Genie all employ generative architectures, while AMI's JEPA approach occupies a technically distinct, non-generative lane with no current commercially deployed competitor in the industrial or healthcare segments. | 中 | SP001, SP004, SP005, SP007 |
| CI001 | AMI Labs closed a $1.03 billion USD (approximately €890 million) seed round, announced on March 10, 2026. | 高 | SI001, SI002, SI003 |
| CI002 | The AMI Labs seed round was set at a $3.5 billion pre-money valuation, implying a post-money valuation of approximately $4.53 billion. | 中 | SI002, SI007 |
| CI003 | PitchBook confirmed the $1.03B AMI Labs raise as Europe's largest seed round on record at time of announcement. | 高 | SI001, SI002, SI003 |
| CI004 | AMI had initially sought approximately €500 million as recently as December 2025; the final close of €890M (~$1.03B) significantly exceeded that target, indicating oversubscription. | 高 | SI004, SI003 |
| CI005 | The AMI Labs seed round was co-led by five investors: Cathay Innovation, Greycroft, Hiro Capital, HV Capital, and Bezos Expeditions. | 高 | SI001, SI002 |
| CI006 | Strategic corporate investors in the AMI seed round include NVIDIA, Samsung, Temasek, Toyota Ventures, and Sea. | 高 | SI001, SI002 |
| CI007 | French institutional and industrial backers include Association Familiale Mulliez, Groupe Industriel Marcel Dassault, Publicis Groupe, Bpifrance Digital Venture, ZEBOX Ventures, Artémis, and Aglaé Lab. | 高 | SI001, SI002 |
| CI008 | Notable angel investors in the AMI seed round include Eric Schmidt, Mark Cuban, Jim Breyer, Tim and Rosemary Berners-Lee, Xavier Niel, and Mark Leslie. | 高 | SI001, SI003 |
| CI009 | CEO Alexandre LeBrun explicitly identified compute and talent as AMI's only two main cost centers for the seed capital, with no mention of marketing, sales, or product spend. | 高 | SI002, SI001 |
| CI010 | AMI Labs has no revenue, no commercial product, and has stated no plans to generate revenue in the near term as of March 2026. | 高 | SI001, SI002, SI003 |
| CI011 | LeBrun stated AMI is 'not your typical applied AI startup' and that 'it could take years for world models to go from theory to commercial applications'. | 高 | SI002, SI003 |
| CI012 | LeCun stated AMI would focus on research and development in its first year, with discussions with corporate partners possible within 6-12 months of the March 2026 announcement. | 中 | SI008, SI009 |
| CI013 | AMI's intended monetization pathway is technology licensing: selling access to domain-specific world model capabilities to industrial and healthcare partners, not consumer or enterprise SaaS. | 中 | SI002, SI014 |
| CI014 | Nabla is AMI's only publicly disclosed commercial partner and receives early research access under undisclosed terms, not a revenue-generating license contract. | 中 | SI021, SI002 |
| CI015 | AMI has committed to publishing open-source code and academic papers, prioritizing community adoption over short-term IP protection. | 高 | SI001, SI002 |
| CI016 | A US SEC Form D was filed on April 3, 2026 for investment vehicles 'Advanced Machine Intelligence I, L.P.' and 'Advanced Machine Intelligence II, L.P.' — managed by Model I, L.P. of San Francisco — which aggregated approximately $16.9 million to co-invest in AMI Labs. | 高 | SI022, SI028 |
| CI017 | AMI's JEPA world models are designed to operate at hundreds of millions of parameters rather than hundreds of billions, which AMI claims would reduce per-training-run compute requirements substantially versus frontier LLMs. | 中 | SI014, SI002 |
| CI018 | Frontier AI model training compute has grown at 4-5x per year since 2020 and the largest known AI data center has a computing capacity equivalent to 800,000 NVIDIA H100-equivalent chips at approximately $24 billion in capital costs. | 高 | SI012, SI013 |
| CI019 | AI chip performance per dollar improved at approximately 37% per year between 2012 and 2025, meaning future compute costs will decline in real terms as better hardware is released. | 中 | SI012, SI013 |
| CI020 | Goldman Sachs estimated aggregate global AI capex at approximately $1 trillion in coming years but found 'this spending has little to show for it so far,' raising systemic questions about AI return on investment. | 高 | SI010, SI011 |
| CI021 | Sequoia Capital's analysis found a gap of $600B between AI infrastructure spending implied by NVIDIA's revenue run-rate and demonstrable AI end-user revenue, flagging risks of capital incineration and GPU commoditization. | 高 | SI011, SI010 |
| CI022 | Comparable frontier AI research labs (Anthropic, OpenAI) are reported to operate with multi-billion-dollar annual burn rates driven by compute and talent, though neither publicly discloses financials. | 低 | SI006, SI011 |
| CI023 | Futurum Group estimated AMI headcount at approximately 12 employees at the time of the seed announcement in March 2026, with active hiring underway across Paris, New York, Montreal, and Singapore. | 低 | SI006, SI003 |
| CI024 | LeBrun stated AMI will 'prioritize quality over quantity' in building its team across four locations, suggesting deliberate hiring rather than rapid headcount expansion. | 中 | SI002, SI005 |
| CI025 | At an estimated burn rate of $150M per year, AMI's $1.03B seed capital provides approximately 6.9 years of runway; at $300M per year, runway falls to approximately 3.4 years. | 低 | SI006, SI012 |
| CI026 | AMI's domain-specific JEPA world models could require substantially less compute than frontier LLM training if the architectural thesis holds; if world model training requires frontier-LLM-scale compute, AMI's runway could compress significantly. | 低 | SI011, SI012 |
| CI027 | AMI's open-source commitment creates tension between IP protection needed for premium licensing and community-adoption goals typical of research platforms, an unresolved model question for investors. | 中 | SI001, SI011 |
| CI028 | Futurum Group described AMI's Series A as 'the first real market test of whether AMI's research output translates to commercial credibility,' citing the absence of current revenue proof. | 中 | SI006 |
| CI029 | No quarterly or annual financial statements, balance sheets, or revenue disclosures exist for AMI Labs as a private French company; all financial metrics beyond the seed round size and valuation are unknown from public sources. | 高 | SI022, SI028 |
| CI030 | Sequoia's analysis found GPUs represent approximately 50% of total AI data center cost-of-ownership, with energy, buildings, backup generators, and networking comprising the other 50%. | 中 | SI011 |
| CI031 | Temasek, NVIDIA, Samsung, and Toyota Ventures' presence as strategic investors suggests AMI's earliest commercial relationships will likely be with large industrial partners in Asia and North America, consistent with a domain-specific licensing model. | 中 | SI001, SI006 |
| CI032 | AMI is recruiting researchers and engineers from OpenAI, Google DeepMind, and xAI, as confirmed by LeCun in a pre-announcement interview, suggesting a competitive talent acquisition strategy and serving as an early proxy for team quality. | 中 | SI023 |
| CI033 | AMI's oversubscription from €500M to €890M implies investors accepted the $3.5B pre-money as fair pricing for a multi-year research program rather than current revenue, a structure atypical even for frontier AI startups. | 中 | SI004, SI003 |
| CI034 | NVIDIA's strategic investment in AMI may include preferential compute access or hardware allocation agreements; such terms have not been disclosed and could materially reduce AMI's net compute spend. | 低 | SI001, SI016 |
| CI035 | The SEC Form D filing for Model I, L.P. aggregated approximately $16.9 million for investment in AMI Labs via 'Advanced Machine Intelligence I, L.P.' and 'Advanced Machine Intelligence II, L.P.' with first sale date March 19, 2026. | 高 | SI022, SI028 |
| CI036 | No formal Series A timeline, size target, or valuation expectation has been disclosed by AMI Labs as of the run date. | 中 | SI006, SI001 |
| CI037 | AI chip performance improvements of approximately 37% per year reduce AMI's long-run compute cost relative to current GPU pricing, broadly supporting the company's claim that JEPA world model training is less expensive than frontier LLM training over the research horizon. | 中 | SI013 |
| CI038 | Goldman Sachs analysts found AI technology 'has little to show for it so far' despite hundreds of billions in infrastructure spend, questioning whether AI revenue payoffs will materialize before sector faces investor pressure to cut capex. | 高 | SI010, SI011 |
| CI039 | AMI is hiring across research, engineering, and operations roles in four cities as of the May 2026 archived job board snapshot, suggesting the team has grown beyond the approximately 12 employees reported at seed close. | 低 | SI003, SI005 |
| CI040 | AMI's research-first mandate means burn will likely accelerate in years 2-3 as experiments scale up and larger training runs are attempted, creating potential burn-rate acceleration risk that could compress runway from the linear estimates. | 中 | SI012, SI011 |
| CE001 | AMI Labs is building world models based on Yann LeCun's Joint Embedding Predictive Architecture (JEPA), a framework first outlined in his June 2022 position paper "A Path Towards Autonomous Machine Intelligence." | 高 | SE001, SE003 |
| CE002 | The JEPA architecture makes predictions in abstract representation space rather than in raw pixel or token space, distinguishing it from generative world models. | 高 | SE003, SE005 |
| CE003 | The non-generative nature of JEPA allows the model to ignore unpredictable details while retaining high-level semantic structure, improving training efficiency. | 高 | SE003, SE005, SE006 |
| CE004 | I-JEPA (Image-based JEPA), published at CVPR 2023, trains a ViT-Huge/14 on ImageNet using 16 A100 GPUs in under 72 hours, achieving strong downstream performance on classification, object counting, and depth prediction. | 高 | SE004, SE006 |
| CE005 | V-JEPA (Video JEPA) is a non-generative model trained to predict missing spatio-temporal regions of video in latent space, released by Meta AI in 2024. | 高 | SE005, SE003 |
| CE006 | V-JEPA demonstrates training and sample efficiency improvements of 1.5× to 6× over prior video representation learning approaches on multiple benchmarks. | 高 | SE005, SE008 |
| CE007 | V-JEPA is capable of "frozen evaluation" — the pretrained encoder is fixed and only lightweight probes are added for new downstream tasks, enabling rapid task adaptation without full fine-tuning. | 高 | SE005, SE009 |
| CE008 | V-JEPA handles video clips up to approximately 10 seconds effectively; longer-horizon temporal reasoning over extended sequences remains an open research problem. | 中 | SE005 |
| CE009 | Meta released V-JEPA under a Creative Commons NonCommercial licence, which precludes direct commercial use by AMI or its partners without a separate arrangement. | 高 | SE005, SE008 |
| CE010 | AMI's official website states: "Action-conditioned world models allow agentic systems to predict the consequences of their actions, and to plan action sequences to accomplish a task, subject to safety guardrails." | 高 | SE001, SE002 |
| CE011 | AMI's core claim is that real-world sensor data is "continuous, high-dimensional, and noisy" and that generative approaches do not work well for such data, motivating the JEPA non-generative design. | 高 | SE001, SE002 |
| CE012 | AMI Labs CEO Alexandre LeBrun confirmed the company plans to publish papers and make code open source as a deliberate community-building strategy. | 高 | SE009, SE008 |
| CE013 | The I-JEPA codebase is publicly available on GitHub under facebookresearch as an official Meta AI Research repository, with full training code and config files. | 高 | SE006, SE004 |
| CE014 | AMI Labs plans to license its world model technology to industry partners for real-world applications rather than building and selling direct-to-customer products. | 高 | SE001, SE008, SE009 |
| CE015 | Nabla, AMI's first disclosed partner, aims to become the first company to bring FDA-certifiable agentic AI systems to healthcare using AMI's world model technologies. | 高 | SE007, SE017 |
| CE016 | The December 2025 Nabla partnership gives Nabla "first access" and "privileged access" to AMI's emerging world model technologies, with CEO LeBrun transitioning from Nabla. | 高 | SE007, SE009 |
| CE017 | NVIDIA Cosmos is a competing generative world foundation model platform trained on 20 million hours of real-world data and 9,000 trillion tokens, released at CES 2025. | 中 | SE013 |
| CE018 | NVIDIA Cosmos uses a generative approach (diffusion and autoregressive transformer models) in direct contrast to AMI's non-generative JEPA architecture. | 高 | SE013, SE003 |
| CE019 | DeepMind Genie 2 (December 2024) is an action-controllable generative 3D world model capable of generating diverse environments for training embodied agents. | 中 | SE019 |
| CE020 | OpenAI Sora, a generative video world model, was discontinued on April 26, 2026, illustrating commercial headwinds for pixel-generative video world models. | 中 | SE020 |
| CE021 | As of June 2026, AMI Labs has no publicly released product, model, or API; the company remains in the foundational research phase. | 高 | SE001, SE009, SE012 |
| CE022 | AMI CEO LeBrun explicitly stated the startup "is not your typical applied AI startup that can release a product in three months" and that it "could take years" for world models to move from theory to commercial applications. | 高 | SE009, SE008 |
| CE023 | AMI's architectural argument that non-generative prediction eliminates hallucinations has not been independently validated or benchmarked against LLM error rates in clinical or industrial settings. | 中 | SE015, SE012 |
| CE024 | AMI plans to contribute to the global academic research community via open publications and open-source code, in addition to its commercial licensing activities. | 高 | SE001, SE009 |
| CE025 | Futurum Group analyst coverage raised skepticism about whether AMI's JEPA-based world models can bridge the gap from academic benchmarks to production-grade commercial deployments given the long research-to-product pipeline. | 中 | SE012 |
| CE026 | AMI's official website identifies five target verticals for its world model platform: industrial process control, automation, wearable devices, robotics, and healthcare. | 高 | SE001, SE002 |
| CE027 | LeCun's 2022 position paper proposes five cognitive architecture modules: a perceptor, a configurable world model, an actor (planner), a configurator (goal setter), and dual memory systems (short-term and long-term). | 高 | SE003, SE006 |
| CE028 | The predictor in I-JEPA serves as a "primitive world model" that predicts semantic content of unseen image regions at a conceptual rather than pixel level. | 高 | SE006, SE004 |
| CE029 | NVIDIA released Cosmos under a permissive open model licence allowing commercial use, making it available to robotics and AV developers regardless of company size. | 中 | SE013 |
| CE030 | The I-JEPA GitHub repository is publicly available and maintained under the facebookresearch organisation, enabling community replication and extension. | 高 | SE006, SE004 |
| CE031 | V-JEPA outperforms prior video models in frozen evaluation on Kinetics-400 and Something-Something-v2 benchmarks, demonstrating superior label efficiency. | 高 | SE005, SE004 |
| CE032 | AMI's strategy of targeting healthcare, industrial, and robotics reflects a deliberate focus on safety-critical domains where LLM hallucinations create unacceptable risk. | 中 | SE001, SE007, SE009 |
| CE033 | Yann LeCun retains his professorship at NYU and continues to supervise PhD and postdoctoral students, creating a direct academic-to-commercial talent pipeline. | 高 | SE008, SE010 |
| CE034 | AMI describes persistent memory and the ability to reason and plan as core system properties of its target architecture, beyond perceptual world understanding. | 高 | SE002, SE001 |
| CE035 | AMI's software-only architecture is hardware-agnostic; the company does not manufacture sensors, robots, or edge hardware, unlike Physical Intelligence or 1X. | 中 | SE001 |
| CE036 | AMI's website explicitly commits to "open publications and open source" as a strategic approach to accelerating world model research and community building. | 高 | SE001, SE008 |
| CE037 | AMI CEO LeBrun predicted that "world models" would become the "next buzzword" with "every company calling itself a world model to raise funding" within six months. | 高 | SE009, SE025 |
| CE038 | Physical Intelligence (π) shipped its π0.7 steerable robot foundation model in April 2026, demonstrating that robotics-specific physical AI models are achieving commercial deployment ahead of AMI's general-purpose world model approach. | 中 | SE021 |
| CE039 | SpAItial's Echo world model for 3D Gaussian Splat environments is an active competitor in the world model space, with its Echo HQ model released in May 2026. | 中 | SE024 |
| CE040 | As of the June 2026 run date, no standalone research papers have been published under the "AMI Labs" banner; the company's technical credibility rests on prior Meta FAIR publications authored by LeCun and team members. | 中 | SE012, SE015, SE023 |
| CE041 | AMI's architecture handles non-visual sensor data in addition to vision; the official website references "cameras or any other sensor modality" as inputs, but no published paper demonstrates multimodal sensor fusion beyond vision and (planned) audio. | 中 | SE001, SE005 |
| CE042 | Epoch AI data shows training compute for frontier AI models growing at 5× per year since 2020, meaning AMI must continuously scale to remain competitive with well-resourced incumbents. | 中 | SE022 |
| CU001 | AMI Labs has no commercial customers, no revenue-generating contracts, and no production deployments as of June 2026. | 高 | SU010, SU011, SU012, SU018 |
| CU002 | CEO Alexandre LeBrun explicitly stated that AMI is "not your typical applied AI startup that can release a product in three months, have revenue in six months, and make $10 million in ARR in 12 months." | 高 | SU010, SU012 |
| CU003 | Yann LeCun stated in March 2026 that discussions with corporate partners "could be held within six to 12 months," implying first partner discussions no sooner than Q3 2026. | 高 | SU014, SU010 |
| CU004 | AMI plans to focus on research and development in its first year before engaging commercial customers, per LeCun's AFP statement. | 高 | SU014, SU018 |
| CU005 | Nabla is AMI's only publicly disclosed partner as of June 2026; no other design-partner agreements have been announced. | 高 | SU009, SU010, SU011 |
| CU006 | AMI's mission statement targets industrial process control, automation, wearable devices, robotics, and healthcare as the intended verticals for future products. | 高 | SU018, SU006, SU013 |
| CU007 | LeCun stated in the MIT Technology Review interview that "Meta might be our first client," citing shared interest in physical-world AI for consumer devices. | 中 | SU012 |
| CU008 | Strategic investors Toyota Ventures, Groupe Industriel Marcel Dassault, Samsung, NVIDIA, and ZEBOX Ventures represent likely early design-partner prospects across automotive, aerospace, devices, hardware, and logistics verticals. | 中 | SU017, SU019, SU016 |
| CU009 | ZEBOX Ventures, the CMA CGM logistics group's early-stage fund, is a strategic investor, suggesting logistics and shipping operations as a target vertical for AMI world models. | 中 | SU017, SU019 |
| CU010 | LeBrun stated that AMI must "put the model in a real-world situation with real data and real evaluations" through partner engagements, confirming the design-partner model as the primary pre-commercial strategy. | 高 | SU011, SU010 |
| CU011 | Nabla announced an exclusive strategic partnership with AMI in December 2025, giving Nabla "first access" to AMI's emerging world model technologies. | 高 | SU009, SU011 |
| CU012 | The Nabla-AMI partnership is structured as a design-partner access arrangement, not a commercial license; no revenue flows to AMI from Nabla's deployments. | 高 | SU009, SU010 |
| CU013 | Nabla's stated goal through the AMI partnership is to become "the first to bring FDA-certifiable agentic AI systems to healthcare," a positioning dependent on future AMI world model technology. | 高 | SU009, SU015 |
| CU014 | As of June 2026, Nabla serves over 190 health organizations and more than 100,000 clinicians across 150+ health systems using its ambient documentation platform. | 高 | SU002, SU001, SU009 |
| CU015 | Nabla has raised $120 million in funding from investors including HV Capital, Highland Europe, and Cathay Innovation, and serves over 150 health systems and provider groups. | 高 | SU009, SU019 |
| CU016 | The AMI-Nabla relationship is structurally non-arm's-length: AMI CEO LeBrun was Nabla's co-founder and CEO, and retains the roles of Chairman and Chief AI Scientist at Nabla. | 高 | SU009, SU011, SU012 |
| CU017 | Yann LeCun has been a Nabla investor and advisor since the company's founding, predating AMI's formation. | 高 | SU009, SU012 |
| CU018 | No timeline has been disclosed for when AMI world model technology will be integrated into Nabla's clinical platform or tested on patient data. | 高 | SU009, SU011 |
| CU019 | Nabla's current clinical AI systems are entirely LLM-based; AMI's world model technology is in research phase and has no production-ready API or product interface. | 高 | SU009, SU015, SU020 |
| CU020 | LeBrun described Nabla as AMI's first disclosed partner "but definitely not the last," signalling undisclosed partner discussions are likely but unconfirmed. | 中 | SU011 |
| CU021 | In healthcare, the relevant AMI buyer role is the hospital system CTO/CMO or a healthcare AI platform (like Nabla) seeking FDA-certifiable autonomous clinical workflows. | 中 | SU009, SU005, SU006 |
| CU022 | In industrial process control and manufacturing, AMI's target buyer is an engineering or operations leader at a sensor-rich manufacturer such as an aircraft engine or chemical plant operator. | 中 | SU006, SU018 |
| CU023 | In robotics, AMI's target buyer is an automation or robotics engineering lead at a manufacturing or logistics firm seeking action-conditioned planning capabilities for physical robots. | 中 | SU018, SU022, SU006 |
| CU024 | In wearable devices, AMI's target buyer is a device OEM (Samsung, Garmin) product or AI team seeking persistent-memory edge intelligence for consumer wearables. | 中 | SU018, SU017 |
| CU025 | Groupe Industriel Marcel Dassault—parent of Dassault Aviation and Dassault Systèmes—invested in AMI's seed round, signalling potential aerospace and industrial design-partner interest. | 高 | SU017, SU019, SU016 |
| CU026 | Toyota Ventures invested in AMI's seed round, signalling potential automotive and robotics use-case alignment, though Toyota Motor (as customer) is a separate entity from Toyota Ventures (as investor). | 高 | SU017, SU019 |
| CU027 | AMI's sovereign-AI positioning (European headquarters, non-US non-Chinese frontier lab) is explicitly designed to attract buyers in EU and Asian markets with AI sovereignty concerns. | 中 | SU015, SU012, SU014 |
| CU028 | Publicis Groupe, a global advertising and communications conglomerate, invested in AMI's seed round, signalling potential media and creative-industry use cases. | 中 | SU017, SU016 |
| CU029 | SEA Group and SBVA (SoftBank Ventures Asia) invested in AMI, signalling Southeast Asia and the Singapore ecosystem as a target commercial development geography. | 中 | SU017, SU025 |
| CU030 | Temasek, Singapore's sovereign wealth fund, invested in AMI, aligning with Singapore's national AI strategy and potentially opening public-sector design-partner pathways in Southeast Asia. | 中 | SU017, SU019 |
| CU031 | Menlo Ventures' 2025 survey found healthcare health systems shortened AI procurement cycles from 8.0 to 6.6 months, but payers lengthened cycles to 11.3 months; these compression benefits apply to production AI tools, not research-stage models. | 高 | SU005, SU004 |
| CU032 | The FDA regulatory pathway for autonomous agentic clinical AI is substantially more complex than the existing 510(k) pathway used for over 900 existing narrow diagnostic AI devices, likely requiring PMA or novel De Novo submission with timelines of 2–7 years. | 中 | SU003, SU015, SU008 |
| CU033 | The FDA has cleared over 900 AI/ML-enabled medical devices as of early 2026, but these are predominantly narrow diagnostic and detection tools, not autonomous agentic systems of the kind Nabla and AMI envision. | 高 | SU003, SU008 |
| CU034 | Enterprise software procurement for complex industrial process control systems typically spans 12–24 months even for established vendors with proven solutions. | 中 | SU015, SU023 |
| CU035 | Goldman Sachs research (2024) found $1 trillion in projected AI capex with "little to show for it so far" in deployed AI applications, a structural observation that applies to any frontier AI vendor seeking enterprise adoption. | 高 | SU023, SU024 |
| CU036 | Sequoia Capital's "AI's $600B Question" analysis (updated 2024) quantified the gap between AI infrastructure spending and demonstrated end-user value, flagging a "$500B hole" in commercial AI revenue to support the current capex level. | 高 | SU024, SU023 |
| CU037 | LeBrun acknowledged that commercial products could take "years" to materialize from AMI's research timeline, establishing that meaningful customer revenue is not expected before 2028 at the earliest. | 高 | SU010, SU014 |
| CU038 | The Futurum Group analyst assessment explicitly noted "healthcare AI development, particularly any pathway to FDA certification, is a long and uncertain process" and that no target vertical can "absorb a world-model-as-a-service offering on a short timeline." | 中 | SU015 |
| CU039 | AMI's LLM-alternative architectural positioning means customers must also adopt a paradigm shift in AI architecture, adding conceptual adoption friction on top of standard enterprise procurement cycles. | 中 | SU015, SU020, SU023 |
| CU040 | Menlo Ventures found healthcare buyers prioritize "maturity of technology" and "production-ready solutions that perform reliably at scale"—standards AMI's research-phase world models cannot yet meet as of June 2026. | 高 | SU005, SU004 |
| CU041 | AMI Labs has zero disclosed revenue, zero ARR, and zero paying customers as of June 2026; these are confirmed absences based on management public statements, not merely undisclosed data. | 高 | SU010, SU011, SU014 |
| CU042 | No signed customer contracts beyond the Nabla design-partner arrangement have been publicly disclosed; the existence of undisclosed agreements is possible but unverifiable. | 高 | SU010, SU011 |
| CU043 | No production deployment of AMI's world model technology in any customer or partner environment has been announced as of June 2026. | 高 | SU009, SU015, SU011 |
| CU044 | Net Revenue Retention (NRR) and Gross Revenue Retention (GRR) are undefined for AMI because no revenue-generating customer relationship exists. | 高 | SU010, SU014 |
| CU045 | Digital health venture funding in 2025 reached $14.2 billion (US), but capital is concentrated in proven production AI startups—not frontier research labs developing architecturally novel systems. | 高 | SU004, SU005 |
| CU046 | Healthcare AI spending of $1.4 billion in 2025 was predominantly in ambient clinical documentation ($600M) and billing automation ($450M)—categories where Nabla competes but AMI's world models have no near-term product. | 高 | SU005, SU004 |
| CU047 | Nabla's 190+ health organization customer base belongs to Nabla; AMI receives no revenue from Nabla's clinical deployments and its world models are not integrated into any Nabla product. | 高 | SU002, SU009, SU011 |
| CU048 | LeCun cited aircraft manufacturer sensor modelling as a hypothetical world model use case in a Wired interview, but no aerospace or manufacturing customer engagement has been confirmed. | 中 | SU006, SU022 |
| CU049 | Major LLM incumbents (OpenAI, Microsoft, Google) are subsidizing AI adoption to gain market share in healthcare and industrial verticals, creating pricing pressure on any future AMI commercial offering. | 中 | SU007, SU004, SU024 |
| CU050 | The BMJ's TRIPOD+AI framework requires external validation, pre-specified calibration, and bias assessment before clinical AI deployment—standards that unvalidated research-phase world models cannot meet. | 高 | SU008, SU003 |
| CR001 | AMI Labs raised $1.03 billion in a seed round in March 2026 at a $3.5 billion pre-money valuation, the largest seed round in European history. | 高 | SR014, SR017, SR021 |
| CR002 | AMI Labs has no revenue and no deployed product as of June 2026, operating on a self-described research-first multi-year timeline. | 高 | SR012, SR020, SR030 |
| CR003 | Estimated annual burn for a frontier AI lab of AMI's stated ambition (compute and talent as primary cost centres) is $300–500M, based on peer lab analogues. | 中 | SR015, SR022, SR019 |
| CR004 | Given 4–5× annual compute cost growth and a multi-year research timeline, AMI will require substantial additional capital beyond its $1.03B seed within approximately 24 months. | 中 | SR011, SR022, SR003 |
| CR005 | Goldman Sachs projected that an estimated $1 trillion in AI capex in coming years has little measurable revenue or benefit to show so far. | 高 | SR015, SR018 |
| CR006 | Sequoia Capital estimated a $600 billion revenue gap between AI infrastructure spending and actual AI-derived revenues as of mid-2024, labelling it 'AI's $600B Question'. | 高 | SR016, SR015 |
| CR007 | AI training compute requirements at the frontier have grown at 4–5× per year from 2010 to 2024, making each successive model training run significantly more expensive. | 高 | SR011, SR022 |
| CR008 | NVIDIA holds a near-monopoly on AI training silicon and is a co-investor in AMI Labs, creating both supply access and alignment risk in AMI's compute dependency. | 高 | SR029, SR011 |
| CR009 | Yann LeCun serves as AMI Labs' executive chairman, not CEO; Alex LeBrun is CEO. LeCun explicitly describes his role as strategic rather than operational. | 高 | SR012, SR014, SR017 |
| CR010 | LeCun retains his NYU professorship, teaching one class per year and supervising PhD students, and remains based in New York while AMI is headquartered in Paris. | 高 | SR012, SR013 |
| CR011 | LeCun described the governance arrangement as 'It's going to be LeCun and LeBrun—it's nice if you pronounce it the French way,' suggesting a symbolic rather than operational executive chairman role. | 中 | SR012 |
| CR012 | LeCun has publicly and repeatedly characterised LLMs as a 'dead end' for achieving general intelligence, making AMI a contrarian bet against the dominant commercial AI paradigm. | 高 | SR012, SR017 |
| CR013 | AMI Labs' founding team—LeBrun (CEO), Saining Xie (CSO), Laurent Solly (COO), Michael Rabbat (VP World Models)—is predominantly composed of former Meta FAIR alumni, creating talent concentration risk. | 高 | SR013, SR017 |
| CR014 | Alex LeBrun simultaneously holds the role of AMI CEO and Nabla chairman and chief AI scientist, creating a dual-role bandwidth and conflict-of-interest risk. | 高 | SR013, SR014 |
| CR015 | No public information is available on any non-compete agreement or IP assignment between Yann LeCun and Meta AI covering JEPA architecture developed under Meta FAIR funding. | 低 | |
| CR016 | LeCun left Meta in November 2025 after 12 years, citing disagreements with Mark Zuckerberg over AI strategy and criticism of Meta's handling of its robotics group. | 高 | SR017, SR012 |
| CR017 | LeCun stated that achieving human-level AI requires 'major conceptual breakthroughs' and is 'not going to happen next year or two years from now.' | 高 | SR012, SR013 |
| CR018 | AMI's JEPA-based world-model architecture has not been validated at commercial scale and no production deployment of AMI technology exists as of June 2026. | 高 | SR012, SR020 |
| CR019 | AMI Labs has no publicly announced commercial product, API, or licensed deployment as of the report run date of June 2026. | 高 | SR020, SR030 |
| CR020 | AMI's world models require training on diverse sensor, video, audio, and proprioceptive data modalities—a data acquisition challenge that exceeds text-only LLM pretraining. | 中 | SR012, SR019 |
| CR021 | Gary Marcus argues that generative AI valuations anticipate trillion-dollar markets that have not materialised, and that frontier AI startups valued in the low billions risk collapse if revenues remain in the hundreds of millions annually. | 中 | SR006, SR035 |
| CR022 | The productisation gap from foundational AI research to commercially deployable licensed product typically takes three to seven years, based on historical AI technology development precedents. | 中 | SR010, SR016 |
| CR023 | AMI Labs' commitment to open-source publication of research and model weights may structurally prevent it from establishing a proprietary licensing moat before better-resourced competitors absorb and implement its findings. | 中 | SR012, SR004 |
| CR024 | Training compute costs growing at 4–5× per year means AMI's first full world-model training run, if delayed 18–24 months, will cost materially more than current estimates allow. | 中 | SR011, SR015 |
| CR025 | The EU AI Act (Regulation 2024/1689) requires GPAI model providers to publish detailed summaries of training data content and comply with EU copyright law, with obligations in force from August 2025. | 高 | SR003, SR002, SR001 |
| CR026 | Under EU AI Act Article 51, a GPAI model is presumed to have systemic risk if its cumulative training computation exceeds 10^25 floating-point operations, triggering mandatory model evaluation, risk mitigation, and incident reporting obligations. | 高 | SR005, SR002, SR001 |
| CR027 | AI applications in healthcare and robotics are classified as high-risk under the EU AI Act, requiring conformity assessments, quality management systems, and ongoing post-market monitoring before deployment in the EU. | 高 | SR001, SR003 |
| CR028 | In the US, AI-enabled medical devices require FDA clearance (510(k) pathway) or premarket approval (PMA) before commercial deployment, with the FDA's AI/ML Action Plan outlining the regulatory framework since January 2021. | 中 | SR028, SR009 |
| CR029 | Healthcare AI regulatory approval in the US typically requires clinical evidence and can take 12 to 36 months from submission, representing a significant timeline and cost barrier for AMI's healthcare AI applications. | 中 | SR028, SR010 |
| CR030 | US Bureau of Industry and Security (BIS) Export Administration Regulations (EAR) restrict the transfer of advanced AI chips and related technology to certain countries and end-users, creating supply-chain risk for AMI's compute infrastructure. | 高 | SR008, SR009 |
| CR031 | GDPR Article 9 restricts processing of special-category health data, including clinical data, in the EU, imposing data processing agreements and data protection impact assessments on any healthcare AI system AMI deploys. | 高 | SR009, SR001 |
| CR032 | AMI Labs' Paris headquarters and GPAI model development activities place it directly within scope of EU AI Act GPAI compliance obligations that began applying from August 2025. | 高 | SR003, SR001, SR002 |
| CR033 | The EU AI Act establishes enforcement mechanisms including fines of up to €35 million or 7% of global annual turnover for violations of prohibited AI practices, with proportionately lower fines for GPAI and other obligations. | 高 | SR003, SR004 |
| CR034 | Google DeepMind (Genie 2), NVIDIA (Cosmos), and Physical Intelligence (π) are directly competing with AMI in world-model and physical-AI research, all with larger balance sheets and production deployments. | 高 | SR029, SR024, SR019 |
| CR035 | Physical Intelligence (π) has raised over $400 million and deployed commercial robotics AI products, establishing a first-mover advantage in physical-world AI that AMI's research-first timeline cannot match for 2–3 years. | 中 | SR017, SR019 |
| CR036 | Meta's continuation of FAIR and its open-source Llama models creates both research competition for talent and an open-source benchmark that could absorb AMI's findings before licensing revenue is established. | 中 | SR013, SR016 |
| CR037 | Open-source Chinese AI models including DeepSeek demonstrate that world-model-adjacent capabilities can be developed and open-sourced by well-resourced but non-Western labs, undermining AMI's sovereign-AI narrative. | 中 | SR012, SR021 |
| CR038 | Hyperscaler AI capex (Google, Microsoft, Meta, Amazon) is projected at approximately $1 trillion in coming years, dwarfing AMI's $1.03B seed by roughly three orders of magnitude. | 高 | SR015, SR016 |
| CR039 | AMI's primary strategic differentiator—sovereign European AI—is a geopolitical narrative rather than a demonstrated technical advantage, and depends on sustained EU policy support, investor appetite, and absence of a superior open-source alternative. | 中 | SR012, SR006, SR010 |
| CR040 | OpenAI and Anthropic have 2–3-year head starts in commercialising AI products and have collectively raised over $15B, giving them customer acquisition and product iteration advantages that AMI cannot close quickly. | 中 | SR014, SR016 |
| CR041 | LeCun explicitly positions AMI as a 'third-path' sovereign European AI alternative to US-dominated and Chinese-dominated models, making the company's strategic case dependent on a sustained EU sovereign AI narrative. | 高 | SR012, SR013 |
| CR042 | EU AI Act GPAI transparency obligations came into force in August 2025; systemic-risk obligations for GPAI models above the 10^25 FLOP threshold apply from August 2026. | 高 | SR003, SR002, SR001 |
| CR043 | US BIS export controls on advanced AI chips could affect AMI's compute supply chain if geopolitical conditions deteriorate or if restrictions on AI chip transfers to European entities tighten. | 中 | SR008, SR030 |
| CR044 | AMI Labs operates across four jurisdictions (Paris, New York, Montreal, Singapore), creating complex regulatory exposure and potential GDPR vs. US data localisation conflicts. | 高 | SR013, SR017 |
| CR045 | If leading European AI researchers migrate to better-funded US labs (as has occurred historically), the EU sovereign AI narrative underlying AMI's strategic positioning could unravel, removing its primary differentiation. | 中 | SR010, SR006 |
| CR046 | French President Macron publicly endorsed AMI Labs' Paris headquarters, creating a political dependency that could expose AMI to French regulatory pressure if political priorities shift. | 中 | SR023, SR013 |
| CR047 | AMI Labs' $3.5B pre-money seed valuation lacks the product or revenue milestones typical of companies at comparable valuations, making it dependent on narrative and team rather than financial metrics. | 中 | SR021, SR016, SR006 |
| CR048 | AMI's commitment to open scientific publication and open-source contributions, combined with its pre-revenue status, means competitors may commercialise AMI's research before AMI itself can license it. | 中 | SR012, SR004, SR020 |
| CR049 | As of June 2026, no public enforcement action by the EU AI Office or US regulatory bodies specifically against any frontier AI lab for GPAI compliance violations has been reported. | 中 | SR003, SR001 |
| CR050 | US BIS export control restrictions on advanced AI hardware principally target China and certain other countries; European entities headquartered in allied nations such as France have not been subject to the most restrictive export licensing requirements as of June 2026. | 中 | SR008, SR009 |
| CV001 | AMI Labs closed its seed round in March 2026 at a $3.5 billion pre-money valuation, implying a $4.53 billion post-money. | 高 | SV010, SV012 |
| CV002 | PitchBook confirmed the $1.03B AMI Labs raise as Europe's largest seed round on record at the time of announcement. | 高 | SV011, SV010 |
| CV003 | World Labs (Fei-Fei Li) raised $230 million at a $1 billion valuation in August–September 2024 in a Series A led by Andreessen Horowitz. | 高 | SV001, SV003 |
| CV004 | According to Dataconomy, World Labs was reportedly in talks to raise at a $5 billion valuation after launching Marble in November 2025. | 中 | SV017, SV019 |
| CV005 | Thinking Machines Lab, founded by former OpenAI CTO Mira Murati, was valued at approximately $12 billion in its seed round. | 中 | SV014 |
| CV006 | DeepSeek reportedly raised approximately $7.4 billion in June 2026 at a valuation exceeding $50 billion, per reporting by The Information and the Wall Street Journal. | 中 | SV002, SV006 |
| CV007 | Odyssey, a world-model startup, led Crunchbase's weekly funding deals for the week of June 18, 2026, indicating a $310 million raise. | 中 | SV005, SV023 |
| CV008 | C3.ai (NASDAQ: AI) reported total revenue of $389.1 million for fiscal year ended April 30, 2025, representing 25.3% year-over-year growth, with subscription revenue of $327.6 million. | 中 | SV007 |
| CV009 | NVIDIA has invested approximately $53 billion across 170 AI startup deals according to PitchBook data cited in press reporting. | 中 | SV017 |
| CV010 | AMI Labs' seed investor syndicate spans US, European, and Asian sovereigns and strategics, including NVIDIA, Samsung, Temasek, Toyota Ventures, and Bpifrance Digital Venture. | 高 | SV010, SV012, SV013 |
| CV011 | Yann LeCun holds the Turing Award (2018) and led Meta's FAIR group for more than a decade, constituting AMI's primary scientific credibility anchor. | 高 | SV012, SV010 |
| CV012 | AMI CEO LeBrun stated publicly that AMI Labs is not a typical applied AI startup with revenue in six months and that commercial products may take years to materialise. | 高 | SV010, SV014 |
| CV013 | The Futurum Group analyst characterises AMI's European positioning as 'strategic autonomy' rather than true technical independence, noting AMI runs on NVIDIA silicon and competes for global talent. | 中 | SV014, SV029 |
| CV014 | Goldman Sachs estimates approximately $1 trillion in planned AI infrastructure capex has generated little measurable revenue or productivity benefit so far. | 高 | SV015, SV016 |
| CV015 | Sequoia Capital updated its '$200B question' to a '$600B question' — the structural gap between AI infrastructure spending and revenues needed to sustain it. | 高 | SV016, SV015 |
| CV016 | AMI's initial fundraising target was approximately €500 million as recently as December 2025; the final close of ~€890M ($1.03B) significantly exceeded that target, indicating strong investor demand. | 高 | SV020, SV010 |
| CV017 | Andreessen Horowitz's investment thesis for World Labs frames spatial intelligence as analogous in impact to the LLM revolution, providing a valuation rationale applicable to all world-model labs including AMI. | 中 | SV001 |
| CV018 | CB Insights' AI 100 for 2026 finds that the most durable AI company businesses are defined by non-textual, rare, or deeply embedded data moats — a standard AMI does not yet meet. | 中 | SV005 |
| CV019 | In the bull scenario, JEPA achieves a demonstrable benchmark advantage by 2028, AMI raises a Series A at $10–15B, and an IPO is plausible by 2031–2033 at $30–50B contingent on $200M+ ARR. | 低 | SV014, SV010, SV016 |
| CV020 | In the base scenario, AMI produces research for 2–3 years, raises a Series A at $5–8B, and exits via hyperscaler acquisition at $8–15B in the 2030–2032 window. | 低 | SV014, SV029 |
| CV021 | In the bear scenario, JEPA fails to differentiate from multimodal LLMs, capital markets tighten for pre-revenue AI research, and AMI faces a distressed sale or wind-down at $1–2B by 2028–2030. | 低 | SV015, SV016, SV018 |
| CV022 | AMI's estimated annual burn rate of $300–500 million is extrapolated from peer frontier-lab analyses and is not confirmed from AMI's own disclosures. | 低 | SV014, SV018 |
| CV023 | Epoch AI documents that frontier AI model training compute grows at 4–5× per year, meaning AMI's fixed compute budget becomes inadequate over a multi-year research horizon without additional fundraising. | 中 | SV018, SV014 |
| CV024 | The Futurum analyst concludes AMI will require additional capital beyond its $1.03B seed, as research labs at comparable scale generally need a deep-pocketed patron or a defined revenue path. | 中 | SV029, SV014 |
| CV025 | The recommended investment stance for AMI as of June 2026 is TRACK: the valuation is stretched relative to fundamental evidence, but the scientific thesis is credible and warrants monitoring. | 中 | SV014, SV015, SV016 |
| CV026 | The valuation stance is STRETCHED: $3.5B pre-money prices in substantial option value on JEPA's commercial breakthrough with no current-period revenue, no product, and no commercial partnerships beyond Nabla. | 中 | SV010, SV012, SV014 |
| CV027 | Confidence in the recommendation is LOW due to the near-total absence of commercial evidence and high uncertainty on the JEPA technology hypothesis as of June 2026. | 中 | SV014, SV027 |
| CV028 | AMI's overall risk rating is HIGH, reflecting key-person concentration on LeCun, technology uncertainty, compute-cost escalation, Big Lab competition, and the need for serial fundraising in a volatile market. | 中 | SV014, SV015, SV016, SV018 |
| CV029 | No dimension in an IC-ready KPI framework (market, tech proof, moat, economics, risk, valuation, evidence quality) exceeds a score of 8 out of 10, with most dimensions scoring 1–4 on current public evidence. | 低 | SV014, SV005 |
| CV030 | AMI is not exit-ready as of June 2026; no revenue, no product, and no precedent for near-term IPO; the most likely liquidity path is a hyperscaler strategic acquisition in the 2029–2033 window. | 中 | SV014, SV029 |
| CV031 | Potential strategic acquirers for AMI include Samsung, Toyota Ventures, Apple, Microsoft, and EU national champions, each with distinct strategic rationale for absorbing the JEPA program. | 低 | SV013, SV014 |
| CV032 | The $1.03B seed provides an estimated 2–3 years of runway at frontier-lab burn rates, meaning AMI will need to raise a Series A in late 2027 or 2028, which will function as the real market-clearing valuation test. | 低 | SV022, SV018, SV014 |
| CV033 | The loss of Yann LeCun as executive chairman is identified as a hard-stop thesis-break trigger, as his brand premium is the primary non-financial valuation input at seed stage. | 中 | SV012, SV014 |
| CV034 | Independent verification of JEPA's performance advantage over multimodal LLMs on a standardised physical-world benchmark is the critical missing evidence for a Series A investment recommendation. | 低 | |
| CV035 | AMI has made no public disclosure of its use-of-proceeds detail beyond 'compute and talent'; a confirmed compute spend rate is required to validate the $300–500M/yr burn estimate. | 低 | |
| CV036 | AMI's founding employee count was approximately a dozen at seed close; no headcount target or hiring plan has been publicly disclosed. | 中 | SV014, SV010 |
| CV037 | SpAItial raised a €13 million seed round — described as unusually large for a European startup — providing a lower-bound comp for European world-model adjacent startups. | 中 | SV010 |
| CV038 | Greycroft's portfolio description of AMI Labs confirms the company is described as a 'research lab developing world-model AI,' consistent with the research-first positioning disclosed in press materials. | 中 | SV008 |
| CV039 | The Futurum report characterises AMI's investor syndicate as 'carefully constructed, not a momentum-driven pile-on,' with co-leads spanning US and European VC and strategic industrials. | 中 | SV014, SV029 |
| CV040 | As of June 2026, no public evidence of AMI's next-round valuation expectations, Series A lead conversations, or JEPA benchmark results is available; these constitute the principal unresolved evidential gaps. | 低 | |
| CV041 | The Futurum analyst explicitly identifies the AMI-Nabla healthcare partnership as the only disclosed early anchor, and notes FDA certification for AI healthcare systems is 'a long and uncertain process.' | 中 | SV014, SV029 |
| CV042 | AMI's valuation of $3.5B compares to World Labs' reported ~$5B in early 2026; this modest premium is consistent with the market applying a discount for AMI's longer research horizon and unproven JEPA commercialisation. | 中 | SV004, SV014, SV017 |
| CV043 | The probability-weighted intrinsic value of AMI across three scenarios (bear $1.5B at 20%, base $11B at 60%, bull $40B at 20%) yields an expected value of approximately $9.1B, above the seed post-money of $4.53B but subject to extreme uncertainty. | 低 | SV014, SV015, SV016 |
| CV044 | AMI's financing optionality includes Series A VC, additional strategic rounds (e.g., from existing NVIDIA or Temasek), sovereign AI public grants (Bpifrance, EU Horizon), and strategic M&A conversations. | 中 | SV014, SV022 |
| CV045 | There is no public evidence of AI research lab valuation compression in 2025–2026; the DeepSeek ($50B), Thinking Machines ($12B), and Odyssey ($310M raise) rounds suggest the frontier AI premium has held or expanded. | 中 | SV002, SV004, SV005 |
| 编号 | 出版方 | 标题 | 引文 |
|---|---|---|---|
| SO001 | AMI Labs | AMI Labs — Real World. Real Intelligence. (Homepage) | AMI is developing world models that learn abstract representations of real-world sensor data, ignoring unpredictable details, and that make predictions in representation space. |
| SO002 | AMI Labs | AMI Labs — Updates (Official Funding Announcement) | 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. |
| SO003 | TechCrunch | Yann LeCun confirms his new 'world model' startup, reportedly seeks $5B+ valuation | His startup is called Advanced Machine Intelligence (AMI) and has hired Alex LeBrun, co-founder and CEO of medical transcription AI startup darling Nabla, as its CEO. |
| SO004 | TechCrunch | Who's behind AMI Labs, Yann LeCun's 'world model' startup | AMI Labs 'is going to be a global company [that's] headquartered in Paris.' The news was welcomed by French President Emmanuel Macron. |
| SO005 | 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. |
| SO006 | Nabla | Nabla Announces Exclusive Partnership With Advanced Machine Intelligence to Pioneer the Next Era of Agentic Healthcare AI | |
| SO007 | MIT Technology Review | Yann LeCun's new venture is a contrarian bet against large language models | |
| SO008 | Forbes | Why Yann LeCun's Startup Advanced Machine Intelligence Is Targeting Healthcare | |
| SO009 | Dataconomy | Yann LeCun's AMI Labs Hits $3.5 Billion Pre-money Valuation | |
| SO010 | Futurum Group | Yann LeCun's AMI Raises $1BN Seed Round — Is the World Model Era Finally Here? | The claim that world models represent a categorical solution to those failure modes — rather than a different set of tradeoffs — is more contested. World models have their own generalization challenges. |
| SO011 | Cathay Innovation | Advanced Machine Intelligence (AMI) is Enabling the Next AI Revolution — Built on Foundational World Models | This investment brings together proven execution and unmatched scientific vision. |
| SO012 | France24 | French AI startup AMI raises $1B to develop 'universal intelligent systems' | |
| SO013 | Euronews | AMI, une startup française de l'IA, annonce une levée de fonds d'un milliard de dollars | |
| SO014 | New York Observer | Yann LeCun's Paris A.I. Startup AMI Labs Raises Record $1B Seed Round | In November, Yann LeCun left Meta after 12 years over disagreements with Mark Zuckerberg over the future of A.I. |
| SO015 | EU Startups | Beyond LLMs: AI pioneer Yann LeCun's new venture AMI raises €890 million to build 'world model' AI systems | |
| SO016 | SiliconAngle | Yann LeCun's new startup AMI Labs raises $1.03B to train world models | |
| SO017 | AI News (TechForge) | The billion-dollar startup with a different idea for AI — AMI Labs | Each module would be trained in ways that relevant to the AI's particular field. |
| SO018 | arXiv / IEEE/CVF ICCV 2023 | Self-Supervised Learning from Images with a Joint-Embedding Predictive Architecture (I-JEPA) | The idea behind I-JEPA is simple: from a single context block, predict the representations of various target blocks in the same image. |
| SO019 | ACM (Association for Computing Machinery) | Fathers of the Deep Learning Revolution Receive 2018 ACM A.M. Turing Award | ACM today named Yoshua Bengio, Geoffrey Hinton, and Yann LeCun recipients of the 2018 ACM A.M. Turing Award. |
| SO020 | Greycroft | Greycroft Portfolio — Advanced Machine Intelligence (AMI) | Advanced Machine Intelligence (AMI) is a research lab developing world-model AI that learns structured representations of the physical world. |
| SO021 | AMI Labs (Jobs) | AMI Labs — Open Positions (Ashby HQ) | |
| SO022 | Yann LeCun (Personal Academic Site) | Yann LeCun's Home Page — NYU / AMI Labs | Executive Chairman, Advanced Machine Intelligence (AMI Labs). Jacob T. Schwartz Professor of Computer Science, Data Science, Neural Science, and Electrical and Computer Engineering, New York University. |
| SO023 | Saining Xie (Personal Academic Site) | Saining Xie — AMI Labs CSO / NYU Faculty | I am the co-founder and Chief Science Officer (CSO) at AMI Labs, and a faculty member at NYU. Before this, I was a research scientist at Google DeepMind. |
| SO024 | Hiro Capital | HIRO Capital — Portfolio (AMI Labs investment, March 2026) | The investment will help drive AMI's vision to develop world models that learn abstract representations of reality, similar to the mental models humans use to reason and guide action in the physical world. |
| SO025 | OpenReview (LeCun et al.) | A Path Towards Autonomous Machine Intelligence — LeCun JEPA Position Paper (2022) | |
| SM001 | Grand View Research | Artificial Intelligence Market Size, Share & Trends Analysis Report, 2033 | The global artificial intelligence market size was estimated at USD 390.9 billion in 2025 and is projected to grow from USD 539.5 billion in 2026 to USD 3,497.3 billion by 2033, at a CAGR of 30.6% from 2026 to 2033. |
| SM002 | MarketsandMarkets | Artificial Intelligence (AI) in Healthcare Market — Global Forecast to 2031 | The global artificial intelligence (AI) in healthcare market is projected to reach USD 194.79 billion by 2031 from USD 36.67 billion in 2026, at a CAGR of 39.7% during the forecast period. |
| SM003 | MarketsandMarkets | Robotics AI Software Market Size, Share & Trends — Global Forecast to 2032 | The global Robotics AI Software Market is estimated at USD 21.0 billion in 2025 and is projected to reach USD 70.8 billion by 2032, growing at a CAGR of 19.0% from 2026 to 2032. |
| SM004 | MarketsandMarkets | Artificial Intelligence (AI) Market by Offering, Technology, Business Function — Global Forecast to 2033 | The Artificial intelligence (AI) market was estimated to be worth USD 601.93 billion in 2026 and is projected to reach USD 3,638.08 billion by 2033, at a CAGR of 29.3%. |
| SM005 | International Federation of Robotics (IFR) | United States Back on Growth Track — Preliminary Robot Installation Results 2025 | The number of industrial robot installations in the United States rose by 11% year-on-year, to reach 38,000 units in 2025. China's annual installations reached 295,000 units in 2024, representing a global market share of 54%. |
| SM006 | Grand View Research | Industrial Automation and Control Systems Market Size, Share & Trends Report, 2030 | The global industrial automation and control systems market size was estimated at USD 206.33 billion in 2024 and is projected to reach USD 378.57 billion by 2030, growing at a CAGR of 10.8% from 2025 to 2030. |
| SM007 | NVIDIA Corporation | Physical AI Solutions — NVIDIA | |
| SM008 | Statista | Artificial Intelligence — Worldwide Market Outlook | |
| SM009 | MIT Technology Review | Yann LeCun's New Venture, AMI Labs | Think about complex industrial processes where you have thousands of sensors, like in a jet engine, a steel mill, or a chemical factory. There is no technique right now to build a complete, holistic model of these systems. A world model could learn this from the sensor data and predict how the system will behave. |
| SM010 | Futurum Group | Yann LeCun's AMI Raises $1Bn Seed Round: Is the World Model Era Finally Here? | The risk is not team quality — it is the structural tension between a research-first mandate and investor expectations calibrated to a billion-dollar raise. |
| SM011 | Cathay Innovation | Advanced Machine Intelligence (AMI) is Enabling the Next AI Revolution Built on Foundational World Models | World models represent a fundamental shift in how AI understands and interacts with reality: systems that reason about cause and effect, enabling intelligence to work in the physical world at scale. |
| SM012 | Forbes | Why Yann LeCun's Hot New AI Startup Is Targeting Healthcare | Healthcare is my baby, and we know what problems we cannot solve today. We hope that this new branch of AI will help us move beyond what we can do today in healthcare. |
| SM013 | TechCrunch | Yann LeCun's AMI Labs Raises $1.03 Billion to Build World Models | |
| SM014 | Nabla | Nabla Announces Exclusive Partnership with Advanced Machine Intelligence to Pioneer the Next Era of Agentic Healthcare AI | |
| SM015 | EU-Startups | Beyond LLMs: AI Pioneer Yann LeCun's New Venture AMI Raises €890 Million to Build World Model AI Systems | AMI claims that it will enhance AI research and develop applications focused on reliability, controllability, and safety, particularly in industrial process control, automation, wearable devices, robotics, healthcare, and more. |
| SM016 | SiliconAngle | Yann LeCun's New Startup AMI Labs Raises $1.03B to Train World Models | |
| SM017 | TechCrunch | Who's Behind AMI Labs, Yann LeCun's World Model Startup? | |
| SM018 | Dataconomy | Yann LeCun's AMI Labs Hits $3.5 Billion Pre-Money Valuation | |
| SM019 | Artificial Intelligence News | The Billion-Dollar Startup with a Different Idea for AI: AMI Labs and Yann LeCun | |
| SM020 | arXiv / Meta FAIR | Self-Supervised Learning from Images with a Joint-Embedding Predictive Architecture (I-JEPA) | We introduce the Image-based Joint-Embedding Predictive Architecture (I-JEPA), a non-generative approach for self-supervised learning from images. The idea behind I-JEPA is simple: from a single context block, predict the representations of various target blocks in the same image. |
| SM021 | France 24 | French AI Startup AMI Raises $1 Billion to Develop Universal Intelligent Systems | |
| SM022 | Euronews | AMI, une startup française de l'IA, annonce une levée de fonds d'un milliard de dollars | |
| SM023 | AMI Labs (Official) | AMI Labs Official Website | |
| SM024 | AMI Labs (Official) | AMI Labs — Updates and Announcements | |
| SM025 | TechCrunch | Yann LeCun Confirms His New World Model Startup, Reportedly Seeks $5B Valuation | |
| SP001 | World Labs | World Labs — Spatial Intelligence Homepage | World Labs is a leading spatial intelligence company, building frontier models that can perceive, generate, reason, and interact with the 3D world. |
| SP002 | World Labs | World Labs Research & Insights Blog | January 21, 2026 — Announcing the World API: A public API for generating explorable 3D worlds from text, images, and video. |
| SP003 | Odyssey | About Odyssey — AI Lab Pioneering General World Models | We're an AI lab pioneering general world models, and believe a new and powerful form of intelligence will emerge from learning all the beauty, physics, and intelligence of our world. |
| SP004 | Odyssey | Odyssey — World Model Products | Odyssey-2: Our most powerful general purpose world model yet, materially advancing the state-of-the-art in physical accuracy of world models. |
| SP005 | SpAItial | SpAItial — Frontier World Models for 3D Space | SpAItial is building physically-grounded world models. Our Echo model family outputs persistent 3D worlds that you can explore in real time. |
| SP006 | SpAItial | SpAItial Blog — Research, Announcements & Tutorials | Introducing SpAItial: We're launching SpAItial: a new AI company building systems that create and understand 3D worlds. Learn about our Spatial Foundation Models. |
| SP007 | Google DeepMind | Genie 2: A Large-Scale Foundation World Model | Genie 2, a foundation world model capable of generating an endless variety of action-controllable, playable 3D environments for training and evaluating embodied agents. |
| SP008 | Google DeepMind | Gemini Robotics — Embodied AI Model | Gemini Robotics models allow robots of any shape and size to perceive, reason, use tools and interact with humans. |
| SP009 | Google DeepMind | Genie: Generative Interactive Environments (Research Publication) | We introduce Genie, the first generative interactive environment trained in an unsupervised manner from unlabelled Internet videos. At 11B parameters, Genie can be considered a foundation world model. |
| SP010 | Meta AI | V-JEPA: The Next Step Toward Advanced Machine Intelligence | V-JEPA is a non-generative model that learns by predicting missing or masked parts of a video in an abstract representation space. |
| SP011 | Meta AI | AI Research: Introducing Muse Spark — New Foundation Model | |
| SP012 | Physical Intelligence | Physical Intelligence (π) — Company Homepage | Physical Intelligence is bringing general-purpose AI into the physical world. |
| SP013 | Physical Intelligence | Our First Generalist Policy — π0 | We are living through an AI revolution… we need to make AI systems embodied so that they can acquire physical intelligence. |
| SP014 | Physical Intelligence | π0.7: A Steerable Model with Emergent Capabilities | π0.7 is a general-purpose model that can perform a wide range of dexterous tasks with the same performance as fine-tuned specialists… it can follow new language commands and perform tasks that were never seen in its training data. |
| SP015 | Physical Intelligence | Physical Intelligence Blog — All Posts | The Physical Intelligence Layer — February 24, 2026: General-purpose physical intelligence models will enable a Cambrian explosion of robotics applications. See how our partners are already solving real-world problems. |
| SP016 | OpenAI | Sora System Card | Sora serves as a foundation for models that can understand and simulate the real world, a capability we believe will be an important milestone for achieving AGI. |
| SP017 | OpenAI | Video Generation Models as World Simulators | |
| SP018 | OpenAI | What to Know About the Sora Discontinuation | The Sora web and app experiences were discontinued on April 26, 2026. The Sora API will be discontinued on September 24, 2026. |
| SP019 | NVIDIA | NVIDIA Robotics Research | GEAR (Generalist Embodied Agent Research) builds foundation models for embodied agents across real and virtual worlds. Research spans vision-language models, world models, general-purpose robotics, large action models, and scalable simulation. |
| SP020 | NVIDIA | NVIDIA Cosmos — Physical AI Foundation Model Platform | Cosmos 3… Use as a vision language model (VLM) to reason over objects, interactions, and intent across complex real-world scenarios. Build Policy Models. Accelerate robot policy learning with NVIDIA Cosmos 3 as the backbone for World Action Models. |
| SP021 | NVIDIA | NVIDIA Isaac Platform — AI Robotics Development | NVIDIA Isaac is the ideal place to start. This open robotics development platform consists of simulation and robot learning frameworks, NVIDIA CUDA-accelerated libraries, AI models, and reference workflows. |
| SP022 | NVIDIA (Blog) | NVIDIA Makes Cosmos World Foundation Models Openly Available to Physical AI Developer Community | Cosmos won Best AI and Best Overall accolades from the Best of CES Awards… trained on 9,000 trillion tokens from 20 million hours of real-world human interactions, environment, industrial, robotics and driving data. |
| SP023 | Wayve | Wayve — Reimagining Autonomous Driving with Embodied AI | Wayve is building a general-purpose driving intelligence that learns from data and scales across vehicles, geographies, and applications. |
| SP024 | Wayve | Wayve AV2.0 Technology — End-to-End Embodied AI | Wayve's AV2.0 Approach — Our innovative approach replaces the modular 'sense-plan-act' architecture of the traditional AV1.0 approach with a single neural network trained on diverse data. |
| SP025 | Meta AI Research (GitHub) | I-JEPA — Official codebase for Image-based Joint-Embedding Predictive Architecture | I-JEPA is a method for self-supervised learning… The predictor in I-JEPA can be seen as a primitive (and restricted) world-model that is able to model spatial uncertainty in a static image from a partially observable context. |
| SP026 | HuggingFace | LeRobot — Making AI for Robotics More Accessible with End-to-End Learning | LeRobot aims to provide models, datasets, and tools for real-world robotics in PyTorch. The goal is to lower the barrier to entry so that everyone can contribute to and benefit from shared datasets and pretrained models. |
| SP027 | arXiv / CVPR 2023 | Self-Supervised Learning from Images with a Joint-Embedding Predictive Architecture (I-JEPA) | From a single context block, predict the representations of various target blocks in the same image. I-JEPA is highly scalable… achieves strong downstream performance across a wide range of tasks. |
| SI001 | AMI Labs | AMI Labs — Updates (Official Funding Announcement) | 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. |
| SI002 | TechCrunch | Yann LeCun's AMI Labs raises $1.03 billion to build world models | Value-add aside, this funding will give AMI Labs some meaningful runway to bankroll its two main cost centers: compute and talent. |
| SI003 | Observer | Yann LeCun's Paris A.I. Startup AMI Labs Raises Record $1B Seed Round | 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. |
| SI004 | TechCrunch | Yann LeCun confirms his new world model startup reportedly seeks $5B valuation | |
| SI005 | TechCrunch | Who's behind AMI Labs, Yann LeCun's world model startup? | |
| SI006 | Futurum Group | Yann LeCun's AMI Raises $1BN Seed Round — Is the World Model Era Finally Here? | Follow-on financing structure: The $1.03 billion will fund an extended research phase, but Series A terms (when they come) will be the first real market test of whether AMI's research output translates to commercial credibility. |
| SI007 | Dataconomy | Yann LeCun's AMI Labs hits $3.5 billion pre-money valuation | |
| SI008 | France 24 | French AI startup AMI raises $1B to develop 'universal intelligent systems' | LeCun…said AMI would focus on research and development in its first year. Discussions with corporate partners could be held within six to 12 months. |
| SI009 | Euronews | AMI, une startup française de l'IA, annonce une levée de fonds d'un milliard de dollars | |
| SI010 | Goldman Sachs | Gen AI: too much spend, too little benefit? | The promise of generative AI technology to transform companies, industries, and societies is leading tech giants and beyond to spend an estimated ~$1tn on capex in coming years, including significant investments in data centers, chips, other AI infrastructure, and the power grid. But this spending has little to show for it so far. |
| SI011 | Sequoia Capital | AI's $600B Question | AI's $200B question is now AI's $600B question…GPU computing is increasingly turning into a commodity, metered per hour…speculative investment frenzies often lead to high rates of capital incineration. |
| SI012 | Epoch AI | Training compute of frontier AI models grows by 4-5x per year | We tentatively conclude that compute growth in recent years is currently best described as increasing by a factor of 4-5x/year. |
| SI013 | Epoch AI | Trends in Artificial Intelligence — Compute, Data Centers, Hardware | AI chip performance per dollar has improved by roughly 40% per year across over 20 AI accelerators released between 2012 and 2025. |
| SI014 | AI News (TechForge Media) | The billion-dollar startup with a different idea for AI: AMI Labs and Yann LeCun | The smaller, focused modules inside AMI Labs' proposed solution could be run on fraction of the GPU power currently necessary for giant LLMs, or even on-device. |
| SI015 | SiliconAngle | Yann LeCun's new startup AMI Labs raises $1.03B to train world models | |
| SI016 | NVIDIA Newsroom | NVIDIA Invests in AMI Labs | |
| SI017 | HV Capital | HV Capital — Portfolio | |
| SI018 | SmartCompany | Yann LeCun's AMI Labs raises $1bn for world model AI | |
| SI019 | Analytics India Magazine | Yann LeCun's AMI Labs raises $1 billion | |
| SI020 | AMI Labs | AMI Labs Official Website | |
| SI021 | Nabla | Nabla Announces Exclusive Partnership with Advanced Machine Intelligence | |
| SI022 | US Securities and Exchange Commission (EDGAR) | Form D — Model I, L.P. / Advanced Machine Intelligence I, L.P. (CIK 0002102055) | |
| SI023 | MIT Technology Review | Yann LeCun's new venture, AMI Labs, wants to build better AI | |
| SI024 | MIT Technology Review | AMI Labs world models funding | |
| SI025 | HV Capital | HV Capital — Companies | |
| SI026 | LessWrong | chinchilla's wild implications | |
| SI027 | Stanford HAI | AI Index Report 2025 | |
| SI028 | US Securities and Exchange Commission (EDGAR) | EDGAR Full-Text Search — Advanced Machine Intelligence Form D filings | |
| SE001 | AMI Labs | AMI Labs: Real World. Real Intelligence. — Official Homepage | "Action-conditioned world models allow agentic systems to predict the consequences of their actions, and to plan action sequences to accomplish a task, subject to safety guardrails." |
| SE002 | AMI Labs | AMI Labs Updates — Funding Announcement | |
| SE003 | Yann LeCun (OpenReview) | A Path Towards Autonomous Machine Intelligence (Version 0.9.2) | "This position paper proposes an architecture and training paradigms with which to construct autonomous intelligent agents. It combines concepts such as configurable predictive world model, behavior driven through intrinsic motivation, and hierarchical joint embedding architectures trained with self-supervised learning." |
| SE004 | arXiv (Assran et al., CVPR 2023) | Self-Supervised Learning from Images with a Joint-Embedding Predictive Architecture (I-JEPA) | "We train a ViT-Huge/14 on ImageNet using 16 A100 GPUs in under 72 hours to achieve strong downstream performance across a wide range of tasks." |
| SE005 | Meta AI | V-JEPA: The next step toward advanced machine intelligence | "V-JEPA is a non-generative model that learns by predicting missing or masked parts of a video in an abstract representation space… proving more efficient than previous models, both in terms of the number of labeled examples needed and the total amount of effort put into learning even the unlabeled data." |
| SE006 | Meta AI Research (GitHub) | facebookresearch/ijepa — Official PyTorch codebase for I-JEPA | "The predictor in I-JEPA can be seen as a primitive (and restricted) world-model that is able to model spatial uncertainty in a static image from a partially observable context." |
| SE007 | 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." |
| SE008 | TechCrunch | Who's behind AMI Labs, Yann LeCun's 'world model' startup | "The startup plans to license its technology to industry partners for real-life applications, but says it also plans to contribute to building the future of AI 'with the global academic research community via open publications and open source.'" |
| SE009 | 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… it could take years for world models to go from theory to commercial applications." |
| SE010 | Yann LeCun | Yann LeCun Personal Academic Website | |
| SE011 | Saining Xie | Saining Xie — Personal Academic Website | |
| SE012 | Futurum Group | Yann LeCun's AMI Raises $1BN Seed Round — Is the World Model Era Finally Here? | "The world model era may be here in theory, but AMI Labs has yet to demonstrate that JEPA can bridge the gap from academic benchmarks to production-grade deployments." |
| SE013 | NVIDIA Blog | NVIDIA Makes Cosmos World Foundation Models Openly Available to Physical AI Developer Community | "Cosmos world foundation models are a suite of open diffusion and autoregressive transformer models for physics-aware video generation… trained on 9,000 trillion tokens from 20 million hours of real-world human interactions." |
| SE014 | AMI Labs (Medium) | AmiLabs — Official Medium Publication | |
| SE015 | AI as Normal Technology (Sayash Kapoor and Arvind Narayanan) | AI as Normal Technology Newsletter — World Models Skeptical Coverage | "World models face a significant gap between academic demonstrations and the scale of domain-specific data, safety validation, and regulatory engineering required for production deployment in safety-critical industries." |
| SE016 | Wayve | Wayve Thinking Blog — Embodied AI and World Models | |
| SE017 | Forbes | Why Yann LeCun's Hot New AI Startup Is Targeting Healthcare | |
| SE019 | 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… trained on a large-scale video dataset and demonstrates various emergent capabilities at scale." |
| SE020 | OpenAI Help Center | What to know about the Sora discontinuation | The Sora web and app experiences were discontinued on April 26, 2026. |
| SE021 | Physical Intelligence | Physical Intelligence (π) — Official Website, April 2026 Update | "π0.7: a Steerable Model with Emergent Capabilities — A steerable robotic foundation model that exhibits a step-change in generalization." |
| SE022 | Epoch AI | Trends in Artificial Intelligence — Training Compute and Model Scale | "Training compute for frontier language models has been growing at 5× per year since 2020. Training costs are climbing by 3.5× annually." |
| SE023 | TechCrunch | AMI Labs tag page — TechCrunch coverage index | |
| SE024 | SpAItial | SpAItial — Frontier World Models for 3D Space (Echo-2 announcement) | "SpAItial is building Echo, a frontier world model for persistent 3D Gaussian Splat worlds." |
| SE025 | TechCrunch | Yann LeCun confirms his new 'world model' startup, reportedly seeks $5B valuation | |
| SE026 | Technology Review | Yann LeCun's New Venture: AMI Labs | |
| SU001 | Nabla | Blog — Nabla Customer Stories (May–June 2026) | Customer Stories: Catalight Expands Access and Equity in Behavioral Health with Nabla (May 2026); LCMC Health Improves Clinician Efficiency and Retention with Nabla (April 2026) |
| SU002 | Nabla | Nabla — Enjoy Care Again (Homepage, June 2026) | Deployed in 190+ health organizations. Loved by 100,000 clinicians. |
| SU003 | U.S. Food and Drug Administration | AI-Enabled Medical Devices (FDA CDRH) | |
| SU004 | Rock Health | 2025 Year-End Digital Health Funding Overview: A Tale of Two Markets | Healthcare is embracing AI at an impressive pace: provider adoption is ballooning, sales cycles are shortening, and agentic AI is being adopted faster in healthcare than many other industries. |
| SU005 | Menlo Ventures | 2025: The State of AI in Healthcare | Health systems have shortened average buying cycles from 8.0 months for traditional IT purchases to 6.6 months. Payers have seen buying cycles lengthen from 9.4 months to 11.3 months. |
| SU006 | WIRED | Yann LeCun Raises $1 Billion to Build AI That Understands the Physical World | LeCun says AMI aims to work with companies in manufacturing, biomedical, robotics, and other industries that have lots of data. For example, he says AMI could build a realistic world model of an aircraft engine and work with the manufacturer to help them optimize for efficiency. |
| SU007 | New York Magazine (Intelligencer) | How Long Will AI's Free-Trial Era Last? | We are in the era of $5 Uber rides anywhere across San Francisco but for LLMs — large incumbents subsidizing AI adoption to grab share, creating price pressure on any new entrant. |
| SU008 | The BMJ | TRIPOD+AI: updated guidance for reporting clinical prediction models using machine learning | External validation, calibration, and bias assessment are required before clinical AI models can be deployed; standards that unvalidated research-phase systems cannot yet meet. |
| SU009 | Nabla | Nabla Announces Exclusive Partnership with Advanced Machine Intelligence to Pioneer Agentic Healthcare AI | Through this partnership, 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. |
| SU010 | TechCrunch | Yann LeCun's AMI Labs Raises $1.03 Billion 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 ARR in 12 months. |
| SU011 | TechCrunch | Who's Behind AMI Labs, Yann LeCun's World Model Startup? | Nabla is the first disclosed partner expecting to access these early models, but definitely not the last. |
| SU012 | 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. |
| SU013 | Observer | Yann LeCun's AI Startup AMI Raises $1 Billion Seed Round | AMI plans to focus first on advancing world model research before pursuing commercial applications, with early use cases likely in factories and hospitals. |
| SU014 | France24 | French AI startup AMI raises $1 billion to develop universal intelligent systems | Discussions with corporate partners could be held within six to 12 months, he added. |
| SU015 | Futurum Group | Yann LeCun's AMI Raises $1BN Seed Round — Is the World Model Era Finally Here? | Healthcare AI development, particularly any pathway to FDA certification, is a long and uncertain process. The company's stated focus on industrial process control, robotics, and wearables as additional verticals provides diversification, but none of these markets can absorb a world-model-as-a-service offering on a short timeline. |
| SU016 | EU Startups | Beyond LLMs: Yann LeCun's new venture AMI raises €890 million to build world model AI systems | |
| SU017 | AMI Labs | AMI Labs Funding Update | AMI is also supported by a group of long-term global investors and strategic backers, including Toyota Ventures, New Legacy Ventures, Temasek, SBVA, NVIDIA, Mark Cuban, Association Familiale Mulliez, Groupe industriel Marcel Dassault, Sea, and Alpha Intelligence Capital. |
| SU018 | AMI Labs | AMI Labs Homepage — 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. |
| SU019 | Cathay Innovation | Advanced Machine Intelligence (AMI) Is Enabling the Next AI Revolution Built on Foundational World Models | AMI is also supported by a group of long-term global investors and strategic backers, including Toyota Ventures, Temasek, SBVA, NVIDIA, Groupe industriel Marcel Dassault, Sea. |
| SU020 | Artificial Intelligence News | The Billion Dollar Startup with a Different Idea for AI: AMI Labs | |
| SU021 | Analytics India Magazine | Yann LeCun's AMI Labs Raises $1 Billion Dollars | |
| SU022 | SiliconANGLE | Yann LeCun's New Startup AMI Labs Raises $1.03B to Train World Models | |
| SU023 | Goldman Sachs | Gen AI: Too Much Spend, Too Little Benefit? | The promise of generative AI technology is leading tech giants and beyond to spend an estimated ~$1tn on capex in coming years, but this spending has little to show for it so far. |
| SU024 | Sequoia Capital | AI's $600B Question | AI's $200B question is now AI's $600B question. The $125B hole is now a $500B hole. OpenAI still has the lion's share of AI revenue while everyone else lags far behind. |
| SU025 | Euronews | AMI: une startup française de l'IA annonce une levée de fonds d'un milliard de dollars | |
| SR001 | European Commission — Digital Strategy | Regulatory Framework for AI — AI Act | "AI safety components in products (e.g. AI application in robot-assisted surgery) and AI-based safety components of critical infrastructures are classified as high-risk." |
| SR002 | EUR-Lex — European Union | Regulation (EU) 2024/1689 — Artificial Intelligence Act (full legal text) | "A general-purpose AI model shall be classified as a general-purpose AI model with systemic risk if it meets any of the following conditions: (a) it has high impact capabilities evaluated on the basis of appropriate technical tools and methodologies." |
| SR003 | European Parliament | Artificial Intelligence Act: MEPs adopt landmark law | "GPAI systems and the GPAI models they are based on, must meet certain transparency requirements, including compliance with EU copyright law and publishing detailed summaries of the content used for training." |
| SR004 | artificialintelligenceact.eu (Future of Life Institute / AI Act analysis) | The Act Texts — EU Artificial Intelligence Act | |
| SR005 | artificialintelligenceact.eu | Article 51 — Classification of GPAI Models with Systemic Risk | "A general-purpose AI model shall be presumed to have high impact capabilities pursuant to paragraph 1, point (a), when the cumulative amount of computation used for its training measured in floating point operations is greater than 10^25." |
| SR006 | Gary Marcus (Marcus on AI, Substack) | What if Generative AI turned out to be a Dud? | "The valuations anticipate trillion dollar markets, but the actual current revenues from generative AI are rumored to be in the hundreds of millions. Even OpenAI could have a hard time following through on its valuation; competing startups valued in the low billions might well eventually collapse, if year after year they manage only tens or hundreds of millions in revenue." |
| SR007 | Gary Marcus (Marcus on AI, Substack) | Nonsense on Stilts | "What is easy for us, like perception and navigation, is hard for computers, and vice versa. LLMs are limited to the discrete world of text. They can't truly reason or plan, because they lack a model of the world. They can't predict the consequences of their actions." |
| SR008 | Bureau of Industry and Security (US Department of Commerce) | Licensing — Do I Need an Export License? | |
| SR009 | OECD | AI Principles | "AI actors should ensure traceability, including in relation to datasets, processes and decisions made during the AI system lifecycle, to enable analysis of the AI system's outputs and responses to inquiry." |
| SR010 | Brookings Institution | How Artificial Intelligence is Transforming the World | |
| SR011 | Epoch AI | Training Compute of Frontier AI Models Grows by 4–5× per Year | "We tentatively conclude that compute growth in recent years is currently best described as increasing by a factor of 4–5x/year." |
| SR012 | 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. I am going to keep my position at NYU. I teach one class per year, I have PhD students and postdocs, so I am going to be kept based in New York." |
| SR013 | TechCrunch | Who's behind AMI Labs, Yann LeCun's 'world model' startup | "LeBrun's transition from Nabla to AMI is part of a partnership announced last December by Nabla, which develops AI assistants for clinical care and to which LeCun has been an advisor. In exchange for 'privileged access' to AMI's world models, Nabla's board supported LeBrun's shift from CEO to chief AI scientist and chairman." |
| SR014 | TechCrunch | Yann LeCun's AMI Labs raises $1.03 billion to build world-model AI | |
| SR015 | Goldman Sachs | Gen AI: too much spend, too little benefit? | "Tech giants and beyond to spend an estimated ~$1tn on capex in coming years, including significant investments in data centers, chips, other AI infrastructure, and the power grid. But this spending has little to show for it so far." |
| SR016 | Sequoia Capital | AI's $600B Question | "In the case of GPU data centers, there is much less pricing power. GPU computing is increasingly turning into a commodity, metered per hour. Without a monopoly or oligopoly, high fixed cost + low marginal cost businesses almost always see prices competed down to marginal cost." |
| SR017 | Observer | Yann LeCun's Paris A.I. Startup AMI Labs Raises Record $1B Seed Round | "In November, Yann LeCun left Meta after 12 years over disagreements with Mark Zuckerberg over the future of A.I. Frustrated with the limitations of large language models (LLMs), the French computer scientist founded AMI Labs." |
| SR018 | New York Magazine / Intelligencer | How Long Will AI's Free-Trial Era Last? | "OpenAI, Google, Meta, Microsoft, and smaller firms like Anthropic are losing massive amounts of money by giving away their AI products or selling them at a loss. We are in the era of $5 Uber rides anywhere across San Francisco but for LLMs." |
| SR019 | Futurum Group | Yann LeCun's AMI Raises $1BN Seed Round: Is the World Model Era Finally Here? | |
| SR020 | AMI Labs (official) | AMI Labs — Updates / Mission Statement | |
| SR021 | U.S. Securities and Exchange Commission (EDGAR) | Form D — Advanced Machine Intelligence I, L.P. / Model I, L.P. | |
| SR022 | Epoch AI | Training Compute of Frontier AI Models (blog, earlier version) | |
| SR023 | France 24 | French AI startup AMI raises $1B to develop universal intelligent systems | |
| SR024 | SiliconAngle | Yann LeCun's new startup AMI Labs raises $1.03B to train world models | |
| SR025 | Dataconomy | Yann LeCun's AMI Labs hits $3.5B pre-money valuation with $1B seed | |
| SR026 | Artificial Intelligence News | The billion-dollar startup with a different idea for AI: AMI Labs, Yann LeCun | |
| SR027 | Analytics India Magazine | Yann LeCun AMI Labs raises $1 billion dollars | |
| SR028 | U.S. Food and Drug Administration | FDA Releases Artificial Intelligence/Machine Learning Action Plan | "This action plan describes a multi-pronged approach to advance the Agency's oversight of AI/ML-based medical software." |
| SR029 | NVIDIA Newsroom | NVIDIA Invests in AMI Labs | |
| SR030 | AMI Labs (official) | AMI Labs — Main Site | |
| SR031 | ACM | Fathers of the Deep Learning Revolution Receive 2018 ACM Turing Award | |
| SR032 | Futurumgroup | Yann LeCun's AMI raises $1BN seed round | |
| SR033 | EU-Startups | Beyond LLMs: AI pioneer Yann LeCun's new venture AMI raises $1 billion for world models | |
| SR034 | Euronews (French) | AMI: une startup française de l'IA annonce une levée de fonds d'un milliard de dollars | |
| SR035 | Gary Marcus (Marcus on AI, Substack) | Marcus on AI — Substack homepage | |
| SV001 | Andreessen Horowitz (a16z) | What's In a World? Investing in World Labs | "We deeply believe that what this team will produce will have as deep an impact as the LLM revolution." |
| SV002 | SiliconAngle | China's DeepSeek reportedly raises $7.4B in funding at $50B+ valuation | "The Information and the Wall Street Journal today cited sources as saying that the company is now worth over $50 billion." |
| SV003 | Dr. Fei-Fei Li (Substack) | From Words to Worlds: Spatial Intelligence is AI's Next Frontier | "World Labs was founded in early 2024 on this conviction: that foundational approaches are still being established, making this the defining challenge of the next decade." |
| SV004 | Thinking Machines Lab | Thinking Machines Lab — About | |
| SV005 | CB Insights | AI 100: The most promising artificial intelligence startups of 2026 | "Physical AI enters the AI 100 as a standalone category for the first time, with 11 companies spanning robotics software, autonomous hardware, and enabling chips." |
| SV006 | The Information | DeepSeek Closes Record $7 Billion-Plus Funding with Unusual Deal Structure | |
| SV007 | U.S. Securities and Exchange Commission (SEC) | C3.ai, Inc. Annual Report on Form 10-K — Fiscal Year Ended April 30, 2025 | "Total revenue was $389.1 million for the fiscal year ended April 30, 2025, representing a 25.3% increase compared to the prior fiscal year." |
| SV008 | Greycroft | AMI Labs Portfolio — Greycroft | "Advanced Machine Intelligence (AMI) is a research lab developing world-model AI that learns structured representations of the physical world." |
| SV009 | AMI Labs | AMI Labs — Updates | |
| SV010 | TechCrunch | Yann LeCun's AMI Labs raises $1.03B to build world models | "AMI Labs, the new venture co-founded by Turing Award winner Yann LeCun after he left Meta, has raised $1.03 billion at a $3.5 billion pre-money valuation." |
| SV011 | Observer | Yann LeCun's Paris A.I. Startup AMI Labs Raises Record $1B Seed Round | "Other key players include Fei-Fei Li's World Labs, which raised $1 billion last month, and General Intuition, which raised $134 million in October." |
| SV012 | WIRED | Yann LeCun Raises $1 Billion to Build AI That Understands the Physical World | "The financing, which values the startup at $3.5 billion, was co-led by investors such as Cathay Innovation, Greycroft, Hiro Capital, HV Capital, and Bezos Expeditions." |
| SV013 | SiliconAngle | Yann LeCun's new startup AMI Labs raises $1.03B to train world models | "AMI Labs is now valued at $3.5 billion." |
| SV014 | Futurum Group | Yann LeCun's AMI Raises $1BN Seed Round — Is the World Model Era Finally Here? | "Former OpenAI CTO Mira Murati's Thinking Machines Lab was valued at $12 billion in its seed round. The AMI raise is not the largest in absolute terms, but it represents a structurally different profile." |
| SV015 | Goldman Sachs | Gen AI: too much spend, too little benefit? | "The promise of generative AI technology... is leading tech giants and beyond to spend an estimated ~$1tn on capex in coming years... But this spending has little to show for it so far." |
| SV016 | Sequoia Capital | AI's $600B Question | "AI's $200B question is now AI's $600B question... The $125B hole is now a $500B hole." |
| SV017 | Dataconomy | Yann LeCun's AMI Labs Hits $3.5 Billion Pre-money Valuation | "NVIDIA has been investing heavily in AI startups, pouring roughly $53 billion across 170 deals according to PitchBook data cited by Forbes." |
| SV018 | Epoch AI | Training compute of frontier AI models grows by 4-5x per year | "We tentatively conclude that compute growth in recent years is currently best described as increasing by a factor of 4-5x/year." |
| SV019 | World Labs | World Labs — Home | |
| SV020 | TechCrunch | Yann LeCun's AMI Labs raises $1.03B to build world models (fundraising history) | "The French AI lab was reportedly seeking just €500 million last December, but ended up raising some €890 million, likely thanks to its team." |
| SV021 | TechCrunch | Who's behind AMI Labs, Yann LeCun's world model startup? | |
| SV022 | EU-Startups | Beyond LLMs: AMI raises €890 million — sovereign AI positioning | |
| SV023 | Odyssey | Odyssey — World Model | |
| SV024 | Cathay Innovation | Advanced Machine Intelligence (AMI) — Cathay Innovation Portfolio | |
| SV025 | Hiro Capital | Hiro Capital — Portfolio | |
| SV026 | EU-Startups | Beyond LLMs: AI pioneer Yann LeCun's new venture AMI raises €890 million to build world model AI systems | |
| SV027 | AI Artificial Intelligence News | The billion-dollar startup with a different idea for AI: AMI Labs and Yann LeCun | |
| SV028 | SmartCompany | AMI Labs, Yann LeCun world models $1 billion funding $3.5 billion valuation | |
| SV029 | Futurum Research | Yann LeCun's AMI Raises $1BN — World Model Era (Analyst commentary on sovereign AI) | "Research labs that have raised at comparable scales… have generally required either a deep-pocketed patron or a defined path to revenue." |
| SV030 | Siliconangle | Yann LeCun's AMI Labs — NVIDIA and Samsung backing context | |
| SV032 | WIRED | Yann LeCun Raises $1 Billion — LeCun on commercial timeline and JEPA |