初创公司尽调
尽调报告 spatial intelligence / 3D generative AI / simulation infrastructure Series B 2026-06-18

World Labs

World Labs 尽职调查报告

World Labs 拥有 AI 领域最强创始团队之一,空间智能早期产品势头清晰;但公开记录仍缺少商业证据, 难以有把握地支撑数十亿美元入场价。

封面要素

成立时间 01
2023
累计融资 02
1230 USD M [CV003]
最新轮次 03
Series B [CV002]
Marble GA 05
2025-11-12 [CO027]
World API 发布 06
2026-01-21 [CO028]

公司概况

World Labs 是一家位于旧金山的空间智能创业公司,由 Fei-Fei Li、Ben Mildenhall、Christoph Lassner 和 Justin Johnson 于 2024 年创立。公司的商业化入口围绕 Marble 展开:这是一个多模态世界模型,可用文本、图像、视频和布局生成持久 3D 环境;World API 则面向程序化生成和下游集成。公开证据显示投资人信心极强——到 2026 年 2 月约已融资 $1.23 billion——但还不足以形成接近申报级别的收入、留存、毛利率或确认估值视图。

官网
www.worldlabs.ai
创始人
Fei-Fei Li, Ben Mildenhall, Christoph Lassner, Justin Johnson
创立地点
San Francisco, California, USA
总部
San Francisco, California, USA
产品
Marble 面向多模态 3D 世界生成,叠加 World API、Spark 流式传输 / 渲染,以及相邻的 Marble Labs 工作流,服务创意、设计和机器人用例
客户
需要可探索 3D 环境的创作者、AI-native 软件平台、建筑与设计团队,以及机器人或仿真开发者
商业模式
自助式 Marble 订阅、API 用量,以及企业或合作伙伴集成的组合;未来可能借助大型软件生态获得战略分发
阶段
Series B
融资情况
2024 年 9 月完成 $230 million 种子轮,随后在 2026 年 2 月完成 $1 billion 融资,累计融资约 $1.23 billion
[CO001, CO002, CO004, CO024, CO028]

执行摘要

主要优势

  • 创始团队覆盖 ImageNet、NeRF、前沿计算机视觉和图形学研究,让公司在空间 AI 上具备稀缺的技术公信力。
  • Marble 和 World API 已跳出研究叙事,成为公开可用的产品入口,并在创作者、设计流程和机器人仿真中出现具名集成。
  • Autodesk、NVIDIA、AMD、a16z 等顶级投资方的战略支持,提高了资本获取能力,也留下潜在分发选项。

主要风险

  • 公开资料没有收入、ARR、留存、单位经济或已确认估值披露,因此 2026 年 2 月隐含价格仍是愿景倍数,不是经营倍数。
  • 商业验证仍早:许多公开引用是展示、合作或偏研究的集成,不是已披露合同金额的规模化企业部署。
  • 大型科技公司和开源阵营在世界模型、3D 生成和物理 AI 上竞争,可能在 World Labs 建立稳固 GTM 优势前压缩定价或收窄差异化。

未决问题

  • Marble 订阅和 World API 的收入、ARR、毛利率与烧钱速度数据,需要经审计或由管理层确认
  • Autodesk 的 $200 million 战略投资若附带商业条款和分发承诺,需披露具体内容
  • 2026 年 2 月融资的已确认投后估值、优先股堆叠和稀释影响
  • 具名生产客户、合同范围、续约证据,以及试点活动和规模化部署的区分

目录

Chapter 01

01公司概览

1.1 身份、产品与商业模式

World Labs(法律实体:World Labs Technologies, Inc.)是一家位于加州旧金山的前沿 AI 研究与产品公司。公司由四位在学术计算机视觉和 3D 图形学领域积累深厚的联合创始人于 2023 年创立,围绕一个科学命题组织起来:空间智能——AI 系统感知、生成、推理并与三维世界交互的能力——将成为大语言模型革命之后,生成式 AI 的下一个主要前沿。公司把自身使命描述为打造「Large World Models」(LWMs),即一类能理解并生成持久、可导航 3D 世界的基础模型,而不是生成扁平图像或文本。 截至 2026 年 6 月,World Labs 的首个、也是主要商业产品是 Marble。Marble 是一个多模态世界模型,可从单张图像、视频片段、360 度全景图、粗略 3D 布局或文本提示生成空间一致、高保真且持久的 3D 环境。Marble 先以限量 beta 发布,并于 2025 年 11 月 12 日在扩展功能集后正式 GA。输出包括 Gaussian splats、多边形网格和视频导出,可接入下游创意、设计和仿真流水线。2026 年 1 月 21 日,公司发布 World API,把 Marble 的世界生成能力以程序化接口开放给开发者,用于嵌入第三方应用、交互系统和仿真工作流。 商业模式以订阅为主,分为四档:Free(有限生成)、Standard($20/month)、Pro($35/month,含商业使用权)和 Max($95/month,开放全部功能)。这个价格带把产品放在 prosumer 创意工具与专业 3D 设计软件之间。公司瞄准创作者、VFX 与游戏工作室、机器人开发者和企业设计工作流。截至 2026 年 6 月,公司处在 late-seed/early-growth 阶段:商业产品已上线、API 已发布、与 Autodesk 的企业合作在推进、累计融资 $1.23 billion,但尚未披露收入、ARR 或客户数。 [CO001, CO003, CO004, CO005, CO027, CO028]

World Labs 快照 KPI 表
指标数值 / 状态日期置信度缺口 / 备注
累计融资额$1.23BFeb 2026由两轮已披露融资计算;无备案确认
种子轮规模$230MSep 2024Reuters 和公司已确认
Series B 规模$1BFeb 2026Reuters 和公司博客已确认
报道估值(未确认)~$5BJan 2026Bloomberg 报道;公司未确认
收入 / ARR未披露Jun 2026N/A缺口——公司尚未披露任何收入指标
员工人数估计 11–50Jun 2026TheOrg 代理值;Reuters 报道 2024 年 9 月退出隐身时为 20 人
客户数量未披露Jun 2026N/A缺口——公司尚未披露
Marble 全面开放日期2025-11-12Nov 2025隐私政策日期与产品发布日期一致
World API 发布日期2026-01-21Jan 2026公司博客公告已确认

估值来自 Bloomberg 报道(2026 年 1 月),公司在 2026 年 2 月融资完成后未确认。累计融资额由 $230M 种子轮 + $1B Series B 计算;没有独立备案确认。员工人数来自 TheOrg(11–50 代理值)和 Reuters(种子轮退出隐身时 20 人);当前数字未验证。收入、ARR 和客户数量都是证据缺口。

[CO016, CO018, CO019, CO022, CO024, CO026]
FO002: World Labs 公司快照逻辑

World Labs 的创始团队、使命、产品、资本和关键合作如何串成当前运营模型。

[CO002, CO004, CO005, CO019, CO032]

1.2 领导层与创始人

World Labs 由四位研究者共同创立,他们的合计成果覆盖了过去十年计算机视觉、神经渲染和 3D 图形学中最有影响力的一批进展。 Fei-Fei Li 是 CEO 兼联合创始人。她担任 Stanford University 计算机科学 Sequoia Professor,也是 Stanford Human-Centered AI Institute(HAI)创始联席主任。2017 至 2018 年,她曾任 Google Cloud AI/ML 副总裁兼首席科学家。Li 是 ImageNet 的设计者;这个大规模视觉数据集催化了计算机视觉的深度学习革命,并被广泛视为现代 AI 的基础要素之一。她入选 Time Magazine「100 Most Influential People in AI」,并获任 UN Secretary-General's Scientific Advisory Board。Li 同时保留 Stanford 的 Sequoia Professorship,形成双重角色安排;她称自己会在创办公司时继续「some of her work at Stanford」,这带来实质性的关键人物集中风险。 Ben Mildenhall 是联合创始人,也是 Neural Radiance Fields(NeRF,arXiv:2003.08934)的主要发明者。NeRF 改写了从多视角图像重建 3D 场景的路径,成为 2020 年代引用最高的 AI 论文之一。NeRF 是 Marble 世界生成能力底层渲染流水线的基础。 Christoph Lassner 是联合创始人,曾在 Meta Reality Labs Research 和 Epic Games 领导研究团队。他在 Tübingen 的 Max Planck Institute for Intelligent Systems 完成博士学位。他是 Pulsar(arXiv:2004.07484)的作者;Pulsar 是一个可微分的基于球体的渲染器,现已作为 PyTorch3D 的球体后端集成。他还在人类姿态估计和神经场景渲染领域发表了大量论文。 Justin Johnson 是联合创始人,现任 University of Michigan(EECS)助理教授,此前曾在 Stanford AI Lab 与 Li 合作。董事会构成、投资人董事席位和治理结构未公开披露,构成治理尽调的实质缺口。 [CO002, CO006, CO007, CO008, CO009, CO010]

领导层与创始人表
姓名角色背景创始人-市场契合关键人物风险
Fei-Fei LiCEO 兼联合创始人Stanford Sequoia Professor;曾任 Google Cloud VP/Chief AI Scientist;ImageNet 创建者;Stanford HAI 创始联合主任现代计算机视觉架构者;25 年以上空间 AI 研究;全球 AI 政策影响力关键——公众面孔、投资者关系,同时担任 Stanford Professorship
Ben Mildenhall联合创始人NeRF 发明者(arXiv:2003.08934);在 3D 场景合成和神经渲染上积累深厚NeRF 是 Marble 渲染栈的基础;2020 年代引用最高的 3D AI 技术高——核心渲染 IP 体现在创始团队成员身上
Christoph Lassner联合创始人曾任 Meta Reality Labs 和 Epic Games 研究负责人;Pulsar renderer 作者(PyTorch3D);Max Planck Institute Tübingen 博士跨游戏和 AR/VR 的神经场景渲染与生产管线经验中——深厚渲染器和工程专长
Justin Johnson联合创始人University of Michigan EECS 助理教授;曾任 Stanford AI Lab 研究员;发表计算机视觉和图形学论文计算机视觉、AI 管线集成和图形系统背景中——偏学术;日常运营暴露较低
董事会 / 治理(未公开披露)任何公开备案或公告均未确认投资者董事席位和独立董事未知——治理结构不透明重大缺口——无法验证独立监督

董事席位和治理结构未公开披露。联合创始人背景来自个人网站、Stanford 教职资料和 Reuters 报道。Justin Johnson 当前在 Michigan 的任职来自其 Stanford 跳转页面。

[CO002, CO006, CO007, CO009, CO010, CO011]

1.3 融资历史与投资人基础

截至 2026 年 2 月,World Labs 在不到 18 个月内完成两轮融资,合计融资 $1.23 billion。 第一轮为 $230 million,于 2024 年 9 月 13 日宣布,时间与公司结束隐身状态同步;该轮由 Andreessen Horowitz(a16z)、New Enterprise Associates(NEA)和 Radical Ventures 共同领投。参投方包括 AMD Ventures、Intel Capital 和 NVIDIA NVentures。World Labs 在公告时拒绝确认估值;Bloomberg 2026 年 1 月的报道援引消息人士称,公司种子轮估值为 $1 billion。 第二轮为 $1 billion,于 2026 年 2 月 18 日宣布,由 Autodesk 以 $200 million 承诺出资锚定。Autodesk 同时担任正式战略顾问,并与 World Labs 签署研究和产品层面的合作协议,把世界模型能力接入 3D 设计工作流,先从媒体与娱乐用例切入;协议明确排除了数据共享。该轮其他投资人包括 AMD、NVIDIA、Emerson Collective、Fidelity Management and Research Company、Sea Group,以及再次参投的 Andreessen Horowitz。种子轮投资人 Intel Capital 未出现在 Series B 投资人名单中。 Bloomberg 2026 年 1 月报道称,World Labs 正在讨论以 $5 billion 估值融资,意味着较种子轮提升五倍。公司未反驳该报道,但融资关闭后也未公开确认该数字,因此估值仍未验证。股权表构成具有战略上的不寻常之处:Autodesk(设计软件)、NVIDIA 和 AMD(AI 计算硬件)、Sea Group(东南亚数字商业)与一线财务投资人同席,显示公司可能借 Autodesk 既有用户基础进入企业分发,并在推理规模上做深硬件集成。公司未披露收入、ARR 或任何商业牵引指标。 [CO016, CO017, CO018, CO019, CO020, CO021]

利益相关方或投资者图谱
利益相关方轮次已确认金额 / 角色战略理由尽调要求
Andreessen Horowitz(a16z,投资方)种子轮(共同领投)+ Series B(参投)共同领投种子轮;回投 Series B;金额未披露一线 AI 聚焦 VC;可能有董事会代表确认董事席位;核实持续治理权利
New Enterprise Associates(NEA,投资方)种子轮(共同领投)共同领投种子轮;金额未披露一线企业科技 VC确认当前持股,以及 Series B 参与状态
Radical Ventures种子轮(共同领投)共同领投种子轮;金额未披露AI 专业 VC;与 Li 网络连接强确认当前持股;评估是否在 Series B 行使按比例跟投权
AutodeskSeries B(锚定)已确认 $200M;顾问角色3D CAD/设计软件;Marble 面向建筑师、工程师、电影工作室的企业分发渠道界定研究合作范围;确认 IP 和收入分成条款
NVIDIA NVentures / NVIDIA种子轮 + Series B两轮均参投;金额未披露算力硬件伙伴;世界模型的推理规模需求驱动方评估优先硬件访问或排他性条款
AMD Ventures / AMD种子轮 + Series B两轮均参投;金额未披露竞争性芯片厂商双重投资,传递硬件多元化策略信号评估任何硬件偏好或授权条款
Fidelity Management & ResearchSeries B参投;金额未披露后期财务投资人;提供流动性信号和公开市场验证了解退出期限预期和二级市场看法
Emerson CollectiveSeries B参投;金额未披露使命驱动投资者(教育、社会影响、新闻)评估空间 AI 论点与 Emerson 社会影响使命的契合
Sea GroupSeries B参投;金额未披露东南亚数字商业和游戏集团;游戏 / 元宇宙方向的潜在东南亚分发澄清授予的任何地域或产品线权利
Intel Capital仅种子轮种子轮参投;Series B 披露中缺席芯片厂商 VC;缺席 Series B 是值得注意的退出调查退出原因;评估是否暗示估值担忧或战略退出

持股比例未披露;Autodesk 的 $200M 是唯一已确认投资金额。Intel Capital 参与种子轮,但未出现在已审阅的任何 Series B 投资者披露中。所有治理权、反稀释条款和信息权均为私有信息。

[CO016, CO017, CO019, CO020, CO021, CO025]
FO003: World Labs 快照 KPI

截至 2026 年 6 月,World Labs 的关键财务和运营指标,并标注置信度和数据缺口。

[CO016, CO018, CO019, CO022, CO024, CO026]

1.4 里程碑、规模与负面信号

从创立到 2026 年 6 月,World Labs 的运营历史约 30 个月。公司以压缩节奏推进:创立、隐身运营、首个产品发布、API 可用,再到一轮重大战略融资。 公司成立于 2023 年;确切注册日期未公开披露,但 Fei-Fei Li 在 2025 年 11 月写道,联合创始人「more than one year ago」创立了 World Labs,因此创立时间不晚于 2023 年末。团队隐身运营到 2024 年中,2024 年 9 月结束隐身时约有 20 名员工。Marble 限量 beta 与「Generating Worlds」研究预览在 2024 年末同步推出;2025 年 11 月 12 日,Marble 正式 GA,距种子轮公告约 14 个月。World API 于 2026 年 1 月 21 日发布,$1 billion Series B 于 2026 年 2 月 18 日完成。公司在 2026 年 6 月 3 日发布了世界模型的功能分类法。 合作与集成信号:Autodesk 是主要公开披露的企业合作伙伴。公司发布了与 OpenArt、Lightcraft、Magnific、Rosebud AI、Viverse 和 Escape 等创意工作流集成方的案例研究,覆盖游戏、VFX 和电影领域,但这些案例未被披露为付费商业合同。 负面与缺口信号:Fei-Fei Li 同时在公司和 Stanford 任职,使 World Labs 面临显著关键人物集中风险。董事会构成、独立董事姓名和治理文件均未公开。公司已融资 $1.23 billion,却未披露任何收入,把投资人的判断完全压在技术命题和团队上,而不是商业牵引上。 EU AI Act(2024 年 3 月通过,2024 年 8 月起分阶段适用)为欧盟市场的通用 AI 模型开发者设定合规义务,包括透明度和版权尽调要求;这些要求与 World Labs 的训练数据实践和 API 输出相关。美国 Bureau of Industry and Security 的出口管制法规(EAR,15 CFR Parts 730-774)也可能适用于 World Labs 具备双用途潜力的 AI 技术。World Labs 的 Terms of Service 和 Acceptable Use Policy 明确限制涉及生物武器开发、欺骗性 deepfakes 和干扰关键基础设施的用途,显示公司已意识到这些监管向量。 [CO033, CO035, CO036, CO037, CO038, CO039]

里程碑表
日期事件类型金额 / 估值 / 状态参与方含义
2023(估计)World Labs 在 San Francisco 创立创立N/AFei-Fei Li、Justin Johnson、Christoph Lassner、Ben Mildenhall 四位联合创始人四位创始人团队拥有深厚 CV/3D 研究背景,围绕空间智能论点组建公司
Pre-Sep 2024隐身运营;组建团队扩张N/A退出隐身时约 20 名员工(Reuters)公开披露前隐身运营约 12 个月
2024-09-13宣布种子轮;退出隐身融资融资 $230M;估值约 $1B(Bloomberg)a16z、NEA、Radical Ventures、AMD、Intel Capital、NVIDIA NVentures 等投资方2024 年已披露 AI 种子轮中最大;产品前融资凸显团队驱动论点
2024-12-02发布 Generating Worlds 研究预览产品N/AWorld Labs 团队首次公开演示可持久、可导航 3D 世界生成能力
2025-10-16RTFM 实时帧模型研究预览产品N/AWorld Labs 团队实时生成式世界模型预览显示研究到产品管线活跃
2025-11-12Marble 全面开放产品Free / $20 / $35 / $95 每月档位World Labs(公开发布)种子轮后 14 个月交付首个商业产品;定价公开披露
2026-01-05Fei-Fei Li 在 CES 2026 发表主题演讲(AMD 舞台)扩张N/AFei-Fei Li, AMD重大曝光事件;Reuters CES 图片说明中将 Li 描述为 CEO
2026-01-21World API 发布产品N/AWorld Labs(开发者发布)程序化世界生成打开直接 Marble 用户之外的开发者和 ISV 市场
2026-01(Bloomberg)报道称 Series B 融资估值约 $5B融资报道估值约 $5B(未确认)Bloomberg 消息源较种子轮隐含估值约 15 个月上涨 5 倍;公司未反驳
2026-02-18$1B Series B 完成融资融资 $1B;估值未确认Autodesk($200M)、AMD、NVIDIA、Fidelity、Emerson Collective、Sea、a16z 等投资方迄今最大空间 AI 融资;Autodesk 锚定投资传递企业分发路径信号
2026-06-03发布世界模型功能分类产品N/AWorld Labs 团队(Fei-Fei Li 及团队)将公司定位为空间 AI 类别定义的思想领导者

创立日期根据 Fei-Fei Li 2025 年 11 月 Substack 中「more than one year ago」的说法估算;确切注册日期未公开确认。RTFM 日期来自公司博客索引。Bloomberg 2026 年 1 月估值来自 TechCrunch 和 StartupHub 二级引用;公司未确认。隐身期 20 人来自 Reuters 2024 年 9 月报道。

[CO001, CO002, CO012, CO016, CO018, CO019]
FO001: World Labs 公司里程碑时间线

World Labs 从创立到 2026 年 6 月的关键事件,覆盖融资轮次、产品发布和战略合作。

Fei-Fei Li 在 2025 年 11 月的表述显示,创立时间估计为 2023 年初;确切月份未确认。

[CO001, CO003, CO016, CO019, CO027, CO028]

1.5 图表

Chapter 02

02市场分析

2.1 市场边界与结构定义

World Labs 以一个称为「空间智能」的概念来定义市场:AI 系统感知、生成、推理并与三维环境交互的能力。这个框架把公司的产品——Marble(生成式世界模型)和 World API(程序化接口)——定位为基础设施,而不是单纯的 3D 美术工具或仿真平台;它让任何下游系统能够理解并生成可导航的空间世界。 该市场的核心纳入支出包括能够生成 3D 环境或基于 3D 环境推理的 AI-native 服务:用于程序化世界创建的 API 调用、云端世界生成平台的订阅访问,以及把空间智能嵌入其他应用的集成工具。排除支出包括 Maya、Blender 或 Houdini 等不使用 AI 生成的传统 3D digital content creation(DCC)工具;传统 CAD/BIM 许可(AutoCAD、Revit),除非嵌入生成式 AI 能力;以及 MuJoCo 或 NVIDIA Isaac Sim 等在没有 AI 生成内容输入时单独使用的物理引擎。边界划在 AI 生成式输出这一点:如果工具靠参数化或艺术家驱动流水线创建 3D 内容,它就是替代品,而不是同一市场的一部分。 与其显著重叠的相邻市场包括机器人仿真基础设施(训练规模下对 sim-ready 3D 环境的需求)、数字孪生平台(AI 生成的世界表征进入镜像与监控工作流)、沉浸式设计可视化(空间 AI 替代静态渲染),以及 AI-native 应用的开发者基础设施(World API 与云 AI API 竞争或互补)。这些相邻市场都可能成为 World Labs 扩大 SAM 的渠道,但也各自带有不同的 incumbent 平台和买方工作流。 各细分市场的现状替代品都很成熟:机器人研究者使用程序化生成器(Infinigen、Isaac Sim 的资产库)、仿真框架(MuJoCo、RoboSuite)和定制环境创作;建筑师使用静态渲染和实时游戏引擎;电影制作人使用实体布景和虚拟制片棚;开发者使用独立 3D 资产库或手工流水线。这些替代品在很多情况下免费或低成本(MuJoCo 开源;Infinigen 以 BSD-3 开源),这给 World Labs 要证明的价值设下很高门槛,才能赢得预算。 World Labs 自己在 2026 年 6 月的分类法文章中指出,「world model」是 AI 中最被过度使用的术语之一,同时覆盖计算机视觉、机器人、游戏研究和神经科学用例。语义范围过宽带来市场分层难题:各垂直领域买家对世界模型能提供什么有不同预期,使现阶段的一致定位和线索筛选变得困难。 [CM001, CM002, CM003, CM004, CM019, CM020]

市场定义表
类别纳入支出排除支出买方 / 付款方与 World Labs 的相关性
空间智能 / 3D 世界生成AI 原生云 API,生成可导航 3D 环境;世界生成平台订阅;SDK / 集成工具2D 图像生成工具(Midjourney、DALL-E);仅文本 LLM;没有 3D 空间输出的仅视频生成开发者、创意工作室、研究实验室 / 产品或 R&D 预算核心市场;World Labs 的直接产品
机器人与具身 AI 仿真可用于仿真的 3D 环境创建服务;域随机化数据管线;机器人训练用合成场景生成不含 AI 内容的物理引擎许可(MuJoCo、Isaac Sim);URDF/CAD 资产库机器人研究者、具身 AI 实验室、硬件 OEM / R&D 和工程预算高相关;驱动近期 API 需求的瓶颈市场
建筑与设计可视化AI 驱动的概念到 3D 空间可视化;沉浸式设计评审工具;建筑世界生成 API静态建筑渲染(V-Ray、Lumion);不含 AI 生成的 BIM 平台(Revit、AutoCAD);仅用于可视化的实时游戏引擎建筑事务所、室内设计师、AEC 软件供应商 / 设计工具和工作室预算高相关;Autodesk 合作传递战略扩张信号
沉浸式媒体与创意制作AI 虚拟布景和外景工具;影视用空间一致 3D 环境生成;营销和内容用交互式 3D实体制作布景;虚拟制作摄影棚(LED 墙);没有 AI 世界生成的传统 VFX 管线独立电影人、VFX 工作室、内容创作者 / 工作室和创作者订阅预算当前早期采用者细分;订阅驱动
开发者 / API 基础设施嵌入第三方应用的程序化世界生成 API 访问;空间 AI 开发者工具包;平台集成通用 3D 资产数据库;非生成式场景图框架;自托管物理仿真基础设施软件开发者、平台公司、SaaS 构建者 / 工程和产品预算新兴;长期潜力最高;World API 于 2026 年 1 月发布

市场边界定义基于 World Labs 产品定位(公司来源)和竞品评审;支出估计为定性判断,未量化。

[CM001, CM002, CM004, CM019, CM020, CM022]

2.2 受证据约束的市场规模测算

本章审阅的分析师报告中,没有任何一份把「空间智能」或「3D generative AI」作为独立市场品类切分。规模测算只能从三个相邻视角拼出:(1)广义生成式 AI 市场,作为上限;(2)数字孪生和仿真基础设施市场,作为空间 AI 仿真细分的功能代理;(3)基于垂直买方渗透分析的自下而上估算。 Global Market Insights(2026 年 1 月)估算,全球生成式 AI 市场 2025 年为 USD 53.7 billion,2026 年升至 USD 83.3 billion,2035 年预测为 USD 988.4 billion,CAGR 31.6%。该数字覆盖所有生成式 AI 产品——文本、图像、音频、视频和多模态——范围太宽,无法作为 World Labs 的实用可服务市场。但它确认了投资人与买方语境:所有生成式 AI 品类都在快速部署并增加支出。 MarketsandMarkets(2024 年,更新至 2025 年中数据)预测,全球数字孪生市场将从 2025 年的 USD 21.14 billion 增长至 2030 年的 USD 149.81 billion,CAGR 47.9%。数字孪生市场可能是 World Labs 基础设施定位最接近的功能类比,因为数字孪生需要与 Marble 提供的同类空间表征和仿真能力。完整数字孪生市场包含硬件、IoT 集成和行业特定应用,并非全部可由 World Labs 触达;但即便 10–15% 的软件 / AI 子细分,也意味着到 2030 年存在数十亿美元机会。 自下而上的买方渗透视角给出的估计更粗:文献估计,机器人和具身 AI 训练数据细分的全球可服务仿真基础设施支出约 $0.5B–$2B;建筑与设计可视化 AI 细分约 $0.3B–$1B;开发者 API 和创意细分仍处早期,没有可靠下限或上限。这些自下而上区间高度不确定——它们取决于机器人 R&D 预算中有多大份额分配给仿真环境创建,而不是其他研究成本——但它们确认 World Labs 现实的近期 SAM 可能低于 $5B。 本节的区间图(FM002)展示了数字孪生市场(目前最佳可用代理)在不同分析师来源和情景下的估算分布。相互矛盾的估算被保留,而不是平均;不确定性本身就是重要分析结论。金字塔图(FM001)展示了市场可触达性的三层 TAM/SAM/SOM 结构。 [CM005, CM006, CM007, CM029, CM030, CM031]

TAM/SAM/SOM 或规模测算镜头表
发布方年份地域市场 / 细分数值(USD)CAGR方法论置信度用于 World Labs 规模测算的局限
Global Market Insights(GMI,市场研究机构)2026全球生成式 AI(所有模态)$83.3B (2026); $988.4B (2035)31.6% (2026–2035)自上而下分析师调查;一手 / 二手研究低(过宽)覆盖包括文本、图像、音频、视频在内的所有生成式 AI;未拆出 3D / 空间份额
MarketsandMarkets2024–2025全球数字孪生(所有行业)$21.14B (2025); $149.81B (2030)47.9% (2025–2030)自上而下一手研究 + 二手研究;供应商调查中(更接近的代理)包括硬件、IoT、行业专用平台;未拆分软件 / AI 子细分
MarketsandMarkets2025全球医疗数字孪生子细分最高 CAGR 细分(52.7%)52.7%数字孪生报告中的细分拆解医疗 DT 是小众领域;与 World Labs 当前产品不直接相关
arXiv survey(Ye 等,学术综述)2026全球具身 AI 的 3D 生成(学术)未引用市场规模;定性记录需求N/A学术文献综述;不是市场规模测算低(定性)确认需求信号,但不提供支出数据
自下而上估计(推断)2026全球机器人 / 具身 AI 仿真环境 SAM$0.5B–$2.0B(估计)估计 ~30–40%根据机器人市场增长率和仿真基础设施份额假设推断低(推断)高度依赖假设;没有公开仿真基础设施支出拆分
自下而上估算(推断)2026全球建筑 / 设计 AI 可视化 SAM$0.3B–$1.0B(估算)约 25–35%(估算)由 AEC 软件市场和 AI 采用率推断低(推断)AEC 中空间 AI 渗透率没有公开数据;以整体 AEC SaaS 增长做代理

所有分析师数据均来自公开可访问的报告落地页;完整方法和模型细节位于付费墙之后。自下而上行是编辑部推断,不是有来源的估算。GMI 和 MarketsandMarkets 的数据应按自上而下生成式 AI 市场报告的通常做法,视为存在 ±50% 不确定性。

[CM005, CM006, CM017, CM030, CM032]
FM001: 市场规模视角

World Labs 空间智能市场的三层 TAM / SAM / SOM 金字塔,采用相邻母市场和估算的垂直渗透率。

SAM 和 SOM 数值来自自下而上的垂直估算,并非基于公司披露收入或独立分析师数据;所有层级的不确定性至少为 ±50%。

[CM036, CM045]
FM002: 市场估算区间

2025 至 2030 年数字孪生市场(空间 AI 基础设施的最佳可得代理)的分析师估算区间,单位为十亿美元。

区间基于单一主来源(数字孪生使用 MarketsandMarkets;生成式 AI 使用 GMI),并由编辑加上对称不确定性带;没有可用于直接对比的第二个独立预测。

[CM005, CM006, CM037, CM041]

2.3 买方、用户与付款方分层

World Labs 的市场不是单一整体:四个已识别买方细分各自拥有不同买方画像、预算所有者、采用触发点和销售动作。要判断 World Labs 当前产品(订阅 SaaS + API)能否同时服务四类客户,还是需要按细分市场分别设计 go-to-market,必须先理解这种异质性。 媒体与创意制作细分包括独立电影人、数字艺术家、VFX 工作室和 AI-native 创意团队。这些买家主要以个人订阅或小型工作室订阅方式使用 Marble;相关价格点是 Pro 档($35/month)。创意电影案例研究记录的用例,是把 Marble 当作持久虚拟外景地,省去每个镜头重建场景,并在 AI 生成视频镜头之间保持空间一致性。采用触发点是生成式视频的一致性问题——这是一个被广泛报道的限制,而世界模型提供的空间锚定正好直接解决它。 机器人与具身 AI 细分包括大学研究实验室、机器人创业公司和硬件 OEM 的 R&D 团队。预算通常来自 R&D grant、风险投资或内部实验室分配。采用触发点是数据瓶颈:World Labs 机器人案例研究明确把「high-quality, diverse simulation data」称为机器人研究中「one of the biggest limiting factors」,并展示 Marble 生成数千个物理准确场景,可导出为兼容 MuJoCo 和 Isaac Sim 的环境。DeepMind 的 Genie 2 和 SIMA,以及 NVIDIA 的 Cosmos 平台,都验证了市场对可扩展、多样化 3D 训练环境的需求,尽管它们覆盖具身 AI 工作流的环节略有不同。该细分的自下而上 SAM 估计为 $0.5B–$2B,但证据稀疏。 建筑与设计(AEC)细分使用 Marble 做从概念到 3D 的沉浸式可视化——把草图和渲染阶段压缩成一个可探索的空间世界。早期合作伙伴 Fenestra 和 Interior AI 为该工作流提供客户证明。相关预算所有者是设计公司负责人,或通过 API 集成 Marble 的工具供应商;Autodesk 合作($200M 投资)显示该垂直领域的战略重要性。付款方通常是 B2B SaaS 席位预算或 API 集成预算。 开发者与 API 细分最广,但渗透最低:任何需要按需生成 3D 环境的软件产品都可以集成 World API。采用触发点是可编程世界生成端点的技术可用性,类似 OpenAI 的 API 如何把 LLM 用例扩展到数千个应用。该细分长期潜力最高,但需要开发者生态、文档成熟度,以及 World Labs 仍在建设中的延迟 / 成本表现。 Luma AI 的交互式 3D 场景(可嵌入、通用可分享,3D 文件 8–20 MB、30 FPS)显示,只要 web 交付性能足够好,开发者 / 创意细分会响应空间 3D 产品;这也为 World Labs 的 Spark 2.0 流式平台设定了基准。 [CM008, CM009, CM010, CM011, CM012, CM013]

细分市场 / 买方地图
细分市场主要买方主要用户付款方核心工作流预算归属采用触发点
媒体与创意制作创意总监、独立电影人、AI 内容工作室电影人、3D 艺术家、AI 视频制作者个人订阅或工作室 / 代理商预算生成可持续复用的虚拟布景和地点,让 AI 视频不同镜头保持空间一致;导出帧作为合成底板个人创作者或制作总监生成式视频工具跨镜头不一致;需要稳定的 3D 参考环境
机器人与具身 AI机器人研究负责人、ML 工程师、具身 AI 实验室负责人机器人专家、仿真工程师、具身 AI 研究人员R&D 实验室预算、研究资助、VC 支持创业公司的工程预算为域随机化生成数千个多样且物理准确的 3D 场景;导出碰撞网格,接入 MuJoCo/Isaac Sim研究总监、实验室 PI 或工程负责人仿真数据短缺卡住大规模机器人训练
建筑与设计(AEC)建筑事务所、室内设计工作室、AEC 软件平台厂商建筑师、室内设计师、设计工具用户设计公司或 SaaS 集成预算;平台厂商(如 Autodesk)的 API 授权将概念图或文本提示转成沉浸式 3D 漫游,用于客户评审和迭代设计工作室负责人、平台产品负责人(如 Fenestra、Interior AI)客户需要沉浸式设计评审;Autodesk 合作扩大了管线
开发者 / API软件开发者、平台公司、创业者应用开发者、集成工程师、产品经理工程或产品预算(API 调用成本)用 World API 端点把世界生成能力以编程方式嵌入应用、游戏或仿真CTO 或工程团队生产级世界生成 API 可用;产品需要 AI 原生 3D 功能

细分地图基于 World Labs 案例研究、产品定位和 Autodesk 合作公告;企业客户数量和按细分市场划分的收入占比未披露。

[CM008, CM009, CM010, CM011, CM024, CM028]
FM003: 买方 / 细分地图

横跨 World Labs 四个市场细分的买方、用户、付款方矩阵,展示采用准备度和预算入口。

采用准备度是定性的编辑判断;World Labs 或独立分析师没有提供客观渗透率数据。

[CM009, CM010, CM011, CM023, CM024, CM038]

2.4 增长驱动因素与采用约束

2026 年,几股结构性力量正在加速空间智能基础设施需求。最强驱动是机器人与具身 AI 的训练数据瓶颈。领先实验室已经开始从手工创作环境转向用于仿真的生成式世界模型;限制因素不再是 AI 生成环境是否有用,而是它们是否足够物理准确,能迁移到真实机器人。World Labs 必须证明 Marble 生成的场景不只是视觉上可信,还能在物理层面 ready for simulation。2026 年一篇关于具身 AI 3D 生成的综述(arXiv:2604.26509)确认了这种需求迁移,同时也记录了技术缺口。 第二个驱动是史无前例的 AI 投资潮。2026 年 Q1,全球有 $242 billion 流向 AI 公司——占当季全部风险投资的 80%——说明资本环境对 AI 基础设施公司极为有利,也包括空间 AI。四个细分市场的买方都面临采用 AI 工具的压力;像 World Labs 这样拥有可信产品(Marble 和 World API)的供应商,可以在市场整合前抢占早期份额。 API 经济是第三个结构性驱动。文本 AI API(OpenAI、Anthropic)已经证明,把基础 AI 能力嵌入第三方产品,会创造复合型生态效应。World Labs 于 2026 年 1 月发布 World API,就是要在空间 AI 中复制这一动态:一旦开发者把世界生成嵌入自己的产品,由此形成的数据飞轮和切换成本将创造可防守的市场位置。 与这些驱动相对的,是三项实质约束。sim-to-real gap 是最尖锐的技术约束:2026 年 arXiv 综述指出,生成的 3D 内容必须满足远超视觉真实感的要求——包括运动学结构、材料属性和交互支持——而持续存在的 sim-to-real 分裂仍是未解决瓶颈。这把 Marble 在机器人训练中的近期价值限制在视觉随机化用例,而不是完整的物理准确仿真。第二项约束是企业集成复杂度:Deloitte 2025 年 AI 领导者调查发现,约 60% 的受访者把遗留系统集成和风险 / 合规担忧列为 physical AI 采用的主要障碍,这与 World Labs 的企业销售动作直接相关。第三项约束是市场早期和买方教育:任何垂直领域都还没有「world models」的成熟采购路径,没有公开 ROI 基准,也没有标准集成模式——这些都会拉长销售周期,并限制企业收入速度。 EU AI Act 和各国 AI 监管框架扩散,也加入了合规维度。美国 NIST 的 AI Risk Management Framework 提供了企业正在采用的自愿治理标准;World Labs 已发布的 AUP 和内容政策,是面向企业合规姿态的早期步骤。但随着空间 AI 进入受监管行业(医疗仿真、自动驾驶),合规负担会继续上升。 [CM015, CM016, CM017, CM018, CM026, CM027]

增长驱动与约束表
驱动 / 约束方向类型时点含义尽调问题
机器人训练数据瓶颈驱动技术需求2024–2028可用于仿真的 3D 世界生成成为关键基础设施;World Labs 被定位为可规模化供给方与手工搭建的仿真环境相比,Marble 达到哪些物理保真度基准?是否已有机器人 OEM(不只是研究人员)在部署 Marble?
AI 投资激增(Q1 2026:全球 AI VC $242B)驱动资本与关注度2025–2027买方认知、工具预算和竞争压力同步上升;spatialAI 可以抢占早期心智资本是否具体流向空间 AI,还是集中在 LLM 和扩散模型基础设施,导致空间 AI 资金不足?
API 经济与开发者生态驱动商业模式2025–2028World API 降低开发者采用摩擦,可以形成类似 LLM API 先发优势的数据飞轮和转换成本World API 的真实开发者采用指标是什么(调用量、注册开发者、第三方集成)?
生成环境的 sim-to-real 缺口约束技术2024–2028Marble 生成的场景目前在机器人领域主要限于视觉随机化;若要物理准确部署,仍需攻克未解决的技术难题World Labs 是否发布过基准数据,对比 Marble 生成环境与手工环境的仿真保真度?
企业集成复杂度与合规约束组织2025–2028约 60% 的 AI 负责人把遗留系统集成和风险 / 合规列为最大障碍;World Labs 的企业销售周期可能较长World Labs 的企业合同规模、销售周期和已签企业协议数量是多少?
EU AI Act 与监管碎片化约束监管2025–2028监管扩散会抬高企业部署的合规成本;空间 AI 还面临合成数据和版权的额外风险World Labs 是否已评估其生成式世界模型输出在 EU AI Act 下的分类?现有 AUP 是否满足企业合规要求?

时点估计为定性判断;「驱动」与「约束」分类反映对 World Labs 近期商业牵引力的净方向性影响。

[CM015, CM016, CM017, CM018, CM031, CM033]
FM004: 采用漏斗或价值链地图

World Labs 面向企业和开发者的采用漏斗,从市场认知到平台锁定,并估算每一阶段的摩擦。

漏斗阶段摩擦评估来自案例研究证据和 Deloitte AI 采用调查数据的定性推断;没有 World Labs 转化率或流失数据。

[CM026, CM028, CM039, CM044]

2.5 规模测算矛盾、估算缺口与开放问题

本章的规模测算分析保留而不是消解了几个重要缺口。最根本的缺口是品类层面:没有第三方分析师把「空间智能基础设施」或「3D generative AI」定义为独立市场品类。因此,规模测算完全依赖代理,而这些代理——广义生成式 AI 和数字孪生——在范围上相差约两个数量级(数字孪生:2025 年 $21B;广义生成式 AI:2025 年 $53B)。两者都不是 World Labs 的「正确」市场;公司的实际可服务市场是二者的子集。 第二个缺口是分析师估算相互冲突。MarketsandMarkets 预测数字孪生市场到 2030 年 CAGR 为 47.9%;其他来源根据品类边界纳入内容不同,暗示更低增长。这些预测背后的方法是自上而下、基于调查的估算,并未得到独立验证。任何单一分析师数字的不确定区间都应至少按 ±50% 看待。 第三个缺口是私有证据:World Labs 未披露收入、ARR、客户数或 API 调用量,因此无法用实际变现数据推导自下而上的 SAM。公司已融资 $1.23 billion,并有 $5 billion 报道估值目标,意味着投资人期待其最终取得市场领导地位;但支撑这些期待的具体收入模型仍未公开。 第四,机器人与具身 AI 的 SAM 对一个假设高度敏感:机器人 R&D 支出中有多少用于仿真环境创建,而不是硬件、软件和运营——公开资料没有拆分。$0.5B–$2B 的自下而上估算来自一般机器人市场数据推断,而不是仿真基础设施支出的直接证据。 最后,关于空间 AI 采用放缓或市场怀疑的独立证据有限。已有来源主要呈现正面信号(Autodesk 投资、客户案例研究、学术验证)。不应把缺少负面证据视为市场热情一致的确认;考虑到 Deloitte 等机构记录的整体 AI hype cycle 动态,这更应被视为研究缺口。 [CM030, CM032, CM033, CM039, CM047, CM048]

2.6 图表

Chapter 03

03竞争对手

3.1 竞争格局概览

World Labs 的 Marble 在四个不同竞争层面展开竞争。第一层是直接的世界模型同行:Google DeepMind 的 Genie 2(2024 年 12 月)是一个基础世界模型,可从单张提示图生成动作可控、可玩的 3D 环境,主要面向 AI agent 训练,而不是商业内容工作流。总部位于 Palo Alto 的 Luma AI 称其使命是打造「unified general intelligence that can generate, understand, and operate in the physical world」,并提供 web 浏览器上 30 FPS 的交互式 3D 场景——直接争夺 World Labs 瞄准的创意生成市场。第二层是基础设施 incumbent:NVIDIA Cosmos(version 3)是开放的 physical AI 基础模型,在 GitHub 和托管目录中免费提供,覆盖视觉 AI 推理、机器人策略训练、世界仿真和合成视频数据生成。NVIDIA Isaac Sim 是建立在 Omniverse 上的开源仿真框架,集成 Cosmos 世界基础模型,用于机器人合成数据增强,并在机器人仿真用例上直接与 Marble 竞争。第三层是相邻创意 AI 平台:Magnific(前 Freepik)、OpenArt 和 Stability AI 的 Stable Zero123,在物体生成和 2D-to-3D 层面有重叠能力,尽管它们都尚未推出与 Marble 等价的场景级生成式世界模型 API。第四层是同时构成替代品的分发伙伴:OpenArt、Magnific、Rosebud AI、VIVERSE(HTC)和 Lightcraft 目前通过集成分发 Marble;但一旦替代或自研世界模型可用,它们都可能降低对 World Labs 的依赖。现状方案——在 Blender、Autodesk Maya 或游戏引擎资产工作流中手工创建 3D——仍是成本有效的替代品,适合不愿把流水线改造为生成式 AI 工具的买家;对资源充足团队而言,内部自建(定制世界仿真代码、Magic123 等开源学术模型)也可行。Google DeepMind 的 SIMA agent 框架增加了互补研究层;如果商业化,它会把 DeepMind 的版图延伸到 World Labs 通过 API 瞄准的 agentic 用例。 [CP001, CP003, CP004, CP005, CP006, CP007]

竞争对手画像矩阵
竞争对手类别规模 / 融资(2026)目标细分市场核心差异化关键限制
World Labs (Marble)直接同业 —— 空间 AI已融资 $1.23 B(种子轮 + Series A);NVIDIA、AMD、Autodesk、a16z 支持创意工具、仿真 API、机器人训练数据、建筑多模态世界生成(文本 / 图像 / 视频 / 3D);Spark 网页渲染器;World API价格设门槛;商业规模和采用指标未公开
Google DeepMind (Genie 2)直接同业 —— 基础世界模型Alphabet(约 $3 T 市值);Genie 未披露单独融资AI 智能体训练与评估;研究社区从单张提示图生成可由动作控制的 3D 环境;智能体仿真环境仅研究论文;截至报告日没有商业 API
Google DeepMind (SIMA)直接同业 —— 通用 AI 智能体Alphabet(同上)3D 游戏环境;具身 AI 研究在多样化游戏场景中遵循自然语言指令研究阶段;不是世界生成器;处在智能体层,而非内容创作层
NVIDIA Cosmos 3基础设施竞争者 —— 物理 AI 基础模型上市公司(NVIDIA);开源,未单独融资机器人研究人员、自动驾驶开发者、物理 AI通过 GitHub 开源权重;覆盖视觉 AI、机器人策略、世界仿真、合成数据聚焦机器人 / 物理 AI;创意 / 消费级用例有限
NVIDIA Isaac Sim既有厂商 —— 机器人仿真基础设施上市公司(NVIDIA);开源机器人仿真、合成数据生成基于 OpenUSD,可扩展,集成 Cosmos;免费开源需要手工搭建世界;原生不提供生成式 AI 世界创建
Luma AI直接同业 —— 3D + 创意 AI私营;截至 2026 年 6 月融资未披露创意团队、品牌、电影人交互式 3D 场景(30 FPS 网页);RAY3.2 视频导演;UNI-1 品牌智能公开文档无法看清世界模型深度;融资和员工数未披露
Stability AI (Stable Zero123)邻近替代 —— 3D 物体生成私营;财务承压(此前有重组报道)研究人员、独立创作者、商业会员单图到 3D 物体生成;Stable Diffusion 技术谱系;开放权重仅物体级(不是场景 / 世界尺度);默认非商业;商业层级有限
OpenArt分发伙伴 / 邻近创意平台私营;融资未披露数字创作者、叙事创作者多模型平台(Veo 3.1、Sora 2、Kling 3.0);角色一致性;免费层没有自有世界模型;3D 世界依赖 World Labs Marble API
Magnific (Freepik)分发伙伴 / 邻近创意平台Freepik 子公司;Freepik 估值约 €900 M(估算);声称有数百万用户品牌 / 营销团队、内容创作者30+ 个 AI 工具;节点式画布;品牌工作流;图像 / 视频 / 音频 / 3D没有自有世界模型;把 Marble 作为多项 3D 能力之一集成
VIVERSE (HTC)分发伙伴 / 元宇宙平台HTC 子公司(母公司上市)元宇宙、XR、基于浏览器的虚拟世界用户浏览器原生虚拟世界;XR 分发网络;社区平台生成式世界模型能力有限;依赖 Marble 生成 AI 世界

融资数据为截至 2026 年 6 月公开披露的轮次;未披露或内部轮次未计入。规模描述为序数型估计。部分竞争对手能力来自公司自述定位,并非独立基准验证。Magnific 估值是根据 Freepik 语境推导的估算值;未由一手来源验证。

[CP001, CP003, CP004, CP005, CP006, CP009]
FP001: 竞争定位图:商业成熟度 vs. 世界生成范围

按商业成熟度(y:仅研究 → 生产 API)和世界生成范围(x:对象 / 场景 → 完整交互世界)绘制竞争对手。World Labs 和 Luma AI 在商业成熟度及世界级输出上领先;NVIDIA Cosmos 已具备生产就绪度,但聚焦机器人;Genie 2 具备世界级能力,但仍停留在研究阶段。

轴位是基于公开产品文档和案例研究的序数判断,并非来自量化基准。x 轴:输出在多大程度上构成完整交互式 3D 世界(而非对象或 2D 媒体)。y 轴:商业产品可用性(0 = 仅研究;1 = 生产 API / 消费产品)。位置为近似值,不应视为精确分数。

[CP030, CP031, CP032, CP039, CP043]

3.2 能力与价格对比

World Labs 的 Marble 以多模态输入广度做差异化:文本、图像、视频、粗略 3D 布局和全景图都可作为生成输入,World Labs 称这是同类首创能力。公司的 web 端 3DGS 渲染器 Spark 2.0 提供 Level-of-Detail 流式系统,可在用户导航时按视角优化 Gaussian splatting 世界细节,让任何带 web 浏览器的设备都能渲染大世界——这一能力在 Luma AI 或 NVIDIA Cosmos 的公开文档中不可见。Luma AI 的交互式场景可在 iOS、Android 和 web 上以 8 MB(物体)和 20 MB(场景)、30 FPS 嵌入;Luma 还发布了用于定向视频的 RAY3.2 和用于品牌特定模型训练的 UNI-1,把竞争面扩展到 3D 捕捉之外。NVIDIA Cosmos 面向 physical AI 开发者,提供开源模型权重,覆盖文本、图像、视频和环境声音输入的世界生成,并专攻机器人与自动驾驶合成数据生成。Magic123 的学术工作(2023)证明,在开源学术语境下,结合 2D 和 3D diffusion priors 可以实现单图到 3D 物体生成,说明底层技术并不锁定在专有系统中。价格是一个实质未知数:World Labs 尚未公布 World API 价格,并采用 gated early-access 模式;NVIDIA Cosmos 开源(无明确许可成本,但基础设施成本仍然存在);Stability AI 的 Stable Zero123 可用于非商业和研究用途,商业用途需要会员(价格未披露);OpenArt 平台提供免费层图像和视频工具,但 OpenArt Worlds 3D 层价格未单独披露。 [CP018, CP019, CP020, CP021, CP022, CP023]

功能 / 能力矩阵
购买标准World Labs MarbleNVIDIA CosmosLuma AIGoogle DeepMind Genie 2Stability AI Zero123
文本到 3D 世界 / 场景生成部分(带世界仿真语境的视频生成)部分(3D 捕捉 + 交互;非生成式场景)是(从单张提示图生成)否(仅物体级)
多模态输入(文本、图像、视频)是(文本 / 图像 / 视频 / 粗略 3D 布局 / 全景图)是(文本 / 图像 / 视频 / 环境声音)是(文本 / 图像 / 视频)部分(图像提示;键盘 / 鼠标动作)有限(单张图像)
基于网页的流式渲染器是(Spark 2.0;LoD 流式;WebGL2)否(SDK/API;未发布网页渲染器)是(30 FPS 网页嵌入;iOS/Android)
商业 API 可用是(World API;设门槛早期访问)是(开源 GitHub + 托管目录)未知(未找到公开 API 文档)否(仅研究论文)有限(商业会员层;权重可用)
机器人仿真支持是(案例研究:机器人训练数据、NVIDIA Isaac Sim 集成)是(核心用例;机器人策略模型骨干)部分(AI 智能体训练环境)
创意电影 / VFX 用途是(案例研究:片场一致性、虚拟制作)部分(合成视频数据;不是创意电影)是(RAY3.2 导演式视频;UNI-1 品牌)
开发者生态 / 展示是(Marble Labs;20+ 个公开演示;Splat World、头像工具)是(NVIDIA Omniverse 生态;Isaac Lab、Isaac Sim)部分(有文档的创意工作流)否(仅研究原型)部分(HuggingFace 模型卡;研究笔记本)

能力评级反映截至 2026 年 6 月公开可得文档和案例研究。标为「未知」的单元格表示未找到公开证据。「是 / 部分 / 否」评级是基于现有证据的编辑部判断,不是独立基准测试。竞争对手能力可能超出公开文档所显示的范围。

[CP019, CP020, CP021, CP022, CP023, CP024]
定价与包装对比
竞争对手访问模式标价 / 单位包含能力已知折扣 / 未知项对 World Labs 的含义
World Labs (World API)设门槛商业 API;早期访问合作未公开披露世界生成;Spark 渲染器;多模态输入;导出到下游工具完整定价不透明;企业定价大概率协商定价不透明限制开发者自助采用;无法做基准对比
NVIDIA Cosmos开源(GitHub);托管目录访问无授权成本(需承担基础设施成本)视频生成;机器人策略模型;世界仿真;合成数据依赖 NVIDIA 算力;企业支持定价未知零授权成本削弱 World Labs 在机器人用例中的定价权
Luma AI消费者 / 专业消费者产品;API 未文档化API 未公开披露;未找到消费者定价交互式 3D 场景;导演式视频(RAY3.2);品牌智能(UNI-1)未知;可能是 SaaS 或按生成计费;未披露企业层级如果 Luma 以更低成本提供 API,创意市场可能出现定价压力
Stability AI (Zero123)开放权重(非商业);商业会员层非商业:免费;商业:会员订阅(金额未披露)3D 物体新视角合成;基于 SD1.5 生成商业层定价未披露;会员模式限制企业销售物体级低成本可能塑造买方对 3D 生成的价格预期
Google DeepMind (Genie 2)研究论文;无商业产品N/A(不可用)可由动作控制的 3D 环境;智能体训练环境若商业化,Alphabet 分发能力可能支持免费增值定价潜在威胁:Alphabet 可能以接近零边际成本规模化商业化
OpenArt免费增值消费者平台;通过 Marble 集成 3D图像 / 视频免费层;3D 层定价未披露图像、视频、音频生成;多模型(Veo 3.1、Sora 2、Kling 3.0)3D 世界层定价未单独披露;集成合同未披露创作者免费入口扩大 World Labs 触达,但可能锚定较低价格预期
NVIDIA Isaac Sim开源(GitHub);通过 NVIDIA 提供企业支持开源层无授权成本基于 USD 的机器人仿真;传感器建模;合成数据生成可获得企业支持合同,但定价未披露机器人仿真的免费基线降低买方为 Marble 机器人用例付费的意愿
Magnific (Freepik)免费增值 / 订阅;Marble 3D 是其中一个功能免费层;来源未披露付费层定价30+ 个 AI 工具;图像、视频、音频、3D;品牌工作流;节点式画布免费增值模式;Marble 集成价格对最终用户不可见Marble 嵌入 Magnific 付费层;定价对最终用户不透明

所有定价均反映截至 2026 年 6 月公开披露的标价或访问模式。World Labs World API 未公开列价。NVIDIA Cosmos 开源访问不排除算力基础设施成本。「未知」表示在已抓取来源中未发现公开商业定价。

FP002: 按竞争对手划分的功能广度:能力覆盖地图

World Labs 与五个关键竞争对手在七个能力维度上的覆盖矩阵,展示 World Labs 领先之处(多模态输入、Spark 渲染器、创意电影)以及与他人重叠的地带(与 NVIDIA Cosmos 的机器人用例重叠、与 Luma AI 的网页 3D 重叠)。

评级(完整 / 部分 / 是 / 否 / 有限 / 增长中 / 成熟)是基于截至 2026 年 6 月公开文档的编辑判断。没有独立基准可用。

[CP019, CP021, CP022, CP023, CP024, CP025]

3.3 护城河、切换成本与分发力量

World Labs 的结构性护城河建立在三根支柱上:创始人可信度和人才吸引力、资本深度,以及 API 嵌入下游平台。联合创始人阵容——Fei-Fei Li(ImageNet、计算机视觉)、Ben Mildenhall(NeRF)、Christoph Lassner 和 Justin Johnson(3D representations)——带来招聘和可信度优势,并体现在研究速度和企业信任建立上。$1.23 B 总融资(已公告种子轮加上 2026 年 2 月由 NVIDIA、AMD 和 Autodesk 支持的融资)提供多年 runway,可投入模型规模、数据基础设施和开发者生态。API 嵌入策略——OpenArt、Magnific、Rosebud AI 和 VIVERSE 等伙伴把 Marble 驱动的功能做进自有产品——会在其终端用户开始使用并依赖 Marble 生成的 3D 环境后,在伙伴层面创造切换成本。但合同锁定条款未公开,多宿主风险真实存在:按标准 API 条款运营的伙伴,可以在数周内把流量转向替代供应商。Marble Labs 展示目录(20 多个公开 demo:VR Gaussian splats、collider builders、avatar tools、建筑空间蓝图等)正在拓宽开发者生态,但单个 demo 的商业牵引未知。NVIDIA 同时是战略投资人和 Cosmos 竞争产品的开发者,这形成内在张力:NVIDIA 可能影响 Cosmos 路线图,吸收 World Labs 的机器人仿真用例,并借 Omniverse 和 Isaac Sim 分发网络把该空间商品化。直接基于 World API 开发的开发者切换成本中等:API 接口新颖,下游集成(伙伴产品功能、仿真流水线)迁移需要工程返工,但目前未披露数据可携带限制或竞业条款。 [CP031, CP032, CP033, CP034, CP035, CP036]

护城河持久性 / 竞争风险登记表
护城河主张威胁严重性缓释 / 尽调问题
多模态输入广度(文本 / 图像 / 视频 / 3D 布局 / 全景图)NVIDIA Cosmos 和学术模型(Magic123)展现了类似输入广度,且不依赖专有 IP核验在等价输入下,Marble 的世界尺度场景生成质量是否显著领先竞争对手;要求提供基准对比
Spark 2.0 网页渲染器与 LoD 流式Luma AI 的 30 FPS 网页嵌入是竞争性渲染器;WebXR 和浏览器技术演进很快中低评估 Spark 是否已申请专利或具备可防御技术 IP;评估第三方渲染器对竞争方案的采用情况
API 平台与伙伴集成(OpenArt、Magnific、Rosebud、VIVERSE、Lightcraft)多归属:按标准 API 条款合作的伙伴可以低转换成本转向 NVIDIA Cosmos 或 Luma AI要求查看合同条款、最低调用量承诺和排他性条款;评估集成深度(功能级还是工作流级)
创始人履历与人才吸引力(Fei-Fei Li、Mildenhall、Lassner、Johnson)DeepMind 和 NVIDIA 可以用更高薪酬吸引同等研究人才审查团队留存指标、股权归属时间表,以及关键研究人才是否受合同约束
资本厚度(已融资 $1.23 B;NVIDIA + AMD + Autodesk 为战略投资方)基础 AI 巨头的资本集中度远超 World Labs 的资本基础;Crunchbase 数据显示,OpenAI/Anthropic/xAI/Waymo 在 Q1 2026 单季融资 $188 B评估现金跑道、烧钱速度和下一轮融资触发点;评估 NVIDIA 战略投资是否限制产品路线图灵活性
NVIDIA 作为战略投资方和生态伙伴(Isaac Sim 集成)NVIDIA 的 Cosmos 产品与机器人仿真用例直接重叠;投资方一致性未必能阻止竞争性产品动作对战略投资协议做独立法律审查;评估是否存在优先购买权、董事席位或路线图影响条款

严重性评级(高 / 中 / 低)是基于公开证据的编辑部判断,并非来自正式风险模型。缓释项是尽调动作,不是已确认的公司实践。本表覆盖了有公开证据支撑的护城河维度;可能存在额外专有护城河因素,但无法用公开来源评估。

FP003: 护城河和竞争准备度 KPI 摘要

截至 2026 年 6 月,World Labs 竞争准备度指标的压缩摘要,展示支撑护城河的优势,也展示构成投资人风险的结构性缺口。

伙伴数量基于公开披露的案例研究和集成公告。模态数量来自公司说法。融资额来自公开公告。开源竞争对手数量反映截至 2026 年 6 月,空间 AI 类别中公开确认的免费 / 开放权重模型。

[CP036, CP037, CP038, CP041, CP042, CP044]

3.4 进入者压力与负面证据

2026 年基础 AI 的竞争资本环境,给 World Labs 构成结构性逆风。Crunchbase 数据显示,2026 年 Q1 基础 AI 创业公司仅 24 笔交易就获得 $178 billion 风险投资;OpenAI、Anthropic、xAI 和 Waymo 合计拿走当季全球 VC 的 $188 billion(65%)。这种集中意味着 World Labs 需要与研究和产品预算远超自身的组织竞争:Google DeepMind 可以通过 Genie 2 和 SIMA 长期资助世界模型研究且无需收费;NVIDIA 可以以零许可成本免费提供 Cosmos 模型权重,借此深化 Omniverse 生态锁定。开源学术轨迹也从下方带来进入者压力:Magic123 和类似学术 3D 生成模型说明,不依赖专有许可也能获得底层 3D diffusion 技术,研究质量与生产质量输出之间的差距每年都在缩小。Stability AI 是独立 AI 模型公司这一原型的负面样本:它曾是生成式图像和 3D 生成(Stable Zero123)的早期领导者,但限制性的非商业发布姿态、商业会员要求和已披露财务困难,显示资本受限的 AI 基础设施公司有多脆弱。NVIDIA Cosmos 的开源可用性明确加速机器人开发者采用;这些开发者本来可能评估 World Labs 的商业 World API,因此 Cosmos 直接削弱 World Labs 在该垂直领域的定价权。最不利情景是双线挤压:NVIDIA 通过免费 Cosmos 模型把机器人仿真市场商品化,同时 Google DeepMind 将 Genie 2 商业化为创意 API,使 World Labs 在两个锚定市场都无法匹配 incumbent 的分发或资本深度。 [CP039, CP040, CP041, CP042, CP043, CP044]

3.5 图表

Chapter 04

04财务

4.1 收入来源与变现界面

World Labs 目前通过两个主要产品界面变现:面向 Marble 世界生成平台的直接订阅服务,以及把同样能力开放给开发者和企业团队的程序化 API。第三个界面——通过 Autodesk 获得的战略合作收入——已经存在,但除 $200 million 投资披露外,商业条款尚未公开。 Marble 订阅分为四档。Free 允许有限世界生成,不含商业权利。Standard 每月 $20,扩大生成预算。Pro 每月 $35,加入明确的商业使用权,是创作者可将 Marble 输出嵌入付费交付物的第一档。Max 每月 $95,提供包括最高保真输出在内的全部功能。这个价格曲线位于 prosumer 创意工具与专业 3D 软件之间:比 Cinema4D 或 Houdini 等高端 DCC 软件席位许可便宜,但明显高于商品化图像生成工具。各档实际收入结构未披露;Free、Standard、Pro 和 Max 用户的相对权重未知。 World API 于 2026 年 1 月 21 日发布,使开发者可以用文本、图像、全景图或视频输入,以程序化方式调用 Marble 世界生成。API 产生一个异步任务,并交付完整可导航的 3D 世界。截至 2026 年 6 月,未观察到 API 的公开价格表;这符合早期企业 API 产品的常见形态,即按客户谈判定价,或置于申请流程之后。API 的设计——提交请求、接收可导航世界——在架构上与按使用量的 token 或 credit 定价一致,但这是推断,并未得到任何公开披露确认。 Autodesk 合作是最显眼的企业变现信号。Autodesk 投资 $200 million,并将担任战略顾问,初始重点放在媒体和娱乐用例。除最初娱乐焦点外,双方未披露合同金额、收入分成条款或产品集成路线图。随着 Autodesk 将 Marble 能力接入其设计与施工工具链,该合作可能发展为经销或嵌入式 API 安排,但目前这仍是前瞻可能性,而非已验证事实。 商业使用权——区分 hobbyist 与专业用途的变现分界线——从每月 $35 的 Pro 档开始提供。这个结构特征对财务建模很重要:任何希望把 Marble 输出用于客户项目、媒体制作或嵌入式产品的订阅者,都必须使用 Pro 或 Max,这会把有商业意图的订阅者集中到两个更高档位,也就是最持久、最可防守的收入。Free 和 Standard 档主要是转化漏斗,还是会产生有意义的流失,仍未知。[CI006, CI007, CI008, CI009, CI010, CI011]

收入流与机制
收入流机制单位当前状态证据质量尽调问题
Marble 订阅(Pro/Max)经常性月度 / 年度订阅;Pro 层起提供商业权利每席位每月已上线 —— 自 2025 年 11 月 12 日起全面可用高(官方定价页和 ToS)披露付费订阅用户数、层级组合和月度订阅 MRR
Marble 订阅(Free/Standard)免费转付费漏斗;生成额度有限;Pro 以下不含商业权利免费增值转化已上线——公开可用高(官方定价页)披露 Free/Standard 向 Pro/Max 的转化率;按层级披露流失率
World API以程序化方式访问 Marble 世界生成;异步请求 / 响应基于用量的点数或按调用计费(定价未公开)已上线——Jan 21, 2026 发布;无公开价目表中(发布公告已确认;缺少定价)发布 API 定价表;披露 API 调用量和收入贡献
Autodesk 战略合作从媒体 / 娱乐切入的共同开发,以及潜在嵌入式 API 或转售安排协商商业条款(未披露)进行中——$200M 投资 + 顾问角色;商业收入条款未知中(Reuters、TechCrunch 交叉印证)披露来自 Autodesk 的任何收入分成或产品集成收入
未来企业授权面向 AEC、机器人仿真、游戏管线嵌入的企业直签合同按年协商;无已披露合同商业化前——仅为潜在机会低(由合作伙伴案例推断)获取 LOI 或已签合同;确认管线价值

所有收入数据均为 null;收入流状态和证据质量基于公开官方与第三方来源。Autodesk 商业条款、API 定价和订阅 MRR 均属承销所需的私有披露项。

[CI006, CI007, CI008, CI009, CI010, CI011]
Marble 定价与货币化
层级月费商业权利生成额度标价 vs. 实收关键备注
Free$0不包含有限仅标价漏斗入口;不可商用
Standard$20/month不包含扩大仅标价面向准专业创作者;不含商业权利
Pro$35/month包含(ToS 明确)更高额度仅标价商业权利门槛;可能覆盖可货币化创作者中的大多数
Max$95/month包含(全部功能)最高额度 + 全部功能仅标价专业 / 工作室层级;定价约为 Pro 的 2.7×
World API未公开列示取决于协商条款按 API 调用或点数计量未公开提供企业 / 开发者层级;截至 Jun 2026 定价未披露

标价来自 worldlabs.ai 和服务条款(Jan 21, 2026 版本)。实收价格、折扣、年度预付和批量费率均未披露。所审阅的所有公开页面均缺少 API 定价。

[CI006, CI007, CI008, CI010, CI034]
FI001: 收入模型桥

Marble 和 World API 上的客户活动如何转化为收入并进入毛利池;核心未解不确定性是每次生成的算力成本。

流程结构有来源支撑,包括层级存在和定价;总收入、GPU 成本和毛利节点为定性推断。未披露任何财务规模。

[CI006, CI007, CI019, CI021, CI022, CI034]

4.2 GTM 路径与企业开发者代理指标

World Labs 的 GTM 走的是混合路线:面向个人创作者,用产品驱动的自助购买;面向企业和开发者伙伴,走定向渠道。两条线互相补位:自助层贡献采用信号、创作者内容和社会证明,反过来支撑企业销售;Autodesk 这类企业合作则打开大规模存量客户池,以 World Labs 当前团队规模,单靠直销自助渠道很难高效触达。 自助订阅漏斗是最清晰的 GTM 动作。Marble 已在 worldlabs.ai 上线,消费者和专业消费者层级不需要销售辅助入门。World Labs 官网案例记录了个人电影创作者、游戏创作者、建筑师和设计师如何在没有企业合同的情况下采用 Marble:电影人 Joshua Kerr 把 Marble 与 Lightcraft Jetset 结合做虚拟制作;Tim Simmons 和 Henrik Vasquez(Theoretically Media)把 Marble 当作虚拟片场,让 AI 生成场景之间的镜头保持一致;Fenestra 和 Interior AI 的建筑师用 Marble 做沉浸式设计可视化。这些用例落在 Standard 到 Pro 的价格带。 开发者和 API 渠道由 World API 以及六个具名第三方平台集成体现:OpenArt(推出「OpenArt Worlds」,作为其创作者平台的持久 3D 环境功能)、Magnific(集成 Marble,用于精准控制产品放置)、Rosebud AI(把 Marble 的世界生成嵌入游戏构建流程)、VIVERSE(用 Marble 生成可在浏览器中交互游玩的 3D 世界),以及 Escape(游戏化和交互体验平台)。第七个集成——NVIDIA Isaac Sim 研究人员用 Marble 生成机器人仿真数据——延伸到企业和研究场景,说明该 API 在高风险机器人和具身 AI 场景中具备技术可信度。 公开资料没有销售效率代理指标。各层级均未披露获客成本、平均合同金额、回本周期或净收入留存。缺少这些指标,再叠加 Autodesk 合作强调「探索」、联合研究与模型协作的表述,说明企业销售动作仍处在早期商业开发阶段,而不是已经规模化、可复制的销售循环。Deloitte 对企业 AI 采用的分析指出,合规复杂度、员工准备度和监管不确定性是普遍拉长 AI 平台公司销售周期的主要阻力——这直接作用于 World Labs 更长期的企业 GTM 目标。[CI011, CI012, CI013, CI014, CI015, CI016]

单位经济状态
指标公开数值置信度重要性尽调要求
月度订阅 MRR未披露订阅业务的基准收入轨迹;任何 DCF 或可比公司分析都需要索取管理账;用支付处理商数据或创作者经济估算交叉验证
API 收入贡献未披露衡量开发者货币化效果,以及 API 优先账户带来的净留存向管理层索取 API 定价表和月度 API 调用量
毛利率(混合)未披露判断订阅定价在规模化后能否覆盖 GPU 推理成本用算力成本基准建模毛利率;再用管理账确认
月度烧钱率未披露计算 runway 和融资依赖必须掌握;也为情景模型划边界索取月度现金流量表和过去 12 个月烧钱趋势
CAC(自助式)未披露消费者 / 准专业漏斗的获客成本效率;决定 LTV/CAC 比值索取数字营销的混合付费获客成本;按层级归因
LTV(Pro/Max 订阅用户)未披露判断订阅经济性是否支撑继续投资漏斗索取按层级划分的平均留存时长和流失率;结合标价估算 LTV
净收入留存未披露显示企业扩张速度;API 优先商业模式的核心指标索取 API 和 Pro/Max 订阅用户前 6 个月队列留存数据

所有单位经济指标均为 null:截至 Jun 2026,公开信息未披露任何财务经营指标。数值完全来自公司内部记录;null 不代表该指标为零。

[CI006, CI010, CI019, CI021, CI022, CI031]
FI002: 单位经济桥

从获客到同期群 LTV 的定性桥,展示证据在哪些环节断裂;所有量化节点在私人披露前均为空。

所有量化节点均为空;流程只使用有来源支撑的结构性证据(层级定价、商业权利门槛、API 发布)。未推断任何数字规模。

[CI006, CI008, CI013, CI031, CI033, CI040]

4.3 成本结构与资本密集度

World Labs 的成本结构符合前沿 AI 研究和云端推理公司的特征:模型训练和实时推理的 GPU 算力占主导,薪酬盘子相对较小,没有硬件制造或库存 COGS。原则上,这是高毛利的软件结构;但生成高保真、持久 3D 世界的算力强度明显高于可比的 2D 图像生成负载。因此,在披露前,其毛利形态更像 GPU 密集型推理服务,而不是轻量 SaaS 部署。 员工规模仍小。The Org 的资料显示,截至 2026 年 6 月,World Labs 有 11 至 50 名员工。Ashby 平台上的研究、工程和产品岗位招聘显示公司仍在扩招,但总员工数和月度薪酬未披露。按区间低端估算,薪酬可能是每月低个位数百万美元;即便按上限,薪酬线相对 $1.23 billion 融资额也很小。 最大成本驱动项是算力基础设施。Marble 能从单张图片或文本提示生成空间连贯、高保真、持久的 3D 环境。每次生成都涉及神经渲染、Gaussian splat 重建和交互环境封装——每个环节都吃 GPU。World Labs 创始人包括 NeRF 发明者和 Pulsar 作者;两人的技术背景合在一起,指向一种训练模型和单次请求推理都需要高端 GPU 集群的渲染基础设施。最可能的算力基础设施来自云服务商或本地 GPU 伙伴,但公司未披露任何合同。 毛利动态:Global Market Insights 的生成式 AI 市场基准把「高基础设施和算力成本」列为该行业的主要挑战;与此同时,随着模型效率提高和算力成本下降,单次生成的边际成本预计会下滑(Stanford HAI 2026 AI Index 也提到这一模式)。World Labs 的长期毛利轨迹高度取决于推理效率提升,以及公司能在多大程度上靠规模拿到更好的算力经济性。公司没有公开任何自身毛利数据。 纯软件交付模式限制了营运资本和资本开支需求。公开资料没有库存、硬件制造扩建或实体基础设施。主要资本投向预计是研发(模型训练运行)和产品基础设施(推理服务),两者都属于经营支出,固定资产强度有限。公司未披露任何债务工具或项目融资义务。[CI019, CI020, CI021, CI022, CI023, CI036]

FI004: 资本配置和成本强度地图

对一家符合 World Labs 画像的前沿 AI 推理公司,定性拆解主要资本投放类别;数值是结构性原型,不是披露数据。

该瀑布图展示的是前沿 AI 推理公司的结构性成本原型,不是 World Labs 实际财务报表的披露或估计。各项数值是示意性比例,来自行业基准和公开可观察的可比公司;不得解读为公司特定财务数据。World Labs 未披露任何成本拆分。

[CI019, CI020, CI021, CI022, CI023, CI036]

4.4 资本充足性与融资依赖

公司概览章节(第 1 章)记录了完整融资时间线;本节聚焦未来资本充足性和融资依赖。World Labs 完成了两轮融资:2024 年 9 月 $230 million 种子轮,以及 2026 年 2 月 $1 billion Series B,合计约 $1.23 billion。种子轮由 Andreessen Horowitz、New Enterprise Associates 和 Radical Ventures 领投,AMD Ventures、Intel Capital 和 NVIDIA NVentures 也参与。Series B 引入 AMD、Autodesk(以 $200 million 作为锚定方)、Emerson Collective、Fidelity Management and Research Company、NVIDIA 和 Sea。Series B 投资者名单中没有 Intel Capital。 投资者构成释放了明确的财务信号。Autodesk 作为单一最大支票方投入 $200 million,这是来自潜在分销伙伴和客户的战略投资,不只是财务资本;它让未来产品集成进 Autodesk 的设计和建筑工具具备激励一致性。AMD 和 NVIDIA 作为战略共同投资者,降低了 World Labs 面临算力供应约束的可能性,因为两家芯片厂商都直接受益于其算力消耗增长。Fidelity Management and Research 的参与,则符合面向多年后公开市场流动性事件的成长股定位。 Bloomberg 2026 年 1 月报道称,Series B 估值讨论约为 $5 billion——较种子轮 $1 billion 估值提高五倍。World Labs 未公开确认或否认该数字,所审阅来源中也没有披露任何条款清单或最终估值证明。 公开资料没有月度 burn rate、现金头寸或 runway 估算。类似阶段和团队规模的算力密集型 AI 公司,月度现金消耗通常从几百万美元到数千万美元不等,取决于训练强度。若以每月 $5 million 的中点 burn(仅作示例;非公司披露)计算,$1.23 billion 约等于二十年 runway——这个数字并不现实,说明实际 burn 明显更高,或资本会投入更激进的模型训练和基础设施扩建。若每月 $20 million,runway 约五年。若每月 $50 million,约两年。这些情景都无法从公开证据验证;列出它们只是为了界定范围,不是估算。公司未披露任何债务额度或项目融资义务。[CI001, CI002, CI003, CI004, CI005, CI025]

资本充足性摘要
项目数值 / 状态来源质量前瞻影响
种子轮(Sep 2024)$230M;由 a16z、NEA、Radical Ventures 领投;AMD Ventures、Intel Capital、NVIDIA NVentures 参投高(Reuters、a16z 投资组合披露)产品前资本;支撑 Marble 开发至 GA(Nov 2025)
Series B(Feb 18, 2026)$1.0B;AMD、Autodesk($200M)、Emerson Collective、Fidelity、NVIDIA、Sea 等投资方高(公司官方公告、Reuters、TechCrunch)GA 后增长资本;支持模型扩展、企业 GTM、API 建设
累计融资额约 $1.23B 累计高(多方来源交叉印证)runway 充足;没有烧钱率披露,精确 runway 未知
估值(未确认)Bloomberg Jan 2026 报道约 $5B;公司未确认中(第三方报道;公司未反驳)较 $1B 种子轮估值上调 >5×;若烧钱率偏高,稀释压力显著
月度现金烧钱未披露无法计算 runway;烧钱情景分析需要私有披露
Runway(计算值)公开数据无法计算公开证据没有基础仅作情景:若每月烧钱 USD 10M,约 10 年;若每月 USD 50M,约 2 年(仅为示例,不是估算)
债务 / 项目融资未披露中(所审阅来源均无债务证据)暗示资本结构干净;需用法律文件 / 409A 审查验证

融资数据来自公司公告、Reuters、TechCrunch、StartupHub 和 a16z 投资组合披露。估值仅来自 Bloomberg 报道。所有 null 值均代表私有数据;runway 区间只是情景示例,不是估算。

[CI001, CI002, CI003, CI004, CI025, CI026]
FI003: 财务估算区间

有来源支撑的融资额和 Bloomberg 报道估值边界;示例性现金可维持年限边界作为情景输入,而非估算。

融资额由多个高信誉来源印证(官方公告、Reuters、TechCrunch)。估值仅来自 Bloomberg 报道。消耗情景是从可比前沿 AI 公司推导的示例边界;不是对 World Labs 实际消耗的估算。

[CI001, CI003, CI004, CI025, CI026, CI027]

4.5 公开牵引力缺口与财务结论

World Labs 于 2025 年 11 月实现产品 GA,并在 2026 年 1 月推出开发者 API。截至 2026 年 6 月——GA 后七个月、种子轮后十七个月——公司未披露收入、ARR、客户数、毛利、月度 burn,或任何单位经济指标。投资规模($1.23 billion)与财务透明度之间的落差,是同阶段私营 AI 公司中最宽的一类;这符合更广泛的空间 AI 投资逻辑:资金押注创始人履历和技术差异化,而不是已被证明的商业牵引力。 财务画像的正面叙事建立在三根支柱上。第一,$1.23 billion 资本由 Autodesk 的战略支票锚定,给公司提供多年运营弹性,不必短期内优化收入。第二,六个具名 API 集成落地在运营平台上(OpenArt、Magnific、Rosebud、VIVERSE、Lightcraft、Escape),再加上 NVIDIA Isaac Sim 机器人仿真用例,说明产品已经商业部署,而不只是实验室测试。第三,订阅定价架构——Pro($35/mo)作为商业使用权入口——形成天然升级漏斗,即使企业合同尚未放量,也应从创作者和专业消费者群体产生一些近期收入。 风险叙事有两个主要财务维度。第一是采用门槛:Deloitte 的企业 AI 采用研究指出,合规复杂度、员工准备度和监管不确定性会拉长企业销售周期并提高获客成本;这些因素同样作用于 World Labs 的企业 GTM 目标。第二是资本集中:Crunchbase 数据显示,基础 AI 融资在 2026 年 Q1 翻倍以上,但越来越集中到少数前沿巨头。对 World Labs 这样的非最前沿玩家而言,这种集中格局意味着,如果企业收入轨迹未能足够快地显现,在当前资本基础被大幅消耗时,一个假设中的第三轮融资可能面临更高稀释或不利市场条件,这不是小风险。 财务结论是:公开证据能可信地支撑资本承诺、团队质量,以及已越过原型阶段的产品化动作。它不能支撑收入质量承保。仅凭公开证据,任何投资者都无法确定仓位规模或建模回报情景。最低尽调要求包括:管理账或审计财报、API 定价表和当前企业合同管线、月度现金 burn 与 runway 预测,以及 Autodesk 合作中除已披露投资之外的财务条款。[CI029, CI030, CI031, CI032, CI033, CI041]

公开财务披露缺口
缺失指标缺口类型投资影响尽调路径
收入 / ARR仅有私有证据无法测算商业化爬坡,或验证订阅货币化假设索取审计财务报表或管理账;用支付处理商数据交叉验证
月度烧钱率仅有私有证据无法计算实际 runway,也无法评估资本是否足以支撑计划索取过去 12 个月月度 P&L 和现金流量表
订阅用户数和层级组合仅有私有证据无法建模贡献毛利或转化漏斗健康度索取 Free / Standard / Pro / Max 的订阅用户拆分,以及月度队列进入量
World API 定价和调用量缺少公开来源无法建模 API 收入轨迹,或评估企业开发者货币化向管理层索取 API 定价表和过去月度调用量
毛利率仅有私有证据无法判断定价在规模化后能否覆盖 GPU 推理成本索取 COGS 拆分,将算力与人员分开;对照 AI 基础设施基准
Autodesk 商业协议条款仅有私有证据$200M 投资条款已披露;嵌入式产品或收入分成条款未披露索取 Autodesk 合作协议;识别任何收入最低承诺
按层级划分的 CAC 和 LTV仅有私有证据无法评估自助漏斗或企业销售的客户经济性索取按渠道划分的营销支出、混合 CAC 和按层级划分的平均留存时长

所有缺口反映的是公开披露缺失,不代表公司经营有缺陷。World Labs 是私营公司,没有义务公开这些指标;所有项目都是商业化阶段 AI 公司的标准尽调要求。

[CI031, CI027, CI010, CI032, CI019]

4.6 图表资料

Chapter 05

05产品与技术

5.1 产品定义与客户工作流

World Labs 的旗舰产品 Marble,把自己定位为生成式空间智能引擎,能把多模态输入直接转成可探索的 3D 世界。从客户视角看,工作流始于用户提供一个创意或信息信号——短文本描述、参考图、视频片段、360 度全景图,或粗略 3D 布局——然后获得一个持久、可导航的 3D 环境。用户可以在网页浏览器中探索、交互式编辑、扩展成更大场景,并导出给下游使用。公司把这描述为「把看见转成行动、把理解转成推理、把想象转成创造」,等于把过去需要专职 3D 艺术家、摄影测量设备或手工场景建模的流程压缩成一个概念化工作流。 Marble 既以 marble.worldlabs.ai 上的直接消费者 / 专业消费者订阅形式提供,也通过 World API 作为可编程服务提供。该 API 于 2026 年 1 月 21 日上线,允许开发者把世界生成嵌入产品和流程,而无需维护 3D 生产栈。API 异步处理请求,并返回一个可完整导航的 3D 世界对象;该对象可在浏览器中渲染,可导出为 Gaussian splat(.spz 或 .rad 格式)、mesh(.glb)或视频,也可直接传入仿真框架。 已记录用例覆盖四个主要垂直场景:(1)电影和 VFX,电影人把 Marble 当作稳定虚拟片场,让场景和镜头之间保持空间一致性;(2)游戏和 AR/VR,开发者从生成环境构建交互式浏览器体验;(3)机器人和具身 AI 仿真,研究人员生成可扩展训练环境,用于域随机化;(4)建筑和设计,Marble 让设计师在绘制平面图前先走进概念,从而加速概念可视化。World Labs 自己也用 Marble 制作发布营销视频,过程中生成了数百个 3D 世界。 产品层级包括 Marble(生成式世界模型)、World API(可编程访问层)、Spark 2.0(渲染引擎)和 Marble Labs(开发者门户与案例中心)。每一层面向不同客户群:Marble 面向个人创作者和专业人士;World API 面向平台开发者和企业团队;Spark 支持第三方运行时渲染;Marble Labs 推动生态采用和开发者赋能。[CE001, CE002, CE003, CE004, CE005, CE006]

产品模块与资产矩阵
模块 / 产品主要用户状态 / 成熟度差异化尽调缺口
Marble(世界模型)个人创作者、设计师、开发者GA(约 Nov 2025)多模态输入;交互式编辑;碰撞体网格导出架构未披露;缺少独立质量基准
World API开发者、平台集成商、企业团队GA(Jan 21, 2026)异步生成;REST 接口;无需专有客户端定价未公开列示;SLA 和速率限制未披露
Spark 2.0 渲染器Web 开发者、游戏开发者、VR 开发者GA(Apr 2026)WebGL2 3DGS;LoD 流式加载;多对象;支持 VR未发布与替代方案对比的渲染性能基准
Marble Labs 门户开发者社区、生态建设者GA(随 Marble GA 发布)案例、教程、展示管线;社区生态枢纽未披露开发者采用指标或 API 调用量
Marble Composer高阶用户、世界组装者可用(Marble UI 功能)单次会话内的多世界组合和空间编辑未正式记录其与基础 Marble 产品的功能边界
碰撞体网格导出机器人研究者、游戏开发者、仿真工程师GA(机器人案例中有记录)无需手工建模,即可在 Isaac Sim、Unity、Unreal 中启用物理仿真尚未用基准对比手工搭建的仿真环境来验证物理保真度
导出格式(splat、mesh、video)创意专业人士、管线工程师GA支持多种格式:.spz、.rad、.glb、video;可接入 DCC 工具大型场景的导出大小限制和保真度上限未记录

状态和成熟度基于截至 June 2026 公司发布的博客、案例和展示页。未观察到独立成熟度审计或第三方认证。尽调缺口指公开来源中未找到的信息。

[CE001, CE002, CE003, CE005, CE011, CE014]
工作流与用例表
用户任务 / 垂直领域既有工作流Marble 方案已记录收益观察到的限制
电影 / VFX:稳定虚拟布景手工 3D 场景建模或外景勘景;逐镜头提示会丢失空间一致性生成单个持久 3D 世界;可在不同场景中探索、取景和复用Tim Simmons 用稳定 3D 环境制作了 48 秒微短片;部分角度仅靠提示无法实现没有直接相机元数据或镜头模拟;视觉保真度与实体 VFX 舞台的对比未评估
机器人:仿真训练环境手工整理环境(仓库、厨房、办公室)——慢、贵,规模化后一致性差生成带深度、光照、几何和碰撞体网格的多样化照片级场景研究者 Yin 和 Joshi 用 Marble 做域随机化;导出物理准确的碰撞体用于交互相对专用仿真器(Isaac Sim、MuJoCo)的物理保真度上限未量化
建筑 / 设计:概念可视化草图 → 静态渲染 → 动画;空间体验推迟到项目后期将概念图输入 Marble;几分钟内得到可行走的 3D 环境Fenestra 和 Interior AI 案例显示,可生成可探索的 3D 概念用于客户评审无 BIM 数据导出;无结构保真度;户外 / 城市场景处理记录有限
游戏 / 浏览器游戏:环境创建手工环境美术 + 关卡设计;单资产制作成本高生成环境、提取碰撞体,并以少量手工工作接入 Unreal Engine 或 UnitySTARSPEED:浏览器中的 100M-splat 科幻环境;Unity 中的 Splat World VR 体验不生成程序化游戏逻辑或 AI NPC 行为;仅限静态环境
VR / AR:沉浸式空间体验手工建模或扫描环境;空间概念迭代慢生成世界、导出 splat,通过 Spark 导入 Meta Quest 3 或 WebXR 会话Splat World VR 体验已在 Meta Quest 3 测试;基于 splat 数据构建了自定义 C# 物理工具未基准测试低端头显上的独立 VR 性能;蜂窝网络延迟未记录

已记录收益基于公司发布的案例和开发者展示材料;与既有工作流的可衡量性能对比来自公司主张或作者报告,未经独立审计。限制由观察到的范围缺口推断。

[CE007, CE008, CE009, CE010, CE012, CE027]
FE002: Marble 客户工作流

端到端客户旅程,从初始输入到世界生成、交互式探索、编辑循环,再到导出至下游用例。

[CE001, CE002, CE004, CE005, CE011]

5.2 架构与技术栈

World Labs 的技术架构由四个紧密耦合的层组成:Marble 生成式世界模型、Spark 2.0 网页渲染器、World API 交付表面,以及导出与集成管线。每一层都建立在创始团队此前研究之上,让 World Labs 从已发表学术 IP 到已部署产品之间拥有异常直接的路径。 公司称 Marble 世界模型是一种多模态架构,能把可用输入信号——文本、图像、视频、空间布局——提升到统一的 3D 世界表征中。模型支持完整迭代编辑,包括元素级修改、扩展,以及把独立生成的世界组合起来。Marble 在导出视觉 splat 数据的同时导出碰撞体 mesh,这对集成基于物理的仿真至关重要。公司在 2026 年 6 月发布分类法博客,把世界模型分为在感知—行动循环中运行的渲染器、仿真器和规划器,显示其有意把架构路线图推进到智能体和仿真能力。 Spark 2.0 是 World Labs 的网页渲染器,使用 THREE.js 和 WebGL2 构建,支持桌面、iOS、Android 和包括 Meta Quest 3 在内的 VR 设备。World Labs 内部开发 Spark,是因为现有基于网页的 3DGS 渲染器一次只能渲染一个 3DGS 对象,缺少动态动画(「4DGS」),且无法在多设备上可靠运行。Spark 2.0 引入 Level-of-Detail 流式系统,会根据用户视角自适应调整 3DGS 细节并实时流式传输数据,让公里级环境能在移动设备和低带宽连接上运行。该渲染器支持实时编辑、重新布光,以及用于动态 splat 效果的 shader graph。 3D Gaussian Splatting(3DGS)是 Marble 和 Spark 使用的核心场景表征,它把场景表示为数百万个半透明彩色椭球体(「splats」),并通过 alpha compositing 排序渲染。该技术发表于 ACM TOG 2023,可在 1080p 分辨率下达到 30+ fps 的实时渲染,这是交互式浏览器和 VR 交付的关键门槛。联合创始人 Ben Mildenhall 在 ECCV 2020 发表原始 NeRF 论文,奠定了支撑现代神经渲染的神经辐射场表征。联合创始人 Christoph Lassner 发表 Pulsar(arXiv 2004.07484),这是一种基于球体的可微渲染器,速度比竞争方法快数个数量级。这些基础工作直接落在 Marble 和 Spark 中,让产品相对基于公开实现搭建的团队拥有第一性原理 IP 优势。[CE013, CE014, CE015, CE016, CE017, CE018]

技术与运营架构表
层 / 组件角色关键依赖主要风险
Marble 世界模型(核心)将多模态输入转换为可导航的 3D 世界表征云端 GPU 算力(供应商未披露);大规模 3D 训练数据架构、训练数据策略和模型质量均未披露;缺少独立基准
Spark 2.0(Web 渲染器)在浏览器中以 LoD 流式加载实时渲染 3DGS;交付 VRTHREE.js;WebGL2 浏览器 API;用于 splat 流式加载的云端数据存储WebGL2 废弃风险;THREE.js 生态依赖;移动带宽约束
World API(REST 交付)异步世界生成端点;接受文本 / 图像 / 视频 / 全景输入Marble 模型推理后端;云基础设施;网络延迟无公开 SLA;定价未披露;速率限制和正常运行时间保证未知
碰撞体网格管线从 splat 世界生成可用于物理的碰撞几何,供仿真使用Marble 模型输出;下游物理引擎(Isaac Sim、MuJoCo、Unity、Unreal)未用真实仿真基准验证物理保真度;可变形 / 铰接物体存在缺口
导出管线(splat、mesh、video)将生成的 3D 世界转换为多种下游兼容格式格式标准(.spz、.rad、.glb);第三方工具兼容性格式互操作性未经独立测试;大型场景文件大小限制未记录
Marble Labs(门户和 SDK)面向案例、教程、文档和展示托管的开发者门户worldlabs.ai Web 基础设施;社区参与无正式开发者支持 SLA;社区贡献集成不由 World Labs 支持
NeRF / 3DGS 神经渲染骨干支撑 Marble 生成的场景表征和新视角合成联合创始人的学术研究(Mildenhall NeRF、Lassner Pulsar、3DGS ACM TOG 2023)技术已发表且被广泛实现,存在竞争复制风险
第三方集成层Unreal Engine、Unity、Isaac Sim、Blender、VR 头显——下游运行环境第三方平台 API 和插件生态;World Labs 导出格式支持插件维护负担;版本兼容;第三方废弃集成 API

架构根据公司博客、展示材料和案例文档重建。未获得独立架构审查或内部工程文档。依赖项由已发布的技术栈描述推断。

[CE013, CE014, CE015, CE016, CE022, CE023]
FE001: World Labs 产品架构栈

五层软件栈,从多模态客户输入开始,经 Marble 世界模型、Spark 2.0 渲染器、交付界面,延伸到下游集成环境。

栈层级根据公司博客、展示材料和案例研究推断。内部模型架构和渲染管线的准确细节未公开披露。

[CE013, CE014, CE015, CE016, CE022]

5.3 部署、集成与路线图

Marble 以浏览器原生产品形态部署,不需要专有桌面客户端。用户在 marble.worldlabs.ai 访问产品;无需特殊设置或安装工具。世界生成通过 World API 异步运行,API 接受标准 HTTP 请求并返回可导航的 3D 世界对象。这个设计让它很容易接入现代 Web 应用、游戏前端和仿真管线:创建世界变成一次可编程 API 调用,而不是一项 3D 管线工程任务。 展示和案例材料记录的第三方集成覆盖了广泛的下游环境。Volinga 插件演示了 Unreal Engine 集成,可导入 Marble 生成的 Gaussian splat 环境,用于实时渲染和场景精修。Splat World 和 Splat Raycasting 案例记录了 Unity 集成,其中自定义 C# 工具操控 splats,用于 VR 游戏玩法和机器人移动。NVIDIA Isaac Sim 集成记录在一个机器人案例中,Marble 生成的场景——包括碰撞体 mesh 导出——被用作机器人学习训练环境。AI 原生 3D 管线展示演示了 Marble 输入一个多系统编排管线,结合 fal、mesh 生成、声音合成和 Unreal Engine。 开发者工具正在 Marble Labs 社区中形成。开源 Third-Person Character Controller 模板展示了 Marble + Spark 2.0 + Rapier3D 物理集成,并以可 fork 的 GitHub 包提供。Collider Builder 工具让用户能在浏览器中直接在 splat 场景上绘制碰撞几何,并导出为 .glb。STARSPEED 浏览器游戏展示了由超过 100 million 个 Gaussian splats 组成的环境也能达到生产级交付,并通过 Colyseus 实现多人网络。 Marble 路线图可从 Marble Labs 展示管线中部分读出:从独立世界生成,推进到支持物理的交互、虚拟化身承载、多人环境和机器人仿真,暗示平台正向交互式和智能体用例扩张。2026 年 6 月的分类法博客专门讨论从渲染器到仿真器再到规划器的路径,释放出公司有意把世界模型延伸到在生成环境中行动的主动智能体。不过,公司没有公开承诺任何具体的前瞻功能时间线或发布计划。[CE026, CE027, CE028, CE029, CE030, CE031]

路线图与发布阶段表
日期 / 阶段功能 / 里程碑状态影响来源
Dec 2, 2024Marble 预览版:基于图像 / 文本生成早期可在浏览器中导航的 3D 世界已发布验证概念;启动 beta 访问通道公司博客:worldlabs.ai/blog/generating-worlds
约 Nov 12, 2025Marble 正式可用:支持多模态输入(文本、图像、视频、全景图、布局)、交互式编辑、扩展 / 合并、 Gaussian splat / mesh / video 导出;Marble Labs 上线已发布产品已可商业使用;开发者社区枢纽上线公司博客:worldlabs.ai/blog/marble-world-model;案例研究:case-studies/bringing-marble-to-life
Jan 21, 2026World API 公开发布:用于世界生成的程序化 REST 接口已发布开发者和企业集成打通;尚未公开定价公司博客:worldlabs.ai/blog/announcing-the-world-api;服务条款:worldlabs.ai/terms-of-service
Mar 3, 20263D-as-code 文章:提出 3D 是物理 / 虚拟世界 AI 的通用接口已发表释放面向智能体世界模型的架构野心;界定竞争定位公司博客:worldlabs.ai/blog/3d-as-code
Apr 14, 2026Spark 2.0:面向大型 3DGS 世界的细节层级流式系统;支持 VR 交付已发布打开移动端和 VR 用例;支持公里级浏览器环境公司博客:worldlabs.ai/blog/spark-2.0
Jun 3, 2026世界模型功能分类:渲染器、模拟器、规划器已发表释放从渲染器延伸到模拟器和规划器能力的意图公司博客:worldlabs.ai/blog/taxonomy-of-world-models
未披露模拟器和规划器能力;real-to-sim 迁移;智能体式世界交互路线图意图(仅为公司表述的方向)可能扩展到机器人和具身 AI 平台;没有承诺时间表公司博客:worldlabs.ai/blog/taxonomy-of-world-models

日期来自博客发布日期和服务条款生效日期。前瞻性项目代表公司在已发表文章中表述的方向性意图;截至 2026 年 6 月,未公开披露已承诺的功能时间表或产品发布时间表。

[CE003, CE008, CE011, CE014, CE023, CE033]
FE003: World Labs 平台关键依赖图

World Labs 关键技术和合作伙伴依赖的有向图,展示 Marble、Spark、World API、云基础设施如何连接下游集成环境和学术 IP 基础。

[CE017, CE018, CE027, CE031, CE036, CE037]

5.4 差异化、IP 与数据策略

World Labs 的主要技术差异化建立在两根支柱上:创始团队直接参与了产品内嵌神经渲染技术的原创工作,以及内部构建了一个没有直接开源等价物的网页渲染器。两者叠加,形成一个难以被「把公开 baseline 拼起来」的团队复制的产品,因为最关键的优化——Spark 的多对象渲染、LoD 流式传输和动态 shader graph——都是为解决 World Labs 自身部署约束而开发,并未开源。 创始团队的学术发表记录与核心产品直接相关:Ben Mildenhall(NeRF、Mip-NeRF 360)、Christoph Lassner(Pulsar)和 Justin Johnson(PyTorch3D、3D 场景理解研究)共同创作了 Marble 和 Spark 所用神经渲染管线的基础技术。这些不是背景专利或防御性申请;它们正是其他 3D AI 公司赖以构建的论文。原作者亲自开发下一代,在模型架构和优化上形成有意义的阶段性优势。 Marble Labs 开发者社区充当分销与数据飞轮。每个公开 showcase 都展示一个新集成表面(Unreal Engine、Unity、NVIDIA Isaac Sim、VR、机器人),进而吸引更多开发者,并产生更多真实世界使用数据,可反哺模型训练和能力路线图。这种有机生态发展符合平台策略:World Labs 投入核心模型和渲染器,社区扩展集成表面。 一个不确定领域是训练数据策略。World Labs 未披露 Marble 使用哪些 3D 数据集训练,训练数据获取受何种数据许可策略约束,或存在哪些专有数据积累优势。Objaverse-XL(来自 Allen AI 的 10M+ 对象)和 Infinigen 程序化生成器这类大规模 3D 数据集,代表了世界级 3D 生成所需的训练基础设施类型;但 Marble 的具体数据管线仍未披露。研究中未发现独立基准把 Marble 输出保真度与竞争系统比较,因此现阶段无法进行第三方质量评估。[CE037, CE038, CE039, CE040, CE041, CE042]

FE004: 产品能力与成熟度矩阵

对十二项 Marble 能力做成熟度和证据评估,显示哪些已普遍可用、哪些只有部分文档,以及证据缺口仍在哪里。

成熟度评级是截至 2026 年 6 月基于公开文档做出的编辑判断。没有独立基准或第三方审计验证这些评级。“已展示”表示有社区展示证据,但 World Labs 未正式标注为发布。

[CE001, CE004, CE005, CE014, CE022, CE033]

5.5 信任、安全、安保与合规

World Labs 发布了三份面向公众的政策文件,定义其信任与合规姿态:服务条款(2026 年 1 月 21 日生效)、隐私政策(2025 年 11 月 12 日生效)和可接受使用政策(AUP,最后修改于 2025 年 11 月 12 日)。AUP 对企业尽调最具操作相关性:它明确禁止生成儿童性虐待材料、设计或协助大规模杀伤性武器、开发用于未授权监控或关键基础设施攻击的工具,以及在未经同意情况下制造伤害个人的欺骗性合成媒体。这些禁令可按 TOS 执行,并可能导致账户终止。 相对企业软件规范,安全姿态明显有限。worldlabs.ai/security 上的安全页面描述了一个负责任披露计划,要求报告发送至 security@worldlabs.ai,但明确表示 World Labs 目前没有公开漏洞赏金计划,也不保证对漏洞报告提供补偿。公开文档没有提到渗透测试频率、基础设施安全架构、数据加密标准或访问控制政策。截至 2026 年 6 月,公司网站上未发现 SOC 2 Type II、ISO 27001 或 AI 专项认证(例如符合 NIST AI RMF 或 EU AI Act 条款)。 隐私政策确认会收集账户信息、使用数据、提示输入、生成内容互动,以及来自第三方集成的数据。该政策没有说明数据保留期限,也没有说明用户提交的图片或提示是否用于模型训练;对企业客户而言,这是关键问题,因为他们的专有设计或仿真输入可能会影响模型。 从监管角度看,研究中未发现与 World Labs 产品相关的执法行动、监管问询或诉讼。NIST 的 AI Risk Management Framework(AI RMF)是负责任 AI 部署的参考标准,而 World Labs 当前文档没有声称符合 AI RMF,也没有描述如何映射到该框架。随着公司进入企业垂直领域——尤其是机器人和建筑——企业采购流程会要求比当前更结构化的安全与合规文档。[CE045, CE046, CE047, CE048, CE049, CE050]

信任、质量与合规表
控制 / 政策 / 认证状态范围缺口 / 尽调要求
Acceptable Use Policy(AUP)已发布(最后修改 Nov 12, 2025)worldlabs.ai 和 marble.worldlabs.ai;所有 API 和订阅用户未描述执行机制和下架 SLA;无透明度报告
Terms of Service(服务条款)已发布(生效 Jan 21, 2026)所有注册用户和 API 客户未观察到企业数据处理附录(DPA);未发布 GDPR 处理者条款
Privacy Policy已发布(生效 Nov 12, 2025)用户账户数据、使用数据、提示输入、生成输出未说明留存期限;未披露提示用于训练数据的规则
负责任安全披露已在 worldlabs.ai/security 发布页面所有 World Labs 系统和服务无漏洞赏金;无 CVD 流程时间表;无致谢计划
SOC 2 Type II未观察到N/A企业部署前索取 SOC 2 报告或同等材料
ISO 27001未观察到N/A采购时索取信息安全管理认证材料
NIST AI RMF 符合性未声称N/A — NIST AI RMF 属自愿指南未见与 AI RMF 风险类别的映射;受监管企业买方应纳入尽调路径
EU AI Act 合规未声称 / 未知作为通用 AI 模型,可能落入适用范围审查 EU AI Act 适用性;向供应商索取合格评定路线图

认证状态基于截至 2026 年 6 月 worldlabs.ai 面向公众的文档。没有某项认证并不证明不合规,只说明 World Labs 尚未公开声称具备该认证。企业尽调应纳入供应商安全问卷。

[CE045, CE046, CE047, CE048, CE049, CE050]

5.6 图表资料

Chapter 06

06客户

6.1 客户分层与买方画像

World Labs 的可触达客户群可分为三个细分市场,差异在于买方如何与产品交互、支付什么,以及提取什么价值。 个人创作者细分市场包括电影人、数字艺术家、建筑师、室内设计师、游戏创作者和 VR/XR 开发者。他们通过自助订阅层级使用 Marble,价格从 Free 到 Standard($20/month)、Pro($35/month)和 Max($95/month)不等。这些客户同时是买方、用户和付款方。已发布案例记录了 Tim Simmons(Theoretically Media,170,000+ YouTube 订阅者)和 Joshua Kerr 等电影人把 Marble 当作虚拟片场或虚拟制作工具;建筑师和室内设计师用 Marble 在正式建模前走进概念;Daniel Skaale 这样的 VR 开发者用 Marble 生成的 Gaussian splat 环境构建交互式 Unity 体验。商业使用权从 Pro 层级开始,把可防御的经常性收入集中到更高层级。各层级总订阅用户数未公开披露。 AI 原生平台和 API 集成商细分市场由一类公司组成:它们通过 World API 把 World Labs 的世界生成能力嵌入自身产品,成为中间买方,终端用户则间接使用 Marble。该细分市场的收入乘数潜力最高,因为每个集成都把 World Labs 的有效触达扩展到集成商既有用户群,而不需要成比例增加直接获客成本。具名集成商包括 OpenArt(推出 OpenArt Worlds,作为持久 3D 环境产品功能)、Magnific(为服务其大型创意平台用户群的 3D Scenes 工具集成 Marble)、Rosebud AI(游戏创建管线)、VIVERSE(基于浏览器的 3D 交互世界)、Lightcraft/Beeble(虚拟制作)和 Escape(2D 到 3D 交互媒体)。API 定价未公开披露;合同金额和收入分成未知。 机器人和具身 AI 研究细分市场由学术研究人员和工程团队组成,他们把 Marble 生成环境用作机器人学习管线的合成训练数据。记录中的例子包括研究人员 Hang Yin(BEHAVIOR-1K/OmniGibson)和 Abhishek Joshi(RoboSuite/Infinigen-Articulated),他们将 Marble 与 NVIDIA Isaac Sim、MuJoCo 和 RoboSuite 搭配使用。该细分市场技术价值高,但现阶段主要由研究小组构成,而不是商业运营方;能否转化为付费企业合同尚未确认。 第四个新兴细分市场是战略企业伙伴层级,由 Autodesk 的 $200 million 投资以及其明确聚焦媒体和娱乐工作流协作来锚定。建筑事务所(Fenestra、xFigura、SHoP Architects)和虚拟制作工具(Preview)补足该细分市场,但所有企业关系的商业条款均未披露。[CU001, CU019, CU022, CU023, CU025, CU041]

按买方、用户、付费方和用例划分的客户分层
客群买方 / 用户 / 付费方主要用例规模代理指标收入 / 战略价值证据缺口
个人创作者买方 = 用户 = 付费方(自助订阅)虚拟电影制作、建筑可视化、游戏创作、VR/XR、AI 艺术不确定:未披露。根据创作者客群定价推断为小 / 中型用户基数单用户价值低至中;订阅 MRR 和套餐结构未披露订阅者数量、套餐结构、转化率和流失率未披露
AI-native 平台 / API 集成方平台公司(买方、付费方);集成方产品的终端用户(用户)将世界生成嵌入创作工具、游戏、互动媒体、AR/VR 平台6+ 个具名集成;集成方平台声称有数百万用户(例如 Magnific)战略价值高 — 每个集成都放大有效触达;API 定价未披露API 定价、调用量、单项集成合同价值和收入分成未披露
机器人 / 具身 AI 研究研究机构或创业公司;付费方 = 实验室预算或科研经费机器人学习的合成训练数据;仿真环境生成;real-to-sim 迁移已记录 2 位具名研究人员;更广泛研究社区规模未知中高价值 — 高算力、高风险工作流;能否转为付费未知该客群付费客户数;从研究试用到商业合同的转化未披露
战略企业伙伴企业(买方、付费方);内部设计 / 制作团队(用户)将世界生成集成到 3D 设计、M&E、AEC、虚拟制作工作流1 个具名战略投资方兼伙伴(Autodesk $200M);具名建筑事务所:Fenestra、xFigura、SHoP Architects若商业条款成型,潜在价值很高;据 TechCrunch 报道,目前仍属探索Autodesk 合同条款、收入分成、集成路线图和已部署用户数未披露

客群规模估计来自已发布案例研究和集成方平台描述的推断;World Labs 没有公开订阅者数量或 API 客户数量。 Magnific 的「数百万用户」来自 World Labs 案例研究对 Magnific 平台规模的引用,并非 Marble 专属用户。

[CU001, CU003, CU004, CU005, CU019, CU022]
FU001: 客户旅程图——客群、触点与扩展循环

World Labs 三个主要客群的客户旅程,从初始认知到试用、采用和扩展,突出创作者、API 集成方和机器人研究者的不同路径。

旅程路径根据公开案例研究叙事和 World API 公告构建。公开来源没有转化率、阶段时长或流失数据。所有路径均为定性推断。

[CU001, CU002, CU009, CU014, CU019, CU023]

6.2 具名客户证明——集成、协作与 Showcase 证据

2025 年 11 月至 2026 年 5 月,World Labs 发布了九个案例研究,并记录了多个开发者 showcase 项目,形成一组具名协作证据。证据质量因集成而异:少数被描述为集成商产品中的已上线产品功能(OpenArt Worlds、Magnific 3D Scenes),其他则明确定位为创意实验、showcase 或研究协作,而不是规模化生产部署。所有案例都没有包含日活用户、API 调用量、错误率或客户满意度分数等量化结果指标;它们都是 World Labs 产出的定性叙事。 OpenArt 被描述为「全球领先的 AI 创意平台之一」,它使用 World Labs 的生成式 3D 世界模型推出 OpenArt Worlds,作为面向创作者群体的具名功能。该案例(日期为 2026 年 5 月 8 日)把集成描述为从单张图像生成持久、可探索的 3D 环境。这是证据集中质量最高的在线生产集成证据,并由 OpenArt 在 openart.ai 上的自有产品存在支撑。 Magnific 在自己的案例中称其服务「数百万」日活用户,包括设计师、营销人员和内容团队;它为其「3D Scenes」产品集成了 Marble。案例明确指出客户问题(AI 生成器缺少产品放置的空间控制),并把解决方案描述为一个 3D 场景构图工具。Magnific 在 magnific.com 的首页确认了该平台作为创意工具的规模。清晰问题陈述、具名产品和庞大既有用户群三者叠加,使其成为较强的集成证明点之一。 Lightcraft 与 Joshua Kerr、Beeble 的合作产出了一个虚拟制作用例:基于 iPhone 的工作流,把 Marble 生成的 3DGS 环境与 Lightcraft Jetset 基于 Emmy 获奖 Previzion 技术的 AR 合成结合起来。这是具名电影人 showcase,而不是平台部署;但它证明 Marble 能与专业工具一起用于真实制作场景。 Rosebud AI(通过自然语言创建游戏)、VIVERSE(HTC 基于浏览器的 3D 世界)和 Escape(2D 到 3D 交互媒体)被记录为研究协作和早期生产实验。Rosebud 案例描述了一个原型游戏;VIVERSE 案例描述实验性世界构建;Escape 在 World API 公告中被提及,称其已经使用该 API 把电影转成可导航 3D 环境。这些从 showcase 到早期生产不等。 机器人细分市场协作涉及具名研究人员(Hang Yin 和 Abhishek Joshi),并记录了与 NVIDIA Isaac Sim、MuJoCo 和 RoboSuite 的集成。World Labs 关于「2-simulate」的案例描述了 Marble 如何加速机器人训练环境生成。这些是研究协作,不是商业合同。 SHoP Architects 是建筑细分市场中唯一具备企业级别的具名客户;World API 公告引用了其 testimonial,内容是缩小设计师与客户之间的沟通差距。公告没有描述合同或部署规模。 Marble Labs showcase(worldlabs.ai/labs)列出至少 18 个开发者和艺术家项目,包括 Matt Workman 的 AI 原生 3D 管线(结合 fal 和开源 IMAGE-BLASTER 工具包)、一个房地产可视化项目,以及 Volinga/Unreal Engine 集成。这些 showcase 条目构成了正式案例研究之外的平台采用开发者信号证据,但不代表商业关系。[CU002, CU003, CU004, CU005, CU006, CU007]

具名客户证明表
客户 / 伙伴客群部署 / 用例状态具名成果或引述证据限制
OpenArtAI-native 平台 / API 集成方OpenArt Worlds:将单张图像转成创作者可持续导航的 3D 环境已发布产品功能(2026 年 5 月 8 日案例研究)World Labs 案例研究:「将世界生成变成创作者、电影制作人和讲故事者都能使用的工作流」未披露 OpenArt Worlds 功能的用户数;所有证据均由 World Labs 自报
MagnificAI-native 平台 / API 集成方3D Scenes:面向产品摆放和营销拍摄的 3D 场景构图与摄影工具已发布产品功能案例研究:「数百万设计师、营销人员和内容团队每天都在使用它」(指 Magnific 平台)Magnific 内 Marble 专属用户数未披露;案例研究由 World Labs 撰写
Lightcraft / BeebleAI-native 平台 / API 集成方虚拟制作 — Marble 3DGS 环境 + Lightcraft Jetset 在 iPhone 上做 AR 合成具名电影制作人展示(2025 年 11 月 12 日案例研究)案例研究展示电影制作人 Joshua Kerr 用 Marble + Lightcraft Jetset 在童年街道上创作僵尸电影单一具名电影制作人;未记录规模化制作部署
Rosebud AIAI-native 平台 / API 集成方游戏创作管线:Marble 世界生成集成进 Rosebud 的自然语言游戏构建器带原型的展示合作(2025 年 11 月 12 日案例研究)Rosebud 团队引述:「借助 Marble,我们可以创造神奇、可探索的世界,并赋能更广泛的游戏制作者社区」仅为原型;无生产用户数或实时集成指标
VIVERSE (HTC)AI-native 平台 / API 集成方互动浏览器 3D 世界:Marble 3DGS 场景在 VIVERSE 生态中精修并变成交互式实验性合作案例研究:「AI 应该增强创作工作流,而不是取代人的愿景」探索性表述;未披露生产指标或活跃用户
EscapeAI-native 平台 / API 集成方互动媒体:用 World API 将 2D 电影转成可导航的 3D 环境,并支持社交观看早期生产使用(World API 公告提及,2026 年 1 月 21 日)World API 博客:「像 Escape.ai 这样的平台正在用 World API 将 2D 电影转成可导航的 3D 环境」仅在 API 公告中提及;无专门案例研究;无指标
NVIDIA Isaac Sim 研究者机器人 / 仿真研究合成训练数据:Marble 生成的 3D 场景用于 Isaac Sim,服务机器人学习和 real-to-sim 迁移研究合作(案例研究 2-simulate)World Labs 案例研究记录 Marble 与 Isaac Sim、MuJoCo、RoboSuite 集成研究合作不是商业合同;未确认付费关系
Hang Yin / Abhishek Joshi(机器人研究者)机器人 / 仿真研究具身 AI:Marble 世界生成用于可扩展、物理准确的 3D 场景,服务机器人仿真数据具名研究合作(案例研究 1-robotics)案例研究记录 Marble 与 OmniGibson、RoboSuite、Infinigen-Articulated 框架配合使用个人研究人员,不是商业客户;使用规模未说明
SHoP Architects战略企业伙伴(建筑)建筑可视化:World API 集成,用于和客户沟通设计早期采用 / 证言(World API 公告)引述:「几秒钟内增加第三个维度令人难以置信。从图像到空间体验,这项技术将大幅缩小建筑设计师与客户之间的沟通差距。」World Labs 营销材料中的单条引述;无合同规模或工作流部署细节
Daniel Skaale(VR 开发者)个人创作者Splat World:Unity VR 第一人称体验,使用 Marble Composer 生成 3DGS 环境,并配套自定义物理和光照工具开发者展示(2025 年 11 月 12 日案例研究)案例研究:开发者构建了一套自定义 C# 工具,可在 Unity 中把 Marble splats 当作动态材质交互个人开发者展示;无商业部署或活跃用户数

所有行均由 World Labs 通过案例研究或 World API 公告博客自报。任何具名客户都没有第三方审计、独立评价或生产规模指标。 状态标签反映现有证据支持的最强表述;实际生产部署成熟度可能不同。

[CU002, CU003, CU004, CU006, CU007, CU008]
FU002: 采用与部署漏斗——从发现到生产

定性采用漏斗展示从 Marble 认知到生产规模部署的相对阶段规模,依据案例研究证据和市场信号;所有阶段规模都是示意性相对比例,不是披露用户数。

阶段数值(100、40、12、4、1)仅为示意性相对比例——不代表已披露用户数、订阅者数或 API 调用量。World Labs 未披露任何量化漏斗指标。漏斗形态根据广泛市场认知与案例研究中记录的具名生产集成之比推断。

[CU015, CU016, CU018, CU020, CU024, CU039]

6.3 采用轨迹与部署状态代理指标

截至 2026 年 6 月,World Labs 未披露任何量化采用指标——订阅用户数、API 调用量、月活用户、账户数或使用率。以下是可用的间接信号。 案例研究发布速度可作为采用代理指标。2025 年 11 月 12 日(Marble GA)至 2026 年 5 月 8 日之间,公司发布了九个案例研究——约每月 1.3 个。这个速度显示,公司有活跃的客户成功或合作伙伴职能在识别并讲述集成故事,符合一家早期商业化公司专注积累社会证明的状态。 World API 于 2026 年 1 月 21 日上线,约在 Marble GA 后十周。公告记录了 API 上线时已经投入使用的三个垂直场景——游戏 / 沉浸式媒体(Escape)、建筑 / 设计(Fenestra、xFigura、Preview)和机器人 / 仿真(Lightwheel)——这意味着公开发布前,精选伙伴已获得预发布 alpha 或 beta API 访问权。这更像受控爬坡,而不是开放 beta 发布。 到 2026 年 6 月,Marble Labs showcase 已增长到至少 18 个项目。这些项目被明确表述为实验性(「blueprints for building」),而不是生产部署;但它们覆盖虚拟制作管线、VR 引擎、游戏引擎、房地产可视化和物理交互艺术,构成有意义的开发者社区信号。 Autodesk 的 $200 million 战略投资(2026 年 2 月 18 日)是最重要的外部采用信号:一家大型 3D 软件既有玩家,对 World Labs 的世界生成能力与其企业设计和娱乐客户群的商业相关性下了大额战略赌注。World Labs 在 Autodesk 实际产品套件中的采用仍是未来可能,而非已有证据。 截至 2026 年 6 月,外部没有关于 Marble 生产部署质量或客户满意度的独立第三方评测、评分或分析师报告。[CU015, CU016, CU017, CU018, CU020, CU021]

客户增长与采用轨迹代理指标
指标数值 / 状态日期来源置信度影响缺失分母
Marble 正式可用发布November 12, 20252025-11-12World Labs 官方商业客户获取期开始n/a
World API 发布January 21, 2026(程序化世界生成的公开接口)2026-01-21World Labs 官方(API 公告博客)支持平台集成方;标志 API 渠道获客启动未披露 API 定价,未披露 API 客户数
具名平台集成(案例研究)6 个具名平台集成方:OpenArt、Magnific、Rosebud、VIVERSE、Lightcraft、Escape2025-11-12 至 2026-05-08World Labs 案例研究语料为开发者 / 平台采用提供社会证明;不是活跃 API 使用或付费合同的度量API 集成方总数未知;部分可能仍处商业化前或不活跃
机器人 / 仿真研究合作2 位具名研究人员;文档记录 Isaac Sim、MuJoCo、RoboSuite 为兼容模拟器2025-11-12 至 2026-01-21World Labs 案例研究(1-robotics、2-simulate)在机器人客群建立技术可信度;商业转化未知使用 Marble 的机器人研究团队数量未披露
案例研究发布速度GA 后约 7 个月发布 ~9 篇案例研究(≈1.3 篇 / 月)2025-11-12 至 2026-05-08World Labs 案例研究页面(计数)伙伴 / 客户成功职能活跃;符合早期商业化阶段特征案例研究由公司自报筛选;不代表完整客户基数
Marble Labs 展示项目截至 2026 年 6 月,worldlabs.ai/labs 记录 ≥18 个项目2026-06-18World Labs Labs 展示中心开发者社区信号;不是商业部署开发者社区总规模未披露
Autodesk 战略投资$200M 投资,为 $1B Series B 的一部分;初期聚焦娱乐用例2026-02-18TechCrunch(105)最大外部采用背书;释放其与企业 3D 工作流具备商业相关性的信号未披露 Autodesk 产品集成时间表或收入承诺
独立客户评价截至 2026 年 6 月,未识别到任何评价(无 G2、Capterra、Gartner Peer Insights 评分)2026-06-18缺乏证据限制对产品质量和满意度的独立验证n/a:该缺口无法从公开来源弥补

所有数值都是从公开公告和案例研究得出的代理证据。没有已披露的订阅者数量、API 调用量、MAU 或收入指标。 置信度评级反映来源质量,而非客户规模的确定性。

[CU015, CU016, CU017, CU018, CU020, CU021]
FU003: 客户证明矩阵——证据质量与部署成熟度

具名集成和合作按证据质量(仅自述 / 合作伙伴网站佐证 / 已上线生产功能)和部署成熟度(展示 / 研究 / 试点 / 已上线功能)定位,显示 World Labs 具名客户群的证明质量分布。

部署成熟度分类基于 World Labs 案例研究和 World API 公告中的定性描述。“已上线产品功能”要求集成方平台上可见具名产品。没有独立验证的生产指标或活跃用户数。

[CU002, CU004, CU006, CU007, CU008, CU009]

6.4 留存、耐久性、扩张与集中风险

公开来源完全缺少 World Labs 客户关系的留存和耐久性证据。任何订阅层级、API 客户或具名集成都没有披露 NRR、GRR、流失率、合同续约数据、cohort 分析或客户满意度分数。缺少这些指标,符合一家早期商业化公司的状态——截至 2026 年 6 月,Marble 公开可用仅七个月——但这意味着具名集成的耐久性无法从公开证据评估:OpenArt、Magnific 或 Rosebud 究竟仍是活跃用户,还是一次性实验性采用者,外部看不清。 最接近的留存代理指标是:(1)2026 年 5 月 8 日 OpenArt 案例,发布于 Marble 2025 年 11 月 GA 约六个月后,至少显示截至该日期仍在持续互动;(2)World API 于 2026 年 1 月上线时,公告中记录的具名伙伴已经立即集成,说明是发布前承诺,而不是发布后流失;(3)Marble Labs showcase 在可观察期间继续收到新项目。这些都不能替代留存指标。 扩张信号存在,但仍早。World API 是关键扩张机制:基于 API 构建的平台集成商,会把 World Labs 的世界生成能力暴露给自身用户群,形成乘数效应增长,而不需要 World Labs 成比例付出直接获客成本。如果 OpenArt 和 Magnific 的集成能在各自用户中获得牵引力,凭借其公开宣称的大用户基数,它们将是扩张潜力最高的账户。地理扩张,以及进入 AEC、国防仿真或数字孪生应用的垂直扩张,仍是未来状态而非已被证实。 集中风险实质存在。截至 2026 年 6 月,具名客户基础集中在创意 / 媒体行业(八个具名集成中有六个是游戏、电影或创意平台公司)。机器人和仿真细分市场由研究协作代表,而不是商业企业账户。AEC 细分市场主要由早期采用者和一位被引用的建筑师代表。如果创意媒体需求走弱,或竞争工具(Luma AI、NVIDIA Cosmos)在创作者市场拿到更大份额,World Labs 的短期客户集中会带来显著收入风险。 企业采购摩擦是 AI 平台公司普遍存在且已有记录的结构性风险。Deloitte 的企业 AI 采用分析指出,合规复杂度、员工准备度和监管不确定性会阻碍采用、拉长销售周期并提高获客成本。对 World Labs 的目标企业垂直领域——AEC、机器人、媒体——这些障碍尤其突出:建筑事务所必须评估生成环境中的数据所有权和 IP;机器人团队必须验证用于安全关键用途的仿真保真度;媒体工作室必须确保版权和内容来源合规。World Labs 的 AUP 和服务条款建立了法律框架,但无法降低这一阶段 AI 工具固有的企业采购摩擦。[CU020, CU021, CU026, CU027, CU028, CU029]

留存、重复使用和满意度 — 证据与缺口
指标数值 / 状态客群置信度尽调请求
净收入留存率(NRR)null — 未披露所有客群n/a向管理层索取 Free-to-Pro 升级路径和 API 客户续约的季度 cohort NRR
总收入留存率(GRR)null — 未披露所有客群n/a索取 Pro / Max 订阅层级的年度流失率和总留存数据
月流失率null — 未披露所有客群n/a从内部分析中索取按订阅层级(Free、Standard、Pro、Max)划分的月流失率
重复世界生成活动(代理指标)叙述中存在:OpenArt Worlds 被表述为创作者会反复返回的「持久」环境;Marble 案例研究在 GA 6 个月后仍发布,暗示持续参与创作者 / 平台集成方低(仅叙述代理)获取 DAU/MAU 或世界生成会话数据,以确认重复使用模式
客户满意度评分(NPS / G2 / Capterra)null — 截至 2026 年 6 月,未在 G2、Capterra 或 Gartner Peer Insights 识别到公开评价所有客群n/a监测 G2 和 Capterra 是否出现 Marble 条目;向管理层索取 NPS 分数
合同期限 / 续约条款按标准 SaaS 结构,订阅层级为逐月订阅;企业 API 合同条款未披露平台集成方 / 企业低(由定价页结构推断)向管理层索取企业 API 最低承诺和续约条款

所有留存指标均为 null;截至 2026 年 6 月,World Labs 没有公开留存、流失或满意度数据。重复使用的叙述代理基于 OpenArt 案例研究中的定性表述。置信度评级反映缺乏定量证据,而非存在留存不佳的负面证据。

[CU020, CU021, CU039]
扩张驱动因素与集中度风险评估
因素描述影响评估尽调路径
World API 乘数效应平台集成方(OpenArt、Magnific)把 Marble 暴露给自身数百万终端用户;每一次集成都能放大有效触达, 又不需要 World Labs 按比例增加直接获客成本高正向影响 — 如果集成方获得用户牵引,World Labs 将受益于嵌入式分发追踪 OpenArt Worlds 和 Magnific 3D Scenes 在各自平台内的采用;向管理层索取 API 调用量
Autodesk 合作扩张空间Autodesk 的 $200M 投资创造了潜在渠道,可进入其庞大装机基础覆盖的建筑、工程、施工和娱乐工作流潜在影响高 — Autodesk 的 3D 软件装机基础是重要分发跑道;时间表和商业条款未知索取 Autodesk 合作路线图和任何已承诺的产品集成里程碑;关注 Autodesk 投资者日披露
创意 / 媒体行业集中度8 个具名集成中有 6 个在游戏、电影或创意平台;机器人和 AEC 的商业代表性不足高风险 — 如果创意媒体 AI 工具需求走弱,或 Luma AI / NVIDIA Cosmos 抢走份额,短期收入会暴露寻找机器人仿真和 AEC 领域的企业合同,分散垂直行业集中度
单一具名企业锚点(Autodesk)Autodesk 是唯一具名的企业战略伙伴;其他具名客户均为较小的平台集成方或研究人员中等风险 — 现阶段企业可信度和分发高度依赖 Autodesk 合作关系确认 Autodesk 在 $200M 投资之外的商业承诺;识别第二、第三个企业设计伙伴
无已记录流失客户未公开披露已终止的集成或流失账户;但没有披露并不能证明零流失信号含混 — 可能说明留存强,也可能只是处于商业化前阶段,流失关系不会公开宣布向管理层索取已终止 API 试用和集成清单;与 World API 发布时的初始伙伴名单对比

扩张驱动因素评估来自公开案例研究、World API 公告以及 TechCrunch 对 Autodesk 合作的报道推断。公开来源没有已确认的收入扩张、 upsell 率或多年合同数据。

[CU007, CU009, CU016, CU019, CU024, CU028]
FU004: 留存证据缺口矩阵——按时间段的集成活动信号

六个具名 World Labs 集成在 GA 后三个时间段的留存证据缺口矩阵;单元格显示定性活动信号(发布案例研究、API 公告提及,或无公开信号),因为没有公开的数值留存或流失数据。

单元格只显示是否存在已发布公开信号(案例研究、博客提及);不代表数值留存百分比。截至 2026 年 6 月,World Labs 未披露 NRR、GRR、流失率或量化 cohort 数据。该图展示留存证据缺口,而不是实际留存表现。之所以使用矩阵类型,是因为没有可公开获知的数值留存数据;原计划的 cohort 类型需要 0–100 的留存百分比。

[CU020, CU021, CU039]

6.5 图表资料

Chapter 07

07风险

7.1 监管与法律风险

World Labs 正站在三条监管战线的交汇处:AI 专项立法(EU AI Act、California SB 1001)、知识产权法(AI 训练数据与输出的版权),以及出口管制(US Export Administration Regulations)。欧洲议会于 2024 年 3 月以 523 票赞成通过 EU AI Act,建立了按风险分层的框架:超过既定算力门槛的通用 AI(GPAI)模型,需要承担更高的透明度和安全义务。按这一定义,World Labs 的 Marble 构成 GPAI 模型;其 AUP 已明确承认这一风险敞口,把 EU AI Act 下的受监管用途列为「高风险用途」。World API 一旦在欧洲市场放量,可能需要做合规评估、技术文档和第三方审计,带来合规成本,也可能拖慢部署。 知识产权方面,US Copyright Office 于 2025 年 5 月以预发布形式公布其 AI 报告第 3 部分,审查训练数据抓取的合法性,以及 AI 生成输出能否获得版权。第 2 部分(2025 年 1 月)已确认,缺乏人类作者身份的输出在美国不能获得版权。World Labs 尚未披露 Marble 训练数据的授权基础;这些数据很可能包括网页抓取图像、3D 扫描和视频,均可能面临相互竞争的权属主张。企业客户要把生成的 3D 环境纳入商业工作流,需要干净的 IP 链条;任何针对 Marble 生成内容的下游侵权索赔,都会影响客户信心,并按服务条款带来赔偿责任。 Bureau of Industry and Security 的 Export Administration Regulations(EAR,15 CFR 730-774)适用于被归类为军民两用的 AI 模型权重和相关技术。World API 的跨境 API 访问可能带来出口管制风险。NIST 的 AI Risk Management Framework 提供自愿性指引,但没有法律强制力,意味着企业客户只能依赖 World Labs 自身的合规姿态——而该姿态尚未公开披露。截至 2026 年 6 月,World Labs 尚未宣布任何第三方安全认证(SOC 2、ISO 27001),也未宣布面向企业部署的监管预批准。 [CR001, CR002, CR003, CR004, CR005, CR006]

监管 / 法律风险登记表
规则 / 许可 / 领域司法辖区状态(2026 年 6 月)可能性严重性缓释措施剩余风险暴露尽调路径
EU AI Act — GPAI / 高风险分类欧盟已生效(2024 年 3 月);实施推进中;GPAI 义务在 2025–2026 年分阶段落地严重AUP 限制高风险用途;需要合规计划高 — 第三方审计义务和透明度要求尚未公开确认已落实确认 Marble 算力是否超过 GPAI 门槛;取得关于分类的法律意见;审查 EU AI Office 指引
美国版权 — 训练数据与 AI 输出所有权美国积极审查中;Part 3 于 2025 年 5 月预发布(训练数据);尚无最终规则未披露训练数据许可计划;TOS 对侵权输出免责声明高 — 暴露于训练数据诉讼;企业客户无法依赖干净 IP索取训练数据来源披露;核实 Marble 输出是否包含赔偿
美国出口管理条例(EAR / BIS)美国已生效;AI 权重可能具备双重用途;跨境 API 访问带来暴露未披露出口管制合规计划;API 面向全球开放中 — World API 的跨境访问可能需要 BIS 出口许可审查确认模型权重是否归入 EAR Commerce Control List;取得出口管制律师意见
隐私 / GDPR / CCPA — 空间数据处理欧盟 / 加州 / 全球2025 年 11 月隐私政策;覆盖空间与地理空间处理;未公开披露 GDPR DPA已有隐私政策和数据处理条款中高 — 在未确认 DPA 的情况下,GDPR 下处理 3D 场景和地理空间数据核实是否提供符合 GDPR 的 Data Processing Addendum;确认 CCPA 合规计划
美国联邦 AI 监管(NIST RMF / FTC 执法)美国自愿采用 NIST AI RMF;FTC 的欺骗性营销执法权适用;尚无行业 AI 法案AUP 禁止欺骗性用途;TOS 采用加州管辖法律和仲裁条款低中 — FTC 可能审查 Marble 空间智能能力的宣传跟踪 FTC AI 执法;确保能力宣传有证据支撑;审查 TOS 仲裁条款可执行性

行按严重程度从高到中排列。可能性和严重程度是分析师基于适用监管文本及公司披露文件作出的判断; 未取得独立法律意见。「状态」反映截至 2026 年 6 月运行日的公开来源。

[CR001, CR002, CR003, CR004, CR005, CR006]
FR001: 风险严重度热力图——World Labs(2026 年 6 月)

按可能性和影响定位 World Labs 识别出的十项主要风险,并按披露缓释措施后的剩余暴露着色。行是影响等级(顶部 = 关键);列是可能性等级。

可能性和影响是截至 2026 年 6 月基于公开证据做出的定性分析师判断。没有可用的公司内部数据。

[CR001, CR003, CR004, CR009, CR014, CR019]

7.2 运营与技术风险

World Labs 的核心产品 Marble 是一条算力密集型生成管线,借助异步处理架构,把文本、图像、视频或多视角输入转换为 3D Gaussian splat 环境。每一次 World API 调用都会触发一次完整的 3D 世界生成任务,资源消耗显著高于文本或图像推理。规模化之后,推理需求会变成运营风险:如果 GPU 供应受限,World Labs 可能出现服务降级或限流,直接压住开发者采用。公司尚未公开披露服务级别协议(SLA)或正常运行时间保证。 安全方面,World Labs 通过 security@worldlabs.ai 维护负责任披露计划,但没有公开漏洞赏金,因此外部研究人员主动挖掘漏洞的经济激励有限。缺少已发布的第三方渗透测试、SOC 2 证明或 ISO 27001 认证,企业买家只能依赖信任,而不是独立审计过的保证。这个缺口在运营上很关键,因为 World API 会处理客户上传的图像、视频和 3D 资产,其中可能包括建筑、国防和工业仿真工作流中的敏感专有场景。 面向机器人和具身 AI 应用时,3D 生成领域长期卡在仿真到现实的保真度问题上。关于具身 AI 3D 生成的 ArXiv 综述(arXiv:2604.26509)把物理标注有限、几何真实感缺口、评估碎片化,以及持续存在的 sim-to-real 鸿沟列为主要瓶颈。对 World Labs 的机器人用例来说,这不是假设:如果 Marble 生成环境的物理精度不足以训练机器人策略,高价值企业用例就会碰到保真度天花板,而拥有物理原生仿真平台的竞争对手(带 Cosmos 的 NVIDIA Isaac Sim)可能先到达。Deloitte 的 2025/2026 AI 采用调查显示,监管合规和劳动力准备度是企业采用 AI 的主要障碍——两者都直接关系到 World Labs 的目标客户。 [CR011, CR012, CR013, CR014, CR015, CR016]

运营 / 质量 / 安全风险清单
失效模式可能性严重程度缓释成熟度剩余暴露未解决缺口
计算成本爆炸 / GPU 供应受限严重高 — 推理需求随用户增长超线性放大;未披露 SLA 或成本上限未披露 GPU 采购合同条款或替代算力策略
API 可靠性下降(宕机 / 限流)中高 — 未披露 SLA;异步流水线增加延迟风险公开资料中没有可用性承诺或事故历史
对抗性 / 有害输出生成中 — AUP 禁止有害用途,但未披露 API 层内容过滤能力未公开说明 API 输出的自动内容审核机制
机器人 / 具身 AI 的仿真到现实保真度缺口低中高 — 3D 生成文献指出,物理标注和仿真到现实瓶颈持续存在机器人案例研究由公司自行发布;Marble 物理仿真准确性未获独立验证
安全漏洞披露(无漏洞赏金)中高 — 未公开漏洞赏金;未披露第三方渗透测试或 SOC 2缺少安全证明会卡住企业采购;未发布安全审计

行按严重程度从严重到高排列。缓释成熟度:低=临时处理;中=已形成文档;高=已审计 / 认证。 剩余暴露反映截至 2026 年 6 月的公开证据。

[CR011, CR012, CR013, CR014, CR015, CR016]

7.3 合作伙伴、依赖与资本风险

World Labs 的合作伙伴生态形成了一种不寻常的依赖结构:两家主要硬件供应商 NVIDIA 和 AMD 同时也是股权投资者。两家公司都参与了 2026 年 2 月的 $1 billion 融资;Autodesk 也参与其中,投资 $200 million,并承担顾问角色。这组战略投资者在商业上有利——既传递可信度,也锁定供应关系——但也集中化了议价权。在竞争张力下(NVIDIA Cosmos 在机器人仿真中直接竞争 Marble),NVIDIA 可能利用供应杠杆或顾问接触,把 World Labs 推向有利于自有平台的结果。TechCrunch 报道确认,Autodesk 首席科学家把这项合作描述为「early days」,商业产品形态「尚未确定」,并明确指出数据共享不属于协议内容。如果 Autodesk 合作到 2027 年仍未产出联合商业工作流或集成产品,Autodesk $200 million 承诺背后的战略理由会变弱。 World Labs 尚未披露其是否运营自有 GPU 数据中心基础设施,还是依赖超大规模云服务商提供训练和推理算力。NVIDIA Cosmos 是一家拥有大得多算力访问权、且已有机器人仿真生态的公司开发的竞争性世界基础模型,在企业仿真细分市场构成直接替代威胁。DeepMind 的 Genie 2 进一步说明,World Labs 正在与算力预算几乎没有上限的组织竞争。 资本层面,Crunchbase 2026 年 Q1 数据显示,全球 AI 风险投资达到 $300 billion,但其中 65% 被四家公司拿走。如果空间 AI 无法在 2026–2027 年证明可触达收入,中型前沿 AI 初创公司的后续融资会收紧。Gartner Generative AI Hype Cycle 2026 暗示,一些生成式 AI 类别正接近膨胀预期峰值,随后的低谷可能压缩估值倍数,并限制 World Labs 未来融资窗口——公司尚未确认传闻中的 $5 billion 估值。 [CR019, CR020, CR021, CR022, CR023, CR024]

合作伙伴 / 依赖风险清单
依赖项交易对手角色集中度失效场景严重程度缓释剩余暴露
GPU / 算力供应NVIDIA, AMD推理和训练硬件;两家公司也是 Series B 股权投资者严重 — 未披露替代方案GPU 配给、定价杠杆,或投资者兼供应商的利益冲突严重投资人利益一致带来短期供应入口;未披露独立算力高 — 单层依赖且没有后备方案;投资者兼供应商的冲突未解决
分发与产品集成Autodesk战略顾问和 $200M 投资方;目标是接入 3D 工作流高 — Series B 最大单笔支票合作未成熟为联合商业产品;Autodesk 转向内部 3D AI若集成成功,Autodesk AEC 用户群提供大规模分发面中高 — 截至 2026 年 Q1,集成形态「尚未确定」
竞争平台(NVIDIA Cosmos)NVIDIA投资者,同时是世界模型 / 物理 AI 仿真领域的直接竞争者高 — NVIDIA Cosmos 瞄准相同的机器人和仿真用例NVIDIA Cosmos 取代 Marble,成为 Isaac Sim 用户首选仿真环境World Labs 通用世界模型的广度形成差异化;投资关系创造渠道高 — NVIDIA 具备全栈优势(硬件、软件、开发者生态)
云基础设施 / 推理托管AWS / Azure / GCP(未确认)Marble API 推理底层算力(若不是自有数据中心)未知 — 未披露基础设施提供商超大云厂商涨价、宕机或锁定效应阻碍利润率改善未缓释 — 未披露基础设施依赖高 — 无法评估;需要尽调
领投方集中度(a16z)Andreessen Horowitz种子轮领投;Series B 继续跟投;推定拥有董事会权利高 — a16z 是唯一横跨两轮的投资者a16z 优先支持其投资组合中的竞争性空间 AI 初创公司Series B 投资者基础多元(Autodesk、Fidelity、AMD、NVIDIA、Sea),降低单一投资者依赖低中 — a16z 持续在场提供声誉锚点,但也带来集中度

行按严重程度从严重到中排列。交易对手集中度基于截至 2026 年 6 月披露的融资和合作来源分类。 由于未披露,云基础设施集中度无法核实。

[CR019, CR020, CR021, CR022, CR023, CR024]
FR002: 风险传导图——风险如何流向收入与连续性

有向无环图展示 World Labs 主要风险如何传导至收入、客户采用、利润率、融资和运营连续性。

[CR020, CR022, CR024, CR026, CR027, CR035]

7.4 人员、执行与财务风险

相对于所处阶段,World Labs 的关键人物集中度极高。Fei-Fei Li 在担任 CEO 的同时,仍持有 Stanford University 的 Sequoia Professorship;她在创立公司时把这种双重安排描述为,在打造初创公司的同时继续「some of her work at Stanford」。联合创始人 Justin Johnson 仍在 University of Michigan 任职。这些分时安排会在当前关键的产品市场契合阶段带来执行风险:公司必须同时拿下企业合同、扩展 API 基础设施,并完成 Autodesk 集成。Li 的公众形象——Wired 封面人物、国会证人、UN Scientific Advisory Board 成员——会放大任何失误的声誉后果。 财务透明度几乎不存在。World Labs 尚未披露员工人数、烧钱速度或任何收入指标。公司两轮合计融资 $1.23 billion,定位为资本密集型前沿 AI 实验室;但消费者订阅定价(免费到 $95/month Max)在结构上无法产生足够收入来抵消前沿模型训练成本,除非企业合同量足够大。StartupHub 分析指出,World Labs 累计融资额已经超过 Google 2014 年收购 DeepMind 的价格,却发生在任何披露收入之前——这凸显了市场押注的是长周期资本向价值转化。 人才获取是结构性风险。World Labs 要与 NVIDIA、Google DeepMind、Meta Reality Labs,以及包括 Luma AI 在内的资金充裕同行争夺机器学习工程和 3D 图形人才。Deloitte 的调查证据显示,劳动力准备度是企业采用 AI 的首要障碍,意味着 World Labs 的企业客户也面临并行的人才约束,可能拖慢集成时间线,并在 2026 年缩小复杂世界模型部署的有效可触达市场。 [CR027, CR028, CR029, CR030, CR031, CR032]

人员 / 执行风险清单
角色 / 职能依赖或缺口可能性严重程度缓释尽调路径
CEO — Fei-Fei Li同时担任 Stanford Sequoia Professor 和 CEO;商业化关键阶段分散时间,带来执行风险低(离职);高(兼职分心)严重创始使命和资本结构提供留任激励;但未披露继任计划确认全职投入时间表;厘清 Stanford 休假安排;索取继任计划
联合创始人 Ben Mildenhall(NeRF 发明者,技术负责人)核心渲染技术主要发明者;若离职,Marble IP 的技术防御力会下降技术深度分布在四位联合创始人之间;NeRF 已发表(公共领域知识)确认雇佣状态和归属时间表;评估 NeRF 衍生技术的专利所有权
联合创始人 Justin Johnson(University of Michigan 教职)兼职教职在产品市场匹配阶段增加执行瓶颈联合创始团队有冗余;Li 和 Lassner 提供纵深厘清目前投入 World Labs 与 Michigan 的时间比例;确认休假安排
ML 工程 / 3D 图形人才获取与 NVIDIA、Google DeepMind、Meta Reality Labs、Luma AI 争夺稀缺 3D AI 研究人员强创始人品牌和使命吸引研究人才;资本支持有竞争力的薪酬审查员工数增长和流失;索取关键人员留任方案细节
产品披露有限但早期曝光高$230M 融资和从隐身到发布的过渡,在 Marble 出货前暴露了竞争情报已发生(已经出现)Marble GA 于 2025 年 11 月发布;暴露属于历史问题监测与发布前技术预览相关的 IP 诉讼;无需进一步尽调

行按严重程度从严重到中排列。可能性反映当前判断,而非历史情况。考虑到融资锁定和使命一致, 所有联合创始人的离职可能性归为低;兼职分心问题仍在持续,且影响重大。

[CR027, CR028, CR029, CR032]

7.5 缓释措施与投资论点失效触发器

应对监管风险的主要缓释措施是 World Labs 的 AUP 架构,其中明确限制引用 EU AI Act、California SB 1001 及可比法规的高风险用途。这会限制直接责任敞口,但不能替机构客户卸下合规负担;客户仍必须评估 Marble 输出是否满足自身监管义务。负责任披露计划覆盖了基础安全底线,但在取得 SOC 2 或同等认证之前,缺少正式第三方安全证明会限制受监管行业的企业采用上限。 面对算力依赖,World Labs 与 NVIDIA 和 AMD 的双重投资者关系可带来优先 GPU 配额——但定价权仍在拥有董事会观察员权利的供应商手中。公司的最佳缓释办法,是把收入加速到足以降低对额外股权融资依赖的水平。如果 Autodesk 合作成熟为集成产品工作流,可能加快企业分发;但 TechCrunch 2026 年 Q1 访谈显示,商业产品形态仍未定义。 应结束投资或从根本上重估价格的论点失效触发器包括:EU AI Act 将 Marble 归类为高风险 GPAI,并强制要求公司无法及时满足的第三方审计;Fei-Fei Li 确认离任 CEO;到 2026 年 Q4 仍未签下任何披露收入的企业合同;失去 NVIDIA 或 AMD 作为基础设施供应商且没有替代 GPU 访问;以及 US Copyright Office 或法院认定 Marble 训练数据构成系统性版权侵权,要求撤回模型或重新训练。 投资者应为每个触发器定义明确阈值和监控节奏,而不是只依赖定期管理层报告。依赖风险图(FR003)显示,所有重大风险最终都会通过算力访问或 Fei-Fei Li 持续领导,传导到公司的运营连续性。 [CR043, CR044, CR045, CR046]

缓释与终止标准表
风险可监测触发项阈值 / 事件行动含义
收入延迟与资本耗尽到 2026 年 Q4 仍没有披露收入的企业合同未完成第三轮融资且跑道降至 12 个月以下;或到 2026 年 Q4 仍未披露 ARR重新定价以反映烧钱换价值风险;任何后续投资承诺前要求披露收入
CEO 关键人物离职(Fei-Fei Li)Li 宣布减少职责、休假,或辞去 CEO 职务Li 全职 CEO 状态发生任何变化触发即时复核;投资论点依赖她的技术愿景和投资者 / 合作伙伴关系
EU AI Act 高风险 GPAI 分类European Commission 发布覆盖 Marble 算力规模的 GPAI 分类指引Marble 被归为高风险;强制第三方审计要求落地加快合规计划;测算欧盟时间线对收入的影响;考虑推迟欧盟发布
算力供应中断(NVIDIA / AMD)NVIDIA 或 AMD 限制 API 或云访问;或价格同比上涨 >50%供应限制被确认,或单次推理成本翻倍且没有收入抵消若 90 天内无法以同等成本锁定替代 GPU 算力,投资论点失效
训练数据被认定侵犯版权法院裁定或 Copyright Office 最终规则认定 Marble 训练数据抓取构成侵权最终不利裁定要求模型下架、重新训练或支付重大损害赔偿退出或寻求人才收购式并购;重新训练成本会吃掉大部分剩余资本
Autodesk 集成失败Autodesk 披露退出顾问安排,或转向内部 3D AI 自研到 2027 年 Q3 仍未宣布联合商业产品;或 Autodesk 宣布竞争产品重新评估分发论点;失去 $200M 战略伙伴会实质削弱企业叙事

触发项用于监测;阈值应由法律顾问确认,并在每次董事会复核时更新。 行动含义反映投资者决策标准,不是管理层指令。

[CR034, CR035, CR038, CR004, CR003, CR043]
FR003: 依赖图——关键伙伴、供应商、平台与监管机构

以 World Labs 为中心,向外映射上游硬件、云、监管和分发关系等关键依赖。

[CR019, CR020, CR021, CR022, CR023, CR004]

7.6 附录

Chapter 08

08估值

8.1 投资论点与反论点

World Labs 的投资逻辑建立在三根互相强化的支柱上:创始人质量、技术差异化和市场时机。Fei-Fei Li 创建了 ImageNet,催化了现代深度学习时代;她的联合创始人包括 Ben Mildenhall(NeRF 发明者,NeRF 是主导性的真实世界 3D 重建技术)、Justin Johnson(Pulsar 主要作者、领先的计算机视觉研究者)和 Christoph Lassner(曾任职 Epic Games 与 Meta)。这可能是空间 AI 领域资历最强的创始团队。领投种子轮的 Andreessen Horowitz 明确阐述了投资论点:语言模型让机器掌握文本;世界模型则是三维空间中的对应物。A16z 把空间智能问题的通用性——横跨 AR/VR、机器人、科学发现和创意工具——列为关键投资信号。 产品论点正在落地。Marble 于 2025 年 11 月全面开放,把文本、图像、视频和全景输入生成可导航的 3D Gaussian Splat 世界。World API 于 2026 年 1 月公开发布,让开发者能以编程方式调用 Marble。已点名的集成包括 OpenArt、Magnific AI、Rosebud AI、VIVERSE、Lightcraft、Escape 和 NVIDIA Isaac Sim,覆盖创意、游戏、企业设计和机器人研究等场景。Autodesk 的 $200 million 战略投资表明,一家拥有数百万专业 3D 工作流用户的公司认为 Marble 具备成为可信企业产品的潜力。 反论点同样需要说清楚。截至 2026 年 6 月,World Labs 未披露收入、ARR、客户数量或单位经济。隐含的 $5 billion 估值(来自 Bloomberg,World Labs 未确认)完全是愿景倍数,而不是收入倍数。Autodesk 合作停留在研究和模型层面,并非商业分销合同。竞争压力实质存在:NVIDIA Cosmos 对某些物理 AI 用例免费且开放,Google DeepMind 的 Genie 和 Luma AI 也都在相邻市场运营,并有强大资金支持。Deloitte 的企业 AI 采用研究指出,合规复杂度、劳动力准备度和监管不确定性是拉长 AI 平台公司销售周期的主要摩擦点——这一逆风直接作用于 World Labs 的企业野心。U.S. Copyright Office 自 2023 年起一直在审查 AI 生成输出的可版权性,这种监管不确定性会影响 Pro 和 Max 订阅层中高价值专业用例的商业可行性。[CV001, CV006, CV009, CV010, CV011, CV012]

投资论点与反论点
论点类型论点什么会改变该观点
论点世界级创始团队:Fei-Fei Li(ImageNet)、Ben Mildenhall(NeRF)、Justin Johnson(Pulsar)——视觉、渲染和空间 AI 专长罕见汇合创始人离职或团队分裂
论点从战略投资者募得 $1.23B:AMD、NVIDIA、Autodesk($200M)、a16z —— 显示硬件、软件和 VC 生态均有信心战略投资者退出,或以更低估值完成降价轮
论点空间 AI 覆盖游戏、电影、机器人、建筑和科学发现中的巨大未满足需求——仅生成式 AI 到 2026 年的可服务市场就超过 $83B空间 AI 被证明只是研究兴趣点,而非商业平台
反论点截至 2026 年 6 月未披露收入、ARR 或单位经济——隐含 $5B 估值完全是愿景倍数公司披露 ARR >$20M,并提供企业合同证据
反论点NVIDIA Cosmos 面向物理 AI 用例免费 / 开放权重;部分细分市场存在零成本商品化压力NVIDIA 和开源替代方案相对 Marble 显示出显著质量差距
反论点Autodesk 合作仍处于研究 / 顾问阶段,不是商业分发合同——企业收入仍属推测Autodesk 转为商业经销商或嵌入式分发协议

论点和反论点基于公开证据;每条论点都有引用来源支撑。本表不为概率赋权——概率见情景表。

[CV001, CV002, CV011, CV019, CV023, CV029]
FV001: 建议逻辑

从规模证据、商业证明、竞争和估值出发,推导 World Labs 当前“跟踪”建议的推理链。

节点权重和边的方向仅作示意;这是一张定性逻辑图,不是量化评分模型。

[CV001, CV006, CV011, CV019, CV022, CV029]

8.2 融资背景与估值锚

World Labs 的融资历史记录清晰:2024 年 9 月完成 $230 million 融资,投后估值 $1 billion,由 Andreessen Horowitz、NEA 和 Radical Ventures 领投,AMD Ventures、Intel Capital 和 NVIDIA NVentures 参投。第二次 $1 billion 融资于 2026 年 2 月 18 日宣布,由 Autodesk 以 $200 million 投资并以战略顾问身份锚定,AMD、NVIDIA、Emerson Collective、Fidelity Management and Research 和 Sea Group 加入。两轮合计融资 $1.23 billion;Reuters、TechCrunch、SiliconANGLE 和 StartupHub AI 的报道口径一致。 隐含估值更模糊。Bloomberg News 在 2026 年 1 月报道称,World Labs 正在以约 $5 billion 估值洽谈融资——较约 16 个月前的种子轮标记上涨 5 倍。Reuters 2026 年 2 月报道这一数字时,World Labs 拒绝确认,但公司也没有反驳 Bloomberg 报道。因此,$5 billion 数字处于中等置信度:第三方报道,且有多处引用佐证,但没有公司确认。在隐含 $5 billion 估值且未披露收入的情况下,收入倍数无法定义——这是一个完全建立在市场相信空间智能论点之上的愿景倍数。 World Labs 的可比公司集合有限。公司仍为私营、收入前阶段,而大多数历史先例都是已经交付显著商业证明的公司。按前沿 AI 融资画像看,最接近的类比是 Anthropic;它在 2026 年 2 月以 $380 billion 投后估值完成 $30 billion Series G——但 Anthropic 已披露可观 ARR。3D AI 公司中,Luma AI 仍为私营,未公开披露规模化估值。Stability AI 曾遭遇财务困难。因此,最相关的参照点是:(1)World Labs 自身种子轮倍数($1B 投后 / $230M 融资),(2)Bloomberg 报道的 Series B 隐含价格,以及(3)2026 年 Q1 更广泛的风险投资市场:AI 公司单季融资 $242 billion,形成异常有利的定价环境。[CV001, CV002, CV003, CV004, CV005, CV007]

建议摘要
维度评估依据
建议跟踪未披露收入;隐含 $5B 价格需要商业化证明,目前尚未出现
信心定性论点强;收入或单位经济数据不足,无法锚定估值
风险评级收入前、资本密集、竞争激烈、估值未确认
估值立场隐含 $5B 承压Bloomberg 报道的 $5B 标记(未确认);现阶段收入倍数无从定义
决策含义观望;在收入里程碑或估值确认后重审买入触发:ARR >$20M 或 Autodesk 商业分发;退出触发:降价轮或创始人离职

建议仅基于公开证据;若能访问私有数据,信心会发生实质变化。所有估值引用均指 Bloomberg 报道的约 $5B 隐含价格,World Labs 尚未确认。

[CV003, CV004, CV005, CV011, CV037, CV038]
可比估值表
可比对象阶段 / 类别估值 / 状态与 World Labs 的相关性局限
World Labs 种子轮(2024 年 9 月)产品前私募轮投后 $1B(Bloomberg 报道);募资 $230M自身历史锚点;种子轮募资额 / 估值比为 4.3x无收入基础;纯愿景倍数
World Labs Series B(2026 年 2 月)早期商业化私募轮隐含约 $5B(Bloomberg,未确认);募资 $1B本分析的主要估值锚点;公司未确认无收入确认;带有投机溢价
Anthropic(2026 年 2 月)商业化生成式 AI,Series G投后 $380B;单轮募资 $30B前沿 AI 融资环境背景;Anthropic 已披露收入Anthropic 已有可观收入;产品阶段不可比
Luma AI3D / 视频生成式 AI,私有公司未公开披露;私有公司最接近的空间 / 视频 AI 可比对象;同属 3D 生成细分无公开估值;财务披露有限
Stability AI生成式图像 AI,私有公司2024–2025 年被报道陷入财务困难;此前估值约 $1B反向可比:未能规模化商业化的竞争者模式(开源)和财务轨迹不同;直接可比性有限
NVIDIA Cosmos(母公司:NVIDIA)物理 AI 世界模型,公开隶属市值约 $3.4T 的 NVIDIA(2026 年 6 月约数);Cosmos 免费 / 开放权重直接竞争可比对象;物理 AI 的零成本参考点不是独立实体;价格为零,不是估值可比对象
Midjourney生成式图像 AI,私有公司报道称在约 $300M ARR 下实现盈利(2024 年报道);估值未披露商业化最成功的生成式 AI 工具公司;可作收入代理2D 图像,不是 3D;用户基础和成本结构不同

所有私有公司估值均来自第三方报道,可能不反映实际交易价格。鉴于缺少与 World Labs 同阶段的公开 3D 空间 AI 可比对象,可比集合必然不完整。NVIDIA Cosmos 和 Midjourney 是定性参考点,不是直接收入倍数。

[CV001, CV004, CV005, CV013, CV014, CV015]
FV003: 估值 / 回报区间——乐观、基准、悲观(2028 年视角,均为估算)

World Labs 在乐观、基准、悲观假设下的情景估值区间,并列出 Bloomberg 报道的隐含估值供参考。所有数字均为估算。

所有情景值都是分析估算。World Labs 未披露任何收入。Bloomberg 报道的 $5B 是第三方数字,公司未确认。区间宽度反映收入轨迹和退出倍数的不确定性,并非统计分布。

[CV004, CV005, CV014, CV026, CV027, CV028]

8.3 情景分析与估值区间

由于 World Labs 未披露收入,本章所有情景估值都明确标注为估计值,基于假设收入结果,以及高增长 AI 基础设施公司的可比市场倍数。这些数字只作为分析框架的情景输入,并非精确预测;全文均明确标注为假设。 乐观情景假设,空间智能在 24 个月内成为广泛采用的 AI 范式;Autodesk 的 $200 million 投资转化为大型分销协议,打开其数百万专业用户装机基础;World API 从机器人、游戏和企业仿真中产生有意义的 API 收入;World Labs 到 2028 年实现估计 $100-$300 million ARR。按成长期 AI 公司 30-40x 远期收入倍数计算,隐含估值约为 $3-12 billion,当前隐含价格位于区间中部。乐观情景依赖尚不存在的商业证明:Autodesk 合作需要从研究走向分销,API 也需要获得大量付费企业开发者。 基准情景假设商业增长受控,Autodesk 关系仍为顾问性质,收入来自创意和产消者市场的直接订阅,以及开发者集成带来的 API 收入。到 2028 年估计 $30-80 million ARR,按 25-30x 远期收入倍数计算,隐含估值约为 $750 million 至 $2.4 billion——明显低于 $5 billion 的隐含入场价。基准情景凸显,如果没有乐观情景中的商业加速,当前价格很难证明合理。 悲观情景假设出现实质性商品化:NVIDIA Cosmos 的开放模型扩大支持用例,开源社区模型缩小质量差距,World Labs 难以在规模化时形成差异化。在该情景下,到 2028 年估计 $5-15 million ARR,按 10-15x 倍数计算,隐含估值约为 $50-225 million,远低于任何已投入资本。资本集中风险进一步放大悲观情景:2026 年 Q1 风险融资中,OpenAI、Anthropic 和 xAI 合计吸收了全球风险投资的 65%,挤压中型空间 AI 玩家可获得的资本。[CV016, CV017, CV018, CV019, CV020, CV024]

牛市 / 基准 / 熊市情景分析
情景关键假设(均为估计)2028 年估计 ARR估计估值(2028 年退出)情景关键风险概率信号
牛市Autodesk 将 Marble 分发给专业用户群;World API 成为企业 3D 仿真的标准;空间 AI 采用速度超过生成式 AI 轨迹。假设:所有数字均为估计。$100M–$300M$3B–$12B(30–40x 远期收入,估计)商业证明未能落地;竞争对手缩小质量差距低 — 需要多个商业催化同时出现
基准来自创意 / 专业消费者和中端企业的直接订阅与 API 收入;Autodesk 仍为顾问角色;商业牵引有意义但不足以改变格局。假设:所有数字均为估计。$30M–$80M$750M–$2.4B(25–30x 远期收入,估计)收入增长慢于假设;AI 市场倍数压缩中 — 鉴于当前商业阶段,这是最可能结果
熊市空间 AI 商品化;NVIDIA Cosmos 和开源替代方案侵蚀定价;World Labs 难以在规模化后保持差异化;资本集中限制未来融资选项。假设:所有数字均为估计。$5M–$15M$50M–$225M(10–15x 远期收入,估计)全面商品化可能比模型假设更快到来低中 — 若开源快速缩小质量差距,则可能发生

本表所有 ARR 和估值数字均为分析用途的情景估计,不是预测。收入倍数参考可比高增长私有 AI 公司;退出时的实际倍数将取决于届时市场条件。World Labs 未披露收入,也未提供财务指引。

[CV019, CV022, CV024, CV026, CV027, CV028]
FV002: 假设 ARR 下的估值敏感性(基准收入倍数 25x,均为估算)

以 2028 年为视角,在固定 25x 收入倍数假设下,估算不同 ARR 水平对应的隐含估值。所有数值都是情景估算,不是预测。

所有 ARR 数字都是情景假设,不是预测。World Labs 尚未披露收入。25x 是参考可比的高增长私有 AI 公司得出的示意倍数;退出时的实际市场倍数会变化。Bloomberg 报道的约 $5B 隐含估值,大致对应 25x 下的 $200M ARR——这意味着公司需要铺开相当规模的企业分销。图中数值以百万美元计。

[CV019, CV026, CV027, CV028]

8.4 退出准备度、论点失效触发器与最终尽调

以当前阶段看,World Labs 最可能的退出路径,是由现有投资者之一进行战略收购。Autodesk 是最明显的收购方:其 $200 million 投资和研究层面的合作创造了期权价值,一旦商业证明成立,就可收购完整平台。NVIDIA 和 AMD 也有战略兴趣,确保世界模型在其硬件上规模化运行。IPO 有可能,但需要更充分的商业发展——AI 公司历史上的 IPO 前基准表明,至少需要 $50 million ARR、持续环比增长和清晰单位经济,公开市场才会按规模给公司定价。基于现有证据,World Labs 尚不具备 IPO 准备度。 建议是跟踪,并设置一组具体的论点失效触发器,让投资判断转向买入或退出。买入触发器是收入证据:如果 World Labs 披露或权威第三方报道 ARR 超过 $20 million,并有企业多年合同结构证据,或 Autodesk 合作从研究转为商业分销,基准与乐观情景在当前隐含价格附近会更可信。退出触发器则相反:降价轮、关键战略合作瓦解、关键创始人流失,或监管行动阻断生成式 3D 内容在商业应用中的使用,都构成论点失效事件。 这个阶段的资本效率画像值得注意:约 11-50 名员工,却已融资 $1.23 billion,组织非常精简,意味着要么人均算力支出极高,要么招聘纪律很强。Gartner Hype Cycle for Generative AI 2026 指出,许多生成式 AI 类别处于或接近膨胀预期峰值——这是影响整个空间 AI 市场的结构性风险,不只针对 World Labs。按当前隐含价格进入的投资者,应把幻灭低谷计入定价情景;如果技术最终兑现承诺,复苏周期可能需要 3-5 年。[CV022, CV031, CV032, CV033, CV034, CV035]

论点失效与终止触发项
触发项阈值 / 事件对论点的传导行动含义
降价轮World Labs 下一轮融资估值低于 Series B 隐含 $5B 标记投资者信心恶化;商业化轨迹论点被削弱立即重新评估;若估值下跌超过 30%,大概率退出
创始人离职Fei-Fei Li、Ben Mildenhall 或 Justin Johnson 离开公司核心技术和声誉护城河受损;招聘管线承压立即重新评估;领导层连续性是主要信念驱动因素
战略合作解体Autodesk 公开退出顾问角色,或 AMD / NVIDIA 降低战略参与度企业分销路径被切断;估值中的战略溢价被拿掉重新审视整个商业化论点;大概率下调至退出观察
监管封锁美国或欧盟监管机构禁止 AI 生成 3D 内容商业化,或施加重大版权责任Pro / Max 订阅层失去商业权利基础;整个 SaaS 模式承压评估监管敞口;可能需要大幅重做产品或退出市场
商品化加速开源项目或 NVIDIA Cosmos 在 12 个月内达到 Marble 质量,定价跌破等值 $20/month收入模型被削弱;竞争护城河不够评估差异化路径;若没有可信的护城河回应,转入退出观察

触发条件和阈值是基于公开证据设定的分析框架。若这些条件出现,当前 TRACK 建议将发生实质性变化。

[CV022, CV025, CV033, CV035, CV036, CV040]
最终尽调问题
主题缺失证据重要性负责人或尽调路径
收入和 ARR未披露收入、ARR 或订阅用户数这是锚定任何估值的基础;也决定 $5B 标记是否站得住管理层数据室;任何成长阶段投资的标准材料
单位经济未披露毛利率、CAC 或回本周期推理算力密集;以当前规模,毛利率可能低于 50%管理层;尽调中要求模型 P&L
Autodesk 商业条款合作被描述为「研究和模型层面」;未确认收入分成或转售协议与 Autodesk 的商业分销协议是牛市情形的核心催化剂尽调中直接向 World Labs 和 Autodesk 询问;若转为商业协议,应有新闻公告
股权结构和优先权栈未披露投资人持股、清算优先权和反稀释条款计算不同退出情形下的真实回报时很关键,尤其是低于 $5B 退出的结果法务尽调;要求股权结构表
IP 归属和训练数据来源未披露 Marble 模型的训练数据授权或合成数据政策美国版权局正在审查 AI 生成内容;IP 风险可能影响 Pro / Max 层级的商业权利IP 尽调;要求披露训练数据来源和授权情况

这些是把建议从 TRACK 推进到有条件买入所需的最低证据项。全部都是成长阶段尽调的标准项目,但 World Labs 尚未公开披露。

[CV008, CV011, CV025, CV033, CV037]
FV004: 投资 KPI 评分卡

截至 2026 年 6 月,从市场、验证、护城河、经济性、风险和估值维度给 World Labs 做投委会可用评分,并标注证据质量。

评分是基于公开证据的定性分析判断。每项评分反映证据质量和商业化验证,不只看公司潜力。若公司披露收入或确认商业里程碑,评分可能大幅改善。

[CV011, CV019, CV022, CV023, CV029, CV037]

8.5 附录

免责声明

本报告基于截至 2026-06-18 可获得的公开资料;做出投资决策前,还应补充管理层尽调、客户访谈和交易文件。

证据索引

结论
编号陈述可信度来源
CO001 World Labs is a frontier AI company headquartered in San Francisco, California; its legal entity is World Labs Technologies, Inc., as identified in its January 2026 Terms of Service. SO096, SO028
CO002 World Labs was co-founded by Fei-Fei Li, Justin Johnson, Christoph Lassner, and Ben Mildenhall. SO002, SO028
CO003 World Labs focuses on building Large World Models (LWMs) that can perceive, generate, reason, and interact with 3D spatial environments, a field the company calls spatial intelligence. SO001, SO002
CO004 World Labs' first product, Marble, is a multimodal world model that generates spatially cohesive, high-fidelity, and persistent 3D environments and was made generally available on November 12, 2025. SO008, SO105
CO005 World Labs describes its mission as advancing spatial intelligence to transform storytelling, creativity, robotics, scientific discovery, and beyond, aiming to lift AI from the 2D plane to full 3D spatial understanding. SO002, SO015
CO006 Fei-Fei Li is CEO and co-founder of World Labs; she holds the Sequoia Professorship in Computer Science at Stanford University and is Founding Co-Director of Stanford's Human-Centered AI Institute (HAI). SO032, SO033
CO007 Fei-Fei Li served as VP and Chief Scientist of AI/ML at Google Cloud from 2017 to 2018 before returning to Stanford. SO032
CO008 Fei-Fei Li created ImageNet, the large-scale visual learning dataset credited as foundational to modern computer vision and the deep learning revolution in AI. SO030, SO054
CO009 Ben Mildenhall is co-founder of World Labs and the primary inventor of Neural Radiance Fields (NeRF, arXiv:2003.08934), a landmark 3D scene synthesis technique that became one of the most-cited AI papers of the 2020s. SO035, SO028
CO010 Christoph Lassner is co-founder of World Labs; he previously led research teams at Meta Reality Labs Research and Epic Games, and completed his PhD at the Max Planck Institute for Intelligent Systems in Tübingen. SO036
CO011 Justin Johnson is co-founder of World Labs and is now an assistant professor at the University of Michigan (EECS), having previously collaborated with Li at Stanford's AI Lab. SO037
CO012 World Labs had approximately 20 employees at the time of its stealth exit in September 2024, per Reuters reporting. SO028
CO013 TheOrg lists World Labs' employee count in the 11–50 range as of June 2026, representing a proxy estimate rather than a confirmed company disclosure. SO055
CO014 Fei-Fei Li was appointed to the UN Secretary-General's Scientific Advisory Board, per Stanford HAI reporting. SO034
CO015 Fei-Fei Li was named to Time Magazine's 100 Most Influential People in AI in 2023, per Reuters reporting. SO028
CO016 World Labs raised a $230 million initial round announced September 13, 2024, led jointly by Andreessen Horowitz, New Enterprise Associates, and Radical Ventures. SO028, SO030
CO017 Additional investors in the September 2024 seed round included AMD Ventures, Intel Capital, and NVIDIA NVentures. SO028
CO018 Bloomberg reported in January 2026 that World Labs had been valued at $1 billion at the seed round; the company had declined to confirm its valuation at the time of the 2024 announcement. SO107, SO105
CO019 World Labs raised $1 billion in a second funding round announced February 18, 2026. SO029, SO006
CO020 Investors in the February 2026 $1 billion round included AMD, NVIDIA, Autodesk, Emerson Collective, Fidelity Management and Research Company, and Sea Group. SO029, SO106
CO021 Autodesk invested $200 million in the February 2026 round and took on a formal advisory role, entering a research and product-level collaboration with World Labs focused initially on media and entertainment use cases. SO105, SO029
CO022 Bloomberg reported in January 2026 that World Labs was in discussions to raise at a valuation of approximately $5 billion; the company did not confirm this figure after the round closed. SO107, SO108
CO023 World Labs declined to confirm its post-Series B valuation following the February 2026 announcement, per Reuters and TechCrunch reporting. SO029, SO105
CO024 Total capital raised by World Labs across both rounds stands at approximately $1.23 billion as of February 2026, computed from $230M seed plus $1B Series B. SO107, SO108
CO025 Andreessen Horowitz returned as an investor in the February 2026 Series B round having co-led the September 2024 seed. SO106, SO108
CO026 World Labs has not disclosed any revenue figures, ARR, customer count, or other commercial traction metrics as of the June 2026 research date. SO107
CO027 Marble launched in limited beta in November 2025 and was made generally available with expanded multimodal capabilities on November 12, 2025, approximately 14 months after the seed announcement. SO008, SO105
CO028 The World API was launched on January 21, 2026, providing a programmatic interface for generating explorable 3D worlds from text, images, panoramas, and video. SO007, SO003
CO029 Marble accepts multimodal inputs including text prompts, photographs, video clips, 360-degree panoramas, multi-view inputs, and coarse 3D layouts to generate 3D worlds. SO008, SO001
CO030 Marble is offered in four subscription tiers: Free (limited generations), Standard ($20/month), Pro ($35/month with commercial usage rights), and Max ($95/month with all features). SO107
CO031 World Labs published a functional taxonomy of world models on June 3, 2026, categorizing world models as Renderers, Simulators, and Planners connected in a loop. SO010, SO003
CO032 The Autodesk partnership is described as starting with media and entertainment use cases, with collaboration at the research and model level; data sharing is explicitly excluded from the agreement. SO105
CO033 World Labs has published case studies featuring creative workflow integrations with OpenArt, Lightcraft, Magnific, Rosebud AI, Viverse, and Escape in gaming, VFX, and film domains. SO001, SO003
CO034 Marble supports output export as Gaussian splats, polygon meshes, and videos for integration into downstream creative and simulation pipelines. SO008
CO035 The EU AI Act, adopted by the European Parliament in March 2024, creates compliance obligations for general-purpose AI model developers in EU markets, including transparency and copyright due-diligence requirements applicable to companies like World Labs. SO100
CO036 World Labs operates under the legal entity 'World Labs Technologies, Inc.' as identified in its Terms of Service last updated January 21, 2026. SO096
CO037 World Labs' Acceptable Use Policy and Terms of Service explicitly restrict uses involving bioweapons development, deceptive deepfakes, and interference with critical infrastructure. SO098, SO096
CO038 The EU AI Act may impose transparency and copyright compliance obligations relevant to World Labs' model training data practices and API outputs for EU-market deployments of its generative AI products. SO100
CO039 U.S. Bureau of Industry and Security export control regulations (EAR, 15 CFR Parts 730-774) may apply to World Labs' AI technology for any dual-use applications under applicable export control categories. SO099
CO040 World Labs has raised $1.23 billion without disclosing any revenue metrics, meaning investor returns depend entirely on future commercial development of a spatial AI business model that has not yet demonstrated monetizable traction. SO107, SO110
CO041 Ben Mildenhall co-authored the NeRF paper (arXiv:2003.08934) in 2020, which became one of the most-cited AI papers in 3D scene representation and transformed neural rendering for view synthesis. SO038, SO035
CO042 Christoph Lassner co-authored Pulsar (arXiv:2004.07484), a differentiable sphere-based renderer now integrated as the sphere-based backend of PyTorch3D. SO040, SO036
CO043 World Labs operates Marble Labs as a creative hub where artists, engineers, and designers showcase real-world workflows and experiments using the Marble world model. SO008, SO001
CO044 World Labs' Privacy Policy was last updated November 12, 2025, consistent with the Marble general availability date, indicating active product and legal operations at that time. SO097
CO045 Bloomberg's January 2026 reporting of a $5 billion valuation target for World Labs' Series B was cited in TechCrunch and StartupHub reporting and was not publicly contested by the company. SO107
CO046 World Labs maintains active job postings across research, engineering, and product roles as of June 2026, indicating ongoing team expansion, but no specific executive leadership hires have been publicly announced. SO056
CO047 Google DeepMind's Genie and SIMA families and Runway's video generation models operate in adjacent world model territory, but World Labs is positioned as the only company with navigable persistent 3D world generation as its primary commercial product focus as of 2026. SO029
CO048 Fei-Fei Li's concurrent role as CEO of World Labs and Sequoia Professor at Stanford creates material key-person concentration risk; Reuters reported at the seed exit that she planned to continue some of her Stanford work while building the startup. SO028, SO032
CM001 World Labs defines spatial intelligence as the capacity for AI to perceive, generate, reason, and interact with three-dimensional environments, positioning it as the next frontier in generative AI after large language models. SM027, SM026
CM002 Marble generates spatially cohesive, navigable 3D environments from single images, video clips, 360-degree panoramas, coarse 3D layouts, or text prompts, and outputs Gaussian splats, polygon meshes, and video exports compatible with downstream tools. SM025, SM024
CM003 The World Labs taxonomy essay (June 2026) identifies four functional categories of world models: renderers, simulators, planners, and the feedback loop connecting them. SM026
CM004 Primary status-quo substitutes to AI-generated 3D world models include: procedural generators (Infinigen), physics engine simulation frameworks (MuJoCo, Isaac Sim, RoboSuite), traditional 3D rendering pipelines (V-Ray, Lumion), BIM tools, physical/virtual production sets, and 3D asset libraries. SM013, SM011, SM014, SM012
CM005 The global generative AI market was valued at USD 53.7 billion in 2025, is forecast to reach USD 83.3 billion in 2026, and is projected to reach USD 988.4 billion by 2035 at a CAGR of 31.6%, according to Global Market Insights. SM019
CM006 The global digital twin market is expected to grow from USD 21.14 billion in 2025 to USD 149.81 billion in 2030 at a CAGR of 47.9%, making it the nearest functional proxy for the spatial AI infrastructure segment, according to MarketsandMarkets. SM020
CM007 The World API enables programmatic creation of navigable 3D worlds from text, images, panoramas, multi-view inputs, and video, with outputs embeddable in third-party applications, interactive systems, and simulation workflows. SM024, SM025
CM008 World Labs' robotics case study explicitly characterizes 'high-quality, diverse simulation data' as 'one of the biggest limiting factors in robotics research,' with manual environment creation described as slow, expensive, and inconsistent. SM001, SM015
CM009 The World Labs robotics case study demonstrates Marble generating physically accurate 3D environments with exportable collision meshes imported into MuJoCo and RoboSuite for robot learning and teleoperation research. SM001, SM003
CM010 Architecture and design customers using Marble via the Fenestra platform can transform concept images or text prompts into explorable 3D worlds within minutes, enabling immersive design review before any floor plan is drafted. SM002
CM011 Filmmaker Tim Simmons used Marble as a persistent virtual backlot for a 48-second AI micro-short, maintaining spatial consistency across multiple shots in a way that text-to-video tools alone could not achieve. SM004
CM012 NVIDIA Cosmos is positioned as the Open Physical AI Foundation Model, enabling robots to build World Action Models (WAMs), simulate controllable physics-grounded environments, and power vision AI reasoning for autonomous systems. SM008
CM013 Google DeepMind's Genie 2 is a foundation world model generating action-controllable, playable 3D environments from single prompt images, aimed at training and evaluating embodied AI agents at unlimited environmental diversity. SM005, SM006
CM014 A 2026 arXiv survey of digital twin AI identifies eleven cross-domain application areas for world-model-enabled digital twins, including healthcare, aerospace, smart manufacturing, robotics, and smart cities — confirming broad expansion potential for spatial AI beyond World Labs' current segments. SM016
CM015 Deloitte's 2025 survey of AI leaders found approximately 60% of respondents cite integration with legacy systems and risk/compliance concerns as the primary barriers to adopting agentic and physical AI at the organizational level. SM018
CM016 Deloitte identifies unclear use cases or business value, lack of technical expertise, and governance/risk concerns as the top three enterprise barriers to physical AI adoption — all directly applicable to spatial AI product deployments. SM018
CM017 The 2026 arXiv survey (arXiv:2604.26509) identifies three roles of 3D generation in embodied AI: data generator (simulation-ready content), simulation environment construction (interactive worlds), and Sim2Real Bridge (digital twin reconstruction and augmentation). SM015
CM018 The 2026 3D generation survey identifies persistent bottlenecks including limited physical annotations, the gap between geometric quality and physical validity, and the sim-to-real divide — confirming that World Labs' simulation use case requires technical advances beyond current visual realism. SM015
CM019 MuJoCo is a free and open-source physics engine widely used for robotics and embodied AI research, optimized for model-based optimization including optimal control and system identification; it is the standard simulation backend for RoboSuite. SM013, SM012
CM020 NVIDIA Isaac Sim is an open-source robotics simulation framework built on the Omniverse platform for physically based simulation, testing, and synthetic data generation; it represents the dominant commercial simulation environment for enterprise robotics. SM011
CM021 RoboSuite is a simulation framework powered by MuJoCo for robot learning, offering benchmark environments with diverse robot embodiments including humanoids and whole-body controllers, widely used in academic robotics research. SM012
CM022 Infinigen is a procedural 3D scene generator developed at Princeton Vision & Learning Lab, producing high-quality training data for computer vision under a BSD-3 open-source license — a direct open-source substitute to commercial world generation APIs for simulation data. SM014
CM023 Google DeepMind's SIMA (Scalable Instructable Multiworld Agent) is a generalist AI agent designed to follow natural-language instructions across diverse 3D virtual environments and video games, demonstrating strong industry demand for scalable multi-environment 3D training. SM007
CM024 Luma AI's interactive 3D scenes deliver 30 FPS web rendering with 8 MB file sizes for objects and 20 MB for scenes, targeting the creative and developer market with universally embeddable 3D across web, iOS, and Android. SM010, SM009
CM025 An MDPI academic review (September 2025) characterizes generative AI as enabling digital twins to become proactive, self-improving cognitive systems capable of simulating complex manufacturing processes, predicting maintenance, and generating novel operational scenarios. SM017
CM026 The World API launch (January 2026) enables worlds to be rendered on the web, exported into downstream tools, or integrated into interactive systems and simulations — making world creation a programmable, on-demand capability. SM024
CM027 World Labs' '3D as code' essay (March 2026) argues that 3D representations are becoming the universal interface for spatial communication between humans and AI, analogous to text as the interface for software via large language models. SM022
CM028 Autodesk's $200 million investment in World Labs (February 2026) is structured as an exploration partnership to integrate world models into Autodesk's 3D CAD workflows, starting with entertainment use cases, according to TechCrunch. SM028
CM029 World Labs' Spark 2.0 renderer adds a Level-of-Detail streaming system for 3D Gaussian Splatting that works on any device with a web browser including desktop, iOS, Android, and VR — enabling large spatial world delivery without custom infrastructure. SM023
CM030 No analyst report reviewed for this chapter explicitly quantifies 'spatial intelligence' or '3D generative AI' as a standalone market category distinct from broad generative AI or digital twin markets. SM019, SM020
CM031 NIST's AI governance mission emphasizes a risk-based approach to AI management with voluntary standards; the AI Risk Management Framework (AI RMF) is the most widely adopted US enterprise standard and will shape compliance requirements for spatial AI deployments. SM029
CM032 MarketsandMarkets and Global Market Insights use top-down analyst survey methodology for their market forecasts; these estimates are not independently verified and typically carry uncertainty of at least ±50% relative to actual market spend. SM019, SM020
CM033 Deloitte's AI adoption analysis documents general enterprise enthusiasm for physical AI constrained by integration complexity, governance uncertainty, and ROI ambiguity — a pattern directly applicable to World Labs' enterprise spatial AI sales motion, representing a buyer education and sales cycle barrier. SM018
CM034 World Labs' taxonomy essay acknowledges that 'world model' is one of the most overloaded terms in AI, used simultaneously by computer vision, robotics, game research, and neuroscience communities — creating market segmentation and buyer education challenges. SM026
CM035 Q1 2026 saw AI companies receive USD 242 billion in venture investment — 80% of the global total of USD 300 billion — marking an all-time high for AI funding concentration, signaling strong capital availability for AI infrastructure including spatial AI. SM030
CM036 World Labs' market addressability is best represented as a three-layer pyramid: the broad generative AI market ($83B TAM in 2026), the digital twin and simulation infrastructure market ($21B SAM-proxy in 2025), and an estimated spatial AI / world generation serviceable market of $1B–$5B across all four buyer segments. SM019, SM020
CM037 Market estimates for the nearest proxies to World Labs' addressable market differ by 5–10× depending on which category boundary is used: the broad generative AI market ($83B in 2026) versus the digital twin market ($21B in 2025) versus bottom-up vertical estimates ($0.5B–$2B for robotics simulation alone), reflecting the absence of a unified spatial AI market category. SM019, SM020, SM015
CM038 World Labs' four buyer segments — media/creative, robotics/embodied AI, architecture/design, and developer/API — have distinct budget owners, adoption triggers, and sales motion requirements, implying segmented go-to-market investment even for a single product. SM001, SM002, SM004, SM024
CM039 The World Labs adoption funnel progresses through five stages: market awareness via research and press coverage, trial via free/pro subscription, API integration via World API, enterprise workflow embedding, and platform lock-in via partnerships and data flywheel. SM024, SM025, SM028
CM040 The global generative AI market in 2026 is forecast at USD 83.3 billion by Global Market Insights, with North America as the largest regional market and Asia Pacific as the fastest-growing region. SM019
CM041 North America accounted for 38% of the global digital twin market revenue share in 2024 according to MarketsandMarkets, making it the primary geographic market for early spatial AI adoption. SM020
CM042 Traditional architectural visualization has followed a sketch-to-render-to-animation pipeline for decades; Marble's generative spatial capabilities represent a compression of this workflow into a real-time, immersive 3D experience that collapses ideation and spatial review into one step. SM002
CM043 Luma AI's interactive 3D scenes achieve 30 FPS web rendering at 8–20 MB file sizes with streaming that starts immediately on load — establishing the delivery performance baseline that any competing spatial AI product must meet for developer and creative adoption. SM010
CM044 Generative AI enables digital twins in manufacturing to simulate complex production lines, optimize operations, predict maintenance needs, and generate training data for rare operational scenarios — representing manufacturing as a large long-term buyer vertical for spatial AI infrastructure. SM017, SM016
CM045 World Labs' '3D as code' framing implies a long-run total addressable market comparable in scope to the LLM market, as 3D spatial representations would become the universal interface for all physical and virtual world interactions. SM022
CM047 World Labs' market framing positions 3D as the universal spatial interface analogous to text as the interface for LLMs, implying a long-run TAM at or above the current LLM market scale (hundreds of billions of dollars). SM022, SM027
CM048 Third-party analyst data for the nearest spatial AI proxies — digital twin market ($21B in 2025) and inferred robotics simulation SAM ($0.5B–$2B) — implies a realistic near-term addressable market an order of magnitude smaller than the LLM market, contradicting the company's long-run TAM framing. SM019, SM020, SM015
CP001 Luma AI declares its mission to be 'unified general intelligence that can generate, understand, and operate in the physical world,' positioning it as a direct peer to World Labs in the spatial intelligence and 3D world model category. SP018
CP002 Luma AI offers interactive 3D scenes achieving 30 FPS on web browsers with 8 MB embedding for objects and 20 MB for scenes, available across iOS, Android, and web platforms. SP019
CP003 Google DeepMind introduced Genie 2 in December 2024 as a foundation world model capable of generating action-controllable, playable 3D environments for training and evaluating embodied agents. SP014, SP015
CP004 Genie 2 generates 3D environments from a single prompt image that can be played by humans or AI agents using keyboard and mouse inputs, with its primary use case being AI agent training rather than commercial content creation. SP014, SP015
CP005 NVIDIA Cosmos 3 is an open physical AI foundation model covering vision AI reasoning, robot policy model building, world simulation, and synthetic video data generation, made available as open-source weights via GitHub and a hosted catalog. SP017, SP027
CP006 NVIDIA Isaac Sim is an open-source reference framework built on NVIDIA Omniverse libraries for robotics simulation, testing, and synthetic data generation in physically based virtual environments. SP027, SP017
CP007 NVIDIA Isaac Sim explicitly integrates Cosmos world foundation models for synthetic data augmentation, deepening the Cosmos-Isaac pipeline as a competing pathway to Marble in the robotics simulation and training data generation use case. SP027, SP017
CP008 World Labs' robotics case study demonstrates that researchers used Marble for scalable robotic simulation scene generation and data collection, directly overlapping with NVIDIA Isaac Sim and Cosmos's target use case. SP005, SP027
CP009 OpenArt describes itself as 'one of the world's leading AI creative platforms' and launched OpenArt Worlds by integrating World Labs' Marble model, making it a distribution partner and downstream channel rather than a competing world model developer. SP010, SP022
CP010 Magnific is described as 'one of the largest creative platforms in the world' with millions of designers, marketers, and content teams using it daily; it integrated Marble for precision 3D environment control in campaign workflows. SP009, SP024
CP011 Rosebud AI enables game creation using natural language prompts without traditional coding and collaborated with Marble on generative 3D world integration, serving as a customer and distribution partner rather than a competing world model developer. SP011, SP025
CP012 VIVERSE (HTC's metaverse platform) collaborated with Marble to blend AI-generated 3D worlds with interactive browser-based experiences, positioning Marble as a generation layer within VIVERSE's existing metaverse distribution network. SP012, SP026
CP013 Stability AI's Stable Zero123 generates novel 3D object views from single images using Stable Diffusion 1.5 as the base, targeting non-commercial and research use with a commercial membership tier, representing a narrower capability (object-level) than Marble's world-scale generation. SP020
CP014 Magic123 (academic, arXiv 2023) achieves high-quality 3D object generation from a single image using combined 2D and 3D diffusion priors, demonstrating that the underlying techniques for 3D generation are accessible in open-source academic form. SP021
CP015 Google DeepMind's SIMA agent follows natural-language instructions across diverse 3D game environments, indicating that DeepMind is developing both world model (Genie 2) and generalist agent (SIMA) capabilities that together could address the agentic use cases World Labs targets via its API. SP016, SP014
CP016 World Labs' Spark 2.0 renderer streams large 3DGS worlds on any device with a web browser using a Level-of-Detail system that automatically optimizes Gaussian splatting detail for viewpoint as users navigate. SP004
CP017 Luma AI has released RAY3.2 for video direction and UNI-1 for brand-specific model training, extending its competitive surface from 3D capture and interactive scenes into multi-modal creative agents. SP018
CP018 The Marble case study with Lightcraft demonstrates integration with virtual production pipelines (Lightcraft Jetset), where Marble provides the 3D world environment layer and Lightcraft provides compositing and live-action integration. SP008, SP023
CP019 World Labs describes Marble as a 'first-in-class generative multimodal world model' accepting text, images, video, and coarse 3D layouts as inputs to produce navigable 3D worlds—a breadth of input modalities that is a company-claimed differentiator. SP003
CP020 The World API enables programmatic on-demand generation of complete 3D worlds for integration into products and workflows, representing a platform distribution strategy of embedding Marble capabilities inside partners' own tools. SP002
CP021 NVIDIA Cosmos provides access to open-source model weights via GitHub and a hosted catalog without explicit licensing cost to developers, presenting a zero-licensing-cost alternative to World Labs' commercial World API for overlapping use cases. SP017, SP027
CP022 Q1 2026 global venture funding reached $300 billion with $242 billion (80%) going to AI startups; OpenAI, Anthropic, xAI, and Waymo collectively raised $188 billion, representing 65% of all global VC in that quarter. SP029
CP023 Venture funding to foundational AI startups doubled in Q1 2026 versus all of 2025, but this growth was concentrated in a small number of large foundational giants, implying smaller specialized AI companies like World Labs face intensifying competition from far better-capitalized peers. SP029
CP024 Genie 2's primary research focus—generating training environments for AI agents—is a complementary rather than directly substitutable trajectory relative to World Labs' commercial API for creative and simulation customers as of the report date. SP014
CP025 World Labs' architectural visualization case study identifies Fenestra and Interior AI as early collaborators using Marble for design ideation, representing a vertical—architecture and interior design—not currently targeted by NVIDIA Cosmos or Google DeepMind's Genie 2. SP006
CP026 World Labs has not published public pricing for the World API as of the report date; pricing is accessed through a gated early-access model with commercial partnerships, making cost benchmarking against competitors not possible from public sources. SP002, SP003
CP027 NVIDIA Cosmos is available through open-source download from GitHub and a hosted catalog at no explicit licensing cost to developers, in contrast to World Labs' gated commercial API model. SP017, SP027
CP028 Stability AI's Stable Zero123 is available for non-commercial and research use without payment, with commercial use requiring a Stability AI membership subscription whose pricing is not publicly specified in the available sources. SP020
CP029 OpenArt's platform offers free-tier access for basic image and video generation; the pricing of the OpenArt Worlds 3D tier, built on Marble integration, is not separately disclosed in the available sources. SP022, SP010
CP030 World Labs' Marble-based partnerships with OpenArt, Magnific, Rosebud AI, VIVERSE, and Lightcraft reflect an API embedding strategy that builds switching costs at the partner level once their end-users engage with Marble-generated 3D environments, though contractual lock-in terms are not publicly disclosed. SP010, SP009, SP011, SP012, SP008
CP031 Marble's multimodal input breadth (text, image, video, coarse 3D layout, panorama) is a company-claimed moat, but Magic123's academic work and NVIDIA Cosmos's video-to-world generation capability demonstrate that input breadth is achievable by well-resourced peers without licensing World Labs' technology. SP003, SP021, SP017
CP032 World Labs' co-founder research heritage—Fei-Fei Li on ImageNet and computer vision, Ben Mildenhall on NeRF, Christoph Lassner and Justin Johnson on 3D representations—provides credibility and talent-attraction advantages that represent a talent-based moat difficult to replicate quickly. SP001
CP033 The Marble Labs showcase catalogue contains over 20 public demos—including VR Gaussian splats, collider builder, concept-to-splat, avatar tools, and spatial blueprints—demonstrating ecosystem breadth, though each demo's commercial traction is not publicly disclosed. SP001, SP013
CP034 NVIDIA's simultaneous role as a strategic investor in World Labs' $1 B round and as the developer of NVIDIA Cosmos, a competing physical AI foundation model, creates a potential conflict of interest that could influence Cosmos product roadmap or World Labs' access to NVIDIA ecosystem resources. SP017, SP029
CP035 The Crunchbase Q1 2026 data shows that capital concentration in large foundational AI labs is accelerating; NVIDIA, Google, and other large platforms can fund competing world model research indefinitely at a scale World Labs cannot match, representing a structural capital asymmetry. SP029
CP036 Third-party developers in the Marble Labs showcase are building production tools—VR locomotion systems, custom avatar generators, splat collider builders—on top of Marble, indicating early formation of a developer community that would increase platform switching costs. SP013, SP001
CP037 The Splat World VR case study shows a developer (Daniel Skaale) building a full real-time VR experience in Unity using Marble-generated Gaussian splats for lighting, animation, physics, and interactive effects—extending Marble's value beyond passive world viewing into interactive game-like experiences. SP013
CP038 NVIDIA Cosmos's open-source positioning—free weights on GitHub and a hosted catalog—could accelerate adoption among robotics researchers who would otherwise evaluate World Labs' commercial World API for simulation use cases, directly undercutting World Labs' pricing power in that vertical. SP017
CP039 Google DeepMind's Genie 2 has not been made available as a commercial API or consumer product as of the report date; it remains a research demonstration, meaning it does not currently address the same commercial market segment as World Labs' World API. SP014, SP015
CP040 Stability AI's Stable Zero123 was released under a non-commercial license with commercial access gated to a membership subscription, positioning it as a research and limited commercial tool rather than an enterprise world model API competitor—Stability AI is not currently pursuing the same enterprise market as World Labs. SP020
CP041 The global generative AI market was valued at USD 83.3 billion in 2026 and is projected to reach USD 988.4 billion by 2035 at a 31.6% CAGR, providing tailwind for multiple competitors to build viable world model products simultaneously. SP028
CP042 World Labs' architectural visualization use case (case study with Fenestra and Interior AI) is outside NVIDIA Cosmos's and Google DeepMind's current commercial focus, indicating World Labs may maintain differentiated territory in design visualization even if robotics simulation is commoditized. SP006, SP017, SP014
CP043 World Labs' Spark 2.0 provides a Level-of-Detail streaming system for large 3DGS worlds on any web browser, a technical capability not described in Luma AI's or NVIDIA Cosmos's public documentation as of the report date. SP004, SP019, SP017
CP044 Open-source academic models like Magic123 and free-access commercial models like NVIDIA Cosmos and Stable Zero123 demonstrate that the technical underpinnings of 3D generation are being commoditized from both above (well-funded incumbents giving away models) and below (academic research), creating a convergent commoditization risk for World Labs' commercial API. SP021, SP017, SP020
CP045 The most adverse competitive scenario for World Labs is a two-front squeeze: NVIDIA commoditizes the robotics simulation market through free Cosmos models distributed through Isaac Sim, while Google DeepMind commercializes Genie 2 into a creative API—leaving World Labs competing without incumbent distribution advantages in either anchor market. SP014, SP017, SP027
CI001 World Labs raised $1 billion in new funding in its Series B, with investors including AMD, Autodesk, Emerson Collective, Fidelity Management & Research Company, NVIDIA, and Sea, as announced February 18, 2026. SI001, SI004
CI002 Autodesk invested $200 million in World Labs' Series B round and will serve as a strategic adviser to the startup. SI028, SI004
CI003 World Labs' total capital raised across both the September 2024 seed and February 2026 Series B rounds is approximately $1.23 billion. SI030, SI028
CI004 Bloomberg's January 2026 reporting indicated World Labs was in funding discussions at a valuation of approximately $5 billion; the company did not publicly confirm or deny this figure. SI030, SI004
CI005 Andreessen Horowitz led World Labs' seed round and lists World Labs as a portfolio company in its public AI portfolio page. SI006, SI005
CI006 Marble subscription pricing spans four public tiers: Free (limited generations, no commercial rights), Standard at $20 per month, Pro at $35 per month with commercial usage rights, and Max at $95 per month with all features. SI037, SI034
CI007 The World API launched January 21, 2026 and provides programmatic access to Marble world generation; no public pricing schedule for the World API has been observed as of June 2026. SI002
CI008 Commercial usage rights are explicitly included starting at the Pro tier at $35 per month per World Labs' Terms of Service, indicating the Free and Standard tiers restrict or exclude commercial use. SI034, SI037
CI009 The World API enables developers to generate navigable 3D worlds from text, images, panoramas, multi-view inputs, and video via an asynchronous request-response model that outputs a complete 3D world. SI002
CI010 No public enterprise pricing, bulk API pricing, or volume discount schedule for the World API or enterprise Marble plans was observable from any reviewed source as of June 2026; enterprise pricing appears to be individually negotiated. SI002, SI037
CI011 The Autodesk-World Labs partnership will initially focus on media and entertainment use cases, with collaboration at the research and model level; data sharing is excluded from the agreement per available reporting. SI028
CI012 Autodesk's investment provides World Labs with a potential distribution channel into architectural, engineering, construction, manufacturing, and entertainment workflows via Autodesk's established enterprise customer base. SI028, SI005
CI013 OpenArt launched "OpenArt Worlds" using World Labs' Marble model, transforming a single image into a fully navigable, persistent 3D environment that creators can revisit and build within over time. SI009, SI018
CI014 Magnific, described as a platform with millions of daily users for designers, marketers, and content teams, integrated Marble to provide a 3D spatial control interface for precise product placement and creative composition. SI015, SI020
CI015 Rosebud AI, a game-creation platform using natural language prompts, collaborated with Marble to integrate generative 3D world creation into its game-building pipeline and explored multiplayer world environments. SI016, SI021
CI016 VIVERSE, an HTC browser-based interactive 3D world platform, collaborated with World Labs to integrate Marble-generated 3D worlds into its interactive worldbuilding workflow. SI017, SI022
CI017 Lightcraft integrated Marble with Lightcraft Jetset and Beeble for virtual production filmmaking workflows, enabling indie filmmakers to create and use AI-generated 3D environments as virtual film sets. SI014, SI019
CI018 NVIDIA Isaac Sim researchers used Marble's generative world technology to create scalable, physically accurate 3D scenes for robotics simulation, training data generation, and real-to-sim transfer experiments. SI012, SI024
CI019 World Labs' cost structure is expected to be dominated by GPU compute for model training and inference, research personnel, and cloud infrastructure, consistent with the cost profile of frontier foundation-model companies. SI037, SI005
CI020 The Org lists World Labs at 11 to 50 employees as of June 2026; at this team size, monthly payroll is modest relative to the $1.23 billion in capital raised. SI007
CI021 Generating spatially cohesive, high-fidelity, persistent 3D worlds from a single input requires a more GPU-intensive pipeline than comparable 2D image generation, implying above-average inference cost per generation. SI037, SI026
CI022 World Labs delivers Marble as a cloud-hosted service with persistent world storage, API serving, and downloadable outputs, implying ongoing cloud infrastructure cost alongside per-generation GPU inference cost. SI034, SI035
CI023 The global generative AI market was valued at $53.7 billion in 2025 and is forecast to grow to $988.4 billion by 2035 at a 31.6% CAGR; high infrastructure and compute costs are identified as a key market challenge in this forecast. SI026
CI024 Stability AI operates in adjacent generative AI product territory with Stable Diffusion and Stable Audio; the breadth of competing generative AI creative tools in the market creates pricing pressure relevant to Marble's subscription tier positioning. SI036
CI025 World Labs raised $230 million in its September 2024 seed round, led by Andreessen Horowitz, New Enterprise Associates, and Radical Ventures, with AMD Ventures, Intel Capital, and NVIDIA NVentures as additional investors. SI003, SI006
CI026 World Labs' $1 billion Series B closed February 18, 2026; Intel Capital — present in the seed round — does not appear in Series B investor lists per available reporting. SI001, SI028, SI029
CI027 No quarterly burn rate, monthly cash consumption, net income, or net loss figures for World Labs have been publicly disclosed as of June 2026; runway cannot be precisely calculated from public evidence. SI004, SI030
CI028 Fidelity Management & Research Company's participation in the Series B is consistent with growth-equity positioning for a company on a potential public-market liquidity trajectory, suggesting a multi-year exit horizon. SI001, SI004
CI029 Deloitte identifies compliance complexity, workforce readiness, and evolving regulatory requirements as primary barriers to enterprise AI adoption that extend sales cycles and increase customer acquisition cost for AI platform companies. SI025, SI027
CI030 Crunchbase data shows that venture funding to foundational AI startups doubled in Q1 2026 versus all of 2025, but is increasingly concentrated in a handful of frontier giants including OpenAI, Anthropic, and xAI, creating a capital-concentration dynamic that disadvantages sub-frontier follow-on rounds. SI033, SI032
CI031 World Labs has not disclosed revenue, ARR, customer count, gross margin, or unit economics metrics as of June 2026 — approximately 17 months after the September 2024 seed round and seven months after Marble's general availability on November 12, 2025. SI030, SI037
CI032 The World API launched January 21, 2026 without a published pricing schedule; developer adoption rates, API call volumes, and API-channel revenue contribution are unknown as of June 2026. SI002
CI033 World Labs has published case studies documenting named integrations with OpenArt, Magnific, Rosebud AI, VIVERSE, Lightcraft, and Escape across gaming, VFX, film, creative, and interactive-experience verticals, plus a NVIDIA Isaac Sim robotics simulation use case. SI009, SI010, SI011, SI012, SI013, SI014, SI015, SI016, SI017
CI034 The Pro tier at $35 per month explicitly includes commercial usage rights per the World Labs Terms of Service, making this the minimum tier for creators embedding Marble outputs in paid deliverables. SI034, SI037
CI035 The World API's asynchronous design — submit a world generation request and receive a navigable 3D world — is architecturally consistent with usage-based credit or per-call pricing, which is a common monetization model for inference APIs, though World Labs has not confirmed its API pricing mechanism. SI002
CI036 Stanford HAI's 2026 AI Index reports declining compute costs year-over-year in the AI sector, a trend that could reduce World Labs' per-generation cost over time, but the report provides no company-specific data. SI027
CI037 World Labs' Privacy Policy includes provisions for data collected through its Sites, Services, APIs, and Large World Models, which is consistent with enterprise-grade data handling requirements for a commercializing API product. SI035
CI038 Escape, an interactive experience and gamification platform, is listed among World Labs' integration partners in the case-study corpus, extending GTM evidence beyond creative media and robotics into enterprise gamification workflows. SI023
CI039 The architectural and interior design use case is documented through case studies with Fenestra and Interior AI as early collaborators, positioning Marble as a design visualization tool in AEC workflows. SI011
CI040 Marble case studies featuring filmmakers Tim Simmons and Henrik Vasquez (Theoretically Media) illustrate a prosumer creative workflow consistent with Standard or Pro subscription tier adoption. SI013
CI041 Q1 2026 global venture investment reached a record $300 billion, with AI companies receiving 80% of global venture funding — a macro environment that remains favorable for World Labs' future capital raises. SI032, SI033
CI042 Andreessen Horowitz framed World Labs as part of a transition from language models toward world models, reinforcing investor expectations that 3D interaction and simulation could become monetizable software primitives. SI038
CI043 The Dwarven Caves showcase demonstrates that Marble can support large connected fantasy environments, expanding the set of premium creator and gaming workflows that could justify paid usage beyond single-scene generation. SI039
CI044 The Musée du Monde showcase ties a February 2026 internal hackathon to an interactive 3D museum experience, suggesting World Labs is still using experimentation-led product development rather than a narrowly fixed SKU catalog. SI040
CE001 Marble accepts text, single images, multi-image sets, 360-degree panoramas, video, and coarse 3D layouts as inputs for world generation. SE001, SE006
CE002 Marble generates spatially consistent, high-fidelity, persistent 3D worlds that users can move through, edit, and export from a web browser. SE001
CE003 The World API was publicly launched on January 21, 2026, the date also reflected in the Terms of Service effective date. SE005, SE012
CE004 Marble supports interactive editing, element-level modifications, expansion into larger scenes, and combination of independently generated worlds within a single session. SE006
CE005 Marble exports generated 3D worlds as Gaussian splats (.spz or .rad), mesh files (.glb), and video; outputs can also be consumed via the World API for downstream integration. SE006, SE005
CE006 World Labs targets film and VFX, gaming and AR/VR, robotics and embodied AI simulation, and architecture and design as primary vertical markets for Marble. SE003, SE011, SE029, SE030
CE007 Filmmaker Tim Simmons (Theoretically Media) created a 48-second AI micro-short using Marble as a stable 3D virtual set, maintaining spatial consistency across all scenes. SE030
CE008 Robotics researchers Hang Yin and Abhishek Joshi used Marble to generate scalable, photorealistic 3D scenes with depth, lighting, geometry, and collider meshes for robot simulation and domain randomization training. SE011
CE009 Architecture case study partners Fenestra and Interior AI used Marble to transform concept images into walkable 3D environments for client review, enabling design exploration before floor plan drafting. SE029
CE010 World Labs created its own launch marketing video using Marble, generating hundreds of 3D worlds during production and rendering them on an LED volume stage. SE015
CE011 The World API processes world generation asynchronously; developers submit a request with their input and receive a fully navigable 3D world; no special client software is required. SE005
CE012 VR developer Daniel Skaale built a first-person VR experience in Unity (Splat World) using Marble-generated Gaussian splat environments, adding real-time physics and player interaction via custom C# tooling. SE031
CE013 Spark is a 3D Gaussian Splatting renderer built by World Labs for the web, integrating with THREE.js and running via WebGL2 on desktop, iOS, Android, and VR devices. SE009, SE001
CE014 Spark 2.0 introduced a Level-of-Detail streaming system that adaptively adjusts 3DGS detail to the viewer's position and streams data in real time, enabling large 3DGS worlds to render on any web-connected device including mobile and VR headsets. SE009, SE001
CE015 Spark was developed internally because existing web 3DGS renderers could render only one 3DGS object at a time, could not dynamically animate splats, and did not reliably support multiple device types. SE009
CE016 World Labs describes Marble as a multimodal world model that lifts whatever input signals are available — text, image, video, or spatial layout — into a unified 3D world representation, analogously to how humans integrate sensory inputs into a mental model. SE006, SE007
CE017 Co-founder Ben Mildenhall is the lead author of NeRF (Neural Radiance Fields for View Synthesis), published at ECCV 2020 as an oral presentation, a foundational paper in neural rendering. SE021, SE002
CE018 Co-founder Christoph Lassner is the lead author of Pulsar (arXiv 2004.07484), a sphere-based differentiable renderer that executes orders of magnitude faster than competing differentiable rendering techniques. SE023, SE002
CE019 3D Gaussian Splatting (arXiv 2308.04079, ACM TOG 2023) achieves state-of-the-art visual quality with real-time rendering at ≥30 fps at 1080p resolution, the first method to reach this threshold for unbounded scenes. SE024, SE009
CE020 Mip-NeRF 360 (arXiv 2111.12077) reduces mean-squared error by 57% compared to mip-NeRF for unbounded scene reconstruction, using non-linear scene parameterization and a distortion-based regularizer. SE022
CE021 World Labs describes Marble as "a first-in-class generative multimodal world model" and characterizes its capabilities as making strides toward fully spatially intelligent AI. SE006
CE022 Marble generates and exports collider meshes alongside visual splat geometry, enabling physical interaction in simulation frameworks such as NVIDIA Isaac Sim, Unity, and Rapier3D. SE011, SE019
CE023 World Labs published a functional taxonomy of world models in June 2026, classifying them into renderers, simulators, and planners operating within a perception-action loop, and positioning this as the conceptual basis for its product roadmap. SE007
CE024 3D Gaussian Splatting represents scenes as millions of semi-transparent colored ellipsoids (splats) each defined by center, XYZ scale, rotation, color, and opacity; rendered by sorting splats back-to-front and applying alpha compositing. SE024, SE009
CE025 Marble Composer is a tool for compositing, aligning, and editing multiple generated 3D worlds within a single workflow, used by developers to assemble complex multi-room or multi-world scenes. SE031, SE003
CE026 Marble is accessible via web browser at marble.worldlabs.ai with no proprietary desktop client required; the same generation capability is exposed programmatically through the World API. SE001, SE005
CE027 Third-party developers have integrated Marble with Unreal Engine 5 via the Volinga plugin, Unity via custom C# tooling, and Three.js for browser-based interactive 3D experiences. SE016, SE031, SE019
CE028 The Collider Builder tool, built by a World Labs product team member, enables users to draw collision geometry (box, sphere, cylinder) directly over Gaussian splat scenes in the browser and export as .glb. SE017
CE029 XR developer Ruben Fro built a walking robot that navigates through Marble-generated splat environments by reading raw splat data via custom raycasting, without requiring pre-built meshes, navmeshes, or baked colliders. SE018
CE030 The AI-native 3D pipelines showcase demonstrates an end-to-end workflow combining Marble, fal, mesh generation tools, sound synthesis, and Unreal Engine, orchestrated via Claude Code. SE016
CE031 The World Labs robotics case study documents integration with NVIDIA Isaac Sim: Marble-generated scenes including collider meshes are used as training environments for robot perception, planning, and control. SE011, SE026
CE032 STARSPEED, a multiplayer browser spaceship game built by James Kane, uses Marble-generated sci-fi environments containing over 100 million Gaussian splats rendered in real time via Spark.js and Three.js. SE020
CE033 The World API launch blog post and the Terms of Service effective date both confirm January 21, 2026 as the World API's public release date. SE005, SE012
CE034 Marble Labs at worldlabs.ai/labs serves as a developer portal hosting case studies, tutorials, technical documentation, and a curated showcase of community-built integrations. SE003
CE035 The open-source Third-Person Character Controller template on GitHub, built on Marble, Spark 2.0, Rapier3D, and Three.js, provides a fork-ready starting point for browser-based interactive 3D experiences. SE019
CE036 Marble's collider mesh export enables physical simulation integration with robotics frameworks including RoboSuite and MuJoCo-based simulators, as well as NVIDIA Isaac Sim, for robot learning workflows. SE011, SE027, SE028
CE037 Ben Mildenhall (co-founder) authored NeRF and Mip-NeRF 360, two of the most widely cited neural rendering papers; both techniques are foundational to Marble's world model architecture. SE021, SE022
CE038 Christoph Lassner (co-founder) authored Pulsar, a differentiable sphere-based renderer architecturally related to 3D Gaussian Splatting and directly relevant to Spark's internal rendering design. SE023, SE002
CE039 Justin Johnson (co-founder) is a co-developer of PyTorch3D and has published research in 3D scene understanding and visual question answering, contributing to the founding team's 3D deep learning depth. SE002, SE025
CE040 World Labs' founding team collectively authored three foundational neural rendering technique families (NeRF/Mip-NeRF 360, Pulsar, and adjacent 3DGS research), all of which are directly instantiated in the Marble and Spark product stack. SE021, SE023, SE024
CE041 Marble Labs community showcases represent developer-led integration expansion across game engines, VR, robotics, and creative toolchains without requiring direct engineering support from World Labs. SE003, SE016, SE017, SE018, SE019, SE020
CE042 Large-scale 3D datasets such as Objaverse-XL (10M+ objects from Allen AI) represent the type of training infrastructure required for generative world models; World Labs has not disclosed its training data composition or data licensing strategy. SE034, SE006
CE043 Infinigen (Princeton Vision Lab, open-source) generates photorealistic 3D scenes procedurally using Blender, in contrast to Marble's neural generative approach; Infinigen requires scene specification while Marble accepts natural language or image prompts. SE035, SE006
CE044 No independent third-party benchmarks or evaluations comparing Marble's 3D world generation fidelity or physical accuracy against competing systems were identified in publicly accessible sources as of June 2026. SE033
CE045 World Labs publishes a security responsible disclosure page directing vulnerability reports to security@worldlabs.ai; the page describes good-faith testing guidelines and commits to no legal action for compliant reporters. SE004
CE046 The World Labs Acceptable Use Policy explicitly prohibits generating child sexual abuse material, facilitating weapons of mass destruction, developing unauthorized surveillance tools, attacking critical infrastructure, and creating deceptive synthetic media harmful to individuals. SE014
CE047 The World Labs Terms of Service became effective January 21, 2026 and applies to both worldlabs.ai and marble.worldlabs.ai, establishing the legal framework for all registered users and API customers. SE012
CE048 The World Labs Privacy Policy (effective November 12, 2025) describes collection of account information, usage data, prompt inputs, generated content interactions, and third-party integration data; it does not specify data retention periods or disclose whether prompt data is used for model training. SE013
CE049 World Labs does not currently operate a public bug bounty program and does not guarantee compensation for security vulnerability reports, per the security page published on worldlabs.ai/security. SE004
CE050 No SOC 2 Type II, ISO 27001, or NIST AI RMF conformance certification was observed on the World Labs website or in any public disclosure as of June 2026; NIST AI RMF guidance represents the applicable voluntary framework for AI risk management. SE004, SE032
CU001 World Labs' Marble platform targets three principal customer segments: individual creators on self-serve subscriptions, AI-native platform companies embedding world generation via the World API, and robotics or embodied-AI researchers generating synthetic training environments for simulation. SU029, SU001
CU002 OpenArt launched "OpenArt Worlds" as a named product feature using World Labs' generative 3D world models, enabling creators to transform a single image into a persistent, explorable 3D environment; the case study (dated May 8, 2026) describes an integration producing a live product feature rather than a prototype. SU002, SU013
CU003 OpenArt describes itself as "one of the world's leading AI creative platforms" in the World Labs case study dated May 8, 2026, indicating the OpenArt Worlds integration reaches a substantial existing creator user base, though the specific user count attributable to the OpenArt Worlds feature is not disclosed. SU002
CU004 Magnific integrated World Labs' Marble for its "3D Scenes" product — a 3D scene composition and photography tool for designers, marketers, and content teams seeking precise product placement and spatial control — and the integration is described as a launched product feature in World Labs' case study. SU009, SU015
CU005 Magnific's case study notes the platform serves "millions of designers, marketers, and content teams" daily, implying the Marble-powered 3D Scenes integration potentially reaches a large existing user base, though the number of users engaging with the Marble-powered feature specifically is not disclosed. SU009
CU006 Rosebud AI, a game-creation platform using natural language prompts, collaborated with World Labs to integrate Marble world generation into its game-building pipeline and explored multiplayer 3D world experiences; the case study describes a prototype game created during the collaboration, not a production deployment. SU010, SU016
CU007 VIVERSE, HTC's browser-based interactive 3D world platform, collaborated with World Labs to integrate Marble-generated 3DGS environments into interactive, playable browser experiences; the collaboration is described as exploratory, testing how AI-generated scenes could evolve into fully interactive experiences. SU011, SU017
CU008 Lightcraft integrated Marble with Lightcraft Jetset (built on the company's Emmy-winning Previzion system) and Beeble for virtual production, enabling filmmaker Joshua Kerr to transform iPhone-captured imagery into a 3D Gaussian Splat film set for an indie zombie movie; the case study is a named filmmaker showcase, not a platform-scale deployment. SU008, SU014
CU009 Escape, an interactive experience platform, is using the World API to turn 2D films into navigable 3D environments, enabling new forms of interactive and social storytelling; this use case is referenced in the World API launch announcement as an early-production integration. SU001, SU018
CU010 Robotics researchers Hang Yin (BEHAVIOR-1K/OmniGibson) and Abhishek Joshi (RoboSuite/Infinigen-Articulated) used Marble to generate scalable, physically accurate 3D scenes for robot simulation data collection, demonstrating research-grade use in the embodied-AI segment. SU003, SU020, SU022
CU011 The World Labs robotics case study documents Marble-generated environments used as inputs to NVIDIA Isaac Sim, MuJoCo, and RoboSuite, establishing technical compatibility across the three leading open robotics simulation frameworks; no commercial licensing relationship with NVIDIA, DeepMind (MuJoCo), or MIT (RoboSuite) is documented. SU003, SU019, SU020, SU021
CU012 Architecture and interior design use cases are documented through case studies with Fenestra and Interior AI as early collaborators for Marble's design visualization workflow, and with xFigura cited in the World API announcement as integrating world generation into professional node-based design workflows used by architecture firms. SU004, SU001
CU013 SHoP Architects is quoted in the World API announcement stating that world generation "will drastically reduce the communication gap between architectural designers and clients," providing the only named enterprise-class architect testimonial in World Labs' public corpus; no deployment size or contract terms are described. SU001
CU014 VR developer Daniel Skaale built "Splat World," a Unity-based VR first-person experience, using Marble Composer to generate 3DGS environments and developing custom C# tools for physics interaction, lighting, and density control — a documented developer adoption of Marble for interactive engine integration beyond Gaussian splat visualization. SU012
CU015 World Labs published at least nine case studies between November 12, 2025 and May 8, 2026, covering six platform integrations and multiple creator/developer showcases, representing an approximate publication velocity of 1.3 case studies per month in the seven months following Marble's general availability. SU002, SU003, SU004, SU005, SU006, SU007, SU008, SU009, SU010, SU011, SU012
CU016 The World API launched January 21, 2026 with named early integrators in gaming (Escape), architecture (Fenestra, xFigura, Preview), and robotics (Lightwheel) already using it at launch, implying pre-launch alpha or beta API access was provided to select partners prior to public release. SU001
CU017 Preview, described in the World API announcement as a professional workspace for filmmakers and AI-native studios, integrates the World API to enable directors to explore scenes, find precise camera angles, and capture 4K stills from Marble-generated immersive environments. SU001
CU018 The Marble Labs showcase at worldlabs.ai/labs lists at least 18 developer and artist projects including AI-native 3D pipelines, real estate visualization, VR locomotion tools, Unreal Engine integration via Volinga, and physics-interactive art — representing a developer-community signal of platform breadth beyond formal case studies. SU030, SU031, SU032, SU033, SU036, SU037, SU038, SU039, SU040, SU041
CU019 World Labs' named collaboration and integration evidence spans at least five distinct vertical segments: creative/film, gaming/interactive media, architectural/interior design, robotics/simulation research, and VR/immersive developer experiences. SU001, SU002, SU003, SU004, SU010, SU011, SU012
CU020 All named integrations and case studies published by World Labs as of June 2026 are described as collaborations, showcases, or experiments; none include quantitative outcome metrics such as daily active users, API call volumes, error rates, or customer satisfaction scores. SU001, SU002, SU003, SU009, SU010, SU011, SU012
CU021 No customer count (subscribers, API customers, named enterprise accounts), NRR, GRR, churn rate, or average contract value has been disclosed by World Labs as of June 2026; retention and expansion dynamics are unobservable from public evidence. SU029, SU034
CU022 World Labs' commercial usage rights begin at the Pro tier ($35/month) per its Terms of Service, meaning only Pro and Max subscribers can embed Marble outputs in paid commercial deliverables; the Free and Standard tiers serve as conversion funnels without generating defensible commercial-rights revenue. SU026, SU029
CU023 The World API serves three explicitly named verticals at launch: gaming and immersive media, robotics and simulation, and architecture and design; named early integrators cover all three verticals in the launch announcement. SU001, SU034
CU024 Autodesk's $200 million investment is described by TechCrunch as a collaboration "to explore how World Labs' models can work alongside Autodesk's tools," with an initial focus on entertainment use cases — explicitly framed as exploratory rather than a committed enterprise deployment or commercial product integration. SU028, SU034
CU025 World Labs' subscription tiers (Free, Standard $20/mo, Pro $35/mo, Max $95/mo) span individual creators through professional studios; the tier architecture creates a natural upsell ladder, but no conversion rates, tier-mix data, or subscriber counts are publicly disclosed. SU026, SU029
CU026 Deloitte's analysis of enterprise AI adoption identifies compliance complexity, workforce readiness, and evolving regulatory requirements as primary barriers to AI adoption that extend enterprise sales cycles and increase customer acquisition costs for AI platform companies broadly. SU023, SU025
CU027 Deloitte explicitly identifies AI model interpretability and trust as key adoption barriers for complex enterprise deployments — a factor directly relevant to World Labs' pursuit of robotics, AEC, and enterprise media customers where simulation fidelity and output auditability matter. SU023, SU025
CU028 Enterprise procurement for AI platforms in regulated or safety-critical industries (AEC, robotics simulation, media production) typically requires compliance verification, data privacy review, IP provenance validation, and vendor risk assessment, creating friction that extends time-to-deployment beyond initial technical evaluation; this applies directly to World Labs' target enterprise verticals. SU023, SU026, SU027
CU029 World Labs' Privacy Policy (effective November 12, 2025) covers personal information collected through Sites, Services, APIs, and Large World Models, confirming the legal infrastructure for enterprise data handling was in place at Marble's GA date. SU027, SU026
CU030 MuJoCo is described as free and open source with no licensing cost, representing a zero-friction integration pathway for robotics researchers using Marble-generated environments in simulation pipelines. SU021
CU031 Robosuite's v1.5 release supports diverse robot embodiments including humanoids, custom robot composition, and photo-realistic rendering, confirming the breadth of the robotic simulation research segment that can incorporate Marble-generated training environments. SU020
CU032 Infinigen, developed by Princeton Vision & Learning Lab, is a procedural 3D scene generator for computer vision research optimized for training data generation; it represents a complementary and competing approach for the robotics synthetic data segment alongside Marble, since Abhishek Joshi worked on Infinigen-Articulated. SU022, SU003
CU033 The global generative AI market is projected to grow from $83.3 billion in 2026 to $988.4 billion by 2035 at a 31.6% CAGR, with integration with existing enterprise software identified as a primary growth opportunity — a trajectory that supports World Labs' API and strategic partnership customer strategy. SU024
CU034 Stanford HAI's 2026 AI Index reports increasing enterprise AI adoption across multiple sectors, providing macro-level context for a growing addressable customer pool for World Labs' platform, though no World Labs-specific adoption data is cited. SU025
CU035 NVIDIA Isaac Sim is described as an open source reference framework for robotics simulation built on NVIDIA Omniverse, confirming technical openness and compatibility with AI-generated world data in the robotics research and commercial simulation segment. SU019, SU003
CU036 The Marble Labs showcase includes a project integrating Marble-generated Gaussian splats into Unreal Engine via Volinga, evidencing developer adoption in game-engine workflows beyond Unity and demonstrating the portability of Marble outputs across commercial real-time rendering environments. SU033
CU037 The Marble Labs showcase features a "Spatial Real Estate" project (spatial-blueprints), indicating adoption in real estate visualization — a vertical not covered in the main case-study corpus — suggesting breadth of early use cases beyond the six named platform integrations. SU032
CU038 An AI-native 3D pipeline project in Marble Labs was built by Matt Workman using fal and the open-source IMAGE-BLASTER toolkit, demonstrating developer integration of Marble into real-time production workflows for games, filmmaking, and interactive environments without any World Labs commercial contract requirement. SU031
CU039 No independently sourced customer reviews, ratings (G2, Capterra, Gartner Peer Insights), or third-party evaluations of Marble's enterprise performance or customer satisfaction have been identified in any publicly accessible source as of June 2026. SU029
CU040 World Labs' Acceptable Use Policy (AUP) restricts outputs used to deceive, harm, or violate intellectual property rights, establishing compliance constraints that enterprise customers in regulated industries must account for during procurement — including IP provenance requirements that may be non-trivial for media production workflows. SU035
CU041 Filmmaker Tim Simmons (Theoretically Media, 170,000+ YouTube subscribers), a former Hollywood documentarian, used Marble as a virtual film set for a 48-second AI micro-short titled "Alarm," testing shot consistency and cinematic framing across AI-generated scenes — a documented prosumer creator adoption in the Standard-to-Pro subscription segment. SU006
CU042 The World Labs team used Marble itself to produce the company's first marketing launch video, generating "hundreds of 3D worlds" over a few weeks for the production — a self-validation of the product for marketing production workflows and an internal customer adoption signal. SU007
CR001 The EU AI Act, adopted by the European Parliament in March 2024 with 523 votes in favour, establishes a risk-based regulatory framework applicable to general-purpose AI (GPAI) models, with enhanced transparency and safety obligations for developers above defined computational thresholds. SR013, SR014
CR002 World Labs' Acceptable Use Policy explicitly classifies regulated uses under the EU AI Act and California SB 1001 as "High-Risk Uses," acknowledging that Marble and the World API may be subject to these regulatory frameworks. SR011, SR013
CR003 The US Copyright Office released a pre-publication version of Part 3 of its AI report in May 2025 examining the legality of AI training data scraping; no final rule on training data copyright has been issued as of June 2026. SR002, SR012
CR004 The US Export Administration Regulations (EAR, 15 CFR 730-774) apply to AI model weights and dual-use technology; cross-border API access to large world models may require export licence review under the Commerce Control List. SR012, SR013
CR005 World Labs' Terms of Service, effective January 21 2026, contains a binding arbitration provision and class-action waiver; this limits customer litigation avenues and creates friction for enterprise procurement in jurisdictions where such clauses are unenforceable. SR009
CR006 World Labs' Privacy Policy (effective November 12 2025) discloses that the company collects spatial measurements, 3D models, geospatial data, and user-uploaded images and video as part of its world-generation services. SR010
CR007 World Labs' AUP prohibits High-Risk Uses including weapons design, surveillance, biometric identification, manipulation of human behaviour, and uses that violate fundamental rights, referencing the EU AI Act and California SB 1001. SR011
CR008 NIST's AI Risk Management Framework provides voluntary guidance for AI governance and does not carry the force of law; US AI-specific federal regulation remains fragmented as of June 2026, creating compliance uncertainty for enterprise deployments. SR014
CR009 World Labs has not publicly disclosed any third-party security certification (SOC 2, ISO 27001) or regulatory pre-approval for enterprise deployment as of June 2026, limiting enterprise procurement in regulated industries. SR001, SR009
CR010 The US Copyright Office Part 2 report (January 2025) established that AI-generated outputs lacking sufficient human authorship cannot be copyrighted in the US; enterprise customers incorporating Marble outputs into commercial workflows face an IP ownership gap. SR002
CR011 World Labs operates a responsible vulnerability disclosure programme via security@worldlabs.ai but explicitly states it does not operate a public bug bounty programme and does not guarantee compensation to security researchers. SR001
CR012 Marble generates compute-intensive 3D Gaussian splat outputs; inference at enterprise scale will require significant ongoing GPU allocations from NVIDIA or AMD infrastructure, creating an operational dependency on hardware suppliers who are also equity investors. SR021, SR031
CR013 The World API (launched January 2026) processes world generation requests asynchronously, meaning each API call triggers a full 3D generation pipeline that is substantially more compute-intensive than text or image API calls. SR022
CR014 Sim-to-real transfer remains an unresolved technical challenge for 3D generation models used in robotics and embodied-AI training; generated environments may lack physical accuracy sufficient for robot policy training. SR023, SR029
CR015 A 2026 arXiv survey on 3D generation for embodied AI identifies limited physical annotations, geometry-realism gaps, fragmented evaluation, and the persistent sim-to-real divide as the primary bottlenecks preventing 3D generation from becoming production-ready for robotic deployment. SR023
CR016 Digital twin AI research identifies real-time synchronisation, data quality, and cybersecurity as the three most persistent operational challenges for AI-based digital twin deployments. SR024
CR017 An MDPI peer-reviewed paper (September 2025) on generative AI in digital twins for Industry 5.0 highlights data heterogeneity, model drift, and integration complexity as key obstacles to enterprise deployment of generative AI in industrial settings. SR003, SR037
CR018 Deloitte's 2025/2026 AI adoption survey identifies regulatory compliance and workforce readiness as the leading barriers to enterprise adoption of agentic, physical, and sovereign AI — both directly relevant to World Labs' enterprise customer segments. SR025
CR019 Autodesk invested $200 million in World Labs' February 2026 round and assumed an advisory role; TechCrunch confirmed that Autodesk's Chief Scientist stated the commercial product shape of the partnership "hasn't been determined yet" and that data sharing is not part of the agreement. SR018, SR016
CR020 NVIDIA and AMD both participated in World Labs' $1 billion February 2026 Series B round as investors, creating a dual role in which the company's primary hardware suppliers also hold equity stakes and potential board access. SR016, SR031, SR015
CR021 Andreessen Horowitz led World Labs' seed round ($230M, September 2024) and participated in the Series B ($1B, February 2026), making it the only investor present across both rounds and creating a single-firm concentration in the investor base. SR017, SR016
CR022 World Labs has not disclosed whether it operates proprietary GPU data-centre infrastructure or relies on hyperscale cloud providers for training and inference compute; this information gap prevents independent assessment of compute cost and resilience.
CR023 World Labs' robotics simulation use case depends on integration with NVIDIA Isaac Sim, creating a platform dependency on NVIDIA that is compounded by NVIDIA's status as a Series B investor and direct competitor via NVIDIA Cosmos. SR030, SR031
CR024 NVIDIA Cosmos, a directly competing world foundation model developed for physical AI and robotics simulation, is built by a company with substantially larger compute budgets, an existing robotics developer ecosystem (Isaac Sim), and GPU supply-chain advantages. SR031, SR032
CR025 The Gartner Generative AI Hype Cycle 2026 indicates that some generative AI categories are approaching the peak-of-inflated-expectations phase; a subsequent trough could compress spatial-AI valuation multiples and constrain World Labs' future fundraising windows. SR004
CR026 Crunchbase Q1 2026 data shows that 65% of global venture investment was captured by four companies; follow-on capital conditions for mid-tier frontier AI startups will tighten if spatial AI fails to demonstrate measurable revenue. SR027, SR028
CR027 Fei-Fei Li holds a concurrent Sequoia Professorship at Stanford and the CEO position at World Labs; Reuters reported at founding that "Li will continue some of her work at Stanford while building the startup," constituting a split-time key-person risk. SR033, SR017
CR028 World Labs co-founder Justin Johnson holds a faculty appointment at the University of Michigan (EECS); a split-time arrangement at a critical commercialisation phase adds execution risk, particularly for technical direction. SR017
CR029 World Labs emerged from stealth in September 2024 with a $230 million raise before shipping any product, creating a period of elevated competitive intelligence exposure during which the company's technical direction was public but unvalidated. SR017, SR033
CR030 World Labs has not disclosed headcount, burn rate, or revenue as of June 2026; the absence of financial transparency makes independent assessment of runway, capital efficiency, and commercialisation pace impossible. SR020, SR021
CR031 StartupHub analysis notes that World Labs' $1.23 billion cumulative raise arrived before any disclosed revenue, placing it in the capital-intensive pre-revenue frontier AI category and underscoring the long-duration capital-to-value conversion bet. SR020
CR032 World Labs competes for ML engineering and 3D graphics talent with NVIDIA, Google DeepMind, Meta Reality Labs, and well-capitalised peers including Luma AI, all of whom offer greater compensation certainty and established research environments. SR033, SR031, SR032
CR033 Deloitte's survey identifies workforce readiness as the top enterprise AI adoption barrier, meaning World Labs' enterprise customers face parallel talent constraints that could slow integration timelines and reduce the effective addressable market for world-model deployments. SR025
CR034 World Labs has raised $1.23 billion in total across a $230 million seed round (September 2024) and a $1 billion Series B (February 2026); the Series B was closed at a rumoured but unconfirmed $5 billion valuation per Bloomberg (January 2026). SR016, SR018, SR020
CR035 No revenue, ARR, customer count, or gross margin has been disclosed by World Labs as of June 2026; the financial model remains entirely speculative and cannot be benchmarked against disclosed operating metrics. SR020, SR030
CR036 World Labs' consumer subscription pricing (Free, $20/month Standard, $35/month Pro, $95/month Max) cannot support frontier model training costs without substantial enterprise contract revenue or API usage fees at scale. SR021, SR022
CR037 Crunchbase data shows foundational AI startup funding doubled in Q1 2026 vs full-year 2025, but the increase was dominated by OpenAI ($122B), Anthropic ($30B), and xAI ($20B); World Labs must compete for follow-on capital against better-capitalised peers. SR027, SR028
CR038 Autodesk's Chief Scientist confirmed in Q1 2026 that the form of the World Labs partnership "hasn't been determined yet" and that no data sharing is involved; if no joint commercial product is announced by Q3 2027, the strategic investment rationale weakens and renewal of the advisory arrangement becomes uncertain. SR018, SR019
CR039 World Labs has not disclosed the licensing basis or data sourcing practices for Marble's world-model training data, which likely includes web-scraped images, video, and 3D scans subject to competing IP ownership claims. SR006, SR021
CR040 The global robotics simulation software market is projected to grow substantially through 2030, but growth depends on hardware capability improvements and operator adoption timelines that remain outside World Labs' control. SR005
CR041 Stanford HAI's 2026 AI Index identifies AI's impact on productivity and labour markets as a significant economic uncertainty; regulatory responses could create overhang for frontier model developers including World Labs. SR026, SR035
CR042 Foundational AI startup funding doubled in Q1 2026 vs all of 2025 (Crunchbase), but the increase was heavily concentrated in a small number of companies; World Labs competes for capital in an environment that rewards disclosed revenue metrics it has not yet provided. SR028, SR027
CR043 World Labs' AUP prohibits malicious uses including weapons design, automated surveillance, disinformation, and content violating fundamental rights; however, the AUP does not describe automated enforcement mechanisms for API-level misuse detection. SR011, SR009
CR044 World Labs' launch video for Marble was produced using Marble itself; while this demonstrates product capability, it is self-referential evidence produced by the company rather than independent commercial validation of enterprise-grade output quality. SR007
CR045 The Objaverse-XL dataset from Allen Institute for AI provides over 10 million open 3D objects for training, substantially lowering the data-moat barrier for competitor world models; World Labs' defensibility depends on model architecture and proprietary training recipes rather than exclusive data access alone. SR008, SR041
CR046 World Labs' Marble Labs showcase extends the product surface to VR, AI-native 3D creation, artistic films, and game-development integrations, expanding the regulatory, IP, and liability surface beyond the core world-generation API into consumer-facing creative applications. SR039, SR040, SR042, SR043
CV001 World Labs raised $230 million in a seed round in September 2024, led by Andreessen Horowitz, New Enterprise Associates, and Radical Ventures, with AMD Ventures, Intel Capital, and NVIDIA NVentures also participating. SV019, SV017
CV002 World Labs raised $1 billion in a second funding round in February 2026, with Autodesk as the anchor investor at $200 million, joined by AMD, NVIDIA, Emerson Collective, Fidelity Management and Research, and Sea Group. SV011, SV012, SV013
CV003 Total capital raised by World Labs across both funding rounds stands at $1.23 billion as reported consistently by multiple independent news sources. SV015, SV012, SV013
CV004 Bloomberg News reported in January 2026 that World Labs was in funding discussions at a valuation of approximately $5 billion. SV012, SV015, SV016
CV005 World Labs declined to confirm the Bloomberg-reported $5 billion valuation figure, per Reuters and Chiang Rai Times reporting on the February 2026 fundraise; the company also did not explicitly contest the figure. SV012, SV016
CV006 World Labs' seed-round investor syndicate included Andreessen Horowitz (lead), NEA, Radical Ventures, AMD Ventures, Intel Capital, and NVIDIA NVentures — spanning tier-1 VC and major semiconductor companies from inception. SV019, SV017
CV007 The February 2026 Series B investors — Autodesk, AMD, NVIDIA, Fidelity, Emerson Collective, and Sea Group — include strategic holders across design software, semiconductor hardware, institutional asset management, and consumer internet, suggesting a multi-dimensional strategic rationale for the round. SV011, SV013
CV008 Autodesk's $200 million investment positions it as a strategic adviser, with the partnership to operate at the research and model level and an initial focus on entertainment use cases; no commercial distribution or revenue-share agreement has been disclosed. SV013, SV012
CV009 Marble was made generally available in November 2025 with subscription tiers at Free, $20 per month (Standard), $35 per month (Pro, commercial rights), and $95 per month (Max). SV027
CV010 The World API was launched publicly on January 21, 2026, providing programmatic access to Marble world generation from text, images, panoramas, and video inputs. SV028
CV011 World Labs had not disclosed any revenue, ARR, customer count, or unit economics as of June 2026, approximately seven months after Marble's general availability and five months after the World API launch. SV015, SV026
CV012 Named commercial integrations as of June 2026 include OpenArt, Magnific AI, Rosebud AI, VIVERSE, Lightcraft, Escape, and NVIDIA Isaac Sim researchers, spanning creative, gaming, enterprise design, and robotics research verticals. SV027, SV003
CV013 World Labs' seed-round post-money valuation was reported at $1 billion by Bloomberg, representing a 4.3x ratio of post-money valuation to capital raised at the seed stage. SV012, SV019
CV014 The Bloomberg-reported Series B implied valuation of approximately $5 billion represents a 5x increase from the $1 billion seed mark in roughly 16 months — among the faster valuation step-ups recorded for frontier AI companies in the 2024–2026 vintage. SV015, SV012
CV015 Anthropic closed a $30 billion Series G funding round at a post-money valuation of $380 billion in February 2026, reflecting a mature commercial AI company with disclosed revenue; this represents the far end of the frontier-AI comparable set. SV021, SV020
CV016 Global venture funding in Q1 2026 reached a record $300 billion, with $242 billion (80%) going to AI companies — the highest single-quarter AI funding total ever recorded. SV020, SV021
CV017 Foundational AI startup funding in Q1 2026 totaled $178 billion, double the $88.9 billion raised in all of 2025, reflecting unprecedented capital velocity into frontier AI companies. SV021, SV020
CV018 OpenAI ($122B), Anthropic ($30B), and xAI ($20B) collectively raised 65% of total global venture investment in Q1 2026, representing extreme capital concentration in three companies out of the entire global venture ecosystem. SV021, SV020
CV019 The global generative AI market was valued at $53.7 billion in 2025 and is projected by Global Market Insights to reach $988.4 billion by 2035 at a CAGR of 31.6%, supporting the large-market thesis behind spatial AI investment. SV022
CV020 The global digital twin market is projected to grow from $21.14 billion in 2025 to $149.81 billion in 2030 at a CAGR of 47.9%, representing an adjacent and highly relevant addressable market for World Labs' world-model platform. SV023
CV021 Stanford HAI's 2026 AI Index Report documents the broad growth of AI investment and enterprise adoption, providing a corroborating macro backdrop for the spatial AI investment thesis. SV005, SV025
CV022 Gartner maintains a Hype Cycle for Generative AI 2026 that tracks the maturity curve of generative AI technologies; the document is paywalled but its existence confirms that Gartner is actively monitoring this category for hype-cycle positioning. SV007
CV023 Deloitte's enterprise AI adoption research identifies compliance complexity, workforce readiness, and regulatory uncertainty as the primary friction points that extend sales cycles for AI platform companies aiming at enterprise customers. SV024
CV024 Crunchbase data shows that capital concentration in three foundational AI companies (OpenAI, Anthropic, xAI) compressed available funding for all other AI startups, creating a structurally more challenging fundraising environment for mid-tier companies like World Labs in their subsequent capital raises. SV021
CV025 The U.S. Copyright Office has been actively examining copyright in AI-generated content since 2023, including Part 2 (January 2025) addressing copyrightability of generative AI outputs — this regulatory process creates uncertainty for the commercial rights basis of World Labs' Pro and Max subscription tiers. SV006
CV026 Bull case assumption (estimate only, not a forecast): if World Labs achieves an estimated $100-$300 million ARR by 2028 via enterprise channels led by Autodesk distribution, implied valuation at 30-40x forward revenue multiples would be approximately $3-12 billion — bracketing the current implied Series B price. SV013, SV022
CV027 Base case assumption (estimate only): if World Labs achieves an estimated $30-$80 million ARR by 2028 through direct subscription and API revenue at 25-30x forward multiples, implied valuation would be approximately $750 million to $2.4 billion — below the current implied Series B price. SV015, SV022
CV028 Bear case assumption (estimate only): if spatial AI commoditises and World Labs achieves only an estimated $5-$15 million ARR by 2028 at 10-15x forward multiples, implied valuation would be approximately $50-225 million, far below the $1.23 billion of capital invested and the implied $5 billion entry price. SV007, SV022
CV029 Andreessen Horowitz's World Labs investment thesis holds that world models represent the spatial-intelligence equivalent of language models — general-purpose AI that will transform every domain requiring 3D reasoning, from robotics to creative content. SV017, SV001
CV030 A16z explicitly cited the technical difficulty of building world models — requiring simultaneous advances in data, graphics, and AI — as a barrier to entry that protects World Labs' market position. SV017
CV031 World Labs used Marble to create its own product launch video, generating hundreds of 3D worlds across multiple scenes from text and image prompts — demonstrating the product's capability without relying on external customer proof points. SV003
CV032 The Spark 2.0 renderer enables streaming of large-scale 3D Gaussian Splatting worlds on any web browser, including mobile and VR, using a Level-of-Detail system — this extends the commercial surface for Marble beyond desktop-only use. SV010
CV033 World Labs' investor composition — including a16z, AMD, NVIDIA, Autodesk, and Fidelity — spans financial, strategic hardware, and enterprise software holders, creating multiple potential exit pathways including M&A by existing strategic investors. SV018, SV012
CV034 The most probable primary exit pathway for World Labs is an acquisition by Autodesk or a semiconductor strategic (AMD/NVIDIA) once commercial revenue proof is established, given these parties' existing investment, advisory relationships, and strategic interest in owning spatial AI infrastructure. SV013, SV018
CV035 A down-round — a fundraise at or below the implied $5 billion Series B mark — would constitute a thesis-break trigger requiring immediate reassessment of the investment case for World Labs. SV015, SV012
CV036 An IPO for World Labs at current development stage — pre-revenue, early commercial traction — is unlikely before the company demonstrates at least $50 million in ARR with consistent quarter-over-quarter growth, based on historical pre-IPO commercial benchmarks for AI companies. SV025, SV020
CV037 The recommended investment stance for World Labs as of June 2026 is TRACK: the technology thesis and founding team are exceptional, but the implied $5 billion valuation is not anchored by disclosed revenue and requires commercial proof — specifically ARR above $20 million or Autodesk commercial distribution — before a buy recommendation can be supported. SV011, SV015, SV022
CV038 Investment confidence in World Labs is rated medium: there is strong qualitative evidence supporting the technology thesis, founder quality, and strategic investor conviction, but no revenue, unit economics, or commercial contract data to anchor the implied $5 billion valuation with precision. SV015, SV026
CV039 Luma AI, which operates in adjacent generative video and 3D AI, is a private company with no publicly disclosed valuation at a scale comparable to World Labs; the absence of a public Luma AI comparable limits the peer-group reference set. SV022, SV021
CV040 NVIDIA Cosmos is available as a free, open-weight physical AI world model for certain embodied and robotics use cases, creating a zero-cost reference point that materially limits World Labs' pricing ceiling in the robotics and physical AI segment. SV008, SV024
CV041 At the implied $5 billion post-money valuation with $1.23 billion raised, World Labs' ratio of total capital raised to post-money is approximately 25%, which is higher than the 10-20% typical for frontier AI companies in comparable rounds — suggesting a less aggressive valuation premium relative to capital deployment. SV015, SV021
CV042 World Labs had approximately 11-50 employees as of June 2026 with $1.23 billion in capital raised; burn rate is not disclosed, but the lean headcount implies either significant compute infrastructure spend or capital being preserved for future operational scaling. SV015
CV043 The Marble Labs showcase catalogue demonstrates third-party developer use of Marble for creative and enterprise use cases — including "Dormant Memories" (merging real-captured spaces with generated alternate realities) and "Mint" (a chat-first 3D creation platform for interactive world publishing) — signalling commercial breadth across artistic and developer segments. SV029, SV030
CV044 Andreessen Horowitz’s September 2024 investment note argued that world models could expand AI from language into interactive 3D environments and simulation, supporting a strategic premium narrative around World Labs. SV031
CV045 World Labs appearing in Andreessen Horowitz’s portfolio provides evidence of continued investor sponsorship, which can support future fundraising even though it does not itself validate valuation. SV032
CV046 IEEE Spectrum’s December 2024 profile of Fei-Fei Li documented live world-model demos and her spatial-intelligence thesis, helping explain why investors may ascribe unusual option value to technical pedigree and ambition. SV033
CV047 The Dwarven Caves showcase indicates that Marble can support large connected gaming-style environments, widening the upside case beyond narrow developer tooling into higher-value immersive content workflows. SV034
CV048 The Musée du Monde showcase suggests creative breadth and internal experimentation velocity, but it also underscores that showcase diversity is not equivalent to recurring commercial demand. SV035
CV049 Christoph Lassner’s TEDAI Vienna talk extends the upside narrative from tooling toward interactive AI-powered entertainment, supporting a larger long-duration option value case for spatial-intelligence platforms. SV036
CV050 The ICARE case study says a three-person team built a large web-native interactive world in roughly two months using Marble and Spark 2.0, suggesting the platform could compress content-production time for prospective customers. SV037
CV051 World Labs publishing a Getting Started with Marble tutorial indicates the company is beginning to package onboarding content rather than relying only on bespoke demos, which can matter for eventual conversion economics. SV038
CV052 The Chisel tutorial shows World Labs expanding from raw world generation into controllable 3D layout editing, which can increase willingness to pay if customers need more deterministic outputs. SV039
来源
编号出版方标题引文
SO001 World Labs World Labs — Spatial Intelligence Homepage World Labs is a leading spatial intelligence company, building frontier models that can perceive, generate, reason, and interact with the 3D world.
SO002 World Labs About | World Labs World Labs was founded by visionary AI pioneer Fei-Fei Li along with Justin Johnson, Christoph Lassner, and Ben Mildenhall.
SO003 World Labs World Labs Blog
SO006 World Labs World Labs Announces New Funding World Labs has raised $1 billion in new funding. We are grateful and excited to partner with our investors, including AMD, Autodesk, Emerson Collective, Fidelity Management & Research Company, NVIDIA, and Sea, among others.
SO007 World Labs Announcing the World API Today, we're launching the World API — a public interface for generating explorable 3D worlds using World Labs' multimodal world model, Marble.
SO008 World Labs Marble: A Multimodal World Model Marble can create 3D worlds from text, images, video, or coarse 3D layouts; Marble also lets you interactively edit, expand, and combine worlds.
SO009 World Labs Spatial Intelligence — TED Talk Feature
SO010 World Labs A Functional Taxonomy of World Models
SO015 World Labs From Words to Worlds: Spatial Intelligence is AI's Next Frontier It's why my cofounders Justin Johnson, Christoph Lassner, Ben Mildenhall, and I created World Labs more than one year ago: to realize this possibility in full.
SO028 Reuters AI 'godmother' Fei-Fei Li raises $230 million to launch AI startup Fei-Fei Li, a leading artificial intelligence researcher, has raised $230 million for a startup she and three colleagues founded to make AI technology that can understand how the three-dimensional physical world works.
SO029 Reuters AI pioneer Fei-Fei Li's World Labs raises $1 billion in funding Autodesk invested $200 million in World Labs and will serve as an adviser to the startup, the design software maker said.
SO030 Andreessen Horowitz What's In a World? Investing in World Labs So when she approached us last year saying now was the time to solve the world-model problem, we were hooked!
SO032 Stanford University Fei-Fei Li — Stanford Computer Science Faculty Profile She is currently a Co-founder/CEO of World Labs, an AI company focusing on Spatial Intelligence and generative AI.
SO033 Stanford HAI Fei-Fei Li | Stanford HAI
SO034 Stanford HAI HAI Co-Director Fei-Fei Li Joins UN Secretary-General's Scientific Advisory Board
SO035 Ben Mildenhall Ben Mildenhall — Personal Research Page
SO036 Christoph Lassner Christoph Lassner — Biography and Research
SO037 Justin Johnson / University of Michigan Justin Johnson — CS Profile
SO038 arXiv NeRF: Representing Scenes as Neural Radiance Fields for View Synthesis (arXiv:2003.08934)
SO040 arXiv Pulsar: Efficient Sphere-based Neural Rendering (arXiv:2004.07484)
SO042 IEEE Spectrum AI Pioneer Fei-Fei Li Has a Vision for Computer Vision
SO054 Wired Fei-Fei Li Is Determined to Make AI Good for Humanity
SO055 TheOrg World Labs — Company Profile
SO056 Ashby / World Labs World Labs Jobs Board
SO096 World Labs World Labs Terms of Service World Labs Technologies, Inc.
SO097 World Labs World Labs Privacy Policy
SO098 World Labs World Labs Acceptable Use Policy
SO099 Bureau of Industry and Security (BIS) Export Administration Regulations (EAR), 15 CFR Parts 730-774
SO100 European Parliament Parliament adopts landmark law on artificial intelligence The regulation establishes obligations for AI based on its potential risks and level of impact.
SO103 NIST NIST Artificial Intelligence
SO105 TechCrunch World Labs lands $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.
SO106 SiliconANGLE World Labs closes $1B investment backed by Nvidia, AMD and Autodesk
SO107 StartupHub AI Fei-Fei Li Company Financial Breakdown 2026 Total capital raised across both rounds stands at $1.23 billion.
SO108 Chiang Rai Times World Labs lands $1 billion
SO110 Crunchbase News Q1 2026: Record-Breaking Global Venture Funding
SM001 World Labs Scaling Robotic Simulation with Marble High-quality, diverse simulation data is one of the biggest limiting factors in robotics research.
SM002 World Labs Reframing Space: How Marble Is Transforming Architectural and Interior Visualization
SM003 World Labs NVIDIA Isaac Sim and Marble
SM004 World Labs Framing Worlds: How Marble Helps Creators Bring Consistency to AI Filmmaking
SM005 Google DeepMind Genie 2: A large-scale foundation world model Genie 2, a foundation world model capable of generating an endless variety of action-controllable, playable 3D environments for training and evaluating embodied agents.
SM006 Google DeepMind Genie: Generative Interactive Environments
SM007 Google DeepMind A generalist AI agent for 3D virtual environments
SM008 NVIDIA NVIDIA Cosmos The Open Physical AI Foundation Model — Develop physical AI faster with leading world foundation models.
SM009 Luma AI Luma | AI Agents for Creative Work
SM010 Luma AI Luma AI - Interactive Scenes 30 FPS on Web browsers; 8 MB for Objects; 20 MB for Scenes.
SM011 NVIDIA Developer Isaac Sim
SM012 robosuite robosuite
SM013 MuJoCo MuJoCo — Advanced Physics Simulation
SM014 Infinigen Home | Infinigen
SM015 arXiv (Ye et al., cs.RO) 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.
SM016 arXiv (Zhou et al., cs.AI) Digital Twin AI: Opportunities and Challenges from Large Language Models to World Models
SM017 MDPI Applied Sciences Role of Generative AI in AI-Based Digital Twins in Industry 5.0 and Evolution to Industry 6.0
SM018 Deloitte AI trends: Adoption barriers and updated predictions According to nearly 60% of the AI leaders and representatives surveyed, their organization's primary challenges in adopting agentic AI are integrating with legacy systems and addressing risk and compliance concerns.
SM019 Global Market Insights (GMI) Generative AI Market Size & Share | Forecast Report 2026-2035 The global generative AI market was valued at USD 53.7 billion in 2025. The market is expected to grow from USD 83.3 billion in 2026 to USD 988.4 billion in 2035 at a CAGR of 31.6%.
SM020 MarketsandMarkets Digital Twin Market size report 2024-2030 The global digital twin market is expected to grow from USD 21.14 billion in 2025 to USD 149.81 billion in 2030 at a CAGR of 47.9%.
SM021 Stanford HAI Economy | The 2026 AI Index Report
SM022 World Labs 3D as code Text became the universal interface for software; 3D is becoming the universal interface for space.
SM023 World Labs Streaming 3DGS worlds on the web (Spark 2.0)
SM024 World Labs Announcing the World API
SM025 World Labs Marble: A Multimodal World Model
SM026 World Labs A Functional Taxonomy of World Models
SM027 World Labs With spatial intelligence, AI will understand the real world
SM028 TechCrunch World Labs lands $1B, with $200M from Autodesk, to bring world models into 3D workflows
SM029 NIST Artificial intelligence — NIST
SM030 Crunchbase News Q1 2026 Shatters Venture Funding Records As AI Boom Pushes Startup Investment To $300B AI shattered records last quarter, with $242 billion — 80% of total global venture funding in Q1— going to companies in the sector.
SP001 World Labs Marble Labs | World Labs
SP002 World Labs Announcing the World API
SP003 World Labs Marble: A Multimodal World Model a first-in-class generative multimodal world model
SP004 World Labs Streaming 3DGS worlds on the web — Spark 2.0
SP005 World Labs Scaling Robotic Simulation with Marble
SP006 World Labs Reframing Space: How Marble Is Transforming Architectural and Interior Visualization
SP007 World Labs Framing Worlds: How Marble Helps Creators Bring Consistency to AI Filmmaking
SP008 World Labs From Backyard to Blockbuster: How Lightcraft and Beeble Empowered a New Kind of Indie Filmmaking
SP009 World Labs From Image to Studio: How Magnific Turned 3D Into a Creative Workflow Magnific is one of the largest creative platforms in the world. Millions of designers, marketers, and content teams use it daily
SP010 World Labs OpenArt Worlds: Directing Stories Inside AI-Generated Worlds
SP011 World Labs From World Generation to Multiplayer: Rosebud AI x Marble
SP012 World Labs Building Worlds Together: How VIVERSE and Marble Empowered Creators to Build Interactive 3D Worlds
SP013 World Labs Splat World: Exploring New Dimensions of Gaussian Splatting in VR
SP014 Google DeepMind Genie 2: A large-scale foundation world model Genie 2, a foundation world model capable of generating an endless variety of action-controllable, playable 3D environments for training and evaluating embodied agents
SP015 Google DeepMind Genie: Generative Interactive Environments
SP016 Google DeepMind A generalist AI agent for 3D virtual environments (SIMA)
SP017 NVIDIA NVIDIA Cosmos Cosmos 3 — The Open Physical AI Foundation Model
SP018 Luma AI Luma | AI Agents for Creative Work Our Mission is to build unified general intelligence that can generate, understand, and operate in the physical world
SP019 Luma AI Luma AI — Interactive Scenes 30 FPS on Web browsers; 8 MB for Objects; 20 MB for Scenes
SP020 Stability AI Introducing Stable Zero123: Quality 3D Object Generation from Single Images This model is being released for non-commercial and research use, and the weights can be downloaded here.
SP021 arXiv Magic123: One Image to High-Quality 3D Object Generation Using Both 2D and 3D Diffusion Priors
SP022 OpenArt OpenArt: AI Art Generator — Free AI Image, Video & Audio Generator
SP023 Lightcraft Lightcraft — A Movie Studio in Your Pocket
SP024 Magnific Magnific (formerly Freepik) | The AI Creative Platform
SP025 Rosebud AI Rosebud AI: Make 3D Games & Worlds with Vibe Coding
SP026 VIVERSE VIVERSE | Browser Games, Videos, 3D Art & Virtual Worlds
SP027 NVIDIA NVIDIA Isaac Sim Isaac Sim is fully extensible, so developers can build custom OpenUSD-based simulators or integrate framework capabilities into existing testing and validation pipelines.
SP028 Global Market Insights Generative AI Market Size & Share | Forecast Report 2026-2035
SP029 Crunchbase News Sector Snapshot: Venture Funding To Foundational AI Startups In Q1 Was Double All Of 2025 That funding is increasingly concentrated in a handful of foundational giants, including OpenAI, Anthropic and xAI... a small number of companies capturing a disproportionate share of global capital.
SI001 World Labs World Labs Announces New Funding World Labs has raised $1 billion in new funding. Investors include AMD, Autodesk, Emerson Collective, Fidelity Management & Research Company, NVIDIA, and Sea, among others.
SI002 World Labs Announcing the World API Today, we're launching the World API — a public interface for generating explorable 3D worlds using World Labs' multimodal world model, Marble.
SI003 Reuters 'AI godmother' Fei-Fei Li raises $230 million to launch AI startup Initial funding for World Labs was led jointly by Andreessen Horowitz, New Enterprise Associates and Radical Ventures. Other investors included AMD Ventures, Intel Capital and NVIDIA's NVentures.
SI004 Reuters AI pioneer Fei-Fei Li's World Labs raises $1 billion in funding Autodesk invested $200 million in World Labs and will serve as an adviser to the startup, the design software maker said.
SI005 Andreessen Horowitz What's In a World? Investing in World Labs Language models changed how we interact with computers because they enabled software to speak and understand natural languages. Spatial intelligence is the next such step.
SI006 Andreessen Horowitz Portfolio | Andreessen Horowitz
SI007 The Org World Labs | The Org
SI008 World Labs (via Ashby) Jobs — World Labs
SI009 World Labs OpenArt Worlds: Directing Stories Inside AI-Generated Worlds
SI010 World Labs Scaling Robotic Simulation with Marble
SI011 World Labs Reframing Space — Marble Architectural and Interior Visualization
SI012 World Labs NVIDIA Isaac Sim and Marble
SI013 World Labs Framing Worlds — Marble for Consistent AI Filmmaking
SI014 World Labs From Backyard to Blockbuster: Lightcraft and Beeble with Marble
SI015 World Labs From Image to Studio: How Magnific Turned 3D Into a Creative Workflow
SI016 World Labs From World Generation to Multiplayer: Rosebud AI x Marble
SI017 World Labs Building Worlds Together: VIVERSE and Marble
SI018 OpenArt AI Art Generator: Free AI Image, Video & Audio Generator
SI019 Lightcraft Lightcraft — A Movie Studio in Your Pocket
SI020 Magnific Magnific (formerly Freepik) | The AI Creative Platform
SI021 Rosebud AI Rosebud AI: Make 3D Games & Worlds with Vibe Coding
SI022 HTC VIVERSE VIVERSE | Browser Games, Videos, 3D Art & Virtual Worlds
SI023 Escape Escape — Interactive Experiences
SI024 NVIDIA Isaac Sim — NVIDIA Developer
SI025 Deloitte AI trends: Adoption barriers and updated predictions Organizations must overcome technical limitations, manage operational complexities, address evolving regulatory and compliance requirements, and ensure their teams are ready for change.
SI026 Global Market Insights Generative AI Market Size & Share | Forecast Report 2026-2035 The global generative AI market was valued at USD 53.7 billion in 2025. The market is expected to grow from USD 83.3 billion in 2026 to USD 988.4 billion in 2035 at a CAGR of 31.6%.
SI027 Stanford HAI Economy | The 2026 AI Index Report | Stanford HAI
SI028 TechCrunch World Labs lands $1B, with $200M from Autodesk, to bring world models into 3D workflows The partnership between World Labs and Autodesk will see the two companies collaborating to explore how World Labs' models can work alongside Autodesk's tools, starting with a focus on entertainment use cases.
SI029 SiliconANGLE World Labs closes $1B investment backed by Nvidia, AMD and Autodesk
SI030 StartupHub Fei-Fei Li's World Labs: $1.23B Raised, Marble Now Shipping World Labs emerged from stealth in September 2024 with $230 million in financing valuing the company at $1 billion. Total capital raised across both rounds stands at $1.23 billion.
SI031 Chiang Rai Times World Labs Lands $1 Billion Round Backed By Nvidia, AMD, And Autodesk
SI032 Crunchbase News Q1 2026 Shatters Venture Funding Records As AI Boom Pushes Startup Investment To $300B The first quarter of 2026 was unlike any other for venture investment. AI shattered records, with $242 billion going to companies in the sector — 80% of total global venture funding in Q1.
SI033 Crunchbase News Sector Snapshot: Venture Funding To Foundational AI Startups In Q1 Was Double All Of 2025 Funding is increasingly concentrated in a handful of foundational giants, including OpenAI, Anthropic and xAI. In 2025 and early 2026, the market saw a shift to a small number of companies capturing a disproportionate share of global capital.
SI034 World Labs Terms of Service This TOS applies to your access to and use of the websites and Services. By affirmatively consenting to use the Services, creating an Account, or accessing or using the Services, you accept the Terms of Service.
SI035 World Labs Privacy Policy
SI036 Stability AI News & Updates — Stability AI
SI037 World Labs World Labs — Spatial Intelligence
SI038 Andreessen Horowitz What’s In a World? Investing in World Labs Language models changed how we interact with computers because they enabled software to speak and understand natural languages.
SI039 World Labs Dwarven Caves | Marble Labs Showcase Dwarf Caves is a massive interconnected fantasy environment created to push the scale and technical limits of Marble.
SI040 World Labs Musée du Monde | Marble Labs Showcase Musée Du Monde is an interactive 3D museum experience that lets users step inside paintings and explore them from the artist’s perspective.
SE001 World Labs World Labs — Spatial Intelligence "World Labs is a leading spatial intelligence company, building frontier models that can perceive, generate, reason, and interact with the 3D world."
SE002 World Labs About | World Labs
SE003 World Labs Marble Labs | World Labs
SE004 World Labs Security | World Labs "We do not currently operate a public bug bounty program and do not guarantee compensation for reports."
SE005 World Labs Announcing the World API "Today, we're launching the World API — a public interface for generating explorable 3D worlds using World Labs' multimodal world model, Marble."
SE006 World Labs Marble: A Multimodal World Model "Marble is the first of its kind - a next-generation world model making strides toward this vision. It can now create 3D worlds from a wide variety of input types, and lets users iteratively edit or expand worlds."
SE007 World Labs A Functional Taxonomy of World Models
SE008 World Labs Generating Worlds
SE009 World Labs Spark 2.0: Streaming 3DGS Worlds on the Web "Spark started out as an internal 3DGS renderer developed by World Labs because existing web renderers had shortcomings that would limit us in the future."
SE010 World Labs 3D as Code
SE011 World Labs Reimagining Robot Simulation with Marble
SE012 World Labs World Labs Terms of Service
SE013 World Labs World Labs Privacy Policy
SE014 World Labs World Labs Acceptable Use Policy "World Labs: Acceptable Use Policy — Last modified: November 12, 2025"
SE015 World Labs Bringing Marble to Life
SE016 World Labs Marble Labs Showcase: AI-Native 3D Pipelines with fal
SE017 World Labs Marble Labs Showcase: Collider Builder
SE018 World Labs Marble Labs Showcase: Splat Raycasting for Robot Locomotion
SE019 World Labs Marble Labs Showcase: Third-Person Character Controller
SE020 World Labs Marble Labs Showcase: STARSPEED
SE021 arXiv / ECCV 2020 NeRF: Representing Scenes as Neural Radiance Fields for View Synthesis "We present a method that achieves state-of-the-art results for synthesizing novel views of complex scenes by optimizing an underlying continuous volumetric scene function."
SE022 arXiv / CVPR 2022 Mip-NeRF 360: Unbounded Anti-Aliased Neural Radiance Fields "Our model reduces mean-squared error by 57% compared to mip-NeRF, and is able to produce realistic synthesized views and detailed depth maps for highly intricate, unbounded real-world scenes."
SE023 arXiv Pulsar: Efficient Sphere-based Neural Rendering "Pulsar executes orders of magnitude faster than existing techniques and allows real-time rendering and optimization of representations with millions of spheres."
SE024 arXiv / ACM TOG 2023 3D Gaussian Splatting for Real-Time Radiance Field Rendering "We achieve state-of-the-art visual quality while maintaining competitive training times and importantly allow high-quality real-time (>= 30 fps) novel-view synthesis at 1080p resolution."
SE025 IEEE Spectrum AI Pioneer Fei-Fei Li Has a Vision for Computer Vision
SE026 NVIDIA NVIDIA Isaac Sim
SE027 robosuite robosuite: A Modular Simulation Framework and Benchmark for Robot Learning
SE028 MuJoCo MuJoCo: Advanced Physics Simulation
SE029 World Labs Architecture and Design: Marble for Spatial Visualization
SE030 World Labs Creative Film: Building an AI Micro-Short with Marble
SE031 World Labs Splat World: Exploring Gaussian Splatting in VR
SE032 NIST NIST Artificial Intelligence Programs
SE033 arXiv 3D Generation for Embodied AI: 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."
SE034 Allen Institute for AI Objaverse: A Universe of 10M+ 3D Objects
SE035 Princeton Vision and Learning Lab Infinigen: Procedural Generation of 3D Scenes
SE036 TechCrunch World Labs lands $200M from Autodesk to bring world models into 3D workflows
SU001 World Labs Announcing the World API "Platforms like Escape.ai are using the World API to turn 2D films into navigable 3D environments. SHoP Architects: 'The addition of a third dimension in seconds is incredible. From images to spatial experiences, this technology will drastically reduce the communication gap between architectural designers and clients.'"
SU002 World Labs OpenArt Worlds: Directing Stories Inside AI-Generated Worlds "OpenArt is one of the world's leading AI creative platforms, built to help anyone bring visual stories to life."
SU003 World Labs Scaling Robotic Simulation with Marble "Hang Yin and Abhishek Joshi … used Marble's world-generation technology to automatically create scalable, physically accurate 3D scenes for simulation and data collection."
SU004 World Labs Reframing Space: How Marble Is Transforming Architectural and Interior Visualization
SU005 World Labs NVIDIA Isaac Sim and Marble
SU006 World Labs Framing Worlds: How Marble Helps Creators Bring Consistency to AI Filmmaking "Tim Simmons (Theoretically Media), known online as Theoretically Media, has built an audience of more than 170,000 YouTube subscribers."
SU007 World Labs Bringing Marble to Life
SU008 World Labs From Backyard to Blockbuster: How Lightcraft and Beeble Empowered a New Kind of Indie Filmmaking
SU009 World Labs From Image to Studio: How Magnific Turned 3D Into a Creative Workflow "Millions of designers, marketers, and content teams use it daily."
SU010 World Labs From World Generation to Multiplayer: Rosebud AI x Marble "With Marble, we can create magical, explorable worlds and empower a wider community of game makers."
SU011 World Labs Building Worlds Together: How VIVERSE and Marble Empowered Creators to Build Interactive 3D Worlds
SU012 World Labs Splat World: Exploring New Dimensions of Gaussian Splatting in VR "VR developer Daniel Skaale wanted to test how far Marble's Gaussian splats could go inside a real-time engine."
SU013 OpenArt AI Art Generator: Free AI Image, Video & Audio Generator
SU014 Lightcraft Lightcraft — A Movie Studio in Your Pocket
SU015 Magnific Magnific (formerly Freepik) | The AI Creative Platform
SU016 Rosebud AI Rosebud AI: Make 3D Games & Worlds with Vibe Coding
SU017 HTC VIVERSE VIVERSE | Browser Games, Videos, 3D Art & Virtual Worlds
SU018 Escape Escape.ai
SU019 NVIDIA Isaac Sim "NVIDIA Isaac Sim is an open source reference framework built on NVIDIA Omniverse libraries for robotics simulation, testing, and synthetic data generation."
SU020 robosuite robosuite "robosuite is a simulation framework powered by the MuJoCo physics engine for robot learning. It also offers a suite of benchmark environments for reproducible research."
SU021 MuJoCo MuJoCo — Advanced Physics Simulation "MuJoCo is a free and open source physics engine that aims to facilitate research and development in robotics, biomechanics, graphics and animation."
SU022 Princeton Vision & Learning Lab Home | Infinigen "Infinigen is a procedural generator of 3D scenes, developed by Princeton Vision & Learning Lab. Infinigen is optimized for computer vision research and generates diverse high-quality 3D training data."
SU023 Deloitte AI trends: Adoption barriers and updated predictions "Organizations must overcome technical limitations, manage operational complexities, address evolving regulatory and compliance requirements, and ensure their teams are ready for change."
SU024 Global Market Insights Generative AI Market Size & Share | Forecast Report 2026-2035
SU025 Stanford HAI Economy | The 2026 AI Index Report | Stanford HAI
SU026 World Labs Terms of Service
SU027 World Labs Privacy Policy
SU028 TechCrunch World Labs lands $1B, with $200M from Autodesk, to bring world models into 3D workflows "The partnership between World Labs and Autodesk will see the two companies collaborating to explore how World Labs' models … can work alongside Autodesk's tools, and vice versa, starting with a focus on entertainment use cases."
SU029 World Labs World Labs
SU030 World Labs Marble Labs | World Labs "Marble Labs is where imagination meets experimentation — where artists, engineers, and designers come together to create new possibilities."
SU031 World Labs AI-Native 3D Pipelines with fal
SU032 World Labs Spatial Real Estate
SU033 World Labs Volinga: Splats in Unreal Engine
SU034 World Labs About | World Labs
SU035 World Labs Acceptable Use Policy
SU036 World Labs Gaussian Splats in VR
SU037 World Labs Image Blaster
SU038 World Labs 360° Photos to Walkable Worlds
SU039 World Labs Memory House
SU040 World Labs Custom 3D Avatar
SU041 World Labs Off-Axis Projection
SR001 World Labs Security — Responsible Disclosure We do not currently operate a public bug bounty program and do not guarantee compensation for reports.
SR002 U.S. Copyright Office Copyright and Artificial Intelligence On May 9, 2025, the Office released a pre-publication version of Part 3 in response to congressional inquiries and expressions of interest from stakeholders.
SR003 MDPI (Applied Sciences) Role of Generative AI in AI-Based Digital Twins in Industry 5.0 and Beyond Generative artificial intelligence plays a crucial role in improving AI-based digital twins, enabling more dynamic, adaptive, and accurate industrial simulations.
SR004 Gartner Hype Cycle for Generative AI, 2026 (Gartner)
SR005 MarketsandMarkets Robotics Simulation Software Market — Global Forecast
SR006 World Labs Generating Worlds Sharing our early progress toward persistent, navigable 3D worlds you can explore in your browser.
SR007 World Labs Bringing Marble to Life When the World Labs team set out to create Marble's first marketing video, they made a bold decision: to build it with Marble itself.
SR008 Allen Institute for AI Objaverse — A Universe of 10M+ 3D Objects Objaverse-XL is 12x larger than Objaverse 1.0 and 100x larger than all other 3D datasets combined.
SR039 World Labs Gaussian Splats in VR — Marble Labs Showcase
SR040 World Labs Mint — AI-Native 3D Creation
SR041 Stability AI / HuggingFace Stable Zero123 — 3D Object Generation Model (HuggingFace)
SR042 World Labs Dormant Memories — Marble Labs Showcase
SR043 World Labs Painted Time — Marble Labs Showcase
SR009 World Labs Terms of Service This TOS contains a binding arbitration provision and class action waiver in Section 10.
SR010 World Labs Privacy Policy Our Services may derive spatial measurements from visual content for 3D world generation and spatial intelligence processing.
SR011 World Labs Acceptable Use Policy High-Risk Uses means uses of AI Technology that pose significant risks to health, safety, fundamental rights, or that are subject to heightened regulatory requirements under applicable law, including as regulated under the EU AI Act.
SR012 Bureau of Industry and Security (BIS) Export Administration Regulations (EAR)
SR013 European Parliament Artificial Intelligence Act: MEPs adopt landmark law The regulation establishes obligations for AI based on its potential risks and level of impact.
SR014 NIST Artificial Intelligence — NIST NIST advances a risk-based approach to maximize the benefits of AI while minimizing its potential negative consequences.
SR015 World Labs World Labs Announces New Funding World Labs has raised $1 billion in new funding including AMD, Autodesk, Emerson Collective, Fidelity Management and Research Company, NVIDIA, and Sea.
SR016 Reuters AI pioneer Fei-Fei Li's World Labs raises $1 billion in funding Autodesk invested $200 million in World Labs and will serve as an adviser to the startup.
SR017 Reuters 'AI godmother' Fei-Fei Li raises $230 million to launch AI startup Li will continue some of her work at Stanford while building the startup.
SR018 TechCrunch World Labs lands $1B, with $200M from Autodesk, to bring world models into 3D workflows Data sharing is not part of the agreement.
SR019 SiliconANGLE World Labs closes $1B investment backed by Nvidia, AMD and Autodesk
SR020 StartupHub AI Fei-Fei Li's World Labs: $1.23B Raised, Marble Now Shipping Total capital raised across both rounds stands at $1.23 billion; it arrived before the company disclosed any revenue numbers.
SR021 World Labs Marble: A Multimodal World Model
SR022 World Labs Announcing the World API World generation runs asynchronously and works across a range of inputs, including text, images, and video.
SR023 arXiv (Cornell) 3D Generation for Embodied AI and Robotic Simulation: A Survey Main bottlenecks include limited physical annotations, the gap between geometric quality and physical validity, fragmented evaluation, and the persistent sim-to-real divide.
SR024 arXiv (Cornell) Digital Twin AI: Opportunities and Challenges from Large Language Models to World Models
SR025 Deloitte AI trends: Adoption barriers and updated AI predictions for 2026 Organizations must overcome technical limitations, manage operational complexities, address evolving regulatory and compliance requirements, and ensure their teams are ready.
SR026 Stanford HAI Economy — The 2026 AI Index Report
SR027 Crunchbase News Q1 2026 Shatters Venture Funding Records As AI Boom Continues Four of the five largest venture rounds ever recorded were closed in Q1 2026, with frontier labs collectively raising $188 billion or 65% of global venture investment.
SR028 Crunchbase News Sector Snapshot: Venture Funding to Foundational AI Startups Doubled in Q1 2026 Funding is increasingly concentrated in a handful of foundational giants, including OpenAI, Anthropic and xAI.
SR029 World Labs Scaling Robotic Simulation with Marble High-quality, diverse simulation data is one of the biggest limiting factors in robotics research.
SR030 World Labs NVIDIA Isaac Sim and Marble
SR031 NVIDIA NVIDIA Cosmos — Physical AI Foundation Model Build Policy Models — Accelerate robot policy learning with NVIDIA Cosmos as the backbone for World Action Models.
SR032 Google DeepMind Genie 2: A large-scale foundation world model Genie 2, a foundation world model capable of generating an endless variety of action-controllable, playable 3D environments for training and evaluating embodied agents.
SR033 Wired Fei-Fei Li's Quest to Make Machines Better for Humanity
SR034 Chiang Rai Times World Labs Lands $1 Billion Round Backed By Nvidia and Autodesk
SR035 Stanford HAI 2026 AI Index Report (Overview)
SR036 arXiv (Cornell) NeRF: Representing Scenes as Neural Radiance Fields for View Synthesis
SR037 MDPI (Applied Sciences) Role of Generative AI in AI-Based Digital Twins — MDPI (Applied Sciences)
SV001 TED / Fei-Fei Li With spatial intelligence, AI will understand the real world AI pioneer Fei-Fei Li says a similar moment is about to happen for computers and robots — gaining spatial intelligence, the ability to process visual data, make predictions and act upon those predictions.
SV002 Fei-Fei Li / Substack From Words to Worlds: Spatial Intelligence is AI's Next Frontier Spatial intelligence will transform how we create and interact with real and virtual worlds — revolutionizing storytelling, creativity, robotics, scientific discovery, and beyond. This is AI's next frontier.
SV003 World Labs Bringing Marble to Life Over the course of a few weeks, hundreds of 3D worlds were imagined, refined, and brought into the stage pipeline. Every scene the camera captured began as a text or image prompt inside Marble.
SV004 Allen Institute for AI Objaverse: A Universe of Annotated 3D Objects
SV005 Stanford HAI AI Index Report | Stanford HAI
SV006 U.S. Copyright Office Copyright and Artificial Intelligence Since launching an initiative in early 2023, the Copyright Office has been examining the copyright law and policy issues raised by artificial intelligence, including the scope of copyright in AI-generated works and the use of copyrighted materials in AI training.
SV007 Gartner Hype Cycle for Generative AI 2026
SV008 World Labs Security | World Labs We take the security of our systems seriously and appreciate responsible disclosure of potential security issues.
SV009 YouTube / World Labs "The Future of AI is Here" — Fei-Fei Li Unveils the Next Frontier of AI
SV010 World Labs Concept to Splat | Marble Labs Showcase
SV011 World Labs World Labs Announces New Funding World Labs has raised $1 billion in new funding. Investors include AMD, Autodesk, Emerson Collective, Fidelity Management & Research Company, NVIDIA, and Sea, among others.
SV012 Reuters 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. World Labs did not disclose a valuation.
SV013 TechCrunch World Labs lands $1B, with $200M from Autodesk, to bring world models into 3D workflows Autodesk's investment is a signal that its product has commercial appeal… the two will collaborate at the research and model level.
SV014 SiliconANGLE World Labs closes $1B investment backed by Nvidia, AMD and Autodesk
SV015 StartupHub AI Fei-Fei Li's World Labs: $1.23B Raised, Marble Now Shipping Total capital raised across both rounds stands at $1.23 billion. That cumulative figure places World Labs above the acquisition price Google paid for DeepMind in 2014, and it arrived before the company disclosed any revenue numbers.
SV016 Chiang Rai Times World Labs Lands $1 Billion Round Backed By Nvidia, AMD, And Autodesk World Labs did not share its new valuation. Still, earlier reports said the company discussed a figure near $5 billion.
SV017 Andreessen Horowitz What's In a World? Investing in World Labs Language models changed how we interact with computers because they enabled software to speak and understand natural languages. Spatial intelligence is the next such step.
SV018 Andreessen Horowitz Portfolio | Andreessen Horowitz
SV019 Reuters 'AI godmother' Fei-Fei Li raises $230 million to launch AI startup Initial funding for World Labs was led jointly by Andreessen Horowitz, New Enterprise Associates and Radical Ventures. World Labs declined to share its valuation.
SV020 Crunchbase News Q1 2026 Shatters Venture Funding Records As AI Boom Pushes Startup Investment To $300B Investors poured $300 billion into 6,000 startups globally in the quarter… AI shattered records last quarter, with $242 billion — 80% of total global venture funding in Q1 — going to companies in the sector.
SV021 Crunchbase News Sector Snapshot: Venture Funding To Foundational AI Startups In Q1 Was Double All Of 2025 That funding is increasingly concentrated in a handful of foundational giants… the market saw a shift to a small number of companies capturing a disproportionate share of global capital.
SV022 Global Market Insights Generative AI Market Size & Share | Forecast Report 2026-2035 The global generative AI market was valued at USD 53.7 billion in 2025. The market is expected to grow from USD 83.3 billion in 2026 to USD 988.4 billion in 2035 at a CAGR of 31.6%.
SV023 MarketsandMarkets Digital Twin Market size report 2024-2030 The global digital twin market is expected to grow from USD 21.14 billion in 2025 to USD 149.81 billion in 2030 at a CAGR of 47.9%.
SV024 Deloitte AI trends: Adoption barriers and updated predictions From compliance to workforce readiness, explore common organizational barriers to adopting agentic, physical and sovereign AI.
SV025 Stanford HAI Economy | The 2026 AI Index Report | Stanford HAI
SV026 World Labs About | World Labs World Labs is a leading spatial intelligence company, building frontier world models that can perceive, generate, reason, and interact with the 3D world.
SV027 World Labs Marble: A Multimodal World Model Today we are making Marble, a first-in-class generative multimodal world model, generally available for anyone to use.
SV028 World Labs Announcing the World API Today, we're launching the World API — a public interface for generating explorable 3D worlds using World Labs' multimodal world model, Marble.
SV029 World Labs Dormant Memories | Marble Labs Showcase Dormant Memories explores how real captured spaces can become the foundation for alternate realities using Marble, Gaussian splats, and real-time browser rendering.
SV030 World Labs Mint: AI-Native 3D Creation | Marble Labs Showcase Mint is a chat-first 3D creation platform for generating, editing, and publishing interactive worlds built from Gaussian splats and 3D assets.
SV031 Andreessen Horowitz What’s In a World? Investing in World Labs What may be most remarkable about the language model revolution is how general the solution is.
SV032 Andreessen Horowitz Portfolio | Andreessen Horowitz Portfolio | Andreessen Horowitz
SV033 IEEE Spectrum AI Pioneer Fei-Fei Li Has a Vision for Computer Vision Fei-Fei Li has a vision for computer vision.
SV034 World Labs Dwarven Caves | Marble Labs Showcase Dwarf Caves is a massive interconnected fantasy environment created to push the scale and technical limits of Marble.
SV035 World Labs Musée du Monde | Marble Labs Showcase Musée Du Monde is an interactive 3D museum experience that lets users step inside paintings and explore them from the artist’s perspective.
SV036 TED What if you could talk to your favorite character in a movie? Imagine watching a movie where the main character turns, looks right at you and asks what to do next.
SV037 World Labs How WithLore Built ICARE: A Web-Native Interactive World with Marble and Spark 2.0 Built in roughly two months by a three-person team, the project combines Marble-generated environments, Spark 2.0 streaming, custom Three.js gameplay systems, and AI-assisted production tools.
SV038 World Labs Getting Started with Marble Getting Started with Marble
SV039 World Labs Chisel: 3D Layout Control Chisel: 3D Layout Control