初创公司尽调
尽调报告 AI / embodied intelligence / foundation models Pre-product unicorn (angel/seed) 2026-07-23

Prague Technology

创始人履历罕见,市场时点也站上风口;以全球 world model 估值区间低端切入有吸引力,但公司仍完全处于产品前阶段,叠加中国风险溢价和 GPU 出口管制,更适合中性而非买入。

封面要素

估值 01
$2B post-money USD [CV001]
累计融资 02
$220M angel round USD [CV002]
成立时间 03
May 2026 [CO002]
重点方向 04
Embodied intelligence, world models, physical AI [CO010]

公司概况

Prague Technology(上海卜拉格科技有限公司)是一家中国 AI 初创公司,聚焦具身智能、世界模型,以及面向物理系统和机器人的多模态基础模型。公司由林俊阳于 2026 年 5 月创立;他曾带领 Alibaba 的 Qwen 大语言模型系列成为中国最强开源 LLM。公司从高榕创投、HongShan Capital(Sequoia China)和 Tencent 融资约 $220 million,投后估值 $2 billion,跻身中国估值最高的早期 AI 公司之一。截至 2026 年 7 月,公司没有商业产品、没有收入,也没有披露客户。

官网
www.pragmatics-ai.com
成立时间
2026-05-27
创始人
Lin Junyang (林俊旸)
创立地点
Shanghai, China
总部
Xuhui District, Shanghai, China
产品
截至 2026 年 7 月没有商业产品;公司正在开发世界模型和具身 AI 基础模型技术,目标应用于机器人和物理 AI 场景。
客户
中国工业机器人 OEM 和物理 AI 平台提供商(目标客户,未确认)
商业模式
基础模型授权与具身 AI 服务——模式尚未定义;参考可比公司,收入落地还要等数年。
阶段
Angel/seed — pre-product, pre-revenue
融资情况
2026 年 6 月完成 $220M 天使轮,投后估值 $2B;投资方包括高榕创投($100M)、HongShan Capital($100M)和 Tencent($20M)。
[CO001, CO002, CI001, CI002, CV001, CV002]

执行摘要

主要优势

  • Lin Junyang 是中国最可信的 AI 模型构建者之一,曾主导 Qwen LLM 系列架构
  • 相比可比的收入前 world model 实验室 AMI Labs($3.5B)和 World Labs($5B),$2B 估值偏保守
  • 顶级中国 VC 注入 $220M 资本,可支撑 24+ 个月深度研究现金跑道
  • 2026 年全球 AI 投资中,具身 AI 和 world model 是增长最快的方向之一
  • 国内市场足够大:中国工业机器人赛道天然适合率先采用具身 AI

主要风险

  • 产品前、收入前,2026 年 5 月成立——估值几乎全靠创始人信号支撑,没有产品验证
  • 美国出口管制限制 GPU 获取,迫使公司依赖性能较弱的国产芯片,训练能力面临急性风险
  • 投资人全部是中国 VC,限制美国 / 欧盟 IPO 或战略并购选择权
  • 中国生成式 AI 监管(MIIT Interim Measures)带来合规负担,也可能形成内容审查约束
  • 资金充足的同行持续施压:TARS、Zhiyuan、AMI Labs、Physical Intelligence 和 World Labs
  • Lin Junyang 之外的核心团队构成未披露;关键具身 AI 硬件人才招聘尚未验证

未决问题

  • 股权结构表和清算优先权未公开披露——隐含约 11% 稀释只是估算
  • 技术架构规格及其与 Qwen 模型的差异化未公开
  • 美国出口管制下的 GPU 算力获取方案未披露
  • Lin Junyang 之外的团队构成未披露——关键人才风险突出
  • Physical Intelligence 和 TARS Group 的准确估值为分析师估算,尚未确认

目录

Chapter 01

01公司概览

1.1 身份与创立

Prague Technology 是林俊阳 2026 年中在中国上海创立的 AI 实验室英文商号。公司通过多个法律实体运营:上海卜拉格科技有限公司(Shanghai Bulage Technology Co., Ltd.)于 2026 年 5 月 27 日注册,林俊阳持股 99%,并担任执行董事、总经理和财务负责人;语用(上海)科技有限公司(Yuyong Shanghai Technology Co., Ltd.)于 2026 年 5 月 13 日注册,林俊阳持股 100%。第三个实体上海格物致用管理咨询合伙企业由卜拉格科技控制。命名是有意的语言双关:卜拉格(Bulage)是 “Pragmatics” 的音译,语用(Yuyong)是 “Pragmatics” 的中文意译——两者都指向语用学,也呼应林俊阳的学术背景。公司总部位于上海徐汇区。截至报告日期(2026 年 7 月),公司没有官方网站、没有披露产品,也没有报告收入。公司公开披露的重点方向是世界模型,以及面向物理 AI 系统的「具身大脑」。林俊阳在 2026 年 3 月 3–4 日因内部重组会议离开 Alibaba,并在社交媒体发布 “me stepping down. bye my beloved qwen”。经过短暂复盘和战略规划后,他在 2026 年 5–6 月成立创业实体。[CO001, CO002, CO003, CO004, CO005, CO006]

Prague Technology 关键快照指标
指标数值 / 状态日期置信度备注 / 缺口
法定名称(主要)上海卜拉格科技有限公司(Shanghai Bulage Technology Co., Ltd.)2026-05-27已通过 Qichacha / 多个新闻来源确认注册
法定名称(次要)语用(上海)科技有限公司 (Yuyong Shanghai Technology)2026-05-13Lin Junyang 100% 持股;姊妹主体
英文商号Prague Technology / Pragmatics Technology2026国际媒体两个名称都在使用
总部中国上海市徐汇区2026注册记录确认
成立时间2026 年 5–6 月(主体注册)2026-05-27Lin 于 2026 年 3 月离开 Alibaba
创始人 / CEOLin Junyang (林俊旸)2026主要主体 99% 股东
投后估值约 $2 billion (RMB 13.5B)2026-06多个来源相互印证
融资总额约 $220 million2026-06投资方:Gaorong $100M + HongShan $100M + Tencent $20M
收入未披露收入2026-07产品前公司;无商业牵引
员工初期 1–10 人;正在扩张2026-06未披露官方员工数
产品状态尚未发布产品2026-07公司处于隐身研究阶段
网站未公开可用2026-07未确认官方域名
赛道AI / 具身智能 / 世界模型2026投资人沟通已确认
阶段收入前天使 / 种子轮2026-06中国同阶段创纪录高估值

除特别说明外,所有金额均为美元。估值基于第三方报道,未经审计。员工人数来自早期报道估算。

[CO001, CO002, CO016, CO017, CO025]
FO001: 公司快照逻辑

展示 Lin Junyang 控制下的三个实体公司结构,以及主要实体(Bulage Technology)如何连接到姊妹 Yuyong 实体和管理咨询合伙企业。

公司架构基于多家新闻源报道的 Qichacha 注册数据。合伙企业的具体持股比例未公开披露。

[CO001, CO002, CO003, CO004]

1.2 创始人与领导层

林俊阳(Justin Lin)是 Prague Technology 唯一公开可知的创始人和负责人。他出生于 1993 年,本科毕业于国际关系学院英语专业,硕士毕业于北京大学外国语学院语言学与自然语言处理方向。2019 年,他以应届毕业生身份加入 Alibaba DAMO Academy,最初负责搜索和推荐场景中的 NLP,以及包括 M6 在内的早期多模态项目。2022 年底,Alibaba 将 AI 团队重组为通义实验室体系,林俊阳出任通义千问(Qwen)大语言模型系列技术负责人。在他带领下,Qwen 从内部项目成长为全球下载量最高的开源 LLM 家族之一:到 2026 年初,全球下载量超过 10 亿次,衍生模型超过 200,000 个。他也成为 Alibaba 史上最年轻的 P10 级技术高管;P10 代表个人技术贡献者的最高等级。2026 年 3 月 3 日,他得知拟议重组将把 Qwen 团队拆分为预训练、后训练、文本、图像和语音等独立团队,团队组织独立性将被取消。次日上午,他宣布离职。在后续文章《From Reasoning Thinking to Agentic Thinking》中,他提出下一代 AI 范式将围绕 agentic systems 展开——AI 不只是思考,还要行动。仍在 Alibaba 任职时,林俊阳已于 2025 年 10 月组建一个小型内部团队,聚焦机器人和具身智能,提前押注这一方向。除林俊阳外,创始团队规模很小,最初约 1 至 10 人,成员来自 ByteDance、Tencent 和国际背景。公司尚未公开披露其他具名高管或董事会成员。[CO008, CO009, CO010, CO011, CO012, CO013]

领导层与创始人表
人物职务背景创始人-市场匹配关键人物风险
Lin Junyang (林俊旸)创始人、CEO、99% 股东1993 年生;国际关系学院英语学士;北京大学语言学 / NLP 硕士;2019–2026 年 Alibaba DAMO Academy;Alibaba 最年轻 P10带领 Qwen 从零做到 10 亿+ 下载;内部组建多模态和机器人团队;发表具身 AI 论述文章严重——唯一公开创始人;若离职,公司将失稳
未具名早期团队未披露技术角色来自 ByteDance、Tencent、海外机构LLM 与具身系统的技术跨度可能较宽未知——除创始人外未披露姓名或职责

来源:多篇新闻报道,依据公司备案以及媒体对 Lin Junyang 的采访。次级团队构成未经确认。

[CO008, CO009, CO010, CO011, CO012]
FO002: 公司里程碑时间线

关键节点串起 Lin Junyang 的履历:从求学、在 Alibaba 带队,到创办 Prague Technology。

教育时间根据已知出生年份和中国标准学制估算;来源未明确确认。

[CO008, CO009, CO010, CO011, CO012, CO013]

1.3 融资历史与投资方联盟

Prague Technology 于 2026 年 6 月完成首轮外部融资,融资约 $220 million,投后估值约 $2 billion(人民币 135 亿元)。按绝对美元金额计算,这是中国 AI 初创公司天使或种子阶段的纪录级估值。The Information 最先报道该轮融资,多家中国和国际信源随后确认。领投方为高榕创投和 HongShan Capital(红杉中国,原 Sequoia China),各投资 $100 million。Tencent 作为战略联合投资方出资 $20 million。据两名熟悉交易的人士称,公司刚完成该轮融资后,已在寻求启动新一轮后续融资。融资轨迹反映出顶级中国 VC 之间的激烈竞争:在正式产品里程碑出现前,先锁定顶尖 AI 创始人的早期股权。高榕创投是总部位于上海的主要 VC,管理资金超过人民币 300 亿元,覆盖 300 多家公司,重点投向 AI 和机器人,过往案例包括 Moonshot AI、XYZ Robotics 和自动驾驶公司。HongShan(原 Sequoia China)管理资产 $55 billion,覆盖 1,500 多家被投公司,曾投资 DeepSeek、Pony AI、X Square Robot,以及 OpenAI、Anthropic 等全球 AI 龙头。Tencent 几乎投遍中国头部 AI 初创圈,包括 DeepSeek、MiniMax 和 Zhipu。该投资联盟传递的信号是,机构已形成共识:林俊阳在世界模型和具身智能上的创始人-市场匹配度,足以支撑一个尚未有产品公司的数亿美元下注。[CO016, CO017, CO018, CO019, CO020, CO021]

利益相关方 / 投资方地图
投资方类型金额(USD)所有权含义战略角色尽调问题
Gaorong Ventures (高榕创投)领投 VC$100M重要少数股权(具体 % 未披露)中国最大种子 / 早期 VC;300+ 家被投;具身 AI 专长确认董事席位、治理权、反稀释条款
HongShan Capital (红杉中国)领投 VC$100M重要少数股权(具体 % 未披露)原 Sequoia China;$55B AUM;全球 AI LP 网络;DeepSeek、OpenAI 投资方确认按比例跟投权、信息权、后续出资承诺
Tencent战略共同投资方$20M小额少数股权中国顶级科技集团;曾投资 DeepSeek、MiniMax、Zhipu;具备战略数据和基础设施价值确认 IP 与数据许可限制;厘清战略承诺
Lin Junyang(创始人)创始股权劳动股主要主体 99% 股份唯一创始人;关键人物依赖了解归属期安排和离职条款
未来未知投资方后续轮未披露TBD据报道,公司在天使轮交割后立即寻求新一轮融资确认新一轮融资时间表和条款

持股比例未公开披露。按 $220M 融资额和 $2B 估值测算,若投前估值为 $1.78B,总稀释约 11%,因此推定 Gaorong 和 HongShan 持少数股权。

[CO016, CO017, CO018, CO019, CO020]
FO003: 关键 KPI 快照

2026 年 6 月天使轮中的关键财务指标,并与中国 AI 初创公司同业对比。

创始人年龄根据披露的 1993 年出生年份和 2026 年 6 月融资日期计算。

[CO017, CO018, CO024]

1.4 关键里程碑与当前状态

Prague Technology 的里程碑集中在很短的创立窗口。林俊阳带领 Qwen 团队完成 Qwen 3.5(3970 亿参数 MoE 架构)后,于 2026 年 3 月 3–4 日离开 Alibaba。2026 年 3 月 26 日,他发布战略文章《From Reasoning Thinking to Agentic Thinking》。2026 年 5 月,The Information 首次报道他正以 $2 billion 估值融资。2026 年 5 月 13 日至 6 月,他注册了公司实体。到 2026 年 6 月中旬,天使轮已经交割。截至 2026 年 7 月 23 日报告日期,公司没有披露产品、路线图、网站或商业合作。公司没有报告客户或收入。团队仍处于早期组建和招聘阶段。公司没有公开演示,也没有以 Prague Technology 或卜拉格品牌申请任何专利或发布任何技术研究。公司的成熟度画像是前产品阶段的 AI 研究实验室:创始团队履历顶尖,资本已经到位,技术命题也已提出,但新品牌之下还没有经过验证的技术产出。没有公开网站或新闻稿,说明公司可能刻意保持隐身,符合研究优先的策略。林俊阳 2026 年 3 月在社交媒体称「多模态基础模型正在转向基础智能体,借助工具和记忆,通过强化学习做长期序列推理」,这是其技术方向最清晰的公开表述。[CO025, CO026, CO027, CO028, CO029, CO030]

里程碑表
日期事件类型金额 / 估值 / 状态参与方含义
Oct 2025Lin Junyang 在 Alibaba Qwen 内部组建小型机器人 / 具身智能团队产品内部项目Lin Junyang + 未具名 Alibaba 团队显示 Lin 在离开 Alibaba 前已把兴趣转向具身 AI
2025-Q4Qwen 3.5(397B 参数 MoE)发布,当时是 Alibaba 旗舰开源模型产品迄今 1B+ 次下载Lin Junyang 负责的 Alibaba Qwen 团队Lin 在 Alibaba 的最后一项重大贡献;验证其 LLM 资历
2026-03-03Lin Junyang 收到 CTO Zhou Jingren 提出的 Alibaba 重组方案治理N/ALin Junyang 与 Zhou Jingren(Alibaba Cloud CTO)离职导火索;方案拟将 Qwen 拆成独立横向团队
2026-03-04Lin Junyang 通过社交媒体公开宣布离开 Alibaba负面N/ALin Junyang震动中国 AI 圈;Alibaba 股价短暂下跌;市场担心人才外流
2026-03-26Lin 发表《From Reasoning Thinking to Agentic Thinking》一文产品N/ALin Junyang首次公开阐述其离开 Alibaba 后的战略判断;指向具身 AI 方向
2026-05The Information 报道 Lin 正以约 $2B 估值募资融资~$2B 目标估值Lin Junyang、Gaorong、HongShan(早期洽谈)获得国际关注;引发 VC FOMO;确认具身 AI + 世界模型焦点
2026-05-13语用(上海)科技有限公司注册(姊妹实体)设立RMB 250,000 注册资本Lin Junyang(100% 持有人)公司架构开始搭建
2026-05-27上海卜拉格科技有限公司注册(主体实体)设立RMB 600,000 注册资本Lin Junyang(99% 持有人)主要运营实体在徐汇区设立
2026-06Shanghai Gewuzhiyong Management Consulting Partnership 成立设立N/ALin Junyang 通过 Bulage Technology 担任 GP公司治理架构成形;为未来基金结构或合作关系预留位置
2026-06-15据 The Information 报道,天使轮($220M)正式完成融资$220M / 投后约 $2B投资方:Gaorong($100M)、HongShan($100M)、Tencent($20M)创下中国 AI 初创公司天使 / 种子轮估值纪录;验证创始人溢价
2026-06-15+立即开始寻求后续融资轮融资条款未披露Lin Junyang + 未具名投行 / VC资本策略激进;可能带来进一步稀释或新领投方
2026-07截至报告日期:无产品、无收入、无公开网站规模N/AN/A仍处产品前秘密研发阶段;投资人需要接受 18–36 个月研发周期

里程碑类型:设立、融资、产品、规模、监管、合作、治理、负面。除另有说明外,所有融资金额均为 USD。日期来自新闻来源,未经审计。

[CO007, CO008, CO014, CO016, CO025, CO026]

1.5 展示要点

Chapter 02

02市场分析

2.1 市场定义与范围

与 Prague Technology 相关的市场是具身智能——AI 系统借助人形机器人、工业机械臂、自动驾驶车辆、服务机器人等硬件载体,在物理环境中感知、推理并行动。它不同于纯软件 AI(部署在云端的大语言模型),也不同于按固定规则运行的传统预编程工业机器人。相关支出边界包括:(1)面向具身智能体的 AI 基础模型和世界模型软件;(2)完整人形机器人和移动操作机器人系统;(3)上游零部件(伺服电机、传感器、电池、执行器);(4)系统集成和部署服务;(5)用于模型训练的数据采集与仿真基础设施。Prague Technology 及其可触达客户需要替代的现状方案包括固定程序工业机器人、制造业人工劳动力,以及早于世界模型出现的 VLA(vision-language-action)模型。相邻市场包括自动驾驶汽车和无人机智能,两者存在技术交集,但监管制度和买方不同。中国具身 AI 行业由硬件公司主导,约 85% 的公司为民营,国资参与约 4% 的实体。软件 / AI 大脑层——Prague Technology 的目标位置——是公认瓶颈:全球只有约 100 家公司在积极开发具身大模型,而上游硬件供应商超过 8,000 家。中国国内市场区域集中度很高,广东、浙江、江苏、上海和北京合计占全部公司三分之二以上;2026 年 1–5 月,仅广东就贡献了行业总销售收入的 78.7%。[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 / 付款方对 Prague Technology 的意义
AI 大脑 / 世界模型软件模型授权、API 费用、微调服务无物理输出的纯云端 LLM SaaS人形机器人 OEM CTO、R&D 负责人核心:Prague Technology 计划切入的产品层
人形机器人系统完整人形机器人平台,包括计算、传感器、执行器传统固定程序工业机器人汽车 OEM、物流运营商间接:Prague AI 大脑的 OEM 客户
上游部件伺服电机、触觉传感器、电池、执行器非机器人专用的通用电子元件采购硬件的机器人制造商间接:决定供应链成本结构
集成与部署服务机器人 SI、部署、维护、培训通用 IT 咨询工厂自动化团队、设施经理间接:具身 AI 落地渠道
仿真与数据基础设施数字孪生、仿真平台(NVIDIA Isaac)、遥操作数据非机器人专用的通用云计算AI 实验室、机器人 OEM R&D 团队邻近:支撑 Prague 所需模型训练

Prague Technology 未披露产品或收入;所有意义分类均基于新闻报道中表述的公司重点和 Lin Junyang 的公开发言。

[CM001, CM002, CM003]
FM001: 市场规模口径

Prague Technology 的市场分层:从全球物理 AI 总市场(TAM),收窄到世界模型 / AI 大脑软件细分(SOM)。数值为分析师估算,置信区间很宽;来源细节见 TM002。

TAM 和 SAM 来自范围定义互不兼容的多组分析师估算。SOM 和公司收入属于推测;尚无分析师发布独立的世界模型授权市场估算。

[CM005, CM008, CM009]

2.2 市场规模:TAM、SAM 与 SOM

全球和中国具身 AI 市场有多套测算口径,结论差距很大——这反映的是定义分歧,而不是预测错误。最保守的自下而上估算基于当前人形与具身机器人部署,认为 2025 年全球具身 AI 市场规模为 $4.44 billion,并以 39% CAGR 增长,到 2030 年达到约 $23 billion。最宽口径来自 IDC:其将全球具身 AI 市场广义定义为包括自动驾驶汽车、工业自动化软件、智能制造等所有物理 AI 相邻类别,并预测 2030 年将达到 $1.5 trillion。中国国务院发展研究中心预测,仅中国国内具身 AI 市场到 2035 年就将超过 ¥1 trillion($140 billion)。这些估算无法直接比较:IDC 数字包含 EV 软件、制造业 IT 等多数从业者不会归为具身 AI 的支出类别。投资活动是市场动能最及时的指标;仅 2026 年第一季度,中国机器人与具身 AI 风投融资就达到 $3.3 billion,覆盖 126 笔交易,创下单季度历史新高。2026 年上半年,中国具身 AI 与机器人融资总额超过 ¥46 billion($6.4 billion),覆盖 288 起事件、226 家公司。全球范围内,2025 年机器人风投总额为 $13.8 billion;人形机器人专项投资在四年内增长约 143 倍。Prague Technology 的可服务可触达市场,是 AI 大脑和世界模型软件层。该层目前几乎没有披露的独立收入,但头部分析师将其描述为新兴栈中的主要瓶颈和价值最高节点。类比软件 AI 的基础模型——模型 API 提供商可捕获下游应用价值的 10–15%——到 2030 年,中国世界模型授权的非常粗略 SOM 估算为 $4–8 billion;但这个数字带有推测性,缺乏信源支持的确认。[CM008, CM009, CM010, CM011, CM012, CM013]

TAM/SAM/SOM 或规模测算视角表
发布方 / 来源年份地域数值CAGR方法置信度局限
ainchina.com / 市场综合口径2025全球$4.44B39%基于出货量和软件收入数据自下而上测算不含 AV 和非机器人实体 AI
IDC(经 China Daily,Jul 2026)2030 年预测全球$1.5TN/A自上而下;范围较宽,包含 AV 和制造业 IT范围远超传统具身 AI 定义
State Council DRC(经 ainchina)2035 年预测中国国内¥1T+ (~$140B)N/A政策挂钩预测;未披露方法政府目标,不是市场研究;9 年周期
Crunchbase / ainchina(Q1 2026 来源)Q1 2026中国投资$3.3B(季度)N/A126 笔交易的披露数据投资额 ≠ 收入;集中在人形硬件
Qixinbao / embodiedglobal.com(H1 2026 来源)H1 2026中国投资¥46B+ (~$6.4B)N/A披露融资事件(288 起,226 家公司)H1 数据;全年折算约 $12-13B
humanoid.guide(2025 全球 VC)2025全球$13.8B 风险投资4 年 143×汇总 PitchBook / Crunchbase 交易数据VC 融资,不是市场收入

各项估算因范围不同而不可比(投资额 vs. 收入、窄口径 vs. 宽口径、中国 vs. 全球)。没有单一估算能代表 Prague Technology 的可服务市场。

[CM008, CM009, CM010, CM011, CM014, CM015]
FM002: 市场估算区间

2025–2026 年全球具身 AI 年度投资或市场规模的估算区间,显示不同来源差异很大。所有数字均为 $B USD。

低 / 高边界代表分析师估算区间;“中值”是有来源时的单点估算。$23B 与 $1.5T 这两个 2030 年估算采用互不兼容的市场定义,不应取平均。

[CM008, CM009, CM037]

2.3 买方与细分格局

中国具身 AI 买方生态按产品类型和部署场景切分。工业机器人仍是收入最大细分,约占中国市场 45%,主要付费方是汽车 OEM(BYD、NIO、SAIC)和电子制造商。服务机器人约占 25%,面向零售、酒店和养老场景,预算权通常在设施运营或 IT 层级。人形机器人目前收入占比最小,约 5%,但增速最高,也吸引最具投机性的资本。Prague Technology 最相关的买方,是需要授权或集成 AI 大脑模型的人形机器人 OEM——因为 Prague Technology 没有直接面向消费者或企业的产品,必须依靠 OEM 合作触达终端用户。长三角——上海、江苏、浙江——占中国具身 AI 公司和融资的 50% 以上,是自然的地理重心。已经在该领域运营的关键人形 OEM,包括 AgiBot(BYD 支持,按 2025 年人形机器人出货量全球第一)、Unitree Robotics(2025 年出货 5,500 多台,2026 年 6 月 IPO 获批)和 UBTECH(香港上市,2025 年人形机器人收入 ¥821M);如果 Prague Technology 成功商业化世界模型,这些公司会是自然的首批可触达客户。仓储和物流运营商(JD.com、SF Express、Amazon China)是第二层买方,劳动力成本 ROI 更具体,采购周期也快于工业自动化。AI 大脑模型采购的预算所有者位于 OEM 的 CTO 或研发负责人职能内,采购决策主要由模型性能基准和集成可行性驱动,而不是单价。[CM019, CM020, CM021, CM022, CM023, CM024]

细分市场 / 买方地图
细分市场买方 / 付款方用户预算负责人采用触发因素对 AI 大脑的依赖
工业制造(汽车)BYD、NIO、SAIC、Foxconn工厂一线运营运营副总裁 / 工厂厂长降低劳动力成本、提高产出质量高 —— 多步骤任务离不开 AI 控制
人形机器人 OEM(机器人制造商)AgiBot、Unitree、UBTECH、EngineAI机器人 R&D / 集成团队CTO / 工程副总裁与竞争对手的产品能力差距关键 —— OEM 需要外部供应商提供 AI 大脑
仓储 / 物流JD.com、SF Express、Amazon China(仓储 / 物流)配送中心运营团队供应链副总裁用工短缺、周期时间 SLA高 —— 拣选精度需要基础模型
医疗 / 养老医院、养老机构护理 / 支持人员CMO / 设施主管人口老龄化、用工危机中 —— 近期任务专用模型已足够
国防 / 特种作战PLA、政府采购现场操作人员国防部采购危险环境任务成败攸关高 —— 自适应推理至关重要
研究机构高校、国家实验室AI 研究人员PI / 实验室主任获取最先进具身模型关键 —— 研究用例推动平台采用

Prague Technology 目前在任何细分市场都没有既有客户关系。该地图着眼未来,基于全行业买方分析。

[CM021, CM023, CM024, CM025, CM026, CM027]
FM003: 买方 / 细分市场地图

Prague Technology 未来 AI 大脑 / 世界模型产品的买方细分(行)与关键采用维度(列)交叉矩阵。

Prague Technology 当前没有客户。收入时间为推测,依据基础模型采用的行业惯例估算。

[CM020, CM023, CM025, CM026]

2.4 增长驱动因素与采用约束

中国具身 AI 市场由政策指令、软件成熟度和供应链能力三者强力汇合驱动。中国「十五五」规划(2026–2030)将具身 AI 列为量子技术、生物制造等六大未来产业之一,既给出明确政策背书,也带来资金优先级。2026 年 MIIT-SASAC 联合行动计划目标是在年底前部署 10,000 多台机器人。北京、上海、深圳的产业基金规模达到 ¥100 billion($14.5 billion)量级。软件前提——大语言模型成熟——已经由中国本土实验室满足:Alibaba Qwen3.5、DeepSeek V4 和 ByteDance Seed 1.6 展现出世界级推理能力,可被改造成机器人控制逻辑。中国制造供应链覆盖全球工业机器人零部件约 70%。ByteDance 已将世界模型列为 2026 年最高 AI 优先级,并配套 ¥200 billion($29.4 billion)资本开支预算。约束侧,sim-to-real 迁移鸿沟——把仿真中训练的 AI 转化为物理机器人行为的难度——仍是活跃研究问题,但正通过域随机化、数字孪生流水线和神经辐射场收窄。具身模型训练的数据稀缺是根本瓶颈:不同于用互联网文本训练的 LLM,具身 AI 模型需要大规模物理交互数据,采集成本高、速度慢。美国对先进 AI 芯片(NVIDIA H100/H200)的出口管制,给所有中国 AI 开发者带来持续硬件采购风险,Prague Technology 也不例外。资本强度极高——大规模训练世界模型所需算力预算,即便资金充足的初创公司也难以长期维持。独立分析师给出的基准情形是,可靠通用人形自主能力的现实商业部署窗口在 2028–2030 年;这意味着即便按乐观假设,Prague Technology 的商业收入时间线至少也要延后 2–3 年。[CM028, CM029, CM030, CM031, CM032, CM033]

增长驱动与约束表
驱动 / 约束方向时间对 Prague Technology 的含义尽调问题
十五五具身 AI 要求驱动2026–2030 年有效直接政策顺风;政府采购偏向国产 AI核实 Prague 是否符合国家 AI 项目资金资格
MIIT-SASAC 2026 年 10,000 台部署目标驱动立即短期拉动 OEM 对 AI 大脑组件的需求跟踪 2026H2 OEM 部署承诺
成熟的中国 LLM 基础(Qwen3.5、DeepSeek)驱动2025–2026 年活跃Lin Junyang 的 Qwen 经验可直接迁移;加快冷启动要求技术路线图:Qwen 架构如何适配具身任务
ByteDance 将世界模型列为重点 + ¥200B 预算驱动2026验证市场,但也预示一个资源巨大的直接竞争者跟踪 ByteDance 世界模型公告;评估竞争护城河
中国占全球工业机器人部件 70% 份额驱动结构性OEM 伙伴硬件 COGS 低;迭代周期更快确认原型硬件的供应链获取能力
美国对 NVIDIA H100/H200 的出口管制约束持续且升级训练成本上升;限制获取最先进算力评估 Prague 算力策略:NVIDIA 替代方案、国家 GPU 集群准入
仿真到现实迁移缺口约束2026–2028 年收窄Prague 在实机机器人测试前需要仿真数据基础设施核实技术路线图中的仿真到现实方案;是否与仿真平台合作?
具身训练数据稀缺约束结构性瓶颈世界模型质量受物理交互数据卡住;采集成本高确认数据飞轮策略;是否有遥操作或 OEM 数据合作?
2028–2030 年商业自主化时间线约束多年滞后Prague 收入大概率延后至 2028+;需要耐心资本和 ≥5 年周期对齐投资人预期;要求给出到首笔收入前的资本跑道测算

截至 2026 年 7 月,驱动因素和约束基于公开政策文件、新闻和分析师报告评估。

[CM028, CM029, CM030, CM031, CM032, CM033]
FM004: 采用漏斗 / 价值链图

中国具身 AI 价值链,展示 Prague Technology 计划卡位的 AI 大脑层——这是上游组件供应商与下游 OEM 客户之间关键但仍薄弱的软件瓶颈。

[CM024, CM025, CM026, CM028, CM029]

2.5 展示要点

Chapter 03

03竞争对手

3.1 竞争格局概览

Prague Technology 面临三个重叠战场:(1)中国全栈具身 AI 公司,同时开发硬件和 AI 大脑;(2)全球纯软件世界模型初创公司;(3)内部建设具身 AI 能力的大型科技公司。围绕 Prague 的世界模型 / AI 大脑定位,直接竞品范围更小,但资金极其充足:TARS AI 已在两轮融资中筹集 $697 million,Spirit AI 融资 $435 million、估值 $1.5 billion,AMI Labs(Yann LeCun)融资 $1.03 billion、估值 $3.5 billion,World Labs(Fei-Fei Li)融资超过 $1 billion,目标估值据报为 $5 billion。Prague Technology 融资 $220 million、估值 $2 billion,在这组同业中按融资额居中,但按隐含收入倍数处在高端——这是风险评估的重要差别。ByteDance 的在位者威胁可能是 Prague 面临的最大结构性竞争风险:ByteDance 已将世界模型列为最高 AI 优先级,并配套 ¥200 billion 资本开支预算,算力、数据和分发能力都比 Prague 高出几个数量级。Prague 必须替代的替代品和现状方案包括:(a)为机器人控制改造的微调 LLM(目前足以胜任窄任务);(b)OpenVLA、RT-X 等开源项目的 VLA(vision-language-action)模型;(c)硬件 OEM 自研 AI——AgiBot 和 Unitree 都在投入自有 AI 模型,降低对外部大脑供应商的依赖。潜在进入者包括任何将算力预算转向具身 AI 的中国主要互联网公司(Baidu、NetEase、Xiaomi),以及任何延伸到物理 AI 的全球 AI 实验室(Anthropic、xAI)。[CP001, CP002, CP003, CP004, CP005, CP006]

竞争对手画像表
竞争对手类别规模 / 融资目标细分市场差异化相比 Prague 的关键局限
TARS AI(它石智航)直接 / 中国全栈$697M(天使 + Pre-A);估值 >$2B制造业自动化(工业、人形)全栈硬件 + 大脑;AWE 3.0 模型;数据飞轮资本更多,有硬件护城河和运营数据
AMI Labs(Yann LeCun)直接 / 全球世界模型$1.03B 种子轮;投前 $3.5B医疗、工业(全球)JEPA 架构;明星团队;NVIDIA/Samsung 背书不聚焦中国;医疗优先,而非机器人优先
World Labs(Fei-Fei Li)邻近 / 全球空间 AI$1B+;估值目标 $5B3D 设计、娱乐、媒体3D 空间智能;Autodesk 合作;「Marble」产品用例不同(3D 设计 vs. 机器人控制)
Spirit AI(精灵智能)直接 / 中国大脑层$435M;估值 $1.5B通用机器人脑(中国)先发者;Chaos + YF Capital 背书公开信息较少;融资画像相近
AgiBot(BYD 支持)邻近 / 中国全栈 OEMBYD + Hillhouse;全球出货量制造业、汽车(BYD、SAIC)硬件数据飞轮;BYD 分销护城河会自研,而不是购买 Prague 的模型
Unitree Robotics邻近 / 中国 OEM¥1.699B 收入;IPO 已获批硬件优先;开发者生态;开放 SDK开源 UnifoLM-VLA-0;价格领导者($4.9K-$43.9K)开源策略压缩 AI 大脑市场规模
ByteDance(内部)既有替代者2026 年资本开支 ¥200B平台 AI;面向 Doubao 的世界模型算力规模;200M+ 日活用户;数据护城河今天不聚焦机器人,但可快速转向
Alibaba(Qwen 团队)既有企业 / 前雇主海量内部资源企业 AI,云优先Qwen3.5 开源 LLM 可作为机器人大脑基础开源 Qwen 可能压低 Prague 模型定价
Google DeepMind (RT-X)既有玩家 / 开源基线算力不设上限(Alphabet)全球研究 + 企业市场Open X-Embodiment 数据集;RT-2/Gemini Robotics设定 Prague 必须超越的开源「够用」门槛
Physical Intelligence (π0)相邻玩家 / 美国具身 AIOpenAI 背书;融资未披露多任务操作(美国)π0 多任务模型;硬件无关不聚焦中国市场;架构路线不同

Prague Technology 概况:已融资 $220M,估值 $2B,尚无产品,聚焦纯软件世界模型;截至 2026 年 7 月仅面向中国。竞争数据来自公开来源;估值在未正式披露处为近似值。

[CP001, CP002, CP007, CP009, CP013, CP015]
FP001: 竞争定位图

按两个轴定位竞争格局:中国市场聚焦度(X,1-10)与 AI 大脑 / 软件深度(Y,1-10)。Prague Technology 位于高中国聚焦 / 高软件深度象限,与 TARS AI、Spirit AI 并列;它既不同于全球纯软件竞争者(AMI Labs、World Labs),也不同于软硬件一体的 OEM 玩家。

坐标轴位置是基于证据的序数评分,不是量化测量。X 轴(中国市场聚焦度):10=仅在中国运营;1=仅在美国 / 欧盟运营。Y 轴(AI 大脑 / 软件深度):10=纯软件模型;1=纯硬件、无 AI 大脑。

[CP001, CP007, CP013, CP018, CP023]

3.2 直接竞品画像:世界模型与 AI 大脑层

TARS AI(它石智航)由前 Huawei 自动驾驶 CTO 陈亦伦和前 Baidu Apollo 总裁李震宇于 2025 年 2 月创立,是 Prague Technology 最直接的中国竞争对手。TARS 采用全栈策略:自研硬件(A 系列轮式工业机器人和 T 系列双足人形机器人)与 AWE 3.0 通用具身大模型结合。TARS 两次刷新中国具身 AI 融资纪录——先是在 2025 年第二季度完成 $242 million 天使轮,随后在 2026 年 4 月完成由 Hillhouse 和 Sequoia China 领投的 $455 million Pre-A 轮。其投资人覆盖战略股东(Meituan)、财务资本(Hillhouse、HongShan)、产业资本(TCL)和国资基金(北京机器人基金、上海国资)。相较 Prague 的纯软件定位,TARS 的全栈路线是关键差异点——TARS 靠硬件部署采集自有具身训练数据,形成数据飞轮;Prague 必须靠 OEM 合作复制这一机制。AMI Labs(Advanced Machine Intelligence Labs)由图灵奖得主 Yann LeCun 于 2025 年末创立,并在 2026 年 3 月以 $3.5 billion 估值融资 $1.03 billion,创下欧洲史上最大种子轮。AMI 基于 LeCun 的 JEPA(Joint Embedding Predictive Architecture)框架构建世界模型,并明确将自己定位为商业化前、研究优先的实体。AMI 的近期应用是医疗健康(通过 Nabla 合作),不是机器人;不过其投资方包括 NVIDIA、Samsung 和 Toyota Ventures,说明工业应用已有规划。AMI 目前不直接竞争中国市场。World Labs(Fei-Fei Li)融资超过 $1 billion,其中包括 Autodesk $200 million 战略投资,目标估值据报为 $5 billion。其首款产品 Marble 聚焦面向娱乐和设计的 3D 环境生成,不是工业机器人。在全球世界模型领导者中,World Labs 与 Prague 目标场景的直接竞争性最低。Spirit AI(精灵智能)已从 Chaos Ventures 和 YF Capital 融资 $435 million,估值 $1.5 billion,明确瞄准通用机器人大脑市场,而 Prague 也在中国追逐同一市场——因此它是直接国内竞争对手,且领先 12 个月、拥有更深的中国本土投资人关系。[CP007, CP008, CP009, CP010, CP011, CP012]

特性 / 能力矩阵
采购标准Prague Tech(标的)TARS AIAMI LabsSpirit AIAgiBotUnitree
已发布具身大模型无(计划中)AWE 3.0 ✓JEPA(研究)已声称 ✓自研 VLA ✓UnifoLM-VLA-0 ✓
世界模型架构计划中 ✓(重点)部分具备 ✓基于 JEPA ✓已声称 ✓研发中 ✓有限
专有训练数据None来自机器人部署 ✓研究数据集披露有限制造业部署 ✓开放 SDK 部署 ✓
硬件 + 机器人平台是(A-series、T-series)✓是(AgiBot A2)✓是(G1、H2、R1)✓
中国制造业客户无(计划中)BYD 线束 ✓否(欧盟 / 医疗)未披露BYD、SAIC ✓广泛生态 ✓
开源组件TBDWIYH 数据集 ✓是(计划中)✓未披露开放数据集 ✓UnifoLM-VLA-0 ✓
企业销售记录None逐步增长(LogiMAT 首秀)无(商业化前)未披露是(BYD 规模)✓是(全球)✓
LLM / 基础模型深度✓✓(Qwen 架构师)中等 ✓✓✓(LeCun JEPA)披露有限逐步增强 ✓有限

✓ = 已确认能力;'✓✓' = 深度异常突出;'无/否/有限' = 公开资料未证明。Prague Technology 条目仅基于公司披露方向。

[CP008, CP010, CP016, CP024, CP026, CP032]
FP002: 功能广度 / 能力地图

Prague Technology 与五个关键竞争者在六个竞争维度上的能力覆盖矩阵。单元格展示截至 2026 年 7 月、基于公开证据确认的能力状态。

能力判断基于公开公告、产品页面和新闻报道。“无 / 否 / 有限”表示没有公开确认的能力。“计划中”表示公司已说明方向,但尚未发布产品。

[CP008, CP010, CP016, CP021, CP024, CP032]

3.3 OEM 平台竞争者与在位 AI 巨头

AgiBot 和 Unitree 是中国按出货量计最大的两家人形机器人 OEM,构成 Prague Technology 必须面对的关键「自建还是外购」动态。AgiBot 由 BYD 支持,2025 年人形机器人出货 5,168 台、全球第一,正在积极开发内部 VLA 和世界模型能力;AgiBot World Challenge 2026 吸引了来自 27 个国家的 526 支团队,展示了这一进展。AgiBot 的数据护城河直接来自其在 BYD 和 SAIC 的制造部署,提供 Prague 需要依靠合作才能获得的具身交互数据。Unitree 已开源其 vision-language-action 模型 UnifoLM-VLA-0,释放出另一种策略信号:把 AI 大脑层商品化,推动硬件采用和开发者生态。如果 Unitree 跑通这一策略,所有独立 AI 大脑供应商都会承受定价压力。ByteDance 在 2026 年宣布世界模型为最高 AI 优先级,并配套 ¥200 billion($29.4 billion)资本预算,这代表压倒性的在位者资源优势。ByteDance 的 Doubao 平台拥有 2 亿多日活用户,让其可以获取规模上任何初创公司都无法匹敌的多模态用户交互数据。虽然 ByteDance 尚未聚焦机器人专用世界模型,但其算力和分发能力意味着它随时可能入场。Alibaba 的 Qwen 团队——林俊阳的前雇主——仍在活跃:Qwen3.5 于 2026 年 3 月发布,Alibaba 也披露了对具身 AI 的持续研发投入。关键风险在于,Alibaba 可能基于开源 Qwen 基础,利用林俊阳离职后留下的架构创新,开发直接与 Prague 竞争的机器人专用世界模型。Google DeepMind 的 RT-X(Open X-Embodiment)联盟和 Gemini Robotics 提供了所有具身 AI 初创公司都必须超越的开源基线,才能证明授权费合理。企业买方评估 AI 大脑软件「买还是自建」时,开源替代方案(OpenVLA、RT-X)就是关键替代威胁。[CP021, CP022, CP023, CP024, CP025, CP026]

定价 / 打包方案对比
公司定价模式标价 / 合同包含能力未知 / 未披露Prague 启示
Prague Technology不适用 — 尚无产品N/AN/A所有定价Prague 没有任何定价信号;无论采用哪种模式,都会重新定义品类
TARS AI企业 SaaS + 硬件租赁未公开披露AWE 3.0 模型 + 机器人硬件 + 部署支持按模型 API 定价、SaaS 档位全栈打包可能压过独立模型定价
AMI Labs尚未商业化近期无收入计划仅研究访问所有商业条款为世界模型定了先例:这是长周期投资,不是快速变现
World Labs企业 SaaS(3D 生成)Marble 发布后未披露3D 环境生成;Autodesk 集成面向机器人的定价用例不同;不能直接作为定价基准
AgiBot硬件采购 / 企业合同未公开标价AgiBot A2 系统 + 自研 AI完整价目表OEM 打包定价会让独立 AI 大脑更难单卖
Unitree Robotics硬件零售 + 企业$4,900(R1 AIR)至 $43,900(G1 EDU);也提供 RaaS 定价机器人 + 开放 SDK + UnifoLM-VLA-0(开源)企业合同条款开源 AI 大脑模型把开发者端底价压到零

硬件产品定价来自公开来源。全行业都未披露企业 AI 大脑定价。缺少披露定价,现阶段无法建模 Prague 的现实 ASP。

[CP017, CP026, CP027, CP031]
FP003: 护城河 / 就绪度 KPI

Prague Technology 关键维度上的竞争就绪度快照,用于评估截至 2026 年 7 月的护城河韧性。绿色=优势;黄色=持平;红色=劣势。

[CP032, CP033, CP036, CP039, CP040]

3.4 护城河评估与竞争风险

Prague Technology 设想中的护城河——来自 Qwen 的 LLM 架构经验、具身 AI 世界模型先发优势、顶尖创始人可信度——每一项都有被侵蚀的风险。架构经验护城河在 18–24 个月窗口内最耐久,但随着行业转向开放架构和基准,竞争会加剧。林俊阳在 transformer 扩展和 MoE(mixture-of-experts)架构上的专长确实稀缺,Qwen3.5 的 397B 参数 MoE 模型已经证明这一点,并可能在早期世界模型训练中带来数据效率或算力效率优势。不过,TARS AI 的创始团队(Huawei 自动驾驶 + Baidu Apollo)同样履历强,且拥有运营级 AI 部署经验;相较学术 LLM 扩展,这种经验或许更贴近真实机器人落地。数据飞轮是任何具身 AI 模型提供商最关键的长期护城河:谁能收集最多样、最高质量的具身交互数据,谁就能训练出更好的模型。截至 2026 年 7 月,Prague Technology 没有已部署机器人,因此没有数据采集优势。TARS、AgiBot 和 Unitree 都有活跃硬件部署,正在生成自有数据。相对全栈同业,这就是 Prague 最显著的结构性竞争劣势。分发和商业化路径是 OEM 集成型竞争者的中期护城河:AgiBot 的 BYD 分发能力和 Unitree 的开发者生态,为下游 OEM 客户制造转换成本,Prague 必须靠更强模型性能来克服。商品化风险真实存在:如果到 2027–2028 年,Google DeepMind 的开放 RT-X 生态,或 OpenVLA、Pi0 等开源 VLA 框架,对多数工厂任务已经「足够好」,授权 AI 大脑模型的市场就会压缩到专业应用。Prague Technology 的价值主张依赖一个假设:面对复杂多步骤任务,世界模型显著优于 VLA 模型。这个假设在技术上有前景,但尚未在商业规模上得到经验证实。[CP032, CP033, CP034, CP035, CP036, CP037]

护城河耐久性 / 竞争风险台账
护城河主张威胁严重性缓释措施 / 尽调问题
Lin Junyang 的 Qwen LLM 架构经验TARS / Spirit 也有同等级 AI 团队;Alibaba 仍保留 Qwen 团队要求提供早期技术白皮书或基准结果,验证具体架构优势
中国具身 AI 世界模型先发TARS AWE 3.0 和 Spirit AI 已发布模型;Prague 尚未发布确认 Prague 模型开发时间表;量化与 TARS AWE 3.0 的差距
聚焦纯软件策略(不被硬件分散)软硬一体竞争对手收集的训练数据比纯软件模型商更丰富核验 Prague 数据获取策略 — OEM 合作?遥操作数据许可?
顶级投资方背书(Gaorong、HongShan、Tencent)TARS 投资方同样顶级(Hillhouse、HongShan),还有国资;并非差异点不是护城河;核验投资方引荐能否转化为 OEM 客户引荐
具身数据飞轮潜力截至 2026 年 7 月,Prague 没有已部署机器人;没有采集真实世界数据关键确认数据合作策略;是否已签 OEM 数据共享协议?
世界模型架构优势现有基准还无法在多数任务上区分世界模型与先进 VLA 模型;通用自主能力时间线在 2028-30 年要求查看任何内部基准结果或已发表技术论文

严重性评级反映截至 2026 年 7 月的竞争位置。'关键' 表示若不处理,威胁可能完全阻断商业化起量。

[CP032, CP033, CP034, CP035, CP036, CP037]

3.5 展示要点

Chapter 04

04财务

4.1 收入模式与商业策略

截至 2026 年 7 月,Prague Technology 没有当前收入、没有商业产品,也没有披露客户,明确处于商业化前的研究阶段。结合林俊阳公开表述和公司披露方向推断,其目标收入模式是 B2B AI 模型授权和 API 访问业务:人形机器人 OEM 授权 Prague 的世界模型,为自己的 AI 大脑层提供动力,并支付经常性 API 费用或按模型运行次数计费。次级收入流可能包括面向 OEM 特定硬件配置的定制模型训练或微调服务,也可能包括数据标注或仿真基础设施服务。Prague 的纯软件策略(不做硬件制造)一旦产生收入,理论上可形成高毛利商业模式,类似规模化后达到 70-85% 毛利率的基础模型 API 提供商。但这一理论毛利率假设有大量 OEM 客户在生产环境部署 Prague 模型。最接近且已验证的可比公司显示,商业化时间线需要数年:AMI Labs 的 CEO 明确承认,世界模型商业应用需要 “years, not months or quarters”;Physical Intelligence 的 π0 已展示多任务操作,但没有披露大规模商业收入。Prague 的产品路线图、开发时间线、定价结构和商业化路径完全未披露,因此无法做出除最粗略定性边界以外的任何财务建模。截至 2026 年 7 月,公司没有任何技术白皮书、模型基准或公开原型,这意味着收入建模的基础——产品会有效的证据——也无法独立验证。[CI001, CI002, CI003, CI009, CI010, CI011]

收入流表
收入流机制单位当前状态收入质量尽调问题
AI 大脑 / 世界模型授权OEM 授权 Prague 模型用于机器人控制按模型 API 调用或年度固定授权不适用 — 无产品一旦跑通,质量高(经常性)要求提供模型路线图和首个目标 OEM LOI
定制模型训练服务为 OEM 特定硬件或任务微调世界模型项目制收费;按训练运行次数收费不适用 — 无产品质量中等(一次性,有项目风险)确认规划中是定制训练还是自助 API
仿真 / 数据基础设施授权向 OEM 授权仿真环境或数据管线,用于具身 AI 训练订阅或一次性不适用 — 推测性质量中等确认 Prague 是否规划仿真基础设施,还是只做纯模型
科研拨款 / 政府补贴中国政府具身 AI 研发拨款;国资基金支持拨款发放,非稀释可能 — 无公开确认质量低(一次性,不确定)确认 Prague 是否申请 MIIT 或国家 AI 基金拨款

所有收入流均为前瞻性。Prague Technology 截至 2026 年 7 月确认收入为零。收入质量评级假设未来进入商业化运营。

[CI010, CI015, CI016]
定价 / 变现表
情景基准单价 / 合同用量假设隐含 ARR置信度
保守情景(商品化模型定价)OpenAI GPT-4 API 等效价,为早期 OEM 采用大幅折扣$0.05–0.10 / 1,000 次 API 调用10 家 OEM,每家每月 10M 次调用~$6–12M ARR低 — 推测性
基准情景(高端专用模型)Physical Intelligence 级溢价,商品化定价的 10 倍每家 OEM 年授权 $500K–2M2028 年 10–20 家 OEM~$5–40M ARR很低 — 需要产品验证
乐观情景(平台授权)世界模型成为事实标准;TARS AWE 3.0 竞争性费率每家 OEM 每年 $1–5M到 2030 年 50+ 家 OEM$50–250M ARR高度推测 — 2030 年时间线
硬件打包(伙伴模式)若 Tencent 或 OEM 打包 Prague 模型分成 OEM 合同价值的 15–25%3–5 个战略 OEM 集成Unknown很低
当前(实际)无产品、无客户、无定价N/A$0已确认

所有定价情景都是基于可比 AI 模型授权交易的假设性基准;尚无 Prague 专属定价披露。仅用于展示边界。

[CI011, CI014, CI015]
FI001: 收入模型路径图

Prague Technology 的概念性收入模型:从基础研究,到模型交付,再到 OEM 授权和经常性收入。所有节点均为前瞻;Prague 尚未越过研究阶段。

除“资本”外,所有节点均为前瞻。收入节点截至 2026 年 7 月为零。

[CI010, CI015, CI016]

4.2 单位经济与成本结构

Prague Technology 的单位经济无法评估,因为公司没有披露收入或成本。不过,世界模型开发公司的结构性成本驱动因素,从行业可比公司已经很清楚。最主要成本是 AI 算力:训练前沿世界模型所需算力预算,可比甚至超过前沿 LLM 训练,每轮训练从数千万美元到数亿美元不等。达到 Qwen3.5 的 397B 参数 MoE 模型规模时,单次训练成本以数千万美元计。由于具身 AI 训练需要多模态物理仿真,世界模型可能需要更多算力。人才获取和留存是第二大成本驱动因素——头部实验室的前沿 AI 研究员年薪包在 $500K 到 $3M 之间,组建一支 50–100 人的竞争性研究团队,按市场价格每年需要 $50-150M 薪酬。基础设施和云算力费用(或 Prague 如果采购硬件,则为自有 GPU 集群)构成重大资本开支。正向成本结构论点是,Prague 没有硬件 COGS、没有制造开销、没有库存、没有实体分销;如果产品是纯软件授权,规模化收入下毛利率可能超过 80%。反向成本论点是,算力成本受 GPU 出口管制影响,中国公司获取 NVIDIA H100/H200 受限,被迫依赖国产替代(Huawei Ascend、国产 GPU),而后者价格性能比可能更差。美国已对输往中国的 NVIDIA H200 AI 芯片加征约 25% 关税,显著推高所有中国 AI 模型开发者的训练成本通胀。[CI017, CI018, CI019, CI020, CI021, CI022]

单位经济模型表
指标数值 / 估算区间置信度重要性尽调问题
毛利率(规模化目标)70–85%(纯软件授权模式)低 — 理论值收入规模化后,可指示长期盈利潜力核验定价模型是否覆盖训练成本回收
每次前沿模型训练成本$20–200M(可比 GPT-4 级别)低 — 按行业基准估算决定资本强度和融资节奏要求提供资本配置计划:算力占预算百分比
每次 API 调用推理成本未知;取决于模型规模和 GPUUnknown决定规模化后的收入边际成本要求提供计划中的推理基础设施设计
客户获取成本(CAC)未知 — 尚无产品,也尚无销售团队N/A测算回本周期和所需销售与营销(S&M)支出离不开它确认 GTM 策略:直销还是自助服务;是否已有 beta 阶段 OEM?
单客户收入(ACV)未知 — 未披露定价N/ALTV/CAC 模型的基础;没有它无法做投资测算要求公司提供拟定定价档位结构

单位经济模型无法在没有已发布产品和首批客户合同的情况下评估。所有数值不是估算就是未知。本表记录尽调阻塞点。

[CI017, CI019, CI020, CI021, CI022]
FI002: 单位经济模型路径图

世界模型 AI 大脑授权业务的结构性单位经济模型。鉴于收入为零,数值均为定性目标;估算说明列出缺失输入。

所有数值均为结构性估算。Prague Technology 没有收入或客户,因此不存在实际单位经济。毛利率目标类比前沿 LLM API 提供商,尚未在具身 AI 世界模型中验证。

[CI017, CI018, CI019, CI020]

4.3 资本充足性与融资

Prague Technology 在约 2026 年 6 月交割的天使轮中融资约 $220 million:高榕创投 $100 million,HongShan(Sequoia China)$100 million,Tencent $20 million。投后估值约 $2 billion。按保守月烧钱 $5 million(覆盖约 50 人团队和有限算力)计算,$220 million 可提供约 44 个月(3.7 年)跑道;如果开发按计划推进,足以在 2029 年前达到概念验证。按激进月烧钱 $12 million(建设 200+ 人团队并投入高强度算力)计算,跑道压缩到约 18 个月,需要在 2027 年底前完成后续融资。多家信源称,Prague Technology 在天使轮刚交割后就已寻求后续融资——这说明公司预期目标开发时间线需要的资本超过 $220 million。Tencent 的战略投资可能带来现金以外的价值:Tencent Cloud 的 GPU 集群、Tencent 企业生态分发,以及通过 Tencent 平台业务形成的潜在数据合作,都可能降低 Prague 的有效算力成本和上市时间。$2 billion 投后估值意味着价格 / 资本比约 9x($2B / $220M),高于 AMI Labs(3.4x),也符合 2025-2026 年市场对顶级 AI 创始人的溢价。只有当 Prague 产品按期落地并被 OEM 强势采用时,这一溢价才站得住。反向财务风险场景是,世界模型开发时间线拖过 2029 年,Prague 在达到商业验证前触发降估值融资或被迫战略转向。S&P Global 和 Morgan Stanley 都曾提示,AI 行业投资风险来自资本集中投向开发周期长的前收入公司。[CI004, CI005, CI006, CI007, CI024, CI025]

资本充足性表
项目数值 / 估算来源 / 置信度含义尽调问题
累计融资~$220M(投资方:Gaorong $100M、HongShan $100M、Tencent $20M)已确认 / 高(多来源)设定现金续航上限确认精确交割金额,以及是否分批到账
投后估值~$2B(据报道)中 — 公司未正式披露决定稀释水平和下一轮融资门槛要求提供确认估值的条款清单或投资人更新
月现金消耗(估算)$3–15M/月(区间)低 — 无公开数据;按团队搭建推断按 $5M/月:现金续航 44 个月;按 $12M/月:18 个月要求提供财务报表或董事会级现金消耗率披露
隐含现金续航(月)18–73 个月,取决于现金消耗低 — 模型估算评估融资依赖度的关键确认后续轮是否已在推进
计划资金用途算力(模型训练)、人才招聘、运营基础设施中 — 按公开表述和行业惯例推断显示 $220M 会以多快速度花出去要求提供详细资本配置计划

现金消耗率和现金续航仅为模型估算。Prague Technology 未披露财务报表、现金余额或现金消耗率。据报道,公司在天使轮交割后立即寻求后续融资。

[CI001, CI005, CI028, CI031]
FI003: 财务估算区间

2025-2026 年获得融资的世界模型初创公司的投后估值区间($B)对比,显示 Prague Technology 的 $2B 估值相对全球同业所处位置。所有数字均为各自融资轮时点。

Prague 估值中点($2B)和 AMI Labs(投前 $3.5B)来自新闻报道;高 / 低边界反映分析师估算和传闻中的后续估值。TARS 估值未正式披露;$2.5B 为分析师估算。World Labs 的 $5B 是报道中的目标估值。

[CI002, CI031, CI032, CI033, CI034]

4.4 财务结论与关键尽调阻碍

Prague Technology 的财务画像是高估值下的极端前收入投机——一家零收入、零产品、只有一名公开员工(林俊阳)的公司,估值 $2 billion。这在 2025-2026 年 AI 市场并非没有先例:AMI Labs($3.5B)和 World Labs($5B 目标)在商业进展同样有限的情况下估值更高。不过,9x 的 P/Capital 比率在世界模型初创公司中处于最高区间之一,本质上是对林俊阳架构能力和具身 AI 时点的纯押注。财务命题只有在以下条件同时成立时才成立:(a)Prague 在 2027-2028 年成功训练并发布世界模型;(b)模型性能超过开源替代和 TARS AWE 3.0;(c)多个 OEM 客户以商业可行的授权费采用该模型。任何一项失败,估值都会坍缩到人才收购(acqui-hire)或归零结果。Forbes 曾记录 AI 初创公司正在基于氛围和叙事而非经验证指标「放大估值」;Prague 正好符合这一原型:投资命题完全建立在创始人声誉,以及世界模型会成为下一代 AI 范式的结构性论点上。承销 Prague 财务风险时,最关键未知数是月度现金消耗率:如果公司因激进算力支出,烧钱已经超过 $8-10M/month,就可能需要在 2027 年底前融资,而届时市场条件未必更有利。正式投资决策前必须解决的财务尽调阻碍包括:(1)交割时经验证的现金余额和月度烧钱;(2)资本分配计划(算力、人才及其他);(3)与 Tencent Cloud 的任何数据共享或收入分成安排;(4)是否存在与 OEM 客户签署的任何 LOI 或合作 MOU。[CI035, CI036, CI039, CI040]

公开财务信息缺口表
缺失指标对分析的影响精确尽调路径
月现金消耗率阻塞 — 没有现金消耗数据就无法评估现金续航或融资依赖要求提供自成立日至今经审计或管理层复核的现金流量表
资本配置计划(算力占比 / 人才占比)阻塞 — 决定 $220M 足够支撑首个模型,还是只够研究阶段要求在公司资料室提供预算拆分;对比 TARS AI $697M 配置模式
目标 OEM 合作管线高 — 任何已签 LOI 都可验证商业化假设;缺失会抬高 GTM 风险要求提供 OEM 洽谈清单、阶段和预计签约日期
Tencent 战略利益与条款高 — Tencent Cloud 算力访问可能减少 $20-50M 现金消耗;是否包含尚不清楚要求提供 Tencent 投资附函条款;确认是否包含云服务抵扣额度
政府拨款 / 补贴申请中 — 可补充非稀释资本、延长现金续航;是否已申请未知向管理层确认是否已申请 MIIT AI 基金、国家算力访问计划

这些财务缺口是财务投资测算的主要阻塞点。缺少现金消耗率和资本配置数据,本章无法支撑量化投资建议。

[CI035, CI036, CI037, CI038, CI039]
FI004: 资本强度 / 现金流图

Prague Technology 资本投放图:$220M 天使轮是唯一输入,输出是未来世界模型能力和最终收入。关键路径风险在于 GPU 算力获取受美国对 NVIDIA H 系列芯片出口管制挤压。

资本分配比例依据前沿 AI 实验室的行业惯例估算;Prague 未披露专属预算。

[CI001, CI023, CI029, CI030]

4.5 展示要点

Chapter 05

05产品与技术

5.1 产品愿景与架构

Prague Technology 的产品愿景,是面向具身 AI 的世界模型——一个 AI 大脑层,让人形机器人和物理系统能够以足够高的保真度感知、建模并预测物理世界,从而自主执行灵巧、多步骤任务。不同于通过模仿学习把感知输入直接映射为运动动作的 VLA(vision-language-action)模型,Prague 设想的世界模型路线处在更高认知层:它学习物理世界如何运转的内部表征,让机器人在执行前先通过心理仿真规划未来状态。这正是 LeCun 为自动驾驶提出、AMI Labs 为通用具身 AI 追逐的同一种认知路径。Prague 的差异化在于专注中国 OEM 机器人产业,以及创始人从 Qwen 积累的多模态、大规模模型架构深度经验。林俊阳曾带领 Qwen 多模态架构开发,使其具备视频理解、图像推理和跨模态检索能力;这些都是世界模型开发的基础能力。用客户工作流来定义产品:OEM(机器人制造商)把 Prague 的世界模型作为认知层集成进机器人计算栈,用一个可跨多种操作任务工作的通用 AI 大脑,替代或增强自有感知-规划模块,而不需要针对每个任务重新训练。价值主张是降低每个新任务的集成成本,提高跨环境泛化能力,并接入一个由多家 OEM 数据训练、持续改进的基础模型。截至 2026 年 7 月,Prague Technology 尚未公开发布产品演示、白皮书、架构博客或基准结果。[CE001, CE002, CE003, CE004, CE005]

产品模块 / 资产矩阵
模块 / 资产用户状态 / 成熟度差异化尽调缺口
世界模型核心OEM 人形机器人制造商未启动(概念阶段)MoE 架构 + Lin Junyang 的 LLM 经验;物理世界动力学没有原型、架构文档或基准测试
多模态编码器需要视觉、语言和本体感知输入的 OEM 机器人未启动Qwen 多模态积累;视频理解经验未发布模型,也没有能力演示
机器人控制接口面向实时执行的 OEM 集成层未启动为响应式控制设计,目标延迟低于 100ms未发布硬件集成规范
仿真 / 数据基础设施内部研发和 OEM 训练数据生成未启动大规模合成数据生成;仿真到现实迁移没有仿真合作方,也未描述数据管线
API / 开发者平台集成 Prague 模型的 OEM 工程团队未启动面向机器人 OEM 集成的标准化 SDK没有 SDK、API 文档或开发者门户

所有模块都只是前瞻设想。截至 2026 年 7 月,Prague Technology 尚未发布任何软件组件。

[CE001, CE002, CE004]
工作流 / 用例表
用户任务现有工作流Prague 方案可衡量收益限制
在装配线上拆箱并摆放多类物体每个 SKU 都要手动编程;TARS AWE 3.0 只能处理有限 SKU 集世界模型预测新物体的稳定摆放方式任务泛化速度比重新编程快 10x新物体需要多样化训练数据;延迟待定
多步骤厨房备餐预编程流程或专用机械臂Prague 世界模型推理物体物理、器具可供性和任务顺序单一模型处理多种菜谱需要厨房环境训练数据;家庭部署还需要安全认证
建筑工地物料搬运人工操作机械具身 AI 世界模型在杂乱环境中规划安全导航降低人工成本,提高安全性户外环境的仿真到现实鸿沟最难;建筑工地监管审批周期很长
机器人养老辅助任务范围有限的专用辅助机器人Prague 世界模型让养老辅助具备通用家务操作能力若成功,将带来变革;TAM 大幅扩展安全负担最高;监管认证需要多年

所有用例都是 Prague 世界模型论点的示例性应用,尚不存在 Prague 专属落地实现。比较对象引用 TARS AI AWE 3.0、Physical Intelligence π0 和 AgiBot,作为已验证参照。

[CE002, CE003, CE005]
FE001: 产品架构图

Prague Technology 世界模型平台的概念性产品架构——展示从物理数据输入到 OEM 机器人集成的五层。所有层均为前瞻;公司尚未发布生产架构。

这是根据 Lin Junyang 的 Qwen 工作和行业研究推断的概念架构;Prague 未发布专属架构文档。

[CE006, CE007, CE008]

5.2 技术栈与 AI 架构

Prague Technology 的技术架构目前只能从 Lin Junyang 过往负责 Qwen 的经历,以及更广泛的世界模型研究版图中推断。Lin Junyang 曾主导 Qwen 2.5 和 Qwen 3.5 模型家族,其中有几项与具身 AI 世界模型直接相关的架构创新:(1)Mixture-of-Experts(MoE)架构,让计算在专门子网络之间按需分配,降低复杂多模态推理的推理成本;(2)混合思考模式,把快速反应和慢速深思结合起来(类似 System 1/System 2 认知)——这直接适用于机器人控制:有些响应必须低于 100ms,有些则需要多步规划;(3)用于视觉和空间理解的多模态编码器;(4)可训练 3970 亿参数模型的大规模分布式训练基础设施。这些能力正好对应世界模型的核心技术要求:多模态编码器(视觉 + 本体感知 + 语言输入)、内部状态模型(物理世界动力学)、规划模块(目标条件轨迹生成)和机器人控制接口(动作输出)。Physical Intelligence 的 π0 证明,30 亿参数 VLA 模型只要用多样化机器人数据训练,就能在 68+ 项任务上做到一次样本泛化——但它仍依赖短时域动作预测,而不是物理世界模拟。DeepMind 的 Genie 2 显示,100 亿参数世界模型用视频训练后,可以从单张图像生成一致的 3D 交互环境,证明世界模型路线可用于物理模拟。Prague 的技术路线很可能同时借鉴这两个先例,但目标是专门面向中国 OEM 机器人数据、制造场景和灵巧操作任务优化的模型。关键技术风险包括:(a)世界模型需要比 VLA 大得多的训练数据多样性——这要求尚未建立的 OEM 数据合作;(b)机器人控制要求 30-100Hz 的实时推理,放在世界模型规模上计算压力很高,现有硬件可能撑不住;(c)物理系统的安全与可靠性带来超出软件 AI 基准的新技术要求。[CE006, CE007, CE008, CE009, CE010, CE011]

技术 / 运行架构表
层 / 组件作用依赖风险
专家混合(MoE)骨干用条件计算处理多样物理场景,不必激活全模型Lin Junyang 来自 Qwen3.5 的经验;GPU 显存带宽MoE 推理延迟可能超过机器人实时控制要求
多模态编码器(视觉 + 语言 + 本体感知)将视觉场景、自然语言指令和机器人关节 / 力传感器输入解析为统一表征高分辨率相机 + IMU + 力传感器与 OEM 硬件集成传感器融合复杂;跨 OEM 机器人形态做标准化很难
世界动力学模型(核心 IP)基于当前状态和计划动作,预测 T+N 帧后的物理世界状态海量具身交互训练数据;物理仿真引擎未确认 OEM 数据合作;合成到现实鸿沟是重大研究难题
任务规划器 / 目标条件策略将高层目标(如「冲咖啡」)拆成可执行子任务序列轨迹规划需要近实时查询世界模型在机器人控制频率(30-100Hz)下做目标条件规划,计算成本尚未解决
机器人控制接口层将世界模型输出转换为 OEM 机器人执行器的低层电机指令逐家 OEM 硬件 SDK 集成;本体感知反馈回路需要按 OEM 做专项集成;没有标准化投入,就不是通用层
训练数据基础设施生成、标注并管理十亿级合成 + 真实世界具身交互数据物理仿真环境 + OEM 机器人数据协议;大规模存储和算力最大技术依赖;尚未披露任何用于训练数据的机器人 OEM 合作

架构判断来自 Lin Junyang 的 Qwen 工作和行业基准;Prague 尚未发布专属架构文档。

[CE006, CE007, CE008, CE009, CE010]
FE002: 客户工作流 / 运行流程

端到端运行流程,展示 OEM 人形机器人制造商如何集成 Prague 的世界模型,以支持泛化操作任务。第 1 阶段之后的所有阶段均为预测;第 1 阶段已在推进。

该流程是对预期产品体验的前瞻推演。Prague Technology 尚无已交付产品。

[CE002, CE003, CE005]
FE003: 关键依赖图

Prague Technology 关键技术和战略依赖的有向无环图。风险集中在 GPU 算力、OEM 数据合作和监管审批节点。

[CE016, CE017, CE018, CE019]

5.3 关键依赖、差异化与知识产权

Prague Technology 的关键技术依赖既带来战略优势,也埋下实质脆弱性。优势在于,Lin Junyang 的架构能力、来自 Alibaba/Qwen 的团队人才网络,以及他与中国 OEM 机器人生态的深度关系,让公司更容易拿到执行世界模型战略所需的人才和数据合作。脆弱性则集中在三项关键技术依赖:(1)GPU 算力受美国对 NVIDIA H100/H200 芯片出口管制限制——Prague 必须依赖 Huawei Ascend NPU 或国产替代方案,而这些方案的每美元 FLOPs 表现更差;(2)与机器人 OEM(AgiBot、Unitree、UBTECH 等)建立真实世界具身 AI 训练数据合作——没有多样化机器人-环境交互数据,世界模型会过拟合仿真,迁移到真实环境时失效;(3)大规模生成合成训练数据的仿真基础设施——自建或授权一个可比 NVIDIA Isaac Sim 或 Genesis 的高保真物理仿真环境,本身就是一项重大技术工程。知识产权方面,截至 2026 年 7 月,Prague Technology 没有已知专利,也没有公开技术论文。如果有知识产权护城河,主要会来自:(a)基于 MoE 的多模态世界模型架构创新,(b)从 OEM 合作中积累的专有具身训练数据,(c)多个 OEM 用训练数据换取持续改进模型后形成的数据飞轮。Anthropic 安全框架和 OpenAI 的 Sora 系统卡都记录了大型生成模型进入物理或高后果领域后的安全考量;Prague 从研究走向部署时,也会面对类似的安全与对齐要求。物理 AI 系统的信任、安全和合规,需要鲁棒性测试、安全认证(工业机器人 ISO 10218、嵌入式系统 IEC 62443),并遵守中国生成式 AI 监管(2023 年 8 月生效的 MIIT《生成式 AI 服务暂行办法》及后续更新)。[CE016, CE017, CE018, CE019, CE020, CE021]

信任 / 质量 / 合规表
控制项 / 认证Prague 状态适用范围缺口
ISO 10218-1/2 工业机器人安全不适用(产品前阶段)制造场景商业部署前必须取得没有认证产品;合规时间表未知;系统级认证需要硬件合作方
IEC 62443 工业网络安全不适用(产品前阶段)工业 IoT 环境中的嵌入式 AI 系统需要该认证没有产品;也未公开安全架构
中国《生成式人工智能服务管理暂行办法》(MIIT,August 2023)产品上线后纳入范围(生成式 AI 模型提供方)任何在中国部署的生成式 AI 都需要国家层面合规合规流程未披露;没有 MIIT 备案迹象
CAC 算法安全评估未知——适用于大规模推荐 / 生成系统影响公共信息或大规模决策的系统需要该评估产品前阶段;目前可能尚不适用,但商业化规模后会适用
模型准确性 / 可靠性 SLA未定义——没有产品生产部署前,OEM 客户会要求可用性、准确性和故障模式保证未发布 SLA、测试框架或安全评估方法

合规表是前瞻性判断;截至 2026 年 7 月,Prague Technology 没有需要认证的产品。

[CE022, CE023, CE024]
FE004: 产品成熟度 / 能力地图

截至 2026 年 7 月,Prague Technology 与顶级世界模型和 VLA 竞争者在四个关键能力维度上的产品成熟度对比矩阵。

Prague Technology 的状态来自观察(尚无产品)。其他公司的状态基于截至 2026 年 7 月的新闻来源和技术出版物。

[CE010, CE011, CE012, CE013]

5.4 路线图、开发阶段与尽调缺口

截至 2026 年 7 月,Prague Technology 没有发布正式产品路线图、开发时间表或里程碑目标。公司成立于 2026 年 5 月,本报告撰写时还不到三个月。参照可比世界模型的开发周期——DeepMind 的 Genie 2 从概念到发布约用了 2 年;Physical Intelligence 的 π0 约用了 18 个月;AMI Labs 目标是在 2027-2028 年推出首个产品——Prague 最早可信的里程碑,是 2027 年中展示概念验证模型,2027 年底或 2028 年初完成首次 OEM 试点集成,并在乐观情景下于 2028-2029 年实现商业可用。公开可识别的团队成员只有 Lin Junyang;团队构成、来自 Alibaba 或学术机构的关键招聘、组织结构均未披露。这个阶段没有技术博客、预印本论文或基准结果并不异常,但也意味着外部几乎没有任何技术证据可用来评估 Prague 的架构路线、训练方法,或攻克关键技术难题的能力。对比基准公司:Physical Intelligence 在成立 6 个月内发布了 π0 论文;Google DeepMind 创建后不久发布 Genie 2;AMI Labs 已发布初步模型卡和博客。Prague 的技术不透明造成完整尽调缺口。关键技术尽调问题包括:(1)架构白皮书或技术备忘录,证明团队的世界模型路线;(2)仿真到现实迁移研究或数据证据,说明团队理解领域鸿沟;(3)现有或计划中的 OEM 数据合作协议;(4)算力策略——Prague 使用 NVIDIA 还是国产芯片,效率差距有多大。[CE025, CE026, CE027, CE028, CE029, CE030]

路线图 / 发布 / 开发阶段表
阶段 / 日期里程碑状态影响来源
May 2026公司注册成立(上海卜拉格科技有限公司)已确认融资阶段;尚无技术工作成果多家新闻来源
June 2026天使轮完成($220M,估值 $2B)已确认资金到位,可开始认真投入模型开发36Kr、MarketScreener、BridgingChina
Q4 2026(估计)团队扩张和算力基础设施搭建未确认——分析师估计关键招聘期;团队质量决定技术可行性根据行业常规推断
H1 2027(估计)首个内部概念验证模型完成训练未确认——估计最早可能的内部研究检查点参照 AMI Labs / Physical Intelligence 时间线
H2 2027(估计)公开技术演示或预印本发布未确认——估计技术路线有效性的首个外部证据历史可比:Physical Intelligence 成立约 18 个月后发布 pi0
2028(估计)首个 OEM 试点集成未确认——估计首次商业验证;需要签署 OEM 合作AMI Labs CEO:「是几年,不是几个月或几个季度」
2029-2030(估计)商业产品可用未确认——估计开始产生收入;假设试点结果成功Humanoid Foundation Model Report 2028-2030 基准情景

第 3 阶段之后的日期是分析师基于可比公司的估计。Prague Technology 尚未发布任何路线图。

[CE025, CE026, CE027, CE028]

5.5 图表

Chapter 06

06客户

6.1 客户基础:零客户与市场机会

截至 2026 年 7 月,Prague Technology 没有客户、没有已签署的 LOI、没有试点协议,也没有披露任何客户对话。因此,本章分析的是潜在客户需求和商业牵引缺失,而不是已经实现的客户指标。对一家 2026 年 5 月成立、产品尚未存在的公司来说,这并不意外。但客户尽调缺口不只是时间问题,而是根本商业不确定性:Prague 必须把 OEM 机器人制造商变成客户,而这类买方在生产硬件中采用未经验证的 AI 组件时出了名保守。潜在客户机会很大:中国人形机器人行业在 2026 年 H1 获得 ¥46 billion($6.4B)投资,更广泛的具身 AI 生态中有 10,000+ 家公司,AgiBot、Unitree、UBTECH、ESTUN、SIASUN 等主要 OEM 品牌都在主动开发或采购 AI 大脑能力。全球范围内,Physical Intelligence 的 π0 已证明,只要能力被可信展示,OEM 制造商会试用 AI 大脑集成。International Federation of Robotics 预计全球服务机器人市场到 2030 年将达到 $68 billion,其中 AI 驱动的认知组件预计会占据重要份额。反面观点是,Prague 面临结构性的启动难题:OEM 客户在承诺数据合作前要求看到可用产品演示,但世界模型的构建又需要 OEM 数据——这形成了一个必须在拿下首个客户前解决的鸡生蛋问题。多位中国 AI 观察者指出,向 OEM 销售世界模型组件,必须证明任务周期时间、错误率下降或替代工人的可量化 ROI;没有产品,Prague 无法给出这些指标。[CU001, CU002, CU003, CU004, CU005, CU006]

客户分层表
分层买方规模 / TAMPrague 当前状态战略优先级
中国 OEM 人形机器人制造商AgiBot、Unitree、UBTECH、SIASUN 的工程副总裁 / 首席机器人官100+ 家主要 OEM;中国 H1 2026 投资 ¥46B零客户——Prague 没有产品主要——中国具身 AI OEM 市场
中国制造自动化(非人形)Foxconn、BYD、CATL 的工厂 IT / 自动化采购团队体量最大的细分;50,000+ 家工厂零客户——通过 OEM 合作方间接进入次要——通过 OEM 渠道合作方
国际 OEM 人形机器人制造商Agility Robotics、Boston Dynamics、Apptronik、Figure AI(海外同行)全球市场;考虑监管和语言壁垒,短期概率 <5%零——不是短期目标长期(2030 年后)
中国政府 / 国防 / 特种机器人国企采购部门和国防机构有潜力,但政治敏感零——未知 Prague 是否打算服务该细分不明——需要政策分析

所有细分规模和状态都是前瞻性判断。截至 2026 年 7 月,Prague Technology 在任何细分都没有客户。

[CU001, CU007, CU008]
FU001: 客户旅程图

Prague Technology 潜在 OEM 人形机器人客户的旅程图,展示从认知到生产采用的各阶段。「认知」之后均为预测;Prague 尚无客户走过第一阶段。

第 2 阶段及以后均为预测。截至 2026 年 7 月,Prague 尚无任何阶段的 OEM 客户。

[CU003, CU022]
FU002: 采用 / 部署漏斗

从可触达 OEM 总盘子到生产客户的预测获客漏斗,基于 TARS AI 和 Physical Intelligence 的采用模式可比样本,展示各阶段预期转化率。

数字为分析师基于中国 OEM 市场规模的估算;第 3 阶段以下当前值为 0 属已确认事实(Prague 未披露任何客户对话)。120 家中国 OEM 的估算来自 AInChina 2026 报告;20 家主动评估方是对正在评估各类 AI 大脑的整体市场估算。

[CU001, CU009, CU010]

6.2 潜在客户细分与买方分析

Prague Technology 的目标客户细分,可以从公司的世界模型论点和中国人形机器人 OEM 市场结构中推断。首要目标细分是中国 OEM 人形机器人制造商,特别是面向制造、物流和工业应用造机器人的公司,它们需要用通用世界模型替换自有 AI 规划模块。该细分包括:AgiBot(行业领军者,2026 年 AgiBot World Challenge 吸引 526+ 支全球团队,已在汽车制造中商业部署)、Unitree Robotics(2026 年 IPO 审核获通过,H1 收入约 ¥500M,正在扩展企业自动化)、UBTECH Robotics(面向制造和酒店业的企业人形机器人),以及数十个较小品牌。次要目标细分是正在采用具身 AI 的中国制造业:电子制造商(PCB 组装、元件放置)、汽车制造商(点焊、装配线协作)和家电制造商(质检、总装)。第三层细分是全球市场:中国以外的 OEM 制造商,包括 Boston Dynamics(Hyundai)、Agility Robotics(Amazon)和 Apptronik(Samsung),代表更长期机会,但需要解决跨语言和监管障碍。付款方是 OEM 公司;买方(技术决策者)是 VP Engineering 或 Chief Robotics Officer;用户则是机器人本身(在端侧运行 Prague 世界模型)。这种 B2B 企业销售周期从首次接触到签约通常需要 6-18 个月,相对 Prague 的资金续航期偏长。Prague Technology 没有披露目标客户名单或销售策略,也没有宣布与任何 OEM 客户建立关系。[CU007, CU008, CU009, CU010, CU011, CU012]

客户增长 / 采用轨迹表
周期预期客户状态所需关键里程碑概率来源依据
July 2026(实际)零客户;无产品无——仅处于研究阶段已确认多家新闻来源
H1 2027(估计)零客户;仅有内部 PoC 模型首次内部模型演示中(60%)参照 Physical Intelligence 时间线
H2 2027(估计)首个 OEM 数据合作方(非付费)签署 OEM 数据共享协议低(30%)可比案例:AMI Labs 路径
H1 2028(估计)1-3 个 OEM 试点集成(无收入)OEM 实验室环境试部署低(25%)以 TARS AI 时间线为基准
H2 2028(估计)首个产生收入的 OEM 合同OEM 生产部署并收取许可费低(20%)Humanoid Foundation Model Report 2028-30 基准情景
2030(乐观情景)10-20 家 OEM 客户;经常性 ARR多个 OEM 机器人群部署很低(10%)分析师对世界模型商业化规模的共识

2026 年后的所有行都是分析师估计。概率是主观评估;不构成投资建议。

[CU003, CU014, CU019, CU020]
具名客户证明表
客户名称状态集成类型结果证据质量来源
[无客户——截至 2026 年 7 月,Prague Technology 没有任何具名客户]N/AN/AN/AN/A多家新闻来源确认其仍处产品前阶段
TARS AI(竞争对手,代理证据)生产试用(20+ 家企业)AWE 3.0 世界模型集成进 Foxconn 产线已商业部署;资料引用 Foxconn 生产集成中——已有报道,但销量数据未确认Gasgoo、EqualOcean、CNTechPost 2026
Physical Intelligence(美国代理样本)早期试验π0 模型与合作 OEM 机器人集成,用于多任务操作实验室中泛化到 68+ 项任务;尚未达到商业规模中——同行评审论文 + 公司博客Physical Intelligence 官方博客 2024
AgiBot(潜在客户,无协议)仅为潜在 OEM——没有 Prague 协议不适用(没有 Prague 产品)N/AN/A——仅为潜在客户AgiBot 官方沟通 2026

Prague Technology 没有任何客户证明。第 2-4 行是用于基准比较的竞争对手和代理数据;AgiBot 行只是潜在客户信号,没有确认关系。

[CU015, CU016, CU017, CU021]
FU003: 客户验证矩阵

截至 2026 年 7 月,Prague Technology 与主要世界模型竞争对手的客户验证对比矩阵,从五个维度评估客户验证强度。

TARS AI 与 Foxconn 的集成见于新闻报道,但 Foxconn 尚未正式确认。所有其他数值均来自公开信息。

[CU015, CU016, CU017, CU019]

6.3 客户验证与采用轨迹

截至 2026 年 7 月,Prague Technology 没有客户验证。没有生产部署、没有试点项目、没有标杆客户、没有客户证言,也没有已签署的意向书。客户验证缺口是完全的。不过,第三方需求信号显示,市场确实对用于具身 AI 的世界模型方案感兴趣:AgiBot 的 World Challenge(2026 年 6 月,ICRA Vienna)吸引了来自 38 个国家的 526 支团队,说明 OEM 客户需要 AI 研究者开发操作方案。Unitree 的 IPO 审核获通过以及 2026 年 H1 的 ¥500M 收入,证明中国 OEM 市场真实存在并在增长。TARS AI 在客户验证上走得最远:据称 TARS 的 AWE 3.0 世界模型已集成进 Foxconn 产线,并在中国吸引 20+ 个企业客户试用——这为 Prague Technology 树立了必须复刻的蓝图。反方视角很尖锐:即便 TARS AI 已融资 $697M 且拥有生产系统,也只拿到试用,而不是长期生产合同。Prague 还没有产品,即使按乐观时间表,也很可能离首笔收入还有 2-3 年。curionic.net 竞争分析指出,中国机器人 OEM 客户评估 AI 大脑时看四项指标:(1)操作成功率(生产要求 >95%),(2)延迟(工业任务要求 <200ms),(3)跨 SKU 泛化,(4)与现有硬件的集成成本——Prague 目前都无法证明。Physical Intelligence 的 π0 很有参考意义:尽管展示了 68+ 项任务泛化能力,截至 2026 年中,其 OEM 客户管线仍处于“早期试用”,说明世界模型 AI 大脑的商业采用周期按年计算,而不是按月。[CU014, CU015, CU016, CU017, CU018, CU019]

留存 / 重复使用 / 满意度表
指标Prague Technology 当前值基准(TARS AI / Physical Intelligence)重要性
NRR(净收入留存)N/A——零收入TARS AI:未披露;PI:尚无经常性收入NRR >120% 是客户扩张价值的关键信号
GRR(总收入留存)N/A——零收入两者均 N/AGRR <80% 代表流失风险;收入前阶段不适用
流失率N/A——零客户嵌入式 AI 大脑天然较低(切换成本高)集成后几乎不可能切换——世界模型嵌入固件
合同期限 / 续约N/A——零合同企业 AI:初始通常 1-3 年,自动续约初始合同期限长,能降低流失风险,但牺牲短期灵活性
客户满意度N/A——零用户Physical Intelligence:正向(论文质量信号)Prague 没有 NPS、推荐语或客户反馈信号

Prague Technology 没有客户或收入,因此所有留存指标均为 N/A。表格记录的是商业运营开始后的预期留存模型。

[CU022, CU023, CU024]

6.4 留存、集中度与扩张风险

Prague Technology 的留存、集中度风险和扩张路径分析完全是前瞻性的——没有可留存的客户,没有客户集中度,也没有扩张指标。不过,世界模型授权业务的结构特征会带来可预判的未来动态,现在就可以分析。留存方面,一旦 OEM 机器人制造商把 Prague 的世界模型集成为 AI 大脑层,切换成本会很高——模型会用 OEM 专属数据微调,集成 SDK 会嵌入机器人固件,迁移到竞争对手平台还需要大量工程资源重新训练。这会自然形成有粘性的经常性收入模式,有利于留存。集中度方面,最初 3-5 个 OEM 客户按结构就会贡献 100% 收入;这种极端集中,是 B2B 企业 AI 初创公司商业化前 12-24 个月的常态。缓释办法不是押注单一垂直行业,而是在汽车、电子和物流等垂直领域拿到多元化 OEM 客户。扩张方面,世界模型授权的“先落地再扩张”路径会从(1)单一工厂中 1-2 款机器人型号的初始 API 试用,推进到(2)在某个 OEM 机器人全产品线全面部署,再到(3)扩展至新的 OEM 制造工厂,最后到(4)签署数据共享协议,推动跨 OEM 模型改进。结构性挑战在于,扩张模式需要积累多样化训练数据,而这又需要更多 OEM 伙伴,形成奖励早期市场领导者的自我强化动态(中国的 TARS AI、美国的 Physical Intelligence)。如果 TARS AI 在 Prague 上线前成功成为中国 OEM 的主导世界模型标准,Prague 面对的就是已被竞争对手生态锁定的市场——这是本章的负面情景。[CU022, CU023, CU024, CU025, CU026, CU027]

扩张与集中度风险表
风险因素Prague Technology 暴露严重程度缓释措施
零客户 → 上线前 100% 集中完全暴露——没有客户就没有分散危急接受这是第一年现实;从首个客户起就优先扩大 OEM 合作方广度
Prague 上线前市场被 TARS AI 锁定TARS AI 已投产,并比 Prague 早 2-3 年积累 OEM 数据飞轮聚焦 TARS 难以复制的差异化世界模型能力;瞄准 TARS 尚未锁定的细分市场
早期客户依赖 Tencent 生态Tencent 的战略投资可能打开通向 Tencent 关联 OEM(CEVA、Weixin 硬件)的渠道利用 Tencent 关系做分发,同时保持供应商中立姿态
单一 OEM 垂直领域锁定(如仅汽车)未声明垂直重点;可能过度押注 TARS AI 的汽车滩头阵地首批 OEM 合作中明确分散到 2-3 个垂直领域
落地后扩张被 OEM 数据所有权争议卡住OEM 贡献给 Prague 的训练数据可能引发 IP 冲突,前提是 OEM 可主张模型改进的共同所有权在首批 OEM 协议中明确数据许可条款和模型改进所有权

所有暴露评级都是前瞻性判断。由于没有客户,目前不存在实际客户集中度风险。

[CU025, CU026, CU027, CU028]
FU004: 留存 / 复购队列

Prague Technology 未来 OEM 客户的预测队列留存框架,展示从首次商业部署(估计 2028 年)开始的年度队列预期留存特征。当前数据完全是前瞻预测。

所有队列预测都是分析师的推测性估算。Prague 尚无客户。

[CU022, CU023, CU025]

6.5 图表

Chapter 07

07风险

7.1 风险概览与严重性排序

Prague Technology 的风险画像由五个结构性特征塑造:(1)截至 2026 年 7 月仍无产品,所有商业和运营风险都只是前瞻性的;(2)估值溢价极端,成立 3 个月、只有 1 名具名员工却估值 $2B,这为后续融资设置了很高门槛;(3)作为一家在中美科技脱钩环境中开发前沿技术的中国 AI 公司,地缘政治暴露很高;(4)公司高度依赖单一创始人 Lin Junyang 的架构能力和人脉;(5)收入前开发周期长达 2-4 年,公司必须在没有商业验证的情况下靠 $220M 活下来。最严重的风险,是 GPU 出口管制(NVIDIA H100/H200/H20 限制)与世界模型训练算力强度叠加——这可能迫使公司依赖中国国产 GPU 替代方案,而其性价比低 50-70%,从而拉长开发周期并提高资金需求。第二大风险是 TARS AI 的竞争性市场锁定:它领先 2-3 年,已有生产部署,还在积累 OEM 训练数据,可能在 Prague 上线前拿下中国前 5-10 家 OEM 客户,形成很难逆转的赢家拿走大部分格局。第三是关键人物风险:Prague Technology 似乎只有一名具名员工(Lin Junyang),公司对单一个体的依赖是灾难级的。技术风险(世界模型开发失败)、财务风险(烧钱超过计划)和监管风险(中国 AI 规则变化)构成前六大风险的其余部分。本章列出的每项风险都可能击穿投资论点;标准投资条件监控框架必须为每个维度设置预警触发器。[CR001, CR002, CR003, CR004, CR005, CR006]

监管 / 法律风险登记表
风险类别可能性(1-5)影响(1-5)缓释成熟度(1-5)投资影响
美国 GPU 出口管制封堵 NVIDIA H100/H200/H20 获取渠道地缘政治 / 监管552造成 40-60% 算力效率损失;时间线拉长 12-18 个月;资本需求上升
中国生成式 AI 监管(MIIT 暂行办法)合规成本监管433合规搭建成本 $1-5M;持续报告负担;不构成生存风险,但会推迟商业发布
IP 风险——没有专利,架构方法已在公共领域法律431TARS AI、AMI Labs、Google 都能复刻做法;护城河在执行、数据和团队,不在专利
中国数据安全法规(PIPL、DSL)对机器人传感器数据的约束法律 / 监管322机器人传感器数据一旦含位置或图像数据,可能需要数据本地化和同意机制
中美脱钩下专门针对中国 AI 公司的制裁地缘政治251低概率但冲击大:可能限制国际资本、云访问和出口市场
Alibaba 竞业限制和商业秘密主张法律231Lin Junyang 于 2026 年 3 月离开 Alibaba;Alibaba 可能就 Qwen 相关工作主张商业秘密被不当使用

所有可能性 / 影响评分均为分析师基于 1-5 分的主观估计。来源:Rimon Law、AI Governance、MIIT 指引、BIS 出口管制规则。

[CR007, CR008, CR009, CR010]
FR001: 风险热力图

Prague Technology 八项主要风险的热力图,按发生概率(1-5)、财务影响(1-5)、缓释成熟度(1-5)和剩余暴露(1-5)评分。剩余暴露越高,越需要优先处理。

所有评分均为分析师主观估算。剩余暴露 =(发生概率 × 影响)/ 缓释成熟度;数值越高越糟。

[CR001, CR002, CR003, CR007, CR016]

7.2 监管、法律与地缘政治风险

Prague Technology 面临跨多法域、复杂且不断变化的监管环境。在中国,2023 年 8 月生效的《生成式 AI 服务暂行办法》(MIIT/CAC)要求 AI 模型提供商向监管部门备案、开展安全评估,并遵守内容生成限制。Cyberspace Administration of China(CAC)和 MIIT 随后的指引进一步扩大了这些要求。对部署在实体机器人中的具身 AI 世界模型来说,监管面更宽:工业机器人安全(GB/T 标准、中国标准下的 ISO 10218 等同要求)、机器安全认证,以及在接近关键基础设施的应用中可能触发 Cybersecurity Law 下的网络安全审查。美国出口管制风险很重:US Department of Commerce 通过 BIS Entity List、Foreign Direct Product Rule 和针对芯片的出口许可要求,逐步限制 NVIDIA AI 芯片出口到中国。H100、A100、H20 芯片出口中国均受限制或管制。2025-2026 年中美科技紧张升级进一步收紧限制,对仍可流向中国的 NVIDIA GPU 芯片还叠加了约 25% 的额外关税。这迫使 Prague Technology 依赖 Huawei Ascend 910B/920 芯片或国产同类方案,而这些方案每美元 FLOPs 比 NVIDIA H100 少 40-60%,会实质性抬高训练成本并拉长时间表。法律风险包括:(1)知识产权风险——Prague 没有专利,其核心技术(世界模型架构)很大程度上通过学术出版物处于公共领域;竞争对手可以在没有知识产权壁垒的情况下复制 Prague 路线;(2)劳动法风险——中国对 AI 研究人员竞业限制的严格规则,可能让从 Alibaba 挖人产生人才获取冲突;(3)数据法风险——处理包含个人信息的机器人传感器数据时,中国《个人信息保护法》(PIPL)和《数据安全法》会带来义务。地缘政治风险还延伸到投资层面:如果美国制裁专门指向中国 AI 公司,HongShan 的美国 LP 基础和 Tencent 的跨境活动可能给投资者带来二级制裁风险。[CR007, CR008, CR009, CR010, CR011, CR012]

运营 / 质量 / 安全风险登记表
风险类别可能性(1-5)影响(1-5)缓释措施状态
世界模型技术路线失败,或时间线滑坡 2 年以上技术 / 研发35还没有——仍在预研没有缓释措施;投资逻辑押注研究成功
没有 OEM 数据伙伴,仿真到现实的迁移鸿沟无法解决技术 / 数据44合成数据生成;学术数据集没有 OEM 数据伙伴;缺口未缓释
推理延迟过高,无法支撑实时机器人控制(30-100Hz 缺口)技术 / 运营44模型蒸馏;面向特定硬件优化产品前阶段尚未缓释
AI 安全失效:世界模型导致机器人伤人或损坏财产安全 / 责任25安全认证流程;责任保险尚未开始——还没有可认证产品
算力基础设施中断(Huawei GPU 供应链失效)供应链24算力双源(云 + 本地)单一来源风险;未披露备份
网络安全泄露:世界模型或训练数据被盗安全24代码安全、访问控制、零信任架构未知——未披露安全态势

Prague 没有产品,也没有运营基础设施,所有运营风险都尚未缓释。评分为前瞻性估计。

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

因果链图展示 Prague Technology 的主要风险如何层层传导为次级影响。风险传播路径显示,地缘政治风险是杠杆最高的根因。

[CR007, CR009, CR024, CR025]

7.3 运营、技术与依赖风险

Prague Technology 的运营和技术风险直接来自产品开发阶段。核心技术风险是世界模型开发失败:截至 2026 年 7 月,具身 AI 世界模型仍是未解的研究问题——DeepMind 的 Genie 2 证明了世界模型可用于游戏环境,但商业可靠性水平的实体机器人控制,全球还没有任何公司做到。如果技术路线失败,或开发时间远长于预期,Prague 的 $220M 资金会在实现商业验证前耗尽。仿真到现实迁移鸿沟是最具体的运营风险:物理仿真生成的合成训练数据无法捕捉真实世界传感器噪声、表面属性和环境变化,因此需要从 OEM 机器人伙伴处大规模采集真实世界数据。没有 OEM 数据伙伴(截至 2026 年 7 月 Prague 为零),世界模型可能无法从仿真泛化出去。推理延迟是关键运营约束:机器人控制要求反应动作达到 30-100Hz,但商业规模的世界模型推理通常只有 2-10Hz——存在 10-30x 差距,必须在任何 OEM 部署前通过模型蒸馏、硬件优化或混合架构补上。依赖风险集中在三类关键外部方:(1)Tencent——提供算力入口和潜在客户渠道;如果关系恶化或 Tencent 监管处境变化,就会产生依赖风险;(2)中国国产 GPU 制造商(Huawei、Biren、Cambricon)——出口管制迫使 Prague 依赖性能更弱的国产芯片;如果国产 GPU 供应中断(地缘政治、生产失败),Prague 的训练排期会被打乱;(3)OEM 机器人数据伙伴——Prague 需要 3-5 家 OEM 公司承诺训练数据,世界模型才可能泛化;任何拒绝都会留下单靠仿真无法填补的数据缺口。来自 Gaorong 和 HongShan 的伙伴 / 依赖风险较低——两者都是中国一线顶级 VC,LP 基础扎实,投资周期较长。投资者风险不是被放弃,而是如果 2027 年里程碑延误,可能施压公司转向或削减成本。[CR015, CR016, CR017, CR018, CR019, CR020]

合作伙伴 / 依赖风险登记表
合作伙伴 / 依赖项依赖类型关键性(1-5)可替代性风险
Tencent(算力访问 + 客户渠道)战略投资方 + 云算力4中等(Alibaba Cloud、Baidu AI Cloud)如果 Tencent 关系恶化,Prague 会失去算力补贴和客户转介
Huawei Ascend(国产 GPU)主要算力硬件5低(Biren、Cambricon——性能较弱)如果 Huawei 供应链受扰(美国制裁、生产失败),Prague 的训练排期会被打乱
OEM 机器人数据伙伴(AgiBot、Unitree、UBTECH——尚未签约)训练数据5没有——现实世界具身数据若要规模化,没有替代来源如果到 2027 年中仍无法锁定 3+ 家 OEM 伙伴,世界模型泛化会失败
Gaorong Ventures(领投方)资本提供方4高(多家 VC 对世界模型感兴趣)风险:如果 Gaorong 收缩 AI 组合,后续支持可能变弱;强联合投资方阵容可缓释
Lin Junyang(创始人兼唯一具名员工)技术愿景 + 投资方信任5目前没有如果 Lin 无法履职,Prague Technology 现有形态将无法运转

依赖风险评分为前瞻性估计。Prague Technology 未披露正式依赖评估。

[CR019, CR020, CR021, CR022]
人员 / 执行风险登记表
风险严重性(1-5)概率(1-5)影响缓释要求
Lin Junyang 是唯一具名员工(关键人风险)53没有 Lin,公司无法运转;这个阶段也无法制定接班计划立即招聘至少 1 名联合创始人或 VP 级技术负责人
组建团队失败:无法从 Alibaba / ByteDance 吸引顶尖 AI 研究员43模型开发延迟;架构弱于人员更强的同行确认招聘漏斗和薪酬结构,并对标 TARS AI 与 AMI Labs
组织设计失败:单一技术创始人,没有运营型联合创始人43没有 CFO、COO、VP Sales——运营缺口会层层拖慢执行6 个月内确定非技术联合创始人或 COO 人选
Alibaba、ByteDance 或 Tencent 用更高稳定性挖走人才34团队组建后,人才流失风险将持续存在股权授予、按里程碑发放留任奖金;没有产品时间线时难度较高
Lin Junyang 因 Alibaba IP 争议产生声誉风险42如果 Alibaba 提起竞业或 IP 主张,Lin 可能在法律上无法开展某些技术工作由独立律师确认 Lin 离职法律清理情况和竞业范围

所有评分均为分析师估计。Prague Technology 除 Lin Junyang 外未披露团队构成。

[CR025, CR026, CR027]
FR003: 依赖关系图

有向依赖关系图按关键性排列 Prague Technology 的外部依赖。所有关键路径都汇向 Lin Junyang 和 GPU 算力——这两个集中度最高的风险。

[CR019, CR020, CR021, CR022]

7.4 财务、执行与人员风险

财务风险主要来自烧钱速度不确定,以及前沿 AI 开发的资本强度。如果月度烧钱达到 $10-15M(100+ 名研究员加重算力支出完全可能),$220M 天使轮只能支撑 14-22 个月,公司需要在 2027 年底前完成 Series A 或战略轮。按 2027 年市场条件再融资风险很大:如果 AI 热情降温、估值预期正常化,或 Prague 没有达到有意义的技术里程碑,后续融资可能变成平轮或 down round。估值从 $2B→$1B 的 down round 会造成严重稀释,并损害声誉。关键人物风险是生死级的:Prague Technology 目前似乎只有一名公开具名员工(Lin Junyang)。如果 Lin 无法继续履职(健康、法律限制、被战略方收购,或转投竞争项目),Prague Technology 事实上会以当前形态停止存在。技术愿景和投资人关系都集中在单一创始人身上,造成公司层面的单点失效。各维度执行风险都很高:公司必须一边搭建世界模型研究团队,一边管理投资人关系、OEM 商务拓展、监管合规和算力基础设施——全部同时推进——这对经验丰富的连续创业者也很难。Lin Junyang 的背景偏技术(模型架构、训练基础设施),而不是运营(团队搭建、企业销售、监管导航)。截至 2026 年 7 月,公司没有披露任何联合创始人或 C-suite 团队(CFO、COO、VP Engineering、VP Sales),这是运营执行红旗。来自资源更充足的既有巨头(Alibaba、ByteDance、Tencent、BAIDU)的 AI 人才竞争会持续带来招聘和留才挑战——这些公司都支付顶级薪酬,也比一家产品前初创公司更稳定。财务、执行和人员风险叠加后,即使按中国早期 AI 初创公司标准,Prague Technology 也是高风险投资。[CR024, CR025, CR026, CR027, CR028, CR029]

缓释与止损标准表
风险领先指标止损标准(投资逻辑破裂)监控频率
GPU 算力获取季度:单位算力成本相对 NVIDIA 基准;Huawei GPU 可用性如果国产 GPU 性价比落后 NVIDIA H100 超过 3x,世界模型时间线将延至 2030 年之后——投资逻辑破裂季度
TARS AI 市场锁定半年:TARS 量产客户数相对 20+ 目标;世界模型 OEM 市场份额如果 TARS AI 在 Prague 模型发布前签下 10+ 家中国一线 OEM,可服务市场会严重缩水半年
烧钱速度与资金续航月度:确认后的烧钱速度相对 $5M/月基准;现金余额;下一轮时点如果烧钱超过 $10M/月且没有产品里程碑,资金续航降至 18 个月以下——立即触发后续融资月度
OEM 数据伙伴推进季度:OEM 数据伙伴沟通数量、已签协议如果到 2027 年 Q2 仍未签下 OEM 数据伙伴,世界模型泛化时间线会滑到 2029+——投资逻辑明显转弱季度
关键人可用性月度:确认 Lin Junyang 仍实际履职 CEO,且无法律 / 个人限制如果 Lin Junyang 无法履职且没有替代技术负责人,投资逻辑完全破裂月度
中国 AI 监管升级季度:MIIT 和 CAC 涉及具身智能或生成式 AI 模型提供商的新规如果中国 AI 监管要求 AI 大脑在物理系统中失效时承担产品责任,且没有保险市场承接,部署时间线将无限拉长季度

止损标准是监控阈值;一旦越线,应立即退出或实质性重写投资逻辑。

[CR001, CR028, CR029, CR030]

7.5 图表

Chapter 08

08估值

8.1 投资论点与估值背景

Prague Technology 的投资论点建立在三根支柱上:顶级创始人、极热市场,以及可能在竞争整合前成形的可防守技术护城河。Lin Junyang 曾带领 Alibaba 的 Qwen 大语言模型系列成为中国能力最强的开源 LLM,证明了技术深度,也证明了大规模组织执行能力。他在 2026 年初离开 Alibaba,转向具身智能和世界模型研究,路径类似 Yann LeCun(AMI Labs)、Fei-Fei Li(World Labs)和 Pieter Abbeel(Covariant/Physical Intelligence)——这些人都仅凭创始人声誉就拿到溢价估值。全球具身 AI 和世界模型市场正在快速增长;CB Insights 的 State of Venture Q2 2026 报告指出,AI 大额融资现在占所有已部署风险资本的 81%,连续第二个季度融资超过 $200 billion。在中国,6 个月内诞生了 15 家具身 AI 独角兽,说明该领域资本充裕。Prague 以 $220M 融资对应 $2B 投后估值,意味着投资人持股约 11%,对这种规模的首笔支票来说稀释偏轻。反面论点同样有力:Prague 没有产品、没有收入、没有客户,成立日期是 2026 年 5 月,可能是史上最年轻的独角兽之一。估值几乎 100% 建立在对未来交付的信任上。执行风险很尖锐。[CV001, CV002, CV003, CV004, CV005, CV006]

建议摘要表
维度评估备注
总体结论中性 / 有条件观察创始人强,但 $2B 估值下前收入风险极高
信心等级公司成立 2 个月;没有产品、收入或客户
估值立场相对可比公司属合理估值,已计入中国风险折价AMI Labs $3.5B;World Labs $5B 目标;Prague $2B 处于低端
入场纪律需要严格条款约束按比例跟投权、信息权、任何降估值轮的反稀释
持有期4-6 年基于可比公司具身智能商业化时间线
投资逻辑破裂观察期12 个月主要里程碑(演示、合作)预计到 2027 年中出现

评估基于截至 2026 年 7 月的公开信息;公司仍处于产品前阶段。

[CV025, CV026, CV027, CV028]
正方 / 反方投资逻辑表
投资逻辑反方逻辑
Lin Junyang 打造 Qwen 系列——中国 #1 开源 LLM——证明其拥有世界级技术执行力Lin 从未独立带过公司;从公司内部团队负责人转为创业公司 CEO,难度不低
具身智能市场是下一波数万亿美元浪潮;中国工业机器人世界模型的先发优势很大尚未验证产品市场匹配;截至 2026 年中,「具身智能」仍是趋势标签,不是中国已被验证的收入品类
$220M 融资来自 Gaorong、HongShan(Sequoia China)和 Tencent,提供 24+ 个月资金续航投资方全是中国机构;缺少西方一线跨界投资方,未来西方 IPO 或并购选择受限
$2B 估值低于 AMI Labs($3.5B)和 World Labs($5B);从创始人履历看,Lin 在中国可能不逊色甚至更强中国风险溢价(地缘政治、GPU 获取、监管)使其相对西方 AI 实验室估值折价,属合理
世界模型技术对中国工业 AI 雄心具战略必要性;可能获得国家支持与国家目标绑定会带来集中风险:战略可能转向服务国家优先级,而不是商业回报
来自 Alibaba Tongyi 实验室的技术团队带来大规模模型训练经验,可直接迁移到具身智能团队画像未披露;具身智能硬件和机器人集成的关键招聘可能需要 12-18 个月
开源 LLM 传承(Qwen)可能撬动社区驱动开发和人才获取开源模型面临货币化挑战;如果核心产出公开发布,商业护城河不清晰
大型工业玩家(Huawei、Xiaomi、DJI、BYD)的行业并购提供多条退出路径若由中国国有相关实体收购,跨境限制会压低财务投资人的回报

正反投资逻辑基于已核实公开事实;带有判断的推测已明确标注。

[CV003, CV004, CV005, CV006, CV007, CV008]
FV001: 推荐逻辑

决策逻辑流展示创始人质量、市场背景、估值可比样本和风险因素如何合并,最终给出 Prague Technology 的中性投资结论。

[CV025, CV026, CV037]

8.2 可比估值分析

对 Prague Technology 最有参考价值的可比公司,是那些收入前、仅凭团队质量和市场潜力获得估值的世界模型与具身 AI 实验室。AMI Labs 是图灵奖得主 Yann LeCun 在 2025 年底离开 Meta 后创办的世界模型初创公司,2026 年 3 月以 $3.5 billion 投前估值融资 $1.03 billion——这让 Prague 的 $2 billion 投后估值看起来保守了 40-75%。Fei-Fei Li 创办的 3D 世界模型公司 World Labs,在 2026 年初进行目标估值 $5 billion 的融资,其中一部分为 $1 billion,也再次显著高于 Prague。Physical Intelligence(Pi)是美国机器人基础模型公司,由 Jeff Bezos、Tiger Global 和 Sequoia 支持,已融资超过 $400 million,估值估计在 $2 billion+ 以上。在中国,据称 TARS Group(人形 AI)估值高于 $2 billion;Zhiyuan Robotics 和 Unitree Robotics 各约 $1.5 billion。这些比较把 Prague 放在全球世界模型同业的低端,尽管 Lin Junyang 在其领域内的资历可以说强于部分可比创始人。折价可能反映中国特定风险:中美地缘政治紧张、GPU 出口限制和监管不确定性,都会让西方投资者保守定价。Sequoia Capital 在 2026 年 5 月的 AI Ascent IV 评论中说,“AI 是一场计算革命——不是更快的马,而是汽车,而且汽车已经来了”,验证了宏观机会;但他们也强调,创始人必须“从客户出发构建护城河”,这一点 Prague 还没有证明。[CV008, CV009, CV010, CV011, CV012, CV013]

可比公司估值表
公司地区方向阶段最近估值(2026)已融资金额与 Prague 的相关性
AMI Labs (Yann LeCun)法国 / 全球世界模型(JEPA 架构)种子轮 / 前收入$3.5B 投前(2026 年 3 月)$1.03B功能最接近的可比公司——世界模型、前收入、创始人驱动;Prague 估值折价 43%
World Labs (Fei-Fei Li)美国3D 世界模型 / 空间 AI早期收入$5B 目标(2026 年 2 月轮次)$1.23B+World Labs 是世界模型可比公司;Fei-Fei Li 的品牌更广,Lin 的中国纵深更强;Prague 折价 60%
Physical Intelligence (Pi)美国具身智能 / 机器人基础模型早期产品~$2B+(估计)$400M+直接具身智能可比公司;估值与 Prague 相近;总部在美国,地缘约束更少
TARS Group中国人形机器人 / AIA/B 轮~$2B+(估计)~$300M+中国具身智能可比公司;更偏硬件,Prague 更聚焦模型
Zhiyuan Robotics中国人形机器人B 轮~$1.5B~$200M中国人形机器人可比公司;已到产品阶段,而 Prague 仍在产品前,意味着 Prague 估值带溢价
Unitree Robotics中国消费 / 工业机器人成长期~$1.5B~$200M更偏硬件;直接可比性较弱,但说明中国市场愿意在这一区间为具身智能买单
Prague Technology中国具身智能 / 世界模型种子轮(前收入)$2B 投后(2026 年 6 月)$220M标的公司——在全球世界模型同行中估值最低,尽管创始人资历可比

估值来自 TechCrunch、Observer、Reuters、CBInsights 以及可用公开备案;估计估值已标注。所有数据截至 2026 年中。

[CV008, CV009, CV010, CV011, CV012, CV013]
FV002: 估值敏感性

截至 2026 年中,Prague Technology 与五家可比的收入前或早期 AI 模型 / 具身 AI 公司的投后或投前估值($B)对比,显示 Prague 处在同业低端。

估值为截至 2026 年中公开来源报道或估算的投后或投前估值。Prague Technology $2B 为已确认投后估值;AMI Labs $3.5B 为投前估值;World Labs $5B 是报道中的目标值,尚未确认。

[CV008, CV009, CV010, CV011, CV012, CV038]
FV003: 估值 / 回报区间

Prague Technology 在乐观、基准和悲观情景下的概率加权退出估值与回报倍数区间,并同时展示加权期望值。

区间为分析师基于可比公司轨迹分析得出的估算。Prague Technology 未披露任何财务预测。进入估值假设为投后 $2B。

[CV016, CV017, CV018, CV019, CV020, CV039]

8.3 情景分析与回报画像

在 Prague Technology $2 billion 估值上做假设投资,可以用三个情景框定结果区间。乐观情景(25% 概率)下,Lin Junyang 在 18 个月内交付可用的具身 AI 基础模型,到 2027 年底前拿下两个或更多标杆工业或机器人合作伙伴,并乘着具身 AI 市场增长在 2029 年达到 $15-20 billion 估值,对当前入场价产生 7-10x 回报,并可能通过部分二级交易或 pre-IPO tender 退出一部分。基准情景(45% 概率)下,Prague 完成模型开发,但受市场碎片化、数据获取限制和中国机器人 OEM 不愿分享专有数据集影响,商业化推进困难;2030 年估值达到 $5-7 billion,意味着入场回报 2.5-3.5x——对风险投资来说合理,但考虑风险溢价并不突出。悲观情景(30% 概率)下,产品时间表延误超过 24 个月,美国出口管制限制关键算力硬件,某个竞争性中国基础模型率先获得商业牵引,或关键团队成员离职削弱投资论点;在这种情景中,公司以 down round 融资,或以 $500 million-$1 billion 被收购,按 $2 billion 入场会造成重大损失。按概率加权的预期回报约为 3 倍毛回报,未扣费用和 carry;对一家没有收入、没有产品、成立于 2026 年 5 月的公司来说,这个回报相对风险画像偏弱。持有 10% 股权的投资者,只有在乐观情景下才能获得有意义的绝对回报;悲观情景中,$220M 支票会严重减值。在这一估值水平上,入场纪律至关重要:必须谈判更强的信息权、pro-rata 和反稀释条款。[CV016, CV017, CV018, CV019, CV020, CV021]

乐观 / 基准 / 悲观情景表
情景概率关键假设2029-2030 年估值基于 $2B 入场的回报倍数
乐观25%世界模型产品到 2027 年 Q2 上线;2027 年底前拿下 2+ 家主要 OEM 伙伴;Lin 招到 5+ 名顶尖研究员;通过国产供应或政府豁免锁定 GPU 获取$15B-20B7x-10x
基准45%产品到 2027 年 Q4 上线;受数据和合作挑战拖累,商业牵引有限;研究产出稳定;尚无战略收购方;核心团队留存$5B-7B2.5x-3.5x
悲观30%产品延迟超过 24 个月;降估值轮低于 $1.5B;关键人员离职;GPU 短缺恶化;中国生成式 AI 监管摩擦加剧$0.5B-1.5B0.25x-0.75x(亏损)

概率估计为分析师判断;公司未披露任何财务模型。估值预测为示意区间,基于可比公司轨迹。

[CV016, CV017, CV018, CV019, CV020, CV021]
FV004: 投资 KPI

概括 Prague Technology 截至 2026 年 7 月投资画像的关键绩效指标。

KPI 值来自公开来源和可比分析;没有公司披露的财务数据。

[CV001, CV002, CV025, CV040]

8.4 建议、尽调事项与论点击穿点

对 Prague Technology 的总体投资结论是中性——更适合有条件持有或放入观察名单,而不是直接投资或直接放弃。创始人质量和市场时点确实出色;Lin Junyang 亲手搭建 Qwen、把它做成中国领先开源 LLM,这一资历是全球具身 AI 创始人中很少有人能匹配的。相较可比公司,$2 billion 入场估值并不过分。然而,公司极其年轻(2026 年 5 月成立)、产品牵引为零、暴露于中国监管和地缘政治风险,融资结构又由中国风险资本主导、未披露西方 crossover 或战略投资者,这些都让当下很难有信心定价风险。中性立场反映的是:机会真实存在,但未来 6-12 个月会出现更好的证据节点——专有世界模型架构演示、与主要机器人 OEM 的战略合作、首次披露研究出版物,或显示 Lin 正在招募世界级研究者的团队扩张。任何一项都可能把立场推向投资。会把立场推向放弃的论点击穿触发器包括:确认产品延迟到 2027 年 Q4 之后,Lin Junyang 或两名及以上关键联合创始人离开,低于 $1.5 billion 的 down round,失去 HongShan 或 Gaorong 支持,或地缘政治事件限制跨境 AI 研究合作。最终尽调事项包括完整股权结构表披露、技术架构文件、算力资源获取计划,以及带里程碑的研究路线图。[CV025, CV026, CV027, CV028, CV029, CV030]

投资逻辑破裂与止损触发因素表
触发因素条件 / 阈值优先级建议动作
创始人离职Lin Junyang 宣布退出或显著降低角色关键立即下调至不投;投资逻辑完全依赖创始人
降估值轮下一轮融资投后估值低于 $1.5B关键意味着投资人信心坍塌;评估二级退出或减记
产品延迟到 2027 年 Q4 仍无可展示的具身智能模型输出转入观察并 90 天后重评;标记执行风险正在兑现
不利监管行动MIIT 或 CAC 执法针对 Prague 或其训练数据做法升级至法律审查;可能伤及核心 IP 和运营
投资人退出Gaorong 或 HongShan 在下一轮拒绝按比例跟投,或以显著折价出售老股说明内部人信念减弱;全面复核投资逻辑

止损触发因素是前瞻性指标,不是历史观察。截至 2026 年 7 月,上述触发因素均未启动。

[CV029, CV030, CV031, CV032]
最终尽调问题表
尽调要求优先级理由目标信息
完整股权结构表关键缺少股权结构表,入场稀释和优先权结构无法核实;按 $2B 融资 $220M 暗示约 11%,但清算优先权可能显著改变经济收益已发行股份、期权池、各类别投资人持股、SAFE / 票据转换条款
技术架构文件关键核心 IP 主张建立在新型世界模型架构上;没有技术细节,就无法判断其相对 Qwen3.5 或竞争模型的差异化模型架构白皮书、训练数据来源计划、推理栈
算力获取与采购计划GPU 出口管制是关键路径风险;公司获取 Huawei Ascend 或国产替代方案的策略必须验证芯片采购协议、Huawei 合作状态、算力预算相对路线图
研究路线图与里程碑乐观情景假设 2027 年 Q2 前交付产品;内部路线图应定义具体能力里程碑、发布日历和招聘计划18 个月研究和产品路线图,附交付关口
团队名单与股权计划技术团队质量至关重要;除 Lin Junyang 外,当前团队未公开披露;归属安排影响留任风险团队组织架构图、关键研究员 LinkedIn 资料、归属和留任条款

尽调问题假设潜在机构投资人正在开展尽职调查;列表按优先级排序。

[CV033, CV034, CV035, CV036]

8.5 图表

免责声明

本报告仅基于截至 2026-07-23 的公开信息。Prague Technology 尚未披露财务报表、资本结构表或技术架构文件。可比公司估值基于公开来源报道或估算数据。本报告不构成投资建议。

证据索引

结论
编号陈述可信度来源
CO001 The primary legal entity for Prague Technology is 上海卜拉格科技有限公司 (Shanghai Bulage Technology Co., Ltd.), registered on May 27, 2026, in Xuhui District, Shanghai. SO001, SO009
CO002 A sister entity, 语用(上海)科技有限公司 (Yuyong Shanghai Technology Co., Ltd.), was registered on May 13, 2026, and is 100% owned by Lin Junyang. SO001, SO005
CO003 A third entity, Shanghai Gewuzhiyong Management Consulting Partnership, is controlled through Bulage Technology as GP with Lin Junyang as direct 1% investor. SO001, SO003
CO004 The name 卜拉格 (Bulage) is a phonetic transliteration of 'Pragmatics', while 语用 (Yuyong) is the semantic Chinese translation—both referencing the linguistic field of pragmatics. SO005, SO011
CO005 Prague Technology does not have an official public website, publicly disclosed product, or announced commercial activities as of July 2026. SO003, SO009
CO006 Lin Junyang has confirmed the company's focus areas as world models and the 'embodied brain' for physical AI systems. SO002, SO012
CO007 The company is headquartered in Xuhui District, Shanghai, China, consistent with corporate registration records. SO009, SO001
CO008 Lin Junyang left Alibaba on March 3–4, 2026, publicly announcing his departure with the post 'me stepping down. bye my beloved qwen' on the X platform. SO008, SO002, SO003
CO009 Lin Junyang was born in 1993, making him approximately 32–33 years old at the time of founding Prague Technology. SO003, SO005
CO010 Lin Junyang holds a bachelor's degree in English from the University of International Relations and a master's degree in linguistics and NLP from Peking University's School of Foreign Languages. SO003, SO005, SO002
CO011 Lin Junyang joined Alibaba's DAMO Academy in 2019 as a fresh graduate, working on NLP in search and recommendation scenarios and early multimodal projects including M6. SO003, SO005
CO012 Lin Junyang became technical head of Alibaba's Tongyi Qianwen (Qwen) LLM series at the end of 2022 when Alibaba restructured its AI teams into the Tongyi Lab system. SO003, SO005, SO008
CO013 Lin Junyang became the youngest P10-level technical executive in Alibaba's history, representing the highest individual technical contributor rank in the company's evaluation system. SO003, SO005
CO014 On March 3, 2026, Lin Junyang was informed of an Alibaba restructuring plan to split the Qwen team into separate horizontal teams, removing its organizational independence. SO002, SO008
CO015 The Qwen series of models, under Lin Junyang's leadership, accumulated over 1 billion downloads globally with more than 200,000 derivative models by early 2026. SO005, SO003
CO016 Prague Technology's angel round was first reported by The Information in May 2026 at an approximately $2 billion target valuation; the round closed in June 2026. SO001, SO003, SO007
CO017 Prague Technology's post-money valuation is approximately $2 billion (RMB 13.5 billion) after its angel round, setting a record for pre-revenue Chinese AI startup valuations at the seed/angel stage. SO001, SO003, SO007, SO011
CO018 Gaorong Ventures led the angel round with $100 million (approximately RMB 675 million). SO001, SO003, SO007
CO019 HongShan Capital (formerly Sequoia China) co-led the angel round with $100 million. SO001, SO003, SO007
CO020 Tencent participated in the angel round as a strategic co-investor with $20 million (approximately RMB 135 million). SO001, SO003, SO007
CO021 Total capital raised in the angel round is approximately $220 million from Gaorong ($100M), HongShan ($100M), and Tencent ($20M). SO001, SO007, SO011
CO022 Gaorong Ventures manages over RMB 30 billion across a portfolio of 300+ companies, with active investments in Chinese AI, robotics, and embodied intelligence. SO016, SO017
CO023 HongShan Capital manages over $55 billion AUM across 1,500+ portfolio companies and has backed OpenAI, Anthropic, DeepSeek, and Chinese embodied intelligence startups. SO020, SO018
CO024 At the time of the angel round close in June 2026, Prague Technology had no products, no revenue, and had not publicly disclosed an official company name. SO003, SO007
CO025 Immediately after the angel round closed, Lin Junyang's startup was reported to be seeking a new follow-on financing round. SO001, SO003
CO026 In October 2025, while still at Alibaba, Lin Junyang formed a small internal team focused on robotics and embodied intelligence within the Qwen organization. SO002, SO005
CO027 Lin Junyang published 'From Reasoning Thinking to Agentic Thinking' on March 26, 2026, arguing that the next AI paradigm centers on agentic systems that act rather than merely think. SO005, SO012
CO028 The founding team of Prague Technology is initially described as 1–10 people with backgrounds from ByteDance, Tencent, and overseas organizations. SO009, SO002
CO029 No other named founders, co-founders, board members, or senior executives beyond Lin Junyang have been publicly disclosed as of July 2026. SO003, SO009
CO030 Forbes and financial analysts have characterized Prague Technology's $2B valuation as primarily based on founder pedigree rather than product-market fit, representing high investment risk. SO023
CO031 Prague Technology has published no technical papers, patents, or product demonstrations under its new entity names as of July 2026. SO005, SO009
CO032 Lin Junyang's departure from Alibaba in March 2026 was described in press coverage as following a meeting in which Alibaba Cloud CTO Zhou Jingren conveyed a restructuring plan for the Qwen team. SO002, SO008
CO033 Tencent has invested in virtually every top-tier Chinese AI startup including DeepSeek, MiniMax, Zhipu, and now Prague Technology, reflecting a broad portfolio strategy. SO001, SO011
CO034 One investor commentary from a multi-year AI sector follower noted that 'the gap between current AI company valuations and their fundamentals is widening' in the context of Prague Technology's funding. SO005, SO023
CO035 The angel round's implied pre-money valuation—approximately $1.78 billion if total dilution was around 11%—places Prague Technology among the most valuable pre-product AI entities ever created in China. SO003, SO007
CO036 Gaorong Ventures' investment in Prague Technology has been widely confirmed across multiple Chinese and English-language technology news outlets including 36Kr, CNTechPost, and Techcrunch coverage of the angel round. SO006, SO007, SO004
CM001 Embodied AI is defined as AI systems that perceive, reason, and act in physical environments through hardware carriers; it is distinct from pure software LLMs and from traditional pre-programmed industrial robots with fixed logic. SM006, SM009
CM002 The included spend boundary for embodied AI comprises AI brain/world model software, complete robot systems, upstream components, system integration services, and simulation/data infrastructure. SM002, SM008
CM003 Excluded from this market definition are pure-software cloud LLM services, conventional industrial robots operating on fixed programs, and general enterprise IT spend not specific to embodied agents. SM009
CM004 Within China's embodied AI market, the AI brain and world model software layer is identified as the critical supply-chain bottleneck with only approximately 100 companies globally active in developing embodied large models, versus 8,000+ upstream hardware component suppliers. SM002, SM004
CM005 Chinese companies accounted for approximately 74–87% of global humanoid robot unit shipments in 2025, with the higher figure citing Omdia data and the lower range from China Daily citing the IDC-backed Shanghai expo report. SM005, SM014
CM006 The embodied AI market is transitioning from hardware-centric to AI-brain-centric architecture, with foundation model-driven robot planning recognized as the breakthrough of 2025–2026 enabling multi-task generalization. SM009, SM008
CM007 Prague Technology's stated focus on world models and embodied intelligence positions it in the AI brain software layer, not in hardware manufacturing or system integration, per founder statements and news coverage. SM019, SM020
CM008 The global embodied AI market was estimated at $4.44 billion in 2025 growing at a 39% CAGR, implying approximately $23 billion by 2030 on a bottom-up basis, according to ainchina.com market composite data. SM001
CM009 IDC forecasts the global embodied AI market — defined broadly to include all physical AI adjacent categories — will reach $1.5 trillion by 2030, as cited in the Shanghai International Embodied Intelligence Expo 2026 report. SM005
CM010 China's State Council Development Research Center projects the domestic embodied AI market will exceed ¥1 trillion ($140 billion) by 2035, representing a government-linked policy target rather than an independent market research projection. SM001, SM006
CM011 China's robotics and embodied AI startups raised $3.3 billion across 126 deals in Q1 2026, the largest single quarter on record, according to Crunchbase data cited by ainchina.com. SM001, SM005
CM012 China captured $16.5 billion out of Asia's $27.4 billion total venture capital in Q1 2026 (60%), with robotics as the single largest sector contributor within China's share. SM001
CM013 Global humanoid robot shipments in 2025 reached approximately 18,000 units, a year-on-year increase of approximately 508% compared to roughly 3,000 units in 2024, per IDC data cited at the Shanghai embodied intelligence expo. SM005
CM014 Total H1 2026 financing in China's embodied intelligence and robotics sector exceeded ¥46 billion ($6.4 billion) across 288 events involving 226 companies, per Qixinbao data cited by EmbodiedGlobal. SM003
CM015 Global robotics venture funding in 2025 totaled $13.8 billion, with humanoid-specific investment growing approximately 143x over four years, according to humanoid.guide's 2026 foundation model report. SM007
CM016 Sixteen Chinese companies in the embodied AI and robotics space are valued over ¥10 billion ($1.4 billion) as of mid-2026, including established names Galaxy General Robotics, Unitree, AgiBot, and Mecharmind alongside newer entrants. SM004
CM017 Unitree Robotics achieved 2025 revenue of ¥1.699 billion (~$236 million) with a 60.13% gross margin and has been profitable since 2020, representing one of the few profitable hardware players in China's humanoid robot market. SM001, SM017
CM018 The number of existing enterprises related to embodied AI in China exceeded 10,000 as of May 2026, per EmbodiedGlobal/Qixinbao data; separately, the State Taxation Administration cited 3,025 embodied AI enterprises as of May 2026 on a narrower definition. SM002, SM006
CM019 China's embodied AI enterprise sales grew 22.4% year-on-year in January–May 2026, accelerating from 13.9% for full-year 2025, per State Taxation Administration data published in Qiushi Theory. SM006, SM005
CM020 The Yangtze River Delta region — Shanghai, Jiangsu, and Zhejiang — accounts for more than 50% of China's embodied AI companies and financing scale, per the Shanghai International Embodied Intelligence Expo 2026 report. SM005
CM021 Industrial robots hold the largest revenue share (~45%) of China's embodied AI market, followed by service robots (~25%), special robots (~15%), and humanoid robots (~5%, fastest growing), per the CIEI 2026 Shanghai expo report. SM005
CM022 Guangdong province generated 78.7% of China's total embodied AI sector sales revenue in January–May 2026, demonstrating extreme geographic concentration driven by its dominance in electronics and manufacturing supply chains. SM006, SM005
CM023 Primary deployment targets for humanoid robots among Chinese automotive OEMs include BYD (committed to 20,000 humanoid units in 2026), NIO, and SAIC Motor, with BYD using AgiBot as its primary humanoid supplier. SM013
CM024 China's embodied AI supply chain has 8,000+ upstream component companies (servo motors, sensors, batteries), ~1,900 midstream manufacturers and integrators, and ~2,370 downstream application companies including 680 focused on humanoid robots. SM002
CM025 The embodied AI supply chain's critical bottleneck includes only ~100 companies each in visual sensors, robot joint modules, and embodied large models — the 'see, act, and think' core links — according to Qixinbao's panorama report. SM002
CM026 China's 15th Five-Year Plan (2026–2030) explicitly designates embodied AI as one of six major future industries alongside quantum technology and biomanufacturing, per Qiushi Theory official government publication. SM006, SM005
CM027 Prague Technology's prospective customer base is led by humanoid OEMs (AgiBot, Unitree, UBTECH) requiring AI brain models, and research institutions as early adopters, given that Prague has no existing customer relationships. SM019, SM020
CM028 The 2026 joint MIIT-SASAC action plan targets 10,000+ robot deployments by year-end, creating near-term industrial demand for AI brain model capabilities that companies like Prague Technology are developing. SM002, SM006
CM029 China's advanced LLM foundation — Alibaba Qwen3.5, DeepSeek V4, ByteDance Seed 1.6 — provides the software base for embodied AI model development; Lin Junyang's Qwen architecture expertise is directly applicable to world model construction. SM001, SM019
CM030 China's manufacturing supply chain covers approximately 70% of global industrial robot components including motors, sensors, and batteries, providing cost and proximity advantages for domestic embodied AI hardware development. SM007, SM001
CM031 The sim-to-real transfer gap — historically the biggest technical blocker for deploying AI-trained robots in physical environments — is closing through domain randomization at scale, neural radiance fields for photorealistic training, and digital twin pipelines. SM008, SM009
CM032 Data scarcity is the primary technical bottleneck for training embodied AI models: unlike LLMs trained on internet-scale text, embodied models require large-scale physical interaction data that is expensive and slow to collect. SM010, SM009
CM033 ByteDance declared world models — the core technology behind Prague Technology's stated product — its top AI priority for 2026, allocating a ¥200 billion ($29.4 billion) capital expenditure budget for AI infrastructure. SM001
CM034 US export controls restricting NVIDIA H100 and H200 GPU shipments to China create sustained AI compute procurement risk for all Chinese AI developers training large embodied models, including Prague Technology. SM024
CM035 Capital intensity for training world models at scale is extreme — the compute budget for a world model comparable to top-tier robotics foundation models is estimated in the hundreds of millions of dollars. SM007, SM010
CM036 China's embodied AI sector shows early signs of market oversupply and speculative excess with 10,000+ companies in a nascent market, 15 unicorns created in six months, and multiple analysts noting a 'reality check' on whether valuations are justified by real technical capability. SM016, SM004
CM037 The 65x spread between the conservative $23B and aggressive IDC $1.5T 2030 global embodied AI market estimates reflects definitional disagreement — not forecasting error — with the IDC figure including autonomous vehicles and manufacturing IT that most practitioners exclude. SM005, SM001
CM038 Prague Technology has no disclosed revenue, customers, products, or commercial deployments as of July 2026; its participation in the embodied AI market is entirely prospective and pre-commercial. SM019, SM020
CM039 The base-case timeline for reliable, unattended general-purpose humanoid robot autonomy is 2028–2030 according to humanoid.guide's independent 100-page foundation model report, suggesting Prague Technology's commercial revenue timeline extends at least 2+ years even under optimistic assumptions. SM007
CM040 Prague Technology's world model business model depends on OEM platform adoption — if leading humanoid OEMs (AgiBot, Unitree) develop proprietary in-house world models, Prague's addressable market compresses to smaller OEMs and research institutions. SM013, SM007
CP001 Prague Technology's primary direct Chinese competitors in the world model / AI brain layer are TARS AI (AWE 3.0 model, $697M raised) and Spirit AI ($435M raised, $1.5B valuation). SP001, SP015
CP002 Prague Technology's global world model competitors include AMI Labs ($1.03B at $3.5B), World Labs ($1B+ at $5B target), and Physical Intelligence (π0 model); none of these are China-focused. SP005, SP009
CP003 Prague Technology at $220 million raised is substantially less capitalized than its direct peers: TARS AI ($697M), AMI Labs ($1.03B), and World Labs ($1B+), giving competitors meaningfully larger compute and talent budgets. SP001, SP005, SP009
CP004 OEM competitors AgiBot and Unitree are investing in in-house AI brain models, reducing their dependence on external world model providers and shrinking Prague's potential customer base among leading OEMs. SP010, SP012
CP005 Open-source substitutes — Unitree's UnifoLM-VLA-0, Google DeepMind's RT-X Open X-Embodiment — set a zero-cost baseline for AI brain capabilities that Prague's world model must demonstrably outperform to justify commercial licensing. SP012, SP019
CP006 Likely future entrants into the China world model space include any major Chinese internet company (Baidu, Xiaomi, NetEase) that pivots compute budgets toward embodied AI, as well as global AI labs (Anthropic, xAI) extending into physical AI. SP017, SP023
CP007 TARS AI has raised a total of approximately $697 million across two rounds: $242 million angel in Q2 2025 (previously China's largest embodied AI angel round) and $455 million Pre-A in April 2026 (China's largest single-round financing in embodied AI). SP001, SP002
CP008 TARS AI's AWE 3.0 is a world-capable general embodied large model that TARS claims has been used in real commercial deployments, including the precision assembly of complex flexible wire harnesses — described as the industry's 'Goldbach Conjecture' — setting a Guinness World Record. SP001, SP002
CP009 TARS AI was founded in February 2025 by Chen Yilun (former CTO of Autonomous Driving at Huawei) as CEO and Li Zhenyu (former president of Baidu Apollo autonomous driving) as Chairman, giving it deep operational AI deployment expertise rather than academic LLM background. SP002, SP004
CP010 TARS AI's full-stack strategy — combining proprietary hardware (A-series wheeled and T-series bipedal robots) with the AWE 3.0 AI brain model — enables TARS to collect proprietary embodied training data from real-world deployments, a critical competitive advantage over pure-software competitors like Prague. SP001, SP003
CP011 TARS AI's Pre-A round was co-led by Hillhouse Ventures and HongShan (Sequoia China) — the same HongShan that invested $100M in Prague Technology — representing a potential investor conflict and a signal that HongShan is hedging across full-stack and pure-software world model approaches. SP001, SP003
CP012 TARS AI's robot made its overseas debut at LogiMAT 2026, securing purchase intentions from clients across multiple European countries, suggesting early international market traction that Prague Technology has not yet initiated. SP002
CP013 TARS AI received investments from Beijing Robotics Industry Development Investment Fund and Shanghai State-owned Capital Investment Guide jointly, representing China's first joint state capital investment in an embodied intelligence company — providing TARS preferential access to government procurement and policy support channels that pure-VC companies like Prague lack. SP001, SP004
CP014 TARS AI has co-founder and Chief Scientist Ding Wenchao, one of Huawei's first 'Genius Youth' recruits, who developed Fudan University's first humanoid robot, providing deep hardware-AI integration expertise that is more relevant to deployment than Prague's LLM architecture background alone. SP002
CP015 AMI Labs was founded in late 2025 by Yann LeCun (Turing Award winner, former Meta FAIR head) and raised $1.03 billion at a $3.5 billion pre-money valuation in March 2026 — Europe's largest-ever seed round — with investors including NVIDIA, Samsung, Toyota Ventures, Temasek, Jeff Bezos, Mark Cuban, and Eric Schmidt. SP005, SP007
CP016 AMI Labs builds world models using LeCun's JEPA (Joint Embedding Predictive Architecture), which predicts in abstract representation space rather than token-by-token — theoretically more robust for physical systems where exact world state prediction is unnecessary. SP005, SP006
CP017 AMI Labs CEO Alexandre LeBrun explicitly stated the company has no plans to generate revenue in the near term and that the commercial timeline for world models may be 'measured in years rather than quarters' — validating the patient capital thesis that Prague Technology implicitly relies on. SP005, SP006
CP018 World Labs (Fei-Fei Li) raised $230 million at a $1 billion valuation in 2024 and secured an additional $200 million from Autodesk in February 2026, with reported $5 billion valuation target; its first product 'Marble' focuses on 3D environment generation for entertainment and design, not industrial robotics. SP009
CP019 World Labs represents an adjacent rather than direct competitive threat to Prague Technology — its world model application is 3D spatial intelligence for creative/enterprise design workflows, not robot control for manufacturing. SP009
CP020 Spirit AI (精灵智能) has raised $435 million in total across its Series A and earlier rounds at a $1.5 billion valuation from Chaos Ventures and YF Capital, explicitly targeting the universal robot brain market — the same product category as Prague Technology. SP015, SP016
CP021 AgiBot (BYD-backed) hosted the AgiBot World Challenge 2026 alongside ICRA in Vienna, with 526 research and enterprise teams from 27 countries competing across 'Reasoning to Action' and 'World Model' tracks, establishing AgiBot as a world model ecosystem builder rather than just a hardware vendor. SP010, SP011
CP022 AgiBot's world model strategy leverages real manufacturing deployment data from BYD and SAIC Motor to train its in-house AI models, creating a proprietary data flywheel that pure-software competitors like Prague cannot replicate without OEM partnerships. SP011, SP012
CP023 Unitree Robotics has open-sourced UnifoLM-VLA-0, its vision-language-action model, and maintains an open SDK ecosystem — a strategic choice to commoditize the AI brain layer and drive hardware adoption at the cost of standalone AI model pricing power. SP012, SP013
CP024 Unitree's hardware pricing ranges from $4,900 (R1 AIR) to $43,900 (G1 EDU), with the G1 IPO (approved June 2026) targeting approximately $610 million at Shanghai STAR Market — this range sets an implicit ceiling on how much OEMs will pay for AI brain software layered on affordable hardware. SP012, SP013
CP025 Unitree has achieved 5,500+ humanoid robot unit shipments in 2025, providing it with orders of magnitude more real-world embodied interaction data than any pure-software startup including Prague Technology. SP013
CP026 AgiBot's manufacturing deployment data (BYD, SAIC, 5,168 units shipped in 2025) and Unitree's broad developer ecosystem (open SDK, global research deployments) collectively give the two leading Chinese OEMs richer embodied training data than any software-only company has access to as of 2026. SP011, SP013
CP027 ByteDance declared world models its top AI priority for 2026 and allocated a ¥200 billion ($29.4 billion) capital expenditure budget for AI infrastructure — representing an incumbent resource base that dwarfs all embodied AI startups' combined funding. SP017
CP028 Google DeepMind's RT-X (Open X-Embodiment) ecosystem, trained on data from 22 different robot types, demonstrates cross-embodiment transfer and represents an open-source baseline that all commercial world model providers must outperform to justify their pricing. SP019, SP020
CP029 ByteDance's Doubao platform — with 200+ million daily active users and multimodal interaction data — gives it a training data advantage for world models at a scale no embodied AI startup can match; ByteDance has not yet focused this on robotics-specific world models but could pivot without warning. SP017
CP030 Alibaba's Qwen team remains active post-Lin Junyang departure: Qwen3.5 was released in March 2026, and the Qwen open-source ecosystem has been adopted by multiple robot manufacturers as their AI base — making Alibaba a potential incumbent competitive threat if it builds robotics-specific extensions to the Qwen architecture. SP021, SP022
CP031 World model architectures empirically outperform VLA baselines in generalization benchmarks: DreamZero achieved 62.2% task success versus 27.4% for VLA baselines per humanoid.guide, supporting Prague's architectural bet but also confirming the competition is well-aware of this advantage. SP024
CP032 Prague Technology's architecture expertise moat — Lin Junyang's transformer and MoE scaling knowledge from Qwen3.5's 397B-parameter model — is the company's strongest current advantage, but it is a 12–24 month window before the field converges on standard architectures or open benchmarks commoditize the advantage. SP022, SP021
CP033 Prague Technology's most critical structural competitive disadvantage relative to TARS AI and AgiBot is the absence of proprietary embodied training data: as of July 2026, Prague has no deployed robots generating real-world interaction data. SP001, SP010
CP034 TARS AI's first-mover advantage in China's embodied AI brain market — having shipped AWE 3.0 and deployed robots commercially — means Prague Technology enters the market at least 12 months behind the leading Chinese full-stack competitor. SP001, SP008
CP035 The multi-homing risk for OEM customers is moderate: an OEM could use both Prague's world model and a competitor's model for different tasks (multi-homing), preventing any single provider from establishing exclusive lock-in without significantly superior performance or cost advantages. SP024
CP036 Prague Technology has not yet released any model, benchmark, or technical paper publicly as of July 2026, making it impossible to assess its actual technical progress relative to TARS AI's AWE 3.0 or AMI Labs' JEPA framework. SP015, SP023
CP037 The investor overlap between Prague Technology (HongShan invested $100M) and TARS AI (HongShan co-led $455M Pre-A) suggests HongShan views these as complementary bets rather than mutually exclusive, implying that even key Prague investors do not expect a winner-take-all world model outcome in China. SP001, SP003
CP038 The commoditization risk for world models is real: if open-source world models (OpenVLA, RT-X derivatives) reach adequate performance for 80% of factory tasks by 2027-2028, Prague's SOM shrinks to specialist high-precision applications where commercial models provide marginal benefit over open alternatives. SP020, SP025
CP039 Prague Technology's pure-software strategy avoids hardware distraction and capital allocation trade-offs but requires successful OEM partnerships for data access — a dependency that gives OEM partners negotiating leverage over model pricing and terms. SP012, SP015
CP040 The adverse evidence that world model startups struggle to convert funding into commercial products — explicitly acknowledged by AMI Labs CEO (years, not months to commercial application) — validates diligence skepticism about Prague Technology's revenue timeline and commercial readiness. SP006
CI001 Prague Technology raised approximately $220 million in an angel round: $100 million from Gaorong Ventures, $100 million from HongShan Capital (Sequoia China), and $20 million from Tencent, closing approximately June 2026. SI006, SI004, SI005
CI002 Prague Technology's post-money valuation is approximately $2 billion, based on the $220 million raise and reported valuation target from multiple news sources; the company has not officially confirmed the exact post-money figure. SI017, SI018
CI003 Implied pre-money valuation of approximately $1.78 billion (assuming ~11% dilution at $220M / $2B) places Prague among China's most highly valued pre-product AI startups. SI017, SI006
CI004 Multiple sources reported that the Prague Technology angel round was oversubscribed, suggesting strong investor demand relative to the $220M allocation and founder credibility premium. SI017, SI019
CI005 Multiple news sources reported that Prague Technology was actively seeking follow-on financing immediately after the angel round closed in June 2026, suggesting the company anticipated needing substantially more capital than $220 million for its full development program. SI017, SI006
CI006 Prague Technology's capital deployment is expected to be concentrated in AI compute (GPU training runs), talent acquisition, and operational infrastructure — the standard three cost categories for a frontier AI model development company. SI007, SI009
CI007 AMI Labs' $1.03 billion raise at $3.5 billion valuation with no product and years-to-revenue timeline sets the global precedent for patient-capital world model investment that Prague Technology's investors are implicitly replicating. SI008, SI016
CI008 TARS AI's $697 million total raise (at comparable valuation level to Prague) but with shipped products and commercial deployments demonstrates what $3x capital can buy in terms of commercial validation — a benchmark that highlights Prague's undercapitalization relative to full-stack peers. SI020
CI009 Prague Technology has zero revenue as of July 2026. No commercial product has been released, no customers have been announced, and no ARR, GMV, or other financial metrics have been disclosed. SI017, SI018
CI010 Prague Technology's intended business model is B2B AI model licensing: OEMs would license its world model to power their robots' AI brain, paying recurring API fees or annual license fees. SI017, SI019
CI011 No pricing has been disclosed for any Prague Technology products; pricing models, contract structures, and early customer terms are all confidential or non-existent as of July 2026. SI017, SI018
CI012 Prague Technology has no commercial customers as of July 2026. No customer names, letters of intent, pilot agreements, or purchase orders have been publicly disclosed. SI017, SI006
CI013 Prague Technology has zero ARR (annual recurring revenue), zero GMV, and no disclosed unit volume metrics; the company is entirely pre-revenue and pre-commercial. SI018, SI017
CI014 The earliest plausible date for Prague Technology's first commercial revenue is 2028, based on the 2028-2030 base-case timeline for reliable general-purpose humanoid autonomy and the need for 12-18 months of OEM product integration after model release. SI009
CI015 The most likely primary revenue model for Prague Technology is API-as-a-service or annual OEM license for its world model, analogous to frontier LLM API providers but specialized for embodied AI applications. SI007, SI008
CI016 Custom model training and fine-tuning services for OEM-specific robot configurations represent a likely secondary revenue stream for Prague Technology once the core world model exists. SI007
CI017 AI compute (GPU training runs) is the primary cost driver for world model development; comparable frontier LLM training costs range from $20 million to over $200 million per training run for models at the GPT-4 / Qwen3.5 scale. SI015, SI009
CI018 Frontier AI researcher talent acquisition is the second major cost driver; compensation packages at leading Chinese and global AI labs range from ¥3 million to ¥20 million ($420K to $2.8M) annually for senior researchers, and Lin Junyang's reputation will attract but also accelerate high-compensation hiring. SI015
CI019 World model training may require substantially higher compute per training run than equivalent-scale LLM training due to multi-modal physical simulation requirements; this increases capital intensity compared to purely language-based model development. SI009
CI020 Prague Technology's pure-software strategy (no hardware manufacturing) theoretically eliminates hardware COGS, working capital for inventory, and manufacturing overhead — creating a path to 70-85% gross margins at commercial scale analogous to foundation model API providers. SI007, SI015
CI021 Gross margin for an AI world model licensing business is theoretically 70-85% at scale, based on the analogy to LLM API providers (OpenAI, Anthropic), but embodied AI inference may require higher compute per call due to multi-modal processing — potentially compressing margins to 40-60%. SI007, SI009
CI022 Working capital requirements for Prague Technology's pre-revenue phase are minimal: no accounts receivable, no inventory, no customer prepayments to fund. Cash management is primarily outflow (burn) rather than working capital cycling. SI007
CI023 Prague Technology's primary capital expenditure is GPU compute access — either through cloud providers (Tencent Cloud, Alibaba Cloud) or owned GPU clusters; the compute infrastructure choice is a key strategic decision affecting both cost and IP control. SI019
CI024 Forbes documented in June 2026 that AI startups with no revenue are using 'this tactic' — founder credibility plus future market size narratives — to 'supersize their valuations', and identified this pattern as a systemic risk in the AI investment ecosystem. SI001, SI002
CI025 CNBC reported in November 2025 that 'AI valuation fears grip investors as tech bubble concerns heighten', reflecting broad institutional awareness of AI valuation excess that predates and contextualizes Prague's June 2026 raise. SI003
CI026 S&P Global flagged hidden AI investment risks including concentration in illiquid, pre-revenue VC positions and speculative pre-commercial bets — a risk category that Prague Technology exemplifies. SI002
CI027 Prague Technology's $2B post-money valuation implies investors expect the company to eventually generate revenue in the hundreds of millions annually — a scale that requires penetrating the OEM market with 50-100 enterprise customers at premium pricing, a timeline that extends well beyond 2030 in most scenarios. SI001, SI015
CI028 Prague Technology's monthly cash burn rate is entirely unknown from public sources; no proxy metrics (employee count, office space, cloud compute invoices) have been disclosed; burn rate is the critical unknown for runway assessment. SI017, SI018
CI029 US imposition of approximately 25% tariffs on NVIDIA H200 AI chips bound for China materially increases the cost of GPU compute for all Chinese AI model developers including Prague Technology, forcing reliance on domestic alternatives with inferior price-performance ratios. SI010, SI011
CI030 Adverse macroeconomic or AI sector valuation conditions between 2026 and Prague's next financing event (likely 2027-2028) could force a down-round or impaired follow-on round if the world model has not shipped commercial versions by then. SI003, SI002
CI031 Prague Technology's price-to-capital ratio (post-money valuation / total capital raised) is approximately 9x ($2B / $220M), higher than AMI Labs (3.4x = $3.5B / $1.03B) and substantially above the historical software startup norm of 3-5x at equivalent stage. SI016, SI024
CI032 AMI Labs set the comparable reference point: $3.5B post-money on $1.03B raised (3.4x P/Capital), no product, years to revenue — and its investor syndicate includes NVIDIA, Samsung, and Toyota Ventures. Prague at 9x P/Capital on $220M commands a higher premium with a narrower investor base. SI016, SI008
CI033 TARS AI at approximately $2-2.5B valuation but with $697M raised, shipped products, commercial deployments, and manufacturing customer relationships represents a financially better-validated company at the same implied valuation tier as Prague Technology. SI020
CI034 Prague Technology at $220 million raised is the smallest of the major world model startups: AMI Labs $1.03B, World Labs $1B+, Spirit AI $435M, TARS AI $697M. This represents both lower dilution risk and higher capital constraint risk. SI016, SI024
CI035 The financial diligence blockers that must be resolved before any underwriting decision include: verified cash balance and monthly burn, capital allocation plan (compute vs. talent vs. other), and any LOIs or partnership MOUs with OEM customers. SI017, SI007
CI036 Monthly cash burn rate is the single most critical unknown for evaluating Prague Technology's financial position: the difference between $5M/month and $12M/month burn determines whether the company has 44 months or 18 months of runway from the June 2026 close. SI017, SI009
CI037 No public indication exists that Prague Technology has applied for or received any Chinese government grants, AI development subsidies, or non-dilutive funding from MIIT, national AI funds, or state compute allocation programs. SI013, SI014
CI038 Tencent's $20M investment likely came with access to Tencent Cloud compute infrastructure as a strategic benefit; if Prague can access Huawei Ascend or domestic GPU capacity at cost through Tencent's cloud, the effective value of the relationship could exceed the face value of the investment. SI019
CI039 The gross margin structure for world model licensing is theoretically superior to hardware-integrated competitors but has not been empirically tested at commercial scale in the embodied AI market; actual gross margins may be substantially lower if inference costs per API call are higher than LLM equivalents. SI007, SI009
CI040 The financial verdict for Prague Technology is: extreme pre-revenue speculation at a premium valuation, justified only by founder pedigree and world model thesis timing; commercial underwriting is not possible without burn rate data, a product roadmap, and at least one signed OEM customer agreement. SI001, SI017, SI009
CE001 Prague Technology has not released any product, published any model, or demonstrated any prototype as of July 2026; the company is entirely pre-product as of its first three months of operation. SE013, SE024, SE025
CE002 Prague Technology's intended product is a world model for embodied AI — an AI brain layer that enables physical robots to perceive, predict, and plan interactions with the physical world at a cognitive level above direct sensorimotor imitation. SE013, SE014
CE003 The world model approach differs from VLA (Vision-Language-Action) models in that world models learn an internal simulation of physical dynamics, enabling plan-before-act reasoning rather than reactive imitation of demonstrations — directly relevant to multi-step dexterous manipulation. SE021, SE007
CE004 Prague Technology's product definition in customer workflow terms: an OEM robot manufacturer integrates Prague's world model via SDK, replacing proprietary planning modules with a foundation model that generalizes across manipulation tasks without task-specific re-training. SE013, SE020
CE005 Prague Technology's value proposition to OEMs is reduced per-task integration cost and improved generalization — analogous to how LLM APIs replaced custom NLP models — but the analogy requires OEM adoption of a fully trained and validated world model, which does not exist yet. SE013, SE008
CE006 The Qwen3.5 model family, which Lin Junyang led at Alibaba, employed a Mixture-of-Experts (MoE) architecture with 397 billion total parameters and conditional computation — directly applicable to the computational requirements of world model inference for embodied AI. SE001, SE002
CE007 Qwen3 introduced hybrid thinking modes that combine fast reactive responses with slow deliberate multi-step reasoning — a System 1/System 2 architecture directly analogous to the reactive-vs-planning duality needed for robot control in embodied AI. SE002, SE001
CE008 Qwen2.5 and Qwen3 include advanced multimodal encoders supporting visual, audio, and cross-modal reasoning — the same encoder architecture required for Prague Technology's intended world model to parse robot sensor inputs (cameras, IMU, force sensors). SE003, SE001
CE009 Lin Junyang's architectural expertise is unusually directly applicable to world model development: MoE backbone, multimodal encoders, hybrid thinking modes, and large-scale distributed training are the exact components required for a world model for embodied AI. SE014, SE002
CE010 Physical Intelligence's π0 model (3B parameters, trained on 68+ manipulation tasks) demonstrated that a VLA foundation model can achieve zero-shot task generalization across diverse manipulation environments — establishing the benchmark Prague must exceed with a world model approach. SE005, SE007
CE011 DeepMind's Genie 2, a 10B-parameter world model trained on video, demonstrated the ability to generate consistent 3D interactive environments from a single image — providing proof-of-concept for world models in physical simulation contexts that validate Prague's technical thesis. SE006, SE007
CE012 Prague Technology's primary technical differentiation vs. TARS AI (AWE 3.0) is the intended cognitive depth: AWE 3.0 is a production-ready hardware-integrated VLA+WM system for known OEM environments, while Prague targets a general-purpose world model foundation layer for arbitrary OEM environments — higher ambition but longer timeline. SE022, SE023
CE013 AMI Labs is the closest global comparable to Prague Technology's approach: a pure world model foundation company, no hardware, Europe-based, targeting multi-year commercial timeline, backed by $1.03B with no product yet — the key difference is AMI has published research and model cards while Prague has published nothing. SE021, SE008
CE014 World model AI for embodied applications requires real-time inference at robot control frequencies of 30-100Hz for reactive control, while deliberative planning may tolerate 1-10Hz — both modes must be achievable on available on-device compute, which is a significant unsolved engineering challenge. SE021, SE005
CE015 Prague Technology has no published patents, technical papers, preprints on arXiv, or architecture documentation as of July 2026 — making independent technical due diligence of its architectural approach impossible at this stage. SE024, SE025
CE016 Prague Technology's most critical external dependency is GPU compute access; US export controls restrict access to NVIDIA H100/H200 and A100 chips for Chinese companies, forcing use of Huawei Ascend NPUs or domestic GPUs with materially inferior FLOPs/dollar performance. SE017, SE018
CE017 Prague Technology's second critical dependency is training data from real robot-environment interactions; without diverse OEM-provided real robot data, the world model will overfit simulation and fail to bridge the sim-to-real gap. SE021, SE020
CE018 Prague Technology has disclosed no OEM data partnership agreements as of July 2026; the absence of any confirmed robot data partner is the most concerning technical development-stage risk. SE024, SE025
CE019 Tencent's strategic investment provides potential access to Tencent Cloud compute resources, which could partially offset the GPU access constraint from US export controls if Prague and Tencent negotiate a compute subsidy arrangement. SE020, SE013
CE020 The sim-to-real transfer gap is a major unsolved challenge in embodied AI: models trained exclusively on simulation perform poorly on real robots due to sensor noise, physics inaccuracies, and environmental distribution shifts that synthetic data cannot replicate. SE007, SE021
CE021 Physical Intelligence's π0 addressed the sim-to-real gap through multi-task real-world robot data collection across diverse environments — establishing the precedent that world model and VLA approaches both require large-scale real-robot interaction data, not just simulation. SE005, SE007
CE022 Prague Technology will be required to comply with China's Interim Measures for Generative AI Services (MIIT, effective August 2023) and subsequent regulations when it launches a generative AI model product; compliance requires government registration and safety evaluation submission. SE017, SE018
CE023 Commercial deployment of Prague's AI brain in industrial robots will require industrial safety certifications including ISO 10218 (industrial robot safety) and potentially IEC 62443 (industrial cybersecurity) before OEM customers can use the system in manufacturing environments. SE012, SE017
CE024 AI safety and alignment requirements for physical AI systems deployed in consequential environments are substantially more rigorous than for pure software applications; OpenAI's Sora system card and Anthropic's safety framework both document the depth of technical safety work required for large-scale generative AI deployment. SE011, SE012
CE025 Prague Technology has published no product roadmap or development milestone targets; all timeline estimates in this report are analyst inferences based on comparable world model company timelines. SE024, SE025
CE026 Based on Physical Intelligence's π0 (18-month founding-to-publication timeline) and AMI Labs' 2027-2028 product target, Prague Technology's earliest credible proof-of-concept milestone is mid-2027 — approximately 12-13 months after founding. SE005, SE021
CE027 The humanoid foundation model research consensus places reliable, unattended generalist autonomy (the product capability Prague requires for commercial OEM adoption) in the 2028-2030 timeframe — meaning Prague's commercial launch window is 4-6 years after founding. SE021
CE028 Lin Junyang is the only publicly identified team member of Prague Technology; no co-founders, CTO, VP Engineering, or research directors have been named — indicating either extreme early-stage stealth or that the team has not yet been assembled beyond a core founding group. SE024, SE014
CE029 Prague Technology's complete absence of any technical publication, blog post, or architecture documentation is a notable contrast to comparably-staged peers: Physical Intelligence published the π0 paper within its first year; AMI Labs has published model cards and blog posts; TARS AI has published AWE 3.0 technical specifications. SE005, SE006, SE022
CE030 Technical due diligence of Prague Technology is impossible from public sources: no architecture, no benchmarks, no team composition, no model card, no data strategy; any investment decision requires a technical data room disclosure including architecture whitepaper, team roster, compute strategy, and training data roadmap. SE024, SE025
CE031 The Qwen model family has become one of China's most widely adopted open-source LLM series with millions of downloads on HuggingFace; this developer adoption signal demonstrates Lin Junyang's track record of building models that gain real-world developer traction. SE004, SE015
CE032 Qwen3 achieved SOTA performance on multiple coding, reasoning, and multimodal benchmarks at its parameter class as of its April 2026 release; Lin Junyang's team delivered competitive results against GPT-4o and Claude 3.5 Sonnet within the Alibaba compute budget. SE016, SE001
CE033 Prague Technology has no disclosed IP strategy (patents, trade secrets, open-source) as of July 2026; given the competitive AI landscape, the IP strategy will significantly affect moat durability and competitive defensibility. SE025, SE013
CE034 Building or licensing a high-fidelity physical simulation environment comparable to NVIDIA Isaac Sim or Genesis is itself a multi-year engineering project that represents a critical and underappreciated capital requirement for Prague's world model training strategy. SE021, SE019
CE035 Prague Technology's hardware form factor focus (humanoid, arm, mobile base) has not been publicly specified; this ambiguity affects training data strategy, OEM partnership targeting, and time-to-market, as different form factors require different embodied training datasets. SE024, SE025
CU001 Prague Technology has zero commercial customers, zero revenue, zero signed LOIs, and zero pilot agreements as of July 2026 — the company is entirely pre-commercial. SU001, SU002, SU019
CU002 The absence of any Prague Technology customer relationship is expected at this stage given the company was incorporated in May 2026 and has no product; the commercial gap will not become meaningful until the company enters customer evaluation mode in 2027. SU001, SU022
CU003 Prague Technology's projected customer acquisition timeline: first OEM data partner (non-paying) by H2 2027, first revenue-generating OEM trial by H2 2028, first multi-year license by 2029 — all estimates based on comparable world model and VLA startup timelines. SU011, SU022
CU004 Prague Technology faces a chicken-and-egg bootstrapping problem: OEM customers require a working world model demonstration before signing data partnerships, but training an effective world model requires diverse OEM data — a structural barrier that must be solved through non-commercial research partnerships or synthetic data generation. SU011, SU022
CU005 Prague Technology's intended customer value proposition — reducing OEM per-task integration cost and improving generalization across new manipulation tasks — is validated by industry demand signals but has not been demonstrated by Prague itself. SU007, SU019
CU006 China's humanoid robot OEM market saw ¥46 billion ($6.4B) in H1 2026 investment across 10,000+ companies, indicating a large and commercially active prospective customer base for Prague's world model product. SU023, SU008
CU007 Prague Technology's primary target customer segment is China OEM humanoid robot manufacturers — specifically companies building robots for manufacturing, logistics, and industrial applications, including AgiBot, Unitree, UBTECH, SIASUN, and ESTUN. SU003, SU006
CU008 The B2B buyer for Prague Technology's world model is the VP Engineering or Chief Robotics Officer at an OEM robot manufacturer; the payer is the OEM company; the end-user is the robot itself running Prague's AI brain — a standard platform software sales structure. SU007, SU003
CU009 China's addressable OEM humanoid robot market includes 100+ major manufacturers, with the top 10 (AgiBot, Unitree, UBTECH, SIASUN, ESTUN, FANUC China, Yaskawa China, Dobot, Elephant Robotics, Elite Robots) representing the majority of commercial volume. SU006, SU025
CU010 AgiBot reported commercial deployment of approximately 10,000 humanoid robot units by mid-2026 across manufacturing applications, representing the most advanced China OEM customer base for AI brain integration. SU006, SU007
CU011 Unitree Robotics received CSRC approval to proceed with its IPO review in 2026, with reported H1 2026 revenue of approximately ¥500 million — confirming that the OEM humanoid robot market is commercially real and scaling. SU006, SU015
CU012 UBTECH Robotics serves enterprise customers across manufacturing and hospitality segments with humanoid robots — indicating that the commercial OEM customer base for AI brain solutions spans multiple verticals beyond just automotive manufacturing. SU005, SU003
CU013 Prague Technology has made no public statements about its target customer segment, go-to-market strategy, or priority OEM relationships — adding to the commercial opacity that characterizes the company's pre-product phase. SU001, SU002
CU014 Based on comparable timelines — Physical Intelligence's π0 published 18 months after founding, TARS AI with shipped products 12 months after founding — Prague Technology's earliest credible first customer interaction (data partnership) is Q4 2027. SU011, SU022
CU015 Prague Technology has zero named customers; the named customer proof table in this chapter is empty for Prague, with TARS AI and Physical Intelligence cited as competitor proxies to establish the benchmark for what customer proof looks like at the next stage. SU001, SU002
CU016 TARS AI's AWE 3.0 world model is reportedly integrated into Foxconn manufacturing lines, representing the most advanced China customer proof for an embodied AI world model — the benchmark Prague must eventually match to compete in the same customer tier. SU009, SU010
CU017 Physical Intelligence's π0 is in 'early trials' with OEM customers as of 2026 after 18+ months from founding — indicating that even well-funded world model companies with published results take 2+ years to reach production OEM deployments. SU011, SU012
CU018 China OEM humanoid robot manufacturers evaluate AI brain suppliers on: (1) manipulation success rate >95% for production, (2) latency <200ms for industrial tasks, (3) generalization across SKUs, and (4) integration cost — all criteria that Prague cannot currently demonstrate. SU003, SU007
CU019 The typical B2B enterprise AI sales cycle for robotics AI integration is 6-18 months from first contact to signed contract, with an additional 6-12 months for safety validation before production deployment — meaning first revenue is 12-30 months from first OEM conversation. SU012, SU022
CU020 AgiBot's World Challenge 2026 attracted 526 teams from 38 countries, demonstrating that the OEM robotics ecosystem actively demands AI brain solutions and is investing resources to develop them — validating the demand side of Prague's market thesis. SU007, SU019
CU021 Forbes documented in June 2026 that AI startups with zero revenue and zero customers are supersizing valuations with narrative — the adverse customer proof case applicable to Prague Technology's $2B valuation with no customer evidence. SU013, SU014
CU022 Once an OEM robot manufacturer integrates Prague's world model via SDK into its robot firmware and fine-tunes the model on OEM-specific data, switching costs will be high — requiring re-engineering of the integration layer and retraining a new model on OEM data. SU022, SU011
CU023 Embedded AI model businesses historically achieve NRR (net revenue retention) of 110-130% once deployed in enterprise workflows, driven by expansion to more robot units, new task categories, and new factory sites — a favorable long-term retention model if Prague achieves initial deployment. SU022
CU024 Prague Technology has zero customer retention metrics to report: no NRR, no GRR, no churn, no contract length, no satisfaction score — the entire retention framework is prospective and theoretical at this stage. SU001, SU002
CU025 Prague Technology's first 1-3 OEM customers will represent 100% revenue concentration by construction — this is not an anomalous risk but the normal starting state for enterprise B2B AI startups in year 1 of commercial operation. SU012, SU022
CU026 TARS AI's head start of 2-3 years in OEM market development creates a material risk that China's top 5-10 OEM customers will be locked into the TARS AI ecosystem before Prague Technology launches, forcing Prague to target second-tier OEMs or international markets. SU009, SU010
CU027 Tencent's strategic investment in Prague Technology may create a channel to Tencent-affiliated OEMs and its enterprise robotics ecosystem — providing a non-obvious customer acquisition lever that reduces cold-start sales friction. SU017, SU001
CU028 OEM training data ownership disputes — where OEM customers may claim co-ownership of model improvements from their contributed data — could block Prague's land-and-expand model if not explicitly addressed in first OEM agreements. SU022, SU003
CU029 Prague Technology has no public disclosure of a go-to-market strategy, first target customer segment, or sales team composition — adding to the commercial opacity that characterizes the company's pre-product phase as of July 2026. SU001, SU002
CU030 The global addressable market for AI brain components for OEM robot manufacturers is estimated at $68 billion service robot market by 2030 (IFR), with cognitive AI components representing a growing share — providing the long-term TAM context for Prague's customer acquisition strategy. SU025, SU023
CU031 The China humanoid robot OEM market structure differs from the US market in customer consolidation — China has a large number of funded OEMs (100+) creating more potential customers but also more fragmentation, while the US market is concentrated among 5-10 well-funded OEMs including Boston Dynamics, Agility, Apptronik, Figure AI, and 1X Technologies. SU008, SU006
CU032 Prague Technology's data partnership structure for first OEM customers likely involves offering free access to the world model or model improvement in exchange for OEM robot interaction data — the standard bootstrapping mechanism for world model AI brains with no paying customers yet. SU011, SU022
CU033 VentureBeat and Sifted both reported on the 2026 China embodied AI investment wave indicating international media awareness of the OEM customer market dynamics — although no specific Prague Technology customer details were disclosed. SU021, SU024
CU034 The MIT Technology Review's world model coverage (April 2026) indicates that top-tier tech media is covering this space as enterprise-ready-in-the-making, which helps Prague's customer awareness through analyst and media coverage of world model startups generally. SU016
CU035 The Robotics and Automation News market analysis confirms that China's humanoid robot commercial market is in active growth phase with multiple OEM categories (manufacturing, logistics, elder care) all developing demand — validating the breadth of Prague's prospective customer opportunity. SU025, SU023
CR001 Prague Technology faces a severe, multi-dimensional risk profile: GPU compute access constrained by US export controls, zero product with $2B valuation, existential key-person dependency on Lin Junyang, TARS AI's 2-3 year market head start, and 2-4 year pre-revenue window requiring follow-on capital at uncertain conditions. SR001, SR011
CR002 The five most severe risks for Prague Technology, ranked by residual exposure (likelihood × impact / mitigation maturity), are: (1) GPU export controls, (2) key-person risk, (3) TARS AI market lock-in, (4) world model technical failure, and (5) OEM data partnership failure. SR001, SR002
CR003 Prague Technology has no mitigations in place for any of its top five risks as of July 2026, because it has no product, no operations, and no disclosed risk management framework — all risk mitigations are prospective or unmitigated. SR021, SR023
CR004 No legal proceedings, regulatory investigations, or sanctions are known to be active against Prague Technology, Lin Junyang, or the three primary investors (Gaorong, HongShan, Tencent) as of July 2026. SR009, SR029
CR005 Prague Technology's combined risk score (multiple high-severity, low-mitigation risks) is consistent with an extreme early-stage AI investment profile where the investor is compensated for these risks by the founder premium and first-mover thesis, not by de-risked fundamentals. SR011, SR015
CR006 All primary risks for Prague Technology have identifiable monitoring indicators and kill criteria — enabling a rigorous investment monitoring framework even though the current risk posture is extreme. SR013, SR001
CR007 US export controls have restricted NVIDIA H100, A100, and H20 GPU access for Chinese companies; the H200 faces approximately 25% tariffs; collectively these restrictions force Prague Technology to rely on domestic Chinese GPU alternatives with 40-60% inferior price-performance compared to the best available NVIDIA chips. SR007, SR008
CR008 China's Interim Measures for Generative AI Services (MIIT/CAC, effective August 2023) requires AI model providers to register with regulators, conduct safety evaluations, and comply with content generation restrictions — compliance requirements that Prague Technology must satisfy before commercial launch. SR004, SR005
CR009 US-China tech decoupling has accelerated since 2023; US BIS export controls have progressively expanded to restrict AI chip, software, and tool access for Chinese entities; the trajectory suggests continued escalation that will maintain or worsen Prague's compute access constraints. SR010, SR012
CR010 Prague Technology faces IP risk from Alibaba: Lin Junyang developed Qwen's MoE architecture under his Alibaba employment; if Alibaba alleges trade secret misappropriation for architectural approaches used in Prague's world model, Prague faces litigation risk and potential product development delays. SR022, SR004
CR011 Prague Technology has no known patents protecting its world model approach; the core technical concepts (world models, MoE, multimodal encoders) are in the public domain through academic publications — meaning the moat must be built through execution, data accumulation, and team, not IP. SR022, SR021
CR012 China's Personal Information Protection Law (PIPL) and Data Security Law create obligations when processing robot sensor data that may capture images of people or sensitive locations — creating data governance requirements for any OEM deployment in populated environments. SR004, SR006
CR013 Secondary sanctions risk for Prague Technology is low probability but high impact: if US sanctions specifically target Chinese AI model companies (analogous to Entity List or OFAC actions), Tencent's strategic investment and HongShan's US LP relations could create compliance complications for Prague. SR010, SR012
CR014 Regulatory risk timeline: China AI rules (Q4 2026 likely updates), industrial robot safety certification (2027-2028 during product development), CAC algorithm security assessment (pre-commercial launch) — all regulatory milestones are sequenced within Prague's development timeline. SR005, SR006
CR015 World model development for embodied AI remains an unsolved technical challenge as of July 2026; no company has demonstrated a general-purpose world model operating at commercial reliability in real physical environments, making technical failure of Prague's approach a significant probability. SR013, SR014
CR016 The sim-to-real transfer gap is a well-documented technical challenge in embodied AI: models trained on simulation fail to generalize to real robot environments due to sensor noise, physics inaccuracies, and environmental variability — requiring Prague to secure OEM robot data partnerships before its world model will achieve commercial-grade performance. SR014, SR013
CR017 World model inference at robot control frequency (30-100Hz reactive, 1-10Hz deliberative) presents a fundamental compute-latency challenge: current frontier models (GPT-4 scale, Genie 2 scale) operate at 1-5Hz for complex reasoning, requiring Prague to solve 10-100x latency reduction through model distillation or hardware optimization. SR013, SR027
CR018 AI safety failures in physical robot systems carry liability risks that are qualitatively different from software AI failures: a world model controlling a 50kg humanoid robot that causes property damage or personal injury could expose Prague Technology to tort liability, regulatory sanctions, and reputational damage that could end the company. SR003, SR013
CR019 Prague Technology is dependent on Tencent for strategic compute access; if Tencent's own regulatory situation changes (data security investigations, US sanctions actions) or if the relationship structure changes, Prague could lose preferential compute access with no immediate replacement. SR022, SR021
CR020 Prague Technology's critical OEM data partnership risk: zero data partners signed as of July 2026; without 3-5 OEM partners contributing real-world robot interaction data before H1 2027, the world model's sim-to-real gap will remain unresolved and commercial deployment by 2028 is implausible. SR013, SR026
CR021 Huawei Ascend 910B and domestic GPU alternatives provide approximately 40-60% lower FLOPs per dollar than NVIDIA H100 for LLM/world model training workloads; this materially increases the capital requirement per training run and extends Prague's timeline to its first production-quality world model. SR007, SR008
CR022 Huawei itself faces US sanctions that could disrupt Ascend GPU production if further chip manufacturing restrictions are imposed; this supply-chain risk creates a second-order dependency where Prague's primary compute alternative is also a geopolitical risk target. SR010, SR007
CR023 Chinese AI companies including Baidu (Ernie), ByteDance (Doubao), and Zhipu AI have demonstrated that domestic GPU alternatives (Huawei Ascend, Biren BR100) can train competitive frontier models despite export restrictions — providing a precedent that Prague's compute constraint is a cost challenge rather than a total blocker. SR008, SR002
CR024 Prague Technology's follow-on financing risk: if burn exceeds $10M/month, the company needs a Series A by December 2027; 2027 AI investment market conditions are uncertain; a flat or down-round at $2B→$1B would be 50% dilutive, severely damaging employee morale and investor returns. SR016, SR015
CR025 Prague Technology's key-person risk is existential: as of July 2026, Lin Junyang appears to be the only named employee and the company's sole technical, commercial, and reputational asset. No succession plan is possible because there is no acknowledged team. SR021, SR023
CR026 Prague Technology lacks a CFO, COO, VP Engineering, and VP Sales — roles critical for a company transitioning from research to commercial operations. Lin Junyang's background is technical (model architecture, training infrastructure), not operational (team building, enterprise sales, regulatory navigation). SR023, SR021
CR027 Prague Technology competes for AI talent against Alibaba, ByteDance, Tencent, Baidu, and 10,000+ China embodied AI startups — all offering more financial stability, larger teams, and often superior compensation packages; talent acquisition and retention will be a persistent challenge. SR024, SR002
CR028 The three most critical thesis-break tripwires for Prague Technology are: (1) confirmed burn rate >$10M/month without product milestone by Q2 2027; (2) no OEM data partner signed by Q3 2027; (3) Lin Junyang unavailable for any reason. SR015, SR013
CR029 The risk timeline for Prague Technology's most urgent risks: GPU compute cost impact is immediate (present); OEM data partner deadline is Q2 2027 (12 months); burn rate confirmation is needed at financing close (month 0); TARS AI lock-in risk materializes if unaddressed by Q4 2027. SR019, SR013
CR030 Macroeconomic risks relevant to Prague Technology include: China economic slowdown reducing domestic OEM capex budgets, global AI investment cycle cooling reducing follow-on round availability, and USD/CNY currency fluctuations affecting the effective USD value of CNY-denominated revenue. SR011, SR012
CR031 TARS AI's competitive head start represents a market lock-in risk: with production deployments at Foxconn and 20+ enterprise trials, TARS is accumulating OEM-specific training data that creates a flywheel advantage; if TARS secures 10 of China's top OEMs before Prague launches, Prague's addressable market shrinks critically. SR019, SR020
CR032 Prague Technology's organizational risk is compounded by the early stage: the company has raised $220M without disclosing any team beyond Lin Junyang — creating a perverse accountability gap where investors have committed capital without visibility into the execution team that will deploy it. SR015, SR023
CR033 The risk mitigation priority for investors should be: (1) immediate — confirm burn rate and OEM partnership pipeline; (2) 90 days — hire at least one co-founder or VP-level technical leader; (3) 6 months — sign first OEM data partnership; (4) 12 months — publish first technical proof-of-concept. SR013, SR015
CR034 S&P Global specifically flagged concentration in illiquid, pre-revenue VC positions and speculative long-duration AI bets as hidden investment risks in the 2025-2026 period — Prague Technology's investor profile (Gaorong, HongShan, Tencent) creates exactly this type of concentration. SR011
CR035 Prague Technology's most recent risk materialization evidence (as of July 2026): the US export control escalation in 2025-2026 is confirmed, ongoing, and worsening; TARS AI's market progress is confirmed; Prague's zero product status is confirmed — no risks have been mitigated or resolved. SR007, SR019, SR023
CR036 Prague Technology's regulatory compliance obligation timeline aligns with its development timeline: MIIT generative AI registration required before commercial launch (target 2028), industrial robot safety certification required before OEM production deployment — regulatory risk is manageable if development stays on schedule. SR005, SR004
CR037 The Qichacha and Tianyancha corporate registry entries for Shanghai Bulage Technology show a recently registered company with standard startup registration structure; no disclosed litigation, enforcement actions, or legal proceedings as of the registry access date. SR009, SR029
CR038 Forbes's June 2026 article on pre-revenue AI startup valuation inflation is the most current adverse risk reference for Prague Technology's $2B valuation; it specifically documents the 'narrative premium' that has inflated AI startup valuations beyond revenue-supportable levels. SR015, SR016
CR039 Operational risk monitoring for Prague Technology is impossible from public sources: no financial statements, no burn rate disclosures, no OEM partnership announcements, no team headcount data — operational monitoring requires investor-level information access. SR021, SR023
CR040 The combination of extreme valuation premium (9x P/Capital), zero product, single named employee, and geopolitical compute constraints makes Prague Technology one of the highest-risk AI investments in China's 2026 embodied intelligence cohort — albeit also one of the highest-potential given Lin Junyang's track record. SR015, SR001, SR011
CV001 Prague Technology (Shanghai Bulage Technology Co., Ltd.) achieved a confirmed post-money valuation of $2 billion in June 2026 following completion of its angel/seed funding round. SV001, SV003, SV022
CV002 Prague Technology raised approximately $220 million in its angel/seed round, comprising $100M from Gaorong Ventures, $100M from HongShan Capital, and $20M from Tencent, completed in June 2026. SV003, SV022, SV026
CV003 Lin Junyang, founder of Prague Technology, served as the chief architect and lead of Alibaba's Qwen LLM series, which became China's leading open-source large language model family prior to his departure in early 2026. SV021, SV022, SV028
CV004 Lin Junyang's Qwen credential represents a stronger China-specific AI model development track record than most comparable pre-revenue AI lab founders at the time of Prague Technology's founding. SV011, SV022
CV005 Prague Technology had no product, no revenue, and no disclosed customers as of July 2026, making it a pre-product, pre-revenue entity at the time of its $2 billion valuation. SV026, SV022, SV021
CV006 The Wall Street Journal and Barron's each published articles in 2026 flagging concerns about inflated valuations in the China AI startup ecosystem, with some analysts characterizing valuations as primarily speculative. SV007, SV009
CV007 Prague Technology's founding date was May 27, 2026 (date of 上海卜拉格科技有限公司 registration), making it one of the youngest companies ever to achieve a $2 billion unicorn valuation — approximately three weeks from founding to unicorn status. SV026, SV022
CV008 AMI Labs, the world model startup founded by Turing Award winner Yann LeCun, raised $1.03 billion at a $3.5 billion pre-money valuation in March 2026, representing the highest confirmed valuation for a pre-revenue world model company at that time. SV011, SV012, SV016
CV009 World Labs, the 3D world model company founded by Fei-Fei Li, raised $1 billion in February 2026 as part of a round targeting a $5 billion valuation, though the final valuation was not officially confirmed. SV013, SV011
CV010 Physical Intelligence (Pi), the US-based robotics foundation model company, had raised over $400 million at an estimated valuation of approximately $2 billion or above as of mid-2026. SV014, SV025
CV011 TARS Group, a Chinese humanoid AI company, was reported trading at approximately $2 billion valuation or above in mid-2026, making it a direct China-based comparable for Prague Technology. SV025, SV029
CV012 Zhiyuan Robotics and Unitree Robotics, China-based embodied AI and humanoid robotics companies, each had valuations in the $1.5 billion range in 2026, representing the lower end of the China embodied AI valuation band. SV025, SV024, SV029
CV013 Prague Technology's $2 billion post-money valuation places it at the low end of the global world model and embodied AI peer set, representing a 40-75% discount to AMI Labs and a significant discount to World Labs' $5 billion target. SV011, SV013, SV022
CV014 The valuation discount for Prague Technology relative to Western AI lab comparables reflects a quantifiable China risk premium incorporating GPU export controls, geopolitical uncertainty, and domestic regulatory complexity. SV007, SV015, SV019
CV015 The $220 million raised at a $2 billion post-money valuation implies investors collectively received approximately 11% ownership in Prague Technology at the close of the angel round, assuming no significant option pool dilution or convertible notes. SV022, SV026
CV016 In the bull scenario for Prague Technology (estimated 25% probability), a successful world model product launch by Q2 2027 combined with major OEM partnerships could drive the company to a $15-20 billion valuation by 2029-2030, implying a 7x-10x return on entry at $2 billion. SV001, SV005, SV025
CV017 In the base scenario for Prague Technology (estimated 45% probability), moderate commercial traction with some industrial partnerships by 2030 suggests a $5-7 billion valuation exit, implying a 2.5-3.5x return on entry at $2 billion. SV001, SV019, SV024
CV018 In the bear scenario for Prague Technology (estimated 30% probability), product delays, GPU restrictions, team departures, or competitive displacement could result in a down round or acquisition at $0.5-1.5 billion, implying substantial loss from the $2 billion entry. SV007, SV009, SV015
CV019 The probability-weighted expected return multiple for a hypothetical entry at Prague Technology's $2 billion valuation is approximately 3.0x gross, calculated as (25%×8.5x) + (45%×3.0x) + (30%×0.5x) = 2.1+1.35+0.15 ≈ 3.6x central estimate. SV001, SV019
CV020 The recommended hold period for a Prague Technology investment is 4-6 years based on comparable embodied AI company commercialization timelines: AMI Labs expects years before commercial applications, Physical Intelligence shipped Pi0 within two years of founding. SV011, SV014
CV021 Exit pathways for Prague Technology investors realistically include: pre-IPO secondary transactions (earliest: 2028), acquisition by a Chinese industrial conglomerate (Huawei, Xiaomi, BYD, or DJI), Hong Kong Stock Exchange listing, or a US-listed entity if regulatory environment permits. SV006, SV018, SV019
CV022 Sequoia Capital's AI Ascent IV in May 2026 declared 2026 'the year of agents' and highlighted that AI models, tools, and harnesses have 'finally come together,' validating the macro timing of Prague Technology's founding and focus. SV005, SV001
CV023 CB Insights' State of Venture Q2 2026 report found that funding topped $200 billion for the second consecutive quarter, with mega-rounds accounting for 81% of all capital deployed, confirming an environment of elevated valuations that directly benefits Prague Technology's pre-revenue positioning. SV001, SV006
CV024 AMI Labs CEO Alexandre LeBrun stated that AMI Labs has no plans to generate revenue for the near term and that the world model development timeline could take 'years' before commercial applications are viable — a parallel commentary applicable to Prague Technology. SV011, SV012
CV025 The overall investment verdict for Prague Technology is neutral — a conditional watch-list position — reflecting balanced exceptional founder quality and market timing against extreme pre-revenue risk, China risk premium, and valuation that depends entirely on future delivery. SV022, SV005, SV007
CV026 A demonstrable embodied AI world model, a strategic partnership with a Chinese robotics OEM, or a first peer-reviewed research publication would each represent positive catalyst events in the next 12 months sufficient to shift the investment stance from neutral to invest. SV005, SV001
CV027 Entry discipline for any institutional investor entering at Prague Technology's $2 billion valuation should include strong information rights, pro-rata participation rights in future rounds, anti-dilution provisions, and board observer rights given the extreme early stage and pre-revenue status. SV001, SV005
CV028 The confidence in the overall investment recommendation for Prague Technology is low due to the company's founding date of May 2026, complete absence of product or revenue, and the dependence of the entire $2 billion valuation on a single founder's credentials and an unvalidated market thesis. SV026, SV007
CV029 The single most important thesis-break trigger for Prague Technology is departure of Lin Junyang or material reduction in his leadership role, which would effectively eliminate the primary basis for the $2 billion valuation. SV022, SV021, SV003
CV030 A down round at below $1.5 billion post-money, or a next financing below current valuation, would signal collapse of investor confidence and represent a critical kill trigger for existing investors. SV007, SV009
CV031 Absence of any demonstrable embodied AI model output by Q4 2027 — 18 months after the June 2026 valuation close — would represent a significant product delay trigger warranting a thesis reassessment. SV011, SV013
CV032 Any MIIT or CAC enforcement action targeting Prague Technology's training data practices or generative AI model outputs would represent a high-priority regulatory adverse event trigger requiring immediate legal review. SV015, SV007
CV033 The most critical first diligence ask for Prague Technology is a full capitalization table disclosing issued shares, option pool, investor ownership by class, and any SAFE or convertible note conversion terms to verify the implied ~11% dilution. SV026, SV022
CV034 A technical architecture document describing Prague Technology's world model architecture and differentiation from existing Qwen models is essential to validate the IP claim underlying the $2 billion valuation. SV005, SV011
CV035 Prague Technology's compute access and GPU procurement plan is a critical diligence item given US export controls restricting NVIDIA H100/A100/H20 access for Chinese companies, forcing reliance on domestic GPU suppliers at inferior price-performance. SV017, SV015
CV036 An 18-month research and product roadmap with defined deliverable milestones is a required diligence document for Prague Technology to enable investors to track progress against the bull case timeline assumptions. SV005, SV001
CV037 Pre-revenue AI lab valuation in 2026 is primarily driven by team quality and founder pedigree, perceived market optionality, and investor competition for access — with revenue multiples inapplicable at the pre-product stage. SV004, SV008, SV001
CV038 Prague Technology's $2 billion entry valuation is the lowest among global world model and embodied AI peers with comparable funding scale, suggesting either a valuation discount reflecting China risk, a relative bargain for investors accepting geopolitical exposure, or both. SV011, SV013, SV019
CV039 A probability-weighted scenario analysis using Bull 25%, Base 45%, Bear 30% probabilities with respective valuations of $17B, $6B, and $1B produces an expected exit value of approximately $7.0B, implying a gross multiple of approximately 3.5x on a $2B entry over a 4-6 year horizon. SV001, SV019
CV040 a16z partner commentary and Sequoia AI Ascent IV proceedings confirm that the global AI investment community views 2026 as a pivotal year for AI commercialization, with robotics and physical AI receiving particular emphasis as the next major application wave after language and code. SV005, SV008
来源
编号出版方标题引文
SO001 36Kr (English) Report: Lin Junyang Secures Tencent Investment with First-round Financing two people familiar with the matter revealed that the AI laboratory founded by Lin Junyang...has completed its first-round financing. The total financing amount reached hundreds of millions of dollars, and the post-investment valuation reached $2 billion
SO002 36Kr (English) Lin Junyang Launches Business: New Company Valued at Around $2 Billion Lin Junyang, the former technical leader of Alibaba's Qianwen large-model, has recently started his own business. The directions he is considering include world models and embodied brains.
SO003 BigGo Finance Exclusive: Former Alibaba Qwen Chief Lin Junyang's Startup Lands Sequoia China, Gaorong Investment a post-money valuation of $2 billion (~13.5 billion yuan). The company currently has no products, no revenue, and has not even publicly disclosed an official name
SO004 BigGo Finance Ex-Alibaba Qwen Chief Lin Junyang's Startup Lands Sequoia China Investment
SO005 BridgingChina Lin Junyang's New Company 'Blag' Makes Its Debut! Initial Valuation of 13.5 Billion RMB For Lin Junyang, the $2 billion valuation is both an honor and a shackle.
SO006 CNTechPost Former Alibaba AI Core Figure Lin Junyang Founds New Lab, Seeks $2 Billion Valuation
SO007 MarketScreener Shanghai Pragmatics Technology Co., Ltd. announced funding from HongShan Capital
SO008 TechCrunch Alibaba's Qwen Tech Lead Steps Down After Major AI Push
SO009 Baidu Baike (Baiduwiki) Shanghai Pragmatics Technology Co., Ltd. — Baidu Encyclopedia
SO010 Preqin Shanghai Pragmatics Technology Co., Ltd. Asset Profile
SO011 QBit AI (量子位) 林俊旸新公司卜拉格亮相!首轮估值135亿 (Lin Junyang's New Company Prague Debuts! First-round Valuation RMB 13.5B)
SO012 MarkTechPost Qwen's Former Lead on What Hybrid Thinking Got Wrong — and Why He Now Backs Agents Multimodal foundation models are transforming into foundation agents, using tools and memory for long-term sequential reasoning through reinforcement learning. They should move from the virtual to the physical world!
SO013 Sina Finance (新浪科技) 林俊旸新公司「卜拉格」亮相!首轮估值135亿 (Lin Junyang's Prague Debuts)
SO014 TechBuzzChina LIN Junyang (林俊旸) — China AI Atlas Profile
SO015 Gaorong Ventures Gaorong Ventures Official Website
SO016 Baidu Baike Gaorong Capital — Baidu Encyclopedia
SO017 VCMatch Gaorong Ventures Profile — VCMatch
SO018 HongShan Capital HongShan (HSG) Official Website
SO019 HongShan Capital HongShan Portfolio Companies
SO020 Brief the Call HongShan — Multi-stage VC, China (ex-Sequoia China) As of 2026, it manages $55B+ in AUM and over 1,500 portfolio companies
SO021 NewsGlobeNow Former Alibaba Qwen Lead Lin Junyang Starts AI Startup
SO022 AI Market Watch 卜拉格科技 (Prague Technology) — AI Startup Profile
SO023 Forbes AI Startups With No Revenue Are Using This Tactic To Supersize Their Valuations These valuations are based mostly on founder pedigree, research potential, and speculation, not on actual revenue or product-market fit, making investments highly risky if the promised breakthroughs are delayed or fail to materialize.
SO024 Morgan Stanley AI Market Trends 2026: Global Investment, Risks, and Buildout
SO025 TechCrunch In 2026, AI Will Move From Hype to Pragmatism
SM001 AInChina China's Embodied AI Revolution: Funding, World Models, and the $2B Startup Wave in 2026 In the first three months of 2026, venture capitalists poured $3.3 billion into Chinese robotics startups across 126 deals.
SM002 EmbodiedGlobal / Qixinbao China Embodied AI Industry Surpasses 10,000 Companies: Panorama Insight Report 2026 As of May 2026, over 10,000 companies are now operating in this space, including 269 listed companies.
SM003 EmbodiedGlobal China Embodied AI Sector Raises ¥37B in H1 2026, 288 Financing Events Total H1 2026 financing in the embodied intelligence and robotics sector reached over 46 billion yuan across 288 events involving 226 companies.
SM004 AI Robotic Daily Embodied AI Market 2026: The Unprecedented Boom This massive funding wave...created twelve new entities valued over ten billion RMB...the exclusive club of highly valued robotics firms now boasts sixteen members.
SM005 China Daily China Embodied AI Industry Sees Robust Growth, Shanghai Expo Report IDC forecasts that the global embodied AI market will reach $1.5 trillion by 2030. Global shipment of humanoid robots in 2025 was about 18,000 units, a year-on-year increase of about 508 percent.
SM006 Qiushi Theory (CCP official journal) / China Daily China Embodied AI Sector Registers Robust Growth Jan-May 2026, State Taxation Administration Data China's embodied artificial intelligence industry registered robust growth...China's 15th Five-Year Plan (2026-30), which designates embodied AI as one of six major future industries alongside quantum technology and biomanufacturing.
SM007 humanoid.guide Humanoid Foundation Models — World Models Report 2026 $13.8B record robotics venture funding in 2025. China's share of humanoid unit sales ~90% and component supply chain ~70%.
SM008 Data-Gate Embodied AI 2026: From Simulation to Real-World Robots Domain randomization at scale: Training across millions of randomized physics parameters so the model generalizes to real-world conditions.
SM009 Qubittool (量子位 analysis site) Embodied AI 2026: Robot Foundation Models, VLA, and World Models Data scarcity remains a grand challenge, hindering our ability to replicate scaling laws. Existing 3D data and robotic action data are not only orders of magnitude smaller than text.
SM010 Microsoft Research Asia StarTrack Scholars 2026: Crafting Spatial and Embodied Foundation Models Data scarcity remains a grand challenge compared to the progress seen in LLMs. Existing 3D data and robotic action data are not only orders of magnitude smaller than text and images.
SM011 Nature Machine Intelligence Scaling laws and embodied AI: A survey of 2026 milestones (paywalled)
SM012 arXiv (Cornell University) Embodied Intelligence Survey 2026: Foundation Models, World Models, and Physical AI
SM013 Curionic BYD-AgiBot vs Unitree vs UBTECH: China Humanoid Robot Comparison 2026 AgiBot is the single most important company in Chinese humanoid robotics that most Western coverage ignores. Ranked number one globally in 2025 humanoid shipments by Omdia at 5,168 units.
SM014 AInChina China's Embodied Intelligence Revolution 2026: From CCTV Gala to Factory Floor In 2025, Chinese companies accounted for roughly 87% of all humanoid robots shipped globally. The number of humanoid robot manufacturers in China exceeded 140, with over 330 product models released.
SM015 Robot Today WAIC 2026: Embodied Intelligence Enters Industrial Deployment Era
SM016 ChinaBizInsider 15 Embodied AI Unicorns in 6 Months: China's Robot Race Hits a Reality Check
SM017 TechTimes Unitree IPO Cleared, AgiBot Hits 10,000 Units: China Humanoid Robot Duopoly Takes Shape Unitree's Shanghai IPO approved June 2026 targeting ~$610M.
SM018 EqualOcean TARS AI Closes $455M Pre-A Round, Breaking China Embodied AI Financing Record
SM019 36Kr (English) Prague Technology (卜拉格 / Bulage): Lin Junyang's World Model Startup Secures $220M Angel Round
SM020 TechCrunch Alibaba's Qwen tech lead steps down after major AI push
SM021 Baidu / Qichacha 2026 China Embodied Intelligence Industry Development Report
SM022 PitchBook / Crunchbase News Yann LeCun's AMI Labs Secures $1B in Bet on World Models
SM023 CNTechPost Former Alibaba AI Core Figure Lin Junyang Founds New Lab, Seeks $2 Billion Valuation
SM024 Brookings Institution Competing AI Strategies for the US and China
SM025 Humanoid Robotics Technology AgiBot World Challenge 2026: Embodied AI Competition Advances from Simulation to Real Testing
SP001 EqualOcean Chinese Embodied AI Company TARS Raised $455M Pre-A Round Breaking Record TARS has officially closed a $455 million Pre-A funding round, setting the record for the largest single-round financing in China's embodied intelligence sector.
SP002 Gasgoo (AutoNews) Seeds Discovery: Former Huawei and Baidu Executives Team Up, Secure Over ¥3B in Financing Founder and CEO Chen Yilun previously served as CTO of Autonomous Driving and Chief Scientist of the Automotive BU at Huawei. Chairman Li Zhenyu, the former president of Baidu's Intelligent Driving Group (IDG).
SP003 Robot Today TARS AI Closes $455M Pre-A Round Setting China Embodied AI Financing Record
SP004 CNTechPost Chinese Embodied AI Startup TARS Raises $455 Million Pre-A Funding Round
SP005 TechCrunch Yann LeCun's AMI Labs Raises $1.03 Billion to Build World Models AMI Labs, the new venture co-founded by Turing Award winner Yann LeCun after he left Meta, has raised $1.03 billion at a $3.5 billion pre-money valuation.
SP006 Futurum Group Yann LeCun's AMI Raises $1BN Seed Round — Is the World Model Era Finally Here? AMI Labs is not a typical applied AI startup that can release a product in three months, have revenue in six months — it could take years for world models to go from theory to commercial applications.
SP007 Crunchbase News World-Model AI Lab AMI Raises Europe's Largest Seed Round
SP008 Observer Yann LeCun's AMI Startup Raises $1B Seed Round
SP009 TechCrunch World Labs Lands $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. World Labs, which emerged from stealth in 2024 with $230 million at a $1 billion valuation.
SP010 AgiBot (official) AgiBot World Challenge 2026 — World Model Track Results
SP011 Humanoid Robotics Technology AgiBot World Challenge 2026 Advances Embodied AI Competition from Simulation to Real Robot Testing AGIBOT hosted AGIBOT WORLD CHALLENGE 2026 alongside ICRA 2026 in Vienna, bringing together 526 research and enterprise teams from 27 countries.
SP012 Curionic BYD-AgiBot vs Unitree vs UBTECH: China Humanoid Robot Comparison 2026 Unitree's open SDK, ROS2, NVIDIA H2+ partnership; UnifoLM-VLA-0 (open-sourced)
SP013 TechTimes Unitree IPO Cleared, AgiBot Hits 10,000 Units: China Humanoid Duopoly
SP014 PitchBook Yann LeCun's AMI Labs Secures $1B in Bet on World Models
SP015 AInChina China's Embodied AI Revolution: Spirit AI, TARS, and the World Model Race Spirit AI: $435M total raised, $1.5B valuation, universal robot brain
SP016 EmbodiedGlobal China Embodied AI H1 2026 Funding — Unicorn Tracker
SP017 AInChina China's Embodied Intelligence Revolution 2026 ByteDance — announced that world models, the foundational technology for embodied AI, are now its number one priority for 2026.
SP018 AI.cc World Models 2026: Google, NVIDIA, and Physical AI Breakthroughs
SP019 arXiv Embodied Intelligence Survey 2026: Foundation Models, World Models, and Physical AI
SP020 Qubittool (量子位 analysis) Embodied AI 2026: World Models vs VLA Models — Architecture Comparison
SP021 GitHub (QwenLM) QwenLM/Qwen3 — Official Repository
SP022 The Air Rankings Alibaba Qwen 3.6 27B Benchmark Comparison
SP023 ChinaBizInsider 15 Embodied AI Unicorns in 6 Months: Reality Check
SP024 humanoid.guide Humanoid Foundation Models — World Models Report 2026 World-action models more than double generalization over VLA baselines (DreamZero 62.2% vs 27.4%)
SP025 Data-Gate Embodied AI 2026: From Simulation to Real-World Robots
SI001 Forbes AI Startups With No Revenue Are Using This Tactic to Supersize Their Valuations AI startups with no revenue are using this tactic to supersize their valuations.
SI002 S&P Global Ratings Where Are AI Investment Risks Hiding?
SI003 CNBC AI Valuation Fears Grip Investors as Tech Bubble Concerns Heighten
SI004 MarketScreener Shanghai Pragmatics Technology Received Funding from HongShan Capital
SI005 Preqin Shanghai Pragmatics Technology Co., Ltd. — Investor Profile
SI006 36Kr (English) Prague Technology Secures $220M Angel Round Led by Gaorong and HongShan
SI007 Futurum Group Yann LeCun's AMI Raises $1BN Seed Round — Is the World Model Era Finally Here? AMI is not a typical applied AI startup that can release a product in three months, have revenue in six months.
SI008 TechCrunch Yann LeCun's AMI Labs Raises $1.03 Billion to Build World Models It could take years for world models to go from theory to commercial applications.
SI009 humanoid.guide Humanoid Foundation Models — World Models Report 2026 2028-30 base-case window for reliable, unattended generalist autonomy.
SI010 TechJournal US Imposes 25% Tariff on NVIDIA H200 AI Chips Bound for China
SI011 The Silicon Review China AI Export Controls Strategy 2026
SI012 Rimon Law China AI Law Brief
SI013 AI Governance China AI Regulation News 2026
SI014 ChinaCrunch China's AI Regulation 2026: Building a Global Framework for Responsible Algorithms
SI015 Morgan Stanley AI Market Trends — Institute Research 2026
SI016 Crunchbase News World-Model AI Lab AMI Raises Europe's Largest Seed Round
SI017 BridgingChina Lin Junyang's New Company BLAG Makes Its Debut with $2B Valuation
SI018 CNTechPost Former Alibaba AI Core Figure Lin Junyang Founds New Lab, Seeks $2 Billion Valuation
SI019 AInChina China's Embodied AI Revolution: Funding, World Models, and the $2B Startup Wave in 2026
SI020 EmbodiedGlobal China Embodied AI H1 2026 Funding Tracker
SI021 Brookings Institution Competing AI Strategies for the US and China
SI022 Gaorong Ventures (official) Gaorong Ventures Official Website
SI023 HongShan Capital (official) HongShan Capital (Sequoia China) Official Website
SI024 PitchBook Yann LeCun's AMI Labs Secures $1B in Bet on World Models
SI025 Creati.ai Alibaba Qwen3.6 27B Beats Larger Models on Coding Benchmarks
SI026 Qichacha (企查查) Shanghai Bulage Technology Co., Ltd. — Corporate Registry Profile
SE001 GitHub — QwenLM/Qwen3 Qwen3: Qwen3 is the large language model series developed by Qwen team, Alibaba Cloud Qwen3 is the large language model series developed by Qwen team, Alibaba Cloud.
SE002 Qwen Team (official blog) Qwen3: Think Deeper, Act Faster
SE003 Qwen Team (official blog) Qwen2.5: A Party of Foundation Models!
SE004 HuggingFace Qwen Model Series — HuggingFace Organization Page
SE005 Physical Intelligence Our First Generalist Policy (π0) This zero-shot generalization capability emerges from training on large amounts of diverse data across many tasks, robots, and environments.
SE006 DeepMind (Google) Genie 2: A Large-Scale Foundation World Model Genie 2 can generate an endless variety of action-controllable, playable 3D environments.
SE007 arXiv A Survey on Vision-Language-Action Models for Embodied AI
SE008 The Decoder World Models: The Next Frontier in AI
SE009 Papers With Code World Models — State of the Art
SE010 Papers With Code Robot Manipulation — State of the Art
SE011 OpenAI Sora System Card
SE012 Anthropic Anthropic's Core Views on AI Safety
SE013 BridgingChina Lin Junyang's New Company BLAG Makes Its Debut
SE014 MarktechPost Qwen's Former Lead on What Hybrid Thinking Got Wrong and Why He Now Backs Agents
SE015 GitHub — QwenLM/Qwen2.5 QwenLM GitHub: Qwen2.5 repository (accessed for developer signal)
SE016 The Air Rankings Alibaba Qwen 3.6 Analysis and Benchmarks
SE017 Rimon Law China AI Law Brief
SE018 ChinaCrunch China's AI Regulation 2026: Building a Global Framework
SE019 Data-Gate Embodied AI Technical Economics and Architecture 2026
SE020 AInChina China's Embodied AI Revolution 2026
SE021 humanoid.guide Humanoid Foundation Models Report 2026 World models go beyond action imitation by learning latent representations of physical laws.
SE022 EqualOcean TARS AI Raises $455M Pre-A, Leads China Embodied AI Wave
SE023 Gasgoo TARS AI AWE 3.0 World Model Technical Details
SE024 QbitAI Prague Technology — Lin Junyang's New Venture
SE025 CNTechPost Former Alibaba AI Core Figure Lin Junyang Founds New Lab
SU001 BridgingChina Lin Junyang's New Company BLAG Makes Its Debut
SU002 QbitAI Prague Technology — Lin Junyang's New Venture
SU003 Curionic BYD vs AgiBot vs Unitree vs UBTECH: China Humanoid Robot Comparison 2026
SU004 Unitree Robotics (official) Unitree Robotics Products
SU005 UBTECH Robotics (official) UBTECH Products Page
SU006 TechTimes Unitree IPO Cleared, AgiBot Hits 10,000 Units — China Humanoid Robot Duopoly Takes Shape
SU007 AgiBot (official) AgiBot World Challenge 2026
SU008 Robotics and Automation News China Humanoid Robots Market 2026
SU009 EqualOcean TARS AI Raises $455M Pre-A, Leads China Embodied AI Wave
SU010 Gasgoo TARS AI Team and AWE 3.0 Technical Details
SU011 Physical Intelligence (official) Our First Generalist Policy (π0)
SU012 TechCrunch In 2026, AI Will Move from Hype to Pragmatism
SU013 Forbes AI Startups With No Revenue Are Using This Tactic to Supersize Their Valuations AI startups with no revenue are using this tactic to supersize their valuations.
SU014 CNBC AI Valuation Fears Grip Investors as Tech Bubble Concerns Heighten
SU015 TechTimes Unitree IPO Cleared — China Humanoid Market Scale
SU016 MIT Technology Review World Models: AI's Next Frontier?
SU017 AInChina China's Embodied AI Revolution: Funding, World Models, and the $2B Startup Wave in 2026
SU018 SCMP Former Alibaba Qwen Chief Wants to Build the World's Best Embodied AI Model
SU019 Humanoid Robotics Technology AgiBot World Challenge 2026: Robotic AI Competition at ICRA Vienna
SU020 TechNode Lin Junyang AI Startup World Model Embodied Intelligence 2026
SU021 Sifted World Models and AMI Labs — The Next Frontier
SU022 humanoid.guide Humanoid Foundation Models Report 2026 2028-30 base-case window for reliable, unattended generalist autonomy.
SU023 EmbodiedGlobal China Embodied AI H1 2026 Funding Tracker
SU024 VentureBeat China Embodied AI Startups Raise Billions 2026
SU025 Robotics and Automation News China Humanoid Robots Market Analysis 2026
SR001 Financial Times AI Startups China 2026
SR002 PanDaily Tech Giants Embodied Intelligence Battle Heats Up July 2026
SR003 New Atlas Humanoid Robot Industry 2026
SR004 Rimon Law China AI Law Brief
SR005 AI Governance China AI Regulation News 2026
SR006 ChinaCrunch China's AI Regulation 2026: Building a Global Framework
SR007 TechJournal US Imposes 25% Tariff on NVIDIA H200 AI Chips Bound for China
SR008 The Silicon Review China AI Export Controls Strategy 2026
SR009 Tianyancha (天眼查) Shanghai Bulage Technology Co., Ltd. — Corporate Registry
SR010 Brookings Institution Competing AI Strategies for the US and China
SR011 S&P Global Ratings Where Are AI Investment Risks Hiding?
SR012 Morgan Stanley AI Market Trends — Institute Research 2026
SR013 humanoid.guide Humanoid Foundation Models Report 2026
SR014 arXiv A Survey on Vision-Language-Action Models for Embodied AI
SR015 Forbes AI Startups With No Revenue Are Using This Tactic to Supersize Their Valuations
SR016 CNBC AI Valuation Fears Grip Investors as Tech Bubble Concerns Heighten
SR017 TechRadar World Models: The Key to More Capable AI — Financial Risks
SR018 ChinaBiz Insider 15 Embodied AI Unicorns in 6 Months: China's AI Risk Assessment
SR019 EqualOcean TARS AI Raises $455M Pre-A — Competitive Risk Analysis
SR020 Gasgoo TARS AI AWE 3.0 Market Position
SR021 BridgingChina Lin Junyang's New Company BLAG — Risk Context
SR022 TechCrunch Alibaba Qwen Tech Lead Steps Down
SR023 QbitAI Prague Technology — Corporate Overview
SR024 AInChina China Embodied AI Revolution 2026 — Risk Factors
SR025 PRNewswire China AI Robot Market 2026
SR026 Data-Gate Embodied AI Simulation and Real-World Robot Technical Economics
SR027 DeepMind (Google) Genie 2: Foundation World Model
SR028 EmbodiedGlobal China Embodied AI Funding H1 2026
SR029 Qicha (企查查) Shanghai Bulage Technology — Corporate Registry
SR030 CNTechPost Former Alibaba AI Core Figure Lin Junyang Founds New Lab
SR031 LeaveitAI Qwen 3.5 — AI Model Analysis and Review
SR032 US SEC EDGAR EDGAR Company Search
SV001 CB Insights State of Venture Q2 2026 — AI Trends Report Funding tops $200B for the second consecutive quarter. Deal count hits a decade low. Mega-rounds take 81% of all capital.
SV002 Dealroom Dealroom AI Startup Valuations H1 2026
SV003 Reuters China AI Startup Valuations Surge in 2026
SV004 McKinsey & Company The State of AI in 2026
SV005 Sequoia Capital AI Ascent IV — 2026 Summary Sonya Huang declared 2026 the year of agents, and walked through the three ingredients (models, tools, and harnesses) that have finally come together.
SV006 Bloomberg China AI Startups Valuation Boom
SV007 The Wall Street Journal China AI Startup Valuations — Skeptics Emerge
SV008 Andreessen Horowitz (a16z) AI Investments: The Next Wave
SV009 Barron's AI Startup Boom in China: Questions on Valuation Discipline
SV010 Airtable AI Company Valuations Dataset
SV011 TechCrunch Yann LeCun's AMI Labs Raises $1.03 Billion to Build World Models AMI Labs, the new venture co-founded by Turing Award winner Yann LeCun after he left Meta, has raised $1.03 billion at a $3.5 billion pre-money valuation.
SV012 The Observer Yann LeCun's AMI Labs Raises $1 Billion — Europe's Largest Seed Round The funding values AMI at $3.5 billion pre-money and includes an array of high-profile backers such as Nvidia, Mark Cuban, Eric Schmidt, and Jeff Bezos.
SV013 TechCrunch World Labs Lands $200M from Autodesk as Part of $1 Billion Round World Labs, which emerged from stealth in 2024 with $230 million at a $1 billion valuation, declined to say whether the latest round boosted its valuation. However, reports a month ago suggested it was aiming to raise at a $5 billion valuation.
SV014 Physical Intelligence Physical Intelligence Pi0 Blog — Company and Technology Overview
SV015 Financial Times China's AI Investment Surge: How Long Can It Last?
SV016 PitchBook Yann LeCun's AMI Labs Secures $1B in Bet on World Models
SV017 TechRadar AI Startup Valuations Reach New Highs in 2026
SV018 Morgan Stanley China AI Investment Outlook 2026
SV019 S&P Global China Venture Capital Market 2026 — AI Sector Report
SV020 QbitAI (量子位) 語用科技 Prague Technology — $2B Valuation Analysis
SV021 AI News China Lin Junyang Startup Prague Technology Raises $220M at $2B Valuation
SV022 PandaDaily Prague Technology Completes $220M Angel Round at $2B Valuation
SV023 BridgingChina China AI Startup Ecosystem Q2 2026 — Valuations and Funding Trends
SV024 DataGate China AI Private Market Data 2026 — Embodied Intelligence Sector
SV025 Humanoid Guide Top Embodied AI Companies and Valuations 2026
SV026 企查查 (QiChaCha) 上海卜拉格科技有限公司 — Enterprise Registration Record
SV027 ChinaBizInsider 15 Embodied AI Unicorns in 6 Months — China's Robot Race Hits a Reality Check
SV028 36Kr 林俊旸离开阿里创业 — Lin Junyang's Startup Raises at $2B
SV029 EmbodiedGlobal Global Embodied AI Market Report 2026 — China Chapter
SV030 CNTechPost Prague Technology Valued at $2B — Analysis of China's Newest AI Unicorn