初创公司尽调
尽调报告 AI / application software Series C 2026-06-15

Decart

实时世界模型势头罕见,但公开披露仍跟不上估值

Decart 在实时世界模型和 AI 基础设施上跑出了少见动能;但公开收入记录还撑不起当前约 $4 billion 估值。

封面要素

最新轮次 01
300 USD M [CO029]
最新估值 02
4000 USD M [CO029]
累计融资 03
450 USD M+ [CO032]
成立时间 04
2023 [CO013]
总部 05
Tel Aviv, Israel [CO002]
已披露战略客户 06
Amazon [CU001]
实时 API 价格 07
0.02 USD / sec [CO005]

公司概况

Decart 是一家位于特拉维夫、纵向一体化的 AI 研究实验室,由 Dean Leitersdorf 和 Moshe Shalev 于 2023 年底创立。公司从 2024 年病毒式传播的 Oasis 类 Minecraft 演示,推进到更宽的平台叙事:Lucy 负责实时视频转换,Oasis 负责世界模拟和 Physical AI,DOS 则是支撑两者的优化推理与训练栈。公司现在向商业、流媒体、游戏以及云 / Physical-AI 工作流销售 API 访问和企业集成,背后有异常战略化的投资人组合,并借 AWS 路径进入市场。

官网
decart.ai
成立时间
2023-10-01
创始人
Dean Leitersdorf, Moshe Shalev
创立地点
Tel Aviv, Israel
总部
Tel Aviv, Israel
产品
Decart 销售一个实时 AI 平台,由用于交互式视频生成与转换的 Lucy、用于实时世界模拟的 Oasis,以及跨 GPU、TPU、Trainium 做硬件感知 AI 优化的 DOS 组成。
客户
云服务商、AI 实验室、超大规模云厂商、游戏和互动媒体开发者、商业与广告平台,以及最终的 Physical-AI / AV 模拟买家。
商业模式
按量计费的 API 定价,叠加企业授权、基础设施优化交易和战略市场进入合作。
阶段
Series C
融资情况
2026 年 5 月以约 $4 billion 估值完成 $300 million Series C,使累计融资超过 $450 million。
[CO001, CO002, CO003, CO004, CO013, CO014, CO015, CO029]

执行摘要

主要优势

  • 实时世界模型栈同时覆盖产品和基础设施,DOS 让 Decart 在成本和延迟上有了可信的差异化叙事。
  • Radical、Nvidia、Sequoia、Benchmark 以及客户关联投资人的战略背书,提高了 Decart 超越普通早期 AI 初创公司的分发概率。
  • AWS/Trainium 与 Amazon 战略客户信号,为一家这么年轻的公司提供了异常强的企业平台验证。
  • 产品表面已经伸向多条变现路径:互动视频、游戏 / 媒体工具,以及 physical-AI 仿真。

主要风险

  • 公司没有公开收入、ARR、毛利率或客户数披露,$4 billion 估值缺少基本面锚点。
  • Oasis 3 仍暴露物理和连贯性限制,可能卡住价值最高的 AV 与机器人用例。
  • 具名客户证据偏薄,Amazon 是唯一明确披露的战略客户,多数企业需求仍只按群组描述。
  • 在 Decart 披露足够规模、能够从容吸收成本前,版权以及欧盟 AI / 版权监管暴露可能变得昂贵。
  • Runway、World Labs、Google DeepMind 等竞争者资源相当或更强,也在切入相近的世界模型地带。

未决问题

  • 经审计的收入、ARR、消耗和毛利率披露。
  • 客户数、留存、集中度和合同期限细节。
  • 自动驾驶或更广泛 physical-AI 仿真中的付费生产采用证据。
  • 完整董事会构成、股权表所有权、优先权和投资人治理权。
  • 2026 年当前员工数,以及创始人与旧金山研发负责人之外的组织扩张情况。

目录

Chapter 01

01公司概览

1.1 身份、运营模式与产品表面

Decart 将自己描述为一家创立于 2023 年、完全纵向一体化的前沿 AI 研究实验室,既构建最先进的实时世界模型,也构建运行这些模型所需的超优化基础设施。公司 2026 年 5 月的融资新闻稿把业务框定为共用同一栈的三条产品线:Lucy 是面向沉浸式视频和交互体验的实时世界模型;Oasis 是面向 Physical AI 和机器人模拟的实时世界模型;DOS 是 Decart Optimization Stack,可将模型权重编译到 NVIDIA GPU、Google TPU 和 Amazon Trainium 芯片上。Decart 通过 platform.decart.ai 按用量销售 API 访问;Lucy 2.1 和 Oasis 3 Preview 等实时模型的定价约为每秒主动生成 $0.02,图像编辑则是单美分级单价,并通过销售联系人提供企业定价。公司运营锚点在特拉维夫,已披露办公室包括以色列北部、旧金山和纽约,并于 2025 年在旧金山开设由 Dr. Kfir Aberman 领导的 R&D 中心。2026 年前的产品表面还包括 Mirage,这是 Decart 的直播扩散模型,公司将其定位为 Lucy 系列实时视频模型的前身。[CO001, CO002, CO003, CO004, CO005, CO006]

快照 KPI 表
指标数值 / 状态日期置信度缺口
成立时间2023 年末2023-12-31
总部 / 主要办公室以色列 Tel Aviv2026-06-15
已披露其他办公室San Francisco、New York、以色列北部2025-08-07来自 Ynet 访谈自述;本次 reviewed 的 decart.ai 页面中,Decart 未发布当前办公室登记表。
公司法律实体Decart.AI, Inc.2025-08-07投资者新闻页列出该法律实体,但公司未在自有网站披露司法辖区或注册日期。
最新披露轮次(USD M)3002026-05-18
最新公开估值(USD B)42026-05-18公司发布称这是 $4B 轮;媒体合作方把该数字描述为「roughly」$4B 或「estimated」$4B 估值。
累计融资总额(USD M)450+2026-05-18Decart 自述累计「over $450 million」;已披露单轮金额合计为 $21M(2024 年 10 月)+ 约 $32M(Ynet 所称 2024-2025 隐身期达到 $53M 的差额)+ $100M(2025 年 8 月)+ $300M(2026 年 5 月),与 $450M+ 说法大体一致。
已报告收入未披露(公司称「significant」)2026-05-18Decart 称其收入来自向云服务商授权 DOS,以及 Lucy/Oasis 客户,但没有发布经审计收入或 ARR。
已报告 lifetime burn低于 $100M2026-06-10Leitersdorf 2026 年 6 月告诉 TechCrunch,Decart lifetime burn「drastically less」than $100M;未经独立审计。
员工数(最近公开口径)~602025-08-07Ynet 2025 年 8 月文章称员工 60 人,高于约 15 人;reviewed 来源未发布 2026 年当前员工数。
开发者社区100,000+2026-06-10Leitersdorf 向 TechCrunch 自述;没有独立验证。
实时 API 定价(Lucy 2.1 / Oasis 3 Preview)每秒有效生成 $0.022026-06-15

估值、轮次规模和定价行可作为外部参考点;收入、员工数和 lifetime burn 都是公司报告口径,不能替代尽调室证据。

[CO005, CO006, CO009, CO012, CO023, CO025]
FO003: KPI 快照

公开可支撑的指标指向一家快速扩张、资金充足的世界模型实验室:开发者拉力强,但员工数、收入和股权透明度仍有明确缺口。

[CO005, CO023, CO031, CO033, CO034, CO041]

1.2 创始人、领导层、治理与关键人依赖

Decart 由 Dr. Dean Leitersdorf(CEO)和 Moshe Shalev(CPO)于 2023 年底创立,两人是在以色列军事情报部队 Unit 8200 服预备役时相识。Leitersdorf 在 24 岁前就在 Technion 取得三个计算机科学学位和博士学位;Shalev 成长于以色列 Bnei Brak 的极端正统派社区,后来服役于 Unit 8200,并加入 Decart 负责运营与产品。公开材料还显示,DreamBooth 扩散调优技术的共同创造者、Snap 和 Google 前高级研究员 Dr. Kfir Aberman,自 2025 年 8 月 Series B 起领导公司旧金山 R&D 中心。Decart 自己 2026 年 5 月的公告仍把 Leitersdorf 作为面向公众的 CEO 和主要发言人;所有已审阅来源中,公司都没有公布独立董事会构成或治理披露。这个模式叠加报道中点名的高管面较窄,说明即便在 Radical 领投 $300 million 之后,公司仍高度依赖 Leitersdorf,治理图景也仍偏薄。Decart 是一家 Delaware C-corp,通过投资人和行业新闻稿以 Decart.AI, Inc. 名义对外营销;公开记录尚未说明外部董事、审计委员会或独立董事长。[CO013, CO014, CO015, CO016, CO017, CO018]

领导层与创始人表
人员职位背景founder-market fit 或职能覆盖key-person 依赖
Dr. Dean LeitersdorfCEO、联合创始人23 岁前取得三个 Technion 计算机科学学位和 Technion PhD;Singapore postdoc;Unit 8200 预备役;来自 Leitersdorf 高科技家族(YL Ventures 创始人 Yoav Leitersdorf 和 Indeni 创始人 Yoni Leitersdorf 是其兄弟)。兼具底层系统研究、GPU / 推理优化可信度、Lucy/Oasis/DOS 公众发言人角色,以及公司主要融资声音。
Moshe ShalevCPO、联合创始人在 Bnei Brak 的 ultra-Orthodox 家庭长大,较晚入伍,在 Unit 8200 与 Leitersdorf 共事,并在 2024 年 10 月 Oasis 发布流量激增时负责 Decart 运营。负责运营、产品和业务的人力系统扩张侧;平衡 Leitersdorf 的研究 / CEO 集中度。
Dr. Kfir AbermanSan Francisco R&D center 负责人(2025 年起)前 Snap 和 Google 研究员;广泛引用的 DreamBooth diffusion-tuning 技术共同创造者;在 2025 年 8 月 Series B 时被招募,负责美国模型 R&D 和招聘。增加 diffusion-model 研究深度,并在 San Francisco 建立美国招聘滩头堡。

本表聚焦创始人和一位公开点名的美国 R&D 负责人;本章 reviewed 的任何来源中,Decart 都未发布完整高管名单或董事会名册。

[CO013, CO014, CO015, CO016, CO017, CO018]

1.3 资本基础、估值路径与利益相关方图谱

Decart 的融资历史压缩得异常快。公司于 2024 年 10 月 31 日走出隐身状态,宣布由 Sequoia 领投、Oren Zeev 的 Zeev Ventures 参投的 $21 million 轮融资,并同时推出病毒式传播的 Oasis 类 Minecraft 演示。Ynet 2025 年 8 月访谈称,公司在跨过独角兽门槛前,两个早期连续轮次合计融资 $53 million,估值约 $500 million。2025 年 8 月 7 日,Decart 宣布以 $3.1 billion 投后估值完成 $100 million Series B,使累计融资达到 $153 million;Sequoia、Benchmark 和 Zeev Ventures 继续跟投,Aleph VC 作为新投资人进入。2026 年 5 月 18 日,Decart 披露由 Radical Ventures 领投、估值约 $4 billion 的 $300 million 轮融资,NVIDIA、eBay Ventures、Adobe Ventures、Toyota Ventures、Atreides Management 和 Valor Equity Partners 与回归的 Sequoia、Benchmark、Zeev Ventures 一同参投;该轮私人支持者包括 OpenAI 联合创始人 Andrej Karpathy、前 Disney CEO Michael Eisner、Nintendo 创始家族成员和游戏投资人 Moritz Baier-Lentz。Decart 自称迄今已融资超过 $450 million;Amazon 公开身份是战略客户,而不是已披露股权投资人。Leitersdorf 于 2025 年 8 月告诉 Ynet,Decart 使用的 $153 million 资本不到 $10 million;这是公司自报消耗,未经过独立审计。[CO023, CO024, CO025, CO026, CO027, CO028]

利益相关方或投资者地图
利益相关方角色控制权或经济重要性尽调问题
Radical Ventures2026 年 5 月 $300M 轮领投方领投该轮,把 Decart 估值从 $3.1B 推到约 $4B;合伙人 Jordan Jacobs 是 Decart 融资发布中的具名投资人声音。要求提供投后 cap table、清算优先权结构、董事会席位分配,以及授予 Radical 的任何 pro-rata 或信息权。
NVIDIA战略投资者和硬件合作伙伴加入 2026 年融资,Decart 公开称其既是投资者,也是 DOS 优化推理中的 accelerated compute 合作伙伴。澄清 NVIDIA 投资是否带有任何路线图或 GPU 分配优先权,从而影响竞争硬件(Trainium、TPU)的商业安排。
Sequoia Capital最早机构投资者(2024 年 10 月);持续跟投领投 2024 年 10 月 $21M stealth-exit 轮,并在 2025 年 Series B 和 2026 年融资中都继续跟投,使 Sequoia 成为多阶段所有权和信息权锚点。确认 Sequoia 持股比例、董事会观察员 / 席位状态,以及 2024 年 seed 条款(anti-dilution、ratchets)是否仍适用。
Zeev Ventures(Oren Zeev)2024 年以来共同投资者;2025 和 2026 年持续跟投Oren Zeev 被公开认定为 2024 年两位原始支持者之一,并参与了截至 2026 年的每一轮后续融资。量化 Zeev 当前持股,以及 2024 年隐身轮授予的任何 side-letter 权利。
Benchmark2025 和 2026 年持续跟投参与 2025 年 8 月 $100M Series B 和 2026 年 5 月 $300M 轮,与 Sequoia、Zeev 并列为三家重复机构投资者之一。确认 preference stack 顺序、任何董事会代表,以及 pro-rata 行权历史。
Aleph VC2025 年 8 月 Series B 新投资者按 The SaaS News 的 Series B 摘要,Aleph 作为唯一新披露的以色列 VC 合作方,加入 $3.1B 估值的 $100M Series B。核实投资规模,以及 Aleph 是否承担任何本地董事会或以色列政府合规角色。
战略投资方:Adobe Ventures / Toyota Ventures / eBay Ventures战略公司投资者(2026 年 5 月)作为公司风投部门加入 2026 年融资,其母公司(Adobe、Toyota、eBay)被 Decart 点名为潜在客户行业(创意工具、汽车、商业)。测试三家企业中是否有任何商业 pilot 或 LOI 与战略投资同步发生。
Atreides Management 与 Valor Equity Partnerscrossover / 成长投资者(2026 年 5 月)作为主要成长 / crossover 名称,与 Radical 和公司 VC 一起加入 2026 年融资。澄清任一方是否在为未来 secondary、pre-IPO 或 late-stage block trade 布局。
个人支持者:Andrej Karpathy、Michael Eisner、Nintendo 家族、Moritz Baier-Lentz战略天使投资者(2026 年 5 月)2026 年融资中具名的个人支持者;即使个人持股很小,也释放了来自 AI 研究(Karpathy)、媒体(Eisner)、游戏(Nintendo family)和 gaming-VC(Baier-Lentz)社区的背书信号。量化个人支票规模(可能较小),以及是否附带任何 carry advisory 或 board-observer 角色。
Amazon / AWS战略客户和 Trainium3 硬件伙伴Decart 公开材料把 Amazon 描述为战略客户,把 AWS 描述为让 Decart 获得 Trainium3 early access 的合作伙伴;Lucy2 部署在 Trainium3 上,DOS 运行在 Trainium silicon 上。判断 Amazon 是否也持有股权(未披露),并量化 Trainium3 商业承诺。

已披露 cap table 不完整:Decart 点名了领投 / 跟投机构投资者,但未发布持股比例、preference stack 或董事会构成。下列行强调已披露商业或融资杠杆,而非确定的股东名册。

[CO024, CO025, CO026, CO027, CO028, CO029]
FO002: Decart 资本路径和利益相关方逻辑

Sequoia / Zeev 种子资金、2025 年 8 月 Series B,以及 2026 年 5 月 Radical 领投轮,供给了一个垂直整合栈(DOS + Lucy + Oasis),再靠 AWS Trainium 和 Bedrock 交付给企业客户。

[CO023, CO028, CO031, CO034, CO037, CO043]

1.4 里程碑时间线、规模信号与证据缺口

公开时间线从 2023 年底创立开始,经 2024 年 10 月 31 日退出隐身和 Oasis 病毒式演示,进入 2025 年扩张期(Mirage 发布、$100 million Series B、旧金山 R&D 中心、员工数从约 15 人增至约 60 人),再到 2026 年战略基础设施推进(DOS 2.0 发布、围绕 Lucy2 的 AWS Trainium3 合作、Amazon 战略客户认定、Radical 领投的 $300 million 轮融资和约 $4 billion 估值,以及 2026 年 6 月 10 日 Oasis 3 Preview 以每秒 $0.02 登陆公共 API)。Decart 公开称 Lucy 响应时间低于 30 毫秒,DOS 栈可交付超过 100-fps 的全高清推理和超过 1,600 tokens-per-second 的智能体吞吐;公司还称服务的开发者社区超过 100,000 名。负面和不确定信号与增长故事并存:TechCrunch 2024 年 10 月对 Oasis 的评测指出未解决的版权问题,因为模型训练使用了 Minecraft 画面,却未披露 Microsoft 授权;TechCrunch 2026 年 6 月对 Oasis 3 的评测称长生成会退化、汽车会穿过其他汽车,物理一致性仍是开放研究问题;经审计收入、确切当前员工数、逐项股权结构和独立董事会构成仍未进入公开记录。因此本章保留的判断是:Decart 是一家扩张快、资本充足、偏基础设施的世界模型实验室;关键尽调变量仍是创始人集中度、未披露财务细节,以及超真实演示与生产级物理之间的差距。[CO038, CO039, CO040, CO041, CO042, CO043]

里程碑表
日期事件类型金额 / 估值 / 状态参与方 / 来源含义
2023-12-31Decart 由 Leitersdorf 和 Shalev 在以色列创立创立2023 年末创立Decart 融资发布;Ctech 和 JNS 报道将公司锚定为 2023 年一代的以色列 AI 初创公司,约在创立后 12 个月内做出首个产品。
2024-10-31隐身期结束并推出 Oasis 病毒式 demo产品从 Sequoia 和 Zeev Ventures 融资 $21MTechCrunch(2024 年 10 月);Decart blog 和 Oasis GithubDecart 成为首家交付实时、可玩的开放世界 AI 模型的公司;病毒式传播(Elon Musk 发推、数日内用户超过 100 万)打响了品牌。
2025-05-01Mirage 实时视频转换模型发布产品实时直播扩散模型Ynet(2025 年 8 月)和 Ctech「From stealth to $3.1B」特写Decart 从游戏演示扩展到通用实时视频转换栈,后来以 Lucy 系列命名。
2025-08-07Series B 公布融资$100M Series B,估值 $3.1BDecart 公告;Ctech;Ynet;SiliconAngle;SaaSNews;Yahoo确认独角兽地位,累计融资增至 $153M,并让 Aleph VC 与既有投资方 Sequoia、Benchmark、Zeev Ventures 一同进入股东名单。
2025-08-07旧金山研发中心开设扩张中心与 Series B 同步启动Ctech 2025 年 8 月 Decart 特写建立由 Dr. Kfir Aberman 领衔的美国工程前哨,支撑以色列以外的招聘。
2025-12-01AWS Trainium 与 Amazon Bedrock 合作披露合作Lucy 针对 Trainium2 / Trainium3 优化;Bedrock 分发AINews 对 AWS re:Invent Trainium-Decart 合作的报道将推理硬件从 NVIDIA 分散出去,并让 Decart 通过 Bedrock 获得企业分发渠道。
2026-05-18$300M 融资完成,估值约 $4B融资$300M,由 Radical Ventures 领投;NVIDIA、Adobe、Toyota、eBay,Decart 公告;JNS;Ynetnews;Ctech;Electronics Weekly总融资超过 $450M,引入 NVIDIA 和战略企业 VC,并把 Decart 重新锚定为实时 AI 基础设施,而不是只讲游戏消费故事。
2026-05-18DOS 2.0 随融资轮同步发布产品DOS 2.0 推理 / 训练栈,覆盖 NVIDIA / TPU / TrainiumDecart 融资公告DOS 从内部优化层改定位为可商业授权的产品,横跨多家硬件厂商。
2026-06-10Oasis 3 Preview 在公开 API 上线产品$0.02/sec API 访问;面向自动驾驶和物理 AITechCrunch(2026 年 6 月 10 日);Decart 文档;Startup FortuneOasis 以付费 API 产品形式向开发者社区开放,正式确认转向物理 AI 和机器人仿真。
2026-06-10TechCrunch 发表带有反向警示的评测反向评测者报告了退化、物理缺口和控制问题TechCrunch(2026 年 6 月 10 日)保留一个反向信号:即便是 Decart 的旗舰 Oasis 3 版本,也带着已有记录的一致性、物理和转向限制发布。
2024-10-31Oasis 训练数据引发版权问题反向TechCrunch 提到未披露 Microsoft 对 Minecraft 素材的许可TechCrunch(2024 年 10 月 31 日)在发布叙事之外,保留一个关于训练数据授权的反向且未解决的尽调信号。

这是第 1 章唯一的里程碑时间线记录,除非后续出现更新证据,后续章节应复用;反向行保留在表内,避免本章显得全是正面叙事。

[CO013, CO023, CO025, CO028, CO031, CO033]
FO001: 公司里程碑时间线

Decart 从 2023 年末创立,走到 2024 年爆红的世界模型演示、2025 年 Series B / 云合作伙伴建设,再到 2026 年在 Radical 领投 $300M 轮融资下转向基础设施和物理 AI。

[CO013, CO023, CO028, CO031, CO034, CO042]

1.5 图表

Chapter 02

02市场分析

2.1 市场边界、纳入支出与现状替代方案

Decart 的可服务市场位于三个松散相连的支出池交汇处,而不是一个成熟品类。第一是游戏中的生成式 AI,独立分析机构把它定义为游戏工作室和创作者经济平台内部,来自程序化内容生成、关卡 / 世界搭建、NPC 行为、叙事生成、自适应个性化和 AI 驱动 QA 的收入。第二是实时生成视频和互动视频,Decart 的 Lucy 2.1 可转换或生成连续视频流,竞争对象包括编辑流程视频模型(Runway、Luma、Adobe Firefly Video),以及 UGC 平台(Fortnite、Roblox、Minecraft)的游戏内视频管线。第三是 Physical-AI / 世界模型模拟,Oasis 3 Preview 面向自动驾驶数据生成,竞争对象包括 Waymo 内部模拟器、NVIDIA Cosmos/DRIVE Sim 栈、World Labs Marble、Google DeepMind Genie 3,以及 Tesla 和 Wayve 的专有 AV 模拟器。买方层面的现状替代方案很关键:手工游戏内容(美术、设计师、过场动画团队)、传统驾驶模拟引擎(Unreal/CARLA),以及传统非实时视频模型(Sora、Veo、Runway Gen),这些模型输出批处理结果,而不是低于 30 毫秒的交互帧。排除的支出包括受监管商业博彩(AGA 的 $51B/month 序列属于另一个池)、GPU 硬件资本开支(输入成本而非软件收入),以及非实时的广播 / 电影 VFX 支出。Decart 每秒 $0.02 的标价把 API 定在开发者工具支出与推理算力之间的边界上,因此公司市场边界最好理解为横向实时生成式 AI 基础设施品类,而不是只属于游戏的一个科目。[CM001, CM002, CM003, CM004, CM005, CM006]

市场定义表
细分 / 类别纳入支出排除支出买方 / 付款方与 Decart 的相关性
游戏生成式 AI(核心 SAM 口径)工作室和创作者在程序化内容、关卡 / 世界生成、NPC 行为、叙事、自适应个性化、AI QA 工具上的支出(TBRC / Research and Markets 定义)GPU 硬件资本开支;博彩收入;非实时 VFX游戏工作室、引擎团队、UGC 平台(CTO / 内容负责人掌握预算)Lucy 和 Oasis 直接瞄准游戏与互动娱乐的实时生成;这是公开规模数据最接近的资金池。
云游戏和实时互动视频纯流媒体玩家,以及在远端渲染并分发画面的混合 / 捆绑式云游戏服务(BCG 定义)设备端主机 / PC 计算;广播级流媒体基础设施;静态 OTT 视频云游戏平台和玩家终端客户(消费者)Decart 低于 30 ms 的响应时间和 DOS 基础设施,切中同一类对延迟敏感的服务瓶颈。
世界模型和物理 AI / AV 仿真面向自动驾驶、机器人和具身 AI 的软件、仿真与合成场景生成(无公开市场规模)车辆硬件;LIDAR 传感器;车载计算;整车层面工程AV / 自动驾驶项目(自动驾驶 / AI 安全负责人)和机器人实验室Oasis 3 Preview 以该细分为定位,并明确获得 Toyota Ventures 和 NVIDIA 背书。
UGC 创作者经济(下游需求池)Fortnite、Roblox、Minecraft 及其他平台上的创作者分成(BCG 称 2025 年超过 $1.5B)托管 / CDN 基础设施;应用商店保留的平台抽成UGC 平台(付款方);个人创作者(用户)Decart 工具作为这些平台创作者的「镐和铲」,沉淀价值。
生成式 AI 推理 / 优化基础设施面向低延迟模型的云转售推理;授权给云厂商的优化栈(Decart DOS on AWS Trainium 是具名例子)裸金属 GPU 资本开支;仅训练支出;非推理服务云平台(付款方);这些云上的企业租户(用户)DOS 2.0 明确在这里竞争;AWS 战略客户关系是已披露的证明点。
相邻批处理 / 编辑式生成视频Sora、Veo、Runway Gen、Luma、Adobe Firefly Video(非实时生成和编辑)静态图像生成;非视频创意套件创意专业人士和后期制作团队替代性支出池:非互动交付物上,客户可能选择批处理输出,而不是实时 API。

“与 Decart 的相关性”基于 Decart 自身产品定位和各细分的买方匹配度作定性判断;纳入 / 排除支出沿用 TBRC、BCG 和 Research and Markets 发布的类别定义。

[CM001, CM002, CM003, CM004, CM005, CM006]
FM003: 买方 / 细分市场地图

矩阵把每类买方映射到 Decart 产品线、付费模型,以及当下最约束采用的监管 / 合同条件。

“最关键监管约束”按当前公开信息,为每类买方选出单一最卡采用的规则;每类买方还面临未列出的叠加义务。

[CM021, CM022, CM024, CM027, CM028, CM030]

2.2 规模测算视角、区间与方法限制

没有单一公开规模测算能覆盖 Decart 的完整机会,因此本章用三个视角锚定。视角 1(游戏生成式 AI,最接近的纯品类):The Business Research Company 估算该市场 2025 年为 $1.79 billion,2026 年增至 $2.21 billion,2030 年达到 $5.09 billion,CAGR 为 23.2%,亚太是 2025 年最大区域;Research and Markets 转载了几乎相同的 2025–2030 年预测,并把序列延伸到 2035 年,包含确定性和非确定性细分。视角 2(云游戏,作为低延迟、服务器渲染互动娱乐的代理):BCG 2026 Video Gaming Report 估算云游戏 2025 年约 $1.4 billion,2030 年扩至约 $18.3 billion,复合年增长率超过 50%;报告还指出,60% 受访玩家尝试过云游戏,其中 80% 反馈体验正面。视角 3(UGC 创作者经济,即 Decart 工具所赋能的平台层):BCG 报告 Fortnite 和 Roblox UGC 创作者分成在 2025 年超过 $1.5 billion,且 40% 玩家消费的 UGC 多于一年前。这些视角都没有测算 Oasis 3 明确瞄准的 Physical-AI / 世界模型模拟细分;McKinsey State of AI report 这个最常被引用的跨行业基准,在本次运行中访问被拒。因此本章并列保留三个已引用区间,把游戏生成式 AI 序列视为最直接相关的 SAM 视角,并标注 Physical-AI 模拟是一个证据受限、尚未发布独立测算的 TAM。[CM010, CM011, CM012, CM013, CM014, CM015]

TAM / SAM / SOM 或规模测算口径表
发布方年份地区价值(USD)CAGR方法置信度局限
The Business Research Company(TBRC,市场研究机构)2026全球$1.79B (2025) → $5.09B (2030)23.2% (2025-2030)按技术和终端用户,自下而上测算生成式 AI 游戏服务市场规模。营销页面只披露方法摘要;完整分项在付费墙后。
Research and Markets(市场研究机构)2026全球$1.79B (2025) → $5.09B (2030);延伸至 203523.2% (2025-2030)转发 TBRC 系列,并拆出确定性 / 非确定性细分。与 TBRC 数字高度相关;不是独立估算。
BCG(Video Gaming Report 2026)2026全球云游戏 $1.4B (2025) → $18.3B (2030)>50%BCG 分析师预测,基于玩家调研、纯流媒体收入,以及一部分混合和捆绑服务收入。云游戏的计量单位是收入,不同于工作室软件支出;不能与 TBRC 直接相加。
BCG(Video Gaming Report 2026)2026全球仅 Fortnite 和 Roblox 的 UGC 创作者分成在 2025 年就超过 $1.5B汇总两大 UGC 平台披露的平台分成。分成不等于平台收入;该数字低估了 UGC 生态的整体经济规模。
Newzoo(2026 年 PC 与主机游戏报告)2026全球PC + 主机游戏市场规模(完整报告在付费墙后)按季度跟踪各平台,是开发者参考装机量和收入的标准资料。本轮中头部数字在付费墙后;这里只验证了报告标题 / 范围。
American Gaming Association(经 TBRC 引用)2025美国商业博彩收入 $51.14B(2025 年 8 月,同比 +8.9%)行业协会按月跟踪美国商业博彩收入。包含受监管的商业博彩——该项被明确排除在 Decart 市场边界之外,但可作为更广义游戏经济的有用分母。
物理 AI / AV 世界模型仿真2026全球审阅来源中没有独立规模测算本轮没有针对实时 AV 仿真 / 世界模型 TAM 的公开分析师预测。存在重大规模测算缺口;Oasis 3 Preview 的收入机会只能作定性定位。

TBRC 与 Research and Markets 系列并不独立,应视作同一套规模测算口径。云游戏(BCG)是玩家时长池的补充口径;UGC 分成和 AGA 商业博彩数据是背景,不应加总进 TAM。物理 AI / 世界模型仿真仍受证据约束。

[CM010, CM011, CM012, CM013, CM014, CM015]
FM001: 市场规模视角

金字塔按覆盖面排序:最宽是互动娱乐背景(云游戏 + 游戏经济),向下到 Decart 可服务切口(游戏中的生成式 AI),再到最早已实现预订(Decart 2026 收入,未披露)。

金字塔按覆盖范围排序,不做加总。云游戏和游戏生成式 AI 口径不同,不应相加。

[CM001, CM003, CM007, CM015, CM020]
FM002: 市场估算区间

三个已量化视角采用分析师报价区间;低 / 基准 / 高值锚定已报告点预测,若只有单一点估计,则用保守 ±10% 区间。

区间由已报告点估计构建;非零区间是保守的分析误差带,不是报告的置信区间。

[CM010, CM011, CM015, CM016, CM017]

2.3 买方、用户、付款方与采用路径

今天为 Decart 技术栈付费或使用的有四类买方。(a)游戏工作室和引擎团队,BCG 估算约 50% 的工作室已经使用某种 AI,并指出到 2025 年中,约 20% 新 Steam 游戏披露使用 AI;预算负责人是 CTO 和内容负责人,采用触发点是生成工具能明确缩短 AAA 游戏内容周期,而这类游戏单款预算可达 $300 million。(b)UGC 平台和创作者,$1.5 billion 的创作者分成池吸引扩散模型和世界模型供应商充当「卖铲人」;付款方是平台(Roblox、Fortnite、Minecraft),用户是创作者。(c)生成式 AI 基础设施客户,云服务商(Amazon/AWS 是 Decart 已点名的战略客户)授权 DOS,以优化自己横跨 NVIDIA、TPU 和 Trainium 芯片的模型服务;付款方是云平台,用户是其企业租户,采用触发点是 GPU/Trainium 效率。(d)Physical-AI 和 AV 模拟买方(Toyota Ventures 和 Nintendo 家族出现在 Decart 股权结构中,构成相邻信号),Oasis 3 Preview 被放在 Waymo、Tesla、Wayve 和 NVIDIA DRIVE Sim 的内部模拟器对面;预算负责人是自动驾驶项目负责人,采用取决于物理保真度和安全论证可接受度。四类客群共同的公开自助采用路径,是 Decart 在 platform.decart.ai 上的 API——每秒 $0.02 的实时价格层,加上美分级图像编辑调用——再叠加企业合同。Decart 2026 年 6 月向 TechCrunch 报告的 100,000 名开发者社区,是付费使用的早期漏斗。[CM021, CM022, CM023, CM024, CM025, CM026]

细分 / 买方地图
细分买方用户付款方工作流预算负责人采用触发因素
游戏工作室(AAA + 独立)引擎团队 / 工作室管理层游戏设计师、关卡美术、NPC 编剧工作室 P&L(通常经中央 R&D 预算)程序化内容生成、NPC 行为、叙事;嵌入 Unity / Unreal 工作流的 AICTO 或内容技术负责人可证明能降低 $300M 级 AAA 开发成本或缩短周期
UGC 平台和创作者平台工程组织(Roblox、Fortnite Creative、Minecraft)在平台上制作体验的个人创作者平台从分成池向创作者付款(BCG 称 2025 年超过 $1.5B)基于 API 的内容生成,嵌入平台创作者工具创作者平台副总裁 / 产品副总裁创作者入驻更快、更便宜,单次会话互动更高
云 / 推理平台公有云提供商(Amazon / AWS 被具名)云平台上的企业租户云提供商(再向企业租户计费)为低延迟视频和世界模型优化推理;DOS 编译到 NVIDIA / TPU / Trainium云产品 / 合作负责人Trainium 或 TPU 利用率提升,客户入驻更快
物理 AI / AV 项目AV 公司(Waymo、Tesla、Wayve)、机器人实验室自动驾驶工程师、安全论证团队、ML 数据运营AV 项目 P&L用于训练和评估的合成场景生成;Oasis 3 Preview 定价 $0.02/sec自动驾驶 / AI 安全负责人物理保真度足以在部分训练数据工作流中替代 Unreal / CARLA
创意工具和开发者(自助 API)独立开发者和创意工作室设计师、视频剪辑师、独立游戏开发者按秒计费的 API 支出platform.decart.ai 自助开发者访问;截至 2026 年 6 月开发者社区超过 100,000 人个人开发者或小型工作室负责人便宜的实时 API、易上手的文档;靠病毒式演示被发现
战略企业 VC(下游需求信号)Adobe、Toyota、eBay、Nintendo 家族、Amazon各战略支持方的内部产品团队各企业母公司的 P&L与 Decart 共同开发试点和基础设施消耗企业风投 / 战略合作负责人进股东名单 + 试点工作被纳入运营预算

买方 / 用户 / 付款方行复刻 Decart 及其分析师报道公开披露的模式;部分采用触发语言,是基于公司定位作出的定性推断,而非买方侧直接披露。

[CM021, CM022, CM023, CM024, CM025, CM026]
FM004: 采用漏斗或价值链地图

采用漏斗从广泛的实时生成式 AI 认知,到 Decart 开发者注册、付费 API 使用,再到企业 / 具名战略客户状态。2026 年公开资料大多不披露付费阶段,因此漏斗从公开的 100,000 名开发者收窄到单一具名企业客户。

Decart 不发布统一买方漏斗,因此漏斗混合了开发者规模、客户数量和资本指标;每一层都是可得的最佳公开代理。

[CM021, CM027, CM029, CM030]

2.4 增长驱动、采用约束与需要保留的矛盾

顺风很具体。延迟敏感工作负载(实时玩法、AV 场景生成、直播视频转换)是批处理导向的视频和图像生成器留下的供给缺口;云游戏 >50% CAGR,以及 2024 到 2025 年披露使用 AI 的 Steam 游戏数量翻倍,都指向实时生成式 AI 基础设施楔子。逆风同样具体。欧洲议会 2026 年 3 月主动报告(报告员 Axel Voss,17-3 投票)要求逐项披露所有受版权保护训练数据,提出在缺乏透明度时推定侵权但允许反驳,并提出对生成式 AI 提供商征收 5–7% 全球营业额的固定版权费;Osborne Clarke 将其称为 EU 监管制度的「分水岭时刻」。EU AI Act 的 General Purpose AI 义务已自 2025 年 8 月 2 日起生效,Digital Omnibus 则推迟了原定 2026 年 8 月生效的高风险义务。在美国,California AI Transparency Act 和 GenAI Training Data Transparency Act 于 2026 年 1 月 1 日生效,Texas Responsible AI Governance Act 同日生效,Trump 政府 2025 年 12 月 11 日行政令指示 DOJ 以联邦优先权为由挑战州级 AI 法律。US Copyright Office 正在发布的报告系列仍是 AI 与版权的主要联邦框架。需要保留的矛盾:(i)云游戏预测和游戏生成式 AI 预测在规模和单位上不一致(玩家小时数 vs. 工作室软件支出);(ii)Physical-AI 模拟尚无已发布 TAM,但 Decart 投资人(Toyota Ventures、NVIDIA、Radical)押注它会成为 Decart 最大细分;(iii)工作室采用 AI 正在上升,但 BCG 自己的调查显示玩家抵触 AI 用于美术 / 动画,提示消费端采用风险;(iv)GDC 2026 State of the Game Industry Report 这个权威开发者调查在本次运行中有付费墙,本文图表未直接引用。[CM031, CM032, CM033, CM034, CM035, CM036]

增长驱动因素和约束表
驱动因素或约束方向时间含义尽调问题
对延迟敏感的工作负载(实时游戏、AV 场景、直播)驱动现在至 2030 年切出一个可防守的楔子:批处理生成器(Sora、Veo、Runway)无法服务低于 30 ms 的闭环。量化 Decart 收入中真正受延迟约束的部分,与可被批处理替代的部分。
云游戏 CAGR >50%(BCG)驱动2025-2030扩大 Decart 栈瞄准的实时服务器渲染互动媒体可寻址池。检验 DOS / Lucy 的商业订单是否跟随云游戏采用曲线。
游戏工作室采用 AI 翻倍(BCG:2025 年新 Steam 游戏中 20% 披露使用 AI)驱动2025-2030工作室侧在生产流水线中使用 AI 逐步常态化。确认 Decart 是否能在 Unity / Unreal 流水线内拿下工作室业务。
EU AI Act GPAI 义务 + 2026 年 3 月版权报告约束2025 年 8 月 2 日起生效;版权解决方案拟于 2026 年 3 月提出逐项训练数据透明度、可反驳的侵权推定、潜在 5–7% 营业额固定费率费用,以及覆盖向欧盟提供服务的非欧盟提供商的地域效力。审计 Decart 的训练数据来源、透明度披露和欧盟合规路线图。
美国州级 AI 法律于 2026 年 1 月 1 日生效(CA SB 942 + AB 2013、TX TRAIGA)约束2026 年 1 月 1 日起增加逐州透明度、训练数据和治理义务;Trump 2025 年 12 月 11 日行政令正在测试联邦优先权。将 Decart 的加州敞口(旧金山研发中心)与 AI Transparency Act 和 AB 2013 义务逐项对照。
GPU / Trainium 计算成本和容量约束持续至 2030 年实时生成受计算约束;DOS 多厂商优化是一道对冲,但算力分配仍是硬约束。按厂商(NVIDIA / TPU / Trainium)量化 Decart 的计算 COGS 和承诺使用合约。
玩家对生成式 AI 艺术 / 动画的情绪约束现在BCG 调研显示,成人对 AI 艺术持负面看法的比例只有约 10%,对 NPC / 任务持负面看法约 5–7%,但消费者反弹风险不为零,尤其对 AAA 品牌。按滚动 6 个月窗口跟踪工作室层面披露和玩家社区情绪。
用 Minecraft 素材训练带来的版权不透明(Oasis 2024 年 10 月 + 2026 年欧盟框架)约束自 2024 年 10 月以来的持续风险这是一个反向信号,会放大欧盟版权压力;还可能影响加州 AB 2013 下未来美国执法的形态。索取 Decart 的训练数据清单,以及围绕 Minecraft 素材使用与 Microsoft / Mojang 或其他授权方的往来函件。

约束行是本章最关键的尽调内容;在承销前,应对照 Decart 的合规姿态和任何未公布授权交易继续追问。

[CM031, CM032, CM033, CM034, CM035, CM036]

2.5 图表

Chapter 03

03竞争对手

3.1 2026 年世界模型版图:直接实验室、既有厂商、邻近玩家与内部自建替代

Decart 并不在一个边界清晰的软件品类中竞争。面向沉浸式体验的 Lucy 模型,竞争对象是 Runway、Luma AI 等视频生成平台,而这些平台本身也在转向世界模型。面向 Physical-AI 模拟的 Oasis 模型,同时面对研究级系统(Google DeepMind 的 Genie 3,2025 年 8 月以研究预览形式发布),以及 Waymo、Tesla、NVIDIA、Wayve 的大型内部自建项目。DOS 推理栈则与超大规模云厂商 GPU 云和专门的推理优化供应商竞争。 三条产品线的共同主线,是低于 30 毫秒延迟的实时、交互式生成——截至 2026 年中,还没有竞争对手在商业可用 API 中匹配这一能力。Genie 3 能以 24 fps、720p 生成帧,但仍是研究预览,没有公开付费 API。Runway 最新的 Gen 4.5 可生成带音频和角色一致性的高清视频,但面向后期制作和创意工作流,而不是实时交互模拟。World Labs 的 Marble 产品生成可下载的 3D 环境,服务设计和娱乐,而不是实时驾驶模拟。Luma AI 于 2025 年 11 月融资 $900 million,估值达数十亿美元,并在扩展到世界模型,但尚未公开披露实时交互 API。 最直接削弱 Decart 长期论点的两类竞争者是:(1)顶级 AV 玩家内部自建,这些玩家拥有 Google 级研究预算和专有传感器数据;(2)已经以相近估值融资的通用视频实验室未来商业化发布。Scenario 和 Inworld AI 分别代表聚焦游戏工具和 AI 角色的邻近替代,但并不直接覆盖实时驾驶模拟。[CP001, CP004, CP005, CP006, CP007, CP019]

竞争对手画像表
竞争对手类别规模 / 融资目标细分差异化局限
Runway直接视频 / 世界模型实验室$315M Series E,估值 $5.3B(2026 年 2 月);约 140 人团队媒体、娱乐、广告;正扩展到游戏和机器人具备物理感知的视频(Gen 4.5)、世界模型 R&D、Adobe 合作没有实时互动 API;生产重点是创意工作流,不是仿真
World Labs直接空间 AI / 世界模型实验室$1B 融资,包括 Autodesk $200M;目标估值约 $5B(2026 年 2 月)设计、娱乐、建筑;与 Autodesk 做 3D CAD 集成Marble 3D 环境创建、Fei-Fei Li 的学术背书产品只能下载 3D,不是实时驾驶仿真;商业范围仍早期
Google DeepMind / Genie 3在位研究实验室 / 内部仿真Alphabet 内部 R&D 预算近乎无限研究社区;内部 Waymo 仿真;未来企业业务待定24 fps / 720p 互动世界生成、可由提示词触发世界事件,被 Waymo 使用仅研究预览,没有公开付费 API;部署时间线未披露
Luma AI相邻视频 / 世界模型实验室$900M Series C,由 Humain 领投(2025 年 11 月);2-GW 沙特 AI 超级集群合作按已披露路线图,面向创意、游戏和 AV 行业Dream Machine 视频生成;已声明世界模型方向,但尚未作为 API 发布截至 2026 年中没有实时互动仿真 API;产品成熟度低于 Runway
Inworld AI相邻 / 游戏实时语音 AI$50M 融资,估值 $500M;总融资超过 $100M消费级陪伴应用、游戏角色 AI;OtherHalf 产品 19 天获 100 万用户实时语音 AI,TTS 低于 $13 / 100 万字符;已从游戏 NPC 引擎重心转向不是世界模型竞争对手;买方细分不同于 Oasis(仿真 vs 角色 AI)
Rosebud AI相邻 / AI 游戏创建种子阶段;Crunchbase 上公开融资数据有限独立开发者用文本提示创建 3D 游戏无代码游戏创建,非工程师也能上手没有实时推理或仿真能力;不是 Oasis 或 Lucy 的替代品
内部 AV 仿真(Waymo / Tesla / NVIDIA / Wayve)现状 / 内部自建顶级 AV 项目近乎无限;NVIDIA DRIVE Sim 预算未披露顶级 AV OEM 和一级 AV 项目,且内部有世界模型 R&D自有传感器数据、物理引擎集成、监管熟悉度外部买方无法获得;可寻址市场只剩自筹项目
现状:没有世界模型仿真(手工采集数据 + 传统仿真器)现状 / 替代品零新增资本开支;使用现有传感器车队和传统游戏引擎仿真器不愿按秒支付生成成本的 AV 项目;更小团队无需新资本;成熟游戏引擎(CARLA、SUMO)具备确定性物理无法规模化生成罕见边缘场景;照片级真实感有限

各行覆盖 Decart 买方会评估的主要竞争类别。内部自建行反映可寻址市场上限,并非缺少内部能力的买方可行替代。融资数据来自截至 2026 年中的最新披露轮次;私营公司可能还有未披露的资本活动。

[CP001, CP004, CP005, CP006, CP007, CP008]
FP001: 竞争定位图

Decart 是唯一同时拥有生产级实时能力和公共 API 的厂商;在位方研究预算更大,但仍停留在仅研究或非互动阶段。

分数是基于公开产品界面、API 可用性和融资证据作出的序数判断,不是直接披露指标。x 轴反映商业 API 就绪度;y 轴反映实时互动仿真能力深度。

[CP001, CP004, CP005, CP006, CP007, CP009]

3.2 竞争对手画像:Runway、World Labs、DeepMind/Genie 3、Luma AI、Inworld 与 Rosebud

Runway 于 2026 年 2 月完成 $315 million Series E,估值几乎翻倍至 $5.3 billion。公司最知名的是具备物理感知的视频生成(Gen 4.5),并明确在为媒体以外的产品和行业预训练「下一代世界模型」。Runway 已与 Adobe 建立合作,并在扩展到游戏和机器人,使其成为 Decart 的 Lucy 模型在创意和媒体工作流中最直接的前沿同业。它与 CoreWeave 有算力合作,2026 年 2 月团队约 140 人。 World Labs 于 2026 年 2 月融资 $1 billion,其中包括 Autodesk 的 $200 million,估值据报目标为 $5 billion。公司由 Fei-Fei Li 创立,Marble 产品可从图像、视频或文本生成可编辑、可下载的 3D 环境。Autodesk 合作最初聚焦媒体和娱乐用例,并把 World Labs 定位为 Decart Oasis 世界在 3D 内容创作工作流中最接近的空间 AI 同业,但不是实时驾驶模拟同业。 Google DeepMind 的 Genie 3 于 2025 年 8 月以研究预览形式发布,可在 24 fps 和 720p 下生成交互世界,支持可提示的世界事件,并已驱动 Waymo 的内部模拟架构。它没有作为公开商业 API 提供,这给 Decart 留出了一个时间窗口,直到 Google 可能向外部变现这一能力。 Luma AI 于 2025 年 11 月融资 $900 million,并在沙特拥有 2-gigawatt AI 超级集群合作。公司主要以 Dream Machine 和视频生成知名,世界模型是已声明的未来方向。Inworld AI 转向面向消费者陪伴应用的实时语音 AI,在 $500 million 估值下 19 天获得 100 万用户,不再是直接模拟竞争者。Rosebud AI 专注文本到 3D 游戏创作,代表开发工具替代,而不是推理基础设施竞争者。[CP002, CP003, CP004, CP005, CP007, CP008]

功能 / 能力矩阵
购买标准Decart(Lucy / Oasis / DOS)Runway Gen 4.5World Labs MarbleDeepMind Genie 3Luma AIInworld AI
实时互动生成(低于 30ms 延迟)强(生产,<30ms TTF)不可用不可用中(24fps,仅研究)不可用不适用
驾驶 / 物理 AI 仿真强(Oasis 3,相机级准确)不可用不可用中(Waymo 内部使用;无公开 API)unknown不适用
创意 / 视频转换强(Lucy 2.1,电商、流媒体)强(Gen 4.5,媒体 / 广告)不适用不适用强(Dream Machine)不适用
3D 环境创建不可用不适用强(Marble,可编辑 3D)中(3D 互动世界,仅研究)不适用不适用
按量付费定价的开发者 API强($0.02/sec,公开)未知(无公开按秒定价)未知(公开定价有限)不可用(研究预览)未知(无公开 PAYG API)强(实时语音 API,$0.10/hr STT)
硬件无关优化(NVIDIA/AWS/TPU)强(DOS 多硬件)unknownunknownunknownunknown不适用
大规模生产部署(>100K 开发者)强(Lucy 2.1 已投产,100K+ 开发者)中(媒体 / 广告生产)部分可用(早期产品)不可用中(Dream Machine 已投产)强(OtherHalf 19 天获 1M 用户)

单元格是基于截至 2026 年 6 月已审阅公开来源给出的有证据支撑的顺序评级。「不可用」表示截至审阅日期,该能力尚未以商业 API 提供。「未知」表示已审阅来源不足以支持判断。Inworld AI 的仿真单元格标为「不适用」,因为它竞争的是语音 AI,而不是世界模型仿真。

[CP003, CP006, CP010, CP012, CP016, CP017]
FP002: 功能广度 / 能力地图

Decart 在实时推理和驾驶仿真领先;Runway 在创意制作领先;World Labs 在 3D 创作领先;Google DeepMind 研究深度领先,但商业访问不足。

序数评级基于已审阅公开来源。“无”表示截至 2026 年 6 月,商业 API 中没有该能力;“未知”表示已审阅来源不足以判断。Inworld AI 实时语音属于与仿真不同的产品垂直,但作为相邻背景纳入。

[CP003, CP006, CP010, CP016, CP017, CP019]

3.3 Decart 差异化:实时延迟、DOS 成本护城河与开发者生态策略

截至 2026 年中,Decart 的核心差异化建立在三根支柱上,竞争对手尚未在生产系统中匹配。第一,实时推理延迟:Lucy 的首帧时间低于 30 毫秒,并通过针对 Amazon Trainium3 优化的 DOS 栈生成最高 100 帧 / 秒,而 Genie 3 是 24-fps 的研究专用管线。第二,成本效率:Decart 称借助 DOS,相比可比系统成本效率提升 100x;DOS 在 Trainium3 硬件上实现超过 80% Model FLOPS Utilization。第三,开发者生态:Oasis 3 从第一天起就以公共 API、每秒 $0.02 定价,Decart 正在构建让 OpenAI 文本 API 变得粘性的同一类开发者生态飞轮——这家初创公司已经有超过 100,000 名开发者在 Lucy 上构建。 DOS 栈的架构差异在于它能运行在任意硬件(NVIDIA GPU、Amazon Trainium、Google TPU)上,而不是绑定单一云服务商。Oasis 3 生成三路同步摄像头画面(正前方加两个侧前方),匹配多数摄像头优先 AV 栈的感知设置。Decart CEO 称公司累计支出「远低于」$100 million,尽管融资超过 $450 million;若属实,说明 DOS 成本效率不仅作用于外部销售,也作用于内部推理经济性。不过,这是公司主张指标,没有独立审计佐证。 关键战略押注是:负担不起 Google-DeepMind 级研究预算的中型 AV 公司、机器人初创公司和无人机项目,会选择以 $0.02/sec 租用 Oasis 模拟,而不是自建内部系统——CEO 估算全球有数十个这样的可服务项目。[CP009, CP010, CP011, CP012, CP021, CP022]

定价 / 包装对比
供应商公开套餐价格 / 单位 / 模型已包含能力折扣或未知项含义
Decart(Lucy 2.1)按量付费 API活跃生成 $0.02/sec(720p);Lucy Restyle 2 为 $0.01/sec实时视频编辑、风格迁移、虚拟试穿、驾驶仿真预览企业定价按用例扩展;Oasis 3 企业费率未披露零售标价透明;企业溢价和批量折扣未公开
Runway套餐层级 + API未披露按秒公开 API;创意工具采用订阅层级Gen 4.5 视频生成,含音频、角色一致性、长视频按秒生成成本未公开暴露;企业条款定制与 Decart $0.02/sec 可比的 API 工作负载看不到透明定价
World LabsMarble 产品未公开披露;Marble 于 2024 年 11 月发布,但审阅来源未找到定价从图像 / 视频 / 文本创建 3D 世界,可编辑导出定价模型不清楚;与 Autodesk 的企业合作仍处研究层面未暴露可比的按用量定价;开发者在成本未知时更难采用
Google DeepMind / Genie 3研究预览(无商业 API)24fps/720p 的互动世界生成;可用提示词触发世界事件无可用定价;Waymo 和研究场景仅内部部署尚未商业化;Google 决定外部定价前,Decart 拥有时间窗口
Luma AIDream Machine(视频)+ 规划中的世界模型视频采用消费者订阅层级;世界模型 API 定价未披露Dream Machine 视频生成;世界模型 API 即将推出审阅的公开来源未找到世界模型 API 定价世界模型能力无法直接做定价对比
现状(传统仿真器 + 传感器车队)CARLA、SUMO、专有游戏引擎流水线定制物理仿真、确定性、已验证资本开支重的传感器车队采集;工程团队开销总拥有成本取决于团队规模和车队规模;没有按秒边际成本

null 值表示审阅的公开来源未找到定价,不表示供应商免费。所有价格均为标价 / 零售价;实际企业定价可能显著不同。

[CP010, CP011, CP015, CP029, CP032]

3.4 护城河耐久度、切换成本与商品化风险

Decart 的护城河真实存在,但很窄,也有时间边界。当 Oasis 3 仍是唯一生产就绪的实时驾驶模拟 API 时,迁移到 Oasis 3 具备粘性:一旦 AV 训练管线集成 Oasis 生成场景,切换成本就包括用不同格式重新收集或重新标注合成数据。不过,这条护城河至少有三处结构性脆弱点。 第一,最有能力构建竞争性实时世界模型的两家公司——Google(通过 DeepMind)和 NVIDIA(虽是 Decart 投资人,但运行自己的 DRIVE 模拟栈)——R&D 预算比 Decart 高几个数量级。Waymo 已经在内部部署 Genie 3,为顶级 AV 梯队自给自足立下先例。第二,Runway、World Labs 和 Luma AI 都在明确投资世界模型,并已获得相当或更大的融资;如果其中任何一家推出有竞争力的实时 API,Decart 的先发开发者生态优势将面临直接压力。第三,通过 AWS Bedrock 的分发渠道虽有利于触达,却让 Decart 依赖 Amazon 的市场进入和定价架构。 Scenario 的聚合器模式提供 50+ 家供应商的 550+ 个模型,说明商品化基础设施市场会如何演化:一旦多个世界模型供应商存在,聚合器就可能压低任何单一供应商的定价护城河。眼下这个风险仍偏前瞻,但它正确框定了 3–5 年竞争耐久度问题。[CP009, CP024, CP030, CP031, CP032, CP034]

护城河耐久性 / 竞争风险清单
护城河主张威胁严重性缓释措施 / 尽调问题
截至 2026 年中,Decart 拥有独特实时延迟(<30ms),没有商业同侪Google DeepMind Genie 3 可能推出商业 API;NVIDIA DRIVE Sim 已能在受控 AV 环境中接近实时运行跟踪 Genie 3 商业化路线图;把 Oasis 3 延迟基准与 2026 年下半年任何新的商业版本对比
DOS 成本栈支撑 Decart 相比可比系统 100x 效率主张Runway、World Labs 和 Luma AI 已为算力扩张融资数亿美元;云服务商可能用标准内核把推理效率商品化要求 100x 主张的独立基准证明;评估 DOS 效率来自硬件协同设计,还是来自可复制的模型架构选择
100K+ 开发者社区在既有玩家发货前形成 API 粘性Genie 3 拥有 Google Developers 生态;Runway 和 World Labs 已有创意社区;若 API 合同不是长期合同,切换成本较低评估开发者留存率和再激活率;理解合同期限和单个开发者平均 API 支出
通过 Bedrock 与 Trainium3 合作接入 AWS 分发,给 Decart 带来企业触达AWS Bedrock 是多供应商平台;Amazon 会在同一市场提供竞争模型;若竞争世界模型进入 Bedrock,分发优势会消失审阅 AWS 排他条款和联合销售承诺;评估 Toyota、Adobe、eBay 的投资人 / 客户关系是否能在 AWS 之外形成多元分发
Oasis 3 是唯一按零售价提供的照片级多摄像头驾驶仿真生产 API物理一致性问题(刚体、实体碰撞)会削弱 AV 公司信心;Waymo 内部使用 Genie 3,为顶级 AV 项目自建树立先例获取 AV 公司试点证据,确认生成场景确实迁移为真实驾驶改进;评估 CEO 2026 年 6 月承认的物理一致性修复时间线

严重性评级是基于已审阅证据的顺序判断,不是正式风险分数。该清单聚焦最可能在 3–5 年内影响竞争耐久性的因素。

[CP002, CP006, CP009, CP011, CP013, CP014]
FP003: 护城河 / 就绪度 KPI

Decart 的 100K+ 开发者社区和 $300M 融资,使其按商业 API 访问计成为领先的世界模型基础设施厂商;但三类结构性威胁限制了耐久性。

开发者数量由公司声称,未经审计。估值来自最新披露融资轮。内部 AV 项目数量是公开来源中具名项目给出的最低下限;实际数量可能更高。

[CP001, CP004, CP005, CP009, CP015, CP028]

3.5 负面动态:物理一致性限制、内部自建与买方谨慎

TechCrunch 在发布时的独立测试发现,Oasis 3 的环境主题完整性会在长时间交互中快速退化——一个 New York City 街道提示最初生成了强场景,但很快变成泛化的西部城市环境,模型也无法在导航中保留地理地标。更关键的是,Oasis 3 中车辆会彼此穿过,说明模型尚未正确模拟刚体物理。Decart CEO 承认这是「我们正在攻克的重大研究问题」,并将其归因于正常驾驶和碰撞场景之间的数据不平衡。 这些限制很重要,因为评估 Oasis 3 用于边缘案例合成数据生成的 AV 公司,在投入训练预算前,需要验证生成场景能否迁移到真实世界行为。不尊重物理的合成数据,可能把分布伪影引入已训练的感知或规划模型。 在内部自建维度,Waymo、Tesla、NVIDIA 和 Wayve 都在内部构建可比的世界模型模拟。因此 Decart 真正可服务市场,是无法自筹世界模型研究努力的第二、第三梯队 AV 项目、机器人实验室和无人机初创公司。可服务客户集中在中端梯队,会压低近期收入天花板;如果大型客户数量很少,也会带来集中度风险。[CP013, CP014, CP015, CP023, CP031, CP032]

3.6 图表

Chapter 04

04财务

4.1 收入模式与定价架构

截至 2026 年中,Decart 经营三条有公开文档的收入流。第一条——也是历史上最早的一条——是向主要云服务商、AI 实验室和超大规模基础设施公司授权 DOS,合同金额达数百万美元。基于这项优化栈的盈利能力,CEO 才能在 2025 年 8 月声称 Decart 在已融资的 $153 million 中花费不到 $10 million,并靠授权收入覆盖数千万美元 GPU 算力成本。第二条是 API 按秒付费:Lucy 2.1(实时视频编辑,720p)每生成秒定价 $0.02,Lucy Restyle 2(风格迁移)每秒 $0.01,Oasis 3 Preview(physical AI 与驾驶模拟)自助访问每秒 $0.02,上一代 Lucy Clip 每秒 $0.15。Oasis 3 和定制部署的企业定价取决于用例,未公开披露。第三条是通过 AWS Bedrock 做企业分发,作为 2026 年 5 月 $300M Series C 的一部分宣布,使 Amazon 企业客户群能够在定制商业条款下访问 Decart 技术,用于媒体、商业、广告和 Physical AI 用例。各收入流的收入确认动态不同:DOS 授权合同可能按交付完成或可摊销授权确认;按秒付费 API 收入随消耗确认;Bedrock 企业合同会遵循协商后的多年交付计划。这些收入流之间的组合及各细分相对规模均未公开披露,构成收入质量的首要尽调缺口。 [CI001, CI003, CI004, CI006, CI007, CI014]

收入流表
收入流机制当前状态收入质量主要尽调问题
向超大规模云服务商授权 DOS以多年合同向云服务商和 AI 实验室授权 Decart Optimization Stack活跃;CEO 称有数百万美元级合同(未披露数量或期限)中高(战略性,可能经常性)合同数量、期限长度、集中度风险
Lucy API(实时视频)按秒付费,$0.02/sec(720p);100K+ 开发者社区活跃;自助和企业两种模式;主要开发者收入流中(可变;取决于开发者使用率)MRR、并发利用率、单开发者 API 收入
Oasis API(物理 AI / 驾驶)按秒付费,$0.02/sec(自助);企业按用例定价2026 年 6 月起活跃;面向 AV 仿真和机器人垂直领域中低(早期市场;Oasis 3 于 2026 年 6 月推出)AV 仿真合同管线;物理缺口缓释时间线
虚拟试穿(Lucy VTON 3)按 API 调用收费;电商虚拟服装试穿活跃;电商客户正在采用中(电商需求增长;用例清晰)ARPU、采用漏斗、零售商合同规模
通过 AWS Bedrock 做企业分发通过 Amazon Bedrock 市场签定制商业合同;联合 GTM活跃;2026 年 5 月作为 $300M 轮战略合作的一部分签署中(合同金额和条款未披露)合同金额、排他条款、Amazon 收入分成条款
视频模型 API(非实时)按生成秒收费,$0.04/sec(480p);旧模型层级活跃(旧版);优先级低于实时模型中低(旧模型;可能低增长)收入贡献;下线或维护计划

所有收入状态评估均来自公司公告和媒体报道;ARR、MRR 或客户数量拆分均未披露。收入质量评级是分析师基于公开可见机制作出的判断,不是审计数据。

定价 / 变现表
产品模型 ID分辨率单位价格计费基础示例成本
Lucy 2.1(实时视频编辑)lucy-2.1720p$0.02/sec按活跃生成秒计费30 秒会话:$0.60
Lucy Restyle 2(风格迁移)lucy-restyle-2720p$0.01/sec按活跃生成秒计费30 秒会话:$0.30
Lucy VTON 3(虚拟试穿)lucy-vton-3720p未披露企业 / 按 API 调用未公开披露
Oasis 3 Preview(世界模型 / 驾驶)oasis-3-preview720p$0.02/sec(自助);企业不同按秒;企业按用例60 秒会话:$1.20
视频编辑模型(480p)lucy-video (480p)480p$0.04/sec按生成秒计费(视频模型)5 秒编辑:$0.20
Lucy Clip(旧版实时)lucy-clip未指定$0.15/sec按活跃生成秒计费(旧版)10 秒会话:$1.50

定价来源为 docs.platform.decart.ai/getting-started/pricing,截至 2026-06-15。VTON 和 Oasis 3 企业层级的企业定价未公开披露。

FI001: 收入模型桥

Decart 收入来自三条主线——DOS 许可、Lucy API 和 Oasis API——并通过 AWS Bedrock 企业渠道汇入披露的总收入。所有美元数值均为公司声称或估算;没有经审计数据。

各收入流金额未公开披露。相对流量基于历史先后顺序作示意,不代表实际收入分成数据。

4.2 GTM 动作与开发者飞轮

截至 2026 年中,Decart 的市场进入策略把自助开发者 API 与通过战略合作伙伴推进的直接企业销售结合起来。开发者主导路径已经形成超过 100,000 名开发者的社区,主要在 Lucy API 上构建电商虚拟试穿和直播转换用例;2026 年 6 月推出的 Oasis 3 则把自动驾驶模拟开发者作为第二个垂直方向。AWS Bedrock 集成是企业分发锚点,让 Decart 能接触 Amazon 销售队伍和云市场信誉,而不必从零搭建直接企业销售组织。这降低了前期获客成本,但也把利润率和控制权交给 Amazon 的商业条款。CEO Leitersdorf 将战略投资人 Toyota、Adobe、eBay 描述为「都可能成为客户」,形成一个被锁定的初始企业管线,但它并不反映独立公平交易下的商业转化。NVIDIA 既是投资人又是硬件伙伴,也进一步模糊了财务支持与商业需求之间的边界。公司未披露按收入层级划分的客户数、净收入留存或销售周期数据。核心 GTM 风险是:Decart 收入集中在少数超大规模云厂商 DOS 授权大单,同时拥有庞大但大多尚未产生收入的开发者基础,由此形成杠铃式收入结构——少数大合同抵消高量但低变现的开发者 API 使用——在没有按收入划分客户明细前,整体收入质量很难评估。 [CI005, CI006, CI007, CI015, CI016, CI022]

4.3 成本结构、毛利驱动与资本强度

Decart 的成本结构由算力和 R&D 薪酬主导;公司采用租用 Amazon Trainium3、NVIDIA GPU 和 Google TPU 容量的轻资产模式,因此没有自有硬件资本开支。规模化托管推理 API 业务的毛利率通常会达到 50–80% 的软件定义服务区间,但 Decart 的实际利润率关键取决于它与超大规模云厂商签订的算力协议条款、Trainium3 相对标准 NVIDIA 实例定价的成本,以及 DOS 授权收入是否比 API 使用拥有显著更高利润率。公开说法中,DOS 优化层在 Trainium3 上实现超过 80% Model FLOPS Utilization——这是 AWS VP Nafea Bshara 引用的硬件效率指标——并把每条视频的生产成本从数百或数千美元降至低于 $0.25。若属实,这些效率提升会相对使用标准推理基础设施的同业,直接转化为毛利改善。不过,这些都是公司主张的基准,没有独立第三方审计。随 $300M 融资一起发布的 DOS 2.0 声称智能体推理达到 1,600+ tokens per second,而行业平均约 200 tokens per second,世界模型最高达到 100 frames per second。如果成本优势确如描述,Decart 的 API 收入毛利率可能显著高于商品化推理 API 利润率。60+ 名员工的 R&D 薪酬(许多具备深厚底层系统和 HPC 专长)增加了显著固定成本基础。公司没有公开披露 COGS、毛利率、员工成本或任何单位经济指标。 [CI009, CI010, CI011, CI021, CI024, CI034]

单位经济表
指标披露值置信度重要性尽调问题
毛利率(混合)未披露N/A$4B 估值下的首要价值驱动因素要求按收入分部提供经审计 P&L
月度烧钱率(2026 年 Q2)未披露N/A资本充足性和现金 runway 评估要求董事会批准预算和实际支出
月度 API 收入(MRR)未披露(CEO 于 2025 年 8 月称“数百万美元收入”)低(公司声称;未验证)收入质量基线要求按产品线拆分 MRR,并提供用户 cohort 明细
DOS 授权平均合同价值每份合同数百万美元(CEO 声称;2025 年 8 月)低(公司声称;无合同数量或期限)集中度和覆盖风险要求合同数量、条款、续约安排
开发者获客成本(CAC)未披露N/APLG 效率和回本周期要求开发者获客和激活指标
单视频计算成本(经 DOS 优化)每个视频 <$0.25(CEO 声称;2025 年 8 月)低(公司声称;无第三方基准)API 收入毛利率替代指标要求独立计算成本基准或 AWS 账单数据

所有 null 值都反映公开披露确实缺失。Series C 阶段最优先尽调的指标是毛利率。

FI002: 单位经济桥

定性单位经济从算力投入,经 DOS 优化流向 API 收入和毛利率,并标注已知数据点和证据缺口。所有毛利率估计均为推断;没有经审计毛利率数据。

50–80% 毛利率区间是对无自有算力的软件定义推理 API 的结构性推断;Decart 实际毛利率未披露。$0.25/video 数字由 CEO 声称,未经过独立基准验证。

4.4 资本充足性与融资依赖

2026 年 5 月 $300M Series C 后,Decart 总资本化超过 $450 million。融资时间线包括约 $53 million 的种子和早期轮次、$100 million Series B(2025 年 8 月,估值 $3.1 billion,由既有投资人 Sequoia、Benchmark、Zeev 领投,加上新投资人 Aleph VC),以及约 $4 billion 估值的 $300 million Series C(2026 年 5 月,由 Radical Ventures 领投,战略参与方包括 NVIDIA、Toyota Ventures、Adobe Ventures、eBay Ventures,Amazon 是战略客户)。JSL Decart.AI Coinvest, L.P.(CIK 0002084011)于 2025 年 9 月提交的 SEC Form D 确认,为 Series B 成立了 Delaware LP 共同投资载体,但文件对投资人构成或估值机制提供的额外细节很少。CEO Leitersdorf 于 2025 年 8 月称,Decart 在 $153 million 资本中花费不到 $10 million,运营成本由 DOS 授权收入覆盖。 2026 年 6 月,他把口径改为累计支出「远低于 $100 million」——与 2025 年 8 月说法一致,但不如前者精确。两项数字都来自 CEO 表述,没有独立审计或董事会证明。若按表面接受累计支出说法,Decart 进入 2026 年中时约有 $350–$440 million 现金,即便月度消耗显著高于 CEO 暗示的低于 $1M/month,也意味着多年资金跑道。战略投资人也参与商业:NVIDIA 作为硬件伙伴和 DOS 分发载体,Amazon 作为客户和 Bedrock 分发渠道,Toyota、Adobe、eBay 则作为潜在客户和企业需求验证者。这制造了融资—商业共依赖:投资条款可能部分基于商业承诺,从而削弱两种信号的独立性。 [CI001, CI002, CI012, CI013, CI017, CI020]

资本充足性表
项目数值 / 估算证据基础置信度尽调问题
已融资总额(所有轮次)>$450M($53M 种子轮 + $100M Series B + $300M Series C)第三方媒体和 Decart 官方公告确认用股权结构表确认精确数字
截至 2026 年 6 月的声称累计支出<$100M(CEO,2026 年 6 月);截至 2025 年 8 月 <$10M(CEO)仅 CEO 声称;无审计或独立确认经审计现金流量表;CFO 证明
推算在手现金(2026 年 5 月 Series C 交割后)$350M–$440M(按融资额减声称支出估算)估算;烧钱数据未验证,区间很宽截至 2026 年 5 月交割时的经审计资产负债表
$300M 融资计划用途R&D 扩张、基础设施、美国招聘(SF/NY R&D 中心)、模型训练公司在 $300M 公告中陈述带里程碑的董事会批准资本配置计划
SEC Form D 共同投资工具JSL Decart.AI Coinvest, L.P.(Delaware LP;2025-09-19 提交)SEC 文件(公开记录)确认 LP 投资人身份和 Series B 条款

在手现金估算假设约 $453M 融资额减去 CEO 声称的低于 $100M 累计支出。该估算完全依赖 CEO 未验证的支出说法,只能视为指示性。

FI003: 财务估算区间

Decart 关键财务估算的低到高区间。所有数值均为推断或公司声称;没有经审计披露支撑这些区间。除非另有说明,数值单位为百万美元。

累计融资采用已确认公开数字(>$450M)。隐含现金为累计融资减去 CEO 声称的支出区间($10M–$100M)。年化 API 收入假设 100K 开发者中 0.1–2% 同时使用,价格 $0.02/sec,每天 8 小时。跑道为 $300M 除以 $4.2M–$12.5M 的月度烧钱。所有估算都高度依赖未验证假设。

FI004: 资本强度 / 现金流地图

投资人资本如何穿过 Decart 轻资产运营模型,从 Series C 资金流向算力访问、R&D、收入生成和再投入。图中展示投资人资本、超大云合作伙伴和收入之间的融资—商业依赖。

流量大小仅为示意;算力、R&D 和再投入之间的实际资本分配未公开披露。从 revenue_generation 到再投入的回路,反映 CEO 关于 DOS 许可收入曾覆盖算力成本的说法。

4.5 收入质量、披露缺口与财务结论

Decart 讲出了有吸引力的资本效率叙事和多收入流架构,但几乎所有财务主张都来自 CEO,无法独立验证。截至 2026 年中,公司没有公开披露 ARR、毛利率、EBITDA、按层级划分的客户数、净收入留存、消耗率或经审计账目。公司是在以色列运营的私营实体,在运营司法辖区没有公开提交账目的法定义务。Decart 唯一公开金融工具,是共同投资载体 JSL Decart.AI Coinvest, L.P. 的 SEC Form D;这是一份程序性文件,确认 Delaware LP 存在,不提供财务表现数据。TechCrunch 对 Oasis 3 的评测(2026 年 6 月)指出重大负面技术限制——物理一致性失败、长会话环境退化、上下文窗口约束——CEO 自己也承认这些是活跃研究问题。这些限制直接影响 Oasis AV 模拟收入流的财务前景:如果无法在规模化下保证物理保真,企业 AV 客户会推迟或限制采用。2026 年 5 月 $300 million 融资——发生在 CEO 声称公司几乎不需要资本约九个月后——暗示公司战略转向激进增长投资(招聘、基础设施、模型训练、美国扩张),这与此前资本效率叙事存在张力。公司是在押注产品市场契合拐点,还是进入增长投资周期,在没有季度管理账目前无法判断。五个阻断性尽调问题是:(1)已确认 ARR 和按收入流划分的收入组合,(2)按产品线划分的毛利率,(3)截至 2026 年 Q2 的月度消耗率和现金头寸,(4)DOS 授权合同条款和集中度,(5)将「数百万美元收入」与 $3.1–4.0 billion 估值之间做调和;按标准 SaaS 倍数,后者隐含 $300M+ ARR。 [CI018, CI019, CI022, CI023, CI025, CI031]

公开财务缺口表
缺失指标对承销的重要性若不披露的影响尽调路径
按产品线拆分 ARR / MRR相对 $4B 估值的收入基线;用于计算 ARR 倍数无法判断隐含 20–40x ARR 倍数是否有支撑向 CFO 要求按收入流拆分 MRR(DOS 授权 / Lucy API / Oasis API / 企业)
毛利率(混合及按收入流)长期企业价值的首要决定因素;计算成本杠杆无法评估定价权或成本竞争力要求经审计 P&L;对比 DOS 授权毛利率与 API 毛利率
月度烧钱率和现金头寸资本充足性和融资风险;runway 评估没有烧钱数据就无法验证多年现金 runway 主张要求 2026 年 Q1 现金流量表和当前银行余额
DOS 合同数量和 HHI收入集中度风险;客户依赖无法评估失去一个超大规模云服务商合同是否会造成重大收入下滑要求 DOS 授权合同数量、规模分布和续约条款
净收入留存(NRR)和流失率收入质量信号;经常性 vs. 交易性;客户满意度没有 NRR 就无法区分经常性收入和一次性收入分别要求开发者 API 和企业收入流的用户 cohort NRR 分析

五个指标都是 AI 基础设施公司 Series C 阶段的标准承销输入。缺失这些指标,意味着承销完全依赖未经验证的 CEO 说法——对 $4B 估值而言,这是非标准风险因素。

4.6 图表

Chapter 05

05产品与技术

5.1 技术栈定义与模块图谱

Decart 将公司定位为纵向一体化研究实验室,而不是单一模型供应商。公开技术栈分成三层:DOS 是优化和推理层;Lucy 是服务商业、流媒体、游戏和直播媒体的实时视觉转换与试穿产品族;Oasis 3 是用于闭环 Physical-AI 模拟的可提示世界模型。产品表面可通过云 API 访问,并提供官方 SDK 和公开定价,这比许多前沿模型发布更具体。与此同时,Decart 仍在快速切换叙事:2024 年 Oasis 发布以类似 Minecraft 的交互演示为中心,随后 Lucy 成为直播视频用例中可见的生产产品,2026 年 Oasis 3 发布又把公司重新框定在 Physical AI 和自动驾驶模拟周围。这种宽度在战略上有吸引力,但也意味着买方仍需要尽调:技术栈中已有多少被产品化,多少仍由研究驱动。[CE001, CE002, CE003, CE004, CE005, CE006]

产品模块和资产矩阵
模块用户 / 买方成熟度 / 状态关键差异化尽调缺口
DOS / 优化引擎云服务商、AI 实验室、Decart 内部模型团队公司称已投产;DOS 2.0 于 2026 年 5 月发布跨 NVIDIA、TPU 和 Trainium 的硬件可移植低延迟推理与训练层没有独立基准包或客户案例研究量化其声称的 8x 速度 / 100x 效率优势
Lucy 2.1 实时及相关 Lucy 应用构建电商、流媒体、游戏和直播媒体应用的开发者公开文档、SDK、定价和示例可用有 SDK 支持和客户端令牌浏览器流程的实时视觉转换生产客户的独立质量和留存数据未披露
Oasis 3 PreviewAV、机器人和物理 AI 开发者2026 年 6 月公开 API 预览已上线可提示的多视角世界模型,带明确动作循环和公开按秒计费公开证据尚未显示激光雷达、分割、确定性回放或长时域验证
API / SDK 表面应用开发者和平台集成方可见通用可用性文档和官方 SDK已发布的 JavaScript、Python、Swift、Android 和专门 Oasis 工具降低了集成摩擦企业支持、SLA 条款和版本保证尚未公开详述
开源开发者资产社区开发者、VR 创作者和实验者GitHub 组织包含 SDK、RL 示例、试穿和 Quest XR 仓库可见的社区示例不只停留在营销页面,也构成开发者信号证明仓库活动可见,但公开 issue 量和生产采用情况未披露

这是截至 2026-06-15 抓取的公开产品表面快照;成熟度标签区分有文档证明的可用性与经独立验证的生产规模。

[CE001, CE004, CE005, CE009, CE011, CE012]
工作流与用例表
用户任务当前工作流Decart 方案可衡量收益限制
实时视频转换开发者把摄像头采集、低延迟推理和实时渲染拼在一起Lucy 实时 API 加官方 SDK 和客户端令牌认证公开文档显示面向生产的浏览器和移动端流程,并采用按用量计费按细分场景独立衡量输出质量和正常运行时间的数据有限
电商虚拟试穿商户通常依赖静态图片或异步渲染Lucy VTON 和试穿指南支持基于摄像头的实时会话官方示例展示产品页试穿流程和 6 个生产就绪示例未披露公开商户转化率或重复使用指标
物理 AI 仿真团队往往内部构建稀有场景数据集或仿真器工具Oasis 3 Preview 根据提示和动作生成多视角驾驶场景公开 API 访问和 $0.02/sec 标价降低了实验门槛当前公开产品以摄像头为先,还不是完整的物理或传感器仿真器
VR / 沉浸式实验XR 构建者通常定制集成透视摄像头和风格化技术栈Decart-XR 开源 Quest 应用演示实时转换透视画面开源样例展示低于 200ms 的声称性能和完整 Quest 管线接线这是开发者演示,不是规模化商业部署的证据

收益仅反映已抓取文档和仓库明确展示的内容,不代表独立衡量的 ROI。

[CE004, CE005, CE006, CE007, CE011, CE013]
FE001: Decart 产品架构图

公开材料展示了一套分层栈:从硬件,经 DOS,到模型 API 和面向开发者的应用。

[CE001, CE003, CE004, CE021, CE022]

5.2 Oasis 3 API 机制与工作流

Oasis 3 文档展示的是一套真实运行工作流,而不是模糊的演示叙事。开发者连接托管 gRPC 端点,用文本提示初始化场景,然后反复发送由四组油门和转向配对组成的固定动作块。服务会为每条已标示的摄像头流返回四帧 RGB,文档将这些摄像头描述为左前、正前和右前。这个接口支撑了一个具体主张:产品为闭环推理而建,不是一次性视频渲染。文档还包含 RL 示例和碰撞风险演示,使产品比缺乏可复现机制的研究预览明显更接近开发者可用。不过,公开证据尚未显示许多 AV 技术栈想要的更丰富模拟模态,例如 lidar、分割或显式物理状态导出,因此当前公开产品最强的用途,似乎是摄像头优先 实验和早期工作流集成,而不是完整替代模拟器。[CE011, CE012, CE013, CE014, CE015, CE016]

技术架构表
层 / 组件角色依赖风险
客户端层浏览器、移动端、Python 和 XR 应用采集媒体、提示和控制信号短期客户端令牌、官方 SDK、设备摄像头访问认证配置错误或客户端密钥处理不当,可能暴露凭证或中断会话
API / SDK 层把请求路由到 Lucy 和 Oasis 端点,并规范集成模式Decart 文档、SDK 维护和端点稳定性版本漂移或企业控制不足,可能提高切换成本
实时推理层用实时响应循环运行 Lucy 编辑或 Oasis 多视角 rollout模型架构加 DOS 运行时效率延迟、连贯性或内存限制会直接拉低用户体验和训练可用性
DOS 优化层在 GPU、TPU 和 Trainium 技术栈上抽象硬件专属优化深度硬件协作,以及持续的编译器 / 运行时调优竞争优势可能难以从外部验证,也可能依赖持续的合作伙伴访问
云 / 硬件层底层算力运行在 NVIDIA、AWS Trainium 和 Google TPU 环境合作伙伴容量、芯片路线图和云分发协议硬件集中或新芯片访问延迟,可能压缩 Decart 的实时优势

架构来自公开文档、仓库和公司发布内容的重建,而非 Decart 发布的正式系统图。

[CE003, CE013, CE014, CE021, CE022, CE023]
FE002: Oasis 3 运行流程

Oasis 文档定义了一个可复用的开发者循环:从 prompt 初始化到返回多视角帧。

[CE011, CE013, CE014, CE015, CE016, CE017]

5.3 性能主张、硬件耦合与差异化

Decart 最强的差异化主张,是实时跑世界模型和视频模型:公司把模型架构和 DOS 基础设施层一起设计。公司材料称,DOS 2.0 在 agentic inference 中可超过每秒 1,600 tokens,可跨 NVIDIA GPU、Google TPU 和 Amazon Trainium 运行,并把全高清视频或世界模型负载推向 100 FPS。独立报道部分支撑了硬件优化叙事:AI News 及合作伙伴报道描述了 Decart 如何在 AWS Trainium 上优化 Lucy,并把 Trainium3 的更低延迟作为目标;TechCrunch 也解释了 Oasis 3 的逐帧生成为什么会很快变成 token 吞吐问题。但成本效率优势、硬件可移植性、多视角一致性,仍主要来自公司或合作伙伴表述,而非第三方基准测试。相比 Genie 3 等其他世界模型路线,Decart 今天的差异化更多在于即时 API 接入、明确控制循环和生产化意图,而不是经独立验证的长时域真实感。[CE021, CE022, CE023, CE024, CE025, CE026]

路线图与发布表
日期 / 阶段功能 / 里程碑状态含义来源
2024-10Oasis 类 Minecraft 交互式演示和代码发布历史发布建立了 Decart 最初的实时世界模型可信度,但也暴露了版权和泛化问题官方 Oasis 项目页;TechCrunch 2024
2025Lucy 商业创意和直播部署推进按文档和合作伙伴报道仍在推进显示公司在转向物理 AI 之前,已有可变现的实时产品文档、AI News、合作伙伴报道
2026-05DOS 2.0 发布和 $300M 公告已公告表明基础设施变现和硬件可移植性是公司论点的核心Decart publication;CTech
2026-06-10Oasis 3 Preview API 发布在线预览 / 公开 APIDecart 从仅靠演示的世界模型定位,推进到公开开发者访问Oasis 文档;TechCrunch 2026;Startup Fortune
前瞻从 AV 扩展到无人机、机器人、海事和人形机器人用例公司路线图表述扩大 TAM,但 AV 相邻叙事之外的公开客户证明仍偏薄Oasis 落地页

日期和阶段反映明确的公开发布或出版标记;前瞻行是公司路线图表述,不代表已交付范围。

[CE020, CE021, CE024, CE025, CE027, CE028]
FE004: 产品成熟度和能力地图

公开证据显示,Lucy 是商业化程度最高的界面;Oasis 3 技术最新,但独立验证最少。

[CE004, CE020, CE022, CE028, CE030, CE031]

5.4 治理、隐私、安全与运营界面

与许多早期前沿模型产品不同,Decart 面向生产用户露出了清晰的运营界面:公开定价、客户端 token 的 FAQ、划分输入和输出所有权的条款、可接受使用政策,以及公开状态页。这些材料对尽调有用,因为它们把实际边界说清楚了。浏览器和移动端集成应使用短期客户端 token,而不是永久 API key;活跃会话在 token 过期后仍会保留;公司也通过专门状态站披露 uptime 历史。法律界面更值得谨慎。条款把输出权利分配给用户,但也为 Decart 保留了广泛权利,可用内容改进平台并用于营销。隐私政策称,公司可能处理输入、输出、生成媒体,以及实时音视频录制,用于产品改进和研究。AUP 禁止欺骗性 deepfake、军事用途,以及故障可能造成严重伤害的安全关键工作流;考虑到 Decart 将营销指向 physical-AI 和自动驾驶用例,这一点尤其值得关注。[CE033, CE034, CE035, CE036, CE037, CE038]

信任、安全与合规表
控制 / 政策状态范围缺口
客户端令牌指南FAQ 和模型文档中已有说明浏览器和移动端集成应使用短期令牌企业 IAM、SSO 或审计日志深度没有公开细节
服务条款已发布,日期为 2026-06-08分配输入所有权、输出转让和 Decart 内容使用权广泛的改进和营销使用权,可能需要为敏感数据设置合同例外
隐私政策已发布,日期为 2026-05-05覆盖输入、输出、生成媒体和实时音视频录制的处理公开材料未说明模型训练或研究复用的默认退出控制
可接受使用政策已发布,日期为 2026-02-12禁止欺骗性深度伪造、军事用途和高危安全关键场景面向物理 AI 和 AV 仿真的营销,与安全关键排除项形成张力
状态可见性公开状态页带历史正常运行时间视图平台服务的运营健康表面未披露公开 SLA、事故复盘标准或支持响应承诺

本表只捕捉可见的公开控制;企业合同控制可能更强,但抓取来源未披露。

[CE033, CE034, CE035, CE036, CE037, CE038]

5.5 独立警示与未解技术风险

公开证据最强的是接口,最弱的是长时域模拟质量。TechCrunch 的上手报道给出了最清晰的独立警示:模型能生成有说服力的写实驾驶场景,但场景身份会随时间漂移,其他车辆也可能互相穿透。这与 Decart 自己的提示一致。2024 年 Oasis 研究文章公开提到域泛化、记忆、远处细节模糊、精确库存或物体控制等问题,并称需要靠规模化解决。2026 年 Oasis 页面本身也说,系统不是物理引擎。这些不是早期世界模型平台的致命缺陷,但会影响采购判断,因为它们缩窄了当前可信部署范围。Decart 今天最可信的定位,是一个可通过 API 接入的实时视觉或模拟层,并带有异常完整的开发者工具;作为成熟物理模拟器的直接替代品,或作为安全关键自动驾驶的充分验证训练基底,证据还不够。[CE020, CE024, CE026, CE027, CE028, CE029]

FE003: 关键依赖地图

Decart 的实时产品论点取决于硬件伙伴、分发渠道和开发者采用能否一起跑通。

[CE021, CE022, CE023, CE024, CE025, CE033]
Chapter 06

06客户

6.1 客户分层与买方地图

已抓取来源指向一个多层客户基础,而不是单一切入口。最高层,Decart 通过 DOS 和合作分发,向云服务商、AI 实验室和超大规模买家出售基础设施和模型容量。公司发布的 $300 million 融资材料和 CTech 报道都把 Amazon 描述为战略客户,同时称 Decart 已从大型云服务商和 AI 实验室合同中产生收入。第二层是商业、媒体、流媒体和广告等企业应用买方,Lucy 支持虚拟试穿、直播效果和动态内容工作流。第三层是使用 Oasis 3 做 camera-first 模拟的 physical-AI 开发者和 AV 团队。企业销售动作之下,还有自助开发者社区,由公开文档、定价、SDK、示例和 GitHub 仓库支撑。这个组合有吸引力,因为它分散了产品界面;但公开记录仍未说明哪一层贡献了最多当前收入。[CU001, CU002, CU003, CU004, CU011, CU012]

客户分层表
细分买方 / 用户 / 付费方用例规模收入 / 战略价值缺口
云服务商和超大规模平台基础设施买方、平台团队、模型团队授权 DOS、加速推理,并分发 Decart 模型公司称与几家大型服务商签有合同,但大多未点名ACV 和战略分发杠杆可能很高公开来源未显示客户数量、合同期限或按服务商拆分的收入占比
AI 实验室和模型构建者研究和推理团队用 DOS 和实时模型基础设施降低算力成本与延迟公司和 CTech 来源称其为活跃合同客户群支撑 Decart 自有模型之外的基础设施护城河论点未披露具名 AI 实验室 logo 或续约数据
企业商业、媒体、流媒体和广告买方产品、增长和内容团队虚拟试穿、直播特效、动态视频、社交体验公开文档展示清晰工作流,但客户 logo 大多未具名显示 Lucy 表面具有广泛适用性未公开按垂直行业拆分的结果和留存
物理 AI 和 AV 开发者仿真、机器人和自动驾驶团队通过 Oasis 3 做提示驱动的多视角驾驶和世界模型仿真公开 API 降低了实验进入门槛可把 TAM 从媒体扩展到训练和评估物理 AI 买方证明大多仍是愿景式且与 Amazon 相关,而非广泛披露的客户名单
自助开发者和开源社区开发者、爱好者和产品构建者API 实验、SDK 集成、XR 演示和试穿原型TechCrunch 称开发者超过 100,000;GitHub 暴露了大量入门资产低摩擦采用管线,有助于放大合作伙伴传播付费账户转化和重复消费未披露

细分合并了具名证据和群组级证明;公开收入拆分缺失,是核心尽调限制。

[CU001, CU003, CU004, CU009, CU011, CU012]
客户增长与采用轨迹表
指标数值日期来源置信度含义缺失分母
公开开发者社区规模超过 100,000 名开发者2026-06-10TechCrunchAPI 分发的强顶部漏斗信号未披露付费占比、MAU 或活跃构建者定义
已披露具名战略客户1 个(Amazon)2026-05 至 2026-06Decart 公开发布 + CTech + JNS证实至少有一个大型企业锚点未披露更广泛的具名客户名单
生产就绪试穿示例数6 个示例2026-06-15电商试穿指南显示公司为商业用例做了具体上手投入这些示例或绑定商户没有使用指标
公开 API 标价Oasis 3 Preview 每秒 0.02 USD2026-06-15定价页 + TechCrunch + Startup Fortune支持自助实验和开发者获客未披露企业折扣或毛利率
已披露公开客户留存指标02026-06-15已抓取公开来源复核凸显采用证明比持久性证明更宽NRR、GRR、流失或续约数据均未公开

行内混合直接指标和披露质量信号,因为公开商业报告仍然稀疏。

[CU001, CU009, CU011, CU028, CU029, CU030]
FU001: 客户旅程地图

公开资料显示,客户旅程大致从发现开始,经由自助原型验证进入企业部署,再走向合作伙伴主导的扩张。

[CU002, CU006, CU011, CU013, CU039]

6.2 采用信号与自助入门

Decart 的获客策略明显依赖开发者优先的入口。定价页给出自助费率,试穿 walkthrough 展示商家如何把实时试穿接入商品页,GitHub 组织公开 SDK 和参考实现,合作伙伴报道反复称 Bedrock 分发降低了 AWS 原生团队的集成摩擦。这一点重要,因为公司试图把 realtime AI 从定制化企业销售,转成可接入的 API 动作。TechCrunch 报道称社区已有超过 100,000 名开发者,许多人在电商和直播场景基于 Lucy 构建;公开 try-on repo 数量则提供了更小但更具体的入门努力证据。不过,公开界面仍没有真正披露漏斗:没有注册开发者到付费账户的转化率,没有使用 try-on 的生产商家名单,也没有公开 churn 或 renewal 指标来证明开发者基础能带来持久变现。[CU005, CU006, CU007, CU008, CU009, CU010]

留存、重复使用与满意度表
指标数值细分置信度尽调问题
开发者社区规模100,000+ 公开社区数字自助开发者要求提供从注册到付费的转化、WAU/MAU,以及按产品家族拆分的群组留存
已披露具名生产客户数1 个具名战略客户企业 / 战略要求提供前十大客户名单、合同类型、年化支出和部署阶段
NRR / GRR / 流失null企业 / 战略要求提供按产品线拆分的 NRR、GRR、logo 流失和总流失
合同期限 / 续约披露null企业 / 战略要求提供平均合同期限、续约率,以及 AWS 关联账户与直销账户之间的扩张行为

Null 值表示公开未披露,而不是留存为零;本表衡量披露质量,不声称内部不存在这些指标。

[CU009, CU028, CU029, CU040]
FU002: 公开采用信号漏斗

公开记录在漏斗顶部最强,在留存端最弱。

[CU007, CU009, CU028, CU029]

6.3 公开客户证明与生态证据

最强的点名证据是 Amazon。多个抓取来源——Decart 自有发布、CTech 和 JNS——都称 Amazon 作为战略客户加入 2026 年融资和合作生态。这个信号比普通集成更强,因为它在公司材料和独立报道中反复出现,并与具体的 Bedrock 或 Trainium 分发语言相连。Amazon 之外,证据基础更多停留在群体层面而非 logo 层面。公司和第三方来源描述了来自大型云服务商、AI 实验室、hyperscaler 的收入,以及商业、流媒体、社交、游戏和 physical-AI 场景部署,但没有发布可比的生产客户名单。开发者侧证明要丰富得多:GitHub 组织、Python SDK、Quest XR 项目、try-on 指南和 AWS 合作伙伴文章,都展示了用户如何实际采用产品。这个分裂对尽调很关键:它说明 top-of-funnel 和生态可信度强,但企业广度的公开证明比头部融资轮暗示的更薄。[CU001, CU002, CU003, CU012, CU013, CU019]

具名客户证明表
客户细分部署 / 用例生产 vs 试点结果限制
Amazon / AWS超大规模平台 / 战略平台客户战略客户关系,加上面向实时模型的 Bedrock / Trainium 上市路径战略客户状态已公开披露;暗示已有生产分发最强的具名证明,也是进入大型企业账户最清晰的路径支出范围、采购产品和收入贡献未披露
使用 Decart 试穿模式的电商商户(未具名)企业商业群组面向产品页、数字镜和造型体验的实时虚拟试穿面向生产的示例和集成指南已公开证明 Lucy 瞄准实用商户工作流,而不只是演示商户名称、转化提升或重复使用统计均未公开
开发者和 XR 构建者群组自助 / 社区群组Python SDK、试穿示例、Quest XR 应用、RL 示例和 Discord 社区实时社区和仓库表面已公开;变现状态混杂显示活跃的开发者获客引擎和真实实现资产这是群组证明,不是具名企业 logo 名单

覆盖范围有意不完整,因为抓取到的公开记录只包含一个明确具名战略客户,以及更广泛但未具名的企业或开发者群组。

[CU001, CU006, CU007, CU013, CU016, CU019]
FU003: 客户证明矩阵

Amazon 的证明质量最强,开发者群体居中,广泛未具名企业类别最弱。

[CU001, CU006, CU019, CU027, CU028, CU038]

6.4 留存、集中度与不利披露缺口

主要客户风险不是表层活动不足,而是公开披露太窄。Amazon 是抓取来源中唯一明确点名的战略客户,其余商业叙事依赖宽泛但未点名的类别,如云服务商、AI 实验室、hyperscaler、商业部署和开发者。这提出了集中度问题:如果近期收入中有相当比例依赖少数基础设施客户,或依赖 Amazon 相关分发,下行可能比公司多元化营销叙事更陡。另一个缺口是耐久性。所有抓取来源都没有发布 NRR、GRR、churn、合同期限,甚至没有准确活跃客户数。公开定价和示例 repo 让入门路径更清楚,但不能证明重复消费或企业标准化。Startup Fortune 是最明确的怀疑来源:它把 Oasis 3 描述为在更大 incumbents 内部化能力之前出租模拟能力的一次押注,并指出当前模态能力相较更成熟内部技术栈仍有限。[CU025, CU026, CU028, CU029, CU030, CU031]

扩张与集中度风险表
扩张驱动集中度风险影响尽调路径
Bedrock 和 AWS 渠道分发Amazon 可能贡献过大的战略杠杆或客户访问可能加速增长,但压缩谈判杠杆和定价独立性要求按直销、AWS 渠道销售和头部客户集中度拆分收入
自助开发者采用庞大注册社区可能难以转化为持久付费使用可能制造顶部漏斗热度,却没有同等留存经济性要求按用例提供从注册到付费使用的群组转化和重复消费
物理 AI 仿真楔子大型既有玩家可以内化仿真技术栈,或要求更丰富模态可能把 Decart 限制在实验预算,而非核心技术栈标准化要求提供具名试点、续约证据,以及 RGB 摄像头输出之外的模态路线图
宽泛垂直行业营销叙事公开具名 logo 披露滞后于公司声称覆盖的行业广度可能说明少数基础设施买方占主导,长尾垂直采用仍早要求提供商业、流媒体、游戏、机器人和 AI 实验室的具名引用客户

风险来自可见上手资产与公开披露中更薄的收入集中度和留存信息之间的落差。

[CU025, CU026, CU028, CU029, CU031, CU038]
FU004: 公开留存披露代理群体

单元格显示各群体可获得的公开留存披露占比,不是实际客户留存率。

0 表示截至 2026-06-15,在已抓取来源中没有找到该群体的公开留存披露;这些是披露代理指标,不是客户行为测量。

[CU029, CU040]
Chapter 07

07风险

7.1 风险概览与严重性框架

Decart 面临多层风险堆栈,相比 $4 billion 估值和早期收入阶段,风险偏高。公司处在三个高风险监管区的交叉点:生成式 AI 版权法(训练数据责任)、EU AI Act(实时 AI 系统),以及 2026 年加速推进的美国州级 AI 治理框架。监管悬顶之外,公司的核心技术——自回归世界模型——存在有记录的可靠性限制,而且这些限制是架构性的,不是表面问题:TechCrunch 对 Oasis 3 的独立评测(2026 年 6 月)发现,生成几分钟内就出现物理一致性失败和上下文窗口退化,CEO 也承认这些是持续研究问题。运营上,Decart 集中使用 AWS Trainium3 基础设施,并依赖 Amazon Bedrock 做 go-to-market,基础设施层存在单点卡脖子模式。竞争上,Google DeepMind(Genie 3)、Waymo(基于 Genie 3 的内部世界模型)、Runway($5.3B、$315M Series E、2026 年 2 月)和 Luma AI($900M Series C)都在推进,资源显著更多,或战略位置更可防守。风险热力图(FR001)按可能性和报告缓释后的剩余影响绘制这些主题。最集中的风险——版权责任和竞争替代——落在高可能性、高影响象限。 [CR001, CR002, CR003, CR004, CR005, CR006]

FR001: 风险热力图

风险热力图按发生可能性(x 轴)和对投资论点的剩余影响(y 轴)绘制 Decart 的关键投资风险。每个单元格列出该象限的主要风险主题。

[CR001, CR007, CR015, CR022]

7.2 监管、法律与 IP 风险

训练数据版权是 Decart 最结构性模糊的法律风险。EU Parliament Legal Affairs Committee 在 2026 年 2 月以 17-3 票通过报告「Copyright and Generative AI — Opportunities and Challenges」,呼吁建立可反驳推定:任何投放 EU 市场的 GenAI 模型,除非满足完整透明度义务,否则推定使用了受版权保护作品训练。该推定一旦写入 EU 法,将把举证责任倒置给 Decart;在正式框架落地前,报告还提出对既往使用征收 5–7% flat-rate turnover fee。另一路,美国 Copyright Office 仍在推进 AI rulemaking(AI Report 第 3 部分),制定 AI 训练许可框架指引;办公室尚未发布有约束力规则,但其指引会影响诉讼姿态和未来法定风险。Osborne Clarke 2026 年 2 月法律分析称,EU 局面是一个「watershed moment」:现有 opt-out 系统不足,新框架很可能约束服务 EU 市场的非 EU AI 提供商(market-of-destination principle)。2025 年 Trump Executive Order on AI(2025 年 12 月)显示联邦层面有意预先排除州级 AI 监管,但风险并未消失——Gunder LLP 的 2026 AI laws update 指出,Colorado、California 等州已制定综合 AI 治理法规,并继续运行,也已影响供应商合同。Decart 的 AUP(2026 年 2 月更新)明确承认 EU AI Act、DSA 和 DMCA 等治理框架,并禁止军事 / 武器、实时生物识别监控和社会评分场景。截至 2026 年 6 月,未发现针对 Decart 的公开 IP 诉讼,但没有诉讼并不能证明训练数据许可合规。尽调需要完整训练数据来源审计,并取得任何受版权保护来源的许可协议。 [CR007, CR008, CR009, CR010, CR011, CR012]

监管 / 法律风险登记表
规则 / 许可 / 案件司法辖区状态可能性严重性缓释剩余暴露尽调路径
EU Parliament Copyright + GenAI rebuttable presumption(拟议)欧盟拟议中;委员会 17-3 投票通过(2026 年 2 月);尚未立法高(立法动能)严重(潜在 5–7% 全球营业额费用 + 诉讼举证倒置)AUP 承认 DSA/DMCA;未披露训练数据透明度高 — 若立法通过将产生追溯责任;训练数据来源未知要求完整训练数据来源审计;提供任何受版权保护来源的许可协议
EU AI Act (AIA) 实时 AI 义务欧盟2025–2027 年分阶段执行;AUP 提及 AIA 合规高(已生效)高(实时 AI 系统面临重大合规义务)AUP 明确提及 EU AI Act;具体技术合规尚未独立审计中 — 合规框架已有部分证据,但未经审计获取独立 AIA 合规审计;确认已满足透明度和人工监督义务
美国 Copyright Office AI 规则制定(Part 3)美国规则制定进行中;预计 2026 年发布指引;尚无具约束力规则中(规则制定进行中)高(可能追溯性地对 AI 训练数据施加授权要求)未发现;未公开披露训练数据授权情况高 — 美国指引可能固化未授权历史训练的责任审查 Copyright Office 第 3 部分预发布指引;评估训练数据风险敞口
美国州级 AI 治理(Colorado、California 综合 AI 法规)美国(多州)已颁布;2026 年开始执行高(法规已生效)中(AI 供应商需承担合规义务;合同规范在变化)未发现州级专项合规文件中 — 供应商合同义务;客户侧合规要求可能向 Decart 传导与律师确认州级合规姿态;审查企业合同条款
美国 AI 行政令(2025 年 12 月)— 试图联邦优先管辖美国2025 年 12 月签署;落地不确定;各州在抵制低(预计会有法律挑战)低(短期内无法消除州级风险)无需行动;仅监控低 — 不会消除短期合规要求监控 DOJ AI Litigation Task Force 和机构指引
IP / 商业秘密 — 训练数据并非来自授权来源全球截至 2026 年 6 月未发现诉讼;风险仍潜伏中(行业诉讼在增加)高(集体诉讼或监管执法可能构成生死风险)未见授权协议证据;AUP 禁止用户下游侵犯 IP高 — 未发现诉讼不等于合规要求完整训练数据清单,逐个数据集列明来源、授权状态和 opt-out 合规

行按严重程度排序(Critical > High > Medium)。所有监管状态反映截至 2026 年 6 月的情况。 可能性评级依据公开监管和法律来源估算;尚未发现专门针对 Decart 的官方政府指引。尽调路径反映 AI 原生公司的标准 IP 尽调。

[CR007, CR008, CR009, CR010, CR011, CR014]

7.3 运营、模型质量与安全风险

Decart 的运营风险主要来自自回归世界模型的根本架构限制,以及对 AWS 基础设施的深度依赖。模型质量方面:TechCrunch 2026 年 6 月对 Oasis 3 的独立测试记录了三类故障——(1)上下文窗口填满后,几分钟内主题一致性退化;(2)物理一致性失败,车辆彼此穿透,因为模型缺少真正碰撞模拟;(3)控制响应问题,无法精确导航。CEO Leitersdorf 承认三者都是持续研究问题,并把上下文退化归因于:30fps 视频生成每帧消耗约 8,000 tokens,在每秒数十万 tokens 的速度下会很快填满上下文窗口。公司在研究更长上下文和记忆压缩技术。对 AV 客户而言,物理失败不是外观瑕疵——合成训练数据可能包含物理上无效的场景,进而劣化下游模型行为。这给 Decart 目标 AV 市场带来直接训练数据质量风险。基础设施方面:Decart 的 Lucy 模型运行在 AWS Trainium3 上,并通过 Amazon Bedrock 分发,形成集中单云依赖。持续 AWS 故障会中断全部推理运营。Trainium3 硬件刚发布,广泛供应有限,规模化阶段存在硬件采购风险。安全方面:作为一个根据客户提示生成实时视频的平台,Decart 面临 prompt injection、绕过 AUP 生成禁止内容(deepfake、CSAM、武器合成)的 jailbreak 风险,以及 API key 泄露风险。AUP(2026 年 2 月)覆盖了禁止用途,但未发现内容审核有效性的独立审计。 [CR015, CR016, CR017, CR018, CR019, CR020]

运营 / 质量 / 安全风险登记表
失效模式发生可能性严重程度缓解成熟度剩余敞口未解决缺口
Oasis 3 物理一致性失效(车辆互相穿透;没有碰撞仿真)高(独立测试和 CEO 承认均已确认)高(AV 客户可能拒绝违反物理规律的合成里程;若用于安全关键训练,可能带来责任)低 — CEO 承认这是「现在正在攻克的重大研究问题」高 — 尚未解决;AV 客户验收标准未知AV 客户是否在合同中为合成训练数据的物理错误向 Decart 提供赔偿保护
Oasis 3 生成数分钟后上下文窗口退化、主题连贯性丢失高(2026 年 6 月独立测试已确认)中(限制实际会话长度;影响 AV 长周期仿真用例)低 — 研究推进中(更长上下文、记忆压缩)中 — 目前实际会话长度不足以支撑完整路线 AV 仿真发布可接受物理精度和连贯性的基准;预期模型版本时间表
AWS Trainium3 单一基础设施依赖中(AWS 可靠性高;但 Trainium3 是新发布硬件)高(若 AWS 宕机或 Trainium3 分配受限,服务会全面中断)低 — 未发现公开多云备援高 — 没有记录在案的故障切换或多云架构确认 SLA 承诺和故障切换架构;多云路线图
Amazon Bedrock 分发渠道集中高(若 Amazon 降低 Bedrock 上第三方模型优先级,分发会坍塌)低 — 依赖属于结构性问题,尚未缓解高 — 未宣布同等规模的替代分发渠道分发协议是否包含独家或排除条款
内容审核有效性(deepfake、CSAM、通过 API 合成武器)高(违反 AUP 可能触发监管行动;CSAM 尤其带有刑事责任)中 — 已发布 AUP(2026 年 2 月)并列明具体禁令;执行机制未知中 — AUP 已存在,但技术执行没有独立审计API 内容审核独立审计;事件响应流程披露
视频生成 API 的提示注入 / 越狱漏洞中(可能绕过 AUP;高曝光滥用会带来声誉风险)低 — 未发现红队测试或安全披露中 — 标准 API 安全风险;严重程度取决于审核韧性红队 / 对抗测试结果;是否存在安全披露计划

按严重程度排序。Oasis 3 的技术限制来自 TechCrunch 2026 年 6 月独立评测和 CEO 承认。基础设施状态来自公开公告以及 AWS/Decart 合作披露。

[CR015, CR016, CR017, CR018, CR019]
FR002: 风险传导图

有向无环图展示根因触发事件如何级联为 Decart 投资的下游破题结果。

[CR007, CR015, CR022, CR033]

7.4 合作伙伴、依赖与客户集中度风险

Decart 的 go-to-market 和基础设施策略形成了多层集中度风险。最实质的单一依赖是 AWS/Amazon:Decart 的 Lucy 模型运行在 AWS Trainium3 上,模型通过 Amazon Bedrock 分发,Amazon 自身又被列为战略客户——Amazon 同时是云服务商、分发渠道和收入来源,形成三重集中。如果 Amazon 开发竞争性内部世界模型能力(考虑到其 AWS AI 基础设施投入,这并非不可能),或降低 Bedrock 对第三方模型的 shelf-space,Decart 的分发渠道会和基础设施一起塌陷。类似地,Nvidia 既是 Decart 的主要替代算力供应方,也是股权投资者(Nvidia 参与了 $100M Series B 和 $300M Series C)。一旦 Nvidia 关系出现重大争议或竞争性转向,Decart 的硬件路线图和 cap table 观感会同时受影响。投资者 / 客户重叠会放大客户集中度:Toyota Ventures、Adobe Ventures、eBay Ventures 都是股权投资者,其母公司也被列为战略客户。这一安排对信号质量有利,但会制造不利激励动态——持股客户如果从平台流失,会向市场释放尤其负面的信号。Startupfortune(2026 年 6 月)指出,Decart 的真实市场是 Waymo/Tesla/Nvidia 层级以下「数十个 AV 项目、机器人实验室和无人机创业公司」——这意味着可变现客户基础更小、资本实力更弱,战略黏性也低于战略投资者 / 客户。开发者社区(100,000+ 注册)规模很大,但 $0.02/second 的 API 收入高度依赖使用量,典型使用模式下每名开发者收入仍不确定。 [CR022, CR023, CR024, CR025, CR026, CR027]

合作伙伴 / 依赖风险登记表
依赖对手方角色集中度失效情景严重程度缓解剩余敞口
AWS Trainium3 算力 + Amazon Bedrock 分发Amazon / AWS主要算力基础设施、分发渠道和战略客户关键 — 算力、分发和核心收入都依赖单一供应商AWS 开发竞品世界模型能力;Bedrock 降低 Decart 优先级;削减硬件分配关键Nvidia 替代算力;AWS Trainium3 合作带来早期访问优先级高 — 没有记录在案的多云或多分发备援
Nvidia GPU / 硬件Nvidia芯片供应商,同时是股权投资者(两轮均参与)高 — Nvidia 既是硬件来源,也是投资者Nvidia 开发竞品世界模型(Cosmos);降低 Decart 硬件优先级Decart 与 AWS Trainium3 的关系提供部分硬件多元化中 — Trainium3 是替代方案,但同样依赖 AWS
Toyota、Adobe、eBay(投资者 / 客户重叠)战略投资方:Toyota Ventures / Adobe Ventures / eBay Ventures战略投资者,同时是付费客户高 — 持股客户流失会释放格外负面的信号持股期间客户流失;Toyota/Adobe 内部自建竞品股权绑定提高转换成本;战略路线图整合中 — 激励一致真实存在,但不能阻止客户离开
开发者社区(100,000+)和 API 分发开发者生态增长引擎和布道渠道中 — 依赖规模;API 收入取决于开发者使用量开发者社区迁移到竞品 API(Google Genie 3、Runway、Luma)$0.02/秒 定价具备竞争力;开发者优先的市场进入策略中 — 开发者锁定需要模型持续迭代
Radical Ventures($300M 轮领投方)Radical Ventures领投方;影响董事会高 — 最近一轮融资的领投方需要下调估值轮或桥接融资;Radical 放弃下一轮多家投资者(Sequoia、Benchmark)提供替代董事会视角中 — 若里程碑未达成,Radical 是否参与下一轮不确定

按严重程度排序。投资者 / 客户重叠关系来自 Decart 2026 年 5 月融资公告和 TechCrunch 报道。AWS 合作来自 AI News 和 Decart 官方渠道。

[CR022, CR023, CR024, CR025, CR026]
FR003: 依赖图

Decart 关键外部依赖的有向无环图。节点代表系统、供应商或实体;边代表依赖关系(上游 → 下游故障路径)。

[CR022, CR023, CR024]

7.5 人员、执行与融资风险

Decart 的执行风险集中在 key-person 依赖、地缘政治暴露,以及 $4 billion 估值与公开收入指标之间的融资落差。Dean Leitersdorf(CEO,27 岁)一直是从 seed 到 Series C 每一轮的主要公众面孔、融资人和技术叙事驱动者。他若离开,会在关键规模化阶段同时造成融资、技术可信度和媒体叙事缺口。除两位联合创始人外,公司的领导团队公开记录不深;Dr. Kfir Aberman(前 Snap 和 Google、DreamBooth 共同创造者)是公开点名的最资深招聘,担任 SF R&D center 负责人。任何联合创始人或 Aberman 离开,都会带来执行风险。以色列运营总部带来地缘风险:Israel-Gaza 冲突或更广泛区域不稳定若持续升级,会影响 Tel Aviv 工程团队运营、全球差旅签证,以及与风险敏感型企业客户的关系。若干大型企业(尤其 defense-adjacent 行业)有采购政策,限制采购活跃冲突地区供应商。融资风险:Decart 已融资 $450M+,但声称 lifetime burn「less than $100M」(CEO,2026 年 6 月)。即便属实,随着公司扩展 AV/physical-AI 模拟基础设施、员工数从 60+ 扩张,并搭建企业销售能力,burn rate 会显著加速。在 $4 billion 估值下,下一轮需要达到 $5–6 billion 或更高,才能避免 flat/down round;这要求公司拿出可证明的收入 traction,而公开证据尚未支持。 [CR028, CR029, CR030, CR031, CR032]

人员 / 执行风险登记表
角色 / 职能依赖或缺口发生可能性严重程度缓解尽调路径
CEO Dean Leitersdorf(联合创始人,27 岁)主要融资人、公众叙事和技术愿景;离任会留下多维缺口低(自愿离任;年轻、使命驱动的创始人)关键 — 很可能触发估值下调风险,并冲击客户信心股权归属、董事会监督;SF R&D 中心提供组织纵深确认归属时间表;评估接班计划和董事会深度
CPO Moshe Shalev(联合创始人)核心技术产品领导;公众曝光低于 CEO,但属于基础角色高 — 联合创始人离任释放团队裂解信号;CPO 职能难以外部替补除股权归属外未发现缓解评估产品领导层深度;确认 CPO 股权归属和留任
Dr. Kfir Aberman(SF R&D 负责人,曾任 Snap/Google)公开点名的最高级外部招聘;DreamBooth 共同创造者;流失会削弱美国研究叙事中(招聘市场竞争激烈;Google/Snap 招聘方活跃)中 — SF R&D 中心连续性存在风险;研究人才信号受损股权薪酬;在 SF R&D 团队扩张中配置股权确认 Aberman 股权份额和雇佣期限;评估 SF 团队深度
以色列总部地缘政治风险Tel Aviv 工程团队;全球差旅签证约束;企业采购政策中(截至 2026 年 6 月,地区不稳定仍在持续)中 — 持续升级可能扰乱运营;部分企业会排除活跃冲突地区供应商SF 和 NY 办公室提供地域多元化;分布式团队模式评估企业客户对以色列总部供应商的采购政策;业务连续性计划
AI 研究人才竞争模型、DOS stack 和 Oasis 3 需要前沿 ML 研究员;人才与 Google/Anthropic/OpenAI 竞争高 — 人才流向超大规模云厂商可能拖慢模型开发速度Unit 8200 网络提供差异化招聘;有竞争力的股权审查员工人数增长轨迹;确认除联合创始人外的研究团队深度

按严重程度排序。领导层信息来自 Decart 融资公告、Calcalis Tech 和 Ynet News 报道。地缘政治风险评估基于以色列科技公司运营环境的公开信息。

[CR028, CR029, CR030, CR031]

7.6 缓释、终止标准与尽调问题

Decart 已对其主要风险采取了有意义但不完整的缓释措施。监管:AUP(2026 年 2 月)明确引用 EU AI Act、DSA 和 DMCA;这显示了监管意识,但不能替代训练数据透明度义务或许可框架。技术:CEO 已公开承认物理一致性和上下文窗口限制,并称有主动研究项目;下一代模型预计会解决视频播种的世界生成,可能改善一致性。基础设施:AWS Trainium3 交易提供性能效率,但加重 AWS lock-in;未发现公开 multi-cloud fallback。人员:SF R&D center 招聘 Dr. Kfir Aberman 带来了地域和人才多元化,但没有解决联合创始人集中度。缓释和 kill-criteria 表(TR005)把每个主要风险映射到可监测触发器和投资 thesis 含义。按紧迫度排序的尽调问题:(1)训练数据许可协议——EU 市场准入最实质的未解项;(2)独立安全和内容审核审计;(3)AV 客户的物理精度验收标准——Decart 的物理模拟是否满足任何 AV 公司的 safety-case 要求;(4)按产品线和客户拆分的收入构成——DOS 授权 vs. Lucy API vs. Oasis 3 enterprise;(5)使用 Lucy 的电商和游戏客户留存指标。 [CR033, CR034, CR035, CR036, CR037]

缓解措施与否决标准表
风险可监控触发项阈值 / 事件行动含义
版权 / 训练数据责任EU Parliament 通过可反驳推定;US Copyright Office 发布有约束力指引EU 法规生效并征收 >2% 营业额费用,或美国法院判定可比 AI 公司败诉重新评估:核查 Decart 训练数据来源;考虑价值风险调整
物理一致性故障阻塞 AV 客户AV 客户公开声明提及物理故障;12 个月内没有 AV 企业合同到 2027 年 Q1 仍无具名 AV 企业客户物理 AI 投资论点被打破;仅按电商 / 游戏重新评估 TAM
AWS / Bedrock 集中风险兑现Amazon 在 Bedrock 上推出竞品世界模型;Decart 失去 Bedrock 上架Amazon 到 2026 年底在 Bedrock 上推出内部世界模型 API分发护城河投资论点被打破;评估 Decart 能否仅靠开放 API 竞争
模型性能相对 Google DeepMind / Runway 停滞第三方基准显示 Genie 3 或 Runway 世界模型以有竞争力价格 / 质量追平2026 年 Q3 前公开基准显示更低价位实现同等性能重新评估技术护城河;提高 DOS 可防守性尽调紧迫性
监管执法行动EU 或美国机构因版权或 AIA 违规对 Decart 采取执法行动任何正式执法行动;EUIPO 投诉;权利人集体诉讼立案重大不利事件;重新评估 EU 市场准入和法律成本风险
估值下调轮或融资失败Series D 前需要桥接融资;下一轮估值下调下一轮估值低于 $3.5B,或收入同比增长低于 100%负面信号;重新评估财务模型和现金消耗轨迹

否决标准是会打破投资论点的事件;监控项是需要跟踪的先行指标。阈值是基于公开证据作出的投资判断估计,并非公司官方指引。

[CR033, CR034, CR035, CR036, CR037]

7.7 图表

Chapter 08

08估值

8.1 融资历史与资本结构

Decart 自 2022 年成立以来已完成多轮融资,累计超过 $450 million,估值轨迹让它跻身以色列 AI 初创公司最快估值上升案例。最新一轮是 2026 年 5 月完成的 $300 million Series C,投后估值约 $4 billion。该轮的战略构成很突出:领投方包括同时也是付费客户的科技公司(Toyota、Adobe、eBay Ventures),形成一种不常见的绑定结构,把资本和商业采用捆在一起。Radical Ventures 领投,Nvidia、Sequoia Capital 和 Benchmark 参投。Series C 之前,Decart 在 2025 年 8 月以 $3.1 billion 估值完成 $100 million Series B,意味着约九个月内估值上升 29%。累计稀释和 pre-money 所有权结构尚未公开。对 Decart 当前商业阶段而言,$450 million+ 现金储备相当充足,即使计算开支上升,也意味着多年 runway;但 burn rate 数据无法验证。战略投资者基础增加了治理复杂度:既是客户又是投资者的一方,可能在产品路线图、企业定价和潜在退出路径上存在冲突激励。 [CV001, CV002, CV003, CV004, CV005]

Decart 融资历史
轮次日期融资金额投后估值领投方累计融资
种子轮2022–2023(估算)~$30M(估算)~$150M(估算)未披露~$30M
Series A/B(早期)2024(估算)~$20M(估算)~$300M(估算)未披露~$50M
Series BAug 2025$100M~$3.1BSequoia、Benchmark、Aleph VC~$150M
Series CMay 2026$300M~$4.0B主要投资人:Radical Ventures、Nvidia、Sequoia、Benchmark、eBay/Adobe/Toyota Ventures~$450M+

种子轮金额和估值根据媒体报道及投资者公告估算;公司披露未确认。其他数字均来自媒体报道。Series C 投后估值来自 Calcalist 报道和 Decart 官方公告。累计融资额为每轮当时的累计数字。

[CV001, CV002, CV003]
FV001: 推荐逻辑链

流程图展示从市场、技术、财务输入到最终 TRACK 建议的证据链,并带有高风险和估值偏高的框架。每个节点代表一个关键证据点或结论,箭头表示逻辑依赖。

[CV017, CV034, CV035, CV027, CV028, CV038]

8.2 可比公司分析

Decart 最接近的公开可比对象,是 AI 视频、世界建模和 AI 游戏基础设施领域的私募市场同行。Runway AI 是功能上最接近的可比公司,2026 年 2 月以 $5.3 billion 估值融资 $315 million,并公开表示正从创意视频编辑扩展到世界模型,因此是未来直接竞争者,估值比 Decart 高 $1.3 billion。World Labs 在 2026 年 2 月融资 $200 million,Autodesk 锚定投资,目标是专业 3D 工作流,隐含估值约 $5 billion;它范围比 Decart 更窄,但在 $25 billion 3D 设计软件市场中商业垂直更清晰。Luma AI 在 2025 年 11 月从 HUMAIN(Saudi Arabia 国家 AI 基金)获得 $900 million Series C,估值约 $900 million——这是一个总融资额(该轮时 $450 million)大致等于投后估值的案例,说明投资者对视频生成子领域相当谨慎。Inworld AI 估值超过 $500 million、融资 $100 million,是主要 AI-native 游戏基础设施同行;其 NPC 角色引擎聚焦更窄,但具有互补性。同行组平均投后估值约 $2.9 billion,Decart 的 $4 billion 比同行均值溢价 38%。如果实时世界模拟证明比后期视频生成是更高价值架构,该溢价可以成立;但在 Decart 商业指标不透明的情况下,目前仍主要由假设支撑。 [CV006, CV007, CV008, CV009, CV010, CV011]

可比估值表
公司最近轮次融资金额投后估值主要重点来源
DecartSeries C(May 2026)$300M~$4.0B世界模型、游戏、AV 仿真Calcalist/官方
RunwaySeries E(Feb 2026)$315M~$5.3BAI 视频生成和创意工作流TechCrunch
World Labs战略轮(Feb 2026)$200M~$5.0B(估算)面向设计工作流的 3D 世界模型TechCrunch/WorldLabs
Luma AISeries C(Nov 2025)$900M~$900M+AI 视频生成CNBC/BusinessWire
Inworld AISeries B(2024)$50M~$500M面向游戏的 AI NPC 引擎GamesBeat

估值为已报道融资轮的投后估算。私营公司(Runway、World Labs、Luma、Inworld)未公开收入;Roblox 是唯一上市可比公司。重点列反映主要商业侧重。比较仅供背景参考;这些都是早期 AI 公司,财务披露很少,估值更多反映投资者情绪,而非基本面倍数。

[CV006, CV007, CV008, CV009, CV010]
FV002: ARR 假设下的估值敏感性

条形图展示在固定 20x 收入倍数下,Decart 的隐含估值如何随不同 2028E ARR 假设变化。图中呈现当前 $4B 价格与情景隐含价值之间的差距。所有 ARR 数字均由假设驱动;没有公开收入数据。

[CV020, CV021, CV022, CV025, CV039]

8.3 总可寻址市场

Decart 覆盖三个重叠市场,成熟度各不相同。BCG Global Gaming Survey 将 2025 年全球游戏市场规模估为约 $210 billion,这是近期切入口:嵌入游戏引擎的模拟基础设施可采用按 session 或 API call 收费模式,瞄准开发者工具预算的 3-5%;若 Decart 完成工具链集成,仅游戏模拟的 serviceable addressable market 就可能达到 $1-3 billion。BCG 特别把生成式 AI 识别为游戏行业的主要结构性变化,并预计 2025 年中约 20% 新 Steam 游戏披露使用 AI(为上一年两倍),游戏行业也正走出三年 post-pandemic 低迷。The Business Research Company 估算,2025 年游戏生成式 AI 市场为 $2.1 billion,到 2030 年以 30% CAGR 增长至 $10.5 billion——Decart 的实时世界模拟是该细分中的高端产品。Research and Markets 同样将 2025 年生成式 AI 游戏市场估为 $2.1 billion,并预计到 2030 年 CAGR 为 24%。第二个板块——自动驾驶模拟——可能是最大价值池,但 Decart 已记录的物理一致性限制让技术风险最高。第三个板块——电商和零售试穿——已通过 Lucy 模型产生 API 收入,是近期 ARR 路径最清晰的部分,尽管 TAM 更小(AI 虚拟试穿约 $500M-1B)。McKinsey State of AI 报告确认,大型组织企业 AI 采用率已超过 72%,验证了企业需求背景;但 Decart 能分到多少仍未量化。Statista 将 2025 年全球移动游戏市场规模估为约 $92 billion,这是 Decart API 可直接触达的总游戏 TAM 子集。 [CV013, CV014, CV015, CV016, CV017, CV018]

TAM 测算 — 世界模型与 AI 游戏市场
市场赛道2025E ($B)2028E ($B)CAGR来源
全球游戏市场(总量)~$210~$260~7%BCG/Newzoo
移动游戏(子市场)~$92~$115~7.5%Statista
游戏中的生成式 AI~$2.1~$6.5~30%Business Research Co. / Research and Markets 市场研究口径
AI 虚拟试穿和电商~$0.8~$2.5~25%分析师估算
世界模型 / AV 仿真(新兴)<$1~$3–5N/A(早期)分析师估算(假设驱动)

市场规模估计来自外部分析师报告。世界模型仿真 TAM 和 AV 仿真子赛道为分析师构建估算,并非独立审计数据。2028E 和 2030E 预测为分析师预测,可能大幅修订。所有数字单位均为十亿美元。

[CV013, CV014, CV015, CV016, CV017, CV018]
FV003: 分情景估值回报区间

区间图展示熊市、基准、牛市情景下 Decart 的隐含股权估值,置信区间反映假设驱动预测的不确定性。当前 $4B 价格位于区间顶部。所有数字均由假设驱动;Decart 没有披露任何财务指标。

[CV020, CV021, CV022, CV026]

8.4 收入情景与财务建模

由于没有任何公开收入数据,估值建模只能靠假设情景构建。主要收入驱动因素包括:(A)电商和直播的 Lucy 模型 API 费用(当前 traction);(B)Oasis 3 游戏模拟 API 费用,定价 $0.02/second(2026 年 6 月推出);(C)潜在 AV 企业合同(尚未宣布,暂无公开证据)。CBInsights 确认 Decart 未公开披露 ARR 或财务指标。牛市情景下,Decart 将 100,000+ 开发者社区转化为付费客户,即便留存适中,也拿下 3-5 个企业游戏或 AV 合同,到 2026 年底达到 $30 million ARR、2028 年达到 $200 million;按 20x forward multiple,可支撑当前估值。基准情景下,采用更慢,Decart 到 2026 年底达到 $15 million ARR,2028 年达到 $80-100 million;按 20x multiple,隐含估值 $1.6-2.0 billion,相比当前 $4 billion 定价显著折价。熊市情景下,物理失败关闭 AV 市场,Runway 和 Google 竞争限制游戏采用;Decart 2028 年 ARR 仅 $25 million,按 20x 估值为 $500 million,较当前定价折价 87.5%。Roblox Corporation 的 SEC 10-K 文件提供了公开公司锚点:Roblox 2025 财年收入约 $3.9 billion,市值接近 $26 billion,隐含 price/revenue multiple 约 6-7x;这说明即便 Decart 达成牛市收入,若按成熟阶段倍数,要支撑 $4 billion 市值也需要异常出色。 [CV020, CV021, CV022, CV023, CV024, CV025]

Decart 收入情景(假设驱动)
情景2026E ARR2028E ARR关键假设隐含估值(20x 2028 ARR)
牛市$25–35M$180–220M游戏 SDK + AV 企业合同规模化;开发者到付费转换率 2–3%$3.6–4.4B
基准$10–15M$75–100MAPI 在游戏和电商中采用;无 AV 收入$1.5–2.0B
熊市$3–7M$20–30M物理限制卡住 AV;Runway 在游戏中替代;企业采用缓慢$400–600M

所有数字都是基于 API 定价、开发者社区规模和市场数据的假设驱动预测。Decart 未披露任何收入数据。这些只是用于说明的情景估计,不应解读为预测。熊 / 基准 / 牛市情景反映不同采用速度和市场成功结果。

[CV020, CV021, CV022, CV023]
估值倍数矩阵 — 当前与情景
估值依据收入估计或参照隐含估值相比当前 $4B
牛市情景(20x 2028E ARR)$200M ARR(2028E,假设驱动)$4.0B持平
基准情景(20x 2028E ARR)$88M ARR(2028E,假设驱动)$1.75B-56%
熊市情景(20x 2028E ARR)$25M ARR(2028E,假设驱动)$500M-87.5%
Roblox 可比(6x 收入,成熟倍数)$3.9B FY2025 收入(SEC 10-K)$23.4B(Roblox 实际)N/A — Roblox 已成熟;Decart 仍处于收入前阶段
私营 AI 视频同业平均未披露收入;同业平均估值 $2.9B$2.9B-27.5%

本表将 Decart 当前 $4B 估值,与不同情景和可比框架下的隐含价值进行对标。倍数扩张 / 收缩假设仅作方向判断。数字用于投资框架;Decart 没有经审计财务数据。Roblox 倍数来自 SEC 10-K 文件(FY2025)。

[CV024, CV025, CV026]

8.5 估值框架、风险因素与 Thesis 破裂点

当前 $4 billion 估值可拆成三部分:(1)已展示的技术 optionality——在当前延迟和保真度规格下独特的实时世界模型架构;(2)基础设施合作溢价——AWS Trainium3 生产交易和 Amazon Bedrock 分发,提供了多数同阶段初创公司没有的规模;(3)战略投资者溢价——客户兼投资者结构意味着早期采用概率较高。反面 thesis 同样实质:TechCrunch 2026 年 6 月独立评测记录了 Oasis 3 物理一致性失败,CEO 承认这是正在攻关的研究问题,说明核心产品限制不是表面瑕疵。CBInsights 因缺乏财务披露将 Decart 评为高风险。物理限制是 AV 垂直的关键 thesis-break 触发器,而 AV 可能占牛市估值的 30-50%。另外三个 thesis-break 触发器是:(A)到 2026 年 Q4 仍无法展示 enterprise-grade ARR;(B)EU 针对 Decart 训练数据发起重大版权执法行动;或(C)Runway 或 Google 在 12-18 个月内推出价格 / 质量有竞争力的实时世界模型。按当前可比估值,AWS 和 Nvidia 战略合作不足以解释相对同行的溢价:Runway 也有类似战略关系(Adobe 潜在收购候选),并披露了商业 traction。任何建议升级到「跟踪」以上之前,尽调会议必须拿到经审计收入、毛利率、burn rate 和客户集中度数据。 [CV027, CV028, CV029, CV030, CV031, CV032]

投资者尽调要求
类别具体要求理由
收入和预订2026 年上半年经审计 ARR、总预订额和季度增长率任何站得住脚的估值倍数都绕不开;目前公开披露为零
客户集中度前 5 大客户占 ARR 比例;按 cohort 划分的流失率战略投资方 / 客户(Toyota、Adobe、eBay)可能意味着集中度过高
毛利率和算力成本当前 API 定价下的毛利率;AWS Trainium3 成本占收入比例GPU 依赖风险;毛利被压缩后,当前定价可能难以持续
训练数据来源训练数据集完整清单,含许可状态和退出合规情况EU 版权可反驳推定会在来源不清时制造追溯责任
物理准确性路线图Oasis 3 物理准确性公开基准;达到 AV 级合规的时间表AV 垂直市场论点完全取决于物理准确性;当前限制已有记录
烧钱速度和跑道已融资 $450M+ 后的月度烧钱速度和现金跑道;资金用途拆分算力密集型基础设施意味着烧钱可能很高;必须验证跑道

尽调要求代表把投资建议从「跟踪」上调到「买入」所需验证的最低信息。截至 2026 年 6 月,所有项目均无法从公开来源获得。理由列解释每项为何对投资论点重要。

[CV034, CV035, CV036, CV037, CV038]
FV004: 投资 KPI — Decart 记分卡

IC 可用的 KPI 记分卡,从六个投资维度评估 Decart:市场、技术证明、竞争护城河、财务证据、风险画像和估值。评分是基于公开来源的证据评估。红 / 黄 / 绿依据截至 2026 年 6 月可获得的证据分配。

[CV011, CV033, CV035, CV036, CV027, CV021]

8.6 建议、尽调问题与披露限制

建议立场:跟踪。置信度:中。风险评级:高。估值立场:偏贵。Decart 的投资案例具有不对称性——技术差异化、AWS 合作和战略投资者质量确实优于同类,但 $4 billion 价格内嵌的执行假设,尚未被任何公开财务披露证明。建议跟踪而非直接放弃,是因为 TAM 足够大、技术 moat 也足够真实,过早「放弃」可能错过代际平台机会;但「买入」建议需要财务验证,目前拿不到。升级为「买入」的触发器是:(1)披露 ARR 达到或超过 $25 million,并有 YoY 增长;(2)至少一个公开 AV 企业合同,且披露 TCO;(3)发布带可衡量里程碑的物理一致性路线图。降级为「放弃」的触发器是:任何重大版权执法行动、确认物理失败阻断 AV 采用,或 Runway/Google 以价格平价推出世界模型。披露限制:本章所有财务预测都由假设驱动;截至 2026 年 6 月,Decart 未公开披露收入、ARR、bookings、burn rate、毛利率、客户数或留存指标。本章只给出情景区间,明确不能支撑具体投资价格建议。投资者应把 $4 billion 估值视为待验证假设,而非已经验证的市场价格。 [CV034, CV035, CV036, CV037, CV038, CV039]

8.7 图表

免责声明

本报告仅供参考。

证据索引

结论
编号陈述可信度来源
CO001 Decart describes itself as a vertically integrated AI research lab building real-time world models and the optimized infrastructure to run them. SO001, SO002
CO002 Decart's public materials anchor the company in Tel Aviv, Israel with additional engineering presence in San Francisco and New York. SO005, SO022
CO003 Decart sells API access to its realtime models through platform.decart.ai on a pay-as-you-go basis with no minimum spend. SO009, SO014
CO004 Decart's three product lines as of May 2026 are DOS (an optimization stack), Lucy (a real-time video world model), and Oasis (a world model for physical AI). SO002, SO004
CO005 Decart prices Lucy 2.1 realtime and Oasis 3 Preview at $0.02 per second of active generation through its documented pricing page. SO009
CO006 Decart's DOS stack runs across NVIDIA GPUs, Google TPUs, and Amazon Trainium silicon. SO002, SO020
CO007 Mirage, Decart's livestream diffusion model, was launched in 2025 as the company's second flagship product and predecessor to the Lucy family. SO005, SO025
CO008 Decart has documented Realtime Integration Paths and JavaScript, Python, Swift, and Android SDKs on its developer documentation. SO008, SO027
CO009 Decart's API platform documents Lucy VTON 3 for virtual try-on, Lucy Restyle 2 for realtime style transfer, and image and video models including Lucy Clip and Lucy Image 2. SO008, SO010
CO010 Decart's developer console at platform.decart.ai requires JavaScript-rendered access for sign-in and console functionality. SO014
CO011 Decart publishes API Terms of Service through Mintlify-hosted documentation at docs.platform.decart.ai with a separate Acceptable Use Policy and Data Processing Agreement. SO029
CO012 Decart's San Francisco R&D center opened in 2025 and is led by Dr. Kfir Aberman, formerly of Snap and Google. SO005
CO013 Decart was founded in late 2023 by Dean Leitersdorf and Moshe Shalev. SO004, SO021, SO022
CO014 Dean Leitersdorf is Decart's CEO and the company's primary public spokesperson in 2026. SO002, SO004, SO007
CO015 Moshe Shalev is Decart's Chief Product Officer and co-founder. SO004, SO021
CO016 Dean Leitersdorf earned three computer-science degrees and a PhD from the Technion before age 24. SO005, SO023
CO017 Moshe Shalev grew up in an ultra-Orthodox family in Bnei Brak and served in Unit 8200 alongside Leitersdorf. SO005, SO023
CO018 Both Decart co-founders are veterans of Israel's Unit 8200 military intelligence unit. SO004, SO021, SO023
CO019 Decart's CPO Moshe Shalev led operational scaling during the October 2024 Oasis viral launch. SO023
CO020 Dr. Kfir Aberman is a co-creator of the DreamBooth diffusion-tuning technique and a former Snap and Google researcher. SO005
CO021 Decart has not published an independent board composition or named outside directors in any reviewed 2026 source. SO001, SO002
CO022 Decart's publicly named senior leadership in 2026 is limited to the two co-founders and the San Francisco R&D head, indicating high key-person dependence on Leitersdorf and Shalev. SO005, SO002, SO022
CO023 Decart raised $21 million in its October 2024 stealth exit, led by Sequoia Capital and joined by Oren Zeev's Zeev Ventures. SO006, SO023
CO024 Within roughly two months of October 2024, Decart closed two rounds totaling about $53 million at a $500 million valuation per Ynet's December 2024 retrospective. SO023
CO025 Decart announced a $100 million Series B at a $3.1 billion valuation on August 7, 2025. SO003, SO018, SO019, SO022
CO026 Sequoia Capital, Benchmark, and Zeev Ventures all rolled over into Decart's August 2025 Series B as existing investors. SO018, SO019
CO027 Aleph VC joined Decart's August 2025 Series B as a new investor per The SaaS News' Series B summary. SO019
CO028 After the August 2025 Series B, Decart's total cumulative capital raised stood at $153 million. SO019, SO022
CO029 Decart announced a $300 million round on May 18, 2026, led by Radical Ventures at a roughly $4 billion valuation. SO002, SO004, SO021
CO030 New investors in Decart's May 2026 round include NVIDIA, eBay Ventures, Adobe Ventures, Toyota Ventures, Atreides Management, and Valor Equity Partners. SO002, SO004
CO031 Returning institutional investors in Decart's May 2026 round include Sequoia Capital, Benchmark, and Zeev Ventures. SO002, SO021
CO032 Decart's May 2026 round brings the company's cumulative capital raised to over $450 million by the company's own disclosure. SO002, SO004
CO033 Private investors in Decart's May 2026 round include Andrej Karpathy, Michael Eisner, members of the Nintendo founding family, and Moritz Baier-Lentz. SO002, SO021, SO030
CO034 Amazon is publicly described by Decart as a strategic customer in the May 2026 announcement rather than as a disclosed equity investor. SO002, SO004
CO035 Leitersdorf told Ynet in August 2025 that Decart had used less than $10 million of its $153 million total funding to date. SO022
CO036 Leitersdorf told TechCrunch in June 2026 that Decart had burned drastically less than $100 million in its lifetime. SO007
CO037 Decart's reported revenue is described as "significant" through DOS licensing to cloud providers and Lucy/Oasis customers but is not audited or quantified in reviewed sources. SO002, SO022
CO038 Decart emerged from stealth with the Oasis viral demo on October 31, 2024 alongside its first announced funding. SO006, SO015
CO039 Decart's Oasis demo reportedly reached more than five million users within one week of launch per Ynet's December 2024 profile. SO023
CO040 Decart grew from roughly 15 employees pre-2025 to about 60 employees by August 2025 per Ynet's Series B coverage. SO005, SO022
CO041 Decart says it has a developer community of more than 100,000 users as of June 2026 per Leitersdorf's TechCrunch interview. SO007
CO042 Decart partners with AWS to optimize Lucy on Trainium2 and Trainium3 chips and to distribute its models through Amazon Bedrock. SO020
CO043 Decart says its DOS 2.0 stack delivers more than 1,600 tokens per second for agentic inference and up to 100 frames per second for full-HD video. SO002
CO044 AWS's Annapurna Labs VP Nafea Bshara is quoted in Decart's May 2026 release saying Lucy2 exceeds 80% Model FLOPS Utilization on Trainium3. SO002
CO045 Decart launched DOS 2.0 alongside its May 2026 funding announcement. SO002
CO046 Decart pledged millions in funding to the Technion as part of a 2025 strategic AI partnership. SO024
CO047 Decart launched the Oasis 3 Preview real-time world model for autonomous-vehicle scene simulation on June 10, 2026 via a public API. SO007, SO011, SO030
CO048 Decart's competitive positioning in 2026 spans Google DeepMind Genie 3, World Labs' Marble, Runway, Luma, and the in-house AV simulators at Waymo, Tesla, NVIDIA, and Wayve. SO007, SO030
CO049 TechCrunch's June 2026 review of Oasis 3 documents long-rollout degradation, missing physics (cars driving through other cars), and unresponsive controls as live product limitations. SO007
CO050 TechCrunch's October 2024 review of Oasis flagged that Decart did not disclose any Microsoft permission to train Oasis on Minecraft footage, leaving copyright as an unresolved adverse signal. SO006
CM001 Decart's product surface in 2026 spans real-time video (Lucy), world models (Oasis), and the DOS optimization stack, positioning the company in real-time generative AI infrastructure. SM001, SM002
CM002 TBRC defines the generative-AI-in-gaming market as revenue earned for procedural content generation, dynamic level design, object placement, and terrain creation, plus related goods and services. SM020
CM003 Decart's flagship realtime models target sub-30-millisecond response loops, which structurally differentiates them from batch-oriented video and image generators like Sora, Veo, and Runway Gen. SM001, SM002
CM004 TechCrunch places Decart's Oasis 3 in competition with Google DeepMind Genie 3, World Labs' Marble, Runway, and Luma in the world-model race as of June 2026. SM007
CM005 TechCrunch describes Oasis 3 Preview as positioned for autonomous-vehicle scene generation alongside the in-house AV simulators of Waymo, Tesla, and Wayve and NVIDIA's DRIVE Sim stack. SM007
CM006 Decart sells API access on a per-second basis through platform.decart.ai with realtime models such as Lucy 2.1 and Oasis 3 Preview priced at $0.02 per second of active generation. SM009, SM002, SM007
CM007 Decart's market boundary is most accurately read as horizontal real-time generative-AI infrastructure rather than a single gaming-only category. SM001, SM002, SM007
CM008 TBRC's market-definition for generative-AI in gaming excludes GPU hardware capex by treating gaming-software revenue as the unit of analysis. SM020
CM009 TBRC's January 2026 report records that the American Gaming Association measured U.S. commercial gaming revenue at $51.14 billion in August 2025, up 8.9% year-on-year. SM020, SM021
CM010 TBRC sizes the generative-AI-in-gaming market at $1.79 billion in 2025. SM020, SM021
CM011 TBRC forecasts the generative-AI-in-gaming market to reach $5.09 billion in 2030 at a 23.2% compound annual growth rate. SM020, SM021
CM012 Research and Markets reports the same $1.79 billion 2025 and $5.09 billion 2030 generative-AI-in-gaming numbers as TBRC, extending the series to 2035. SM021
CM013 TBRC's 2026 report states the generative-AI-in-gaming market will grow from $1.79 billion in 2025 to $2.21 billion in 2026 at a 23.1% CAGR. SM020, SM021
CM014 TBRC names Asia-Pacific as the largest generative-AI-in-gaming region in 2025. SM020, SM021
CM015 BCG forecasts global cloud-gaming revenue to grow from roughly $1.4 billion in 2025 to about $18.3 billion in 2030 at a compound annual growth rate above 50%. SM023
CM016 BCG reports that UGC creator payouts for Fortnite and Roblox alone exceeded $1.5 billion in 2025. SM023
CM017 BCG's 2026 Video Gaming Report finds that 60% of surveyed players have tried cloud gaming and 80% of those reported a positive experience. SM023
CM018 Newzoo publishes the PC & Console Gaming Report 2026 as a standard industry tracker for installed base and platform revenue. SM022
CM019 McKinsey QuantumBlack's State of AI page was access-denied in this run, preventing direct citation of its 2026 cross-industry generative-AI adoption benchmarks. SM024
CM020 No reviewed analyst source publishes an independent market size for the AV / physical-AI / world-model simulation segment that Oasis 3 Preview targets as of June 2026. SM007, SM002
CM021 BCG estimates that approximately 50% of game studios were using AI as of mid-2025 per its 2026 Video Gaming Report. SM023
CM022 BCG's Steam metadata analysis found that around 20% of new games disclosed AI use as of mid-2025, double the level a year earlier. SM023
CM023 BCG cites AAA game development cost reaching $300 million per title as the cost pressure motivating AI adoption inside studios. SM023
CM024 Decart's May 2026 cap-table additions include Toyota Ventures, NVIDIA, Adobe Ventures, eBay Ventures, and angels Andrej Karpathy and Michael Eisner alongside members of the Nintendo founding family, signalling demand from creative tools, automotive, commerce, and gaming verticals. SM002, SM004
CM025 Amazon is publicly described by Decart as a strategic customer of DOS in the May 2026 funding release, with AWS giving Decart early Trainium3 access. SM002, SM019
CM026 Decart's DOS optimization stack is compiled to NVIDIA GPUs, Google TPUs, and Amazon Trainium silicon, signalling a multi-cloud inference posture. SM002, SM013
CM027 Decart says it serves a developer community of more than 100,000 users as of June 2026 per Leitersdorf's TechCrunch interview. SM007
CM028 Decart's published self-serve adoption path is its platform.decart.ai console with per-second realtime pricing layered over enterprise contracts. SM014, SM009
CM029 Decart does not publish paid-customer counts, ARR, or named enterprise customers beyond Amazon in reviewed 2026 sources. SM002, SM007
CM030 Decart's Oasis 3 Preview opened to developers on June 10 2026 via a public API at $0.02 per second, marking the first paid touchpoint for AV/physical-AI buyers. SM007, SM011
CM031 The European Parliament's March 2026 own-initiative report on copyright and AI calls for full transparency of training data, an itemised list of all copyright-protected works used, and a rebuttable presumption of infringement when transparency is not met. SM026, SM027
CM032 Osborne Clarke characterises the March 2026 EU Parliament copyright resolution as a "watershed moment" that proposes a 5-7% global-turnover flat-rate copyright fee pending a permanent licensing solution. SM027, SM026
CM033 The EU AI Act's General Purpose AI obligations have been in force since 2 August 2025 per the Presenc AI 2026 policy tracker. SM028
CM034 The California AI Transparency Act (SB 942) and the GenAI Training Data Transparency Act (AB 2013) both took effect on 1 January 2026. SM028
CM035 The Texas Responsible AI Governance Act (TRAIGA) took effect on 1 January 2026 with the broadest sectoral scope of any U.S. state AI law. SM028
CM036 A Trump administration executive order dated 11 December 2025 directed the Department of Justice to challenge state AI laws on preemption grounds, with three states (Colorado, New York, Illinois) under active federal litigation by May 2026. SM028
CM037 The U.S. Copyright Office maintains an ongoing Artificial Intelligence portal as the federal framing for AI and copyright as of June 2026. SM025
CM038 TechCrunch's October 2024 review of Oasis flagged that Decart did not disclose Microsoft permission to train on Minecraft footage, leaving a pre-existing copyright-licensing question that compounds 2026 EU regulatory exposure. SM006
CP001 Decart raised $300 million in May 2026 at a valuation of approximately $4 billion, led by Radical Ventures with participation from NVIDIA, Sequoia Capital, Benchmark, and strategic investors Toyota Ventures, Adobe Ventures, and eBay Ventures. SP019, SP027
CP002 Decart's DOS optimization stack delivers over 1,600 tokens per second for agentic inference, compared to an industry average of approximately 200 tokens per second. SP019
CP003 Lucy model generates video with a time-to-first-frame of under 30 milliseconds and achieves up to 100 frames per second on Amazon Trainium3 hardware via the DOS stack. SP019, SP016
CP004 Runway raised $315 million in a Series E round in February 2026, nearly doubling its valuation to $5.3 billion, led by General Atlantic with participation from NVIDIA, Adobe Ventures, and others, to pre-train the next generation of world models. SP002, SP005
CP005 World Labs raised $1 billion in new funding in February 2026, including $200 million from Autodesk, with investors including AMD, Emerson Collective, Fidelity, and NVIDIA, to accelerate its world-model and spatial-intelligence mission. SP003, SP023
CP006 Google DeepMind released Genie 3 as a research preview in August 2025, capable of generating interactive worlds at 24 fps at 720p with promptable world events and environmental consistency for several minutes, but without a public commercial API. SP001
CP007 Luma AI raised $900 million in a Series C led by Humain in November 2025, with a partnership on a 2-gigawatt AI supercluster in Saudi Arabia. SP010, SP021
CP008 Inworld AI raised $50 million at a $500 million valuation from Lightspeed Venture Partners, giving it over $100 million in total funding, focused on AI game characters and voice AI. SP007
CP009 Decart has a community of more than 100,000 developers building products on its Lucy model API, primarily in e-commerce, livestreaming, and creative applications. SP012, SP019
CP010 Oasis 3 is priced at $0.02 per second of active generation via a public API, making it the only publicly priced real-time world-model API for driving simulation as of mid-2026. SP015, SP012
CP011 Decart claims the DOS stack delivers more than a 100x improvement in cost efficiency over comparable systems and runs real-time AI 8x faster than any comparable system on equivalent hardware. SP019, SP024
CP012 Oasis 3 generates three synchronized camera feeds — one front-facing and two side-facing — to match the perception setup of most camera-first autonomous vehicle stacks. SP012
CP013 TechCrunch's independent June 2026 testing of Oasis 3 found that the model's thematic integrity degrades rapidly — a New York City street became a generic urban environment within minutes and did not preserve location landmarks across navigation. SP012
CP014 Oasis 3 does not simulate physics correctly — vehicles drive through each other — which Decart's CEO acknowledged as "a major research problem that we're cracking now" in June 2026. SP012
CP015 Waymo unveiled its own generative world-model simulation architecture in early 2026, built on top of Google DeepMind's Genie 3, positioned as a core training tool for the Waymo Driver, producing both camera and lidar output from plain language prompts. SP022
CP016 Runway's Gen 4.5 model outperformed video-generation offerings from Google and OpenAI on several benchmarks, earning Runway credibility in the industry and likely factoring into its Series E investor interest. SP002
CP017 Runway is expanding beyond its historical media and entertainment base into gaming and robotics, and plans to use Series E capital to rapidly grow its team across research, engineering, and go-to-market from approximately 140 people in February 2026. SP002, SP005
CP018 World Labs' first commercial product, Marble, allows users to create editable and downloadable 3D environments from images, video, or text, primarily targeting media, entertainment, and architecture use cases. SP003, SP023
CP019 Scenario offers access to 550+ AI models from 50+ providers as an aggregator platform for gaming and creative teams, trusted by over 15,000 customers, representing a multi-model aggregation substitute model distinct from Decart's single-vendor inference approach. SP009
CP020 Inworld AI repositioned as "The #1 Realtime Voice AI" for consumer companion applications, achieving 1 million users in 19 days on OtherHalf, and no longer focuses on world-model or AV simulation capabilities that compete with Decart Oasis. SP006
CP021 Decart's Lucy2 model exceeds 80% Model FLOPS Utilization (MFU) on Amazon Trainium3, meaning more than 80% of the chip's raw computing capacity is doing productive inference work, according to AWS VP Nafea Bshara. SP019
CP022 Oasis 3 uses an auto-regressive architecture that generates approximately 8,000 tokens per frame, consuming hundreds of thousands of tokens per second, which causes memory context to fill rapidly and contributes to long-session environment degradation. SP012
CP023 Decart's CEO states that extending Oasis 3's context window to store millions more tokens and compressing memory into fewer tokens are active research problems that may partially resolve environment degradation in a future model version. SP012
CP024 Decart's strategic investors include NVIDIA (which also runs its own DRIVE Sim stack), Toyota Ventures, Adobe Ventures, and eBay Ventures — all potential customers but also entities with adjacent or competing simulation and media capabilities. SP019
CP025 Rosebud AI focuses on AI-generated 3D game creation from text descriptions using a no-code approach targeting indie developers, with limited disclosed funding from Crunchbase data. SP008, SP025
CP026 Luma AI's $900M Series C included a partnership to build a 2-gigawatt AI supercluster in Saudi Arabia through Humain, representing a significantly larger compute infrastructure investment than Decart's AWS Trainium3 arrangement. SP010, SP021
CP027 World Labs focuses on 3D spatial AI and downloadable 3D world creation (Marble), while Decart focuses on 2D real-time video transformation (Lucy) and real-time driving simulation (Oasis), representing distinct architectural and market approaches to world models. SP003, SP004, SP012
CP028 Decart's primary Oasis 3 market is mid-tier AV companies, robotics labs, and drone startups below the top tier (Waymo, Tesla, NVIDIA, Wayve) that cannot afford Google-scale world-model R&D programs. SP022, SP012
CP029 Decart's CEO stated in June 2026 that the company has burned through "drastically less" than $100 million in its lifetime despite raising over $450 million, a claim consistent with the August 2025 statement that it had spent less than $10 million of investor capital. SP012, SP026
CP030 Runway signed a deal with CoreWeave to expand its compute capacity for world-model pre-training, while Decart uses Amazon Trainium3 and distributes through AWS Bedrock — representing different cloud-infrastructure strategies for world-model compute. SP002, SP019
CP031 Waymo, Tesla, NVIDIA, and Wayve are all building comparable world-model simulation in-house, making the top tier of AV programs unavailable as Oasis 3 customers and limiting Decart's addressable market to the mid-tier. SP022, SP012
CP032 Google DeepMind's Genie 3 is a research-preview product with no public commercial API as of mid-2026, giving Decart a temporal first-mover window before Google decides whether to monetize world-model capabilities externally. SP001, SP022
CP033 Decart claims its DOS stack achieves more than 100x improvement in cost efficiency and runs real-time AI performance 8x faster than any comparable system on equivalent hardware, representing the primary technical basis for competitive cost differentiation. SP019, SP024
CP034 Runway plans to rapidly expand its approximately 140-person team across research, engineering, and go-to-market using Series E capital, and has existing partnerships with Adobe and CoreWeave that provide distribution and compute advantages. SP002
CP035 Both Runway and World Labs explicitly view gaming as an initial go-to-market for world models, making Decart's gaming and interactive entertainment customer base directly contested by two well-funded competitors building world models. SP002, SP003
CI001 Decart's total capital raised exceeds $450 million across three rounds: approximately $53 million in seed funding, $100 million Series B at a $3.1 billion valuation (August 2025), and $300 million Series C at approximately $4 billion (May 2026). SI021, SI001
CI002 CEO Dean Leitersdorf stated in August 2025 that Decart spent less than $10 million of its investors' money despite raising $153 million in eleven months, with operating costs covered by licensing revenue. SI023, SI026
CI003 As of June 2026, Decart's pay-per-second pricing is: Lucy 2.1 at $0.02/sec (720p realtime), Lucy Restyle 2 at $0.01/sec, Oasis 3 Preview at $0.02/sec (720p world model), and the legacy Lucy Clip at $0.15/sec. SI009, SI022
CI004 Decart's first revenue stream was licensing its GPU optimization technology to major AI laboratories and cloud providers through multi-million-dollar contracts before the public API launch. SI007, SI023
CI005 Decart had more than 100,000 developers on its platform as of June 2026, primarily building on the Lucy API for e-commerce virtual try-on and livestreaming transformation use cases. SI022, SI027
CI006 Decart's pay-per-second pricing makes per-unit economics observable for self-serve API users, but enterprise contract pricing for Oasis 3 and AWS Bedrock deployments varies by use case and is not publicly disclosed. SI009, SI022
CI007 Decart signed a commercial agreement with Amazon Web Services for joint go-to-market distribution via Amazon Bedrock, targeting enterprise customers in media, commerce, advertising, and physical AI. SI021, SI001
CI008 NVIDIA participated in Decart's $300M Series C as both a financial investor and technology partner, enabling DOS deployment on NVIDIA GPU hardware for enterprise customers. SI021, SI001
CI009 DOS 2.0, announced in May 2026, claims 1,600+ tokens per second for agentic inference versus an industry average of approximately 200 tokens per second, and supports up to 100 frames per second for world model generation. SI021, SI010
CI010 Lucy 2 running on Amazon Trainium3 achieved greater than 80% Model FLOPS Utilization (MFU) according to a public statement by AWS VP Nafea Bshara, indicating near-peak hardware efficiency. SI002, SI003
CI011 Decart's DOS optimization reduced per-video production cost from hundreds or thousands of dollars to less than $0.25 per video, per the CEO and corroborated by independent press reporting. SI007, SI023
CI012 CEO Leitersdorf stated in June 2026 that Decart had burned "drastically less" than $100 million in its lifetime, implying total cumulative spend under $100 million through at least mid-2026. SI022, SI027
CI013 CEO Leitersdorf stated in August 2025 that Decart is "already generating millions in revenue" and "nearing profitability," having used less than $10 million of total funding to date. SI023, SI026
CI014 As of the $300M raise, Decart had active contracts with "several of the world's largest cloud providers, AI laboratories, and hyperscale companies," with Amazon joining as a strategic customer. SI021, SI001
CI015 Decart's revenue streams as of mid-2026 include: DOS licensing to hyperscalers, Lucy API pay-per-second, Oasis API pay-per-second, virtual try-on API, and enterprise distribution via AWS Bedrock. SI009, SI021, SI022
CI016 Toyota Ventures, Adobe Ventures, and eBay Ventures participated in Decart's $300M round as strategic investors and are described by the CEO as potential customers, creating commercial alignment with investment participation. SI001, SI022
CI017 With more than $450 million raised and company-claimed sub-$100M lifetime spend through mid-2026, Decart's implied remaining cash position is in the $350–$440M range, supporting multi-year runway even at materially elevated burn rates. SI001, SI022
CI018 Decart has not publicly disclosed ARR, gross margin, monthly burn rate, customer count by revenue tier, net revenue retention, EBITDA, or any audited financial statements as of mid-2026. SI022, SI028
CI019 TechCrunch's hands-on review of Oasis 3 (June 2026) found physics consistency failures — cars driving through each other — environmental degradation over long sessions, and described the experience as "dream-like, disjointed." The CEO acknowledged physics simulation as "a major research problem we're cracking now." SI022, SI027
CI020 The SEC Form D filed September 19, 2025 by JSL Decart.AI Coinvest, L.P. (CIK 0002084011, Delaware LP) confirms the existence of a co-investment vehicle for the Decart Series B but provides no revenue, burn, or valuation detail beyond confirming the LP structure. SI029
CI021 Decart's asset-light compute model — using leased Amazon Trainium3, NVIDIA GPU, and Google TPU capacity rather than owned hardware — limits capital expenditure intensity compared to AI companies building proprietary data center infrastructure. SI002, SI021
CI022 Decart's AWS commercial partnership was structured as a go-to-market agreement enabling enterprise distribution through Amazon Bedrock and technical integration using Trainium3 accelerators for high-efficiency inference. SI002, SI001
CI023 Decart's $0.02/sec API price implies hourly revenue of $72 per active concurrent stream; at a realistic concurrent utilization of 0.1–2% of its 100K+ developer base, estimated annual API throughput revenue runs $4M–$90M before compute costs. SI009, SI022
CI024 Decart's software-only inference-delivery model means primary COGS are third-party compute billing and R&D payroll, implying a gross margin profile structurally similar to a managed API provider at 50–80% at scale — though actual margins are not disclosed. SI021, SI010
CI025 Decart's $300M raise in May 2026 — approximately nine months after the CEO claimed the company barely needed capital — is either a strategic growth-investment pivot or is in tension with the prior capital-efficiency narrative; both claims cannot simultaneously be fully accurate. SI001, SI022, SI023
CI026 The AWS-Decart Trainium3 commercial integration was operationally live and generating revenue before the $300M raise (as reported by AI News in December 2025), confirming that partnership revenue preceded the Series C round close. SI002, SI004
CI027 Strategic investors NVIDIA, Amazon, Toyota, Adobe, and eBay represent commercial distribution or hardware supply relationships; if commercial contracts were influenced by investment terms, reported revenue from these entities may not reflect true arms-length demand. SI001, SI021
CI028 Decart's Acceptable Use Policy and Terms of Service prohibit commercial use for non-consensual deepfakes, CSAM, and other harmful applications, establishing the enterprise-grade compliance posture required for cloud marketplace distribution via AWS Bedrock. SI011, SI020
CI029 Radical Ventures led Decart's $300M Series C at approximately $4 billion valuation, with Andrej Karpathy, former Disney CEO Michael Eisner, members of the Nintendo founding family, and gaming investor Moritz Baier-Lentz among angel participants. SI001, SI024
CI030 Decart's privacy policy and platform legal terms establish the contractual framework for enterprise API usage, consistent with the compliance requirements of large cloud-marketplace distribution through AWS Bedrock. SI019, SI020, SI011
CI031 Decart has not published quarterly reports, investor letters, audited accounts, or any formal financial disclosure to the public as of mid-2026; as a private Israeli company it has no statutory obligation to do so absent an IPO process. SI022, SI029
CI032 Decart's financing dependency for its current growth rate — 60+ employees, multi-site R&D, model training, and infrastructure build-out — cannot be rigorously assessed without burn-rate and cash-position disclosure. SI007, SI026
CI033 As of mid-2026, all of Decart's financial performance claims — "millions in revenue," "nearing profitability," "spent less than $10M" — originate from CEO statements transmitted through press coverage; no board attestation, auditor sign-off, or third-party financial verification has been disclosed publicly. SI022, SI028
CI034 DOS 2.0 performance benchmarks (1,600+ tokens/sec, 100fps, >80% MFU) are company-published metrics; as of mid-2026 no independent third-party audit or peer-reviewed validation of these performance figures has been published. SI021, SI009
CI035 Decart's vertical integration — DOS optimizing the same hardware stack that powers Lucy and Oasis — means internal model-training and inference cost savings accrue to COGS rather than being passed externally, potentially supporting gross margin expansion as scale increases. SI021, SI010
CE001 Decart publicly presents DOS, Lucy, and Oasis as distinct but integrated product layers. SE012, SE019
CE002 The public platform exposes realtime and batch model workflows through a hosted API. SE001, SE014
CE003 Decart publishes official SDK references for JavaScript, Python, Swift, and Android. SE013, SE024, SE025
CE004 Use-case docs position Lucy for live streaming effects, virtual try-on, character transformation, and creative video workflows. SE004, SE014
CE005 New accounts receive free credits for evaluation according to the public overview and pricing surfaces. SE014, SE020
CE006 Public pricing is usage-based for realtime, video, and image generation rather than seat-based. SE020, SE014
CE007 Oasis 3 Preview is listed at $0.02 per second and the pricing page illustrates a 60-second session cost of $1.20. SE020, SE010
CE008 The Lucy 2.1 pricing example shows a 30-second session cost of $0.60. SE020
CE009 The GitHub organization lists public SDK, try-on, RL example, and XR repositories that act as visible developer signals. SE024, SE025, SE026
CE010 The Decart-XR repository describes an open-sourced Quest app for realtime world transformation with a Discord community link. SE026
CE011 The Oasis 3 docs describe the product as a real-time promptable world model. SE003, SE017
CE012 Oasis 3 uses a dedicated Python gRPC SDK rather than only the generic realtime SDK paths. SE003, SE025
CE013 The default hosted Oasis endpoint is presented as a managed service so developers do not host the model themselves. SE003
CE014 The Oasis initialize flow advertises three output streams named left_forward, front, and right_forward. SE003, SE017
CE015 Each Oasis inference call takes exactly four throttle-and-steering action pairs and returns four frames per stream. SE003
CE016 The Oasis docs include an RL example and a collision-risk demo loop built on live API calls. SE003, SE024
CE017 The Oasis browser demo is exposed publicly through a dedicated preview domain. SE003
CE018 Company materials say Oasis 3 is initially targeted at autonomous-vehicle developers and broader physical-AI use cases. SE017, SE010, SE022
CE019 Company materials extend Oasis 3 positioning from AV into drones, maritime, humanoid, and robotics workflows. SE017
CE020 The Oasis page says the system is not a physics engine. SE017, SE018
CE021 Decart claims Oasis 3 can deliver under-200ms end-to-end latency. SE017, SE026
CE022 Decart claims Oasis 3 runs at 22 FPS at 512 by 768 by 3 resolution. SE017
CE023 Decart claims DOS 2.0 can exceed 1,600 tokens per second for agentic inference. SE019
CE024 Decart claims DOS 2.0 can run across NVIDIA GPUs, Google TPUs, and Amazon Trainium. SE019, SE009, SE028
CE025 Decart claims DOS 2.0 can support full-HD video and world-model inference up to 100 FPS. SE019, SE021
CE026 AI News reports that Lucy achieved a 40ms time-to-first-frame on AWS Trainium2. SE021
CE027 AI News reports that Trainium3 should enable outputs up to 100 FPS with lower latency for Lucy. SE021, SE029
CE028 TechCrunch observed that Oasis 3 can generate compelling driving scenes but scene identity can drift over time. SE010
CE029 TechCrunch observed that vehicles can pass through one another in Oasis 3, indicating imperfect collision physics. SE010
CE030 TechCrunch reported that every Oasis frame is roughly 8,000 tokens, making long context management a core challenge. SE010
CE031 The 2024 Oasis project page openly described limits in domain generalization, memory, precise control, and distant-detail fidelity. SE018, SE011
CE032 Google DeepMind presents Genie 3 as a real-time interactive world model, making Decart part of a fast-moving competitive set rather than a category of one. SE030, SE010
CE033 The FAQ instructs browser and mobile developers to use short-lived client tokens instead of permanent API keys. SE015, SE002
CE034 The FAQ says an expired client token blocks new connections but does not disconnect an active realtime session. SE015
CE035 The terms say users own their input and Decart assigns output rights to users subject to law and the terms. SE016
CE036 The terms also say Decart may use customer content to develop and improve the platform and for marketing or promotional purposes. SE016
CE037 The privacy policy says Decart may process inputs, outputs, generated media, and live audio or video recordings to improve products and conduct research. SE006, SE016
CE038 The privacy policy lists government-issued identification as a possible data category for identity verification. SE006
CE039 The acceptable use policy bans deceptive deepfakes, military or ITAR-related uses, and high-harm safety-critical applications. SE008
CE040 The public status page showed all systems operational with a 90-day historical uptime view at the time of capture. SE007
CE041 The authentication documentation specifies that permanent API keys carry a dct_ prefix and that browser and mobile apps must use short-lived client tokens (ek_ prefix) instead to avoid exposing credentials in frontend bundles. SE031
CE042 Decart publishes a dedicated API Terms of Service (last updated May 28 2026) that establishes binding developer obligations distinct from the general Terms of Service, including attribution, usage limits, and compliance requirements. SE032
CE043 The SDK-direct integration path lets applications connect end users to Decart WebRTC endpoints without any proxy infrastructure; the documented session lifecycle exposes connect, set, setPrompt, setImage, disconnect, and getConnectionState methods, plus event callbacks for connectionChange, generationTick, and error. SE033
CU001 Decart and multiple independent reports describe Amazon as a strategic customer. SU001, SU007, SU022
CU002 Decart says go-to-market collaborations with AWS are already underway across the ecosystem. SU001, SU007
CU003 Decart says it is generating significant revenue through contracts with cloud providers, AI labs, and hyperscalers. SU001, SU007
CU004 Decart says Lucy powers live deployments across commerce, virtual try-on, dynamic in-video advertising, live streaming, social platforms, and gaming. SU001, SU003
CU005 The public use-cases page highlights live streaming effects, virtual try-on, and batch content pipelines as current application patterns. SU003
CU006 The ecommerce try-on guide shows how merchants can add a webcam-based Try it on button to product pages using garment images and descriptive prompts. SU002, SU003
CU007 The try-on guide says the tryon-examples repository contains six production-ready examples. SU002, SU029
CU008 The try-on guide says client tokens last 10 minutes and active sessions continue working after expiry. SU002, SU026
CU009 TechCrunch reports that Decart already has a community of more than 100,000 developers. SU006
CU010 TechCrunch says many of those developers are building on Lucy in e-commerce and livestreaming. SU006
CU011 The public pricing surface keeps Oasis 3 Preview at $0.02 per second while enterprise pricing depends on use case. SU004, SU006, SU008
CU012 CTech reports that enterprise customers will be able to use Decart for AI applications in media, commerce, advertising, and physical AI. SU007
CU013 Independent partner coverage says Decart models are or will be available through Amazon Bedrock. SU010, SU032, SU033
CU014 AI News reports Lucy achieves a 40 millisecond time to first frame on AWS Trainium2. SU010
CU015 Independent partner coverage says early access to Trainium3 should enable higher FPS and lower latency for Lucy. SU010, SU032
CU016 The GitHub organization lists SDK, try-on, Oasis 3 RL example, Android realtime example, and AI SDK provider repositories. SU029
CU017 The decart-python repository presents itself as a Python SDK for Decart models. SU030
CU018 The decart-python repository includes a Gradio-based interactive test UI and a realtime synthetic example. SU030
CU019 The Decart-XR repository is released as an open-source Quest developer project. SU031
CU020 The Decart-XR repository claims sub-200ms latency for live Quest 3 video transformation. SU031
CU021 The Decart-XR repository links to a public Discord community for developers. SU031
CU022 Tech Startups says Lucy on Bedrock is intended to bring real-time AI video to every industry and market at scale. SU032
CU023 TechBriefly says customers can access Lucy on AWS Trainium through Amazon Bedrock. SU033
CU024 bonega.ai argues that Bedrock distribution lowers vendor-relationship and security-review friction for enterprises already on AWS. SU034
CU025 Startup Fortune frames Oasis 3 as a bet that rented simulation can win before larger players rely only on in-house stacks. SU008, SU024
CU026 Startup Fortune says Oasis 3 currently provides three synchronized camera feeds but not lidar output. SU008
CU027 TechCrunch says Oasis 3 initially targets autonomous-vehicle companies and broader physical-AI applications. SU006, SU008
CU028 Across the fetched public sources, Amazon is the only clearly named strategic customer while most other customer references remain cohort-level or unnamed. SU001, SU007, SU022, SU006
CU029 No fetched public source discloses exact active customer count, NRR, GRR, gross churn, or renewal rate. SU001, SU007, SU014, SU026
CU030 Official materials disclose usage-based self-serve pricing but not the structure of enterprise contracts. SU004, SU006, SU026
CU031 CTech reports that Decart already generates revenue through contracts with several of the world's largest cloud providers, AI laboratories, and hyperscale companies. SU007, SU001
CU032 JNS independently repeats that Amazon joined Decart as a strategic customer in the 2026 round. SU022, SU007
CU033 CTech reported in 2025 that Decart had licensed its GPU optimization stack to major cloud providers in multimillion-dollar deals. SU019
CU034 Decart’s 2026 publication says the upcoming Lucy 2.5 is focused on gaming, e-commerce, streaming, and advertising. SU001, SU011
CU035 The ecommerce guide positions live try-on for product pages, digital mirrors, and styling apps. SU002, SU003
CU036 The use-cases page also positions Decart for post-production and content-pipeline automation. SU003
CU037 AWS and partner sources emphasize production-readiness, scale, reliability, and compliance as key adoption criteria for enterprise AI buyers. SU009, SU034
CU038 Public customer proof is materially stronger for ecosystem and developer cohorts than for a broad roster of named enterprise logos. SU001, SU007, SU029, SU031
CU039 The GitHub organization shows adoption assets spanning web, mobile, Python, and VR, implying a broad developer-acquisition strategy rather than a single integration path. SU029, SU030, SU031
CU040 Public retention evidence is weak because none of the fetched sources provide renewal, contract-length, or cohort-retention data. SU001, SU007, SU014, SU026
CU041 The platform docs describe a client-token flow where developers mint short-lived keys (60-second default TTL, configurable up to 3600 seconds) on the backend and pass them to browser or mobile frontends, enabling secure self-serve onboarding without exposing permanent API keys. SU035
CU042 Decart publishes an HTTP signaling proxy integration path that lets enterprise platforms white-label the Decart API behind their own stateless HTTP endpoints, providing full control-plane visibility while media flows directly via WebRTC. SU036
CU043 Decart also documents a WebSocket signaling proxy path for platforms with existing WebSocket infrastructure, offering white-labeled endpoints and full control-message visibility for enterprise buyers. SU037
CU044 Decart publishes an official async Python SDK (pip install decart) supporting the full realtime, queue, and process APIs, broadening the accessible developer base beyond JavaScript. SU038
CU045 Decart publishes an official Android SDK available via JitPack, indicating active investment in native mobile developer channels and physical-device deployments. SU039
CU046 Decart provides a documented open-source Expo mobile app example showing end-to-end integration of realtime AI camera transformation on mobile devices, lowering the barrier for mobile-first developer onboarding. SU040
CR001 Decart raised $300 million at an estimated valuation of approximately $4 billion in May 2026, led by Radical Ventures with participation from Nvidia, Sequoia, Benchmark, eBay Ventures, Adobe Ventures, and Toyota Ventures. SR005, SR013
CR002 Decart raised $100 million at a $3.1 billion valuation in August 2025, with existing investors Sequoia Capital, Benchmark, and Zeev Ventures participating alongside new backers Aleph VC. SR004, SR014
CR003 Decart's total capital raised as of May 2026 exceeds $450 million, bringing lifetime capital to more than $450 million after the $300 million Series C. SR005, SR013
CR004 Decart has approximately 100,000+ developers in its community, primarily building on the Lucy real-time video model for e-commerce and livestreaming use cases as of June 2026. SR003
CR005 Decart's Oasis 3 world model is priced at $0.02 per second via public API, with enterprise pricing varying by use case, as of its June 2026 launch. SR003, SR030
CR006 Decart operates across three product lines — DOS (Decart Optimization Stack), Lucy (real-time video model), and Oasis 3 (physical AI world model) — as of June 2026. SR012, SR013
CR007 The EU Parliament's Legal Affairs Committee voted 17-3 in February 2026 to propose a rebuttable presumption that any generative AI model placed on the EU market has used copyrighted works for training unless full transparency obligations are met, reversing the burden of proof onto AI providers. SR006, SR027
CR008 The EU Parliament's explanatory statement proposed a flat-rate copyright fee of 5–7% of global turnover for retroactive licensing of past training-data uses where a licensing market could not yet be established. SR006, SR027
CR009 The US Copyright Office's ongoing AI rulemaking (Part 3) is developing licensing guidance for AI training data; no binding rule has been issued as of June 2026 but the guidance shapes litigation posture and creates future statutory risk. SR001, SR028, SR035
CR010 Osborne Clarke's February 2026 legal analysis characterizes the EU copyright situation for generative AI as a "watershed moment" where the principle of territoriality must be adapted so that EU law applies even when training takes place outside the EU. SR006, SR027, SR036
CR011 Gunder LLP's 2026 AI laws update notes that Colorado and California have enacted comprehensive AI governance statutes with enforcement beginning in late 2025 and 2026, and companies should continue to comply with state laws despite the December 2025 Trump Executive Order. SR007, SR008, SR037
CR012 Decart's Acceptable Use Policy (updated February 12, 2026) explicitly acknowledges obligations under the EU AI Act (AIA), Digital Services Act (DSA), and DMCA, and prohibits military, warfare, nuclear, espionage, and autonomous weapons uses of its Offers. SR010, SR009
CR013 No public IP litigation, copyright enforcement actions, or regulatory proceedings against Decart have been identified in public records as of June 2026. SR003, SR018, SR038
CR014 Decart has not publicly disclosed its training data sources, provenance, or licensing agreements for any of its models (Lucy, Oasis, DOS) as of the research date June 2026. SR010, SR009
CR015 TechCrunch's June 2026 independent review of Oasis 3 documented that vehicles drive through other vehicles, meaning the model does not simulate physics properly in the driving environment — a failure the CEO acknowledged as a "major research problem we're cracking now." SR003, SR018
CR016 Oasis 3 experiences thematic decoherence and context degradation within minutes of generation; each frame consumes approximately 8,000 tokens and generating at tens of frames per second fills context windows at hundreds of thousands of tokens per second, per CEO Leitersdorf's June 2026 admission. SR003
CR017 Decart's Lucy model runs on AWS Trainium3 hardware and is distributed through Amazon Bedrock, creating a concentrated single-cloud dependency for both inference compute and primary distribution channel. SR017, SR023
CR018 AWS Trainium3 provides Decart with 4x faster frame generation at half the cost of GPUs, per CEO Leitersdorf's statement; the hardware was newly announced and early access was obtained as part of the AWS partnership. SR017
CR019 Decart's AUP prohibits use for deepfakes without verifiable consent, CSAM/CSEM, military and weapons development, real-time biometric surveillance, and social scoring, as of the February 2026 update. SR010, SR009
CR020 No independent security audit, SOC 2 certification, or ISO 27001 certification for Decart has been identified in public records as of June 2026. SR009, SR011
CR021 Lucy achieves a time-to-first-frame of 40ms and can generate video at up to 30fps, with Trainium3 enabling up to 100fps and lower latency in the announced roadmap, per AI News reporting on the AWS partnership. SR017
CR022 Amazon (AWS) is simultaneously Decart's primary compute infrastructure provider, primary distribution channel (Amazon Bedrock), and a strategic customer — creating three-way concentration in a single counterparty. SR005, SR017
CR023 Nvidia is both an equity investor in Decart (participated in both the $100M Series B and $300M Series C) and a primary hardware supplier, creating correlated investor-supplier concentration risk. SR005, SR032
CR024 Toyota Ventures, Adobe Ventures, and eBay Ventures are equity investors in Decart's May 2026 round whose parent companies are also listed as strategic customers, creating investor/customer overlap that amplifies concentration and conflict-of-interest risk. SR005, SR013
CR025 Waymo has built its own world model on Google DeepMind's Genie 3, using it as a core simulation tool with camera and lidar output generation — representing a direct competitive alternative to Oasis 3 for the primary-tier AV market segment. SR018, SR026
CR026 Startupfortune (June 2026) characterizes Decart's addressable AV market as "everyone below the top tier" — mid-tier AV programs, robotics labs, and drone startups — after Waymo, Tesla, and Nvidia implement their own internal world models. SR018
CR027 Decart's developer community of 100,000+ primarily builds on Lucy for e-commerce and livestreaming; API revenue at $0.02/second is highly volume-dependent and revenue per developer at typical usage patterns is not publicly disclosed. SR003, SR030
CR028 Dean Leitersdorf (CEO, 27 years old) is the primary public face, fundraiser, and technical narrative driver for Decart; no formal succession plan or alternative CEO candidate has been publicly identified. SR004, SR005
CR029 Dr. Kfir Aberman, co-creator of DreamBooth (formerly at Snap and Google), heads Decart's San Francisco R&D center, serving as the most senior publicly named external hire as of June 2026. SR004, SR015
CR030 Decart's Israeli headquarters and primary engineering team in Tel Aviv creates geopolitical risk; the ongoing Israel-Gaza conflict and regional instability may affect enterprise procurement decisions by defense-policy-sensitive customers. SR004, SR022
CR031 CEO Leitersdorf stated in June 2026 that Decart has burned "drastically less" than $100 million in its lifetime — a claim materially inconsistent with the earlier August 2025 claim of less than $10M spent against $153M raised, but both are unverified company claims. SR003, SR004
CR032 Decart had approximately 60 employees as of August 2025; headcount as of June 2026 following the $300M Series C has not been publicly disclosed but hiring is expected to accelerate. SR004, SR015
CR033 Decart's AUP (February 2026) and public compliance references demonstrate regulatory awareness but do not constitute an independent audit of training-data transparency obligations under the EU AI Act or EU copyright frameworks. SR010, SR027
CR034 The next version of Oasis 3 is planned to accept video input as the starting prompt rather than image input, which CEO Leitersdorf believes may partially address consistency issues, per his June 2026 TechCrunch interview. SR003
CR035 Decart's DOS optimization stack provides inference efficiency claimed to be more than 10x cheaper than competitors — a technology advantage that would be eroded if Nvidia's own inference optimization software or AWS-native optimization achieves comparable efficiency. SR020, SR017
CR036 The EU Parliament's report calls for the EUIPO to manage a new opt-out exclusions register for AI training data and to facilitate voluntary sector-based licensing, potentially providing a compliance pathway for AI providers including Decart. SR006, SR027
CR037 Decart co-founders Dean Leitersdorf and Moshe Shalev met while serving in the reserves of Israel's Unit 8200 and founded Decart in late 2023; Unit 8200 network is a key recruiting advantage but also concentrates talent in Israeli cyber/intelligence alumni. SR004, SR022
CR038 Decart claims its Lucy model achieves a 40ms time-to-first-frame and can match the quality of OpenAI's Sora 2 and Google's Veo-3 at up to 30fps on Trainium3 — unverified company claims with no independent benchmark. SR017
CR039 Presenc.ai's 2026 AI policy tracker documents a rapidly shifting global regulatory landscape for AI systems in 2026, with multiple new jurisdictions enacting or strengthening AI governance frameworks. SR008
CR040 Google DeepMind released Genie 3, described as "a new frontier for world models," in research preview, with Waymo building its proprietary simulation system on top of Genie 3 per Startupfortune's June 2026 analysis. SR026, SR018
CR041 Decart CEO Leitersdorf has described the company's ambition as building a "before-and-after company like the iPhone" and said early investors may see 10,000x returns — language that signals a high-growth, high-risk positioning inconsistent with measured enterprise infrastructure scaling. SR004
CR042 Decart's Oasis 3 generates three synchronized camera feeds (one front-facing, two side-facing) but does not produce the lidar output that Waymo's internal world model generates — a capability gap for perception-stack-complete AV simulation. SR018, SR024
CR043 Decart's revenue is derived from DOS licensing to cloud providers and AI labs, Lucy API usage by e-commerce and livestreaming developers, and Oasis 3 enterprise AV/physical AI contracts — but revenue figures, composition, or ARR have not been publicly disclosed. SR004, SR013
CR044 The US Trump Executive Order on AI (December 2025) signals federal intent to preempt state AI regulation but does not immediately invalidate existing state AI laws; Gunder LLP advises companies to continue complying with state laws until courts and agencies clarify the EO's reach. SR007, SR001
CR045 The US Copyright Office's July 2024 report on Digital Replicas (Part 1 of its AI series) analyzed liability frameworks for AI-generated likenesses of real people; this is directly relevant to Decart's Lucy real-time video transformation products, which apply AI models to live camera feeds containing human subjects. SR039
CR046 The US Copyright Office's January 2025 report on Copyrightability (Part 2 of its AI series) held that AI-generated outputs without sufficient human authorship are not copyrightable under current US law; this affects Decart's terms that assign output rights to users, since those outputs may lack copyright protection in the US unless sufficient human creative input is present. SR040
CV001 Decart raised $300 million in a Series C round in May 2026 at an implied post-money valuation of approximately $4 billion, led by Radical Ventures with participation from Nvidia, Sequoia Capital, Benchmark, and strategic investors eBay Ventures, Adobe Ventures, and Toyota Ventures. SV001, SV003
CV002 Decart raised $100 million at a $3.1 billion post-money valuation in August 2025 (Series B), with participation from Sequoia Capital, Benchmark, Aleph VC, and Zeev Ventures, implying a 29% step-up to the $4B Series C valuation in approximately nine months. SV004, SV023
CV003 Decart's total capital raised exceeded $450 million as of May 2026, making it one of the best-funded Israeli AI startups and placing it in the top tier of private AI company capital formation globally. SV001, SV013
CV004 The Series C round included technology company investors who are simultaneously paying customers, including Toyota Ventures, Adobe Ventures, and eBay Ventures, creating an atypical investor-customer overlap structure. SV003
CV005 Decart's official Series C announcement characterized lead investors as "tech leaders" who chose to invest because they are users of the product, confirming the dual customer-investor role and strategic backing thesis. SV003
CV006 Runway AI raised $315 million at a $5.3 billion post-money valuation in a Series E round in February 2026, positioning it as the highest-valued direct comparable to Decart in the AI video and world model space. SV005, SV006
CV007 World Labs secured $200 million from Autodesk in February 2026 at an estimated valuation near $5 billion, with a specific commercial focus on 3D design workflows; this is a more commercially anchored investment than Decart's round despite similar valuation levels. SV007, SV008
CV008 Luma AI raised $900 million in a Series C round from HUMAIN (Saudi Arabia's state AI fund) and partners in November 2025, implying a post-money valuation close to its total capital raised — a relatively thin premium that signals investor caution on pure video generation business models. SV009, SV010
CV009 Inworld AI achieved a $500 million post-money valuation after raising $50 million from Lightspeed Venture Partners in 2024, giving it a 10x price-to-capital ratio but limited absolute scale — it is the primary AI-native gaming comparable to Decart's gaming vertical. SV011
CV010 The peer group of Runway ($5.3B), World Labs (~$5B), Luma AI (~$0.9B), and Inworld AI ($0.5B) produces a simple average post-money valuation of approximately $2.9 billion; Decart's $4B represents a 38% premium to this peer average. SV005, SV007, SV009, SV011
CV011 The global gaming market generated approximately $196 billion in revenues in 2024 according to Newzoo, with the 2025 total estimated at approximately $210 billion after recovery from the post-pandemic slump. SV021, SV014
CV012 BCG's 2026 Global Gaming Survey found that approximately 20% of new Steam games disclosed AI use in mid-2025, doubling the prior year's rate, confirming generative AI as a structural growth driver in the gaming industry. SV014, SV021
CV013 The Business Research Company estimates the generative AI in gaming market at approximately $2.1 billion in 2025, growing to $10.5 billion by 2030 at a CAGR of approximately 30%, placing real-time world model simulation as a premium category within this segment. SV015
CV014 Research and Markets separately estimates the generative AI gaming market at $2.1 billion in 2025 growing at a 24% CAGR through 2030, corroborating the Business Research Company estimate and confirming a multi-analyst consensus on this TAM. SV016, SV015, SV014
CV015 McKinsey's State of AI report finds enterprise AI adoption has crossed 72% among large organizations, validating broad enterprise demand for AI tools including simulation and video generation platforms like Decart. SV018
CV016 Statista sizes the global mobile games market at approximately $92 billion in 2025, representing the most immediately accessible portion of the gaming TAM for Decart's real-time video generation via the Lucy model. SV019
CV017 BCG identifies generative AI, user-generated content expansion, cloud gaming, and app store liberalization as the four structural trends reshaping the gaming industry over the next five to ten years — all of which are favorable tailwinds for Decart's world model platform. SV014, SV019
CV018 The e-commerce AI virtual try-on market is estimated at approximately $800 million in 2025, providing a near-term ARR pathway for Decart's Lucy model that does not require physics-accurate simulation. SV014, SV015
CV019 Decart's serviceable addressable market in the near term (2026–2028) is estimated at $150–500 million, combining e-commerce try-on, gaming API fees, and potential early AV synthetic data contracts — well below the total gaming TAM of $210 billion. SV014, SV015, SV016
CV020 In a bull scenario, Decart could achieve $25–35 million ARR by end of 2026 and $180–220 million ARR by 2028, requiring a 2–3% conversion rate of its 100,000+ developer community combined with 3–5 enterprise contracts; at 20x ARR, this sustains a $3.6–4.4 billion valuation. SV026, SV014
CV021 In a base scenario, Decart reaches $10–15 million ARR by end of 2026 and $75–100 million ARR by 2028, implying a valuation of $1.5–2.0 billion at 20x forward ARR — a 50–62% discount to the current $4 billion price. SV014, SV026
CV022 In a bear scenario, physics limitations close the AV market and competitive pressure limits gaming adoption; Decart reaches only $20–30 million ARR by 2028, implying an 87.5% discount to current valuation at 20x multiple. SV012, SV014
CV023 Decart publicly prices its Oasis 3 world model API at $0.02 per second as of June 2026; at this rate, achieving $30 million ARR requires approximately 47 billion seconds of API usage annually — implying either a very large developer community or substantial enterprise contracts. SV026, SV029
CV024 Roblox Corporation's SEC 10-K annual report shows FY2025 revenue of approximately $3.9 billion at a market capitalization of approximately $26 billion, implying a 6–7x price/revenue multiple for a mature gaming platform — providing a long-term valuation ceiling benchmark for Decart. SV017
CV025 At Roblox's mature 6–7x revenue multiple applied to Decart's bull-case 2028E ARR of $200 million, the implied valuation is approximately $1.2–1.4 billion — significantly below the current $4 billion, suggesting the current pricing embeds a technology-platform premium well above mature gaming infrastructure multiples. SV017, SV014
CV026 The private AI video peer group (Runway, World Labs, Luma, Inworld) trades at implied revenue multiples of approximately 20–100x based on estimated ARRs, consistent with the early-stage technology premium Decart also commands, but none has disclosed audited revenue that would anchor these multiples. SV005, SV007, SV009
CV027 TechCrunch's June 2026 independent review of Decart's Oasis 3 documented physics-consistency failures including vehicles passing through each other and loss of navigational control; the CEO acknowledged these as "a major research problem we're cracking now," confirming core product limitations that are not cosmetic. SV012, SV028
CV028 CBInsights confirms that Decart has not publicly disclosed any revenue, ARR, bookings, or profitability metrics as of June 2026, making the $4 billion valuation entirely assumption-led from a fundamental analysis perspective and representing a significant diligence gap. SV020, SV001
CV029 Runway's $5.3 billion valuation is supported by publicly indicated commercial traction across its video editing and AI video generation products; Decart's $4 billion valuation has no equivalent public revenue evidence, creating a fundamental comparative disadvantage in diligence. SV005, SV020
CV030 World Labs' $5 billion valuation is anchored by a strategic investment from Autodesk, a public company with direct commercial synergy in 3D design software; Decart's comparable investor-customer structure exists but is concentrated in gaming/e-commerce rather than a single dominant enterprise vertical. SV007, SV003
CV031 Decart's reliance on AWS Trainium3 for production inference creates gross margin exposure — compute costs that scale with API usage could compress margins materially below the 70–80% range typical of software-as-a-service, particularly at current API pricing of $0.02/second. SV027, SV026
CV032 The strategic investor-customer overlap (Toyota, Adobe, eBay) introduces governance complexity — these investors may have conflicting incentives around product roadmap, enterprise pricing, and exit routes, a risk structure that is atypical for a $4 billion private company. SV003, SV020
CV033 The world model AI space has no single dominant platform as of June 2026; Runway, World Labs, and Google DeepMind Genie 3 are all advancing with substantially more resources or strategic positioning, creating high market-structure uncertainty for the Decart investment thesis. SV005, SV007
CV034 Decart's technical differentiation — real-time autoregressive world model architecture at sub-100ms latency — is genuinely unique among announced products as of June 2026, providing a defensible technical moat for at least 12–18 months before comparable capability is likely replicated. SV028, SV005
CV035 The AWS Trainium3 production partnership and Amazon Bedrock distribution create scale and go-to-market leverage that are atypical for a startup at Decart's commercial stage, representing a material structural advantage over peers without equivalent cloud platform integration. SV027, SV003
CV036 No public source — media, investor, or company — has disclosed any specific revenue, ARR, or financial metric for Decart as of June 2026; all financial projections in this chapter are assumption-led and must be treated as scenario analysis, not fundamental forecasts. SV020, SV001
CV037 The primary thesis-break triggers for Decart are (A) confirmed physics failures preventing AV enterprise adoption, (B) failure to demonstrate enterprise ARR crossing $15 million by Q4 2026, (C) material copyright litigation or EU enforcement action, or (D) Runway or Google shipping a competitive real-time world model within 12 months. SV012, SV005
CV038 The recommended investment stance is TRACK with medium confidence and high risk; the upgrade trigger to BUY requires audited ARR at or above $25 million with demonstrated YoY growth and at least one public AV enterprise contract; the downgrade trigger to PASS is any of the four thesis-break triggers materializing. SV020, SV014, SV012
CV039 Decart's $4 billion valuation implies a 25% discount to Runway's $5.3 billion despite broadly comparable technical claims and investor base quality; this discount likely reflects Decart's earlier commercial stage and greater geographic concentration (Israel-headquartered team). SV005, SV001
CV040 The gaming simulation TAM is potentially $5–15 billion if world models displace traditional game engine simulation components over a 7–10 year horizon, but this displacement thesis has significant execution risk and is not evidenced by any current customer commitment. SV014, SV015
CV041 Strong momentum indicators as of June 2026 — including three major product launches, a $300 million strategic raise, and public backing from Toyota, Adobe, Nvidia, and AWS — support a track rather than pass recommendation, preserving optionality on a potentially generational world model platform. SV003, SV027, SV001
CV042 EU AI Act and copyright regulation exposure (detailed in Chapter 7) could add meaningful compliance and litigation costs that are not currently provisioned for in any public Decart financial disclosure, creating a downside financial risk that is difficult to quantify but material at scale. SV012, SV020
CV043 The combination of $450 million+ in raised capital, a defensive AWS infrastructure partnership, and a multi-vertical product suite covering gaming, e-commerce, and AV creates an asymmetric return profile — in the bull case the upside is substantial; in the base case the current $4B price embeds 50%+ downside risk. SV001, SV014
CV044 Alpha Partners' published investment memo for Decart characterizes the company as targeting "physical AI's simulation layer," framing the TAM as displacing incumbent physics simulation software across games, AV, and robotics; this framing is credible technically but the commercial timeline is highly uncertain. SV022, SV029
来源
编号出版方标题引文
SO001 Decart Decart AI Lab | Real-Time World Models
SO002 Decart Decart Raises $300M: Tech Leaders Back the Company as Both Customers and Investors With funding led by Radical Ventures, Decart is building the infrastructure layer for the next generation of low-latency AI systems, through three product lines: DOS … Lucy … and Oasis.
SO003 Yahoo Finance / Reuters Exclusive: Decart raises $100 million at a $3.1 billion valuation, chasing the future of real-time creative AI
SO004 Ctech (Calcalist) Nvidia backs Israeli AI unicorn Decart in $300 million funding round at $4 billion valuation Decart, founded in late 2023 by Dr. Dean Leitersdorf (CEO) and Moshe Shalev (CPO), both veterans of Israel's elite Unit 8200, plans to launch new versions of its three core product lines.
SO005 Ctech (Calcalist) From stealth to $3.1 billion in less than a year Decart … has secured $100 million in new funding at a $3.1 billion valuation, just two years after it was founded. It now counts 60 employees and a growing global presence.
SO006 TechCrunch Decart's AI simulates a real-time, playable version of Minecraft Decart doesn't say that it got Microsoft's blessing to train on footage of Minecraft. (Microsoft owns Minecraft.) Is Oasis basically creating an unauthorized copy of Minecraft?
SO007 TechCrunch Decart's new world model can simulate hours of photorealistic driving — with some caveats The car will just drive through other cars, meaning the model doesn't simulate physics properly … the environment looked less like New York and more like a standard version of any urban, Western city.
SO008 Decart Models | Decart API Platform
SO009 Decart Pricing | Decart API Platform $0.02/sec … Oasis 3 Preview is billed at $0.02 per second.
SO010 Decart Lucy 2.1 | Decart API Platform
SO011 Decart Oasis 3 | Decart API Platform
SO012 Decart Decart platform llms.txt
SO013 Decart Decart Optimization Engine
SO014 Decart Decart API Platform (developer console)
SO015 Decart and Etched Oasis: A Universe in a Transformer We're excited to announce Oasis, the first experiential, realtime, open-world AI model. … Oasis takes in user keyboard input and generates a real-time experience.
SO016 Alpha Partners Alpha invests in Decart, a developer of real-time generative AI
SO018 SiliconANGLE Decart raises $100M on $3.1B valuation to grow real-time AI video platform
SO019 The SaaS News Decart Raises $100M Series B at $3.1B Valuation Sequoia Capital, Benchmark, and Zeev Ventures—all existing investors—participated in the round, with Aleph VC, a new investor, also joining. This brings Decart's total funding to $153 million.
SO020 Electronics Weekly Decart raises $300M (Trainium AINews coverage) Decart is essentially going all-in on AWS … the company will also make its models available through the Amazon Bedrock platform.
SO021 JNS (Jewish News Syndicate) Israeli AI firm Decart raises $300 million at $4 billion valuation Decart, based in Tel Aviv and founded in 2023 by former Israeli military intelligence Unit 8200 veterans Dean Leitersdorf and Moshe Shalev …
SO022 Ynetnews From 15 to 60 employees and a $3.1B valuation, Decart raises $100M Decart has grown from 15 to more than 60 employees, with offices in northern Israel, San Francisco, and New York. … having used less than $10 million of its total funding to date.
SO023 Ynetnews Decart - the Leitersdorf-Shalev profile and Oasis launch retrospective Within just a week of its launch, the company reached 5 million users with its generative AI demo Oasis.
SO024 Ctech (Calcalist) Decart pledges millions to Technion in strategic AI push
SO025 Ctech (Calcalist) Decart launches real-time AI tool for live video transformation (Mirage)
SO026 Decart Oasis - Decart's first world model
SO027 Decart Decart API Platform Overview
SO028 Decart Decart API Platform FAQ
SO029 Decart Decart API Terms of Service
SO030 Startup Fortune Decart opens its world model to developers for two cents a second betting it can own physical AI's simulation layer Decart launched Oasis 3 on June 10, a real-time world model that renders photorealistic driving environments on demand via public API, priced at $0.02 per second.
SM001 Decart Decart AI Lab | Real-Time World Models
SM002 Decart Decart Raises $300M: Tech Leaders Back the Company as Both Customers and Investors Decart is building the infrastructure layer for the next generation of low-latency AI systems, through three product lines: DOS … Lucy … and Oasis.
SM003 Yahoo Finance / Reuters Exclusive: Decart raises $100 million at a $3.1 billion valuation, chasing the future of real-time creative AI
SM004 Ctech (Calcalist) Nvidia backs Israeli AI unicorn Decart in $300 million funding round at $4 billion valuation
SM005 Ctech (Calcalist) From stealth to $3.1 billion in less than a year
SM006 TechCrunch Decart's AI simulates a real-time, playable version of Minecraft Decart doesn't say that it got Microsoft's blessing to train on footage of Minecraft.
SM007 TechCrunch Decart's new world model can simulate hours of photorealistic driving — with some caveats Decart joins Google DeepMind (Genie 3), World Labs (Marble), Runway and Luma in the race to build real-time world models … the car will just drive through other cars, meaning the model doesn't simulate physics properly.
SM008 Decart Models | Decart API Platform
SM009 Decart Pricing | Decart API Platform $0.02/sec
SM010 Decart Lucy 2.1 | Decart API Platform
SM011 Decart Oasis 3 | Decart API Platform
SM012 Decart llms.txt | Decart Platform Docs
SM013 Decart Decart Optimization Stack | Decart.AI
SM014 Decart Decart Platform Console
SM015 Oasis Model (GitHub Pages) Oasis - The first playable AI-generated open-world game
SM016 Alpha Partners Alpha invests in Decart - a developer of real-time generative AI
SM017 SiliconAngle Decart raises $100M at $3.1B valuation to grow real-time AI video platform
SM018 The SaaS News Decart raises $100M Series B at $3.1B valuation
SM019 Electronics Weekly Decart raises $300m
SM020 The Business Research Company (TBRC) Generative AI In Gaming Market Report 2026 Generative AI In Gaming market size has reached to $1.79 billion in 2025 … expected to grow to $5.09 billion in 2030 at a compound annual growth rate (CAGR) of 23.2%.
SM021 Research and Markets Generative AI in Gaming Market Size & Forecast to 2030 It will grow from $1.79 billion in 2025 to $2.21 billion in 2026 at a CAGR of 23.1% … grow to $5.09 billion in 2030 at a CAGR of 23.2%.
SM022 Newzoo Games Market Reports and Forecasts | The PC & Console Gaming Report 2026
SM023 BCG (Boston Consulting Group) Video Gaming Report 2026: The Next Era of Growth We project that global revenues for cloud gaming will grow from around $1.4 billion in 2025 to roughly $18.3 billion in 2030 … above 50% CAGR.
SM024 McKinsey QuantumBlack The State of AI
SM025 U.S. Copyright Office Copyright and Artificial Intelligence
SM026 European Parliament (EPRS) Copyright and generative artificial intelligence (At a Glance, March 2026) The own-initiative report on copyright and AI calls for the respect of existing rules and the adoption of new rules that require AI developers to be fully transparent about their use of copyrighted works.
SM027 Osborne Clarke EU copyright law and generative AI: a watershed moment Pending the introduction of an appropriate provision, the Commission should establish an immediate, simple, flat-rate copyright fee of 5 to 7% of global turnover.
SM028 Presenc AI AI Policy and Regulation Tracker 2026 California and Texas state laws took effect on 1 January 2026, the EU AI Act's General Purpose AI obligations have been operational since August 2025.
SM029 GDC / Informa 2026 State of the Game Industry Report What impact is generative AI having on game development? … The 2026 State of the Game Industry Report answers these questions and more.
SP001 Google DeepMind Genie 3: A new frontier for world models Genie 3 is our first world model to allow interaction in real-time, while also improving consistency and realism compared to Genie 2.
SP002 TechCrunch AI video startup Runway raises $315M at $5.3B valuation, eyes more capable world models Runway has raised a $315 million Series E round, nearly doubling its valuation to $5.3 billion.
SP003 TechCrunch World Labs lands $1B, with $200M from Autodesk, to bring world models into 3D workflows Fei-Fei Li's World Labs has secured a $200 million investment from software design giant Autodesk as part of a larger $1 billion round.
SP004 World Labs World Labs — World Models
SP005 Runway New Funding to Scale World Simulation This capital lets us pre-train the next generation of world models and bring them to new products and industries.
SP006 Inworld AI Inworld AI — The Status reached 1M users in 19 days.
SP007 GamesBeat Inworld AI raises $50M round at $500M valuation for AI game characters
SP008 Rosebud AI Rosebud AI — Make 3D Games and Worlds with Vibe Coding
SP009 Scenario Scenario — The Creative AI Infrastructure Trusted by 15,000+ customers and the world's top creative teams; 550+ Models.
SP010 CNBC Luma AI raises $900 million in funding round led by Saudi AI firm Humain
SP011 Decart Decart — Real-Time AI Platform
SP012 TechCrunch Decart's new world model can simulate hours of photorealistic driving — with some caveats By letting you generate a world for so long, the model also degrades significantly.
SP013 TechCrunch Decart's AI simulates a real-time playable version of Minecraft
SP014 Decart API Platform Models — Decart API Platform
SP015 Decart API Platform Pricing — Decart API Platform Lucy 2.1 $0.02/sec Realtime video editing (latest)
SP016 Decart API Platform Lucy 2.1 — Decart API Platform
SP017 Decart API Platform Oasis 3 Preview — Decart API Platform
SP018 Oasis Model Oasis — A Universe in a Transformer
SP019 Decart Decart Raises $300M: Tech Leaders Back the Company as Both Customers and Investors Decart achieves more than a 100x improvement in cost efficiency and delivers real-time, live AI performance 8x faster than any comparable system on equivalent hardware.
SP020 SiliconAngle Decart raises $100M at $3.1B valuation to grow real-time AI video platform
SP021 BusinessWire Luma AI Raises $900 Million Series C Led by HUMAIN and Partners on 2 Gigawatt AI Supercluster in Saudi Arabia
SP022 Startup Fortune Decart opens its world model to developers for two cents a second Waymo already has a world model. The company unveiled its own generative simulation architecture in February, built on top of Google DeepMind's Genie 3, and positioned it as a core training tool for the Waymo Driver.
SP023 World Labs World Labs Announces New Funding World Labs has raised $1 billion in new funding.
SP024 Decart Decart Optimization Stack
SP025 Crunchbase Rosebud AI — Crunchbase Company Profile and Funding
SP026 Yahoo Finance / Fortune Exclusive: Decart raises $100 million at a $3.1 billion valuation, chasing the future of real-time creative AI despite raising $153 million in the last 11 months, Decart has spent less than $10 million of its investors' money.
SP027 Electronics Weekly Decart raises $300M (May 2026)
SI001 Calcalist / CTech Decart raises $300M; adds Amazon and NVIDIA as customers and investors
SI002 AI News Decart uses AWS Trainium3 for real-time video generation
SI003 Amazon Web Services AWS and NVIDIA deepen strategic collaboration to accelerate AI from pilot to production
SI004 NVIDIA NVIDIA and AWS expand strategic partnership at re:Invent
SI005 JNS Israeli AI firm Decart raises $300 million at $4 billion valuation
SI006 Gunder 2026 AI laws update: key regulations and practical guidance
SI007 Calcalist / CTech From stealth to $3.1 billion in less than a year While many AI companies burn investor cash on compute, Decart claims to have spent less than $10 million of its $153 million in capital. Revenue from GPU acceleration and video licensing pays for the rest.
SI008 SaaS News Decart raises $100M Series B at $3.1B valuation
SI009 Decart Decart Platform Pricing — pay-as-you-go pricing Decart uses pay-as-you-go pricing. Realtime models: billed per second of active generation. Lucy 2.1 $0.02/sec. Lucy Restyle 2 $0.01/sec. Oasis 3 Preview $0.02/sec.
SI010 Decart Decart Platform Overview
SI011 Decart Decart Platform Terms of Service
SI012 Decart Decart Platform FAQ
SI013 Decart Decart Platform llms.txt machine-readable index
SI014 Decart Decart Developer Platform Home
SI015 Decart Decart Oasis — World Model for Physical AI
SI016 Decart Decart Platform Use Cases
SI017 Decart Decart Platform E-commerce Virtual Try-On Guide
SI018 Decart Decart Platform Quickstart Guide
SI019 Decart Decart Platform Privacy Policy
SI020 Decart Decart Platform Acceptable Use Policy
SI021 Decart Decart raises $300M — tech leaders back the company as both customers and investors The company said it is already generating revenue through contracts with several of the world's largest cloud providers, AI laboratories, and hyperscale companies.
SI022 TechCrunch Decart's new world model can simulate hours of photorealistic driving — with some caveats The startup's models are so efficient, per Leitersdorf, that it has burned through "drastically less" than $100 million in its lifetime. [...] The car will just drive through other cars, meaning the model doesn't simulate physics properly in the environment.
SI023 Ynet News Decart CEO interview: less than $10M spent, millions in revenue, nearing profitability Leitersdorf said the company is already generating millions in revenue and is nearing profitability, having used less than $10 million of its total funding to date.
SI024 Ynet News Decart raises $300M at $4B valuation with NVIDIA and Amazon as investors
SI025 Calcalist / CTech Decart raises $100M at $3.1B valuation — third round in eleven months
SI026 Fortune / Yahoo Finance Exclusive: Decart raises $100 million Series B; CEO says company barely spent investor money Despite raising $153 million to date, the company says it has used less than $10 million of investor capital. The company is funding its tens of millions in GPU compute costs through revenue.
SI027 StartupFortune Decart opens its world model to developers for two cents a second
SI028 SiliconAngle Decart raises $100M on $3.1B valuation to grow real-time AI video platform
SI029 U.S. Securities and Exchange Commission SEC Form D — JSL Decart.AI Coinvest, L.P. (CIK 0002084011) JSL Decart.AI Coinvest, L.P. — Delaware limited partnership. Filed September 19, 2025.
SI030 Electronics Weekly Decart raises $300M
SE001 Decart API Platform Models - Decart API Platform
SE002 Decart API Platform Lucy 2.1 Realtime - Decart API Platform
SE003 Decart API Platform Oasis 3 Preview Realtime - Decart API Platform
SE004 Decart API Platform Use Cases - Decart API Platform
SE005 Decart API Platform Quickstart - Decart API Platform
SE006 Decart API Platform Privacy Policy - Decart API Platform
SE007 Decart Status Decart Status
SE008 Decart API Platform Acceptable Use Policy - Decart API Platform
SE009 Decart AI Optimizations Engine
SE010 TechCrunch Decart's new world model can simulate hours of photorealistic driving — with some caveats
SE011 TechCrunch Decart's AI simulates a real-time, playable version of Minecraft
SE012 Decart AI Decart AI Lab | Real-Time World Models
SE013 Decart API Platform Decart API Platform
SE014 Decart API Platform Overview - Decart API Platform
SE015 Decart API Platform FAQ - Decart API Platform
SE016 Decart API Platform Terms of Service - Decart API Platform
SE017 Decart AI Decart AI Lab | Oasis
SE018 Decart and Etched Oasis: A Universe in a Transformer
SE019 Decart AI Decart Raises $300M: Tech Leaders Back the Company as Both Customers and Investors
SE020 Decart API Platform Pricing - Decart API Platform
SE021 AI News Decart uses AWS Trainium3 for real-time video generation
SE022 Startup Fortune Decart opens its world model to developers for two cents a second betting it can own physical AI's simulation layer before the big players build one themselves
SE023 SiliconANGLE Decart raises $100M on $3.1B valuation to grow real-time AI video platform
SE024 GitHub DecartAI organization
SE025 GitHub DecartAI/decart-python: Python SDK for Decart AI
SE026 GitHub DecartAI/Decart-XR: The First RealTime World Transformation App for Quest
SE027 AWS Machine Learning Blog AWS and NVIDIA deepen strategic collaboration to accelerate AI from pilot to production
SE028 CTech Nvidia backs Israeli AI unicorn Decart in $300 million funding round at $4 billion valuation
SE029 Tech Startups AWS and Decart team up to unlock the potential of real-time AI video
SE030 Google DeepMind Genie 3: A new frontier for world models
SE031 Decart API Platform Authentication - Decart API Platform
SE032 Decart API Platform API Terms of Service - Decart API Platform
SE033 Decart API Platform SDK Direct - Decart API Platform
SU001 Decart AI Decart Raises $300M: Tech Leaders Back the Company as Both Customers and Investors
SU002 Decart API Platform E-commerce Virtual Try-On - Decart API Platform
SU003 Decart API Platform Use Cases - Decart API Platform
SU004 Decart API Platform Pricing - Decart API Platform
SU005 Decart API Platform Models - Decart API Platform
SU006 TechCrunch Decart's new world model can simulate hours of photorealistic driving — with some caveats
SU007 CTech Nvidia backs Israeli AI unicorn Decart in $300 million funding round at $4 billion valuation
SU008 Startup Fortune Decart opens its world model to developers for two cents a second betting it can own physical AI's simulation layer before the big players build one themselves
SU009 AWS Machine Learning Blog AWS and NVIDIA deepen strategic collaboration to accelerate AI from pilot to production
SU010 AI News Decart uses AWS Trainium3 for real-time video generation
SU011 Decart AI Decart AI Lab | Real-Time World Models
SU012 Decart API Platform Overview - Decart API Platform
SU013 SiliconANGLE Decart raises $100M on $3.1B valuation to grow real-time AI video platform
SU014 The SaaS News Decart Raises $100M Series B at $3.1B Valuation
SU018 Electronics Weekly Decart raises $300m
SU019 CTech From stealth to $3.1 billion in less than a year
SU020 Decart and Etched Oasis: A Universe in a Transformer
SU021 TechCrunch Decart's AI simulates a real-time, playable version of Minecraft
SU022 JNS Israeli AI firm Decart raises $300 million at $4 billion valuation
SU023 Ynet The Israeli AI unicorn that wants to replace Netflix, YouTube and TikTok
SU024 Google DeepMind Genie 3: A new frontier for world models
SU025 CTech Decart hits $3.1 billion valuation on $100 million raise to power real-time interactive video
SU026 Decart API Platform FAQ - Decart API Platform
SU028 Decart API Platform Terms of Service - Decart API Platform
SU029 GitHub DecartAI organization
SU030 GitHub DecartAI/decart-python: Python SDK for Decart AI
SU031 GitHub DecartAI/Decart-XR: The First RealTime World Transformation App for Quest
SU032 Tech Startups AWS and Decart team up to unlock the potential of real-time AI video
SU033 TechBriefly Decart partners with AWS to boost real-time video AI capabilities
SU034 bonega.ai AWS and Decart Build the First Real-Time AI Video Infrastructure
SU035 Decart API Platform Client Tokens - Decart API Platform
SU036 Decart API Platform HTTP Signaling - Decart API Platform
SU037 Decart API Platform WS Signaling Proxy - Decart API Platform
SU038 Decart API Platform Overview - Decart API Platform (Python SDK)
SU039 Decart API Platform Overview - Decart API Platform (Android SDK)
SU040 Decart API Platform Realtime Mobile App - Decart API Platform
SR001 U.S. Copyright Office Artificial Intelligence and Copyright The Copyright Office is examining questions about the use of copyrighted works to train AI models and the copyright status of AI-generated works.
SR002 TechCrunch Decart's AI simulates a real-time playable version of Minecraft
SR003 TechCrunch Decart's new world model can simulate hours of photorealistic driving — with some caveats In my testing, the car will just drive through other cars, meaning the model doesn't simulate physics properly in the environment. Leitersdorf calls this a 'major research problem that we're cracking now.'
SR004 Calcalist Tech Decart hits $3.1 billion valuation on $100 million raise to power real-time interactive AI
SR005 Calcalist Tech Amazon joins as strategic customer as Decart raises $300M at $4B valuation
SR006 Osborne Clarke EU copyright law and generative AI — a watershed moment The EU Parliament now calls for a direct solution — the principle of territoriality must be adapted so that the use of European content is subject to EU law even when training takes place outside the EU.
SR007 Gunder LLP 2026 AI Laws Update: Key Regulations and Practical Guidance Several states have enacted or finalized broad AI governance statutes that impose affirmative risk management, documentation, and oversight obligations for certain high-impact AI systems, with enforcement beginning in late 2025 and 2026.
SR008 Presenc.ai AI Policy and Regulation Tracker 2026
SR009 Decart Terms of Service
SR010 Decart Acceptable Use Policy This AUP is designed to ensure safe and responsible online environment, to protect individuals and users right, to uphold applicable legal obligations, including without limitations the EU Digital Service Act (DSA), the EU AI Act (AIA), Digital Millennium Copyright Act (DMCA) and any other applicable regulatory requirements.
SR011 Decart Privacy Policy
SR012 Decart Decart AI — Official Website
SR013 Decart Decart Raises $300M — Tech Leaders Back the Company as Both Customers and Investors The funding round was led by Radical Ventures and included participation from chip giant NVIDIA. Additional investors included eBay Ventures, Adobe Ventures, Toyota Ventures, Atreides Management, and Valor Equity Partners, alongside existing investors Sequoia Capital, Zeev Ventures, and Benchmark.
SR014 Yahoo Finance / Reuters EXCLUSIVE: Decart raises $100 million in funding
SR015 SiliconANGLE Decart raises $100M on $3.1B valuation to grow real-time AI video platform
SR016 Electronics Weekly Decart raises $300M in 2026
SR017 AI News Decart uses AWS Trainium3 for real-time video generation Decart is essentially going all-in on AWS, and as part of the deal, the company will also make its models available through the Amazon Bedrock platform.
SR018 Startup Fortune Decart opens its world model to developers for two cents a second Waymo already has a world model built on top of Google DeepMind's Genie 3. Decart's real market is everyone below that top tier: the dozens of AV programs, robotics labs, and drone startups that want frontier-grade simulation without a DeepMind-sized research budget attached.
SR019 GitHub Pages (Oasis Project) Oasis — A Universe in a Transformer
SR020 Decart DOS — Decart Optimization Stack
SR021 Decart Platform Docs Models Overview
SR022 JNS Israeli AI firm Decart raises $300 million at $4 billion valuation
SR023 AWS AWS and NVIDIA deepen strategic collaboration to accelerate AI from pilot to production
SR024 Decart Platform Docs Oasis 3 Model Reference
SR025 Decart Platform Docs Lucy 2.1 Model Reference
SR026 Google DeepMind Genie 3: A new frontier for world models
SR027 European Parliament Research Service Copyright and Generative Artificial Intelligence — Opportunities and Challenges The Commission should be called on to propose the establishment of a rebuttable presumption that, for any generative AI model placed on the EU market, works and other subject matter protected by copyright have been used for training, where the transparency obligations set out in the resolution have not been fully complied with.
SR028 U.S. Copyright Office Copyright and Artificial Intelligence — Part 3 Pre-Publication
SR029 Calcalist Tech Decart company profile and strategy
SR030 Decart Platform Docs Pricing
SR031 The SaaS News Decart Raises $100M Series B at $3.1B Valuation
SR032 Nvidia Blog AWS Partnership Expansion: Reinventing AI Infrastructure
SR033 Decart Platform Docs Platform Status
SR034 AIAAIC AIAAIC Repository — AI, Algorithmic, and Automation Incidents and Controversies
SR035 WIPO Intellectual Property and Artificial Intelligence — WIPO Policy Overview
SR036 European Commission EU AI Act Policy Page
SR037 National Law Review 2026 AI Regulatory Landscape — Key Developments Enterprises Must Know
SR038 Wired The Copyright Fight Over AI Training Data
SR039 U.S. Copyright Office Copyright and Artificial Intelligence, Part 1: Digital Replicas
SR040 U.S. Copyright Office Copyright and Artificial Intelligence, Part 2: Copyrightability
SV001 Calcalist Decart Raises $300 Million at $4 Billion Valuation
SV002 Electronics Weekly Decart Raises $300M in May 2026
SV003 Decart Decart Raises $300M — Tech Leaders Back the Company as Both Customers and Investors
SV004 SiliconAngle Decart Raises $100M at $3.1B Valuation to Grow Real-Time AI Video Platform
SV005 TechCrunch AI Video Startup Runway Raises $315M at $5.3B Valuation, Eyes More Capable World Models
SV006 Runway Runway Series E Funding
SV007 TechCrunch World Labs Lands $200M from Autodesk to Bring World Models into 3D Workflows
SV008 World Labs World Labs 2026 Funding Announcement
SV009 CNBC Luma AI Raises $900 Million in Funding Led by Saudi AI Firm HUMAIN
SV010 BusinessWire Luma AI Raises $900 Million Series C Led by HUMAIN
SV011 GamesBeat Inworld AI Raises New Round at $500M Valuation for AI Game Characters
SV012 TechCrunch Decart's New World Model Can Simulate Hours of Photorealistic Driving — With Some Caveats
SV013 Yahoo Finance Exclusive — Decart Raises $100 Million
SV014 Boston Consulting Group Video Gaming Report 2026: The Next Era of Growth
SV015 The Business Research Company Generative AI in Gaming Global Market Report
SV016 Research and Markets Global Generative AI in Gaming Market Report
SV017 U.S. Securities and Exchange Commission EDGAR Full-Text Search — Roblox 10-K Annual Reports
SV018 McKinsey and Company The State of AI (QuantumBlack / McKinsey Global Survey)
SV019 Statista Mobile Games — Worldwide Market Outlook
SV020 CBInsights Decart — Products, Competitors, Financials, Employees
SV021 Newzoo Global Games Market Will Generate $196B+ in 2024
SV022 Blog.Amy.VC Decart's $100M Funding Round
SV023 Calcalist Decart Raises $100M at $3.1B Valuation
SV024 Calcalist Decart Background and Context Article
SV025 Decart Decart Company Website
SV026 Decart Platform Docs Platform Pricing
SV027 AI News Decart Uses AWS Trainium3 for Real-Time Video Generation
SV028 TechCrunch Decart's AI Simulates a Real-Time Playable Version of Minecraft
SV029 Startup Fortune Decart Opens Its World Model to Developers for Two Cents a Second
SV030 Ynetnews Israeli AI Firm Decart Raises $300M
SV031 World Labs World Labs Website
SV032 Inworld AI Inworld AI Website
SV033 Decart Platform Docs Gaming Examples
SV034 Tech in Asia Nvidia Joins Decart's $300M, $4B Valuation Round
SV035 AWS Startups Decart Real-Time AI Simulation