General Intuition
数据故事有差异化的前沿 AI 实验室,但估值跑在公开验证之前
技术可信的前沿 AI spinout,靠差异化游戏玩法数据立论;但在公开客户与经济性证据出来之前,当前 $2.3B 估值已经跑在前面。
封面要素
公司概况
General Intuition 是从 Medal 拆分出的前沿 AI 实验室,Medal 是 Pim de Witte 创办的游戏片段平台。公开报道把公司同时指向纽约运营中心、荷兰法律和 IP 架构,以及 从 2025 年启动轮到 2026 年 6 月宣布 $320M Series A、估值 $2.3B 的快速融资跃迁。产品论点是用游戏视频和动作标签训练行动模型和世界模型,再通过面向游戏、仿真和机器人方向的选择性合作伙伴 API 商业化。MIRA 及配套研究资产已经公开,对一家年轻公司而言,公开技术验证异常强;但客户验证、收入质量和经济性披露仍很薄。
- 成立时间
- 2025-10-01
- 创始人
- Pim de Witte, Eloi Alonso, Adam Jelley, Vincent Micheli
- 创立地点
- New York, NY, USA
- 总部
- New York, NY, USA
- 产品
- 行动模型和世界模型栈,以 Medal 游戏片段和动作标签为底座;面向游戏、仿真和机器人提供选择性商业 API,并通过 MIRA 和 Nerve 数据采集界面展示可见技术验证。
- 客户
- 游戏、仿真和机器人领域技术能力较强的合作伙伴;他们能共同开发用例,并最终把选择性部署转成更广泛的生产使用。
- 商业模式
- 高触达、合作伙伴主导的商业化:选择性 API 访问、定制集成和共同开发,同时扩展数据采集 / 标注界面;目前还不是广泛自助式 SaaS 路线。
- 阶段
- Series A (private, venture-backed)
- 融资情况
- 2026 年 6 月宣布 $320M Series A,投后估值 $2.3B;自 2025 年 10 月启动轮以来,已披露融资总额约 $454M。
执行摘要
主要优势
- Medal 关联训练语料差异化,且以带动作标签的游戏玩法数据支撑智能体 AI 逻辑。
- 投资人阵容强,资本也够厚,能支撑高算力研究和商业化。
- MIRA 及相关研究成果已经可见,因此这家年轻实验室的公开技术证明强于常态。
主要风险
- 在 $2.3B 估值下,还没有公开收入、毛利、留存或具名生产客户证明。
- 业务资本强度高,又依赖 Medal 数据和 CoreWeave 支撑的算力,执行对合作伙伴与基础设施风险很敏感。
- DeepMind、OpenAI、NVIDIA、World Labs、Physical Intelligence 的竞争可能在大规模发布前压缩差异化。
未决问题
- 股权结构表、清算优先权以及下一轮是否有结构性保护尚未披露。
- 算力成本曲线、毛利率路径、烧钱速度 / 现金跑道仍未知。
- 具名客户名单、合同结构和续约证据尚未公开。
- 训练数据权利链和按用例划分的监管映射仍需私下尽调。
目录
01公司概览
1.1 身份、论点和产品姿态
General Intuition 公开把自己定位成一个让系统在空间和时间中行动的前沿实验室,而不是又一家文本或图像生成创业公司。公司官网把核心技术栈拆成两个相连系统:行动模型决定下一步做什么, 世界模型预测行动后的结果。官方叙事认为,单靠语言预训练不足以形成具身或智能体能力,因为真实世界能力依赖意图、动作和后果的连续序列。在这个框架里,游戏玩法不是面向消费者的副业, 而是数据引擎。General Intuition 称,自己的底座来自 Pim de Witte 创办的游戏片段平台 Medal;用户每年上传数十亿条游戏片段,这些片段可以配上精确动作标签,而不只是原始视频。 产品姿态对一家这么早期的公司来说也很具体。TechCrunch、TNW 和公司网站都说,General Intuition 主要把世界模型用作训练环境,智能体本身才是最终产品。这个区分有商业意义: 收入应来自已部署的决策系统或 API,而不只是仿真软件。官网称,公司已经让游戏、仿真和机器人领域的首批合作伙伴接入选择性商业 API;但公开材料仍没有给出伙伴名称或使用量。 结果是一家公司拥有清晰战略叙事和选择性合作伙伴计划,但在产品成熟度和变现上公开披露仍有限。[CO001, CO002, CO003, CO004, CO005, CO016]
| 指标 | 数值 / 状态 | 日期 | 置信度 | 缺口 |
|---|---|---|---|---|
| 创立 / 分拆时间 | 2025 年创立;2025 年 10 月从 Medal 分拆 | 2025-10 | 中 | |
| 最新公开估值(USD B) | 2.3 | 2026-06 | 中 | |
| 最新披露轮次 | $320M Series A 轮,由 Khosla Ventures 领投 | 2026-06 | 中 | |
| 披露融资总额 | $454M | 2026-06 | 中 | |
| 公开运营中心 | 纽约实验室 / 多篇报道称基地在纽约 | 2026-06 | 中 | 公开来源与提示给出的旧金山总部相冲突;法律结构另为荷兰。 |
| 法律 / IP 基础 | 荷兰公司,数据和 IP 据称位于 Naarden | 2026-06 | 中 | 母子公司架构和董事会权利仍未公开。 |
| 训练数据来源 | 数十亿条带动作标签的 Medal 游戏片段 | 2026-06 | 中 | |
| 商业化姿态 | 筛选式商业 API;首批伙伴覆盖游戏、仿真、机器人 | 2026-07 | 中 | 伙伴名称、合同条款和使用指标未公开。 |
| Medal 月活用户 | 公开引用为 10M 至 17M | 2026-06 | 低 | 公开报道对当前 MAU 数字不一致。 |
| 收入 / ARR | 未公开披露 | 2026-07 | 中 | 没有公开收入、ARR 或毛利率数据。 |
公开公司层面指标有方向性参考价值,但若干运营和法律数据点仍只部分披露,且不同来源存在冲突。
[CO006, CO007, CO021, CO022, CO024, CO028]公开可见规模集中在融资和数据,而不是收入;商业化仍处于选择性发布模式。
数据量行反映公开来源估计;收入行刻意呈空值态,因为公司尚未披露商业指标。
[CO007, CO019, CO028, CO031, CO034]1.2 创始人、研究可信度与治理不透明
公开报道一致认为,General Intuition 由 Pim de Witte 与 Eloi Alonso、Adam Jelley、Vincent Micheli 共同创办。De Witte 是运营重心:媒体报道、投资方资料页和会议简介都把他列为 General Intuition 创始人兼 CEO,也列为 Medal 创始人或前 CEO。他的经历结合了游戏创业和人道主义工作;他也借此阐述公司围绕武力使用的野心和边界。Eloi Alonso 的个人网站提供了研究侧 创始人-市场匹配的直接证据:他在日内瓦大学完成强化学习和世界模型方向 PhD 工作后,成为 General Intuition 联合创始人。 创始团队能发表前沿技术工作的最强独立证据,是 2026 年 MIRA 项目。Adam Jelley、Eloi Alonso、Vincent Micheli 和 Pim de Witte 都出现在这篇多人世界模型论文及开源发布的作者名单中,项目由 Kyutai 和 Epic Games 共同参与。这帮助验证,公司不只是一个持有数据的壳。即便如此,治理披露仍薄。公开页面列出创始人和投资方,却没有正式董事会名单、独立董事、保留事项权利或股权结构控制条款。 TechFundingNews 称,公司以在荷兰合法注册的公益公司形式运营;DutchNews 则称数据和 IP 放在 Naarden 的一家荷兰公司里。这些事实方向上可以并存,但母子公司图谱和董事会结构仍解释不足。[CO011, CO012, CO013, CO014, CO015, CO023]
| 人物 | 角色 | 背景 | 创始人-市场匹配或职能覆盖 | 关键人依赖 |
|---|---|---|---|---|
| Pim de Witte 创始人 | CEO 兼联合创始人 | Medal 创始人或前 CEO;此前从事游戏创业和人道主义领域运营 | 串起专有游戏数据集、公司叙事、投资人关系和运营策略 | 高 |
| Eloi Alonso | 联合创始人 | 强化学习与世界模型研究者;日内瓦大学 PhD 背景 | 提供直接的世界模型和仿真研究深度 | 高 |
| Adam Jelley | 联合创始人 / 技术人员 | 公司报道中被列为联合创始人,并在 MIRA 中列为 General Intuition 贡献者 | 把研究产出接到面向生产的世界模型系统 | 中 |
| Vincent Micheli | 联合创始人 | 融资报道中被列为联合创始人,并在 MIRA 中列为 General Intuition 贡献者 | 为基于扩散模型的仿真和世界模型工作增加技术可信度 | 中 |
本列举聚焦公开具名、且对产品方向和论点可信度最重要的创始人;并非完整高管名单。
[CO011, CO012, CO013, CO014, CO015, CO034]1.3 融资、投资方图谱和运营足迹
公开记录里最清晰的硬数据,是 2026 年 6 月的融资披露。TechCrunch、The SaaS News、DutchNews、The Robot Report、AI Insider 和 TechFundingNews 都指向同一组数字: $320 million Series A、估值 $2.3 billion;叠加 2025 年 10 月此前启动轮后,已披露融资约 $454 million。Khosla Ventures 被一致列为领投方,General Catalyst、Jeff Bezos、 Eric Schmidt 或 Hillspire、Nico Rosberg 则出现在不同辛迪加披露中。公开投资方页面提供了间接佐证:General Catalyst 将 General Intuition 列为投资组合公司,Backed VC 则显示该公司在 2025 年获得种子轮支持。 这套公开材料没有完全对齐的是时间和地理。DutchNews 称 Series A 在 2026 年 1 月完成,只是在 6 月对外公布;TechCrunch 6 月 18 日预告文章则说公司当时正在以略高于 $2 billion 的估值融资约 $300 million。这个顺序可以解释为一轮融资从市场沟通到交割、再到官宣的过程,但也提醒投资人不要过度解读单个时间戳文章。地点披露同样混杂:TechCrunch 和 TechFundingNews 多次把 General Intuition 描述为总部在纽约或围绕纽约实验室运转,DutchNews 则强调公司和 IP 仍属荷兰,同时在日内瓦、伦敦和巴黎也有办公室。最稳妥的读法是:General Intuition 运营重心在纽约,法律锚点在荷兰,并拥有分布式研究足迹。[CO006, CO007, CO008, CO009, CO020, CO021]
| 利益相关方 | 角色 | 控制权或经济重要性 | 尽调问题 |
|---|---|---|---|
| Khosla Ventures | 领投方 | 领投 Series A,并在多轮融资中反复被描述为主要支持者 | 厘清持股比例、董事会权利、按比例跟投权和任何结构化条款。 |
| General Catalyst | 既有投资人 | 出现在公开财团披露中,并列为组合公司投资人 | 确认 GC 是否持有治理权,还是仅有经济参与。 |
| Jeff Bezos / Bezos Expeditions | 战略性财务支持者 | 个人资本和信号价值提高与大型伙伴合作的选择权 | 确认该投资是否附带任何信息权或后续投资权。 |
| Eric Schmidt / Hillspire | 战略性财务支持者 | 除资本外,还增加 AI 生态可信度和网络触达 | 厘清参与是个人名义、通过 Hillspire,还是两者都有。 |
| Medal | 关联数据平台 | 提供支撑论点的专有游戏玩法和动作标签语料 | 审查数据所有权、公司间许可、排他性和少数股东权利。 |
| CoreWeave | 算力供应商 | 据称大部分新融资将通过 CoreWeave 交易投向算力容量 | 审查承诺支出、期限、预付款和集中度风险。 |
公开股权结构表不完整;这张图谱只隔离出对数据访问、资本形成和运营杠杆最重要的投资人和交易对手。
[CO008, CO009, CO020, CO026, CO027]General Intuition 将 Medal 源数据、世界模型训练、选择性 API 商业化和算力融资串成一个运营闭环。
[CO003, CO016, CO017, CO020, CO041]1.4 里程碑、早期商业化和未解风险
公司的公开历史被压缩得异常紧。Medal 2015 年成立;据报道 OpenAI 在 2024 年底曾尝试收购其数据集;General Intuition 于 2025 年 10 月拆分成立;到 2026 年 6 月, 公司已经宣布 $2.3 billion 估值,并通过 MIRA 拿出公开研究产出。TechCrunch 融资后画像补充了运营细节:同一模型家族被展示用于驱动游戏智能体和四足机器人,管理层称只需要几分钟真实世界数据即可微调。 公司还推出 Nerve,这是一个用于收集游戏玩法、标注和遥操作数据的市场,并称更广泛 API 可用性目标是 2026 年夏末。 这些进展并没有消除核心尽调风险。公开来源证明 General Intuition 拥有差异化数据源和快速推进的研究组织,但尚未证明游戏玩法预训练能够大规模迁移到稳健的真实世界机器人,也没有证明持续商业需求。 MIT Technology Review 2026 年 4 月的世界模型综述提供了最有用的怀疑侧对照:文章认为当前 AI 在物理世界中仍然吃力,世界模型论点还更像承诺,而不是已经坐实的能力。公开报道也在 Medal 当前月活用户数上互相冲突,分别引用了约 1000 万和约 1700 万用户。再叠加收入、董事会构成、客户名称和精确法律结构均未披露,公司概览应被视为一个有真实动能证据的高信念论点,而不是已经充分去风险的运营故事。[CO010, CO018, CO019, CO028, CO029, CO030]
| 日期 | 事件 | 类型 | 金额 / 估值 / 状态 | 参与方 | 含义 |
|---|---|---|---|---|---|
| 2015 | Medal 创立 | 创立 | 游戏玩法平台上线 | Pim de Witte / Medal 关联 | 形成后来成为 General Intuition 护城河的长期数据资产。 |
| 2024 年末 | OpenAI 据报提出收购 Medal | 反向 | $500M 据报报价 | TNW 引述 The Information | 表明外部实验室在分拆前已经看到了数据价值。 |
| 2025-10 | General Intuition 推出 / 从 Medal 分拆 | 治理 | 据报启动轮 $133.7M | General Intuition 创始人、Khosla、General Catalyst | 把 AI 实验室从消费者剪辑平台中分离出来,并为首笔外部融资铺底。 |
| 2026-01 | Series A 据报已私下完成 | 融资 | 据报以 $2.3B 估值完成 $320M | DutchNews / FD 摘要 | 暗示融资在公开宣布前数月已基本完成。 |
| 2026-06-18 | 公开宣布前的融资谈判被报道 | 融资 | 洽谈中:约 $300M,估值 >$2B | TechCrunch | 提供从轮次形成到最终披露之间的公开中点。 |
| 2026-06-25 | Series A 公开宣布 | 融资 | $320M,估值 $2.3B;披露融资总额 $454M | TechCrunch、TechFundingNews、AI Insider | 让公司在异常早期就成为融资最高的世界模型初创公司之一。 |
| 2026-06 | Nerve 数据采集市场被披露 | 产品 | 游戏玩法 / 遥操作工作平台 | TechCrunch / TechFundingNews | 把数据飞轮扩展到被动 Medal 片段之外。 |
| 2026-03 至 2026-06 | 与 Kyutai 和 Epic Games 发布 MIRA 技术成果 | 产品 | 发布 5B 参数多人世界模型 | General Intuition、Kyutai、Epic Games 生态 | 展示公开技术产出,以及大规模世界模型研究可信度。 |
| 2026 年夏末(目标) | 计划扩大 API 可用性 | 规模化 | 筛选式发布目标 | TechCrunch / 主页 | 标记报告日期之后仍待完成的首个公开商业化里程碑。 |
这条时间线保留了从数据平台起源,到 2026 年融资和首批可见商业化步骤的公开记录。
[CO006, CO007, CO010, CO019, CO021, CO032]公开时间线从 Medal 积累游戏玩法数据开始,随后快速拆分,2026 年 1 月完成交割,并在 2026 年 6 月公开融资公告中绑定选择性 API 发布。
若干项目区分了私下交割日期和后续公开公告日期,因为公开来源同时报道了两者。
[CO006, CO010, CO019, CO021, CO032, CO037]1.5 展品
02市场分析
2.1 市场边界、纳入支出和现状替代品
General Intuition 容易被误分,因为它的叙事同时触及游戏、机器人、合成数据和世界模型研究。公司材料把它定义为构建跨空间和时间行动系统的实验室,并称首批合作伙伴覆盖游戏、 仿真和机器人。这意味着真实品类既不是单纯的“游戏里的 AI”,也不是单纯的“机器人软件”。更准确的边界,是一个横跨世界模型训练平台、合成数据与仿真基础设施,以及面向必须感知、 预测和行动的系统的智能体 API 的软件层。Kaiso 对 AI 世界模型的市场定义最接近这一直接品类,因为它明确纳入基础世界模型、仿真引擎、合成数据平台,以及用于机器人、自动驾驶、工业自动化和数字孪生的开发框架。 纳入支出因此包括世界模型 API、具身智能体训练平台、合成数据工具和可用于训练或评估智能体的仿真基础设施。排除支出同样重要。通用 LLM 订阅、纯机器人硬件、传统创意或游戏引擎工具, 不属于同一市场,除非它们直接提供智能体训练或世界模型层。MuJoCo、robosuite、Isaac Sim 和 Infinigen 是尤其关键的现状替代品,因为它们已经用开放或低成本工具链解决了部分任务。 现有工具意味着,General Intuition 必须证明游戏玩法训练出的世界模型栈能带来买方无法用现有仿真软件和内部工程低成本复刻的东西。[CM001, CM002, CM003, CM004, CM005, CM025]
| 细分 / 类别 | 纳入支出 | 排除支出 | 买方 / 付款方 | 与 General Intuition 的相关性 |
|---|---|---|---|---|
| AI 世界模型 | 基础世界模型、仿真引擎、合成数据平台、开发框架 | 商品化 LLM 订阅;不含模型层的通用云算力 | AI 实验室、机器人开发者、AV 团队、工业软件供应商 | 最接近的直接类别,因为它明确包含世界模型软件和平台层 |
| AI 驱动仿真与数字孪生 | 数字孪生平台、仿真软件、AI 运营优化、3D 建模工具 | 大多数硬件、IoT 设备和非 AI 企业系统 | 制造、物流、资产密集型工业运营方 | 可作为 GI 仿真价值主张的功能代理,但比 GI 当前软件范围更宽 |
| 机器人 AI / 物理 AI | 面向 AI 机器人的感知、规划、控制、机队和策略软件 | 机器人硬件、传感器、执行器、非 AI 机械自动化 | 机器人 OEM、仓储运营方、医疗和制造用户 | 具身智能需求的大邻近市场,但远宽于 GI 当前筛选式 API 姿态 |
| 游戏生成式 AI / 游戏 AI | NPC 智能、场景生成、关卡创建、创作者工具、游戏 AI 系统 | 传统游戏引擎、静态资产管线、非 AI 内容工具 | 游戏工作室、开发者、设计师、创作者 | 之所以有用,是因为 GI 的数据起点和早期产品故事都与游戏相连,但这不是完整目标市场 |
| 现状仿真工具链 | MuJoCo、robosuite、Isaac Sim、Infinigen 等开放或低成本工具,以及内部工作流 | 超出点状工具使用的完整专有智能体平台 | 已经运行仿真管线的研究人员和工程团队 | 这些替代品定义了 General Intuition 必须在保真度、速度或数据杠杆上打败的现有基线 |
公司的真实品类是这些细分的交集,而不是任何单一市场报告桶。
[CM001, CM002, CM003, CM004, CM005]2.2 证据约束下的 TAM、SAM 和 SOM
公开市场数据没有给“游戏玩法优先的智能体基础设施”提供一个干净的分析师分类,所以本章必须从多个视角拼出规模判断。最直接贴近的品类是 AI 世界模型,Kaiso 估计 2025 年为 $1.8 billion,到 2035 年增至 $52.7 billion,CAGR 为 40.2%。第二个视角是 AI 驱动的仿真和数字孪生:The Business Research Company 估计该市场 2026 年为 $6.89 billion,Fortune Business Insights 使用更宽定义,给出 2026 年数字孪生 $33.97 billion。第三个视角是机器人 AI,Grand View 估计 2025 年市场为 $20.4 billion,CAGR 32%,这意味着一个比 General Intuition 当前可服务范围大得多的邻近市场。最后,游戏专属市场明显更小:The Business Research Company 将 2026 年生成式 AI 游戏市场定为 $2.21 billion,广义游戏 AI 为 $3.4 billion。 关键尽调点在于,最大的数字不等于最贴合的数字。General Intuition 并不出售整个数字孪生栈、全部机器人硬件或完整的游戏 AI 软件宇宙。它当前的产品姿态是选择性 API 访问和合作伙伴主导的实验。 这把现实的当前 SAM 限制在世界模型软件、合成数据基础设施和具身智能体训练预算的交集——更可能是低个位数十亿美元,而不是数百亿美元。近期 SOM 更小,因为机器人和工业软件的企业采用仍然慢、 技术门槛高,并且高度依赖具体预算负责人。市场无疑大到值得重视,但估值承销应锚定狭窄交集市场,而不是最宽的母品类幻灯片。[CM006, CM007, CM008, CM009, CM010, CM011]
| 视角 | 2025 或 2026 年数值 | 预测 / CAGR | 实际衡量对象 | 置信度 | 局限 |
|---|---|---|---|---|---|
| AI 世界模型 | 2025:$1.8B | 2035:$52.7B;40.2% CAGR | 基础世界模型、仿真引擎、合成数据平台、开发框架 | 中 | 最佳直接类别,但仍新且由供应商定义 |
| AI 驱动仿真与数字孪生 | 2026:$6.89B | 2030:$21.33B;32.6% CAGR | AI 驱动仿真和数字孪生软件 / 服务 | 中 | 仍比 GI 宽,因为它横跨许多行业工具和服务 |
| 广义数字孪生市场 | 2026:$33.97B | 2034:$384.79B;35.4% CAGR | 跨行业所有数字孪生技术 | 中 | GI 承销用它会过宽;包含许多非智能体工作流 |
| 机器人 AI | 2025:$20.4B | 2033:$182.7B;32.0% CAGR | AI 赋能机器人硬件和软件市场 | 中 | 大型邻近市场,而非 GI 近期实际可服务的纯软件市场 |
| 机器人仿真 | 2026:$7.58B | 2032:$13.90B;10.56% CAGR | 面向机器人的物理仿真和验证 | 中 | 无法捕捉 GI 的游戏数据和动作模型角度 |
| 游戏生成式 AI | 2026:$2.21B | 2030:$5.09B;23.2% CAGR | 游戏内容、NPC、场景和创作者工具 | 中 | 与数据起点和早期游戏用例相关,但不是全部具身 AI 上行空间 |
| 游戏 AI(广义) | 2026:$3.4B | 2030:$6.73B;18.6% CAGR | 用于游戏的所有 AI 技术 | 中 | 包含许多与 GI 世界模型战略无关的游戏 AI 类别 |
| General Intuition SAM(估算) | 当前:~$1B–$3B | 一旦证明可迁移,可能扩大 | 世界模型软件、仿真工具和具身智能体训练预算的重叠部分 | 低 | 编辑估算;没有独立报告单独拆出这个精确类别 |
| General Intuition SOM(估算) | 近期:< $0.25B | 取决于 API 推出和伙伴转化 | 公司在广泛生产部署前可能合理服务的范围 | 低 | 没有公开客户或定价数据支撑更窄估算 |
广义市场数字可作为母市场天花板;GI 专属 SAM 和 SOM 是编辑层面受约束的估算,因为没有公开报告单独拆出公司的精确类别。
[CM006, CM007, CM008, CM009, CM010, CM011]General Intuition 的可寻址机会,从广义物理 AI 相邻市场到匹配其当前产品姿态的较小重叠市场,迅速收窄。
SAM 和 SOM 是编辑基于重叠市场视角得出的估算;没有公开分析师报告单独切出 General Intuition 的精确品类。
[CM013, CM014, CM015]不同 2026 年市场视角给出的天花板差异极大;更窄、偏软件的类目才最适合用来衡量 General Intuition。
如果来源只发布单一端点估算,低 / 高区间就是编辑部给出的不确定性范围,并非独立分析师预测。
[CM006, CM007, CM008, CM009, CM010, CM012]2.3 买方、用户、付款方和采用路径
买方图谱拆成四个实际集群。第一类是游戏开发商和 AI 原生游戏工具团队,任务是比传统流程更快生成场景、NPC 行为、关卡内容或沙盒环境。在这个细分里,用户通常是玩法、技术美术或 AI 系统团队,付款方是开发工具或平台预算。第二类是机器人和具身 AI 团队,包括创业公司、OEM 和研究实验室。他们的任务是用合成训练环境、策略学习基础设施或世界模型生成的边缘案例, 降低对稀缺真实世界数据的依赖。第三类是数字孪生和工业软件厂商,它们的价值主张是预测性维护、虚拟调试或设施优化,而不是类似游戏的智能体。第四类是自动驾驶和出行开发者,稀有场景生成和环境仿真比创意工具更重要。 这些细分共享技术底层,但预算、销售周期和证明要求不同。游戏买方可以快速试用,只要输出有用,就能容忍创意上的不完美;机器人和工业买方更看重保真度、安全性和工作流集成。 General Intuition 当前证据显示,公司在机器人和工业采用旅程中的进度,仍落后于叙事搭建的成熟度。公开材料称游戏、仿真和机器人领域已有合作伙伴,但公司没有公布客户名称或 ROI 基准。因此,采用路径大概率从研究和原型预算开始;只有公司能在每个垂直领域证明迁移、延迟、成本和运行可靠性后,才会进入更深的平台嵌入。[CM016, CM017, CM018, CM023, CM024, CM025]
| 细分 | 买方 | 用户 | 付款方 / 预算所有者 | 采用触发因素 | General Intuition 必须证明什么 |
|---|---|---|---|---|---|
| 游戏工作室和 AI 原生工具团队 | 游戏工作室管理层、AI 工具供应商、技术美术组织 | 游戏玩法 AI 工程师、技术美术、内容团队 | 开发工具、平台或内容预算 | 需要更快的场景生成、NPC 行为或交互式沙盒 | 带动作标签的游戏数据能让智能体或工具优于现有游戏 AI 工作流 |
| 机器人初创公司和 OEM | CTO、研究副总裁、机器人平台负责人 | 策略学习、仿真和自主系统工程师 | R&D、平台或风投支持的工程预算 | 需要更便宜、覆盖更广的训练环境和稀缺边缘案例数据 | GI 的预训练能迁移到机器人场景,且成本与保真度权衡优于现有合成数据管线 |
| 工业数字孪生和仿真厂商 | 产品负责人、工业软件团队、数字化转型部门 | 仿真、运营和分析工程师 | 转型、运营或软件平台预算 | 需要与真实运营绑定的仿真、预测和场景测试 | GI 能接入工业工作流,安全性、延迟和可靠性达到可接受水平 |
| 自动驾驶 / 移动出行开发者 | 自主系统平台负责人、AV/AMR 软件负责人 | 感知、规划和验证团队 | 自主系统项目和安全验证预算 | 需要合成边缘案例和可控环境测试 | GI 的模型不止能用于游戏,还能支撑真实验证体系 |
| 定向 API 开发者 | 基于 GI 模型构建应用或研究工具的开发者 | 产品和研究工程师 | 工程和实验预算 | 在完整自建内部模型前,需要可编程访问世界模型或动作模型 | 文档、延迟和封装足以支持第三方集成 |
不同细分市场的经济性差异很大;游戏买家试用更快,机器人和工业买家通常需要更深的保真度和集成证明。
[CM025, CM026, CM027, CM028, CM029, CM041]每个客群要解决的任务不同,预算来源不同,也要求 General Intuition 提供不同的证明包。
预算推进速度和证明要求是编辑部综合各客群来源描述后的定性判断,不是 GI 披露的管线数据。
[CM025, CM026, CM027, CM028, CM029, CM035]最难的一步不是让市场知道 GI,而是证明游戏训练出的直觉能在游戏之外创造可衡量的运营价值。
漏斗数值是编辑部给出的相对强度标记,不是实测转化指标;公司未披露管线数据。
[CM022, CM034, CM040, CM041]2.4 增长驱动、约束和可能放慢采用的因素
几股 2026 年力量明确支撑需求。DeepMind 明确说,丰富训练环境供应一直是具身智能体的瓶颈;NVIDIA 则把 Cosmos 定位为机器人学习、世界仿真和合成数据生成的基础设施。 AI 驱动仿真、数字孪生、机器人和游戏 AI 的市场报告都指向同一方向:买方想要更多自动化、更多模拟场景和更多软件定义工作流。360iResearch 的机器人仿真展望尤其有用, 因为它显示该品类正从离线工程工具,走向与数字孪生、AI 验证和虚拟调试绑定的一体化数字工程生态。在这个世界里,如果 Medal 的游戏玩法语料确实能降低构建稳健智能体行为的成本,差异化数据源就会有价值。 刹车也同样真实。关于具身 AI 的 3D 生成 arXiv 综述称,具备物理基础、可交互的内容仍难以生产,sim-to-real 鸿沟还没有解决。MIT Technology Review 对世界模型的怀疑性综述给出同样的大方向: 这个领域有前景,但在真实世界环境中仍不可靠。Deloitte 补上企业采用视角——遗留系统集成、安全、合规、基础设施准备度、员工技能和 ROI 不确定性都会拖慢部署。Fortune 的数字孪生市场报告还加入了安全和互操作性担忧。合在一起,这组约束意味着 General Intuition 所在市场可以快速增长,却不一定同样快速转化为收入确认。公司所在市场真实、战略重要且在扩张, 但仍足够早,买方教育和价值证明可能与原始模型能力同样重要。[CM019, CM020, CM021, CM022, CM030, CM031]
| 因素 | 类型 | 方向 | 时间 | 影响 | 尽调问题 |
|---|---|---|---|---|---|
| 具身智能体训练环境稀缺 | 驱动因素 | 正向 | 2026–2028 | DeepMind 和 NVIDIA 都验证了市场需要更富仿真的智能体训练环境 | GI 路线图有多少直接提升训练环境质量,而不只是做世界模型演示? |
| 数字孪生和仿真软件扩张 | 驱动因素 | 正向 | 2026–2030 | 工业软件预算越来越愿意为场景测试、优化和预测系统付费 | 哪些现有数字孪生或工业厂商最可能成为渠道伙伴? |
| 游戏 AI 商业化 | 驱动因素 | 正向 | 2026–2030 | 游戏是更近的细分市场,GI 的数据来源也最容易向买家解释 | 在更广泛的机器人采用成熟前,GI 能否拿下具名游戏工作室或工具客户? |
| 自动化和机器人 AI 增长 | 驱动因素 | 正向 | 2026–2033 | 具身 AI 融资面扩大,会提高市场测试新训练和策略基础设施的意愿 | GI 与物流、制造、自动驾驶等高支出机器人细分场景贴合多紧? |
| 仿真到现实鸿沟 | 约束 | 负向 | 2026–2029 | 视觉效果惊艳但不能改善真实结果的系统,会被市场惩罚 | 哪些基准或客户证据证明游戏玩法预训练能迁移到机器人或自动驾驶技术栈? |
| 开源和既有仿真厂商竞争 | 约束 | 负向 | 当前 | 在许多研究和早期工程用例中,免费工具会压低付费意愿 | GI 有哪一层专有能力强到足以替代 MuJoCo、Isaac Sim 或内部工具链? |
| 集成、安全和合规负担 | 约束 | 负向 | 当前 | 工作流触及安全关键或受监管运营时,企业采用会放慢 | 面向工业或移动出行客户,GI 提供哪些治理、安全和部署控制? |
| 算力强度和数据稀缺 | 约束 | 负向 | 当前 | 世界模型和具身 AI 平台更利好拥有大量 GPU 预算和专有数据的实验室 | Medal 数据集相对于开发它所需资本,能带来多少结构性杠杆? |
时间判断是方向性而非数字化估计;本表强调哪些力量最可能影响 GI 未来两年的采用和估值支撑。
[CM016, CM017, CM020, CM022, CM035, CM038]2.5 展品
03竞争格局
3.1 格局:直接同行、巨头、邻近玩家和替代品
General Intuition 并不在一个干净的单一产品桶里竞争。在直接同行层,World Labs 是最清晰的创业公司类比,因为它推广的世界模型产品 Marble,可以从文本、图像、视频和布局生成持久 3D 世界。在前沿实验室巨头层,Google DeepMind 的 Genie 2、Genie 3 和 SIMA 代表一条资本更充足的研究路径,目标是可控 3D 环境以及能在其中行动的智能体。NVIDIA 低一层竞争,但杠杆极大:Cosmos 和 Isaac Sim 把世界模型工具、合成数据基础设施和现有机器人分发界面组合在一起。OpenAI 明确把视频模型框定为世界模拟器,这一点重要, 因为它表明最大的基础模型实验室把类似仿真的生成系统视为战略要地,即便商业包装仍在流动。 替代品集合和具名创业公司集合同样重要。Luma 和 Rosebud 从更易上手的创意工具角度,切入邻近的创作者和交互式世界工作流;Unity ML-Agents、MuJoCo、robosuite、Habitat、 ManiSkill 和 Infinigen 则让技术成熟团队能用开放或低成本方式自行拼出大部分栈。对机器人或仿真买方来说,真正替代方案往往不是另一家炫目的创业公司,而是由现有仿真器、开放环境和内部工程拼出的工具链。 这意味着 General Intuition 不仅要赢在模型新颖性上,还要在成本、集成、可控性,以及游戏玩法训练出的行动先验能否比开放基线更好迁移到客户工作流上赢。[CP001, CP002, CP003, CP004, CP005, CP006]
| 竞争者 | 类别 | 规模 / 融资背景 | 目标细分市场 | 核心差异化 | 主要限制 |
|---|---|---|---|---|---|
| General Intuition | 直接同类 / 动作模型 | 私营公司;2026 年 6 月宣布 $320M Series A | 游戏、仿真、机器人、具身 AI 开发者 | 来自游戏玩法的动作数据和定向商业 API | 未披露公开定价、客户名称和基准优势 |
| World Labs | 直接同类 / 空间智能 | 私营初创公司;融资充足且已有公开产品 | 创意团队、仿真用户、世界构建工作流 | 持久可编辑 3D 世界、多模态输入、可导出输出 | 公开牵引力和定价仍有限;机器人深度不如 NVIDIA 明确 |
| Google DeepMind | 既有前沿实验室 | Alphabet 支持的研究项目 | 具身智能体研究、世界模型研究、未来平台用户 | Genie 2/3 加 SIMA 带来前沿世界和智能体研究广度 | 商业化封装不如自助式初创公司工作流直接 |
| NVIDIA Cosmos + Isaac Sim | 基础设施既有厂商 | 上市公司生态,采取开源姿态 | 机器人开发者、物理 AI 实验室、合成数据用户 | 开放平台叠加机器人工作流里的仿真分发 | 面向创意和游戏的封装不如创作者中心工具有吸引力 |
| OpenAI 世界模拟器布局 | 潜在进入者 / 相邻实验室 | 拥有大规模算力基础的前沿模型实验室 | 关注视频、仿真、智能体环境的开发者 | 明确的世界模拟器叙事,由大规模视频生成研究支撑 | 公开产品重心一直在变化,尚未定位为与 GI 等同的工作流软件 |
| Luma | 相邻竞争者 / 创意 AI | 私营创意 AI 平台 | 视频、图像、营销活动和互动创意团队 | 快速多模态创意智能体,以及物理世界使命叙事 | 对机器人或具身智能体仿真预算的优化不够明显 |
| Rosebud AI | 相邻替代品 / 游戏创作 | 私营游戏创作平台 | 创作者、独立游戏开发者、互动内容用户 | 低门槛 AI 游戏创作和互动世界生成 | 面向严肃机器人或仿真客户的基础设施叙事更浅 |
| 开源 / 内部自建 | 现状替代品 | 开放或低成本工具,加上内部工程投入 | 研究实验室、高阶开发者、成本敏感团队 | 在仿真和训练中获得控制力、可扩展性和低许可成本 | 需要自行拼装,且不提供 GI 式封装好的专有模型层 |
本组画像覆盖了截至 2026 年公开英文来源中可见的主要直接、既有、相邻和替代路径;它具有代表性,但并不穷尽。
[CP001, CP002, CP003, CP004, CP005, CP006]General Intuition 介于前沿实验室的野心和创业公司的聚焦之间;但生态触达落后于在位者,公开工作流包装也落后于产品化最强的同行。
坐标轴是编辑部综合公开材料对生态触达和差异化定位的序数判断,不是市场份额或基准分数。
[CP017, CP018, CP019, CP029, CP030, CP031]3.2 能力、包装,以及各类对手最强的位置
公开产品呈现显示,这是一个分裂市场,而不是赢家通吃赛道。World Labs 强调空间一致性、持久性、可编辑 3D 世界和可导出结果;Luma 强调视频、图像、音频和品牌感知内容上的快速创意执行; NVIDIA 强调物理 AI、合成数据和仿真基础设施;DeepMind 强调研究级世界生成和智能体,而不是广泛商业包装。General Intuition 自己的公开姿态更窄:选择性发布的商业 API、 覆盖游戏、仿真和机器人的早期合作伙伴,以及一个围绕游戏玩法数据落地行动模型的产品故事。这个故事有差异化,但公开包装程度不如 World Labs 的 Marble 工作流或 NVIDIA 的开发者生态。 整个领域价格透明度都弱。General Intuition 不公布价目表,大多数直接世界模型对手也不公布。这会削弱直接价格比较,把尽调重点转向包装姿态:开源生态降低进入成本,选择性 API 释放稀缺和实验信号,创作者工具即便模型新颖性不深,只要今天能让工作流更容易,也能赢得采用。竞争含义是,买方重视具体的“从游戏玩法到行动数据”论点时,General Intuition 可能最强;但如果买方主要想要成熟顺手的创意世界构建工具、完全开放的仿真栈,或巨头生态带来的采购安全感,它就较弱。[CP015, CP016, CP017, CP018, CP020, CP021]
| 购买标准 | General Intuition | World Labs | DeepMind | NVIDIA | Luma | 开源基线 |
|---|---|---|---|---|---|---|
| 可用动作控制的 3D 世界 | 是(公司声称) | 是 | 是 | 部分 | 部分 | 部分 |
| 公开可见的机器人 / 物理 AI 定位 | 是 | 部分 | 是 | 是 | 有限 | 是 |
| 创意世界构建用户体验 | 部分 | 强 | 有限 | 有限 | 强 | 有限 |
| 开源或低成本访问路径 | 未公开自助入口 | 未公开自助入口 | 否 | 强 | 有限 | 强 |
| 有文档记录的开发者生态 | 公开证据有限 | Marble Labs 触点在扩张 | 研究论文 | 强 | 中等 | 强 |
| 导出 / 工作流集成表述 | Unknown | 是 | Unknown | 是 | 是 | 是 |
| 已审阅页面上的具名定价 | 否 | 否 | 否 | 否 | 否 | 基本是 / 内部承担基础设施成本 |
是 / 部分 / 否 / 强等标签是基于已审阅公开界面的编辑判断,而非基准性能测试。“未知”表示审阅材料不足以作出公平判断。
[CP016, CP017, CP018, CP020, CP021, CP022]| 厂商 / 类别 | 公开价格可见性 | 访问模式 | 页面强调内容 | 影响 |
|---|---|---|---|---|
| General Intuition | 未公开标价 | 定向商业 API 和伙伴主导导入 | 世界模型,以及覆盖游戏、仿真和机器人的首批合作伙伴 | 稀缺性可支撑筛选式导入,但会让买家更难比较 |
| World Labs | 未公开标价 | 以 Marble 和 Marble Labs 产品化的工作流 | 创作、编辑、导出和空间一致性 | 即便透明定价尚未出现,更清晰的封装也可能推动采用 |
| DeepMind | Genie 或 SIMA 没有公开价目表 | 研究发表和实验室叙事 | 世界模型和智能体能力前沿 | 战略信号强,但当前直接采购路径弱 |
| NVIDIA | 情况不一;产品页强调平台访问,而非简单价目表 | 开放平台叠加仿真生态 | 物理 AI、合成数据、机器人学习、开放框架 | 开放生态会挤压机器人工作流中初创公司的定价权 |
| Luma | 已审阅页面未看到明确企业价目表 | 创意平台 / 应用访问 | 快速端到端创意执行 | 可赢下更看重速度和简单性、而非深层基础设施主张的用户 |
| 开源基线 | 许可成本通常很低或为零 | 仓库、文档或云基础设施自助拼装 | 控制力、实验空间和可扩展性 | 任何想为早期技术评估收费的初创公司,都需要拿出更强证明 |
已审阅公开页面很少披露企业级单位经济;本表比较的是封装姿态和价格透明度,而非已实现合同价值。
[CP015, CP020, CP021, CP022, CP023, CP024]不同买方要完成的任务不同,对同一批竞争者的排序也会很不一样;当游戏玩法衍生的行动智能比创作者 UX 或开源经济性更重要时,General Intuition 最强。
“强 / 中等 / 有限”标签是基于公开页面和代码库证据的编辑部判断,不是独立审计的客户结果。
[CP014, CP015, CP016, CP017, CP021, CP022]3.3 护城河耐久性、切换成本和多供应商并用风险
General Intuition 最可信的公开护城河主张,不是当前分发规模,而是输入数据的独特性。公司称,它正在用 Medal 的数十亿条游戏片段训练模型,并把这个动作密集语料转成可以跨虚拟和物理环境感知、 预测和行动的模型。如果这些数据真的能产生更好的行动先验或规划行为,就可能很重要。但公开证据还没有显示硬性锁定。公司没有披露具名生产客户、相对开放或巨头栈的基准差异,也没有披露足以让切换变痛的工作流依赖。 相比之下,NVIDIA 已经通过 Isaac Sim 和配套工具嵌入机器人开发者工作流,World Labs 也让自己的产品界面对创意和仿真用户更清楚。 这会形成高度多供应商并用的环境。开发者可以为一个用例评估 General Intuition,为另一个用例评估 World Labs,同时保留开源仿真器作为基线工作流。即便在同一组织里,世界生成、智能体训练和生产部署也可能来自不同供应商。 在这种背景下,选择性 API 访问可以保留稀缺性并支持谨慎导入,但它本身不创造切换成本。持久优势需要证据证明,General Intuition 的数据、工具或集成层能产出明显优于买方从开源栈或巨头平台获得的结果。[CP027, CP028, CP029, CP030, CP031, CP032]
| 护城河主张 | 支持证据 | 威胁 | 严重程度 | 缓释 / 尽调问题 |
|---|---|---|---|---|
| 来自游戏玩法的动作语料 | Medal 的游戏玩法历史,以及公司关于富动作训练数据的叙事 | 竞争者可能靠合成数据、游戏合作或大规模视频训练复现类似行为 | 高 | 要求提供基准证据,证明源自 Medal 的数据相较开源基线显著改善迁移或规划 |
| 筛选式伙伴导入 | 公司称首批合作伙伴覆盖游戏、仿真和机器人 | 筛选式访问本身不能创造转换成本;买家试点期间仍可多平台并用 | 中高 | 要求试点转生产的转化数据、留存和独家条款 |
| 从虚拟到物理的跨域论点 | 公司用一个模型家族同时面向游戏和真实世界环境 | 仿真到现实的跃迁可能失败,或弱到不足以支撑付费生产使用 | 高 | 要求客户案例,证明机器人或仿真结果有可衡量提升 |
| 早期品类时点 | General Intuition 在市场成熟前入场 | 价值一旦被证明,算力和分发更强的既有厂商可以吸收这个品类 | 高 | 跟踪路线图速度、招聘深度,以及 GI 能否在既有厂商标准化技术栈前建立狭窄滩头阵地 |
| 初创公司敏捷性 | 小公司能比业务宽泛的既有厂商更快聚焦狭窄论点 | 开源框架会降低市场为专有实验层付费的意愿 | 中 | 询问 GI 如何把模型能力转成部署、工具或数据飞轮,让开放工具难以快速复制 |
严重程度反映未来 12–24 个月的竞争耐久度风险,而非破产风险。
[CP027, CP028, CP029, CP030, CP031, CP032]公司资金充足,叙事有差异化,但公开可见的竞争就绪度仍显得更早期,不及同类最佳在位者和最可见的产品化同行。
[CP014, CP015, CP023, CP027, CP028, CP030]3.4 反向证据:商品化、巨头反击和品类波动
最强的反证是结构性的。MIT Technology Review 认为世界模型仍不可靠,这一点重要,因为品类热度可能跑在企业价值创造前面。MuJoCo、Habitat、ML-Agents 和 ManiSkill 等开放工具持续改进;NVIDIA 正在开源物理 AI 栈的大部分;DeepMind 和 OpenAI 这类前沿实验室能用更强算力和分发,把邻近研究推入产品。这意味着 General Intuition 同时被上方巨型实验室和下方商品化开发者工具挤压。 品类本身也很波动。OpenAI 的世界模拟器叙事验证了这个方向的战略重要性,但围绕视频生成世界的产品包装已经快速变化;创意 AI 买方也常常可以用 Luma 或 Rosebud 等更简单工具解决即时需求, 而不必押注更深的世界模型平台。因此,反向竞争案例很直接:General Intuition 可能在品类方向上判断正确,但如果市场标准化在开放基础设施上、如果巨头把最佳功能吸收到更大生态里, 或如果客户认为游戏玩法优先论点有意思但并非任务关键,它仍可能丧失经济权力。[CP035, CP036, CP037, CP038, CP039, CP040]
3.5 展品
04财务
4.1 收入模式、变现界面和当前披露水平
General Intuition 的公开商业化故事仍然很薄,但并非看不见。公司官网称,已经让游戏、仿真和机器人领域的首批合作伙伴接入商业 API。这一点重要,因为它指向谈判式企业或合作伙伴主导收入路径, 而不是广泛自助式开发者平台。TechCrunch 6 月 25 日报道强化了这一解读:公司仍需要把 API 放到更多客户手里来测试用例。换句话说,公司看起来已经启动商业活动,但仍处早期。 缺失的是定价层,投资人无法据此判断收入质量。没有找到公开价目表、使用量指标或最低承诺。因此有四个合理变现界面:选择性 API 合同、付费试点、定制集成或微调工作, 以及更广泛发布后可能出现的经常性平台授权。对前沿 AI 实验室来说,这些路径都可信,但毛利率和可预测性含义差异很大。财务读数是,General Intuition 有可见收入路径, 但还没有公开路径指向可预测的软件收入。[CI001, CI002, CI003, CI004, CI005, CI034]
| 收入流 | 机制 | 计费单位 | 当前价值 / 状态 | 质量 | 尽调问题 |
|---|---|---|---|---|---|
| 定向商业 API | 面向游戏、仿真和机器人伙伴,协商访问 General Intuition 模型 | 企业合同或按量使用安排 | 处于筛选式伙伴模式;无公开价目表 | 中 | 要求活跃合同数量、ACV 和用量定价逻辑 |
| 伙伴试点 / 概念验证工作 | 围绕特定工作流或数据集的定制技术试点 | 试点 SOW 或里程碑费用 | 筛选式导入暗示存在;合同金额未公开 | 低 | 要求试点转化率、周期和扩展条款 |
| 模型微调 / 定制集成 | 面向客户环境的适配、部署或集成支持 | 一次性服务费或经常性平台费 | 未公开披露;对早期企业销售动作而言合理 | 低 | 要求 SOW 占比与经常性软件收入对比 |
| 未来平台授权 | 更广泛发布后,模型或平台按经常性订阅收费 | 年费或按量软件合同 | 仅属前瞻;尚未公开推出 | 低 | 要求路线图、定价假设和续约机制 |
| Medal 相邻数据 / 生态杠杆 | 源自 Medal 资产的潜在交叉销售或数据商业杠杆 | Unknown | 未公开披露变现 | 低 | 澄清 Medal 贡献直接收入、仅贡献数据优势,还是两者兼有 |
公开证据支持变现触点和姿态,但不支持实际收入值。每条收入流都缺少披露的规模、定价实现或利润率数据。
[CI001, CI002, CI003, CI004, CI005, CI030]| 触点 | 公开价格 | 证据基础 | 缺失内容 | 影响 |
|---|---|---|---|---|
| 商业 API | 主页和 TechCrunch 确认 API 访问,但未确认定价 | 价目表、用量单位、最低承诺额、折扣 | 缺少合同细节,早期收入很难建模 | |
| 选择性合作伙伴计划 | 公开资料只确认合作伙伴和试点 | 试点费用、生产费用、转化机制 | ACV 可能很高,但收入不连续,且服务占比重 | |
| 定制集成 / 微调 | 从企业化姿态和用例宽度推断 | SOW 费率、人员配置模型、利润率影响 | 服务占比过高会稀释软件利润率 | |
| 未来更广泛发布 | 公司称更广泛发布还在后面,目前尚未上线 | 打包方式、自助定价、续约条款 | 产品化质量决定收入可预测性 | |
| 股权融资驱动增长 | $320M 已披露融资轮 | 融资公告和媒体报道 | 按职能和时间维度拆分现金投放计划 | 资金到位情况可见;变现质量不可见 |
本表把可见融资和不可见变现拆开。融资是公开信息;定价和实际收入不是。
[CI001, CI003, CI016, CI017, CI018, CI034]General Intuition 早期商业路径可能是从选择性技术访问走向协商式生产合同,而不是立刻靠可规模化的自助漏斗起量。
节点反映的是推断出的商业化阶段,因为公司尚未发布完整定价或收入架构。
[CI001, CI002, CI005, CI030, CI034]4.2 组织扩张和可能的成本结构
最好的公开成本线索来自招聘页。General Intuition 和 Medal 正在招聘财务、基础设施、安全、数据平台、游戏集成和高级技术岗位。许多岗位的基本薪资区间集中在 $180,000 到 $300,000,Member of Technical Staff 岗位在多个城市的区间为 $250,000 到 $450,000 加股权。这些数字没有透露人数,但强烈指向高端研究和基础设施薪酬结构,而不是轻量应用创业公司的成本底座。 算力是另一大成本中心。世界模型和具身 AI 系统训练与服务成本都高,公开基础设施定价也支持这个直觉。CoreWeave 列出的 HGX H100 按需价格为每小时 $49.24,AWS 的 P5 系列围绕 H100 或 H200 GPU 集群和超高带宽网络构建。这些基准不是公司具体账单,但说明为什么薪酬加云算力很可能是主导费用组合。好处是,General Intuition 看起来更偏 opex 而非 capex:没有公开证据显示公司需要制造库存或重固定资产。[CI006, CI007, CI008, CI009, CI010, CI011]
| 指标 | 值 / null | 置信度 | 为什么重要 | 尽调要求 |
|---|---|---|---|---|
| 付费客户数 | 低 | 用来区分叙事热度和商业牵引力 | 分别提供活跃付费客户、试点客户和已流失试点 | |
| 平均合同价值 | 低 | 决定销售动作能否撑住前沿模型成本结构 | 按合作伙伴试点、生产客户和开发者账户拆分 ACV | |
| 毛利率 | 低 | 算力强度让毛利率形态成为核心测算变量 | 按产品线提供毛利率,并列明算力和支持成本分摊 | |
| 年度薪酬强度 | $25M–$60M 情景 | 低 | 薪资区间显示,即便尚未扩招,人才成本底盘也很贵 | 提供实际人数、现金薪酬和股权激励消耗 |
| 年度算力强度 | $15M–$80M 情景 | 低 | 前沿模型训练和服务可能主导运营开支 | 提供云账单、预留容量和模型训练节奏 |
| CAC / 回本周期 | 低 | 即便技术强,漫长企业销售周期也会吃掉效率 | 按细分市场提供销售周期、CAC 和队列回本周期 |
表中数值是情景区间或 null,不是公司披露。公开证据太薄,无法支撑更细的单位经济模型。
[CI004, CI008, CI009, CI010, CI011, CI012]公司单位经济性大概率取决于企业合同价值能否比人力、算力和支持成本涨得更快。
这座桥是结构性分析,不是数字模型;公开来源未披露已实现的单位经济输入。
[CI005, CI015, CI030, CI034]4.3 融资事实、备案背景,以及官方记录真正证明了什么
最具体的公开财务事实关乎融资,而不是经营表现。TechCrunch 6 月 18 日报道称,General Intuition 正洽谈以约 $2 billion 估值融资 $300 million。到 6 月 25 日, DutchNews 和其他媒体报道称,公司披露了 $320 million 融资、估值 $2.3 billion,投资方包括 Jeff Bezos、Eric Schmidt、Khosla Ventures 和 General Catalyst。 DutchNews 还报道称,自 2025 年 10 月以来总融资为 $454 million。这足以确认主要投资方承诺和估值快速上升,但几乎不说明收入质量。 官方备案证据比媒体报道窄。直接获取的 SEC Form D 属于 AVSF - General Intuition 2026, LLC,这是一个 Delaware 集合投资基金载体,向 93 名投资者售出 $4.497 million 的发行额。它对财务分析有用,因为它显示了与本轮融资相邻的载体,但不是运营公司披露。它不揭示 General Intuition 的收入、现金、烧钱速度或合同质量。因此, 本章把该备案用作融资机制的佐证,而不是公司经营指标证明。[CI016, CI017, CI018, CI019, CI020, CI021]
4.4 资本充足性、烧钱情景和可能资金用途
公司披露融资但不披露现金消耗,因此资本充足性必须用情景框住。低情景年度烧钱约 $60 million,意味着单独 $320 million 融资可支撑约 64 个月;中情景年度烧钱 $120 million,约 32 个月;高情景年度烧钱 $180 million,约 21 个月。这些不是公司披露,也没有计入此前现金余额、交易成本和战略储备。但它们是有用边界, 因为招聘数据和公开 GPU 定价都显示,前沿模型公司可能很快放大费用。 即便没有内部账目,最可能的资金用途也相当清楚:训练和服务模型、招聘昂贵人才、强化数据 / 安全 / 合规基础设施,并把选择性技术合作转成生产收入。公开证据没有显示债务、库存融资或工厂建设。 因此,核心财务问题不是公司是否资产沉重;而是其经营现金消耗能否保持在 $320 million 购买的窗口内,直到商业转化足够可见,支撑下一轮融资或内部可持续。[CI024, CI025, CI026, CI027, CI028, CI029]
| 指标 | 公开值 / 状态 | 置信度 | 为什么重要 | 尽调要求 |
|---|---|---|---|---|
| 最新披露融资轮 | $320M Series A,估值 $2.3B | 中 | 界定当前资本底盘和投资人预期 | 核对交割日期、分期结构和扣费后到账资金 |
| 2025 年 10 月以来总融资 | 报道为 $454M | 中 | 勾勒近期融资启动以来公司可用资源总量 | 确认其中多少是一级发行经营资金,多少属于旁侧载体或二级转让活动 |
| 在手现金 | 低 | 融资额不等于扣除既往消耗和预留后的可用现金 | 提供最近月末银行现金和受限现金 | |
| 年度消耗情景 | $60M / $120M / $180M | 低 | 资金可支撑期的敏感性取决于算力节奏和招聘速度 | 提供实际月度消耗、预算和相对计划偏差 |
| 仅按 $320M 融资计算的资金可支撑期 | 64 / 32 / 21 个月 | 低 | 约束公司在出现有意义收入前是否还需要再融资 | 提供董事会资金可支撑情景和下一轮融资触发条件 |
| 债务 / 项目融资 | 未发现公开披露 | 中 | 隐性义务会实质改变风险 | 确认所有债务、授信额度和有担保供应商承诺 |
本表把已披露融资事实与情景化资金可支撑期测算合在一起。情景行是编辑估算,不是管理层指引。
[CI016, CI017, CI018, CI019, CI021, CI022]情景分析显示,跑道对烧钱速度高度敏感,因为实际成本结构的公开证据很薄。
图中每个数值都是编辑部情景,不是公司披露。
[CI010, CI011, CI012, CI013, CI027, CI028]资产强度看起来像软件公司,运营强度则像前沿 AI 公司。
矩阵标签是基于公开证据的方向性判断,不是管理层报告的成本分配。
[CI014, CI015, CI024, CI025, CI030, CI031]4.5 财务判断和尽调阻塞点
公开证据支持一个简单但重要的结论。General Intuition 资金充足、技术野心大、组织成本高。它还不能按传统软件指标公开承销。没有可见收入基础,没有 ACV,没有毛利率披露, 没有烧钱速度,没有签约待交付订单,也没有客户集中度数据。就连法律和实体图景也比用户提示更复杂,荷兰报道指向 Naarden 持有的 IP 架构和跨国运营足迹。 这并不意味着公司财务弱,而是财务不透明。正向案例是,$320 million 新资本和 $2.3 billion 估值给了公司时间去证明一个狭窄滩头阵地。风险案例是,如果选择性 API 路线不能足够快转成粘性的生产收入,算力强度、高端人才成本和企业采用摩擦可能跑赢商业化。下一步尽调的最低要求是私有证据:管理账、定价材料、合同分组和董事会续航期材料。[CI033, CI034, CI035, CI038, CI039, CI040]
| 缺失指标 | 影响 | 具体尽调路径 |
|---|---|---|
| 收入 / ARR / 年化收入 | 无法判断估值支撑和经营杠杆 | 要求按产品和客户类型提供月度收入桥 |
| 已签订单积压和试点转化 | 无法判断销售管线是真实需求还是探索性接触 | 要求按阶段拆分销售管线,并提供试点到生产的队列转化 |
| 毛利率和算力分摊 | 无法判断使用经济性随规模扩大改善还是恶化 | 要求提供收入成本政策和算力分摊模型 |
| 现金、消耗和董事会资金可支撑情景 | 无法判断融资依赖时点和下行情形下的稀释风险 | 要求提供月度现金瀑布和基准 / 悲观资金可支撑情景 |
| 客户集中度 | 无法评估客户风险或续约依赖 | 提供前 10 大客户 / 试点敞口和已签约续约时间 |
| 股权结构表和荷兰实体架构 | 无法完整评估法律、税务或控制权影响 | 提供股权结构表、实体架构图和 Naarden 实体的 KVK 摘录 |
本表每一项都是投资测算的具体阻断项,不是锦上添花的指标。
[CI004, CI021, CI023, CI024, CI026, CI033]4.6 展品
05产品与技术
5.1 当前产品形态:选择性 API 加研究资产背书
General Intuition 的公开产品呈现窄于它的野心。公司官网描述的行动模型决定下一步做什么,世界模型预测行动结果,并把两者连接到面向游戏、仿真和机器人合作伙伴的选择性商业 API。 这一点重要,因为公司不只是在推抽象研究;它已经呈现一个商业入口。但周边证据仍显示,这更像高触达、早期企业销售动作,而不是包装好的开发者产品。合作伙伴门户是一个表单,询问另一家公司正在构建什么、 应该共同构建什么,而不是以文档为先的工作流;没有公开密钥、SDK 下载、使用示例或定价。连网站条款都明确说网站仅供信息参考,且可能随时更改,提醒投资人不要把营销语言误认为承诺路线图。 从实际客户工作流看,今天的产品似乎是专有模型访问、协作式合作伙伴导入,以及扎根 Medal 和 Nerve 的增长数据引擎。[CE001, CE002, CE003, CE004, CE005, CE006]
| 模块或资产 | 主要用户 | 当前状态 | 差异化 | 尽调缺口 |
|---|---|---|---|---|
| 选择性商业 API | 游戏、仿真和机器人合作伙伴 | 已上线,但选择性开放 | 在广泛发布前,把公司研究直接接入合作伙伴工作流 | 需要文档、认证模型、支持的端点和定价 |
| 动作模型 | 应用开发者和智能体构建者 | 核心论点;已公开描述 | 不只是预测帧,而是基于观察优化下一步动作 | 需要相对基线策略和竞品栈的基准测试差值 |
| 世界模型 | 内部研究和合作伙伴演示 | 已通过 MIRA 和公司信息公开演示 | 从像素和动作中学习环境动态 | 需要证明能迁移到受限游戏领域之外 |
| 源自 Medal 的动作数据集 | 模型训练团队 | 已在使用的战略资产 | 把视频与玩家输入里的意图信号配对 | 需要权利、治理和长期数据共享稳定性 |
| Nerve 数据采集平台 | 研究运营和未来合作伙伴 | 近期上线 | 把数据护城河延伸到付费标注和遥操作 | 需要规模、质量控制和贡献数据的经济性 |
状态仅反映公开表面。内部模块可能比公开网站显示的更丰富。
[CE001, CE003, CE004, CE005, CE035, CE040]| 用户任务 | 当前工作流 | General Intuition 方案 | 可衡量收益信号 | 限制 |
|---|---|---|---|---|
| 在类玩家约束下训练游戏智能体 | 手写脚本机器人或窄策略 | 用人类游戏过程输入训练动作模型 | 可能产出与玩家处在同一感知限制下推理的智能体 | 没有公开生产基准测试或具名工作室证明 |
| 安全原型化具身智能体 | 采集稀缺真实世界机器人数据 | 先用世界模型和游戏过程预训练,再做小规模真实世界微调 | TechCrunch 报道了 8 分钟机器人微调演示 | 从演示到生产的迁移仍未证明 |
| 在仿真环境中评估策略 | 定制模拟器或任务专用沙盒 | MIRA 这类交互式世界模型 | 公开视频显示可控多智能体轨迹展开 | 演示领域狭窄且合成 |
| 采集新动作数据 | 临时标注或外包网络 | 用于标注和遥操作的 Nerve 市场 | 可能降低新增动作数据的边际成本 | 公开规模和质量指标缺失 |
| 企业对新具身形态的实验 | 研究合作或定制试点 | 选择性 API 和合作伙伴申请表 | 让 GI 能嵌入敏捷内部团队 | 没有自助路径或广泛文档 |
收益信号是方向性的,来自公开演示或管理层表述,不是经审计的客户 ROI。
[CE005, CE006, CE030, CE031, CE034, CE035]当前使用路径似乎是从合作伙伴资格筛选走向选择性技术集成,而不是自助注册漏斗。
该流程抽象自管理层关于客户筛选和公开入口机制的表述。
[CE005, CE006, CE030, CE031, CE036]5.2 架构、训练数据,以及技术谱系为什么重要
观察 General Intuition 技术栈的最清晰公开窗口是 MIRA,这是与 Kyutai 和 Epic Games 共建的 Rocket League 世界模型发布。MIRA 重要,因为它把公司论点变成可检查的东西: 5B 扩散 Transformer、600M 视频编解码器、同步多人动作流、公开代码,以及解释优势和限制的技术报告。模型以每秒 20 帧运行,从像素加动作而非特权状态出发,并且只在评估中记录物理信息。 这是强证据,说明团队能构建可控、交互式模拟器,而不只是发布模糊的前沿实验室文案。同样重要的是,代码库和邻近项目显示了团队来处。IRIS、Δ-IRIS 和 DIAMOND 展示了一条从 token 化世界模型,走向基于扩散、更高视觉保真、可交互系统的演进。合在一起,公司架构故事变得清楚:General Intuition 不是从零发明技术栈, 而是在产品化一个已经探索过离散 token、扩散和控制条件世界建模的研究家族。[CE011, CE012, CE013, CE014, CE015, CE016]
| 层或组件 | 作用 | 公开证据 | 依赖 | 风险 |
|---|---|---|---|---|
| Medal 游戏过程语料 | 预训练数据底座 | 官网和报道描述了数十亿段游戏片段 | Medal 平台连续性和数据权利 | 数据治理或平台切分可能削弱护城河 |
| 动作标签 / 玩家输入 | 教模型理解意图和控制 | 公司和 MIRA 材料强调动作与帧配对 | 准确同步和日志记录 | 更大语料没有公开质量统计 |
| 世界模型核心 | 根据过往观察和动作预测未来观察 | MIRA 论文 / 博客 / 仓库 | 大规模训练算力 | 受限领域的成功可能无法泛化 |
| 视频表征编解码器 | 把帧压缩进生成式潜在空间 | MIRA 博客和仓库 | 用 DINOv3 权重训练最佳编解码器变体 | 第三方受限权重增加摩擦 |
| 评估探针 | 衡量物理保真度和可控性 | MIRA 论文使用动作跟随和状态探针 | 可靠内部监测工具 | 没有横跨竞争对手的公开标准化基准测试 |
| 商业 API / 合作伙伴层 | 向外部团队开放模型 | 官网和合作伙伴门户 | CoreWeave 容量和 GI 支持团队 | 企业就绪度和 SLO 仍不透明 |
本表反映可推断或直接观察的公开架构;它不是完整的内部系统图。
[CE003, CE004, CE010, CE012, CE013, CE014]公开技术栈从行动密集的数据采集出发,进入潜在世界模型,再进入选择性合作伙伴交付。
这是基于公开材料拼出的概念图,不是内部系统图。
[CE003, CE004, CE012, CE018, CE035]当前技术栈依赖专有数据、外部算力、协作者和第三方模型组件。
依赖项仅限公开来源明确可见的部分。
[CE004, CE018, CE029, CE034]5.3 部署现实:依赖关系看得见,信任层薄,推出节奏分阶段
公开部署证据是混合的。正面看,公司在公开互联网上有足够技术深度,可以越过基本可信度门槛:开放的 MIRA 代码、公开论文、明确算力依赖,以及与 Kyutai 和 Epic Games 合作的证据。 路线图也有一些真实具体性。多方消息称,大部分新资本将用于算力,CoreWeave 是具名基础设施依赖;更广泛 API 访问目标是 2026 年夏末。负面看,这些说法背后的公开运营细节很少。 没有找到公开 API 文档、公开可用性数据、模型卡、安全认证或信任中心。隐私声明和条款聊胜于无:它们展示了 Delaware 公司、纽约办公室、EU/UK 代表、 传输用标准合同条款,以及网站数据和 Medal 治理的游戏玩法数据之间的明确切分。但这些是治理基础,不是商业 API 已经为企业部署加固的证明。结果是, 公司技术核心看起来比公开可靠性和合规界面更成熟。[CE009, CE010, CE017, CE018, CE031, CE034]
| 控制项或问题 | 公开状态 | 范围 | 为什么重要 | 缺口 |
|---|---|---|---|---|
| 网站隐私通知 | 已有 | 网站数据和治理框架 | 显示基本数据治理姿态和跨境传输机制 | 不能证明 API 或训练管线的合规成熟度 |
| 网站使用条款 | 已有 | 仅信息网站 | 明确营销网站限制和 IP 主张 | 不能替代产品合同或 SLA |
| Medal / General Intuition 数据切分 | 明确记录 | 游戏过程数据与网站数据 | 有助于追踪哪个实体管理训练数据 | 未找到公开 DPA 或详细处理图 |
| 安全认证或信任中心 | 未公开出现 | 商业 API / 企业控制 | 客户会关注访问控制、可审计性和可用性纪律 | 未找到 SOC 2、ISO 27001 或信任中心证据 |
| 模型卡或公开安全报告 | 未公开出现 | 模型行为和风险控制 | 可帮助评估预期用途、限制和安全边界 | 未找到商业 API 的公开模型卡 |
这里的缺失表示在已审阅公开材料中未找到,不等于证明不存在。
[CE008, CE009, CE010, CE039, CE040]| 日期或阶段 | 功能或里程碑 | 公开状态 | 含义 | 来源 |
|---|---|---|---|---|
| 2026-06 公开公司网站 | 选择性商业 API | 已面向首批合作伙伴上线 | 显示一定商业化,但仍有闸门 | 官网 / Coalition / GamesBeat |
| 2026-06 融资报道 | 2026 年夏末前开放更广泛 API 访问 | 计划中 | 暗示研究密集期后分阶段推出 | TechCrunch / InvestGame |
| 2026-07 MIRA 发布 | 开源演示、论文和仓库 | 已发布 | 最强的公开技术执行证明 | MIRA 博客 / 仓库 / 论文 |
| 2026 年 Nerve 上线 | 数据采集市场 | 已发布 | 把产品边界延伸到数据获取 | Coalition / TechCrunch |
| 未披露的未来发布 | 更广泛模型可用性 | 尚未公开 | 产品化时间点仍是关键尽调变量 | 合作伙伴门户和新闻报道 |
状态标签区分已发布公开产物和计划中的商业化里程碑。
[CE005, CE017, CE024, CE031, CE034, CE035]公开证据在研究能力上最强,在企业级加固细节上最弱。
成熟度等级总结的是公开证据,不是内部就绪度指标。
[CE005, CE017, CE031, CE036, CE039, CE040]5.4 相对品类的差异化,以及仍未证明的部分
General Intuition 最可防守的技术差异化,并不是只有它相信世界模型,而是它把世界模型研究与大规模动作标注游戏玩法数据结合在一起。这让它不同于 DeepMind 的提示驱动 Genie 2 环境、Physical Intelligence 的机器人优先 π0 栈、Wayve 的结构化驾驶仿真、World Labs 的空间智能产品化,以及 NVIDIA 的广义物理 AI 平台。公司优势在于 Medal 数据同时包含观察和意图:视频,加上导致下一步结果的准确玩家输入。如果大型行动模型成为智能体的关键控制层,这可能非常重要。即便如此,公开证据还没有证明持久优越性。MIRA 很亮眼但范围很窄, 其作者也承认回放失败、隐藏状态问题,以及 sim-to-real 迁移到底能扩展多远仍不确定。与此同时,资本更充足的对手已经露出更多公开产品细节、更广生态或更清晰的终端用户工作流。因此, 产品技术读数对人才和研究执行有利,但对广泛产品成熟度和可防守部署优势保持谨慎。[CE022, CE023, CE024, CE025, CE026, CE027]
5.5 展品
06客户
6.1 产品看起来服务谁,以及销售动作如何运转
General Intuition 的公开客户故事更多由细分标签定义,而不是具名客户标识。官网、GamesBeat、Coalition、InvestGame 和 TechCrunch 中反复出现同样三个目标群体:游戏、仿真和机器人。 这种一致性很重要,因为它说明公司不是仍在寻找市场叙事。但它也暴露了进入市场动作仍有多早。公司没有公开文档、定价和自助导入,而是提供一个合作伙伴门户,询问另一家公司在构建什么、 未来应该共同构建什么。这是典型的高触达、企业主导导入模式。公开证据还暗示,公司希望客户能贡献有用的具身数据,并与 General Intuition 研究人员紧密共创。换句话说, 今天的买方大概率不是随手试用的开发者,而是有技术能力、愿意共同开发用例的合作伙伴,换取对模型栈的早期访问。[CU001, CU002, CU003, CU004, CU005, CU006]
| 细分市场 | 买方 / 用户 / 付费方 | 当前公开用例 | 战略价值 | 主要缺口 |
|---|---|---|---|---|
| 游戏工作室 / 游戏团队 | 买方可能是工作室或平台团队;用户是开发者和机器人设计师;付费方未披露 | AI 角色、机器人行为、感知世界的游戏系统 | 鉴于 Medal 数据和游戏原生的创业故事,这是自然的首个市场 | 没有具名付费工作室或生产部署 |
| 仿真团队 | 买方可能是仿真或数字孪生团队;用户是研究人员和操作员 | 在合成或镜像环境中测试智能体 | 把游戏预训练接到更广企业用途 | 没有具名仿真客户或结果指标 |
| 机器人开发者 / 运营方 | 买方可能是机器人公司或运营方;用户是机器人工程师 | 四足机器人导航、危险环境评估、具身智能体 | 如果迁移跑通,这是最有野心的市场 | 没有具名机器人客户或生产证明 |
| 数据贡献者 / Nerve 工作者 | 不是典型软件客户,而是数据市场参与者 | 标注、游戏过程贡献、遥操作 | 可加深数据护城河,并降低采集成本 | 不能证明模型收入具有经常性 |
| 基础设施 / 生态合作伙伴 | CoreWeave、Kyutai、Epic 这类协作方 | 算力供给、研究协作、展示型演示 | 验证生态兴趣,也帮助补齐能力 | 不等同于终端客户变现 |
公开证据更多按细分领域展开,而不是围绕具体客户 logo;公司未披露买方 / 付款方时,字段为推断。
[CU001, CU012, CU017, CU018, CU022, CU023]| 指标 | 公开值 | 日期 | 置信度 | 含义 | 缺失的基准分母 |
|---|---|---|---|---|---|
| 公开具名客户数 | 未找到具名付费客户(0 个) | 2026-07 | 高 | 披露远落后于融资规模 | 仍可能存在未公开客户 |
| 公开客户描述 | 少数客户来自游戏、仿真和机器人领域 | 2026-06 | 中 | 已有一些外部真实使用 | 未给出确切数量 |
| 商业 API 状态 | 选择性上线 | 2026-06 | 中 | 公司已不只是实验室姿态 | 没有公开端点或使用指标 |
| 更广泛的 API 可用性 | 计划在 2026 年夏末前开放 | 2026-06 | 中 | 采用仍是分阶段推出模式 | 没有时间线细节或里程碑门槛 |
| 具名生态合作伙伴 | CoreWeave、Kyutai、Epic Games | 2026-07 | 中 | 外部协作比终端客户采用更容易验证 | 合作伙伴证据不等于付款方证据 |
本表区分已经公开的信息和单纯缺席的信息;这里的零表示没有找到具名公开证据,不表示不存在私下活动。
[CU003, CU004, CU006, CU017, CU018, CU028]可见旅程从定向合作伙伴触达,走向嵌入式实验;若成功,则在 2026 年晚些时候扩大推出。
阶段是从公开合作伙伴措辞和路线图表述推断出来的,不是披露的销售打法。
[CU002, CU003, CU006, CU026, CU035]公开证据显示漏斗顶部兴趣很强,但到具名生产客户阶段,公开证明有限。
最后一个节点标记的是公开证据缺口,不是已知的私下运营为零。
[CU004, CU006, CU028, CU037, CU038]6.2 今天什么算客户证明,以及为什么仍然偏弱
今天最强公开证据是公司存在真实外部接触,而不是存在具名生产部署。TechCrunch 称这家创业公司在游戏、仿真和机器人领域有少数客户。GamesBeat 和 Coalition 描述了同样细分里的首批合作伙伴。 InvestGame 称商业 API 已经上线。合在一起,这些是有意义信号,说明公司做的不只是内部研发。但在已审阅来源中,没有任何一个具名付费游戏工作室、仿真公司或机器人运营方。 出现的具名外部实体——CoreWeave、Kyutai 和 Epic Games——是生态或合作验证,而不是传统 SaaS 意义上的客户验证。这个区分很重要。合作伙伴和协作者可见度可以确认市场兴趣, 但不能告诉投资人部署是否进入生产、合同是否续约,或用户是否获得可衡量结果。结果是一组真实但低分辨率的证明。[CU003, CU004, CU005, CU006, CU017, CU018]
| 具名证据 | 细分领域 | 部署或角色 | 生产还是试点 | 结果信号 | 局限 |
|---|---|---|---|---|---|
| CoreWeave | 基础设施 / 合作伙伴 | 支撑模型扩展和更广泛 API 推出的算力合作伙伴 | 生产级基础设施关系 | 多个外部来源具名提及 | 不是付费模型客户 |
| Kyutai | 研究 / 协作方 | 共同参与 MIRA 发布,产出公开技术证据 | 已落地的研究协作 | 显示外部伙伴愿意一起发布代码和论文 | 未披露为客户 |
| Epic Games | 游戏生态协作方 | MIRA 模拟 Rocket League,并标注与 Epic 协作 | 演示协作 | 让 GI 连上一个可识别的游戏场景 | 未披露为付费 API 客户 |
| 未具名游戏伙伴 | 游戏 | 公司和媒体称其获得选择性商业 API 访问 | 可能是试点 / 早期部署 | 多个来源反复提到它们存在 | 没有客户 logo、合同或结果 |
| 未具名机器人伙伴 | 机器人 | 公司和媒体称其获得选择性商业 API 访问 | 可能是试点 / 早期部署 | 细分领域被反复提及,能提供支撑 | 没有客户 logo、合同或结果 |
覆盖并不完整。公开披露稀疏,所以多行是生态或未具名伙伴证据,而不是传统意义上的具名付费客户。
[CU003, CU004, CU005, CU017, CU018, CU019]| 缺口 | 当前公开状态 | 重要性 |
|---|---|---|
| 具名付费客户 | 未找到 | 没有名字,就无法测试客户背书质量 |
| 结果案例研究 | 未找到 | ROI 和部署深度仍未知 |
| 留存指标 | 未找到 | 无法承销持久性 |
| 定价 / 合同模型 | 未找到 | 无法区分软件收入和服务很重的试点 |
| 渠道或引擎合作伙伴 | 未找到 | 分发杠杆尚未验证 |
每一行都是判断客户质量时的尽调阻断点,不是小遗漏。
[CU019, CU020, CU021, CU029, CU030, CU038]外部证明质量在生态合作上最高,在客户具体成果或留存上最低。
矩阵把关系存在的证据,与变现或续约证据分开。
[CU017, CU018, CU019, CU020, CU021, CU031]6.3 持续性、扩张和集中度大多还是公开未知数
公开留存质量是本章最薄的一环。审阅过的来源没有披露 NRR、GRR、流失率、续约时间、合同期限或客户满意度,也没有找到 G2 或 Gartner Peer Insights 这类第三方评价入口。公司私下留存可能很好,但公开证据看不出来。因此,客户质量分析只能退回到结构判断。如果公司如今确实只有少数客户, 集中度风险大概率偏高,扩张很可能取决于早期项目能否变成可重复工作流。Medal 的社区和内容网络长期看可能把漏斗顶端喂起来,类似 Nerve 的数据采集也可能带来更多生态锁定, 但这些都是未来可能性,不是当前经常性软件收入的证明。眼下,客户基础应视为有希望但脆弱:筛选严格、技术要求高,公开披露太少,支撑不了耐久性判断。[CU012, CU013, CU014, CU015, CU028, CU029]
| 指标 | 数值或状态 | 细分领域 | 置信度 | 重要性 |
|---|---|---|---|---|
| 净留存率(NRR) | null | 全部细分领域 | 高 | 能看出早期部署在初始接入后是否扩张 |
| 总留存率(GRR)/ 流失 | null | 全部细分领域 | 高 | 用来判断技术试点是否持续 |
| 合同期限 | null | 全部细分领域 | 高 | 区分一次性实验和可持续收入 |
| 满意度 / 评价入口 | 未找到公开评价信号 | 全部细分领域 | 中 | 客户背书质量目前不可观察 |
| 续约 / 重复部署证据 | 未找到公开证据 | 全部细分领域 | 高 | 没有续约,客户质量仍是推测 |
空值表示审阅材料未公开披露,不代表业绩为零。
[CU016, CU029, CU030, CU038]| 驱动因素或风险 | 当前信号 | 影响 | 重要性 |
|---|---|---|---|
| 选择性接入合作伙伴 | 强 | 能加深产品贴合度,但拖慢规模化 | 产品尚不成熟时,可能需要高服务强度的接入 |
| 少数客户集中度 | 可能较高 | 任一试点停滞,都可能放大收入波动 | 少数客户会让早期经济性起伏大 |
| 数据共享要求 | 可能 | 可能改进模型,但抬高采购摩擦 | 客户可能需要提供有价值的真实世界数据 |
| Medal 生态杠杆 | 中 | 可能支撑未来漏斗扩张 | 社区供给是优势,但不是当前收入证据 |
| 跨细分领域扩张 | 看似可行但未验证 | 可能把一个模型家族带进多个垂直领域 | 投资逻辑取决于模型能否跨具身形态复用 |
信号来自公开姿态推断,应通过销售管线和队列数据确认。
[CU026, CU027, CU033, CU034, CU036, CU038]6.4 客户结论
因此,General Intuition 的客户判断呈现不对称。它比纯隐身故事更好,因为多家独立媒体相互印证:公司已经在游戏、仿真和机器人领域推进筛选式伙伴合作。 它又弱于常规企业软件故事,因为同样的报道没有点名客户、没有证明生产使用,也没有披露留存指标。公司做客户开发最强的资产不是客户名单,而是供给侧生态: Medal 的游戏社区、差异化训练语料,以及愿意共创前沿演示的合作者。这些可能足以敲开成熟技术买家的门。但还不足以断定 General Intuition 已经降低客户质量或产品市场匹配风险。 现实的承销姿态是:把客户开发视为一个带有鼓励性主动需求的实时实验,而不是已经验证的经常性收入引擎。下一步实质尽调很简单,但只能私下拿: 客户 logo 清单、部署阶段、合同结构和客户访谈。公开证据仍然很薄。[CU001, CU004, CU007, CU022, CU031, CU037]
6.5 证据
07风险
7.1 核心论点拥挤、烧钱,且仍处早期
第一重风险在于,General Intuition 想赢的赛道已经被规模大得多或资本更厚的玩家验证。DeepMind、OpenAI、NVIDIA、Physical Intelligence 和 World Labs 都在围绕世界模型、具身控制或物理 AI 基础设施发布相邻叙事。这验证了品类需求,也压缩了新进入者把技术新意转化为持久定价权的窗口。 General Intuition 的公开发布姿态仍显筛选式:少数伙伴、分阶段 API 推出,没有广泛的生产使用公开证明。也就是说,公司必须同时解决两个硬问题—— 证明模型能超出狭窄演示场景泛化,并赶在竞争对手或平台把类似功能做得更容易采购之前完成。由此带来的时间线风险不只是技术问题。一旦发布延误, 商业化和融资风险会迅速叠加,因为公司已经走上重算力扩展路线。[CR001, CR002, CR003, CR004, CR005, CR006]
| 失效模式 | 公开证据 | 可能性 | 严重性 | 缓释成熟度 | 剩余暴露 | 未解缺口 |
|---|---|---|---|---|---|---|
| 世界模型质量无法泛化到小范围演示之外 | MIRA 是强证明点,但缺少公开生产基准和广泛客户结果 | 高 | 高 | 中 | 高 | 需要真实客户任务上的基准结果,而不只是演示场景 |
| 具身智能体行为在物理场景中不安全或脆弱 | 面向机器人的叙事已经存在,但未描述公开安全控制和评估护栏 | 中 | 高 | 低 | 高 | 需要安全论证、红队结果和部署约束 |
| 安全姿态弱于企业买家预期 | 未找到公开信任中心、SOC 2 披露或 SLA 文档 | 中 | 中高 | 低 | 中高 | 需要安全计划、事件响应流程和保证材料 |
| 更广泛 API 推出滑期 | 公开材料仍描述选择性访问,更广泛发布尚未落地 | 中 | 中 | 中 | 中 | 需要发布标准、路线图负责人和可靠性阈值 |
| 训练数据质量或标注循环退化 | Medal 支持和消费者平台现实意味着上游数据输入噪声大且不一致 | 中 | 中 | 中 | 中 | 需要数据集质量保证(QA)指标、片段过滤政策和溯源工具 |
运营风险聚焦可能挡住广泛商业化的失效模式,即便研究方向本身仍有吸引力。
[CR001, CR005, CR012, CR016, CR017, CR018]残余风险最高的部分集中在差异化、数据权利、算力强度和早期客户集中度。
评级概括的是公开证据,而非公司私下可能已有的内部控制。
[CR003, CR015, CR023, CR026, CR031, CR033]产品延迟、证据薄弱会很快传导到更高消耗、融资需求和估值压力。
该图抽象出公开证据可见的主要因果链,不覆盖每一种缓释手段。
[CR005, CR015, CR032, CR033, CR037, CR038]7.2 数据权利与监管风险能否可控,取决于尚未公开的私下控制
第二组风险是法律和监管。General Intuition 自己的隐私声明罕见地明确:用于研究和模型开发的 Medal 数据由 Medal 政策和单独的公司间安排约束。 这有帮助,但也意味着,对训练语料最关键的权利链并不公开透明。Medal 的条款强调遵守法律和第三方权利,却不能单独回答每一种下游训练或商业化用途是否都有合同保护。 与此同时,政策压力正从多个方向推进:先进 AI 芯片的出口管制规则持续变化;欧盟 AI Act 即使在简化规则,落地仍然很重;美国围绕 AI 训练和敏感数据的版权、 生物识别制度也在继续演化。没有哪一项能证明短期会断裂。合在一起,它们构成的合规栈远比当前公开信任材料显示的更重。因此,监管准备度是销售前置条件, 不是后台清理任务。[CR007, CR008, CR009, CR010, CR011, CR022]
| 风险 | 公开证据 | 司法辖区 | 可能性 | 严重性 | 当前缓释措施 | 剩余暴露 | 尽调路径 |
|---|---|---|---|---|---|---|---|
| 训练数据版权和知识产权(IP)链条 | 版权局仍在分析 AI 训练使用受版权保护材料的问题;Medal 条款要求合法使用,并遵守第三方权利要求 | 美国 / 全球 | 中 | 高 | General Intuition 和 Medal 发布了法律页面;公司间安排已被承认 | 高 | 要求提供游戏片段、许可、赔偿和退出流程的完整权利链 |
| 游戏玩法隐私和跨境转移 | General Intuition 隐私通知将 Medal 数据处理指向单独政策,并对欧盟 / 英国转移使用 SCC 式保障 | 美国 / 欧盟 / 英国 | 中 | 高 | 已发布隐私通知,列明欧盟 / 英国代表,并有转移条款 | 中高 | 要求提供数据流图、DPA 文件包、留存政策和删除流程 |
| 未来数据集中的生物识别或语音暴露 | FTC 政策和 Illinois BIPA 显示,如果产品范围扩大,语音声纹和面部几何会触发更高敏感度 | 美国州级 / 联邦 | 目前低到中 | 高 | 未披露面向生物识别的公开控制 | 中 | 核实是否收集、脱敏或排除语音、面部或遥操作视频 |
| 欧盟 AI Act 部署义务 | 欧盟委员会称部署按风险分层,实施仍在演进 | 欧盟 | 中 | 中高 | 未找到公开 AI Act 映射 | 中 | 将每个目标用例映射到可能的 AI Act 类别和负责人 |
| 先进芯片出口管制和地域限制 | BIS 继续收紧先进计算项目的尽调和地域挂钩出口规则 | 美国 / 全球 | 中 | 高 | 以美国为中心的算力伙伴和充足融资有帮助,但政策仍不稳定 | 中高 | 审查算力合同、芯片获取假设和非美国客户限制 |
各行按严重性排序,反映截至 2026-07-09 公开可见的主要法律与监管暴露。
[CR007, CR009, CR011, CR022, CR023, CR024]7.3 对 Medal、CoreWeave 和创始团队的依赖会很快传导为融资风险
第三组风险是依赖集中。公开证据把 General Intuition 与三根尤其重要的外部支柱绑在一起:Medal 是可见的数据飞轮,CoreWeave 是具名算力伙伴, 创始团队看起来仍承担大部分商业和技术叙事。这些依赖没有哪一项天然糟糕。事实上,每一项都是多头论点的一部分。问题在于,它们压缩了缓冲。 Medal 访问权限变化、算力成本飙升、基础设施可用性延迟,或创始人层面扰动,都可能在公司拥有广泛公开客户基础来吸收冲击之前打到产品进展。 由于更广泛发布仍未到来、客户披露稀疏,任何执行滑坡更可能先表现为烧钱加重和证明力变弱,而不是短暂小麻烦。投资人需要把组织厚度、算力经济性和数据控制耐久性列为关键尽调项, 不应当成次要问题。[CR012, CR013, CR014, CR016, CR031, CR032]
| 依赖 | 对手方 | 角色 | 集中度 | 失效情景 | 严重性 | 缓释措施 | 剩余暴露 |
|---|---|---|---|---|---|---|---|
| 游戏玩法数据飞轮 | Medal | 片段、社区行为和未来标注循环的来源 | 高 | 访问收窄、平台策略变化,或数据共享经济性恶化 | 高 | 共同起源和明确政策关联 | 高 |
| 训练和推理算力 | CoreWeave | 具名的规模化基础设施伙伴 | 中高 | 产能、价格或路线图变化拖慢推出和利润率改善 | 高 | 大额现金余额和公开伙伴协同 | 中高 |
| 选择性设计伙伴 | 未具名游戏 / 仿真 / 机器人伙伴 | 早期证明、反馈和背书 | 高 | 试点在更广泛发布前停滞或失败,削弱市场证明 | 高 | 选择性接入能改善贴合度 | 高 |
| 外部技术协作方 | Kyutai / Epic / MIRA 生态 | 研究证明和演示迭代速度 | 中 | 协作放慢或终止,降低公开交付节奏 | 中 | 已有开放成果 | 中 |
| 合规工具和流程栈 | 未公开披露的供应商 / 内部控制 | 规模化企业销售和国际部署所必需 | Unknown | 客户索要保证时,治理栈显得不成熟 | 中高 | 未看到公开缓释措施 | 高 |
最深的依赖不是单一供应商,而是数据、算力和早期伙伴证明必须同时成熟。
[CR001, CR013, CR014, CR032, CR033, CR034]| 角色 / 职能 | 依赖或缺口 | 可能性 | 严重性 | 缓释措施 | 尽调路径 |
|---|---|---|---|---|---|
| 创始人 / CEO 领导力 | 对外叙事和融资仍高度依赖创始人 | 中 | 高 | 从 Medal 到 General Intuition 的创始人-市场匹配可信 | 索取决策权地图、接班框架和领导梯队材料 |
| 研究领导层 | 世界模型和动作模型人才稀缺,难以替代 | 高 | 高 | 大额 Series A 轮支撑招聘 | 索取组织架构图、留任计划,以及按系统拆分的贡献者集中度 |
| 基础设施 / 平台运营 | 快速扩张可能跑在可靠性、成本控制和可观测性前面 | 中 | 高 | 已具名算力伙伴和招聘计划 | 索取 SRE 归属、事故指标和容量规划 |
| GTM / 解决方案工程 | 选择性共创打法未必能规模化为可复制销售 | 高 | 中-高 | 伙伴优先的导入能沉淀深度案例 | 索取销售管线结构、扩张计划和背指标团队建设方案 |
执行风险仍高,因为公司还在把研究可信度转化为面向客户的可复制运营系统。
[CR002, CR003, CR006, CR036, CR037, CR041]General Intuition 目前公开可见的技术栈要靠 Medal 数据、外部算力、选择性合作伙伴和监管放行,才能扩大规模。
这里只展示公开材料中明确可见的依赖项。
[CR007, CR014, CR023, CR026, CR034, CR035]7.4 风险结论
General Intuition 看起来不是鲁莽,而是尚未完成。公司有钱、有研究可信度,也有可见的伙伴势头。公开层面缺的是足够证据,证明训练数据权利、 安全控制、安全态势、客户耐久性和演示后的产品经济性都能支撑规模化。在普通软件初创公司里,这个缺口没那么致命;在这里更重要,因为每一层缺失的控制都会和其他层相互作用。 一个里程碑错过,就可能先变成产品问题,再变成烧钱问题,最后变成估值问题。因此,投资案取决于私下尽调能否足够快地补上技术承诺与运营证明之间的缺口, 跑赢这个品类的竞争和监管漂移。实务上,这是一家公司必须重尽调的故事,不是只靠公开材料就能形成确信的故事。[CR030, CR031, CR037, CR038, CR040, CR041]
| 风险 | 可监控触发项 | 阈值 / 事件 | 行动含义 |
|---|---|---|---|
| 产品差异化 | 没有具名生产客户,也没有外部可信基准胜出 | 下一个主要发布窗口后仍缺席 | 从主动尽调转入观察名单 |
| 训练数据权利 | 公司无法证明训练和商业化使用有干净的权利链 | 缺少许可、来源控制或退出路径 | 暂停尽调,或要求强赔偿保障并折价 |
| 算力经济性 | 无法给出可接受单位经济性的可信路径 | 训练 / 推理成本仍不透明,或结构上不经济 | 大幅下调估值 |
| 客户集中 | 太多验证依赖过少试点 | 一个伙伴主导案例价值或试点收入 | 将收入质量视为脆弱且不可复制 |
| 领导层集中 | 创始人或核心研究负责人在平台成熟前离任 | 意外离职或角色扰动 | 从头重新研判投资论点 |
| 监管阻断 | 出口管制或 AI Act 变化实质限制目标部署 | 关键地域或工作流在运营上受阻 | 实质下调 TAM 和时间线假设 |
这些触发项设计成可在尽调和未来 12 个月内监控,而不是抽象的长期警告。
[CR023, CR024, CR030, CR037, CR038, CR041]7.5 证据
08估值
8.1 建议与估值框架
General Intuition 更像一个价格敏感的尽调故事,而不是仅凭公开证据就能干净给出买入的标的。公司所处品类显然受到 2026 年资本市场奖励: Stanford AI Index 显示私人 AI 融资规模巨大,World Labs、Physical Intelligence 等可比前沿实验室也拿到了数十亿美元级估值。但 General Intuition 自身公开记录在估值纪律最关键的地方仍很薄。报道仍只描述少数客户、筛选式 API 访问,以及围绕算力的重投入模式。没有公开收入基数、毛利率画像、 股权结构表或客户留存数据,足以支撑精确倍数测算。因此,正确框架应从简单叙事热情转向情景承销。在这个基础上,当前 $2.3 billion 投后估值可以成立, 但支撑并不宽裕。最能站得住的判断是继续研究 / 跟踪:把公司留在漏斗里,但不要按商业证明和资本效率已经解决的价格来付钱。[CV001, CV002, CV003, CV004, CV005, CV006]
| 维度 | 评估 | 信心 | 风险评级 | 估值立场 | 决策含义 |
|---|---|---|---|---|---|
| 总体建议 | 继续研究 / 跟踪 | 中 | 高 | 昂贵 | 继续尽调,但不要按商业化证据已经坐实来付价 |
| 当前价格支撑 | $320M Series A 轮、投后估值 $2.3B,在 2026 年 AI 语境下说得通,但公开运营证据支撑不足 | 中 | 高 | 偏高至昂贵 | 把这轮视为高溢价前沿下注,而不是已去风险的软件入口 |
| 上行支撑 | 大品类顺风、强研究叙事、Medal 关联数据故事和顶级投资人背书 | 中 | 中 | 如果证据落地,可能支撑溢价 | 与管理层保持接触,推进私下尽调 |
| 阻碍买入的因素 | 未披露公开收入、利润率、股权结构、定价、留存或具名客户 | 高 | 高 | 证据不足 | 没有私下证据前,不要把兴趣升级为确信 |
| 改变判断的证据 | 具名生产客户、更广 API 发布、算力经济性、干净权利链和优先权结构清晰度 | 中 | 中 | 可能走向合理 | 只有证据补上当前尽调缺口,才上调判断 |
建议有意保持价格敏感。核心问题不是 General Intuition 是否令人兴奋,而是以今天可见的公开记录看,当前价格是否已经计入过多成功。
[CV001, CV003, CV006, CV036, CV037, CV038]从品类顺风、技术潜力,到价格纪律和投资建议的决策流。
该流程把细致的尽调简化为本章最关键的两个闸门问题:证据质量和价格支撑。
[CV011, CV012, CV035, CV036, CV039, CV050]8.2 融资背景、私下参照与公开市场区间
关于价格支撑,最强论点来自背景,而不是公司本身:前沿 AI 资本依然充裕。Stanford 记录了创纪录的融资背景,World Labs 和 Physical Intelligence 也表明,在广泛商业化之前,私人投资人仍愿意给世界模型和具身 AI 故事打出数十亿美元估值。这解释了 General Intuition 为什么能拿到 $2.3 billion, 却不能证明这轮便宜。World Labs 通过 Marble 和 Autodesk 看起来更接近可见产品化;Physical Intelligence 已经融到更多资金,并瞄准更大的数字。 公开可比公司同样更多是警示,不是安慰。按公开市值看,C3.ai 低于 General Intuition 的估值标记,Unity 远高于它——但两家公司都有 10-K 级披露, 能让投资人判断现实,而不是主要靠叙事承销。因此,可比集合拉宽了合理区间,却没有取消入场纪律。[CV007, CV008, CV009, CV010, CV011, CV012]
| 维度 | 正方论点 | 反方论点 | 什么会改变判断 |
|---|---|---|---|
| 品类顺风 | 2026 年 AI 资本市场仍积极奖励前沿实验室 | 火热融资市场不保证这个具体入场价有吸引力 | GI 正在把品类顺风转化为可持续客户证据的证明 |
| 技术叙事 | 世界模型和动作模型叙事具战略重要性 | 大型在位者和资金更足的同行也在追逐相邻技术栈 | 基准胜出或生产部署,证明差异化可持续 |
| 数据护城河 | Medal 关联若能复利成训练优势,价值可能异常高 | 权利链、可转移性和长期排他性公开上仍不透明 | 干净的数据权利尽调,以及数据优势改善结果的证明 |
| 商业化 | 选择性伙伴显示真实需求和审慎发布 | 选择性访问也意味着收入证明仍薄 | 具名客户、公开文档和可复制部署证据 |
| 估值 | 私募市场语境让数十亿美元估值有可解释性 | 基准情形下的公开证据并不说明 $2.3B 便宜 | 更低入场价或实质更好的披露 |
正方论点成立;反方论点主要围绕价格、披露和证据落地时点。这些问题恰好把有意思的公司和可投资的轮次区分开。
[CV006, CV023, CV027, CV028, CV035, CV039]| 可比对象 / 信号 | 指标 | 倍数 / 估值 / 状态 | 参考意义 | 局限 |
|---|---|---|---|---|
| General Intuition Series A 轮 | 当前私募轮次 | 融资 $320M,投后估值 $2.3B | 公司自身最硬的当前锚点 | 除规模和头部估值外,轮次条款未公开 |
| World Labs(2026 年 2 月) | 私有世界模型融资 | $1B 轮融资;估值未公开披露,但报道提到约 $5B 的讨论 | 最接近的私有世界模型参照,且有可见产品化 | 产品成熟度和生态可见度不同 |
| Physical Intelligence(2026 年 3 月) | 私有具身 AI 融资报道 | 据称正洽谈以 >$11B 估值融资约 $1B;此前估值 $5.6B | 显示投资人对具身 AI 平台上限的胃口 | 资本基础大得多,商业化仍偏轻 |
| C3.ai(公开) | 公开 AI 软件市值 | 截至 2026 年 7 月市值 $1.39B;最新 10-K 于 2026 年 2 月 27 日提交 | 具有完整披露纪律的公开市场下沿参考 | 产品组合不同,且受公开市场折价影响 |
| Unity(公开) | 公开游戏技术平台市值 | 截至 2026 年 7 月市值 $13.40B;最新 10-K 于 2026 年 2 月 11 日提交 | 可作为规模化游戏相邻平台的公开市场上沿参考 | 已安装基础、分发能力和披露深度大得多 |
| 公开机器人 / AI 篮子 | 板块估值基准 | 2025 年 Q4 收入倍数中位数 3.4x;高端 24.0x | 显示公开市场中位数与溢价估值的差异 | General Intuition 没有公开收入分母,无法直接换算 |
这张表有意保持局部视角,因为 General Intuition 没有完美的纯公开可比公司;最接近的私有同行也只披露了有限商业化信息。
[CV001, CV007, CV009, CV013, CV015, CV016]面向投委会的评分卡,覆盖市场、证明、护城河、经济性可见度、风险和估值支撑。
分数是为投委会框架作出的分析判断,不是标准化外部评级。分数越高越好。
[CV006, CV011, CV023, CV027, CV028, CV037]8.3 乐观、基准、悲观情景估值区间
由于没有可靠公开收入分母,情景测算应基于里程碑。悲观情景 $0.6 billion 至 $1.2 billion,假设更广泛 API 推出延误、客户证明仍薄, 下一轮融资受更严的公开可比纪律约束。基准情景 $1.4 billion 至 $2.2 billion,假设技术论点仍可信、伙伴活动继续,但仍要折价处理经济性缺失、 治理不透明和监管风险。乐观情景 $3.0 billion 至 $5.0 billion,假设更广泛发布、具名生产客户、来自 Medal 衍生数据的护城河证据更清晰, 并且相对 World Labs、DeepMind、OpenAI、NVIDIA 和 Physical Intelligence 有足够差异化,值得拿顶级前沿溢价。在这个框架下,当前 $2.3 billion 轮并不荒唐,但它高于基准情景中点,因此要求投资人在公开证据完全支撑之前,就承销一条明显偏乐观的路径。[CV020, CV021, CV022, CV023, CV024, CV025]
| 情景 | 核心假设 | 估值 / 回报逻辑 | 主要风险 | 概率信号 |
|---|---|---|---|---|
| 悲观 | 更广发布延后,客户证据仍稀疏,公开市场纪律下融资条款恶化 | 当前价值约 $0.6B-$1.2B;下行主要由证据不足和融资重置风险主导 | 降价轮、算力消耗、监管拖累、采用疲弱 | 当前价格下不可忽视的实质尾部风险 |
| 基准 | 技术进展延续,伙伴活动仍真实,但经济性和护城河证明仍不完整 | 当前价值约 $1.4B-$2.2B;相对一般 AI 软件有溢价,相对证据更强的故事有折价 | 缺少收入可见性、无具名客户群、竞争压缩 | 当前公开证据最能支撑的区间 |
| 乐观 | 更广发布成功,具名生产客户出现,GI 证明从游戏玩法到具身的护城河异常坚固 | 当前价值约 $3.0B-$5.0B;支撑相对当前轮的显著上行 | 执行滑坡可能快速压垮溢价 | 可能发生,但依赖多个里程碑一起落地 |
区间是分析性的当前价值估计,不是管理层指引或退出预测。区间有意拉宽,因为证据缺口也很宽。
[CV031, CV032, CV033, CV034, CV041, CV042]当前价值的敏感性取决于里程碑兑现,不取决于一个事实上未知的公开收入倍数。
数值单位为百万美元,是按里程碑调整的分析估计,不是管理层指引,也不是收入倍数模型。
[CV031, CV032, CV033, CV041, CV042]由当前公开证据支撑的悲观、基准和乐观当前价值区间。
数值单位为百万美元,代表当前价值区间而非未来退出价值,因为稀释和结构并未公开。
[CV041, CV042]8.4 最终尽调问题、论点破裂触发器与退出姿态
能改变这项建议的因素很直接。如果私下尽调能证明真实客户名单、合同耐久性、可信的算力成本曲线、干净的训练数据权利链,以及没有惩罚性优先权尾巴的 股权结构表,估值讨论会有利得多。如果这些答案拿不出来,公司更适合作为跟踪中的前沿资产,而不是立即高确信度入场。论点破裂来自具体事件, 而不是抽象不适:错过更广泛发布窗口、继续没有具名生产用户、重大出口管制或 AI Act 摩擦,或融资重置释放内部信心变弱的信号。退出准备度低, 因为公开记录仍远不像公开市场披露规范。眼下,正确姿态是保持接触,但把价格纪律和尽调纪律紧紧绑在一起。在此之前,承销应偏向保留选择权,而不是被迫建立确信。 价格纪律仍然必要。[CV027, CV028, CV043, CV044, CV045, CV046]
| 触发项 | 阈值 | 对论点的传导 | 行动含义 |
|---|---|---|---|
| 更广发布未到来 | 选择性访问持续到下一个计划扩张窗口 | 乐观情景削弱,基准情景压缩 | 从主动尽调转入观察名单 |
| 具名生产客户未出现 | 追加产品周期后,仍没有公开或尽调获得的案例集 | 商业证明太薄,支撑不了溢价 | 要求更锋利的价格折扣,或后退一步 |
| 算力经济性仍不透明或缺乏吸引力 | 管理层无法展示可信的成本曲线和利润率路径 | 当前价格失去支撑,因为消耗主导价值捕获 | 大幅下调估值 |
| 监管摩擦实质拖慢部署 | 出口管制、AI Act 映射或权利问题阻断关键工作流或地域 | 情景上行收窄,时间线拉长 | 重新测算 TAM 和时间线 |
| 下一轮重置价格或结构 | 平轮 / 降价轮 / 结构化轮次显示市场支撑走弱 | 当前估值不再能作为可信锚点 | 除非下行调整后的条款实质改善,否则暂停 |
触发项设计成可在未来 6-18 个月内监控,并能直接转化为投资动作,而不是模糊担忧。
[CV022, CV033, CV041, CV043, CV044, CV050]| 主题 | 缺失证据 | 重要性 | 负责人或尽调路径 |
|---|---|---|---|
| 客户证明 | 具名生产客户、合同规模、续约行为 | 决定当前价格反映真实采用,还是只反映叙事 | 管理层、客户访谈和销售管线审查 |
| 算力经济性 | 训练成本、推理成本、利润率路径、CoreWeave 承诺 | 资本强度决定上行是否归属股权 | 财务负责人、基础设施负责人和供应商合同 |
| 数据权利 | Medal 剪辑和元数据进入模型训练与商业化的完整权利链 | 可能是最高的隐性法律风险和护城河风险来源 | 法律尽调、DPA 审查和公司间协议 |
| 资本结构 | 股数、清算优先权、投资人保护条款和任何附函 | 决定头部估值能否转化为普通股价值 | CFO 或律师,以及完整股权结构审查 |
| 监管地图 | 出口管制假设、AI Act 分类、按用例拆分的隐私治理 | 承销时间线和国际部署风险时必须掌握 | 政策律师、产品负责人和合规工作流 |
这些要求有意保持机械化。没有这些材料,围绕建议和估值立场的争论很难超出有信息支撑的叙事判断。
[CV029, CV043, CV045, CV046, CV047]8.5 证据
免责声明
本报告仅供参考,基于截至 2026-07-09 的公开来源,不构成投资建议。作出任何投资决定前,应独立核验财务和经营结论。
证据索引
| 编号 | 陈述 | 可信度 | 来源 |
|---|---|---|---|
| CO001 | General Intuition publicly describes itself as a frontier lab for acting in space and time. | 中 | SO001 |
| CO002 | The official website says the company trains models on action-labeled video datasets across many environments. | 中 | SO001 |
| CO003 | General Intuition says it builds on Medal, where players upload billions of gameplay clips every year. | 中 | SO001 |
| CO004 | The company states that its current frontier work spans action models and world models. | 中 | SO001 |
| CO005 | The homepage says the company has onboarded first partners across games, simulation, and robotics to a selective commercial API. | 中 | SO001 |
| CO006 | Independent June 2026 coverage converges on a $320 million Series A at a $2.3 billion valuation. | 中 | SO003, SO006, SO007, SO008 |
| CO007 | The post-announcement total disclosed funding stands at roughly $454 million after the prior launch round. | 中 | SO003, SO007, SO008 |
| CO008 | Khosla Ventures is consistently identified as the lead investor in the Series A. | 中 | SO003, SO005, SO006 |
| CO009 | General Catalyst, Jeff Bezos, Eric Schmidt or Hillspire, and Nico Rosberg appear across the publicly named investor syndicate. | 中 | SO003, SO005, SO006, SO009 |
| CO010 | General Intuition was spun out of Medal after outside AI labs reportedly tried to acquire the gameplay-data asset. | 中 | SO004, SO020 |
| CO011 | Public reporting identifies Pim de Witte, Eloi Alonso, Adam Jelley, and Vincent Micheli as General Intuition co-founders. | 中 | SO003, SO005, SO020 |
| CO012 | Pim de Witte is the founder and CEO of General Intuition and the founder or former CEO of Medal. | 中 | SO003, SO014 |
| CO013 | Conference biographies and interviews describe de Witte as a gaming entrepreneur with prior humanitarian-sector work and earlier startup projects such as Highlight. | 中 | SO003, SO014, SO015 |
| CO014 | Eloi Alonso says he is a co-founder at General Intuition and that his prior work centered on reinforcement learning and world models during his Geneva PhD. | 中 | SO016 |
| CO015 | The MIRA paper and repository list Adam Jelley, Eloi Alonso, Vincent Micheli, and Pim de Witte as General Intuition contributors, supporting the team’s research credibility. | 中 | SO017, SO018 |
| CO016 | General Intuition’s public pitch is that world models are the training ground and agents are the eventual product. | 中 | SO003, SO004, SO020 |
| CO017 | The company’s claimed edge is that Medal clips embed action labels such as button presses and movement decisions, not only video frames. | 中 | SO003, SO004, SO008 |
| CO018 | TechCrunch described an internal demo where the same model family powered a game-playing agent and a quadruped after only minutes of real-world fine-tuning data. | 中 | SO003 |
| CO019 | Management says broader API availability is targeted for the end of summer 2026. | 中 | SO001, SO003, SO008 |
| CO020 | Multiple sources say most of the new capital is earmarked for compute scaling, including through CoreWeave. | 中 | SO003, SO005, SO010 |
| CO021 | DutchNews says the Series A was completed in January 2026 but not publicly announced until June 2026. | 中 | SO007 |
| CO022 | DutchNews says the company’s data and intellectual property are held through a Dutch company based in Naarden. | 中 | SO007 |
| CO023 | Tech Funding News says General Intuition operates as a public-benefit corporation legally registered in the Netherlands. | 中 | SO005 |
| CO024 | Public media coverage repeatedly describes General Intuition as centered on a New York lab or as New York-based. | 中 | SO003, SO004, SO005, SO006, SO008 |
| CO025 | Public reporting also points to offices in Geneva, London, and Paris in addition to the New York operating hub. | 中 | SO005, SO007 |
| CO026 | Backed VC shows General Intuition as a seed-stage portfolio company backed in 2025 and describes it as a gaming-AI frontier research lab spun out from Medal. | 中 | SO013 |
| CO027 | General Catalyst publicly lists General Intuition in its portfolio, corroborating its participation as an investor. | 中 | SO012 |
| CO028 | Public coverage consistently describes the Medal data supply as roughly 2 billion gameplay clips or videos per year. | 中 | SO004, SO005 |
| CO029 | TechCrunch’s June 18 article said Medal had more than 10 million monthly active users. | 中 | SO004 |
| CO030 | Tech Funding News and AI Insider instead cited Medal at about 17 million monthly active users. | 中 | SO005, SO009 |
| CO031 | Because credible public sources cite both roughly 10 million and roughly 17 million monthly active users, current Medal MAU should be treated as a conflicted datapoint rather than a hard fact. | 中 | SO004, SO005, SO009 |
| CO032 | TNW reported that OpenAI had previously offered $500 million to acquire Medal for its gameplay data. | 中 | SO020 |
| CO033 | Startup Fortune and Andrew.ooo say the company is already discussing a Series B shortly after the Series A, but that should be treated as directional rather than confirmed. | 低 | SO010, SO011 |
| CO034 | Public sources still do not disclose a board roster, revenue base, ARR, customer count, or named production customers. | 中 | SO003, SO005, SO007, SO013 |
| CO035 | The official website says the company wants to collaborate with creatives and the gaming industry rather than compete with them. | 中 | SO001 |
| CO036 | De Witte has publicly said the company will not pursue lethal military applications. | 中 | SO003, SO005 |
| CO037 | MIRA is a 5-billion-parameter multiplayer world model that runs in real time at 20 frames per second on a single Nvidia B200 GPU. | 中 | SO017, SO018, SO019 |
| CO038 | MIRA is a collaboration among General Intuition, Kyutai, and Epic Games, showing the company can ship public technical work with major partners. | 中 | SO017, SO018 |
| CO039 | MIT Technology Review argues that today’s AI remains unreliable in the physical world and that the world-model thesis is still an open path rather than a solved capability. | 中 | SO021 |
| CO040 | Public sources still do not independently prove that General Intuition’s gameplay-first pretraining approach transfers at scale into real-world robotics outcomes. | 中 | SO003, SO011, SO021 |
| CO041 | The public source set supports a selective-partner commercialization stage rather than a broad, self-serve software rollout. | 中 | SO001, SO003, SO013 |
| CM001 | General Intuition is best analyzed at the intersection of AI world models, simulation/digital twins, AI in robotics, and gaming AI rather than in only one of those categories. | 中 | SM001, SM002, SM004, SM006, SM008 |
| CM002 | The official company narrative is about systems that act across space and time, not generic chat or image generation. | 中 | SM001, SM022 |
| CM003 | Included spend should focus on world-model software, synthetic-data and simulation infrastructure, and embodied-agent APIs rather than all adjacent AI spending. | 中 | SM001, SM004, SM006, SM011 |
| CM004 | Excluded spend includes pure robot hardware, generic LLM subscriptions, and non-agentic tools such as physics engines or game engines used in isolation. | 中 | SM005, SM015, SM016, SM017, SM018 |
| CM005 | Status-quo substitutes include MuJoCo, Isaac Sim, robosuite, Infinigen, and bespoke internal simulation/data-collection workflows. | 中 | SM015, SM016, SM017, SM018 |
| CM006 | Kaiso’s AI world-models category was valued at $1.8 billion in 2025 with a 40.2% CAGR to 2035. | 中 | SM004 |
| CM007 | The Business Research Company sizes AI-powered simulation and digital twins at $6.89 billion in 2026. | 中 | SM006 |
| CM008 | Fortune Business Insights uses a far broader digital-twin definition and reaches a 2026 market size of $33.97 billion. | 中 | SM007 |
| CM009 | The Business Research Company sizes generative AI in gaming at $2.21 billion in 2026. | 中 | SM008 |
| CM010 | The Business Research Company sizes AI in games broadly at $3.4 billion in 2026. | 中 | SM009 |
| CM011 | Grand View estimates AI in robotics at $20.4 billion in 2025 and $182.7 billion by 2033, a much larger adjacency than General Intuition’s current software-only product scope. | 中 | SM005 |
| CM012 | 360iResearch estimates robotics simulation at $7.58 billion in 2026, providing a narrower proxy for General Intuition’s simulation angle. | 中 | SM012 |
| CM013 | The most relevant sizing lens for General Intuition is the overlap of world-model software, synthetic-data infrastructure, and embodied-AI tooling rather than the largest parent-market ceiling. | 中 | SM004, SM005, SM006, SM007, SM012 |
| CM014 | A reasonable current SAM for General Intuition’s product posture is low single-digit billions, not tens of billions, because only part of digital twins, AI-in-robotics, and gaming AI maps to agentic world-model software. | 低 | SM004, SM005, SM006, SM007, SM008, SM009, SM012 |
| CM015 | General Intuition’s near-term SOM is narrower than its SAM because the company is still in selective-partner mode and has not disclosed broad production deployment metrics. | 中 | SM001, SM002, SM003 |
| CM016 | Google DeepMind says the availability of sufficiently rich and diverse training environments has been a bottleneck for embodied-agent progress. | 中 | SM010 |
| CM017 | NVIDIA positions Cosmos as infrastructure for robot learning, synthetic data generation, and closed-loop world simulation across robotics, AVs, and industrial vision. | 中 | SM011 |
| CM018 | MIRA’s public technical materials frame playable game world models as a stepping stone to physical AI because real-world data is scarcer and riskier than game data. | 中 | SM021, SM025 |
| CM019 | The 3D-generation survey says embodied AI requires physically grounded, interaction-ready content rather than merely visually realistic output. | 中 | SM014 |
| CM020 | The same survey identifies limited physical annotations, fragmented evaluation, and the persistent sim-to-real divide as core blockers. | 中 | SM014 |
| CM021 | MIT Technology Review argues that current AI remains unreliable in the physical world despite rising enthusiasm for world models. | 中 | SM020 |
| CM022 | Deloitte reports that legacy integration, risk/compliance, infrastructure, cost, safety, and workforce readiness all materially slow agentic and physical-AI adoption. | 中 | SM013 |
| CM023 | Manufacturing and industrial automation are major demand centers in AI-powered simulation and digital twins. | 中 | SM006, SM007 |
| CM024 | AI-in-robotics demand already spans manufacturing, logistics, healthcare, e-commerce, and service-robot deployments. | 中 | SM005 |
| CM025 | General Intuition’s homepage says the company already has partners across games, simulation, and robotics, implying at least three initial buyer clusters. | 中 | SM001 |
| CM026 | Gaming-AI buyers focus on content generation, NPC behavior, scenarios, and creator tooling rather than real-world robotics transfer. | 中 | SM008, SM009 |
| CM027 | Robotics and embodied-AI buyers focus on synthetic data, simulation fidelity, and policy learning. | 中 | SM011, SM014, SM016, SM017 |
| CM028 | Digital-twin and industrial-software buyers focus on predictive maintenance, virtual commissioning, and operational optimization. | 中 | SM006, SM007, SM019 |
| CM029 | Budget owners differ by segment: game tools come from development budgets, robotics from R&D/platform budgets, and digital twins from transformation or operations budgets. | 中 | SM006, SM008, SM013, SM019 |
| CM030 | Asia-Pacific was the largest region in generative AI in gaming in 2025. | 中 | SM008 |
| CM031 | North America was the largest region in the broader AI-in-games market in 2025. | 中 | SM009 |
| CM032 | Asia-Pacific accounted for more than 45% of AI-in-robotics revenue in 2025. | 中 | SM005 |
| CM033 | AI-powered simulation and digital twins were largest in North America in 2025 while Asia-Pacific was the fastest-growing region. | 中 | SM006, SM007 |
| CM034 | General Intuition’s gameplay-first data moat is differentiated, but buyers outside gaming still need proof that the learned intuition transfers into operational outcomes. | 中 | SM002, SM003, SM020, SM021, SM022, SM024 |
| CM035 | Open and incumbent simulation stacks keep willingness to pay under pressure because many target users can prototype without buying a proprietary model API. | 中 | SM015, SM016, SM017, SM018 |
| CM036 | The absence of standardized world-model evaluation makes procurement slower and more bespoke because buyers cannot compare vendors on a shared benchmark. | 中 | SM004, SM014, SM020 |
| CM037 | Kaiso says cloud deployment dominates initial world-model procurement, while Grand View shows on-premise led AI robotics in 2025, so deployment preference varies by customer type. | 中 | SM004, SM005 |
| CM038 | High compute requirements and multimodal data scarcity create structural barriers that favor well-capitalized labs and strong cloud/GPU partners. | 中 | SM004, SM011, SM013, SM024 |
| CM039 | Core demand drivers include automation pressure, digital-twin adoption, synthetic-data needs, and the search for richer training environments for embodied agents. | 中 | SM005, SM006, SM010, SM011, SM012 |
| CM040 | Public evidence still lacks standardized ROI benchmarks for world-model APIs in games, simulation, or robotics. | 中 | SM001, SM013, SM014, SM020 |
| CM041 | General Intuition’s public use cases today center on gaming AI, simulation environments, and early quadruped/robotics experiments rather than a mass-market application suite. | 中 | SM002, SM003 |
| CM042 | TNW and Startup Fortune reinforce that major investors prize General Intuition’s gameplay dataset because it could reduce dependence on scarce real-world behavior data. | 中 | SM022, SM024 |
| CM043 | Fortune Business Insights highlights data security/privacy concerns and the absence of universal standards as restraints on the broader digital-twin market. | 中 | SM007 |
| CM044 | 360iResearch says robotics simulation is evolving from offline engineering tools into integrated digital-engineering ecosystems tied to digital twins, AI validation, and virtual commissioning. | 中 | SM012 |
| CP001 | General Intuition competes simultaneously with direct world-model startups, incumbent frontier labs, creator-tool adjacencies, and open-source simulation stacks. | 中 | SP001, SP004, SP006, SP009, SP011, SP014 |
| CP002 | World Labs markets itself as a spatial intelligence company building frontier models that can perceive, generate, reason, and interact with the 3D world. | 中 | SP004 |
| CP003 | World Labs says Marble generates spatially consistent, high-fidelity, persistent 3D worlds from multimodal inputs. | 中 | SP005 |
| CP004 | DeepMind says Genie 2 can generate action-controllable playable 3D environments for training and evaluating embodied agents. | 中 | SP006 |
| CP005 | DeepMind positions SIMA as a generalist AI agent for 3D virtual environments, making it more agent-layer competition than pure world-generation competition. | 中 | SP007 |
| CP006 | Genie 3 extends DeepMind’s public world-model push into 2026 and increases incumbent pressure on startups pursuing controllable environments. | 中 | SP008 |
| CP007 | NVIDIA markets Cosmos as an open physical-AI platform built around world foundation models, data processing, training, and evaluation frameworks. | 中 | SP009 |
| CP008 | Isaac Sim gives NVIDIA a strong distribution position in robotics simulation, testing, and synthetic data generation. | 中 | SP010 |
| CP009 | OpenAI publicly argues that scaling video generation models is a promising path toward building general-purpose simulators of the physical world. | 中 | SP013 |
| CP010 | Luma frames itself as a creative-AI platform while also claiming a mission to build intelligence that can operate in the physical world. | 中 | SP011 |
| CP011 | Rosebud AI positions itself as an AI game maker, making it an adjacent substitute for game-centric interactive-world use cases. | 中 | SP015 |
| CP012 | Unity ML-Agents allows games and simulations to serve as environments for training intelligent agents. | 中 | SP014 |
| CP013 | MuJoCo, robosuite, Infinigen, Habitat, and ManiSkill collectively show that sophisticated teams can assemble much of a simulation or embodied-AI workflow from open or low-cost tools. | 中 | SP016, SP017, SP018, SP019, SP021, SP022 |
| CP014 | General Intuition’s clearest public differentiation claim is its gameplay-derived action corpus from Medal and related action-model thesis. | 中 | SP001, SP003 |
| CP015 | General Intuition publicly describes a selectively released commercial API with early partners rather than a broad self-serve product. | 中 | SP001, SP002 |
| CP016 | World Labs is more visibly productized than General Intuition on public surfaces because Marble emphasizes creation, editing, exporting, and case-study workflows. | 中 | SP004, SP005 |
| CP017 | NVIDIA and DeepMind have stronger public ecosystem reach and distribution power than General Intuition. | 中 | SP006, SP008, SP009, SP010 |
| CP018 | OpenAI and Meta demonstrate that likely entrant pressure extends beyond named startups because large labs can repurpose video or predictive-model research into adjacent products. | 中 | SP013, SP020 |
| CP019 | For sophisticated buyers, the most credible substitute is often internal build on top of open frameworks rather than a single rival startup. | 中 | SP014, SP016, SP019, SP021 |
| CP020 | World Labs publicly emphasizes multimodal inputs, editable persistent worlds, and exportable outputs as key workflow features. | 中 | SP004, SP005 |
| CP021 | NVIDIA’s open-platform and simulation-ecosystem posture can pressure startup pricing in robotics-facing workflows. | 中 | SP009, SP010 |
| CP022 | Luma’s public packaging emphasizes fast end-to-end creative execution across video, image, audio, and text rather than a robotics-first stack. | 中 | SP011, SP012 |
| CP023 | No public General Intuition rate card was found in the reviewed materials. | 中 | SP001, SP002 |
| CP024 | Public pricing visibility is weak across most reviewed direct rivals, making packaging posture more comparable than list price. | 中 | SP004, SP005, SP006, SP009, SP011 |
| CP025 | Open-source and low-cost infrastructure raises the burden of proof for any startup trying to charge for early technical evaluation. | 中 | SP014, SP016, SP017, SP021, SP022 |
| CP026 | Customer multi-homing risk is high because buyers can mix proprietary models, incumbent platforms, and open-source simulators across the same workflow. | 中 | SP010, SP014, SP016, SP021 |
| CP027 | General Intuition’s moat thesis depends heavily on the idea that gameplay-derived action data produces better agent behavior than generic world-model training inputs. | 中 | SP001, SP003 |
| CP028 | Public evidence does not yet show named production customers, benchmark superiority, or hard lock-in for General Intuition. | 中 | SP001, SP002, SP003 |
| CP029 | World Labs has a more legible public workflow package for world creation than General Intuition currently exposes. | 中 | SP004, SP005 |
| CP030 | NVIDIA has the strongest public channel power in robotics among the reviewed competitors because it combines world models with Isaac Sim and broader Omniverse distribution. | 中 | SP009, SP010 |
| CP031 | DeepMind can sustain competitive pressure without near-term startup-style monetization because Genie and SIMA sit inside Alphabet-backed research programs. | 中 | SP006, SP007, SP008 |
| CP032 | OpenAI’s world-simulator framing suggests frontier labs can quickly collapse the line between video generation, simulation, and agent-training categories. | 中 | SP013 |
| CP033 | Luma and Rosebud show that some game or creative buyers can solve immediate needs with simpler creator tools instead of a deeper action-model platform. | 中 | SP011, SP012, SP015 |
| CP034 | Selective API access can improve curation and scarcity but does not itself create switching cost. | 中 | SP001, SP002 |
| CP035 | MIT Technology Review’s skeptical treatment of world models is adverse evidence that category enthusiasm can outrun reliability. | 中 | SP025 |
| CP036 | Improving open tooling means General Intuition is exposed to commoditization from below even if the overall category grows. | 中 | SP014, SP016, SP017, SP021, SP022 |
| CP037 | Category packaging in world-model and video-generation products is still volatile, which weakens confidence that today’s interface or monetization approach will persist. | 中 | SP013, SP025 |
| CP038 | General Intuition can be strategically correct about action models and still lose pricing power if incumbents or open ecosystems standardize the stack. | 中 | SP009, SP010, SP025 |
| CP039 | The strongest likely entrants to monitor beyond the named startup set are large foundation-model labs such as Meta and OpenAI. | 中 | SP013, SP020 |
| CP040 | There is no public benchmark set in the reviewed materials that directly compares General Intuition against World Labs, DeepMind, or NVIDIA on common tasks. | 中 | SP001, SP004, SP006, SP009 |
| CI001 | General Intuition’s only clearly public monetization surface is a selectively released commercial API rather than a broad self-serve product. | 中 | SI001, SI005 |
| CI002 | The company publicly says it has first partners across games, simulation, and robotics, implying a partner-led or enterprise-led early go-to-market motion. | 中 | SI001 |
| CI003 | No public General Intuition API price list or contract schedule was found in the reviewed materials. | 中 | SI001, SI005 |
| CI004 | No public revenue, ARR, customer-count, gross-margin, or burn-rate figure was found in the reviewed materials. | 中 | SI001, SI004, SI005, SI007, SI008 |
| CI005 | The selective API posture suggests early revenue is more likely to come from negotiated pilot or enterprise arrangements than from high-volume self-serve usage. | 中 | SI001, SI005 |
| CI006 | The careers surface includes a Financial Controller role, indicating the finance and control function is being built ahead of broader public financial disclosure. | 中 | SI002 |
| CI007 | The careers surface also includes infrastructure, data platform, security, and game integrations roles, which is consistent with a compute- and integration-heavy operating model. | 中 | SI002, SI003 |
| CI008 | Multiple roles on the careers page carry base-salary bands around $180K to $300K plus equity. | 中 | SI002 |
| CI009 | The Member of Technical Staff role lists a $250K to $450K salary band plus equity across New York, Geneva, London, and Paris. | 中 | SI003 |
| CI010 | Public salary bands imply that frontier research and infrastructure payroll is likely a major cost bucket even before considering employer taxes and equity expense. | 中 | SI002, SI003 |
| CI011 | CoreWeave’s pricing page lists NVIDIA HGX H100 on-demand pricing of $49.24 per hour and A100 pricing of $21.60 per hour. | 中 | SI013 |
| CI012 | AWS says P5 instances provide up to eight NVIDIA H100 GPUs with up to 640 GB of GPU memory and 3,200 Gbps of networking. | 中 | SI016 |
| CI013 | Public compute benchmarks indicate that training or serving frontier multimodal world models can become expensive quickly even before a company reaches scaled revenue. | 中 | SI013, SI016, SI018 |
| CI014 | Because General Intuition appears to be software-only, its cost structure is more likely dominated by payroll, cloud compute, data infrastructure, and security rather than hardware inventory. | 中 | SI001, SI002, SI003, SI016 |
| CI015 | Deloitte identifies legacy integration and risk/compliance concerns as leading barriers to agentic-AI adoption, which is relevant to General Intuition’s enterprise monetization path. | 中 | SI017 |
| CI016 | TechCrunch reported on June 18, 2026 that General Intuition was in talks to raise $300 million at around a $2 billion valuation. | 中 | SI004 |
| CI017 | DutchNews reported on June 25, 2026 that the disclosed financing became public at $320 million and a $2.3 billion valuation. | 中 | SI007 |
| CI018 | The move from the June 18 report of $300 million at roughly $2 billion to the June 25 report of $320 million at $2.3 billion suggests final round terms improved during disclosure. | 中 | SI004, SI007 |
| CI019 | DutchNews says total funding since October 2025 reached $454 million. | 中 | SI007 |
| CI020 | Public financing coverage names Jeff Bezos, Eric Schmidt, Khosla Ventures, and General Catalyst among investors in the latest disclosed round. | 中 | SI007, SI008 |
| CI021 | The SEC Form D is for AVSF - General Intuition 2026, LLC, a Delaware pooled investment fund vehicle, not a direct operating-company financial statement. | 中 | SI009 |
| CI022 | The SEC filing shows a total offering amount and total amount sold of $4,497,475 with 93 investors. | 中 | SI009 |
| CI023 | The SEC filing is financially relevant as evidence of a financing vehicle around the round, but it does not disclose General Intuition operating revenue, burn, or cash. | 中 | SI009, SI010, SI011 |
| CI024 | The DutchNews article says the company’s data and intellectual property are held through a Dutch company based in Naarden. | 中 | SI007 |
| CI025 | DutchNews says export-control concerns were given as a reason to maintain the Dutch company structure. | 中 | SI007 |
| CI026 | No public cash-on-hand, monthly-burn, runway, or debt figure was found in the reviewed materials. | 中 | SI004, SI005, SI007, SI008, SI009 |
| CI027 | A low-case annual burn assumption of roughly $60 million would imply about 64 months of runway from a standalone $320 million round before considering other cash needs. | 低 | SI002, SI003, SI013, SI016 |
| CI028 | A mid-case annual burn assumption of roughly $120 million would imply about 32 months of runway from a standalone $320 million round. | 低 | SI002, SI003, SI013, SI016 |
| CI029 | A high-case annual burn assumption of roughly $180 million would imply about 21 months of runway from a standalone $320 million round. | 低 | SI002, SI003, SI013, SI016 |
| CI030 | The most likely uses of the new capital are model training, inference infrastructure, hiring, security/compliance, and partner onboarding rather than hard-asset expansion. | 中 | SI001, SI002, SI003, SI013, SI016 |
| CI031 | No public debt facility, project-finance obligation, or inventory financing disclosure was found for General Intuition. | 中 | SI004, SI005, SI007, SI009, SI010 |
| CI032 | The next funding trigger is likely to depend more on proving production use cases and revenue conversion than on raising awareness, because the company already has large-capital backing. | 中 | SI001, SI005, SI007 |
| CI033 | Missing private metrics such as paying-customer count, ACV, gross margin, burn, and pipeline conversion block revenue-quality underwriting. | 中 | SI004, SI005, SI007, SI008 |
| CI034 | A selective API motion can create high-value initial contracts but usually produces lumpier and less predictable early revenue than self-serve SaaS. | 中 | SI001, SI005, SI017 |
| CI035 | The company’s public financing and partnership narrative is stronger than its public monetization evidence. | 中 | SI001, SI004, SI005, SI007, SI008 |
| CI036 | CoreWeave advertises reserved compute discounts of up to 60% versus on-demand pricing, implying that procurement optimization can materially change GPU economics. | 中 | SI013 |
| CI037 | AWS says P5 instances can reduce model-training cost by up to 40% versus the previous generation, showing that compute efficiency is a meaningful but not sufficient lever. | 中 | SI016 |
| CI038 | MIT Technology Review’s skepticism toward world-model reliability is adverse evidence that commercialization may lag capital deployment. | 中 | SI021 |
| CI039 | General Intuition’s financial outlook improved materially in 2026 because the company converted funding talks into a disclosed $320 million Series A at a $2.3 billion valuation. | 中 | SI004, SI007, SI008 |
| CI040 | Public evidence supports strong capitalization and ambition but does not support a defensible revenue multiple, margin forecast, or payback model. | 中 | SI004, SI005, SI007, SI008, SI009, SI017, SI021 |
| CE001 | General Intuition publicly positions itself as a lab building action models and world models rather than a conventional text-first AI application. | 中 | SE001, SE020 |
| CE002 | The company says action models decide what actions to take, while world models predict the outcomes of actions. | 中 | SE001 |
| CE003 | General Intuition says its models learn from action-labeled video datasets rather than from text alone. | 中 | SE001, SE022 |
| CE004 | General Intuition says Medal users upload billions of gameplay clips each year, forming the raw substrate for its product thesis. | 中 | SE001, SE024 |
| CE005 | General Intuition says it has onboarded first partners across games, simulation, and robotics to a commercial API. | 中 | SE001, SE024, SE025 |
| CE006 | The public partner portal asks for company role, website, what the company is building, and what should be built together, which indicates a high-touch enterprise intake flow rather than self-serve onboarding. | 中 | SE004 |
| CE007 | General Intuition’s public site does not expose a public rate card, open docs portal, or self-serve SDK download. | 中 | SE001, SE004, SE003 |
| CE008 | The site terms explicitly say the site is informational only and may change without notice, which weakens any attempt to treat website copy as a binding roadmap. | 中 | SE003 |
| CE009 | The privacy notice states that General Intuition is a Delaware corporation and gives a New York office address, plus EU and UK representatives in Naarden. | 中 | SE002 |
| CE010 | General Intuition’s privacy notice distinguishes site data from Medal gameplay data and says Medal-governed processing applies where General Intuition uses Medal data for research and model development. | 中 | SE002 |
| CE011 | MIRA is a playable multiplayer world model for Rocket League that runs in real time at 20 fps. | 中 | SE006, SE008 |
| CE012 | MIRA is described as a 5B-parameter diffusion transformer paired with a 600M-parameter video representation codec. | 中 | SE006, SE007 |
| CE013 | MIRA publicly claims to operate without a physics engine, rendering engine, or explicit 3D representation at inference time. | 中 | SE006 |
| CE014 | MIRA is trained on about 10,000 match-hours of bot-generated 2v2 Rocket League data rather than on human gameplay. | 中 | SE006 |
| CE015 | The public MIRA dataset release is smaller than the full training set: a 4,000-hour slice named Rocket Science at 720p with action streams and physics states. | 中 | SE007, SE008 |
| CE016 | The technical report says MIRA uses logged physics only for evaluation and not for training, which keeps the core world model grounded in pixels plus actions. | 中 | SE006, SE007 |
| CE017 | The public GitHub repo exposes installation, exploration, training, and evaluation commands, giving outside developers direct evidence of engineering maturity beyond marketing copy. | 中 | SE008 |
| CE018 | Codec training in the MIRA repo depends on a gated DINOv3-L/16 encoder from Meta, which makes part of the highest-fidelity training stack dependent on third-party weights. | 中 | SE008 |
| CE019 | IRIS shows the team’s earlier world-model lineage in discrete autoencoders plus autoregressive transformers. | 中 | SE009 |
| CE020 | Δ-IRIS extends that lineage into more efficient world models with context-aware tokenization. | 中 | SE010 |
| CE021 | DIAMOND demonstrates a diffusion-based world model lineage that emphasizes visual fidelity and interactive simulation, including CS:GO rollouts. | 中 | SE011 |
| CE022 | Wayve’s GAIA-2 shows a contrasting architecture aimed at controllable, multi-camera driving simulation with structured conditioning for weather, lanes, traffic, and actions. | 中 | SE012 |
| CE023 | DeepMind says Genie 2 generates action-controllable playable 3D environments from a single prompt image and can support agent training. | 中 | SE013 |
| CE024 | DeepMind says SIMA is a generalist agent that follows natural-language instructions across 3D virtual environments. | 中 | SE014 |
| CE025 | Physical Intelligence’s π0 uses broad robot data plus a vision-language-action architecture to emit low-level motor commands, representing a robotics-first alternative to General Intuition’s gameplay-first approach. | 中 | SE015 |
| CE026 | World Labs publicly emphasizes spatial intelligence, editable persistent 3D worlds, and exportable outputs, making its public packaging more workflow-specific than General Intuition’s current surface. | 中 | SE016 |
| CE027 | OpenAI explicitly frames video generation models as promising paths toward general-purpose simulators of the physical world, which validates the category but intensifies competition. | 中 | SE017 |
| CE028 | NVIDIA Cosmos markets an integrated physical-AI stack around world foundation models, data processing, training, and evaluation frameworks. | 中 | SE018 |
| CE029 | Kyutai is described as an open-science AI lab and is a named collaborator on MIRA, indicating that General Intuition is willing to collaborate externally on flagship technical releases. | 中 | SE008, SE019 |
| CE030 | TechCrunch says the startup currently has only a handful of customers in gaming, simulation, and robotics, which implies product maturity is still early relative to the size of the financing. | 中 | SE020 |
| CE031 | TechCrunch says the majority of the Series A proceeds will go to compute and that broader API availability is targeted for the end of summer 2026. | 中 | SE020, SE023 |
| CE032 | The Robot Report says the company uses billions of Medal gameplay clips instead of collecting large quantities of real-world robotics data or synthetic simulation first. | 中 | SE021 |
| CE033 | SiliconANGLE reports that the company combines world models with action models and ties the founding team to DIAMOND, IRIS, and GAIA-2 research threads. | 中 | SE022 |
| CE034 | InvestGame says General Intuition’s compute scaling is tied to a CoreWeave partnership and that broader commercial API access is planned by the end of summer 2026. | 中 | SE023 |
| CE035 | Coalition says the company has launched Nerve, a data-collection platform, which expands the product surface beyond a model API into supervised data acquisition. | 中 | SE024 |
| CE036 | GamesBeat says the company has onboarded first partners across games, simulation, and robotics but will still work selectively with only a few companies ahead of a broader model release. | 中 | SE025 |
| CE037 | The MIRA blog explicitly says the demo is a stepping stone to physical AI and that sim-to-real transfer remains something the authors do not yet know how far it can take them. | 中 | SE006 |
| CE038 | The MIRA materials disclose concrete limitations, including replay failures, single-player hidden-state challenges, and the possibility that narrow-domain stability may not generalize to messy real-world video. | 中 | SE006 |
| CE039 | No public source reviewed exposed a security certification, formal model card, or public trust center for the commercial API. | 中 | SE001, SE002, SE003, SE004 |
| CE040 | The public evidence supports a technically credible research stack with real code and demos, but it does not yet support the claim that the broader platform is productized beyond selective partner onboarding. | 中 | SE001, SE004, SE008, SE020, SE025 |
| CU001 | General Intuition publicly says it has first partners across games, simulation, and robotics on a commercial API. | 中 | SU001, SU005 |
| CU002 | The partner portal asks companies what they are building and what could be built together, implying consultative intake rather than self-serve signup. | 中 | SU002 |
| CU003 | GamesBeat reports that General Intuition has onboarded first partners across games, simulation, and robotics and will still work selectively with only a few companies before broader release. | 中 | SU003 |
| CU004 | TechCrunch reports that the startup currently has only a handful of customers in gaming, simulation, and robotics. | 中 | SU004 |
| CU005 | Coalition Capital describes first commercial partners but does not name them. | 中 | SU005 |
| CU006 | InvestGame says a commercial API for gaming, simulation, and robotics partners has launched and broader access is planned by the end of summer 2026. | 中 | SU006 |
| CU007 | Tech Funding News says investors backed the research trajectory rather than a commercial product, which is an adverse signal on customer maturity. | 中 | SU007 |
| CU008 | Axios frames the company around gaming-derived training rather than around a documented customer roster or case-study set. | 中 | SU008 |
| CU009 | The Verge frames the company as a big bet on Medal data and world models rather than as a business with disclosed customer deployments. | 中 | SU009 |
| CU010 | The SaaS News says broader capital deployment still centers compute and research hiring, not public customer expansion metrics. | 中 | SU010 |
| CU011 | The Robot Report says API availability is still expected to broaden in summer 2026, implying current access remains constrained. | 中 | SU011 |
| CU012 | Medal is a broad gaming clip platform rather than a named General Intuition customer; its relevance is as upstream data and community supply. | 中 | SU012, SU013 |
| CU013 | Medal says it works with every game and supports instant sharing, which increases ecosystem breadth but does not prove that game studios pay General Intuition today. | 中 | SU012, SU013 |
| CU014 | Medal support and product pages show an existing user community and support surface, but not a General Intuition customer success or review surface. | 中 | SU013, SU014 |
| CU015 | Medal terms and privacy show a robust user-generated-content platform with its own rights and privacy regime, reinforcing that Medal is infrastructure for General Intuition rather than customer proof for it. | 中 | SU015, SU016, SU017 |
| CU016 | General Intuition’s site and legal pages do not disclose pricing, contract length, or customer-count metrics beyond broad partner statements. | 中 | SU001, SU017, SU018 |
| CU017 | CoreWeave is a named ecosystem partner for compute, not a disclosed end-customer of General Intuition’s models. | 中 | SU006, SU019 |
| CU018 | Kyutai and Epic Games are named collaborators on MIRA, but public materials do not describe them as paying API customers. | 中 | SU020, SU021, SU025 |
| CU019 | No reviewed public source names a specific paying game studio customer. | 中 | SU001, SU003, SU004, SU005, SU006 |
| CU020 | No reviewed public source names a specific paying robotics customer. | 中 | SU001, SU003, SU004, SU006, SU011 |
| CU021 | No reviewed public source names a specific paying simulation customer. | 中 | SU001, SU003, SU004, SU006 |
| CU022 | The customer base that is publicly inferable today is segmented more by target workflow—games, simulation, robotics—than by named account list, geography, or revenue band. | 中 | SU001, SU003, SU004, SU006 |
| CU023 | The gaming segment is the strongest publicly evidenced beachhead because both the training data and several use-case examples are drawn from games. | 中 | SU001, SU003, SU004, SU012 |
| CU024 | Simulation is presented publicly as a target segment for testing agents in digital twins or synthetic environments, but no named simulation buyer is disclosed. | 中 | SU001, SU004, SU006 |
| CU025 | Robotics is presented publicly through demos such as quadruped navigation and hazardous-environment use cases, but no named robotics operator is disclosed. | 中 | SU004, SU011, SU003 |
| CU026 | The go-to-market motion appears enterprise-led because access is selective, partner intake is bespoke, and public evidence emphasizes embedded collaboration rather than mass developer self-service. | 中 | SU001, SU002, SU003, SU004 |
| CU027 | The partner-first motion also implies procurement friction, because integration likely requires joint evaluation, data-sharing, or internal research collaboration. | 中 | SU002, SU004 |
| CU028 | Public adoption metrics stop at “handful of customers” plus broad partner-category language; no public account count, usage volume, or deployment-location metric was found. | 中 | SU004, SU005, SU006 |
| CU029 | No public source reviewed disclosed NRR, GRR, churn, renewal rates, contract duration, or satisfaction scores. | 中 | SU001, SU004, SU018 |
| CU030 | No public G2, Gartner Peer Insights, or comparable review signal was found for General Intuition. | 中 | SU001, SU018 |
| CU031 | The named-proof set is strongest on partner or collaborator visibility—CoreWeave, Kyutai, Epic—rather than on named paying users. | 中 | SU017, SU018, SU019, SU020, SU021, SU025 |
| CU032 | Because public customer proof is sparse, the named customer proof table in this chapter is necessarily partial and includes ecosystem proof rather than confirmed payer proof in several rows. | 中 | SU003, SU004, SU006, SU025 |
| CU033 | Concentration risk is likely high if the company is currently serving only a few customers while customizing integrations for each. | 中 | SU003, SU004 |
| CU034 | Expansion depends on whether selective partner projects become reusable workflows that can generalize across many embodiments rather than staying services-heavy. | 中 | SU002, SU004, SU006 |
| CU035 | The roadmap for broader API availability suggests the company is still in the proof-building phase of commercialization rather than in scaled deployment. | 中 | SU004, SU006, SU011 |
| CU036 | Medal’s broad gamer ecosystem and Nerve-style data collection could create a future funnel of developers and data suppliers, but that is not the same thing as current recurring customers. | 中 | SU012, SU013, SU005 |
| CU037 | The strongest public evidence of customer relevance today is not ROI proof but repeated third-party confirmation that outside organizations in games, simulation, and robotics are already in selective engagement. | 中 | SU001, SU003, SU004, SU005, SU006 |
| CU038 | The public customer verdict is that segment fit looks plausible, but durability, reference quality, and production scale remain unproven because no named paying accounts, renewal metrics, or outcomes are disclosed. | 中 | SU003, SU004, SU006, SU007, SU009 |
| CR001 | General Intuition publicly says it has first partners across games, simulation, and robotics but remains in a selective pre-broad-release phase. | 中 | SR001, SR004, SR005 |
| CR002 | The partner portal asks counterparties what they are building and what could be built together, implying bespoke co-development rather than commodity self-serve onboarding. | 中 | SR004 |
| CR003 | TechCrunch reports that the startup still has only a handful of customers in gaming, simulation, and robotics. | 中 | SR006 |
| CR004 | Tech Funding News frames the financing as a bet on research trajectory rather than on a mature commercial product. | 中 | SR008 |
| CR005 | InvestGame says broader API availability is planned by the end of summer 2026, implying current rollout is still staged. | 中 | SR007 |
| CR006 | GamesBeat also describes the company as selectively working with only a few companies ahead of broader release. | 中 | SR005 |
| CR007 | General Intuition’s privacy notice says Medal data used for research and model development is governed by the Medal privacy policy and the arrangement between General Intuition and Medal. | 中 | SR002 |
| CR008 | The same privacy notice distinguishes the marketing site from Medal-platform processing, which means key training-data governance lives outside the main General Intuition site disclosures. | 中 | SR002, SR003 |
| CR009 | General Intuition’s site privacy notice names EU and UK representatives and says international transfers rely on contractual safeguards, showing the company already faces cross-border privacy-compliance work. | 中 | SR002 |
| CR010 | General Intuition’s terms describe the site as informational only and provide no public warranties that the site or systems are uninterrupted, error-free, or secure. | 中 | SR003 |
| CR011 | Medal’s terms require users to comply with applicable law and third-party rights, which underscores that gameplay-clip rights and downstream training rights are legally distinct questions. | 中 | SR014 |
| CR012 | Medal support surfaces active user-support and community processes, which is useful operationally but also shows a UGC platform whose data quality and policy enforcement can affect upstream training inputs. | 中 | SR013 |
| CR013 | Medal’s product pages show it can capture and organize clips across many games, reinforcing that General Intuition’s data moat is tied to a broad but externally facing consumer platform. | 中 | SR011, SR012 |
| CR014 | CoreWeave is the publicly named compute partner behind model scaling and rollout, creating visible infrastructure concentration. | 中 | SR007, SR015 |
| CR015 | TechCrunch says most of the Series A proceeds will go to compute, a strong public signal of capital intensity. | 中 | SR006 |
| CR016 | The Robot Report says General Intuition is using billions of Medal gameplay clips rather than collecting equivalent volumes of real-world robotics data first, which sharpens both its data advantage and its transfer-risk profile. | 中 | SR010 |
| CR017 | DeepMind’s Genie 2 is a public large-scale foundation world model for action-controllable 3D environments, showing that a global incumbent is shipping directly into the same conceptual category. | 中 | SR016 |
| CR018 | Physical Intelligence’s π0 is a vision-language-action system emitting low-level robot actions, representing a robotics-first substitute path for embodied control. | 中 | SR017 |
| CR019 | World Labs publicly markets spatial-intelligence products, and TechCrunch reports it has raised more than $1 billion, increasing competitive pressure from a better-capitalized peer in adjacent world-model workflows. | 中 | SR018, SR030 |
| CR020 | OpenAI explicitly frames video-generation models as promising world simulators, validating the category while increasing the probability that foundational-model leaders compress differentiation. | 中 | SR019 |
| CR021 | NVIDIA Cosmos packages world foundation models with data processing, training, and evaluation infrastructure, which raises the risk that platform vendors bundle capabilities General Intuition hopes to sell independently. | 中 | SR020 |
| CR022 | BIS’s January 2025 semiconductor-control update adds broader license requirements and more due-diligence obligations around advanced chips. | 中 | SR021 |
| CR023 | BIS guidance published in May 2026 says licenses are required for advanced-computing exports to entities headquartered in Country Group D:5 or Macau even when located elsewhere, complicating global customer and supply-chain planning. | 中 | SR022 |
| CR024 | The May 2025 rescission of the AI Diffusion Rule and promise of a replacement rule show that AI-chip policy is still moving, which makes long-range infrastructure planning less stable. | 中 | SR023 |
| CR025 | NIST’s AI RMF and GenAI profile show that frontier-AI vendors are increasingly expected to document trust, governance, and risk-management controls even when the framework is voluntary. | 中 | SR024 |
| CR026 | The European Commission says the AI Act uses four risk levels and that implementation was still being simplified in May 2026, meaning European go-to-market requirements remain material and evolving. | 中 | SR025 |
| CR027 | The U.S. Copyright Office is still analyzing how copyright law applies to AI training on copyrighted materials, leaving meaningful legal uncertainty around training-data doctrine. | 中 | SR026 |
| CR028 | The FTC has issued a policy statement on biometric information under Section 5, which matters if General Intuition expands from gameplay clips toward richer human, voice, or face-linked datasets. | 中 | SR027 |
| CR029 | Illinois BIPA still defines biometric identifiers to include voiceprints and scans of face geometry, highlighting how embodied or human-video expansion can import state-law exposure. | 中 | SR028 |
| CR030 | General Intuition’s visible public mitigations on privacy are stronger than its public mitigations on training-data rights, model safety, or security assurance. | 中 | SR002, SR003, SR014, SR024 |
| CR031 | No public trust center, SOC 2 disclosure, dedicated security page, or uptime/SLA documentation was found on the reviewed public surfaces. | 中 | SR001, SR003, SR004 |
| CR032 | Because the commercial API remains selective and lightly documented, product reliability and deployment maturity are hard to verify from public evidence. | 中 | SR001, SR005, SR006, SR007 |
| CR033 | Having only a handful of customers and no named paying logos makes concentration risk difficult to size but likely high. | 中 | SR005, SR006, SR007 |
| CR034 | The company’s training and product narrative is tightly coupled to Medal, so any separation, sale, policy change, or slowdown at Medal would weaken the core data flywheel. | 中 | SR002, SR011, SR012, SR013 |
| CR035 | Public evidence also suggests that Medal is more than historical provenance: it remains the most visible route through which gameplay volume, community behavior, and future labeling loops can flow into the model stack. | 中 | SR001, SR011, SR029 |
| CR036 | Public narrative around General Intuition remains founder-centric, with CEO Pim de Witte carrying much of the external storytelling burden in coverage and interviews. | 中 | SR006, SR009 |
| CR037 | Capital-intensity risk is amplified because the company is funding frontier research, compute expansion, and specialized hiring before broad commercialization is visible. | 中 | SR006, SR007, SR008, SR015 |
| CR038 | If broader release slips while compute costs stay high, the risk transmits directly from product delay to burn, financing need, and valuation pressure. | 中 | SR005, SR006, SR007, SR015 |
| CR039 | Expansion from gaming into robotics or other physical-world use cases would likely raise the combined burden of safety, privacy, and regulatory compliance relative to the current public disclosure set. | 中 | SR010, SR024, SR025, SR027, SR028 |
| CR040 | The strongest visible mitigations today are capital availability, a named compute partner, legal/privacy pages, and credible research output, but none of those fully resolves commercialization or governance risk. | 中 | SR002, SR007, SR015, SR024, SR029 |
| CR041 | The most important remaining diligence asks are a customer list, compute-cost curve, training-data rights chain, security program evidence, and a use-case-by-use-case regulatory map. | 中 | SR003, SR006, SR014, SR024, SR025 |
| CR042 | Overall, General Intuition’s risk profile is investable only if an investor accepts frontier-model uncertainty and can verify private mitigations that are not visible in public materials. | 中 | SR006, SR008, SR019, SR030 |
| CV001 | TechCrunch and GamesBeat report that General Intuition raised a $320 million Series A at a $2.3 billion post-money valuation in June 2026. | 中 | SV001, SV002 |
| CV002 | Using the disclosed post-money valuation and round size, the implied pre-money valuation is about $1.98 billion. | 中 | SV001, SV002 |
| CV003 | Public commercialization evidence is still early: TechCrunch says the company has only a handful of customers and broader API access is not yet fully open. | 中 | SV001, SV003, SV006 |
| CV004 | InvestGame and The Robot Report say broader API availability is planned rather than already generalized, which keeps near-term monetization timing uncertain. | 中 | SV003, SV006 |
| CV005 | Tech Funding News says investors backed the research trajectory rather than a commercial product, which is adverse evidence for paying peak-like narrative prices. | 中 | SV004 |
| CV006 | Because the current public record shows selective customer proof, the valuation already embeds substantial success that has not yet been publicly demonstrated. | 中 | SV001, SV002, SV004 |
| CV007 | World Labs raised a $1 billion round in February 2026, with TechCrunch and Reuters both tying market discussion to about a $5 billion valuation level even though the company did not publicly confirm an exact mark. | 中 | SV013, SV014 |
| CV008 | World Labs also had a released product, Marble, and an Autodesk partnership aimed at commercial workflows, which gives its private-mark discussion somewhat more visible productization than General Intuition currently shows. | 中 | SV013 |
| CV009 | Physical Intelligence was reported in March 2026 to be discussing a $1 billion raise at a valuation above $11 billion, roughly double its $5.6 billion mark from four months earlier. | 中 | SV015 |
| CV010 | The same TechCrunch report says Physical Intelligence had no timeline for commercialization and still believed there was effectively unlimited compute it could deploy, showing that frontier embodied-AI rounds can run far ahead of revenue proof. | 中 | SV015 |
| CV011 | Stanford’s 2026 AI Index says U.S. private AI investment reached $285.9 billion in 2025 and that billion-dollar funding events nearly doubled. | 中 | SV020 |
| CV012 | The same report describes frontier AI valuation events such as OpenAI at $300 billion and Anthropic at $183 billion, proving that 2026 capital markets remained willing to finance AI labs at extraordinary prices. | 中 | SV020 |
| CV013 | Finerva says the median public robotics and AI revenue multiple rose to 3.4x by Q4 2025. | 中 | SV021 |
| CV014 | Finerva also says the high end of the same public cohort still reached 24.0x revenue, meaning premium outcomes exist but are the exception rather than the median. | 中 | SV021 |
| CV015 | CompaniesMarketCap reports C3.ai at a roughly $1.39 billion public market cap as of July 2026. | 中 | SV022 |
| CV016 | SEC EDGAR shows C3.ai had a current 10-K filing dated February 27, 2026, which means investors can evaluate it against much richer public disclosure than General Intuition offers. | 中 | SV023 |
| CV017 | CompaniesMarketCap reports Unity at a roughly $13.40 billion public market cap as of July 2026. | 中 | SV024 |
| CV018 | SEC EDGAR shows Unity had a current 10-K filing dated February 11, 2026, again highlighting the disclosure advantage public comps have over General Intuition. | 中 | SV025 |
| CV019 | General Intuition’s $2.3 billion post-money valuation already exceeds C3.ai’s public market cap while remaining well below Unity’s, so the public-comp bracket is broad but not obviously supportive of calling the current round cheap. | 中 | SV001, SV022, SV024 |
| CV020 | CoreWeave is the named compute partner, linking valuation support directly to continued access to large-scale infrastructure. | 中 | SV003, SV012 |
| CV021 | TechCrunch says most of the Series A proceeds will go to compute, reinforcing that this is a capex-like AI software bet rather than a light, self-serve SaaS story. | 中 | SV001 |
| CV022 | BIS export-control updates, NIST governance expectations, the EU AI Act, and the Copyright Office’s ongoing AI-training review all enlarge the discount rate an investor should apply to forward scenarios. | 中 | SV026, SV027, SV028, SV029 |
| CV023 | DeepMind’s Genie 2, OpenAI’s world-simulator framing, NVIDIA Cosmos, World Labs, and Physical Intelligence collectively validate the category but make differentiation expensive and fragile. | 中 | SV013, SV015, SV017, SV018, SV019 |
| CV024 | DeepMind’s Genie 2 is a public foundation world model for action-controllable 3D environments, demonstrating that a deep-pocketed incumbent is shipping adjacent capability. | 中 | SV017 |
| CV025 | OpenAI explicitly frames video-generation models as world simulators, which increases the risk that foundational-model leaders collapse category novelty into broader platforms. | 中 | SV018 |
| CV026 | NVIDIA Cosmos packages world models with surrounding tooling, raising the odds that infrastructure vendors bundle functionality that startups hoped to sell as stand-alone products. | 中 | SV019 |
| CV027 | General Intuition’s public surface still lacks public pricing, API docs, customer case studies, or SLA-style disclosure, which limits the evidence basis for a high-conviction Buy call. | 中 | SV007, SV009, SV010 |
| CV028 | The company also lacks named public paying-customer proof, which weakens any attempt to defend the mark with conventional commercial traction arguments. | 中 | SV001, SV002, SV003 |
| CV029 | Because revenue, margins, and retention are not publicly disclosed, a precision EV/revenue or DCF-style underwriting model would be false precision. | 中 | SV001, SV009, SV021 |
| CV030 | A milestone-adjusted scenario framework is therefore more defensible than a single-point multiple model. | 中 | SV003, SV021, SV026 |
| CV031 | The bull case requires broader API release, named production customers, clear evidence that Medal-derived data creates a durable moat, and enough competitive separation to deserve a premium closer to leading private world-model comps. | 中 | SV003, SV007, SV013, SV015 |
| CV032 | The base case assumes the company remains technically credible but commercially selective, making a valuation near the current round difficult to call cheap. | 中 | SV001, SV003, SV004, SV021 |
| CV033 | The bear case is that rollout slips, compute burn stays high, customer proof remains thin, and the next financing happens on less favorable terms. | 中 | SV001, SV004, SV012, SV026 |
| CV034 | At the current price, expected return looks more dependent on near-bull-case execution than on base-case delivery. | 中 | SV001, SV021, SV013 |
| CV035 | That asymmetry makes the correct recommendation price-sensitive: the company may be exciting, but the round does not yet look comfortably underwritten on public evidence. | 中 | SV001, SV004, SV021 |
| CV036 | The most supportable recommendation is Research-More / Track rather than Buy. | 中 | SV001, SV004, SV021, SV026 |
| CV037 | Confidence should be medium, not high, because the price is explicit while the revenue model, unit economics, and cap-table detail are not. | 中 | SV001, SV009, SV021 |
| CV038 | Risk rating should be high because valuation depends on hard-to-verify product, data, competitive, regulatory, and financing milestones landing together. | 中 | SV001, SV012, SV026, SV028, SV029 |
| CV039 | Valuation stance should be expensive rather than fair, since the current round looks above a conservative public-evidence base case and only moderately below aggressive frontier-AI peer marks. | 中 | SV001, SV013, SV015, SV021 |
| CV040 | The best current use of public comps is to bound scenarios, not to prove that $2.3 billion is a bargain. | 中 | SV013, SV015, SV021, SV022, SV024 |
| CV041 | A reasonable current underwriting range is roughly $0.6B-$1.2B in bear, $1.4B-$2.2B in base, and $3.0B-$5.0B in bull, expressed as today's value rather than a future exit. | 中 | SV001, SV013, SV015, SV021, SV022, SV024 |
| CV042 | The current $2.3B round sits above the base-range midpoint and closer to the upper end of what can be justified without clearer private diligence. | 中 | SV001, SV021, SV013 |
| CV043 | What would change the call is tangible rather than rhetorical: customer names, contract shape, compute-cost curve, data-rights chain, and cap-table / preference disclosure. | 中 | SV001, SV008, SV009, SV012, SV030 |
| CV044 | The thesis breaks if broader API release misses again, no named production users emerge, export or compliance friction materially slows deployment, or the next round resets price. | 中 | SV003, SV006, SV026, SV028 |
| CV045 | Exit readiness is low because the company is nowhere near public-market disclosure norms on revenue, margins, customer concentration, or governance detail. | 中 | SV009, SV023, SV025 |
| CV046 | Filings from C3.ai and Unity are useful mainly because they show what disclosure-rich comparables look like; they cannot close the core General Intuition information gap. | 中 | SV023, SV025 |
| CV047 | The anti-thesis is not that world models lack value; it is that the current price asks investors to pay now for proof that remains private or future-dated. | 中 | SV001, SV013, SV015, SV020 |
| CV048 | The positive counterargument is that 2026 AI capital markets continued to reward frontier labs very aggressively, so a multibillion mark for General Intuition is not anomalous in context. | 中 | SV013, SV015, SV020 |
| CV049 | Even so, World Labs and Physical Intelligence were either more richly financed or already associated with clearer product or category leadership cues, limiting the read-through that General Intuition is automatically underpriced. | 中 | SV013, SV015, SV016 |
| CV050 | The final valuation verdict is therefore to keep tracking the company, but demand either a lower entry price or materially better private evidence before upgrading the recommendation. | 中 | SV001, SV004, SV021, SV026 |
| 编号 | 出版方 | 标题 | 引文 |
|---|---|---|---|
| SO001 | General Intuition | General Intuition | The frontier lab for acting in space and time. | We have onboarded our first partners across games, simulation, and robotics to our commercial API and will be selectively working with a few companies ahead of the broader release of our model. |
| SO002 | Medal | Medal FAQ / platform overview | |
| SO003 | TechCrunch | General Intuition’s $2.3B bet that video games can train AI agents for the real world | General Intuition said it raised $320 million at a $2.3 billion valuation, bringing total disclosed funding to $454 million after the $134 million round it raised at launch last October. |
| SO004 | TechCrunch | General Intuition in talks to raise $300M at around $2B valuation | The startup trains embodied AI and world models using Medal’s dataset of 2 billion videos per year from 10 million monthly active users. |
| SO005 | Tech Funding News | General Intuition bags $320M Series A at $2.3B to build the AI that learns to act from gamers | General Intuition operates as a public-benefit corporation, a legal structure that requires the company to consider broader social impact alongside profit, and is legally registered in the Netherlands. |
| SO006 | The SaaS News | General Intuition Raises $320M Series A | General Intuition, a New York-based AI lab focused on training models to act in the real world using gameplay data, has raised $320M in a Series A round. |
| SO007 | DutchNews | Dutch AI firm General Intuition raises $320 million in new round | The latest funding round, completed in January but only now made public, means the company is now valued at $2.3 billion. |
| SO008 | The Robot Report | General Intuition raises $320M to use video game data to train robots | The New York-based company said its Series A brings its valuation to $2.3 billion. |
| SO009 | The AI Insider | Spatial AI Training Startup General Intuition Valued at $2.3B After $320M Series A Funding Round | General Intuition trains large action foundation models on action-labeled gameplay clips from Medal’s 17 million monthly active users. |
| SO010 | Startup Fortune | General Intuition raises $320 million on the thesis that video game footage is the most underrated training data in robotics | The company is already in conversations for a Series B and has rejected multiple acquisition approaches, with the majority of new funding going toward compute via a deal with CoreWeave. |
| SO011 | Andrew.ooo | General Intuition $320M Series A: Gameplay AI (June 2026) | The transfer-to-real-world question is the open empirical risk. |
| SO012 | General Catalyst | General Intuition | General Catalyst Portfolio | |
| SO013 | Backed VC | General Intuition | Backed portfolio | General Intuition is a frontier research lab dedicated to gaming AI, spun out from Medal.tv. |
| SO014 | Slush | Pim de Witte — Slush speaker profile | Pim de Witte is the CEO of General Intuition ... He is also the co-founder & former CEO of Medal. |
| SO015 | ai-PULSE | Pim de Witte speaker profile | |
| SO016 | Eloi Alonso | Eloi Alonso personal site | Hello! I’m a researcher and co-founder at General Intuition. Before that, I worked on reinforcement learning and world models during my PhD, in François Fleuret’s group at the University of Geneva. |
| SO017 | MIRA authors | MIRA: Multiplayer Interactive World Models with Representation Autoencoders | We introduce the first multiplayer world model for highly dynamic environments governed by complex physical interactions. |
| SO018 | GitHub | mira-wm/mira | MIRA is a real-time world model of Rocket League ... a 5B-parameter latent diffusion model. |
| SO019 | MIRA | MIRA blog post | The project is a stepping stone to physical AI, where data is messier and scarcer. |
| SO020 | The Next Web | General Intuition is raising $300 million to train AI agents on the video game data OpenAI tried to buy | OpenAI reportedly offered $500 million to acquire Medal ... Instead he spun out General Intuition in October 2025. |
| SO021 | MIT Technology Review | World models | Today’s AI is still unreliable. |
| SO022 | Stanford HAI | The 2026 AI Index Report | |
| SO023 | CoreWeave | CoreWeave announces agreement with OpenAI to deliver AI infrastructure | |
| SO024 | General Intuition | generalintuition.ai landing attempt | |
| SO025 | General Intuition | General Intuition about page attempt | |
| SM001 | General Intuition | General Intuition | The frontier lab for acting in space and time. | We have onboarded our first partners across games, simulation, and robotics to our commercial API and will be selectively working with a few companies ahead of the broader release of our model. |
| SM002 | TechCrunch | General Intuition’s $2.3B bet that video games can train AI agents for the real world | The company says once it gets its API into more customers’ hands, it would be able to test its mettle with a variety of use cases. |
| SM003 | The Robot Report | General Intuition raises $320M to use video game data to train robots | General Intuition uses billions of gameplay clips uploaded to Medal to build AI models that can perceive, predict, and act in virtual and physical environments. |
| SM004 | Kaiso Research / MarketResearch.com | Global AI World Models Market Size, Opportunity Analysis and Forecast | The Global AI World Models market was valued at USD 1.8 billion in 2025, and is projected to reach USD 52.7 billion by 2035, growing at a CAGR of 40.2% from 2026 to 2035. |
| SM005 | Grand View Research | Artificial Intelligence in Robotics Market | The global artificial intelligence in robotics market size was estimated at USD 20,433.0 million in 2025 and is projected to reach USD 182,705.1 million by 2033, growing at a CAGR of 32.0% from 2026 to 2033. |
| SM006 | The Business Research Company | Artificial Intelligence (AI)-Powered Simulation And Digital Twins Market Report 2026 | The artificial intelligence (AI)-powered simulation and digital twins market size will grow from $5.18 billion in 2025 to $6.89 billion in 2026. |
| SM007 | Fortune Business Insights | Digital Twin Market Size, Share & Growth Report | The global digital twin market size was valued at USD 24.48 billion in 2025 and is projected to grow from USD 33.97 billion in 2026 to USD 384.79 billion by 2034. |
| SM008 | The Business Research Company | Generative AI In Gaming Market Report 2026 | The generative AI in gaming market size will grow from $1.79 billion in 2025 to $2.21 billion in 2026. |
| SM009 | The Business Research Company | Artificial Intelligence (AI) In Games Market Report 2026 | The artificial intelligence (AI) in games market size will grow from $2.87 billion in 2025 to $3.4 billion in 2026. |
| SM010 | Google DeepMind | Genie 2: A large-scale foundation world model | Today we introduce Genie 2, a foundation world model capable of generating an endless variety of action-controllable, playable 3D environments for training and evaluating embodied agents. |
| SM011 | NVIDIA | NVIDIA Cosmos | Develop physical AI faster with leading world foundation models and open data processing, training, and evaluation frameworks. |
| SM012 | 360iResearch | Robotics Simulation Market - Global Forecast 2026-2032 | The Robotics Simulation Market size was estimated at USD 6.88 billion in 2025 and expected to reach USD 7.58 billion in 2026. |
| SM013 | Deloitte | AI trends: Adoption barriers and updated predictions | According to nearly 60% of AI leaders, the primary challenges in adopting agentic AI are integrating with legacy systems and addressing risk and compliance concerns. |
| SM014 | arXiv | 3D Generation for Embodied AI and Robotic Simulation: A Survey | Embodied AI and robotic systems increasingly depend on scalable, diverse, and physically grounded 3D content for simulation-based training and real-world deployment. |
| SM015 | MuJoCo | MuJoCo — Advanced Physics Simulation | MuJoCo is a free and open source physics engine that aims to facilitate research and development in robotics. |
| SM016 | NVIDIA Developer | NVIDIA Isaac Sim | Isaac Sim is an open source reference framework built on NVIDIA Omniverse libraries for robotics simulation, testing, and synthetic data generation. |
| SM017 | robosuite | robosuite | robosuite is a simulation framework powered by the MuJoCo physics engine for robot learning. |
| SM018 | Princeton Vision & Learning Lab | Infinigen | Infinigen is a procedural generator of 3D scenes optimized for computer vision research and diverse training data. |
| SM019 | StartUs Insights | Digital Twin Report 2026: Scaling Toward a USD 70B+ Infrastructure | The digital twin market is transitioning from experimentation to decision-grade infrastructure. |
| SM020 | MIT Technology Review | World models | Today’s AI is still unreliable. |
| SM021 | MIRA | MIRA blog post | The project is a stepping stone to physical AI, where data is messier and scarcer. |
| SM022 | The Next Web | General Intuition is raising $300 million to train AI agents on the video game data OpenAI tried to buy | General Intuition builds world models to train agents, making the agents the product and the world model the training ground. |
| SM023 | The AI Insider | Spatial AI Training Startup General Intuition Valued at $2.3B After $320M Series A Funding Round | General Intuition is building large action foundation models trained on billions of action-labeled gameplay clips collected through Medal. |
| SM024 | Startup Fortune | General Intuition raises $320 million on the thesis that video game footage is the most underrated training data in robotics | The market the company is entering is large and moving fast. |
| SM025 | MIRA authors | MIRA: Multiplayer Interactive World Models with Representation Autoencoders | We introduce the first multiplayer world model for highly dynamic environments governed by complex physical interactions. |
| SP001 | General Intuition | General Intuition | The frontier lab for acting in space and time. | We have onboarded our first partners across games, simulation, and robotics to our commercial API. |
| SP002 | TechCrunch | General Intuition's $2.3B bet that video games can train AI agents for the real world | The company says once it gets its API into more customers’ hands, it would be able to test its mettle with a variety of use cases. |
| SP003 | The Robot Report | General Intuition raises $320M to use video game data to train robots | General Intuition uses billions of gameplay clips uploaded to Medal to build AI models that can perceive, predict, and act in virtual and physical environments. |
| SP004 | World Labs | World Labs | World Labs is a leading spatial intelligence company, building frontier models that can perceive, generate, reason, and interact with the 3D world. |
| SP005 | World Labs | Marble | Marble, our first product, generates spatially consistent, high-fidelity, and persistent 3D worlds that you can move through, edit, and inhabit. |
| SP006 | Google DeepMind | Genie 2: A large-scale foundation world model | Today we introduce Genie 2, a foundation world model capable of generating an endless variety of action-controllable, playable 3D environments for training and evaluating embodied agents. |
| SP007 | Google DeepMind | A generalist AI agent for 3D virtual environments | SIMA is a generalist AI agent for 3D virtual environments. |
| SP008 | Google DeepMind | Genie 3: A new frontier for world models | Genie 3 is a new frontier for world models. |
| SP009 | NVIDIA | NVIDIA Cosmos | Develop physical AI faster with leading world foundation models and open data processing, training, and evaluation frameworks. |
| SP010 | NVIDIA Developer | Isaac Sim | Isaac Sim is an open source reference framework built on NVIDIA Omniverse libraries for robotics simulation, testing, and synthetic data generation. |
| SP011 | Luma | Luma | AI Agents for Creative Work | Our Mission is to build unified general intelligence that can generate, understand, and operate in the physical world. |
| SP012 | Luma | Creative agents that make you prolific | Agents research, generate, and refine across video, image, audio, and text. |
| SP013 | OpenAI | Video generation models as world simulators | Our results suggest that scaling video generation models is a promising path towards building general purpose simulators of the physical world. |
| SP014 | Unity Technologies / GitHub | GitHub - Unity-Technologies/ml-agents | The Unity Machine Learning Agents Toolkit (ML-Agents) is an open-source project that enables games and simulations to serve as environments for training intelligent agents. |
| SP015 | Rosebud AI | Rosebud AI Game Maker | Create Games with AI | Create Games with AI. |
| SP016 | MuJoCo | MuJoCo — Advanced Physics Simulation | MuJoCo is a free and open source physics engine that aims to facilitate research and development in robotics. |
| SP017 | robosuite | robosuite | robosuite is a simulation framework powered by the MuJoCo physics engine for robot learning. |
| SP018 | Princeton Vision & Learning Lab | Home | Infinigen | Infinigen is a procedural generator of 3D scenes optimized for computer vision research and diverse training data. |
| SP019 | AI Habitat | AI Habitat | AI Habitat. |
| SP020 | Meta | V-JEPA: The next step toward advanced machine intelligence | V-JEPA is a joint-embedding predictive architecture for video. |
| SP021 | GitHub / FAIR | GitHub - facebookresearch/habitat-lab | A modular high-level library to train embodied AI agents across a variety of tasks and environments. |
| SP022 | ManiSkill | ManiSkill | ManiSkill. |
| SP023 | Hugging Face / Stability AI | stabilityai/stable-zero123 · Hugging Face | stabilityai/stable-zero123. |
| SP024 | Amazon Web Services | Racing Simulator Software - DeepRacer | Racing Simulator Software - DeepRacer. |
| SP025 | MIT Technology Review | World models | Today’s AI is still unreliable. |
| SI001 | General Intuition | General Intuition | The frontier lab for acting in space and time. | We have onboarded our first partners across games, simulation, and robotics to our commercial API. |
| SI002 | General Intuition | General Intuition & Medal Jobs | Financial Controller General Intuition & Medal • New York City • Full time • On-site $180K – $250K • Offers Equity. |
| SI003 | Ashby | General Intuition & Medal Jobs | Member of Technical Staff General Intuition & Medal • New York City; Geneva; London; Paris • Full time • On-site $250K – $450K • Offers Equity. |
| SI004 | TechCrunch | General Intuition in talks to raise $300M at around $2B valuation | General Intuition is in talks to raise $300 million at around a $2 billion valuation. |
| SI005 | TechCrunch | General Intuition's $2.3B bet that video games can train AI agents for the real world | The company says once it gets its API into more customers’ hands, it would be able to test its mettle with a variety of use cases. |
| SI006 | The Robot Report | General Intuition raises $320M to use video game data to train robots | General Intuition uses billions of gameplay clips uploaded to Medal to build AI models that can perceive, predict, and act in virtual and physical environments. |
| SI007 | DutchNews | Dutch AI firm General Intuition raises $320 million in new round - DutchNews.nl | The latest funding round, completed in January but only now made public, means the company is now valued at $2.3 billion. |
| SI008 | The SaaS News | General Intuition Raises $320M Series A | General Intuition has raised $320 million in Series A funding. |
| SI009 | Securities and Exchange Commission | SEC FORM D | Name of Issuer AVSF - General Intuition 2026, LLC ... Total Offering Amount $4,497,475 ... Total Amount Sold $4,497,475. |
| SI010 | Securities and Exchange Commission | EDGAR Entity Landing Page | EDGAR Entity Landing Page. |
| SI011 | Securities and Exchange Commission | SEC.gov | EDGAR Full Text Search | EDGAR Full Text Search. |
| SI012 | Kamer van Koophandel | Zoeken bij KVK | KVK | Zoeken bij KVK. |
| SI013 | CoreWeave | CoreWeave Cloud Pricing | CoreWeave | NVIDIA HGX H100 ... On-Demand Price: $49.24 / Hour. |
| SI014 | DigitalOcean / Paperspace | Pricing | DigitalOcean | Pricing | DigitalOcean. |
| SI015 | Amazon Web Services | Instance Types | Instance Types. |
| SI016 | Amazon Web Services | Amazon EC2 P5 Instances | P5 instances provide up to 8 NVIDIA H100 GPUs with a total of up to 640 GB HBM3 GPU memory per instance. |
| SI017 | Deloitte | AI trends : Adoption barriers and updated predictions | According to nearly 60% of AI leaders, the primary challenges in adopting agentic AI are integrating with legacy systems and addressing risk and compliance concerns. |
| SI018 | OpenAI | Video generation models as world simulators | Our results suggest that scaling video generation models is a promising path towards building general purpose simulators of the physical world. |
| SI019 | Backed VC | General Intuition | General Intuition. |
| SI020 | General Catalyst | Portfolio | General Catalyst | Portfolio | General Catalyst. |
| SI021 | MIT Technology Review | World models | Today’s AI is still unreliable. |
| SI022 | Kaiso Research / MarketResearch.com | Global AI World Models Market Size, Opportunity Analysis and Forecast | The Global AI World Models market was valued at USD 1.8 billion in 2025, and is projected to reach USD 52.7 billion by 2035. |
| SI023 | NVIDIA | NVIDIA Cosmos | Develop physical AI faster with leading world foundation models and open data processing, training, and evaluation frameworks. |
| SI024 | PitchBook | General Intuition 2026 Company Profile: Valuation, Funding & Investors | PitchBook | General Intuition 2026 Company Profile: Valuation, Funding & Investors. |
| SI025 | FormDs.com | AVSF - General Intuition 2026, LLC | Most recent fund raising on April 6, 2026 raised $4,497,475 in Equity. |
| SE001 | General Intuition | General Intuition | The frontier lab for acting in space and time. | We have onboarded our first partners across games, simulation, and robotics to our commercial API and will be selectively working with a few companies ahead of the broader release of our model. |
| SE002 | General Intuition | Privacy Notice - General Intuition | Where General Intuition processes Medal data for research and model development, that processing is governed by the Medal Privacy Policy and the arrangement between General Intuition and Medal described there. |
| SE003 | General Intuition | Terms of Use - General Intuition | The Site provides general information about General Intuition, our research, and open roles. Content is provided for informational purposes only and may change without notice. |
| SE004 | General Intuition | Partner with General Intuition | Tell us what becomes possible if we build it together. |
| SE005 | Medal | Record, Edit, and Share Your Game Clips & Gameplay - Medal | Medal works with every game, allowing you to capture clips from the most popular games to the smallest indies. |
| SE006 | MIRA | MIRA - Blog post | It's a 5B-parameter diffusion transformer paired with a 600M-param video representation codec. |
| SE007 | MIRA | MIRA technical report | We publicly release Rocket Science, a 4,000-hour slice of this data ... paired with the action streams and physics states: everything you need to train your own model. |
| SE008 | GitHub / mira-wm | GitHub - mira-wm/mira | MIRA is a real-time world model of Rocket League: a 5B parameters latent diffusion model ... a full 2v2 match can be played inside the model at 20 FPS on a single GPU. |
| SE009 | GitHub / Eloi Alonso | GitHub - eloialonso/iris | The world model is composed of a discrete autoencoder and an autoregressive Transformer. |
| SE010 | GitHub / Vincent Micheli | GitHub - vmicheli/delta-iris | Efficient World Models with Context-Aware Tokenization. ICML 2024 |
| SE011 | DIAMOND | Diffusion for World Modeling: Visual Details Matter in Atari (DIAMOND) | DIAMOND achieves a mean human normalized score of 1.46 on the competitive Atari 100k benchmark; a new best for agents trained entirely within a world model. |
| SE012 | Wayve | GAIA-2: Pushing the Boundaries of Video Generative Models for Safer Assisted and Automated Driving | GAIA-2 combines a latent diffusion architecture with extensive domain-specific conditioning to enable precise control over multi-camera video generation. |
| SE013 | Google DeepMind | Genie 2: A large-scale foundation world model | Today we introduce Genie 2, a foundation world model capable of generating an endless variety of action-controllable, playable 3D environments for training and evaluating embodied agents. |
| SE014 | Google DeepMind | A generalist AI agent for 3D virtual environments | The SIMA agent is designed to complete tasks in a range of 3D game worlds by following natural-language instructions. |
| SE015 | Physical Intelligence | Our First Generalist Policy | Over the past eight months, we’ve developed a general-purpose robot foundation model that we call π0. |
| SE016 | World Labs | World Labs | World Labs is a leading spatial intelligence company, building frontier models that can perceive, generate, reason, and interact with the 3D world. |
| SE017 | OpenAI | Video generation models as world simulators | Training videos on the internet, alongside effectively utilizing these models, can be a promising path towards building general purpose simulators of the physical world. |
| SE018 | NVIDIA | NVIDIA Cosmos | Develop physical AI faster with leading world foundation models and open data processing, training, and evaluation frameworks. |
| SE019 | Kyutai | kyutai: open-science AI lab | Kyutai is an open-science AI lab. |
| SE020 | TechCrunch | General Intuition's $2.3B bet that video games can train AI agents for the real world | Today, the startup has a handful of customers in gaming, simulation, and robotics. |
| SE021 | The Robot Report | General Intuition raises $320M to use video game data to train robots | Instead of gathering hundreds or thousands of hours of real-world data or generating simulated data, the company uses billions of gameplay clips uploaded to Medal. |
| SE022 | SiliconANGLE | Game-clip AI startup General Intuition in talks to raise $300M at $2B valuation | The startup combines world models with action models, systems that generate the next likely action taken by a player or agent. |
| SE023 | InvestGame | General Intuition: $320m Series A to Train AI Agents on Gameplay Data | Proceeds will fund compute scaling through a partnership with CoreWeave, with a portion earmarked to broaden commercial API access by the end of summer 2026. |
| SE024 | Coalition Capital | Coalition Capital Backs General Intuition's $320M Series A to Define the Next Frontier in AI | General Intuition has onboarded its first commercial partners across games, simulation, and robotics, and has launched Nerve, its own data collection platform. |
| SE025 | GamesBeat | General Intuition raises $320M at $2.3B valuation for AI frontier models based on gameplay | exclusive interview | The company already has a lot of player data uploaded from Medal, but it has also onboarded its first partners across games, simulation, and robotics to its commercial API and will be selectively working with a few companies ahead of the release of the model. |
| SU001 | General Intuition | General Intuition | The frontier lab for acting in space and time. | We have onboarded our first partners across games, simulation, and robotics to our commercial API and will be selectively working with a few companies ahead of the broader release of our model. |
| SU002 | General Intuition | Partner with General Intuition | Tell us what becomes possible if we build it together. |
| SU003 | GamesBeat | General Intuition raises $320M at $2.3B valuation for AI frontier models based on gameplay | exclusive interview | The company already has a lot of player data uploaded from Medal, but it has also onboarded its first partners across games, simulation, and robotics to its commercial API and will be selectively working with a few companies ahead of the release of the model. |
| SU004 | TechCrunch | General Intuition's $2.3B bet that video games can train AI agents for the real world | Today, the startup has a handful of customers in gaming, simulation, and robotics. |
| SU005 | Coalition Capital | Coalition Capital Backs General Intuition's $320M Series A to Define the Next Frontier in AI | General Intuition has onboarded its first commercial partners across games, simulation, and robotics, and has launched Nerve, its own data collection platform. |
| SU006 | InvestGame | General Intuition: $320m Series A to Train AI Agents on Gameplay Data | A commercial API for gaming, simulation, and robotics partners has launched, with broader access planned by the end of summer 2026. |
| SU007 | Tech Funding News | General Intuition bags $320M Series A at $2.3B to build the AI that learns to act from gamers | The pace of fundraising reflects something investors don’t often say out loud: they backed the research trajectory, not a commercial product. |
| SU008 | Axios | General Intuition raises $320 million to develop AI from gaming | GI's bet is that gaming — both gameplay video and the player inputs that produced it — can help build both world models and large action models faster and cheaper than by other training techniques. |
| SU009 | The Verge | Why world models are the next big thing in AI | It’s a pretty big bet. |
| SU010 | The SaaS News | General Intuition Raises $320M Series A | General Intuition plans to use the capital to scale its compute capacity, specifically through a deal with CoreWeave, and to fund further research, model development, and hiring for AI researchers and infrastructure engineers. |
| SU011 | The Robot Report | General Intuition raises $320M to use video game data to train robots | General Intuition also hopes to make its API more broadly available this summer, according to TechCrunch. |
| SU012 | Medal | Record, Edit, and Share Your Game Clips & Gameplay - Medal | Medal works with every game, allowing you to capture clips from the most popular games to the smallest indies. |
| SU013 | Medal | Medal Features - Record, Edit, and Share PC Games Instantly | Medal handles uploads for free. |
| SU014 | Medal Support | Medal TV Support | |
| SU015 | Medal | Terms of Service | Medal allows you to post content, including video (clips), comments ... |
| SU016 | Medal | Privacy Policy | |
| SU017 | General Intuition | Privacy Notice - General Intuition | Where General Intuition processes Medal data for research and model development, that processing is governed by the Medal Privacy Policy and the arrangement between General Intuition and Medal described there. |
| SU018 | General Intuition | Terms of Use - General Intuition | The Site provides general information about General Intuition, our research, and open roles. |
| SU019 | CoreWeave | The Essential Cloud for AI | CoreWeave | CoreWeave Cloud is an AI-native platform purpose-built for AI. |
| SU020 | Kyutai | kyutai: open-science AI lab | Kyutai is an open-science AI lab. |
| SU021 | Epic Games | Home - Epic Games | |
| SU022 | CNBC | General Intuition CEO Pim de Witte on training AI on gamers | |
| SU023 | TechCrunch | General Intuition in talks to raise $300M at around $2B valuation | |
| SU024 | DutchNews | Dutch AI firm General Intuition raises $320 million in new round | |
| SU025 | MIRA | MIRA - Blog post | We train a model to simulate Rocket League, Epic Games' car-football game. |
| SR001 | General Intuition | General Intuition | The frontier lab for acting in space and time. | We have onboarded our first partners across games, simulation, and robotics to our commercial API and will be selectively working with a few companies ahead of the broader release of our model. |
| SR002 | General Intuition | Privacy Notice - General Intuition | Where General Intuition processes Medal data for research and model development, that processing is governed by the Medal Privacy Policy and the arrangement between General Intuition and Medal described there. |
| SR003 | General Intuition | Terms of Use - General Intuition | The Site provides general information about General Intuition, our research, and open roles. |
| SR004 | General Intuition | Partner with General Intuition | Tell us what becomes possible if we build it together. |
| SR005 | GamesBeat | General Intuition raises $320M at $2.3B valuation for AI frontier models based on gameplay | exclusive interview | The company already has a lot of player data uploaded from Medal, but it has also onboarded its first partners across games, simulation, and robotics to its commercial API and will be selectively working with a few companies ahead of the release of the model. |
| SR006 | TechCrunch | General Intuition's $2.3B bet that video games can train AI agents for the real world | Today, the startup has a handful of customers in gaming, simulation, and robotics. |
| SR007 | InvestGame | General Intuition: $320m Series A to Train AI Agents on Gameplay Data | A commercial API for gaming, simulation, and robotics partners has launched, with broader access planned by the end of summer 2026. |
| SR008 | Tech Funding News | General Intuition bags $320M Series A at $2.3B to build the AI that learns to act from gamers | The pace of fundraising reflects something investors don’t often say out loud: they backed the research trajectory, not a commercial product. |
| SR009 | CNBC | General Intuition CEO Pim de Witte on training AI on gamers | |
| SR010 | The Robot Report | General Intuition raises $320M to use video game data to train robots | General Intuition also hopes to make its API more broadly available this summer, according to TechCrunch. |
| SR011 | Medal | Record, Edit, and Share Your Game Clips & Gameplay - Medal | Medal can video capture any PC game that you are running, and regularly adds support for new games so you can browse clips and organize by game. |
| SR012 | Medal | Medal Features - Record, Edit, and Share PC Games Instantly | Medal handles uploads for free. |
| SR013 | Medal Support | Medal TV Support | |
| SR014 | Medal | Terms of Service | You must ensure that your use of the Services is in accordance with applicable law and with any third party rights. |
| SR015 | CoreWeave | The Essential Cloud for AI | CoreWeave | CoreWeave Cloud is an AI-native platform purpose-built for AI. |
| SR016 | Google DeepMind | Genie 2: A large-scale foundation world model | Today we introduce Genie 2, a foundation world model capable of generating an endless variety of action-controllable, playable 3D environments for training and evaluating embodied agents. |
| SR017 | Physical Intelligence | Our First Generalist Policy | Over the past eight months, we’ve developed a general-purpose robot foundation model that we call π0. |
| SR018 | World Labs | World Labs | World Labs is a leading spatial intelligence company, building frontier models that can perceive, generate, reason, and interact with the 3D world. |
| SR019 | OpenAI | Video generation models as world simulators | Training videos on the internet, alongside effectively utilizing these models, can be a promising path towards building general purpose simulators of the physical world. |
| SR020 | NVIDIA | NVIDIA Cosmos | Develop physical AI faster with leading world foundation models and open data processing, training, and evaluation frameworks. |
| SR021 | Bureau of Industry and Security | Updates to Prior Controls on Advanced Semiconductors Provide Additional Safeguards and Guidance for Chip Manufacturers | These updates are necessary to maintain the effectiveness of these controls, close loopholes, and ensure they remain durable. |
| SR022 | Bureau of Industry and Security | Guidance Regarding Enforcement of License Requirements for Advanced Computing Items for Entities Headquartered in Country Group D:5 and Macau | A license is required to export advanced computing items to entities headquartered in Country Group D:5 or Macau, even if the entities themselves are located outside Country Group D:5 or Macau. |
| SR023 | Bureau of Industry and Security | Department of Commerce Announces Rescission of Biden Era Artificial Intelligence Diffusion Rule and Strengthens Chip-Related Export Controls | BIS plans to publish a regulation formalizing the rescission and will issue a replacement rule in the future. |
| SR024 | NIST | AI Risk Management Framework | The profile can help organizations identify unique risks posed by generative AI and proposes actions for generative AI risk management that best aligns with their goals and priorities. |
| SR025 | European Commission | The EU’s approach to artificial intelligence | The AI Act introduces a clear, easy-to-understand approach based on 4 different levels of risk. |
| SR026 | U.S. Copyright Office | Copyright and Artificial Intelligence | The Office is issuing a Report in several Parts analyzing the issues, including the use of copyrighted materials in AI training. |
| SR027 | Federal Trade Commission | Policy Statement on Biometric Information and Section 5 of the FTC Act | Policy Statement on Biometric Information and Section 5 of the FTC Act. |
| SR028 | Illinois General Assembly | Public Act 103-0769 | "Biometric identifier" means a retina or iris scan, fingerprint, voiceprint, or scan of hand or face geometry. |
| SR029 | MIRA | MIRA - Blog post | We train a model to simulate Rocket League, Epic Games' car-football game. |
| SR030 | TechCrunch | World Labs lands $1B, with $200M from Autodesk, to bring world models into 3D workflows | The startup has now raised more than $1 billion in total funding, including a new $200 million tranche led by Autodesk. |
| SV001 | TechCrunch | General Intuition's $2.3B bet that video games can train AI agents for the real world | Today, the startup has a handful of customers in gaming, simulation, and robotics. |
| SV002 | GamesBeat | General Intuition raises $320M at $2.3B valuation for AI frontier models based on gameplay | exclusive interview | The company already has a lot of player data uploaded from Medal, but it has also onboarded its first partners across games, simulation, and robotics to its commercial API and will be selectively working with a few companies ahead of the release of the model. |
| SV003 | InvestGame | General Intuition: $320m Series A to Train AI Agents on Gameplay Data | A commercial API for gaming, simulation, and robotics partners has launched, with broader access planned by the end of summer 2026. |
| SV004 | Tech Funding News | General Intuition bags $320M Series A at $2.3B to build the AI that learns to act from gamers | The pace of fundraising reflects something investors don’t often say out loud: they backed the research trajectory, not a commercial product. |
| SV005 | CNBC | General Intuition CEO Pim de Witte on training AI on gamers | |
| SV006 | The Robot Report | General Intuition raises $320M to use video game data to train robots | General Intuition also hopes to make its API more broadly available this summer, according to TechCrunch. |
| SV007 | General Intuition | General Intuition | The frontier lab for acting in space and time. | We have onboarded our first partners across games, simulation, and robotics to our commercial API and will be selectively working with a few companies ahead of the broader release of our model. |
| SV008 | General Intuition | Privacy Notice - General Intuition | Where General Intuition processes Medal data for research and model development, that processing is governed by the Medal Privacy Policy and the arrangement between General Intuition and Medal described there. |
| SV009 | General Intuition | Terms of Use - General Intuition | The Site provides general information about General Intuition, our research, and open roles. |
| SV010 | General Intuition | Partner with General Intuition | Tell us what becomes possible if we build it together. |
| SV011 | Medal | Record, Edit, and Share Your Game Clips & Gameplay - Medal | Medal can video capture any PC game that you are running, and regularly adds support for new games so you can browse clips and organize by game. |
| SV012 | CoreWeave | The Essential Cloud for AI | CoreWeave | CoreWeave Cloud is an AI-native platform purpose-built for AI. |
| SV013 | TechCrunch | World Labs lands $1B, with $200M from Autodesk, to bring world models into 3D workflows | World Labs, which emerged from stealth in 2024 with $230 million at a $1 billion valuation, declined to say whether the latest round boosted its valuation. However, reports suggested it was aiming to raise at a $5 billion valuation. |
| SV014 | Reuters / U.S. News | AI pioneer Fei-Fei Li's World Labs raises $1 billion in funding | Bloomberg News reported in January that the startup was in funding discussions at a valuation of about $5 billion. |
| SV015 | TechCrunch | Physical Intelligence is reportedly in talks to raise $1B, again | Physical Intelligence ... is in discussions to raise about $1 billion in new funding at a valuation exceeding $11 billion. The deal would effectively double the company's $5.6 billion valuation in just four months. |
| SV016 | Physical Intelligence | Our First Generalist Policy | Over the past eight months, we’ve developed a general-purpose robot foundation model that we call π0. |
| SV017 | Google DeepMind | Genie 2: A large-scale foundation world model | Today we introduce Genie 2, a foundation world model capable of generating an endless variety of action-controllable, playable 3D environments for training and evaluating embodied agents. |
| SV018 | OpenAI | Video generation models as world simulators | Training videos on the internet, alongside effectively utilizing these models, can be a promising path towards building general purpose simulators of the physical world. |
| SV019 | NVIDIA | NVIDIA Cosmos | Develop physical AI faster with leading world foundation models and open data processing, training, and evaluation frameworks. |
| SV020 | Stanford HAI | AI Index Report 2026 | U.S. private AI investment reached $285.9 billion in 2025, and billion-dollar funding events nearly doubled. |
| SV021 | Finerva | Robotics & AI 2026 Valuation Multiples | The median revenue multiple rose steadily from 2.5x in the first quarter to 3.4x by Q4 2025. |
| SV022 | CompaniesMarketCap | C3 AI (AI) - Market capitalization | As of July 2026 C3 AI has a market cap of $1.39 Billion USD. |
| SV023 | U.S. SEC | EDGAR Search Results for C3.ai 10-K filings | 10-K ... Acc-no: 0001801170-26-000057 ... Filing Date 2026-02-27. |
| SV024 | CompaniesMarketCap | Unity Software (U) - Market capitalization | As of July 2026 Unity Software has a market cap of $13.40 Billion USD. |
| SV025 | U.S. SEC | EDGAR Search Results for Unity Software 10-K filings | 10-K ... Acc-no: 0001810806-26-000011 ... Filing Date 2026-02-11. |
| SV026 | Bureau of Industry and Security | Updates to Prior Controls on Advanced Semiconductors Provide Additional Safeguards and Guidance for Chip Manufacturers | These updates are necessary to maintain the effectiveness of these controls, close loopholes, and ensure they remain durable. |
| SV027 | NIST | AI Risk Management Framework | The profile can help organizations identify unique risks posed by generative AI and proposes actions for generative AI risk management that best aligns with their goals and priorities. |
| SV028 | European Commission | The EU’s approach to artificial intelligence | The AI Act introduces a clear, easy-to-understand approach based on 4 different levels of risk. |
| SV029 | U.S. Copyright Office | Copyright and Artificial Intelligence | The Office is issuing a Report in several Parts analyzing the issues, including the use of copyrighted materials in AI training. |
| SV030 | Medal | Terms of Service | You must ensure that your use of the Services is in accordance with applicable law and with any third party rights. |