初创公司尽调
尽调报告 AI Semiconductors / Neural Processing Units Private / Pre-IPO 2026-05-20

Rebellions

韩国 AI 芯片冠军:强背书与极端的收入估值缺口

Rebellions 在韩国 AI 芯片创业公司中战略位置最强,但当前 $2.34B 价格更多靠主权 AI 期权和投资人背书支撑,而不是公开披露的商业规模。

封面要素

最近一轮估值 01
2340 USD M [CV002]
累计融资 02
850 USD M [CV005]
最近披露收入 03
2.1 USD M (FY2023) [CV011]
首次生产部署 04
kt cloud (May 2023) [CO025]
IPO 路径 05
KOSPI preparation reported [CV008]
据报道的海外订单 06
U.S., Japan, Thailand [CU014]

公司概况

Rebellions 是一家总部位于 Seongnam 的无晶圆厂 AI 半导体公司,创立于 2020 年;2024 年 12 月与 SAPEON Korea 合并后,它成为后者的合并继承者,也拼出了韩国首个 AI 芯片独角兽。公司销售 AI 推理硬件和软件,产品覆盖第一代 ATOM 加速器、REBEL chiplet 平台、Rebel100/RebelCard,以及机架级 RebelRack/RebelPOD 系统;Samsung 是核心制造伙伴,Red Hat 是企业软件生态伙伴。截至 2026 年 3 月,Rebellions 已融资 $850M,正在进入日本、沙特阿拉伯和美国,但收入、客户数和利润率的公开披露仍很稀疏。

官网
rebellions.ai
成立时间
2020-01-01
创始人
Sunghyun Park
创立地点
Seongnam, Gyeonggi Province, South Korea
总部
Seongnam, Gyeonggi Province, South Korea
产品
AI 推理加速器和系统:用于实时数据中心部署的 ATOM 与 ATOM-Max 产品;REBEL chiplet、Rebel100 平台、RebelCard 加速卡、RebelServer,以及 RebelRack/RebelPOD 机架级系统;另有 Rebellions SDK 和 OpenShift AI 集成层。
客户
电信运营商、云 / 数据中心运营商、主权 AI 基础设施项目,以及采用大模型推理系统的企业。
商业模式
硬件 + 软件销售模式,把推理芯片、加速卡、服务器、机架级系统和部署软件打包销售;采用往往由合作伙伴和客户定制验证项目推动。
阶段
Private / Pre-IPO
融资情况
2026 年 3 月完成 $400M pre-IPO 轮,投后估值约 $2.34B;累计披露融资为 $850M。
[CO001, CO002, CO017, CO021, CO022, CO025, CO029, CO030]

执行摘要

主要优势

  • 韩国资本最充足的 AI 芯片平台,累计融资 $850M,并获得 Arm、Samsung、Mirae 和 SK 相关投资人支持
  • 以推理为核心的产品路线图可信,从 ATOM、REBEL 延伸到机架级 RebelRack/RebelPOD 系统
  • kt cloud 和 SK Telecom 相邻部署提供真实生产证据,另有来自沙特、日本和美国的订单报道
  • Samsung 代工与封装关系,加上 Red Hat OpenShift AI 集成,提升商业化可信度
  • 与韩国主权 AI 议程高度同频,强化获取伙伴和资本的能力

主要风险

  • 估值与已披露收入差距极大:最后一个公开收入数字仅为 FY2023 约 $2.1M
  • 缺少 FY2024-FY2025 审计财务、客户数披露和留存指标,无法验证商业规模
  • 与 Nvidia 的基准差距仍然关键:Rebellions 仍缺少 MLPerf 提交等标准化公开证据
  • 对 Samsung 代工、封装和先进内存生态的单一栈依赖,造成供应链集中
  • IPO 时点和披露风险高,因为韩国上市会迫使公司交出目前尚未公开的财务透明度
  • 客户集中度看起来偏高,最清晰证据集中在投资人伙伴和主权 / 电信账户

未决问题

  • FY2024 和 FY2025 审计收入、毛利率和现金消耗报表
  • 具名头部客户名单、前三大客户收入占比,以及合同期限 / 续约数据
  • REBEL 相比 Nvidia H100/H200/B100 和 AMD MI300X 的标准化基准证据
  • RebelRack 和 RebelPOD 已确认的生产交付及收入贡献
  • IPO 承销商阵容、上市时间表,以及 pre-IPO 轮带来的任何优先权 / 稀释压力

目录

Chapter 01

01公司概况

1.1 身份、总部、创立与产品模式

Rebellions 是一家无晶圆厂 AI 半导体公司,总部位于韩国 Gyeonggi Province Seongnam(Bundang district)。公司由五名韩国工程师在 2020 年共同创立,CEO Sunghyun Park 是公开资料中确认的领导者。公司正式注册地址为 3F, 6 Jeongjail-ro 156beon-gil, Bundang-gu, Seongnam-si, Gyeonggi-do, Korea。公司只做 AI 推理——把训练好的神经网络放到生产规模运行——而不进入 Nvidia 和 AMD 主导的训练市场。 Rebellions 自称在打造「面向特定用途的 AI 加速器,以重塑大规模 AI 时代的能效和扩展能力」。它的 chiplet、互连和软件栈面向超大规模部署设计。公司向云运营商、电信运营商和主权 AI 项目销售硬件(NPU 芯片、服务器系统)及配套软件(Rebellions SDK)。商业模式是硬件 + 软件销售:芯片由 Samsung 以先进制程代工,组装成服务器平台(ATOM-Max Server、RebelServer、RebelRack、RebelPOD),并随 SDK 一起用于模型部署。 2024 年 12 月合并后,法律上的存续主体是 SK Telecom 子公司 SAPEON Korea,但公司完全以 Rebellions 名称和领导层运营。因此,SK Telecom、SK Square 和 SK Hynix 通过其 SAPEON 持股成为合并后实体的战略投资方。截至 2026 年 5 月,Rebellions 已在日本(2025 年 2 月设立)和沙特阿拉伯(2025 年 8 月设立)运营子公司,并设立由首席商务官 Marshall Choy 领导的美国实体,以支撑北美扩张。 公司的产品路径从 ATOM 开始:第一代基于 GDDR6 的推理加速器,2022 年 6 月流片,2023 年 5 月首次出货;随后是 REBEL,一款采用 HBM3E 内存的 UCIe-Advanced AI chiplet,2024 年 11 月流片,公司称其为全球首款同类产品;再到当前的 Rebel100 平台、RebelCard 加速卡,以及新推出的 RebelRack 和 RebelPOD 集成机架级系统。2025 年 12 月与 Red Hat OpenShift AI 的软件集成,把可触达的企业客户基础从亚洲扩展出去。 [CO001, CO002, CO003, CO009, CO024, CO025]

KPI 快照表
指标数值 / 状态日期置信度缺口
成立年份20202020
总部韩国京畿道城南(盆唐)2026-05-20
法定运营实体SAPEON Korea(合并后更名为 Rebellions)2024-12-02
当前阶段Pre-IPO2026-03-30IPO 交易所、申报日期和价格区间尚未公布
最新投后估值(USD M)23402026-03-30公司披露的近似值;未经独立审计
累计融资(USD M)8502026-03-30
最近一轮融资Mirae Asset 和 Korea National Growth Fund 领投的 Pre-IPO $400M 融资2026-03-30
Series C 估值(USD B)1.42025-09-30
员工数未披露2026-05-20私营公司;无公开员工数
收入 / ARR(USD M)未披露2026-05-20私营公司;无公开财务数据
核心产品ATOM(第 1 代)、REBEL(第 2 代)、Rebel100 / RebelCard / RebelRack / RebelPOD2026-03-30
关键制造伙伴Samsung Electronics(ATOM 使用 5nm,REBEL 使用 4nm)2024-10-05晶圆厂产能分配未确认

估值和累计融资来自 Rebellions 官方 Pre-IPO 新闻稿(2026 年 3 月 30 日),并由 CNBC 和 PRNewswire 佐证。员工数和收入未披露;单元格写「未披露」而不是 null,以区别于零值。Series C 估值来自 2025 年 9 月 30 日官方新闻稿。

[CO001, CO002, CO020, CO021, CO022, CO024]
FO002: 公司快照 — 身份、产品、资本和合作关系

Rebellions 以半导体设计为核心,依托 Samsung 产能,部署到云和电信客户;资本结构由战略投资者和 财务投资者两层支撑。

[CO003, CO009, CO024, CO026, CO031, CO032]

1.2 创始人、领导层、关键人物风险与治理

CEO Sunghyun Park(截至 2025 年 4 月 Forbes 画像为 40 岁)是公司最主要的公开人物,也是被引用最多的创始人。Park 拥有 MIT Computer Science and Artificial Intelligence Laboratory(CSAIL)的硕士和博士学位;在创办 Rebellions 前,他先后在 Intel(芯片工程)、SpaceX(系统工程)和 Morgan Stanley(金融行业)工作。他带领 Rebellions 完成五轮融资,并亲自谈判 SAPEON 合并。深科技硬件创始人通常难以同时具备技术深度和资本市场经验,Park 的履历正是公司关键人物集中风险的主要来源。 本章审阅的公开资料没有逐一披露另外四名联合创始人的姓名;这是一个尽调缺口,后续章节应当标出,因为更广泛创始团队的治理和继任安排仍未知。2024 年 12 月合并后,Park 被确认为合并后实体的 CEO。2025 年 11 月,公司任命 Marshall Choy 为首席商务官,常驻硅谷;他拥有二十多年企业 AI 和系统经验,负责 Rebellions 新设美国实体和全球商业化拓展。 公司仍是私人公司,治理透明度有限。合并后的董事会构成没有公开披露。SK Telecom 具备战略投资方身份,并可能拥有董事会席位,但公开材料没有确认确切董事会构成和投票权。这是重要尽调缺口:如果 SKT 拥有董事会控制权或重大否决权,Rebellions 在韩国以外全球市场的战略决策独立性就会受到影响。合并条款规定,SKT、SK Square 和 SK Hynix 在交割前出售其 SAPEON 持股的 3%,以确保 Rebellions 管理层成为合并后实体的多数股东;这部分缓解了治理被俘获的风险,但剩余持股分布没有披露。 [CO004, CO005, CO006, CO007, CO008, CO009]

领导层与创始人表
人物角色背景创始人-市场匹配 / 职能覆盖关键人物依赖
Park Sung-hyun (Sunghyun Park)CEO 兼联合创始人MIT 博士(CSAIL);Intel(芯片工程)、SpaceX(系统)、Morgan Stanley(金融)深厚半导体设计经验 + 资本市场经验;主导全部五轮融资关键——公开资料中唯一具名、推动战略、融资和 IPO 准备的高管
Marshall Choy首席商务官(2025 年 11 月任命)20+ 年硅谷企业 AI 与系统经验;负责新设美国实体负责全球商业扩张和 GTM;补上美国销售负责人缺口高——美国收入取决于 Choy 的人脉和执行
[四位联合创始人——公开来源未具名]角色未知审阅的公开来源均未披露未知;Rebellions 由五名韩国工程师共同创立未知——创始团队治理和接班计划未见文件披露

审阅的公开来源中,只有 Park Sung-hyun 和 Marshall Choy 具名。其他四位联合创始人仅在 Korea Times 和 Korea JoongAng Daily 中被称为「五名韩国工程师」。合并后的董事会构成未公开披露。本表只列出部分人员;实际领导层深度很可能比这里记录的更广。

[CO004, CO005, CO006, CO007, CO008, CO040]

1.3 融资历史、投资方与估值轨迹

截至 2026 年 3 月 30 日,Rebellions 已累计融资 $850 million,在韩国 AI 半导体创业公司中遥遥领先。资本轨迹在截至 2026 年 3 月的六个月内急剧加速:公司仅在这一窗口就融资 $650 million(超过累计融资的 75%),说明计划 IPO 前投资人信心很强。 融资历史早于 Series B。公司 2024 年 1 月 Series B 新闻稿披露,当时累计融资「超过 $200 million」,意味着 Series B 之前的种子轮和 Series A 约为 $76 million;这些更早轮次没有在公开来源中单独公告。Series B 于 2024 年 1 月 30 日完成,金额 $124 million,由 KT(韩国主要数据中心运营商)领投,Temasek 旗下 Pavilion Capital(跟投)、KDB(Korea Development Bank)、Mirae Asset Venture Investment、IMM Investment、KT Investment、SV Investment、Korelya Capital(法国)、DGDV(日本)和多家韩国机构投资方参与。2024 年 7 月 23 日,公司又完成 $15 million 的 Series B 延展轮,全部来自 Saudi Aramco 旗下风险投资机构 Wa'ed Ventures——这是该基金首次投资韩国创业公司。 2025 年 9 月 30 日的 Series C 融资 $250 million,投后估值 $1.4 billion;Arm 成为战略财务伙伴,Samsung Ventures、Pegatron VC、KDB(跟投)和 Korelya Capital(跟投)也参与。2025 年 11 月,Kindred Ventures 和 Top Tier Capital Partners 加入 Series C 延展轮,硅谷风险资本首次进入股权结构表。 2026 年 3 月 30 日,Mirae Asset Financial Group 和 Korea National Growth Fund 领投 $400 million pre-IPO 轮,投后估值约 $2.34 billion。本轮融资确认,2024 年 12 月合并后达到的独角兽估值里程碑(合并公司企业价值超过 1 trillion Korean won)到 pre-IPO 时已大幅抬升。截至 2026 年 5 月,IPO 交易所、申报日期和价格区间均未公开确认。 [CO011, CO012, CO013, CO014, CO018, CO019]

利益相关方或投资人图谱
利益相关方角色控制权或经济重要性尽调问题
Mirae Asset Financial GroupPre-IPO 领投方(2026 年 3 月)韩国最大资管公司;锚定 Pre-IPO 轮,显示强机构信心确认董事席位、锁定期条款和 IPO 承销关系
Korea National Growth FundPre-IPO 联合领投方(2026 年 3 月)政府支持的战略资本;让 Rebellions 与韩国 AI 主权议程对齐明确政府基金投资附带的任何条件或战略义务
SK Telecom战略投资人(经 SAPEON 合并进入);主要商业客户在 SAPEON 原股东中持有合并后最大股份;ATOM-Max 首个商业 NPU 客户确认确切持股、董事会代表和商业采购量承诺
SK HynixSAPEON 股东 → Rebellions 股东;HBM3E 内存供应商为 REBEL chiplet 提供 HBM3E 内存;战略供应链对齐确认供货协议条款、优惠定价和排他性条款
SK SquareSAPEON 股东 → Rebellions 股东SK Group 金融部门;合并后实体少数股东明确当前持股,以及 Pre-IPO 前是否在减持
Samsung Electronics晶圆代工制造伙伴;Samsung Ventures 投资人(Series C)关键:唯一确认在量产规模支持 ATOM(5nm)和 REBEL(4nm)的晶圆厂伙伴确认产能分配、晶圆厂优先级和晶圆定价条款
Arm战略伙伴和 Series C 投资人AGI CPU 由 Arm 与 Rebellions 联合开发,用于主权 AI 服务器;存在 IP 授权关系确认授权范围、如有排他性则确认条款,以及 AGI CPU 交付时间表
KT / kt cloud 客户Series B 领投方;最早上线客户首个在真实数据中心部署 ATOM 的外部客户;KT 关联投资人在 Series B 中参投评估持续 ATOM 部署规模和收入贡献
Wa'ed Ventures(Saudi Aramco)投资方Series B 延展轮投资人($15M)打开沙特阿拉伯市场;触发沙特子公司设立(2025 年 8 月)明确沙特阿拉伯商业进展和任何合同管线
Korea Development Bank (KDB)Series B 和 Series C 跟投方国家政策银行;多轮参与显示与政府方向一致的战略兴趣确认 KDB 投资是否附带任何本土生产条件

所有投资人的持股比例(%)均未公开披露。利益相关方重要性评级根据轮次规模、战略角色和公开公告推断。SKT、SK Hynix 和 SK Square 经 SAPEON 合并继承 Rebellions 股份;其合并后确切比例未公开。Pre-IPO 投资人的锁定期条款未披露。

[CO011, CO012, CO013, CO014, CO015, CO018]
FO003: 截至 2026 年 5 月的 KPI 快照

截至 2026 年 5 月,Rebellions 累计融资 $850M,IPO 前估值 $2.34B,并拥有三个活跃地域子公司; 但收入和员工数均未公开披露。

所有财务数字来自 Rebellions 官方新闻稿。2026 年 3 月新闻稿将估值表述为「约 $2.34B」。 收入和员工数确实未披露;使用「未披露」而不是 null,以区别于零值。

[CO020, CO021, CO022, CO029, CO030]

1.4 产品演进、合作伙伴与公司里程碑

Rebellions 的里程碑时间线,是后续章节可以直接复用、无需反复论证的事实主轴。公司在 2020 年创立时押注:AI 推理——在生产规模运行训练好的模型——会从训练中分化成独立硬件市场;能效更高、面向特定用途的 NPU 可以在推理负载上胜过 Nvidia 的通用 GPU。ATOM 的商业成功和 2024 年以来快速累积的资本,已经验证了这一判断。 关键产品里程碑:ATOM 于 2022 年 6 月流片(Samsung 基于 GDDR6 的制程);2023 年 5 月首次出货给 kt cloud,成为首款投入实时数据中心使用的韩国自研推理芯片;目标是在 2024 年上半年采用 Samsung 5nm 量产。Samsung Electronics 于 2023 年 10 月成为正式共同开发伙伴,承诺以 Samsung 4nm 制程并集成 HBM3E 生产 REBEL。REBEL SoC 于 2024 年 11 月流片,被称为全球首款采用 144GB HBM3E 的 UCIe-Advanced AI chiplet。REBEL-Quad 在 Hot Chips 2025(2025 年 8 月)亮相,目标是 Blackwell 级性能。Rebel100 平台、RebelCard 加速器、RebelRack 和 RebelPOD 随 2026 年 3 月 pre-IPO 轮一同发布,标志着 Rebellions 从只卖芯片转向集成 AI 基础设施系统。 关键合作里程碑:SKT-Arm-Rebellions MOU(2026 年 4 月)约定共同开发主权 AI 推理基础设施,把 Arm 的 AGI CPU 与 Rebellions 的 RebelCard 结合,先在 SKT 数据中心验证,再推向全球商业化。Red Hat 合作(2025 年 12 月)围绕由 Rebellions NPU 驱动的 Red Hat OpenShift AI 展开。NTT DOCOMO Innovations(DOCOMO)合作(2025 年 4 月)用于 AI 加速验证。SAPEON 合并公告(2024 年 6 月)、最终协议(2024 年 8 月)和交割(2024 年 12 月)共同造就韩国首个 AI 芯片独角兽。日本子公司(2025 年 2 月)和沙特阿拉伯子公司(2025 年 8 月)构成区域落点。 SKT 正在核心 AI 服务中测试配备 Rebellions NPU 的服务器,包括 A.Dot 通话摘要、PASS 垃圾信息过滤、PASS 金融助手和 X Caliber 服务。这是 ATOM 在生产规模获得商业验证的主要证据。kt cloud 是最早投入使用的客户。Wa'ed Ventures 的投资打开了沙特阿拉伯市场。根据 REBEL-Quad Hot Chips 演示(2025 年 8 月),ATOM 也已在日本和沙特阿拉伯部署,但部署规模没有公开量化。 [CO015, CO016, CO024, CO025, CO026, CO027]

里程碑表
日期事件类型金额 / 估值 / 状态参与方含义
2020公司由五名韩国工程师创立创立私营公司;Series A 前累计融资未披露Sunghyun Park + 四名未具名联合创始人确立只做推理的 NPU 策略;早期 Samsung 和 KT 关系启动
2022-06ATOM SoC 基于 Samsung GDDR6 工艺流片产品第一代推理加速器Samsung Foundry韩国首款进入流片阶段的 AI 推理芯片;确认与 Samsung 的晶圆代工关系
2023-05ATOM 首次发货给 kt cloud(真实数据中心部署)规模化商业部署kt cloud(KT 子公司)首款韩国开发的推理芯片投入生产;KT 成为锚定客户
2023-10Samsung Electronics 战略合作——REBEL 采用 4nm 并配 HBM3E合作联合开发 MOURebellions + Samsung Electronics下一代产品路线落定;Samsung 的资本与制造绑定加深
2024-01-30Series B 完成 $124M;累计融资超过 $200M融资$124M / 累计 >$200MKT(领投)、Pavilion Capital(Temasek)、KDB、Mirae Asset、IMM、KT Investment、Korelya Capital、DGDV当时韩国半导体初创公司最大融资;全球投资人基础建立
2024-07-23Wa'ed Ventures(Saudi Aramco 子公司)投入 Series B 延展轮 $15M融资$15MWa'ed Ventures进入沙特阿拉伯市场;Wa'ed 首次投资韩国初创公司;沙特子公司管线打开
2024-08-18与 SAPEON Korea 签署最终合并协议;股权比例 2.4:1(Rebellions:SAPEON)治理股权比例 2.4:1参与方:Rebellions + SAPEON Korea + SK Telecom韩国 AI 芯片行业整合;合并后实体将以 Park 领导下的 Rebellions 命名
2024-11REBEL SoC 流片——全球首款 UCIe-Advanced AI chiplet,配 144GB HBM3E产品第二代旗舰Samsung FoundryChiplet 架构突破;为 petascale 推理平台铺路
2024-12-02与 SAPEON Korea 合并完成;形成韩国首个 AI 芯片独角兽治理>1 trillion KRW 企业价值参与方:Rebellions + SKT + SK Hynix + SK Square达成独角兽里程碑;人才和 IP 合并沉淀;SKT 成为战略投资人
2025-02Rebellions 日本子公司成立规模化首个海外分支Rebellions进入 APAC 企业市场;日本 AI 数据中心合作纳入范围
2025-04与 SK Telecom 和 DOCOMO Innovations 签署 AI 加速 MOU合作MOURebellions + SK Telecom + DOCOMO Innovations(NTT DOCOMO 子公司)电信 AI 基础设施获得验证;ATOM 服务器在 SKT NPU farm 测试
2025-08-27REBEL-Quad 在 Hot Chips 2025 发布;沙特阿拉伯子公司成立产品下一代 chiplet 系统;目标 Blackwell 级性能披露方:Rebellions(Hot Chips symposium,Palo Alto)首次在西方公开技术披露;Aramco 支持的沙特落点正式建立
2025-09-30Series C 完成 $250M;估值 $1.4B;Arm 成为战略伙伴融资$250M / $1.4B 投后Arm(新战略投资人)、Samsung Ventures、Pegatron VC、KDB(跟投)、Korelya Capital(跟投)新一轮确认独角兽估值;Arm 合作巩固 CPU+NPU 路线图
2025-11Series C 延展轮;Kindred Ventures 和 Top Tier Capital Partners 加入;Marshall Choy 出任 CBO融资延展轮(金额未披露)投资方:Kindred Ventures + Top Tier Capital Partners硅谷 VC 进入;公司聘请美国负责人推动北美扩张
2025-12-11与 Red Hat OpenShift AI 合作,在 Rebellions NPU 上跑企业 AI合作联合产品发布Rebellions + Red Hat企业 AI 分发路径打开;开源生态集成里程碑
2026-03-30Pre-IPO $400M 完成,估值 $2.34B;RebelRack 和 RebelPOD 发布融资$400M / 投后约 $2.34B;累计 $850MMirae Asset Financial Group(领投)+ Korea National Growth FundIPO 准备得到确认;六个月融资 $650M;集成式机架级系统进入市场

日期来自 Rebellions 官方新闻稿和 Rebellions 关于页面时间线。关于页面时间线确认 2022-06 流片和 2023-05 首次发货。Samsung 合作日期来自 2023 年 10 月新闻稿。合并日期来自 2024 年 8 月协议和 2024 年 12 月完成合并新闻稿。Series C 和 Pre-IPO 日期来自 2025 年 9 月和 2026 年 3 月官方新闻稿。DOCOMO MOU 日期来自 2025 年 4 月新闻稿。

[CO001, CO011, CO012, CO014, CO015, CO016]
FO001: 公司里程碑时间线

Rebellions 从 2020 年创立走到 IPO 前阶段,六年累计融资 $850M;产品创新(ATOM→REBEL→Rebel100)、 公司整合(SAPEON 合并)和地域扩张三条线并行推进。

[CO001, CO011, CO015, CO016, CO018, CO020]

1.5 反向因素、竞争与证据缺口

根据 The Investor 在 SAPEON 合并公告时的报道,2023 年 Nvidia 约占 AI 芯片市场 94% 份额。这个竞争现实是最核心的反向因素:每一次产品决策、客户赢单和融资,都必须放在一个根基极深的市场领导者面前评估;后者在软件生态(CUDA)、规模和品牌认知上都占优。Rebellions 只做推理、刻意避开训练市场,这既是聚焦式差异化,也是在押注推理负载会足够大、足够独立,让专业厂商拿到有意义份额。 公司没有公开发布收入、年经常性收入(ARR)、毛利率或客户数。披露只停留在定性部署状态(SKT 测试、kt cloud 上线、进入沙特阿拉伯和日本市场)。仅靠公开来源无法做财务尽调,这是本报告最重要的证据缺口。没有披露财务数据符合私人公司常态,但考虑到公司已表示准备 IPO,这一点尤其值得关注——投资人应预期 IPO 关闭前会看到正式申报文件或招股书。 合并后的整合风险不小。把 Rebellions 和 SAPEON Korea 两支芯片开发团队、各自工程文化和重叠投资方放进同一公司结构,是复杂的组织挑战。法律存续主体是 SAPEON Korea(重组为 Rebellions),这可能在 IP 所有权、团队留任和文化磨合上带来摩擦。公司没有披露合并后的任何人员流失数据。 Park Sung-hyun 的关键人物集中,是首要领导层风险。公开记录中的每一轮融资和谈判,Park 都是核心人物;计划 IPO 能否推进,取决于他能持续投入。董事会构成和继任计划没有公开文件记录。治理缺口叠加收入披露缺失、IPO 时间表不明确,构成本章为后续分析标出的剩余尽调负担。 [CO046, CO047, CO050, CO051, CO052, CO053]

1.6 证据要点

Chapter 02

02市场分析

2.1 AI 推理芯片市场——边界、相邻市场与现状替代品

本报告分析的市场是 AI 推理加速芯片市场:面向特定用途的芯片——NPU、ASIC 和针对推理优化的 GPU——部署在数据中心和边缘环境中,用于在生产规模运行训练好的神经网络。纳入支出的范围包括芯片采购、配套服务器平台和推理专用软件授权。排除在市场边界外的是:(a)未使用 AI 加速的通用 CPU 推理,(b)以训练为主的加速器(仅用于训练部署的 Nvidia H100/H200),(c)AI 网络芯片(NVLink、InfiniBand),以及(d)面向消费者的边缘推理芯片(智能手机 NPU)。相邻市场——AI 训练加速器——目前是主要支出类别,但 Rebellions 明确不参与,因此单独跟踪。 Nvidia 的 Compute & Networking 分部收入,是全球 AI 加速计算市场最好的可用代理指标。FY2026(截至 2026 年 1 月 25 日的年度),Nvidia 披露 Compute & Networking 收入 $193.5 billion,较 FY2025 的 $115.9 billion 增长 67%。Nvidia 总收入为 $215.9 billion(同比 +65%)。尽管这些数字横跨训练和推理负载,行业共识和 Nvidia 自身产品路线图都表明,推理是增长最快的组成部分:面向智能体 AI 的 Blackwell Ultra,以及 Rubin 后续平台(预计 H2 FY2027 投产),都瞄准推理效率和 token 吞吐。专用推理芯片的现状替代品,是在多用途数据中心 GPU 上做基于 GPU 的推理——2026 年仍是主流部署方式——Rebellions 必须在 TCO 和能效上将其替代。第二个替代品是面向低吞吐负载的 CPU-only 推理,在小 batch size 下仍有竞争力。CY2023 全球 AI 推理芯片市场约为 $34.3 billion(主要是 Nvidia GPU 收入),当时 Nvidia 约占 94% 市场份额。此后市场大幅扩张:Nvidia FY2025 C&N 收入 $115.9 billion,意味着 CY2024 总加速计算约 $120–130 billion,仍由 Nvidia 主导,但 AMD MI300X 和超大规模云厂商自研 ASIC(Google TPU、AWS Trainium/Inferentia)正在拿到更大比例。 [CM001, CM002, CM003, CM004, CM006, CM008]

市场定义表
细分市场 / 类别纳入支出排除支出买方 / 付款方对 Rebellions 的意义
AI 推理芯片(NPU、推理优化 GPU、自定义 ASIC)NPU/ASIC 硬件、推理服务器平台、推理 SDK 授权仅训练计算、通用 CPU 推理、消费级移动 NPUCSP、电信运营商、企业数据中心、主权 AI 项目核心目标市场——Rebellions ATOM、REBEL、REBELRACE 都是专为推理打造的 NPU
AI 训练加速器(Nvidia H100/H200、AMD MI300X、Google TPU)用于模型训练的 GPU/TPU 硬件、训练云实例仅推理部署(即便同一硬件也可能被使用)AI 模型开发者、超大规模云厂商(Google、Amazon、Microsoft、Meta)相邻市场——Rebellions 明确不在这里竞争;不计入 SAM
AI 网络和互连(NVLink、InfiniBand、自定义 fabric)GPU 集群高带宽网络、交换芯片、DPU计算芯片;网络是支撑基础设施,不是 Rebellions 可服务市场与计算相同的 CSP 买方;通常与 Nvidia GPU 系统一起打包排除——Rebellions 的 NPU 产品线无法服务
消费级和移动边缘推理芯片(智能手机 NPU、IoT)嵌入移动设备和边缘摄像头的 SoC 推理引擎数据中心推理芯片;设计点不同(功耗、形态)消费电子 OEM(Apple、Qualcomm、Samsung Mobile)排除——Rebellions 面向数据中心规模工作负载,不是移动边缘
韩国主权 AI 基础设施(SKT、KT Cloud、Samsung SDS)本土 AI 芯片采购、云推理容量、主权 AI 计算不符合主权 AI 条件的全球来源 GPU 计算韩国科学技术信息通信部、SKT、KT、Samsung SDS、NAVER Cloud最高优先级 SAM 细分市场——Rebellions 拥有正式 SKT 合作和 Samsung 支持

细分市场和边界基于 Nvidia FY2026 10-K 的产品分部描述,以及 Rebellions 对 AI 推理的产品聚焦来界定。训练和网络分部相邻,但不属于可服务市场。

[CM001, CM003, CM006]
TAM/SAM/SOM 或规模测算视角表
发布方 / 来源年份 / 时期地域数值CAGR / 增长方法置信度局限
Nvidia 10-K(SEC 文件,FY2026)FY2026(截至 2026 年 1 月 25 日的年度)全球$193.5B 计算与网络收入+67% 同比披露的分部收入;可作为 AI 加速计算市场底线代理高——主要监管文件包含训练、汽车和网络;高估纯推理市场
Nvidia 10-K(SEC 文件,FY2025 推导)FY2025(截至 2025 年 1 月 26 日的年度)全球$130.5B 总收入;约 $115.9B C&N(由 FY2026 +67% 基数估算)总收入 +65% 同比(FY2025 vs FY2024)披露的总收入;C&N 由 FY2026 披露反推估算总收入置信度高;C&N 置信度中(推导)C&N 数字是估算值,FY2025 未直接披露
The Investor / Korea Herald(CY2023 市场报告)2023 日历年全球约 $34.3B AI 芯片市场未说明;鉴于 FY2024/2025 轨迹,假设快速增长第三方市场报告引用分析师数据;称 Nvidia 市占率 94%中——二级市场报告;来源方法不清楚相对 2025-2026 数据已过时;由于 AMD/ASIC 增长,Nvidia 当前份额可能更低
基于 Nvidia 数据推断(CY2025 研究员估算)CY2025(Rebellions 研究估算)全球~$200-230B AI 加速计算(所有厂商:Nvidia + AMD + 自定义 ASIC)基于 Nvidia 数据推导,2023-2025 隐含 ~40-50% CAGR由 Nvidia FY2026 C&N + AMD MI300X 估算 + 超大规模云厂商自定义 ASIC 收入估算推导低——分析师估算,不确定性大自定义 ASIC 收入(Google TPU、AWS Inferentia/Trainium、Maia)未披露
Rebellions SOM 代理(研究员前瞻估算)2026-2028(前瞻估算)韩国 + 亚太电信运营商未公开披露;韩国主权 AI 计算 TAM 在数亿美元区间公开文件未说明基于 SKT 数据中心规模、MSIT AI 项目、NTT Docomo 合作低——无公开合同金额;$400M+ 投资意味着投资人对收入有信心收入确认时点和政府合同敲定情况未在公开来源确认

市场规模测算锚定 Nvidia FY2026 10-K 披露数字(SM001)。第三方分析师估算(Gartner、IDC、Grand View Research)在研究期内无法访问。CY2025 SAM 和 SOM 估算由研究员推导,不确定性高。

[CM001, CM002, CM003, CM004]
FM001: 市场规模测算视角

Rebellions AI 推理芯片市场机会的 TAM/SAM/SOM 金字塔。TAM 用 Nvidia FY2026 C&N 收入作为 可观测下限;SAM 是其中推理相关且符合主权采购资格的子集;SOM 是 Rebellions 未来 3 年有可信机会拿下的 韩国和亚太电信 / 主权云市场。

[CM001, CM006, CM026]
FM002: 市场估计区间

以 Nvidia Compute & Networking 收入代理全球 AI 加速算力市场,给出 FY2025(CY2024)、FY2026(CY2025) 和分析师 FY2027 预测的低 / 基准 / 高估计区间。所有数值为 $B USD。FY2025 和 FY2026 为已报告实际值; FY2027 为推断估计。

FY2025 C&N 是派生数字(FY2026 ÷ 增长率),并非直接披露。FY2027 区间是场景估计,不确定性很高。 韩国 SOM 是一阶估计;没有公开采购合同金额可用。

[CM001, CM004, CM024, CM034]

2.2 买方、用户与付款方分层——预算归属与采用路径

AI 推理芯片市场可按四类主要买方划分,它们的采购模式、预算权和 AI 芯片采用触发点各不相同。 超大规模云厂商(CSP——AWS、Azure、GCP,加上新兴 AI 原生云)代表最大采购量。它们拥有或租赁计算基础设施,每年向 GPU 和 ASIC 供应商投入数十亿美元 capex。Nvidia 明确将 CSP 视为其主要数据中心客户,所有主要云服务商都在使用其平台。超大规模云厂商采购由企业 API 客户和消费者 AI 产品带来的模型推理需求驱动。预算权在基础设施 / 平台工程部门。替代芯片的采用触发点,是规模化场景下的 TCO 和能效。 电信运营商(SK Telecom、KT、NTT Docomo 及国际同类公司)构成 Rebellions 的战略重要客群。电信运营商正在建设 AI 原生网络,并在边缘和区域数据中心部署推理。SK Telecom 与 Rebellions 和 Arm 合作开发主权 AI 推理芯片,建立了正式的电信采购路径。NTT Docomo Innovations 加入 Rebellions/SKT 基础设施合作,说明亚洲电信运营商兴趣正在扩大。电信 AI 芯片预算通常由网络技术和基础设施采购管理;在韩国,主权 AI 政策又创造出一个与政策挂钩的预算类别。 企业买方(银行、医疗、政府机构、制造业)通常通过云 API 或本地服务器采购间接采用 AI 推理芯片。Red Hat OpenShift AI 认证 Rebellions NPU,打开了企业本地推理部署渠道。Samsung Electronics 的战略投资和共同开发合作,为 Samsung 关联公司成为企业推理买方创造路径。 韩国本土云服务商(Samsung SDS、KT Cloud、NAVER Cloud)是重要的可服务市场(SAM)细分。IDC 数据确认,Samsung SDS 是韩国第一大托管云服务商(MSP 份额 23.9%)和第二大本土 CSP(份额 11.0%),使 Samsung Group 成为韩国主导性云买方;而 Samsung 对 Rebellions 的投资,又创造出潜在的锁定采购关系。 [CM009, CM010, CM011, CM012, CM015, CM016]

细分市场 / 买方图谱
细分市场买方 / 决策者用户付款方AI 推理工作流预算负责人替代芯片采用触发因素
超大规模 CSP(AWS、Azure、GCP、Alibaba Cloud)基础设施 VP / 芯片工程负责人AI 应用与 API 团队;终端用户经由 AI 服务使用CSP 资本开支预算;通过 API 定价向下游传导LLM 推理、图像生成、推荐系统、搜索 AI基础设施平台管理层,需董事会层面批准资本开支规模化后 TCO 优势 >20%;受限数据中心里的功率密度优势;供应链多元化
电信运营商(SKT、KT、NTT Docomo、Deutsche Telekom)网络技术 CTO / AI 基础设施负责人内部企业 AI 服务;B2B AI 即服务租户网络资本开支 / 主权 AI 项目预算(韩国 / 日本政策挂钩)AI 原生 5G 网络功能、面向客户的 AI 服务、网络智能电信运营商 CTO / Ministry of Science and ICT(主权 AI 赛道)主权 AI 政策要求;国产芯片偏好;SKT-Rebellions 战略合作
企业数据中心(银行、医疗、制造)IT 基础设施负责人 / 首席 AI 官内部 ML/AI 团队,部署模型推理流水线IT 资本开支 / AI 转型预算文档 AI、欺诈检测、临床 NLP、预测性维护CIO 或专门的 AI 基础设施团队本地部署的数据主权;Red Hat OpenShift AI 认证降低集成风险
韩国及亚洲本土云厂商(Samsung SDS、KT Cloud、NAVER Cloud)云基础设施 VP / Samsung Group 战略协同韩国企业租户、政府机构云平台资本开支;政府 AI 基础设施补助云端推理 API 服务;主权云 AI 工作负载Samsung Group 战略方向;韩国政府采购政策Samsung Electronics 投资 Rebellions;Samsung SDS 是天然客户;MSIT 政策

买方细分根据 Nvidia FY2026 10-K 客户描述、SKT/Rebellions/Arm 合作公告,以及 IDC 韩国云市场数据推断。预算数字为估算值;没有可用的一手采购数据。

[CM009, CM010, CM011, CM012, CM015, CM016]
FM003: 买方 / 细分市场地图

买方细分矩阵展示四类主要 AI 推理芯片买方,并按交易规模、采购速度、竞争强度和 Rebellions 当前定位打分。

[CM009, CM010, CM012, CM015, CM016, CM030]

2.3 增长驱动与采用约束——时点、预算和估值含义

AI 推理市场最主要的驱动因素,是超大规模云厂商的资本开支承诺。Nvidia FY2026 C&N 收入同比增长 67%,确认云基础设施建设是市场的定义性引擎。训练负载启动了 GPU 支出浪潮,但随着生成式 AI 从开发走向生产部署,面向数十亿终端用户运行 AI 模型的推理,正在推动下一段 capex。Epoch AI 研究显示,2009 年至 2022 年,ML 训练成本每年增长 0.49 个数量级;随着模型规模扩大,市场对更高效推理芯片的需求被推高。 第二个结构性驱动是地缘政治供应风险。根据 Nvidia 自身 10-K 文件,美国出口管制到 FY2026 已实质上让 Nvidia 退出中国数据中心计算市场。这会在重视 AI 供应链独立性的市场中创造替代芯片供应商需求,包括韩国和更广泛的亚太市场。韩国由 Ministry of Science and ICT 牵头的主权 AI 芯片政策,并体现在 SKT 与 Rebellions 的采购策略中,是对这一动态的直接政策回应。 主要采用约束是 CUDA 生态锁定。Nvidia 的 CUDA 开发者平台已经积累十多年模型优化、工具链集成和开发者熟悉度。把现有 GPU 推理负载切换到 NPU 替代方案,成本实质存在:团队必须迁移模型代码、验证准确率,并重建 MLOps 流水线。Rebellions 的 RBLN SDK 和 Red Hat OpenShift AI 集成在应对这一约束,但还没有证明能在超大规模云厂商尺度上支撑广泛生产迁移。第二个约束是电力可得性:Nvidia 将数据中心电力约束列为需求逆风,这反过来有利于 Rebellions chiplet 这样的高能效替代方案在受限部署中胜出。第三个约束是 benchmark 不透明:Rebellions 的 REBEL Quad chiplet(Hot Chips 2025)声称 TPS/W 有突破,但截至研究日期尚未发布独立第三方 benchmark 结果。 [CM006, CM018, CM019, CM020, CM021, CM023]

增长驱动因素与约束表
驱动因素 / 约束方向时点给 Rebellions 的影响尽调追问
超大规模云厂商 AI 资本开支扩张(Nvidia FY2026 C&N +67%)驱动因素 - 上行当前(2025-2026 持续)验证总市场规模;Rebellions 必须在扩大的蛋糕里拿到份额超大规模云厂商 AI 算力资本开支中,有多少比例专用于推理而非训练?
美国出口管制推动 AI 供应链地缘政治化驱动因素 - 上行(利好替代芯片)当前(按 Nvidia 10-K,FY2026 生效)在韩国、日本及重视主权的市场,催生非美国供应链需求确认韩国政府政策是否明确偏向本土芯片厂商而非进口产品
韩国主权 AI 政策(MSIT 和 SKT 要求)驱动因素 - 上行近期(2025-2026 正式化)提供政策托底需求;降低首批电信运营商订单的获客成本韩国 AI 半导体计划已承诺的预算和时间表是什么?
CUDA 生态锁定效应(Nvidia 开发者护城河)约束 - 下行持续存在(多年开发者切换成本)将 Rebellions 限在绿地推理部署,或需要大量开发者投入的迁移已有多少生产工作负载从 CUDA 迁移到 Rebellions RBLN SDK?
数据中心电力约束混合(短期约束;中期利好高效芯片)近期(Nvidia 在 FY2026 10-K 中提到电力约束)在电力受限部署中,可能利好高能效 NPU;但也会限制 AI 建设节奏Rebellions REBEL Quad chiplet 是否有独立验证的 TPS/W 优势,超过 H100?

驱动因素和约束来自 Nvidia FY2026 10-K 披露、SK Telecom 主权 AI 合作公告、Epoch AI 算力成本研究,以及韩国 MSIT 政策报道。影响幅度为定性判断;未能获取独立分析师对市场驱动因素的排序数据。

[CM006, CM018, CM019, CM021, CM023, CM024]

2.4 韩国主权 AI 市场——政策、电信采购与 Rebellions 本土定位

韩国主权 AI 芯片计划,是 Rebellions 近期最重要的需求向量。政策语境很明确:韩国政府在 Ministry of Science and ICT 支持下,把本土 AI 半导体发展列为国家战略优先事项。SK Telecom 通过与 Rebellions 和 Arm 合作共同开发面向 SKT 自身数据中心基础设施的主权 AI 推理芯片,把政策转成商业合同。HPCwire 和 Convergedigest 对该计划的报道,将其描述为把 Rebellions 芯片定位成韩国主权 AI 电信数据中心的推理层。SKT 已发布正式新闻稿确认合作。 对 Rebellions 来说,韩国市场的结构性优势是垂直整合:Samsung Electronics 既是代工伙伴(制造 Rebellions 芯片),也是战略投资方;Samsung SDS 又是韩国主导性云服务商。SK Telecom 是战略 AI 芯片客户。这形成了芯片设计、制造、云部署和 AI 负载拥有者之间异常紧密的闭环;美国以外的大多数 AI 芯片创业公司无法接触这种结构。NTT Docomo 的参与把网络延伸到日本,这是中国之外亚洲第二大 AI 芯片市场。 韩国主权 AI 细分里的关键尽调风险,是采购节奏。政策公告往往比实际硬件采购早 12–24 个月;本报告审阅的公开来源没有完全披露韩国政府 AI 投资时间表中哪些仍是规划、哪些已变成承诺合同。Rebellions 已生产并出货 ATOM、REBEL 和 REBELRACE 芯片,建立了生产可信度,但 SKT 实际下单的主权 AI 芯片规模与宣示性目标之间的差距,需要一手来源核验。Arm 和 Samsung 支持的 $250 million 融资(2025 年)以及 $400 million pre-IPO 轮(2026 年)说明投资人相信收入规模可以达成,但来自主权 AI 合同的确认年经常性收入没有公开披露。 [CM022, CM024, CM026, CM027, CM028, CM029]

FM004: 采用漏斗 / 价值链地图

AI 推理芯片采用漏斗,展示从政策 / 市场认知到生产部署的采购旅程。阶段值代表相对机会集,不是 Rebellions 具体管线数量。

[CM024, CM026, CM027, CM037]

2.5 证据要点

Chapter 03

03竞争格局

3.1 竞争格局概览

Rebellions 所在的赛道,是半导体行业竞争最激烈的细分之一:面向数据中心部署的专用 AI 推理加速器。竞争格局至少分四层。第一层也是主导层是 Nvidia,其 Blackwell 架构(GB200 NVL72)的 LLM 推理速度比上一代 H100 快 30x,每瓦性能高 25x,为所有挑战者设下必须追上、或在总拥有成本上可信压低的标杆。Nvidia H100 相比 A100 已经把 LLM 推理提升 30x;连续的阶跃变化意味着每一代都会重置竞争门槛。 第二层是 GPU 通才:AMD Instinct MI300/MI325X 系列以 256 GB HBM3E 内存和 6 TB/s 带宽,提供 Nvidia H100/H200 的直接替代方案,并由 ROCm 开源软件栈支撑。AMD 同时瞄准训练和推理,竞争面很宽,这一点是 Rebellions 只做推理的产品无法匹配的。 第三层——与 Rebellions 最直接竞争——由 AI 推理芯片创业公司组成:Groq(LPU/GroqCloud,$750M 融资后估值 $6.9B)、SambaNova(Dataflow RDU、SambaCloud)、Cerebras(Wafer Scale Engine 3)、Tenstorrent(Wormhole,开源 RISC-V)和 FuriosaAI(RNGD/Renegade,512 TFLOPS,180W)。其中 FuriosaAI 是韩国市场最直接竞争对手,用相似的 180W 风冷形态瞄准同一批电信 / 主权 AI 客户。Groq 已展示最大商业牵引力,但在 2025 年 12 月转向授权 Nvidia 技术,创始人兼总裁加入 Nvidia——这是该领域创业公司独立性遭遇的重大反向信号。 第四层是会压缩 Rebellions 可服务市场的替代方案:Google Cloud TPU(Ironwood,每 pod 42.5 ExaFlops)和 AWS Trainium(Trainium3,3nm,相比上一代效率 4x)是超大规模云厂商自用芯片,不对外销售,但会限制 Rebellions 进入超大规模云客户。Hailo 是边缘加速器公司,不参与数据中心推理竞争。内部自研(超大规模云厂商超出上述清单的自研 ASIC)是第五类,是最大 AI 支出方的现状选择。 [CP001, CP026, CP029, CP030, CP031, CP038]

竞争对手概况表
竞争对手类别规模 / 融资目标客群核心差异化相对 Rebellions 的局限
NvidiaGPU 在位者$3T+ 市值;FY2026 C&N 收入 ~$193B全部 AI 工作负载;数据中心;企业;HPCCUDA 生态护城河;H100/GB200 性能领先;全栈(DGX、网络、软件)绝对在位者;每家 NPU 初创都要对标 CUDA;Blackwell 重置性能标杆
AMDGPU 通用厂商大市值上市公司;Instinct MI300/MI325X 已量产数据中心 AI 训练和推理;云 CSP;HPCROCm 开源;256GB HBM3E(MI325X);广泛 CDNA3 生态从 Nvidia 转向 AMD 的切换成本低于 NPU;GPU 优先的工程团队更偏好 AMD,而不是重新集成 NPU
Groq推理初创公司(LPU)~$1.39B+ 融资;$6.9B 估值(Sept 2025);Samsung、Cisco 投资方开发者;API 优先的云端推理;主权 AI(MENA)LPU 专为推理打造;GroqCloud API 拥有 2M+ 开发者;单 token 超低延迟创始人和总裁离职加入 Nvidia(Dec 2025);GroqCloud 是 API 层,不是硬件销售竞争者
SambaNova推理初创公司(RDU)私营;截至 2023 年已融资 $676M+;SambaCloud企业 AI;智能体工作流;云端推理 API数据流架构;多模型同时执行;MiniMax M2.7 上 435 tok/s云 API 模式;硬件未大规模对外销售;销售落地聚焦美国
Cerebras推理 / 训练初创公司(WSE)上市公司(IPO 2025/2026);Sunnyvale, CA大模型训练和推理;研究;企业Wafer Scale Engine 3:4T 晶体管、125 petaflops;无内存带宽瓶颈需要液冷;功耗 / 散热边界与 Rebellions 不同;重心偏训练
Tenstorrent训练 / 推理初创公司$693M D 轮(Samsung、LG 投资方)研究;云端训练;推理;开源 AI开源 RISC-V;广泛生态策略;Samsung/LG 背书聚焦西方市场;开源模式可能压低单芯片利润率;同样与 Samsung 投资方重叠
FuriosaAI韩国推理初创公司(NPU)计划 IPO 轮融资 $500M(Jan 2026);LG CNS 合作伙伴韩国电信运营商;企业 AI;主权 AI;亚洲数据中心RNGD:512 TFLOPS、48GB HBM3、180W 风冷;自 Jan 2026 起出货;Tensor Contraction Processor韩国直接同业,目标客户几乎相同;同一主权 AI 预算的 IPO 竞品
Google Cloud TPU超大规模云厂商自用 ASICGoogle 内部投资;Ironwood GA May 2026Google 内部 AI(Gemini、Search、Maps);仅限 Google Cloud 客户Ironwood 第 7 代:42.5 ExaFlops/pod;专为智能体 AI 打造;1B+ 用户规模Google Cloud 之外不可用;不对外销售芯片;是替代威胁,不是直接竞争者
AWS Trainium超大规模云厂商自用 ASICAWS 内部投资;Trainium3 3nm(2026)Amazon 内部;AWS EC2 Trn3 客户;Anthropic 合作Trainium3:3nm、2.52 PFLOPS FP8,能效为 Trn2 的 4x;Neuron SDK 生态仅限 AWS;不对外销售;压缩 Rebellions 在超大规模云厂商中的可服务市场
Hailo边缘 AI 厂商私营;Hailo-8/10H/15 已量产;以色列公司边缘设备:摄像头、汽车、机器人、工业低功耗边缘推理(1–10W);Hailo-15 AI ISP 支持 4K30 摄像头;成本效率高仅做边缘;不参与 Rebellions 所在的数据中心推理市场

规模 / 融资数据来自截至 May 2026 的官方新闻稿和新闻室来源;私营公司估值为最后披露轮次的投后估值。AMD、Nvidia、Google、Amazon 的收入数字来自公开披露。“局限”单元格反映作者的分析判断,基于公开定位推断,并非竞争对手确认。SambaNova 累计融资估计汇总截至 2023 年公开报道轮次;所审阅公开来源未确认后续轮次。

[CP001, CP002, CP008, CP015, CP016, CP018]
FP001: 竞争定位图 — 推理聚焦度 vs. 生态规模

八家推理芯片竞争者在两轴上的序数定位:推理专业化程度(x:通用 GPU 到只做推理的 NPU)和生态 / 软件规模 (y:新兴到成熟)。位置是有证据支撑的序数估计,不是数值基准测试分数。

所有位置均为作者基于本次研究审阅公开证据给出的序数评分(1–10);并非来自基准测试数据或第三方排名。 x 轴:1=通用 GPU(训练 + 推理),10=只做推理的 NPU。y 轴:1=新兴开发者生态,10=拥有 1M+ 开发者的 成熟生态。具体数值不应视为精确测量。

[CP001, CP021, CP029, CP039, CP040]

3.2 直接推理 NPU 竞争对手

Rebellions 最接近的对手,是一批同样聚焦推理的芯片公司;它们和 Rebellions 一样,把自己定位成 Nvidia 通用 GPU 架构之外的专用替代方案。 **Groq**(2016 年成立,Mountain View, CA)首创 LPU(Language Processing Unit),即专为推理构建的芯片架构。截至 2026 年 5 月,Groq 已在多轮融资中募集超过 $1.39B,最近一次是 2025 年 9 月以 $6.9B 投后估值融资 $750M,支持方包括 Samsung、Cisco 和 Disruptive。Groq 在北美、欧洲和中东运营数据中心,通过 GroqCloud 服务 2M+ 开发者。其沙特阿拉伯布局以 $1.5B 主权 AI 承诺为锚点,与 Rebellions 的中东扩张目标形成地理重叠。2025 年 12 月,Groq 与 Nvidia 签订非独家授权协议;关键在于,Groq 创始人 Jonathan Ross 和总裁 Sunny Madra 作为该安排的一部分离开 Groq、加入 Nvidia。这是迄今为止最强的反向信号,指向独立推理芯片公司面对 Nvidia 软硬件生态时的生存难度。 **FuriosaAI**(首尔,韩国)是 Rebellions 在韩国和亚洲市场最直接的竞争威胁。FuriosaAI 的 RNGD(Renegade)芯片提供 512 TFLOPS(8 个处理单元,每个 64 TFLOPS FP8)、48 GB HBM3 内存和 180W TDP,瞄准风冷数据中心;这与 Rebellions 的目标热设计包络相似。FuriosaAI 于 2026 年 1 月开始出货 RNGD,2026 年 2 月拿下 LG CNS 企业 AI 合作,并计划在自身 IPO 前融资 $500M。两家公司都来自韩国,都尚未实现规模化收入,也都瞄准同一批韩国电信和主权 AI 客户。至少在早期韩国数据中心采购中,二者竞争是零和的。 **SambaNova**(Palo Alto, CA)用 Reconfigurable Dataflow Unit(RDU)架构形成差异化;该架构把模型执行直接映射到处理器上,以减少数据移动——这种推理效率思路与 Rebellions 的 NPU 聚焦在理念上相近,但架构不同。SambaNova 的 SN50(第 5 代)支持智能体 AI 负载,并可同时运行多个模型。SambaCloud 报告其在 API 层上 MiniMax M2.7 达到 435 tokens/second,DeepSeek-V3.1 达到 200 tokens/second。SambaNova 主要在云 API 层竞争,而不是裸芯片供应商;这意味着它与 Rebellions 在企业 AI 上重叠,但商业化模式不同。 **Cerebras**(Sunnyvale, CA)采用晶圆级路线:WSE-3 在单片晶圆上集成 4 trillion 个晶体管,提供 125 petaflops。截至 2026 年 5 月,Cerebras 已完成 IPO。Cerebras 同时瞄准训练和推理,但晶圆级架构需要液冷,目标性能 / 功耗包络也不同于 Rebellions 的 chiplet 路线;在多数数据中心部署中,它们更多是间接竞争而非直接竞争。 **Tenstorrent**(Toronto/Austin)已完成由 Samsung Securities 和 LG Electronics 领投的 $693M Series D,并以开源 RISC-V 路线定位 Wormhole AI 加速器,试图建立比专有 NPU 公司更广的开发者生态。Tenstorrent 同时瞄准训练和推理;Samsung/LG 背书让它与 Rebellions 自身 Samsung 制造和 LG 生态关系产生潜在重叠。 [CP001, CP002, CP003, CP004, CP005, CP006]

功能与能力矩阵
采购标准Rebellions(REBEL/Rebel100)Nvidia H100/GB200AMD MI325XGroq(LPU)SambaNova(SN50)Cerebras(WSE-3)Tenstorrent(Wormhole)FuriosaAI(RNGD)
主要工作负载重心仅推理 NPU训练 + 推理(GPU)训练 + 推理(GPU)仅推理(LPU)仅推理(RDU)训练 + 推理(WSE)训练 + 推理仅推理(NPU)
内存容量 / 带宽HBM3E(REBEL,具体待定)内存:80GB HBM2e / 3.35 TB/s(H100);H200 141GB内存:256GB HBM3E / 6 TB/s(MI325X)仅片上 SRAM;无 DRAM多层级(片上 SRAM + HBM + 外部)晶圆上(无外部 DRAM 瓶颈)HBM2e(Wormhole)内存:48GB HBM3 / 1.5TB/s
软件生态软件:Rebellions SDK;Red Hat OpenShift AI(Dec 2025)CUDA + cuDNN(成熟,4M+ 开发者)ROCm(开源,3.7+ million 个应用)GroqCloud API(兼容 OpenAI);2M+ 开发者SambaCloud API + SambaNova SDKCerebras SDK + PyTorch 支持TT-BUDA + RISC-V 开源Furiosa SW Stack(编译器、运行时、性能分析器)
商业化模式硬件 + SDK 销售(企业)直销 + OEM + 云;主导渠道直销 + OEM + 云;AMD 合作伙伴计划云 API(GroqCloud);数据中心部署云 API(SambaCloud)+ 企业本地部署云 + 本地部署硬件销售硬件 + SDK;开源生态硬件 + SDK;企业 + 云部署
制造 / 晶圆厂资源Samsung Foundry(韩国);HBM3E 来自 SK HynixTSMC(Blackwell 使用 N4/N3)TSMC(MI300 系列使用 N5/N4)TSMCTSMCTSMCTSMC/SamsungSamsung Foundry(韩国)
地理目标 / 护城河韩国;亚太;MENA 主权 AI全球(处处占优)全球;美国 / 欧洲云 CSP 强势全球 API;美国 + 沙特 / MENA 主权 AI美国企业;SambaCloud 全球 API美国;全球研究 / 企业全球;开放生态韩国;亚洲数据中心;LG CNS 企业客户

Rebellions 称 REBEL chiplet(UCIe-Advanced,Nov 2024 流片)的内存规格为 “HBM3E”;产能和精确带宽未公开披露。标注 “(具体待定)” 的单元格代表公开不可得或未经确认的规格。Nvidia H100 数据来自官方规格表;GB200 数据为机架级口径,不能直接与单芯片指标相比。AMD 相对竞争对手 1.3x 的说法为公司自称。SambaNova 内存架构为多层级;细分拆解未完全公开。所有软件生态成熟度评估均为作者基于公开开发者活动信号作出的判断。

[CP001, CP009, CP010, CP012, CP021, CP026]
FP002: 功能广度与能力地图

截至 2026 年 5 月,八家竞争者在六个关键维度上的能力覆盖快照。单元格反映基于官方产品页和新闻来源的、 有证据支撑的评估。

「未知」表示研究中未找到公开证据的单元格。「部分」表示能力已确认但不完整。「主权 AI 客户」反映截至 2026 年 5 月,与政府或国家级 AI 项目的已确认合作或部署。Nvidia H100 液冷说明指数据中心规模部署 (GB200 NVL72 需要液冷);部分较小 H100 PCIe 形态可风冷。

[CP001, CP009, CP010, CP021, CP029, CP039]

3.3 GPU 在位者与超大规模云厂商自研 ASIC

Nvidia 和 AMD 占据 GPU 在位者层。Nvidia 的优势同时来自硬件和软件:H100 的 LLM 推理速度相对 A100 提升 30x,GB200 NVL72 又在 H100 基础上再提升 30x,这意味着 Nvidia 每一代都会重置性能基线。CUDA 软件生态约在 2006 年建立,拥有 4M+ 开发者和 3,800+ GPU 加速应用,形成结构性切换成本;目前还没有挑战者在规模上系统性突破。每一家推理芯片创业公司,包括 Rebellions,都必须提供的不只是更好硬件,还包括迁移工具、运维支持和模型兼容性,足以匹配 CUDA 用户已经拥有的能力。 AMD Instinct MI325X(CDNA3 架构)提供 256 GB HBM3E 和 6 TB/s 带宽;AMD 声称其 AI 性能是竞争加速器的 1.3x。AMD 的 ROCm 开源软件栈已经成熟,并越来越兼容 PyTorch 和主流 LLM 框架。对不想依赖 Nvidia 的云运营商和企业来说,AMD 是摩擦最低的替代方案:熟悉的 GPU 编程模型、广泛生态和多供应商采购。AMD 对 Rebellions 的威胁不一定是取代 Nvidia,而是填补可信 GPU 替代品的角色,从而挤出推理 NPU 评估周期。 超大规模云厂商自用 ASIC——Google Cloud TPU(Ironwood,第 7 代,每 pod 42.5 ExaFlops)和 AWS Trainium(Trainium3,3nm 制程,2.52 PFLOPS FP8,相比 Trainium2 能效 4x)——是大型云部署中最强的替代威胁:这些芯片不对外销售,只会缩小 Rebellions 和其他外部芯片供应商可触达的云推理市场。Google TPU 支撑 Gemini 和 1B+ 终端用户;AWS Trainium3 在 EC2 上支持 Anthropic 和其他主要 AI 客户。二者不算 Rebellions 的直接竞争对手,因为它们不争夺同一个采购决策,但都会降低任何外部 AI 芯片供应商争取超大规模云客户时的可服务市场(SAM)。 [CP026, CP027, CP028, CP029, CP030, CP031]

定价与打包对比
竞争对手销售模式定价 / 单位(公开)合同 / 最低承诺包含能力给 Rebellions 的定价启示
Nvidia H100 SXM直销;OEM;云租赁$30K–$40K / GPU(现货;截至 late 2024)云端无最低承诺;OEM 批量折扣CUDA、cuDNN、Triton、NVLink;完整生态Rebellions 要赢,TCO(不是芯片价格)必须低于 Nvidia;GB200 可能进一步拉大 TCO 差距
Nvidia GB200 NVL72仅云租赁(初期)云租赁;硬件价格未发布仅限超大规模云厂商 OEM;不零售完整 Blackwell 技术栈;推理性能为 H100 的 30xNvidia 新一代产品持续重置 TCO 标杆
AMD MI325X直销;OEM;云市场价未官方发布;估计 $25K–$35KOEM 批量;云端按小时ROCm、CDNA3、256GB HBM3E;广泛框架支持AMD 价格竞争可能削弱高端 NPU 在价格敏感买家中的定位
Groq (GroqCloud)仅云 API按 token 定价;相对云 GPU API 有竞争力无硬件最低采购;提供开发者套餐兼容 OpenAI 的 API;覆盖主要开放模型API 层竞争者;不卖硬件;与 Rebellions 硬件销售渠道重叠有限
SambaNova (SambaCloud)云 API + 企业本地部署API 定价未公开详细披露企业本地部署走销售周期SambaCloud API;多模型执行争夺企业 AI 预算;定价未知
Cerebras硬件 + 云WSE-3 硬件价格未公开列示;提供云租赁企业;云租赁按小时Cerebras SDK;PyTorch;晶圆级计算功耗 / 冷却要求不同;对多数 Rebellions 目标部署并非直接替代
Tenstorrent硬件(板卡 / 系统)Wormhole n150s ~$1.3K(社区价格);数据中心系统待定开发者板卡无最低采购;企业议价开源工具;RISC-V 固件开发者获取门槛低;企业定价未披露
FuriosaAI (RNGD)硬件 + SDK未公开披露企业销售周期Furiosa SW Stack;LG CNS 集成韩国电信 / 主权 AI 的直接价格竞争者;定价未披露

定价数据不完整,主要来自截至 May 2026 的现货市场报道、社区论坛和云端按需定价页面。多数 NPU 初创公司(Rebellions、FuriosaAI、Cerebras 企业版)的硬件标价未公开披露,需要直接接洽供应商。Nvidia H100 现货价格波动很大;所列区间反映 late 2024 广泛报道的市场数据;GB200 硬件标价未公开。Tenstorrent Wormhole 开发者板卡价格来自公开可得的社区来源。

[CP001, CP021, CP026, CP029, CP040]

3.4 地理与相邻动态

Hailo(以色列无晶圆厂半导体公司)处在边缘 AI 层,不与 Rebellions 在数据中心推理竞争。Hailo 的 Hailo-8、Hailo-10H 和 Hailo-15 处理器瞄准嵌入式边缘平台——摄像头、汽车 ECU、机器人割草机、工业视觉系统——功耗区间为 1–10W。随着 AI 模型变小,边缘市场可能增长;但在 2026 年,Hailo 和 Rebellions 服务的是完全不同的买方和负载。 地理维度更关键。Rebellions 身处韩国主权 AI 生态,Samsung、SK Telecom 和国家政策要求带来西方竞争者难以复制的优先准入。FuriosaAI 是这一韩国语境中唯一可信的同类对手。Groq 在沙特阿拉伯的存在,以及 Rebellions 2025 年与 SK Telecom 和 Arm 围绕 Saudi/MENA 主权 AI 基础设施的合作,构成了一个高增长地区里的真实重叠。Tenstorrent 的 Samsung Securities 和 LG Electronics 支持方,与 Rebellions 自身的 Samsung 代工和 SK Telecom 战略投资方共享机构关系——这可能让两家公司竞标同一批韩国企业 AI 合同。 很多潜在 Rebellions 客户的现状替代方案,是继续依赖 Nvidia GPU:已经集成 H100 集群的 CSP 和企业,在 CUDA 软件、运维工具和应用兼容性上投入了大量沉没成本。从任何供应商切换到 NPU——不只是 Rebellions——都会产生迁移成本,拖慢采用周期,并给 Nvidia 带来复利式先发优势。内部自研(定制 ASIC 开发)主要是超大规模云厂商的替代方案;Rebellions 的目标客户(电信运营商、主权 AI 项目、中型 CSP)通常没有自研芯片所需的工程能力和规模,这为 Rebellions 等外部 NPU 供应商保留了位置。 [CP017, CP018, CP019, CP020, CP037, CP038]

3.5 护城河耐久度与被替代风险

Rebellions 的竞争护城河若存在,主要是地理和关系层面的,而不是技术层面的。Samsung 制造合作让它能以本土条件获得先进制程,外国芯片创业公司很难在韩国复制;SK Telecom 战略投资方关系为电信 AI 部署提供优先首客路径;韩国 MSIT 的 AI 芯片政策要求(在主权 AI 项目中部署本土芯片)给了 Rebellions 一个事实上的受保护细分市场。上述护城河都不是永久的:Samsung 也可以为 FuriosaAI 代工;SK Telecom 的持股可能被稀释或转让;政策要求也会变化。但在韩国市场,它们能提供 2–4 年竞争隔离期。 被替代风险真实存在,且来自多个方向。Nvidia 永不停步的硬件路线图(H100 → H200 → GB200 → Rubin,每一代都压缩与 NPU 挑战者的性能差距)是主要风险。如果 Nvidia 能以有竞争力的总拥有成本同时交付更高性能和更好能效,NPU 的价值主张就会收窄。2025 年 12 月 Groq-Nvidia 交易——Groq 创始人加入 Nvidia——是最直接证据,说明即使不进行完整收购,Nvidia 也能通过吸收关键人才和技术来中和 NPU 挑战者。AMD 带来的商品化风险(ROCm 成熟、MI300X 部署)也会侵蚀「NPU 是独特高效替代方案」的叙事。最后还存在多归属风险:云运营商和企业可以运行混合机群(Nvidia + NPU),这会降低采用 NPU 的紧迫性,并限制合同规模。Rebellions 与 Red Hat OpenShift AI 的集成(2025 年 12 月)是降低迁移摩擦的相关一步,但任何 NPU 挑战者要达到 CUDA 生态平价,仍需数年。 [CP004, CP040, CP041, CP042, CP043]

护城河耐久性与竞争风险登记表
护城河或风险方向严重性缓释措施 / 尽调追问
Nvidia CUDA 软件生态锁定效应风险(所有挑战者)Rebellions SDK 必须实现 CUDA 兼容,或让迁移几乎无摩擦;Red Hat OpenShift AI 集成只能部分缓释,距离 CUDA 对等能力仍有多年差距
FuriosaAI 争夺韩国主权 AI 预算风险(直接)两家公司瞄准同一批韩国电信 / 政府客户;Rebellions 的 SK Telecom 战略投资者股权是一道护城河,但 FuriosaAI 的 LG CNS 合作也打出一条平行渠道
Groq-Nvidia 授权信号:初创公司人才 / IP 向 Nvidia 迁移风险(生态)中高Groq 创始人 Jonathan Ross 和总裁 Sunny Madra 加入 Nvidia(Dec 2025);Nvidia 可以系统性收购初创公司的技术和人才,作为竞争防御机制
Samsung 制造资源(Rebellions)护城河Rebellions 的 Samsung Foundry 晶圆厂和共同开发协议,提供西方竞争者拿不到的韩国本土晶圆厂资源;但 Samsung 也可能为 FuriosaAI 和其他韩国公司代工
韩国 MSIT 主权 AI 要求护城河韩国政府偏好国产 AI 芯片的政策,给 Rebellions(和 FuriosaAI)一个受保护的本土市场;风险在于政策反转,或授权范围扩大到 FuriosaAI
Tenstorrent 与 Samsung/LG 投资方重叠风险Tenstorrent 融资 $693M,由 Samsung Securities 和 LG Electronics 领投——这套机构生态也包括 Rebellions 的 Samsung 晶圆厂伙伴;可能造成机构渠道冲突
Nvidia 硬件路线图压缩(Rubin 下一代)风险每一代 Nvidia(A100→H100 30x;H100→GB200 30x)都会重置 NPU 性能 / TCO 价值主张;Rebellions 的 REBEL 必须保持领先一代的效率优势,否则可能在量产前就被超越
AMD ROCm 生态成熟风险AMD ROCm 与 PyTorch/HuggingFace 的兼容性越来越高;如果 ROCm 接近 CUDA 兼容,买家可能更偏好熟悉的 AMD GPU 模式,而不是新型 NPU
SK Telecom 战略投资者治理风险(治理)中低合并后 SK Telecom 仍是战略投资者;其董事会影响力范围,以及与全球扩张方向冲突的可能性,尚未公开披露
客户多宿主 / 混合 NPU+GPU 集群风险企业和云客户可能运行混合集群(Nvidia H100 + NPU 补充);这会把 Rebellions 限在次级工作负载,而不是替代主力推理集群

严重性评级是作者基于公开证据作出的分析判断;没有公司确认或排序这些风险。“方向”指该因素是利好 Rebellions(护城河)还是威胁 Rebellions(风险)。缓释路径是尽调建议,不是已确认的公司计划。

[CP004, CP008, CP040, CP041, CP042]
FP003: 护城河与竞争准备度 KPI

截至 2026 年 5 月,Rebellions 关键竞争耐久指标的紧凑摘要,覆盖融资、护城河来源和竞争威胁的交叉点。

[CP002, CP004, CP008, CP021, CP030, CP031]

3.6 证据要点

Chapter 04

04财务情况

4.1 收入模式与公开牵引力可见度

Rebellions 采用硬件加软件的收入模式。芯片(ATOM、REBEL-Quad/Rebel100)由内部设计,交由 Samsung Foundry 在先进制程节点制造,并以独立加速卡(RebelCard)、整机服务器(RebelServer)和机架级系统(RebelRack、RebelPOD)销售。软件层是一个贴近开源生态的 SDK,支持 vLLM、PyTorch、Triton、Hugging Face 和 Red Hat OpenShift AI;截至本次评审时点,公司把它随硬件打包提供,而不是单独变现。 公开收入证据很少。Forbes Asia 2025 年 4 月人物稿审阅的韩国公司登记文件显示,FY2023 收入约 2.7 billion KRW(按 2023 年当时平均汇率约 1,300 KRW/USD,约 $2.1 million)。这代表 ATOM 芯片带来的早期商业牵引;该芯片 2023 年 5 月进入量产,并已部署给 SK Telecom 数据中心部门、Kakao、Naver 等客户。截至本报告日期,公司没有公开披露 FY2024 或 FY2025 收入;ARR、GMV 或芯片出货量也没有发布。 CEO Sunghyun Park 在 2025 年 4 月 Forbes 人物稿中给出 2025 年收入目标 100 billion KRW(约 $68 million),意味着相对最后一个已知公开基线约跳升 37×。本章审阅的任何来源都没有确认该目标是否达成。Series C 和 pre-IPO 新闻稿提到公司在日本、沙特阿拉伯和美国面向“企业和政府”部署,并称基于 ATOM 的芯片“支撑韩国最大的商业 AI 服务”——这些表述能佐证客户牵引,但没有量化收入。 定价没有公开价目表。公司把价值主张表述为推理工作负载下,相比 Nvidia H100 有更好的每瓦性能和总拥有成本;这意味着公司主打性价比定位,而不是公开标价模型。与 Samsung 的 turnkey 合作覆盖芯片制造、HBM3E 内存和封装,Pegatron 与 Penguin Solutions 负责系统组装合作,说明渠道和伙伴经济性已经嵌进成本结构,但没有可用的利润分成数据。 [CI001, CI002, CI003, CI004, CI005, CI006]

收入流与变现机制
收入流机制单位 / 定价基础当前状态(May 2026)收入质量尽调追问
ATOM 芯片 / 服务器销售硬件销售(芯片 + ATOM-Max 服务器)按单元计价;未披露标价已量产;已在韩国、日本、沙特阿拉伯、美国部署硬件业务,毛利率能见度低确认出货量和 ASP
REBEL-Quad / Rebel100 芯片销售硬件销售(基于 chiplet 的 NPU)按单元计价;未披露标价已宣布扩大量产;已承诺 IPO 前资本开支硬件业务,尚处规模化前阶段索取单颗毛利率;Samsung 晶圆成本
RebelCard 加速卡硬件销售(插卡)按卡计价;未披露标价随 IPO 前融资同步发布(March 2026)硬件渠道利润率确认与 Pegatron / Penguin Solutions 的渠道经济性
RebelRack / RebelPOD 集成系统系统销售(整机柜 / 集群)按系统计价;未披露标价截至 March 2026 已可供货ASP 更高的系统产品;利润率结构未知索取系统级毛利率,并与纯芯片销售对比
Rebellions SDK / 软件栈随硬件捆绑;无独立定价未单独定价贴近开源策略;捆绑提供目前披露的软件收入为零确认是否存在 SaaS 或支持合同收入
合作伙伴 / 政府 AI 项目基于合同(主权 AI、电信运营商)按项目 / 合同计价Saudi Aramco、SKT、NTT DOCOMO Innovations 合作政府背书收入;留存风险取决于合同续签获取合同金额、期限、排他条款

所有当前状态陈述均来自截至 May 2026 的官方新闻稿和新闻报道;标价、ASP 或客户合同金额均未公开披露。收入质量评级根据类似商业模式推断。等同 null 的条目(“未披露标价”)表示已确认缺少公开定价数据,不代表收入为零。

[CI001, CI002, CI003, CI004, CI005, CI006]
定价与变现背景
产品 / 服务定价基础标价 / 实际成交价已披露折扣 / 条款来源
ATOM / ATOM-Max 芯片按单元计价(硬件)未公开披露未披露公司新闻稿;Forbes Apr 2025
REBEL-Quad(Rebel100)按单元计价(硬件)未公开披露未披露官方 Series C 轮和 IPO 前融资新闻稿
H100 GPU(Nvidia 基准)~$25,000–$35,000 单台标价(行业参考)标价被广泛引用;实际成交价不一批量折扣常见;不适用于 Rebellions行业媒体;仅作基准
RebelRack / RebelPOD 系统按系统计价(硬件 + 集成)未公开披露未披露IPO 前融资新闻稿(March 2026)
软件 SDK捆绑提供;无独立价格未确认独立变现N/A公司文档;开源策略

Rebellions 产品定价数据完全未披露。Nvidia H100 参考项是 Rebellions 在公开定位主张(TCO 对比)中使用的基准锚点。本表仅反映公开可得信息;客户合同的实际成交价格可能与任何标价有实质差异。

[CI007, CI008]
FI001: 收入模型桥 — 从客户活动到收入和毛利润

展示 Rebellions 如何把 AI 推理需求转化为芯片和系统销售,以及当前阶段毛利转正的障碍。

毛利润节点是定性判断;没有公开 COGS 或毛利率数据。流程反映官方新闻稿和 Forbes 报道描述的商业模式机制, 而非财务报表数据。

[CI001, CI002, CI003, CI005, CI009]

4.2 融资历史、投资人结构与资本基础

截至 2026 年 3 月 30 日,Rebellions 已通过五轮融资累计募集 $850 million,成为韩国融资额最高的 AI 芯片创业公司,也是韩国芯片行业首个独角兽。融资节奏尤其压缩:2025 年 9 月至 2026 年 3 月的六个月里,公司募集 $650 million,占累计融资超过 75%。 逐轮来看,最早是未披露金额的种子轮和 Series A,合计约 $76 million(根据 2024 年 1 月披露的“已融资超过 $200 million”,扣除 $124 million Series B 推算)。$124 million Series B 于 2024 年 1 月 30 日完成,由 KT Corporation 领投;KT 是韩国主要数据中心运营商,也是战略客户。Korelya Capital(法国)、Korea Development Bank、Samsung Ventures 等参投。2024 年 7 月 23 日,公司又获得 $15 million Series B 延展轮,投资方为 Saudi Aramco 旗下风险投资机构 Wa'ed Ventures;这是 Wa'ed 对韩国创业公司的首笔投资,也直接打开了沙特市场关系。 $250 million Series C 于 2025 年 9 月 30 日完成,投后估值 $1.4 billion。Arm 以战略投资人身份领投——这一点重要,因为它带来 ARM 架构路线图和生态关系入口——Samsung Ventures 与 Pegatron VC 也参与其中;后者总部在台湾,是关键制造伙伴。老股东 Korea Development Bank 和 Korelya Capital 继续跟投。2025 年 11 月 10 日,Series C 又获得延展,Kindred Ventures(首次投资韩国创业公司)和 Top Tier Capital Partners 加入,为 IPO 前增加美国风险资本背书。 $400 million pre-IPO 轮于 2026 年 3 月 30 日完成,投后估值 $2.34 billion。Mirae Asset Financial Group 领投;它是韩国最大资产管理机构,自 Series A 起支持 Rebellions。Korea National Growth Fund 也参投,这是其在 K-Nvidia 国家 AI 半导体计划下的首笔投资。本轮释放出强机构信心和隐性政府背书,但也带来公共基金投资常见义务,包括报告要求和募资用途的战略约束。 投资人结构横跨战略半导体伙伴(Arm、Samsung、通过 SAPEON 股份进入的 SK Telecom)、电信 / 数据中心客户(KT)、主权能源资本(Wa'ed/Aramco)、韩国本土金融机构(Korea Development Bank、Mirae Asset、Korea National Growth Fund)和美国风投(Kindred、Top Tier Capital Partners)。这种广度降低了对单一资本方的依赖,也减少下一轮 rollover 风险;但战略和政府关联资本占比较高,可能带来纯财务投资人不会施加的治理或市场准入义务。 [CI010, CI011, CI012, CI013, CI014, CI015]

单位经济性 — 关键指标与证据缺口
指标数值 / 估计置信度重要性尽调要求
FY2023 收入(KRW)收入:~2.7 billion KRW (~$2.1 M)中 — 来自 Forbes 审阅的韩国申报文件合并和大规模扩张前最后一个确认收入基准获取经审计的 FY2024 和 FY2025 数据
FY2023 净亏损(KRW)净亏损:~13.7 billion KRW (~$10.5 M)中 — 来自 Forbes 审阅的韩国申报文件显示合并前运营亏损很深;意味着烧钱主要投向 R&D与合并后实体的 P&L 对比
FY2022 净亏损(KRW)净亏损:~8.1 billion KRW (~$6.2 M)中 — 来自 Forbes 审阅的韩国申报文件亏损扩大趋势;发生在 ATOM 量产前索取 FY2024 亏损趋势,用于判断轨迹
2025 收入目标(CEO 表述,KRW)收入目标:~100 billion KRW (~$68 M)低 — 仅为 CEO 表述;未获确认较 FY2023 跳升 37×;若实现,可验证规模化逻辑;若落空,估值显得偏高询问 FY2025 实际收入与目标的差异
毛利率(%)未披露无可用信息评估硬件经济性和资本效率的关键索取 Samsung Foundry 发票中的 COGS 拆分
ATOM 芯片毛利率未披露无可用信息决定 ATOM 在 REBEL 扩量前是否跑通单位经济性索取 ATOM / ATOM-Max 单颗 / 单台成本和 ASP
客户数(FY2025)未披露无可用信息没有客户数,无法评估集中度风险披露前三大客户收入占比
ARR / 收入运行率(2026)未披露无可用信息IPO 关键指标;缺失会限制可比公司分析获取合并后过去 12 个月收入
月度现金消耗未披露无可用信息计算现金跑道的关键索取管理账,展示 Series C 轮前后月度现金消耗
现金跑道(估计)估计 18–24 个月(根据累计融资 $850 M 和此前烧钱信号推断)低 — 估计高度不确定IPO 必须在现金跑道耗尽前完成;存在时间风险确认截至 March 31, 2026 的在手现金
每次流片 NRE 成本未披露;sub-5 nm EUV 光罩行业区间为 $20–80 M低 — 仅为行业代理值;Rebellions 具体数据未知每一代芯片都会吃掉一轮融资中的相当一部分披露 R&D 资本化政策和历次流片成本

所有 KRW 金额按约 1,300 KRW/USD(2023 平均汇率)折算,仅供参考;实际经济性必须用管理账验证。“未披露”条目表示已确认缺少公开数据,不代表数值为零。NRE / 光罩成本的行业代理值来自公开可得的半导体分析师评论,并非 Rebellions 特定披露。

[CI009, CI022, CI023, CI024, CI025, CI026]
融资轮次时间线与投资者构成
轮次日期金额(USD)投后估值主要投资者战略意义
种子轮 + Series A(合并)2020–2023(约)~$76 M(推断)未披露Kakao Ventures;早期天使投资人概念验证;ATOM 芯片开发
Series B 轮January 30, 2024$124 M~$680 M(880 B KRW 隐含)KT Corp(领投);Korelya Capital;Korea Development Bank;Samsung Ventures战略客户 KT 领投;Korea Development Bank 降低本土资本风险
Series B 延伸轮July 23, 2024$15 M未披露投资方:Wa'ed Ventures(Saudi Aramco)打开沙特阿拉伯市场;Wa'ed 首次投资韩国公司
SAPEON 合并December 2024全股票(非现金)~1.3 T KRW(合并时 ~$1 B)SK Telecom;SK Square;SK Hynix(经由 SAPEON)获得 SK Hynix 的 HBM3E 供应;SKT 作为锚定客户
Series C 轮September 30, 2025$250 M$1.4 B 投后Arm(战略领投);Samsung Ventures;Pegatron VC;Korea Development Bank(跟投);Lion X VenturesArm 作为战略投资者;Pegatron 作为制造合作伙伴
Series C 延伸轮November 10, 2025包含在 $250 M 中$1.4 B(同一轮)投资方:Kindred Ventures;Top Tier Capital PartnersKindred 首次投资韩国初创公司;美国 VC 在 IPO 前背书
Pre-IPO 轮March 30, 2026$400 M~$2.34 B 投后Mirae Asset Financial Group(领投);Korea National Growth Fund(首笔 K-Nvidia 投资)与政府方向一致的机构资本;释放 IPO 路径信号

种子轮和 Series A 金额为推断值:January 2024 的 Series B 新闻稿称当时累计融资“more than $200 million”,意味着 Series B 前资本约 ~$76 million。SAPEON 合并列为非现金全股票交易;多方消息报道合并时合并实体估值约 $1B。Series C 延伸轮没有改变 $1.4B 估值。KRW 到 USD 的折算为近似值。

[CI010, CI011, CI012, CI013, CI014, CI015]
FI003: 财务估计区间 — 收入、烧钱、估值和现金跑道

有来源支撑的关键财务参数边界估计;宽区间反映截至 2026 年 5 月 Rebellions 公开数据有限。

收入区间由 CEO 所称目标(上限)和保守落空情景(下限)界定;暂无实际值。烧钱区间是基于 FY2023 亏损数据的 粗略放大;合并后员工数和全球扩张很可能显著推高月度烧钱。现金跑道为推断,未披露。毛利率区间仅作示意, 用于标出尽调缺口。

[CI009, CI022, CI023, CI031, CI032, CI033]

4.3 成本结构、资本强度与 Fabless 模式经济性

Rebellions 是一家 Fabless 半导体公司。该模式省掉了 Samsung 或 Intel 这类 IDM 需要投入的数十亿美元制造资本开支,但换来另一种资本强度:每次流片都要向 Samsung Foundry 支付反复发生的 NRE(一次性工程)和掩膜版成本,还要为芯片设计中使用的第三方 IP 持续支付版税或许可费。 Samsung Foundry 在先进制程节点上制造 Rebellions 芯片。ATOM 芯片采用领先制程并配 GDDR6 内存;REBEL-Quad 使用芯粒架构和 HBM3E 内存。Samsung Foundry 公开能力覆盖 14/10/8/5/4nm FinFET 和 3nm GAA,5nm 起使用 EUV,并提供适配 Rebellions 芯粒路线的 3D/2.5D 集成封装方案。在带 EUV 的 sub-5nm 节点上,单次流片仅掩膜版成本就可能达到 $20–80 million(行业分析师估计,并非 Rebellions 专属数据),意味着每一代新芯片在首片晶圆出货前,就会消耗一轮融资的相当大一部分。 与 Samsung 的 turnkey 关系覆盖硅片制造、HBM3E 内存供应(通过 SK Hynix 和 Samsung Electronics 获取)和封装,简化了供应商管理,但也集中供应链风险。Samsung Foundry 的任何良率问题、产能分配变化或 HBM3E 供应约束,都会直接影响 Rebellions 履约客户订单的能力。Forbes 报道称,Park 承认“foundry 产能短缺”和“HBM 短缺”是结构性风险。这些是全行业约束,并非 Rebellions 独有;但相较 TSMC 或 Samsung 的超大规模客户,小客户在产能优先级上的议价能力更弱。 成本端看,Rebellions 这一阶段的 Fabless AI 芯片创业公司通常承担三类成本:(1)研发是最大运营费用,由芯片设计工程师人数驱动;(2)销货成本(COGS)包括晶圆、内存、封装,以及 Pegatron 带来的系统组装成本;(3)销售与营销似乎正在增加,因为 Chief Business Officer Marshall Choy 领导下的美国团队在扩张,日本和沙特阿拉伯子公司也已开设。FY2023 净亏损 13.7 billion KRW(约 $10.5M),对应收入 2.7 billion KRW,说明当时经营利润率深度为负;此后的扩张——尤其是 SAPEON 合并(2024 年 12 月)和 REBEL-Quad 量产——应当显著推高收入和支出,但没有 FY2024–2025 财务数据可验证。 成本结构的反向情景是:AI 芯片创业公司在硬件销售毛利转正所需规模到来前,需要持续融资来支撑反复流片周期、设计团队扩张和 go-to-market 投入。公司未公开披露毛利率,外部无法确认 ATOM 芯片是否跑通正向单机经济性。公司强调总拥有成本定位,意味着硬件毛利可能很薄,或者未来需要软件 / 服务补贴。 [CI022, CI023, CI024, CI025, CI026, CI027]

资本充足性评估
项目数值 / 状态来源置信度
累计融资(成立至今)$850 million官方 IPO 前融资新闻稿(March 30, 2026)
最近一轮规模$400 million IPO 前融资(March 30, 2026)官方新闻稿;PRNewswire
最近一轮领投方领投方:Mirae Asset Financial Group;Korea National Growth Fund官方新闻稿
投后估值(最近一轮)~$2.34 billion官方新闻稿
在手现金(截至 March 2026)未公开披露无可用来源无可用信息
月度现金消耗未公开披露无可用来源无可用信息
估计现金跑道估计 18–24 个月(定性)根据累计融资和已披露扩张计划推断
下一轮触发因素IPO(目标:2026 或之后;未披露日期或交易所)CEO 表述;IPO 前融资新闻稿措辞
IPO 前融资计划用途美国市场扩张;Rebel100 扩量生产;IPO 准备官方 IPO 前融资新闻稿
债务义务 / 项目融资未公开披露无可用来源无可用信息
政府背书资本Korea National Growth Fund($400 M 轮中金额,占比未披露)官方新闻稿

在手现金、烧钱速度和债务义务均无公开信息,无法从新闻稿或新闻报道中判断。18–24 个月 现金跑道是定性估计,建立在一个假设上:按 Rebellions 已披露增长轨迹,$400 M 融资意味着资本大约足够支撑两年;如果 SAPEON 合并后烧钱明显加速,这一估计可能大错。“未披露”条目表示已确认缺少数据。

[CI010, CI011, CI012, CI031, CI032, CI033]
成本结构与资本密集度驱动因素
成本类别性质估计重要性证据依据风险因素
R&D / 工程人员固定 / 半固定;这一阶段 fabless 公司的主导 OpEx高(AI 芯片初创公司的 OpEx 通常 50–70%)行业类比;Forbes 提到严格招聘标准;合并后员工数 ~400+合并整合后的人才流失风险
NRE / 光罩套件成本(Samsung Foundry)按流片发生;不连续的大额资本开支高:sub-5 nm EUV 节点行业区间为 $20–80 MSamsung Foundry 已披露能力(3 nm GAA、5 nm FinFET EUV);行业分析师代理值良率风险;晶圆代工产能分配风险
晶圆 / COGS(单颗)可变;与出货量绑定未知 — 未披露无可用 COGS 数据出货量不足会削弱毛利覆盖
HBM3E 内存(SK Hynix / Samsung)按单元发生;行业供给受限高 — HBM3E 稀缺且昂贵Forbes 引述 Park:“shortage of HBM”供给约束限制产量上限
系统组装(Pegatron)按系统发生;可变 COGS未知 — 未披露合作于 November 2025 宣布;无利润率数据合作伙伴利润会压低 Rebellions 系统级毛利率
销售、营销和全球扩张持续增长;美国团队由 Marshall Choy 负责;在日本、沙特阿拉伯设有子公司OpEx 上升;由人员增长驱动IPO 前融资新闻稿;高管任命新闻稿韩国以外市场的获客成本未知
IPO 准备成本一次性;审计、法律、投行费用同等规模公司通常为 $5–15 M行业惯例;无 Rebellions 特定数据加快 2026 年现金消耗

R&D 员工占比和 NRE / 光罩成本的重要性估计,来自公开资料中无晶圆半导体经济模型的行业代理值, 不是 Rebellions 自身披露。HBM3E 供应短缺直接来自 Forbes 对 CEO Park 的采访。 系统组装伙伴利润率按典型合同电子制造商经济模型估计。

[CI022, CI023, CI024, CI025, CI026, CI027]
FI004: 资本强度地图 — 按类别划分的累计资本投放

展示一家处于 Rebellions 阶段的 Fabless AI 芯片公司,累计 $850M 资本可能如何投向关键支出类别。

所有数字均为估计,来自 Fabless 半导体初创的行业可比和公开披露事实(轮次规模、团队扩张、代工合作)。 Rebellions 未披露实际资本投放数据。类别仅用于说明该阶段资本通常如何消耗;实际分配可能大幅不同。

[CI010, CI011, CI028, CI029, CI030, CI033]

4.4 资本充足性、现金跑道与 IPO 路径

基于公开信息,Rebellions 未来 18–24 个月的资本充足性看起来相对稳健,核心锚点是 2026 年 3 月 30 日完成的 $400 million pre-IPO 轮。不过,公司没有公开披露实际现金余额、月度烧钱速度和预计现金跑道;没有内部财务数据,就无法做出精确的充足性判断。 定性信号偏正面:公司在六个月内(2025 年 9 月至 2026 年 3 月)募集 $650 million,说明投资人相信近期收入增长。Korea National Growth Fund 把 Rebellions 定为其首个 K-Nvidia 投资,意味着公司可能与韩国政府支持计划有所对齐,并获得额外补贴或软贷款入口。Mirae Asset 自 Series A 起支持公司,也说明投资人延续性较强,不是可能快速退出的新进入者。 计划中的 IPO 是最关键的前瞻资本事件。新闻稿提到“为未来 IPO 做准备”,但截至报告日期,没有披露交易所、申报日期、价格区间或承销商细节。若 2026 年 IPO,公司需要披露财务报表——通常是两到三年的审计结果——IPO 流程本身会暴露目前公开来源缺失的收入、毛利率和烧钱数据。在此之前,公司仍按私募市场披露规范运作。 首要资本风险是执行缺口:如果 $68 million 的 2025 年收入目标大幅落空,$2.34 billion 投后估值隐含的倍数就会显得偏高,进而可能拖累 IPO 定价。Fabless 芯片公司通常需要 3–5 年收入增长,才能支撑接近软件公司的估值倍数;Rebellions 仍处在这条收入曲线的早期。次要风险包括 REBEL-Quad 量产潜在延迟、HBM3E 供应约束,以及美国市场渗透慢于预期。 [CI031, CI032, CI033, CI034, CI035, CI036]

公开财务缺口与尽调路径
缺失指标对判断的影响具体尽调路径
FY2024 和 FY2025 经审计收入阻断性 — 无法验证 $2.34 B 估值倍数或 2025 目标达成情况向 CFO(Sungkyue Shin)索取经审计财务报表
毛利率(硬件和系统)阻断性 — 无法评估单位经济性或盈利路径要求 COGS 拆分:晶圆、内存(HBM3E)、封装、组装、物流
月度经营现金消耗阻断性 — 无法计算现金跑道或下一轮时点索取最近 6 个季度的管理账
客户数与集中度重大 — 无法评估收入多元化或流失风险披露前三大客户收入占比;活跃客户总数
ARR 或收入运行率(Q1 2026)重大 — IPO 可比公司分析必需获取过去 12 个月收入,供 IPO 前审计师审阅
在手现金(March 31, 2026)重大 — 验证资本充足性和现金跑道说法索取最近一轮融资交割日的资产负债表
芯片出货量(ATOM 和 REBEL-Quad)重大 — 验证新闻稿措辞之外的市场牵引力按地区和客户分群索取出货量
流片和 NRE 成本历史较小 — 帮助理解 R&D 资本化和烧钱构成索取 FY2023–2025 R&D 费用拆分
IPO 准备状态(承销商、交易所、时间表)重大 — IPO 是公司表述的下一次资本事件询问承销商委任、目标交易所(KRX vs. NASDAQ 双重上市)、时间表
债务义务和契约条款较小 — 未披露债务,但任何契约条款都可能限制运营灵活性索取债务明细,以及关键 IP 或设备上的任何质押 / 留置权

所有缺口描述均反映截至 May 2026 Rebellions 已确认缺少公开披露,而非分析不确定性。尽调路径需要直接接触公司管理层;这些缺口均无法仅靠公开来源解决。

[CI037, CI038, CI039, CI040]
FI002: 单位经济桥 — 输入项与证据缺口

梳理评估 Rebellions ATOM 和 REBEL-Quad 芯片线单位经济所需的已知与未知输入项,突出尽调必须补齐的缺口。

除结构关系外,所有节点均未知或未披露。该图是尽调地图,不是财务预测。证据基础:官方新闻稿、Forbes 访谈 (CEO 承认 HBM 短缺)、Samsung Foundry 公开能力页面。

[CI022, CI023, CI024, CI025, CI026, CI031]

4.5 财务缺口、披露限制与尽调阻碍

Rebellions 是韩国私营公司,在提交 IPO 招股说明书之前,收入、毛利率、经营亏损、现金余额、烧钱速度、客户数和 ARR 等重要财务指标没有公开披露义务。公司选择不披露任何财务数据,除韩国公司登记文件中的内容,以及个别高管在媒体采访中分享的信息之外。 这给依赖公开来源的投资人或分析师制造了结构性尽调缺口。最后一个确认收入数字是 FY2023 的 2.7 billion KRW,早于 SAPEON 合并(2024 年 12 月)、REBEL-Quad 发布(2025 年 8 月)、Series C(2025 年 9 月)和 pre-IPO 轮(2026 年 3 月)。公司最关键增长阶段的整整两个财年完全不透明。即便已披露的估值路径(2025 年 9 月 $1.4 billion,2026 年 3 月 $2.34 billion),也无法用外部收入或利润倍数验证。 客户数量和客户集中度缺失是一项具体风险:如果公司只有一两个大客户(韩国的 SK Telecom/Kakao/Naver,或海湾地区的 Saudi Aramco),收入增长可能夸大了可服务市场中的真实牵引。Series C 新闻稿称 ATOM 芯片“支撑韩国最大的商业 AI 服务”但没有点名,暗示单一锚定客户可能主导早期收入。 对尽调而言,填补这些缺口的主要路径是:(1)直接索取 FY2024–2025 审计财务报表;(2)要求披露客户集中度(前 3 大客户收入贡献占比);(3)获得 ATOM 和 REBEL-Quad 芯片出货量;(4)审阅 Samsung Foundry 供应协议,确认 take-or-pay 承诺或最低采购量要求;(5)取得 IPO 准备时间表和承销商选择状态。 [CI037, CI038, CI039, CI040]

财务结论——收入质量、利润率路径与尽调阻断项
维度评估支撑证据阻断级别
收入质量能见度低;合并前 FY2023 是唯一公开数据点FY2023 为 2.7 B KRW;FY2024–2025 未披露阻断投资判断
收入增长轨迹方向上为正但未获确认;37× 目标未验证CEO 提出 2025 年 100 B KRW 目标;尚无确认重大关注项
毛利率路径未知;未披露 COGS无公开数据;硬件优先公司早期通常低于 50%阻断投资判断
资本密集度对无晶圆 AI 芯片公司偏高;NRE 反复投入,并依赖 HBM 供应Samsung Foundry 先进制程;CEO 提及 HBM 短缺重大风险因子
烧钱速度 / 现金跑道烧钱速度未知;现金跑道定性估计约 18–24 个月最新一轮 $400 M;未披露在手现金重大——投资前必须确认
资本充足性(未来 18–24 个月)考虑近期融资,大概率够用;但无法确认累计 $850 M;过去 6 个月 $650 M在假设成立下可控
估值支撑度若收入未确认,估值偏紧;倍数取决于 FY2025 实际数$2.34 B 投后估值,对应收入尚未确认重大关注项
IPO 准备度已完成 Pre-IPO 融资;未披露申报日期、交易所或承销商只有 Pre-IPO 新闻稿措辞不确定性高

本表是对本章全部证据的定性综合评估。“阻断”表示该缺口无法靠公开资料关闭, 必须取得直接数据访问后,才能形成高置信度投资判断。“重大关注项”表示该问题会影响分析, 但可由其他证据部分缓释。

[CI037, CI038, CI039, CI040]

4.6 证据项

Chapter 05

05产品与技术

5.1 产品组合与客户工作流概览

Rebellions 提供两代 AI 推理加速硬件,以加速卡、服务器和机架级系统销售,并由自研软件栈支撑。第一代(ATOM)基于定制推理 NPU SoC,采用成熟制程节点制造,以多 die 配置封装并配备 GDDR6 内存。第二代(REBEL / REBEL-Quad)采用 Samsung 4nm SF4X 制造,使用带 UCIe-Advanced die-to-die 互连的四芯粒架构,并通过 Samsung CoWoS-S 先进封装搭载 HBM3E 内存。产品有四种形态:加速卡(PCIe 接入)、加速服务器(多卡机架式服务器)、mini-POD(多服务器集群)和机架级系统(RebelRack/RebelPOD)。Rebellions 还为低功耗部署提供两款单芯片 ATOM 变体:RBLN-CA21(<75 W,无外接电源接口)和 RBLN-CA22(最高 90 W)。 客户工作流从模型准备开始,使用 RBLN Compiler(RBLN SDK 的一部分)把 PyTorch 或 TensorFlow 模型转换为 Rebellions 运行时格式。模型随后部署在 ATOM-Max 或 RebelServer 硬件上,可直接通过 RBLN Runtime API,也可借助 vllm-rbln Python 包走 vLLM 兼容接口。Kubernetes 集成由 Rebellions NPU Operator 提供;截至 2025 年 12 月,该 Operator 已获 Red Hat OpenShift AI 认证,Helm chart 通过 OCI registry 分发。Rebellions model zoo 列出 300+ 个受支持的 PyTorch/TensorFlow 模型,覆盖 LLM、视觉 transformer 和经典推理工作负载。企业客户的工作负载运行在 ATOM-Max Server(4U、8 卡、Ubuntu/RHEL/AlmaLinux)或 RebelServer(5U、8 卡、AMD EPYC 9355 host CPU、400G 网络)上。机架级客户可使用 ATOM-Max POD(8 服务器 mini-cluster、400 GB/s RDMA fabric)或 2026 年 3 月新发布的 RebelRack/RebelPOD。 [CE001, CE002, CE003, CE004, CE007, CE013]

产品模块与资产矩阵
产品 / SKU代际形态核心算力规格内存部署状态主要用例
ATOM-Max 卡(RBLN-CA25)ATOM 第 1 代PCIe Gen5 x16 卡,FHFL 双槽,350 W算力:128 TFLOPS FP16 / 512 TOPS INT8 / 1024 TOPS INT4内存:64 GB GDDR6,1024 GB/s2024 年上半年出货LLM 推理、云端 AI 服务
ATOM-Max 服务器ATOM 第 1 代4U 机架服务器,8× ATOM-Max 卡合计 1024 TFLOPS FP16内存:512 GB GDDR6,8 TB/s2024 年上半年出货多卡 LLM 推理、企业数据中心
ATOM-Max PODATOM 第 1 代小型集群,8 台服务器、64 个 NPU、400 GB/s RDMA~8 PFLOPS FP16(估计)~4 TB GDDR6(估计)2024 年出货机架规模分布式推理
RBLN-CA21(单芯片)ATOM 第 1 代PCIe,<75 W,无外接供电每颗芯片 32 TFLOPS FP1616 GB GDDR6已出货边缘推理、低功耗
RBLN-CA22(单芯片)ATOM 第 1 代PCIe,最高 90 W每颗芯片 32 TFLOPS FP1616 GB GDDR6已出货中端推理服务器
RebelCard(RBLN-CR 系列)REBEL 第 2 代PCIe Gen5 双 x16每卡最高 2 PFLOPS FP8(估计)内存:144 GB HBM3E,4.8 TB/s截至 2026 年已出货高吞吐 LLM 推理
RebelServerREBEL 第 2 代5U,8× RebelCard,AMD EPYC 9355,典型 4–6 kW最高 2 PFLOPS FP8(8 卡估计)合计 1.15 TB HBM3E(估计)截至 2026 年已出货企业 AI 推理机架单元
RebelRack / RebelPODREBEL 第 2 代机架规模集群,多台服务器多 PFLOPS FP8(机架规模)TB 级 HBM3E(机架规模)2026 年 3 月发布;网站标注“Coming Soon”超大规模 AI 数据中心集群

标注(估计)的算力由单卡规格外推;REBEL 代际的官方单系统 PFLOPS 数据尚未公开披露。 RBLN-CA21/CA22 据 Chips and Cheese SC2024 报道为单芯片 ATOM 卡。 RebelRack/RebelPOD 可用性差异已在 CE015 标出。

[CE001, CE002, CE003, CE004, CE007, CE013]
FE002: 客户部署工作流

工作流来自 RBLN SDK 文档和 vLLM 集成指南。内部 CI/CD 步骤可能因客户环境而异。

[CE016, CE017, CE018, CE020, CE021]

5.2 ATOM 代际:架构、规格与部署

ATOM SoC 是 Rebellions 第一代 AI 推理芯片,2022 年 6 月完成流片,并自 2024 H1 起在 ATOM-Max 产品家族中出货。每颗 ATOM 芯片标称 32 TFLOPS FP16、128 TOPS INT8 和 256 TOPS INT4,配备 16 GB GDDR6 内存,单芯片带宽 256 GB/s(128-bit bus 上 16 Gbps)。旗舰 ATOM-Max 卡(型号 RBLN-CA25)在一个多 die 封装中集成四颗 ATOM 芯片,在 350 W TDP 下提供 128 TFLOPS FP16、512 TOPS INT8 和 1024 TOPS INT4,总计 64 GB GDDR6,聚合内存带宽 1024 GB/s。该卡使用 PCIe Gen5 x16 host interface,占用全高全长双槽形态。Chips and Cheese 在 Supercomputing 2024 独立确认了这一架构。 ATOM-Max Server 在 4U 机箱中集成八张 ATOM-Max 卡,合计 1024 TFLOPS FP16、512 GB GDDR6 和 8 TB/s 总内存带宽。服务器典型功耗 3.4 kW、最高 4.3 kW,落在标准 5 kW 机架供电包络内。服务器运行 Ubuntu、RHEL、AlmaLinux 和 Rocky Linux,并开箱支持 vLLM、Triton、Kubernetes 和 Docker。ATOM-Max POD 扩展到机架级:一个 8 服务器 mini-POD 提供 64 个 NPU、512 GB GDDR6(8 TB/s)和用于节点间通信的 400 GB/s RDMA fabric,可支撑分布式张量并行推理。ATOM 在 ISSCC 2024 与 AMD、Intel 论文同场亮相,并被描述为已具备量产条件。实地部署上,ATOM 卡已部署在阿联酋 ECOPEACE(相对 GPU 基线每瓦性能 2×,TTA 认证)和蒙古海关(相对 GPU TPS/Watt 2.7×,公司引用),公司也报告了沙特阿拉伯和日本部署。 [CE001, CE002, CE003, CE004, CE005, CE006]

各产品层级硬件规格
规格ATOM-Max 卡ATOM-Max 服务器ATOM-Max PODREBEL-Quad 卡RebelServer(8× 卡)
制程节点未披露(ATOM SoC)制程:Samsung SF4X 4nm制程:Samsung SF4X 4nm
架构MDP 内 4× ATOM SoC8× ATOM-Max 卡8 台服务器 / 64 个 NPU4 芯粒 UCIe-Advanced8× RebelCard
峰值算力(FP16)128 TFLOPS1024 TFLOPS~8 PFLOPS(估计)未披露单卡数据~2 PFLOPS FP8(标称)
内存容量64 GB GDDR6512 GB GDDR6~4 TB GDDR6(估计)144 GB HBM3E~1.15 TB HBM3E(估计)
内存带宽1024 GB/s8 TB/s~64 TB/s(估计)4.8 TB/s~38 TB/s(估计)
TDP / 功耗350 W3.4 kW 典型 / 4.3 kW 最大~27 kW(估计)未披露4–6 kW 典型 / 7 kW 最大
主机接口主机接口:PCIe Gen5 x16互连:400 GB/s RDMAPCIe Gen5 双 x164× 400G 网络
已发布基准ISSCC 2024(量产就绪)56.8 TPS LLaMA 70B(ISSCC 2026)

标注(估计)的数值由单卡规格外推;官方集群级规格尚未公开披露。REBEL-Quad 单卡 FP16 TFLOPS 未发布;按官方规格表,FP8 标称值(2 PFLOPS)对应完整 RebelServer 系统。

[CE001, CE002, CE003, CE005, CE009, CE010]

5.3 REBEL 代际:四芯粒架构、HBM3E 与 ISSCC 结果

REBEL Gen 2 芯片在量产形态中命名为 REBEL-Quad,采用 Samsung Foundry 4nm SF4X 制程制造。其定义性架构特征是四芯粒设计:四颗 REBEL compute ASIC 通过 UCIe-Advanced die-to-die 链路互连,单 lane 速率 16 Gbps,IP 由 Alphawave Semi 提供。封装在 Samsung CoWoS-S substrate 上组装,带四组 HBM3E memory stack(每组 36 GB、9.6 GT/s),提供 144 GB HBM3E 总容量和 4.8 TB/s 聚合内存带宽。四个集成硅电容(ISC)完成封装。Host interface 为双 PCIe Gen5 x16。在已出货 AI 芯片中使用 UCIe-Advanced 很值得注意,因为这是首个公开营销的量产 AI 加速器 UCIe 芯粒互连案例,ServeTheHome 在 Hot Chips 2025 的独立评测也予以佐证。 性能数据由 ISSCC 2026 同行评审论文(IEEE Xplore document 11409003)锚定:"A Quad-Chiplet AI SoC with Full-Chip Scalable Mesh Over 16Gb/s UCIe-Advanced Die-to-Die Interface for Large-Scale AI Inferencing." 论文报告,在 LLaMA v3.3 70B 上,输入 2,048 token、输出 2,048 token 序列时达到 56.8 tokens per second。这是目前可信度最高的性能数据点,来自同行评审的 IEEE 会议论文,而非厂商宣传材料。在 Hot Chips 2025(2025 年 8 月),Rebellions 现场演示 LLaMA 3.3 70B 运行,并展示了执行 Qwen3 235B MoE 的能力。RebelServer 系统(5U、8x RebelCard)标称最高 2 PFLOPS FP8,配双 AMD EPYC 9355 host CPU、1.5 TB DDR5 host memory 和 4× 400G 网络端口,典型功耗 4–6 kW、最高 7 kW。 [CE008, CE009, CE010, CE011, CE012, CE013]

FE001: Rebellions 产品架构栈

架构层基于公开文档、ISSCC 论文和 SDK 发布说明。除 ISSCC 论文外,内部 REBEL-Quad 微架构细节未公开披露。

[CE008, CE009, CE011, CE016, CE022, CE033]

5.4 软件生态:RBLN SDK、vLLM 集成与模型库

Rebellions 软件平台(RBLN SDK)在 2026 年 5 月达到 0.10.3 版本,通过公司自有 PyPI mirror(pypi.rbln.ai)分发,也发布在公共 Python Package Index(pypi.org)上。SDK 包含三个包:rebel-compiler==0.10.3(把 PyTorch/TensorFlow 模型转换为 RBLN 二进制格式的 ahead-of-time compiler)、optimum-rbln==0.10.3(用于 LLM 模型准备的 HuggingFace Optimum 集成)和 vllm-rbln==0.10.3.post1(面向 Rebellions NPU 的 vLLM backend plugin)。vllm-rbln 包于 2026 年 5 月 18 日发布在 pypi.org,支持 vLLM v0.18.0。vllm-rbln 已实现 paged attention 和 continuous batching,能够支撑生产级 LLM serving。官方 vLLM 文档把 Rebellions NPU 列为受支持硬件后端,与 Google Cloud TPU、Intel Gaudi、AMD Instinct 和其他加速器并列,使其获得与成熟 NPU 厂商相同的一线集成层级。 SDK 精度支持覆盖 FP32、FP16、FP8、FP6 和 FP4,覆盖领先 LLM 使用的完整范围。Kubernetes 中的设备命名遵循 rebellions.ai/ATOM(用于 RBLN-CA* ATOM 设备)和 rebellions.ai/REBEL(用于 RBLN-CR* REBEL 设备)模式。NPU Operator v0.4.0 于 2025 年 12 月获得 Red Hat OpenShift AI 认证,其 Helm chart 通过 OCI registry 分发,方便企业部署。Rebellions model zoo 包含 300+ 个受支持模型,覆盖 PyTorch 和 TensorFlow 框架。不过,vllm-rbln 仍缺少若干 GPU 对齐功能:speculative decoding、distributed KV cache(支持跨节点内存共享)和 prefill/decode disaggregation(支持异构 serving 架构)被列为开发中,而非已可用。截至 2026 年 5 月,GitHub organization(rebellions-sw)只包含 archived 或 internal repository,没有公开模型代码或 SDK 源码。 [CE016, CE017, CE018, CE019, CE020, CE021]

软件栈与集成架构
层 / 组件包 / 工具版本(2026 年 5 月)功能集成点状态
编译器rebel-compiler0.10.3将 PyTorch/TF 模型转换为 RBLN 二进制格式(预先编译)Python API;CLI生产可用
HuggingFace 集成optimum-rbln0.10.3用于 LLM 模型准备的 HuggingFace Optimum 后端接口:from_pretrained() API生产可用
vLLM 插件vllm-rbln0.10.3.post1让 RBLN NPU 支持分页注意力 + 连续批处理的 vLLM 后端兼容 vLLM v0.18.0生产可用(与 GPU 仍有差距:无推测解码 / 分布式 KV)
运行时RBLN Runtime0.10.3设备端执行引擎;管理 RBLN-CA* 和 RBLN-CR* 设备内核驱动;设备插件生产可用
Kubernetes OperatorRebellions NPU Operator0.4.0在 k8s 中暴露 rebellions.ai/ATOM 和 rebellions.ai/REBEL 设备资源Red Hat OpenShift AI 认证;Helm / OCI生产可用 / 已认证
模型库docs.rbln.ai 模型库N/A支持 300+ 个 PyTorch/TF 模型,覆盖 LLM、视觉和传统推理SDK 安装指南配套资源生产可用

所有版本均通过 pypi.org 发布历史和 docs.rbln.ai 发布说明确认(2026 年 5 月)。按 SDK 文档, vllm-rbln 缺口功能(推测解码、分布式 KV 缓存、prefill / decode 解耦)仍在开发中。

[CE016, CE017, CE018, CE019, CE020, CE021]

5.5 制造、供应链与战略合作

Rebellions 的硅供应链集中在 Samsung。REBEL-Quad 的逻辑采用 Samsung Foundry 4nm SF4X 制程,内存使用 Samsung HBM3E(每个封装四组 36 GB stack),先进封装采用 Samsung CoWoS-S interposer。逻辑、内存、封装三项关键供应输入都由单一供应商承担,构成结构性集中风险:Samsung Foundry 的任何产能中断、HBM3E 良率问题或封装瓶颈,都会直接约束 Rebellions 产品供应,且没有披露替代来源。ATOM-Max 产品家族使用 GDDR6 内存(另行采购),因此 Gen 1 产品线风险较低;但 REBEL-Quad 的完整供应链只有 Samsung。 REBEL-Quad 中的 UCIe-Advanced 芯粒互连 IP 由 Alphawave Semi 提供;这是一家加拿大半导体 IP 公司,专注高速有线连接。Alphawave IP 让 REBEL-Quad 封装内部实现 16 Gbps per-lane die-to-die 带宽。Rebellions 于 2025 年 10 月加入 Arm Total Design 生态,显示其计划把 Arm Neoverse compute subsystem(CSS)集成到未来的 Rebellions AGI CPU 产品中。Arm 也对 Rebellions 的 $250M Series C(2024 年 12 月)进行了战略投资。Red Hat 于 2025 年 12 月通过 OpenShift AI 认证成为生态伙伴,支持企业在 Rebellions 硬件上进行 Kubernetes 部署。SK Telecom 参与 2026 年 4 月宣布的主权 AI 数据中心验证和联合开发。Samsung 既是共同投资方(通过 Samsung Securities),也是 foundry,供应链与战略资本高度一致,但议价筹码集中在单一交易对手手中。 [CE026, CE027, CE028, CE029, CE030, CE040]

供应链、制造与合作伙伴依赖
供应投入供应商 / 合作伙伴范围集中度风险战略一致性
逻辑芯片制造(4nm)Samsung Foundry(SF4X)REBEL-Quad 计算芯粒唯一来源高——未披露备选晶圆厂;未使用 TSMCSamsung 同时是战略投资方和 HBM 供应商
HBM3E 内存SamsungREBEL-Quad 4× 36 GB HBM3E 堆叠唯一来源高——公开资料显示未引入 SK Hynix 或 Micron据报道,Samsung HBM3E 首个客户供货对象是 Rebellions
先进封装(CoWoS-S)Samsung(CoWoS-S 中介层)REBEL-Quad 封装基板,承载 4 个芯粒 + 4 组 HBM3E高——仅 Samsung 封装栈与 Samsung 的晶圆代工和 HBM 供应整合
芯粒互连 IP(UCIe-Advanced)Alphawave Semi芯粒间 UCIe-Advanced SerDes IP,16 Gbps/lane中——UCIe IP 来自成熟供应商;架构可迁移无股权关系;商业 IP 授权
CPU 架构(计划中)伙伴:Arm(Total Design / Neoverse CSS)计划中 AGI CPU 产品的 Arm CPU 子系统低(未来产品)——Arm 也是战略投资方Arm 投资 Rebellions Series C;2025 年 10 月加入 Total Design
企业 k8s 认证Red Hat(OpenShift AI)面向 OpenShift AI 平台的 NPU Operator v0.4.0 认证低——认证带来价值增量,不受供应约束2025 年 12 月战略合作;扩大企业 GTM

供应链数据来自官方产品页、新闻稿和独立硬件评测(ServeTheHome、Chips and Cheese)。 Samsung 在逻辑芯片、HBM 和封装上形成单一供应商风险,是主要供应链脆弱点。

[CE026, CE027, CE028, CE029, CE040]
FE003: 关键技术依赖图谱

依赖数据来自产品页面、ISSCC 论文、新闻稿和独立硬件评测。内部供应合同条款未公开披露。

[CE026, CE027, CE028, CE029, CE031]

5.6 性能证据、基准缺口与技术风险

Rebellions 硬件最强的独立性能证据,是 REBEL-Quad 的 ISSCC 2026 同行评审论文(LLaMA v3.3 70B、2k/2k sequence 下 56.8 TPS),以及包括 ServeTheHome 在内的行业媒体观察到的 Hot Chips 2025 现场演示。对 ATOM Gen 1 来说,所有每瓦性能说法都来自公司引用或 TTA(Telecommunications Technology Association,韩国认证机构)认证:蒙古海关相对 GPU 的 2.7× TPS/Watt 优势,以及阿联酋 ECOPEACE 的 2× 每瓦性能,均由 Rebellions 呈现,没有独立公开的基准方法。截至 2026 年 5 月,Rebellions 没有在 mlcommons.org 向 MLPerf Inference Datacenter benchmark 提交结果。MLPerf 是行业最主要的第三方推理基准,所有主要竞争对手(NVIDIA、AMD、Google、Intel)都会发布。缺少 MLPerf 数据,使外部无法基于公开来源独立比较 Rebellions 与 H100、MI300X、Gaudi 3 的性能。 REBEL-Quad 的双 PCIe Gen5 x16 host interface 可能落后于 NVIDIA GB300;后者目标是 PCIe Gen6 连接。如果客户工作负载卡在 host interface,二者会形成代际 I/O 带宽差异。软件端,vllm-rbln 缺少 speculative decoding、distributed KV cache 和 prefill/decode disaggregation——这些已经是大规模 GPU vLLM 部署中的标准功能,用来支持成本优化的混合 serving 架构。RebelRack 和 RebelPOD 系统在 2026 年 3 月 30 日新闻稿和投资人沟通中被宣布为“现已可用”,但截至 2026 年 5 月研究日期,Rebellions 网站导航仍显示这些产品“即将推出”。这种不一致是生产就绪度的反向信号:要么产品在供应确认前就被宣布,要么网站没有更新可用状态。RSD(Rebellions Scalable Design)框架支持从单卡(128 TFLOPS FP16 / 64 GB)到可配置机架系统(512–7168 TFLOPS FP16,256 GB– 3.5 TB GDDR6)的张量并行推理,并支持 Ubuntu、RHEL、Kubernetes 和 OpenStack。 [CE010, CE015, CE024, CE025, CE031, CE033]

软件路线图与功能状态
功能 / 里程碑状态(2026 年 5 月)影响依赖 / 风险时间窗口
pypi.org 上的 vllm-rbln v0.10.3.post12026 年 5 月 18 日发布将开发者触达从自有 PyPI 镜像扩展出去;安装路径标准化vLLM 社区接受度;与 GPU vLLM 的版本一致性当前——已发布
vllm-rbln 中的推测解码开发中——尚不可用降低自回归 LLM 解码的单 token 延迟;GPU 服务中已是标准能力RBLN NPU 上的推测执行架构支持2026 年下半年(估计)
分布式 KV 缓存开发中——尚不可用支持跨节点共享 KV 缓存;通过 prefill 卸载服务超大模型时必需多节点 RBLN 运行时;RDMA fabric 集成2026 年下半年或更晚(估计)
Prefill / decode 解耦开发中——尚不可用支持异构服务架构(prefill 与 decode 节点分离);规模化时优化成本vLLM 解耦协议;多节点 RBLN 运行时2026–2027(估计)
RebelRack / RebelPOD2026 年 3 月宣布“available now”;2026 年 5 月网站仍写“Coming Soon”面向超大规模 AI 工作负载的机架规模 REBEL-Quad 集群REBEL-Quad 芯粒量产爬坡;供应链就绪状态不清——活跃尽调项
AGI CPU(Arm Neoverse CSS 集成)流片前——2025 年 10 月宣布合作Rebellions 设计的 AI 计算 + Arm 服务器 CPU 放进同一平台;区别于纯 NPUArm Total Design 成员资格;未披露硅片流片时间2027+(估计)

开发中功能的路线图时间由分析师估计;Rebellions 尚未发布带承诺日期的公开 SDK 路线图。 RebelRack/POD 在新闻稿与网站之间的状态差异,是尚未解决的反向信号。

[CE015, CE017, CE019, CE021, CE028]
FE004: 产品成熟度与能力评估

成熟度评级是分析师判断,依据截至 2026 年 5 月公开产品数据、经同行评审的基准测试和独立评测。

[CE001, CE010, CE024, CE025, CE031, CE037]

5.7 证据项

Chapter 06

06客户情况

6.1 客户基础与细分结构

理解 Rebellions 客户故事,按买方工作流看比简单数 logo 更清楚。官方解决方案和产品页显示,公司把推理基础设施卖给电信、主权 AI、企业 AI 和数据中心环境;经济买方通常是云或基础设施运营商,而不是应用团队。公开记录随后分成三层证据。第一层是韩国 KT/kt cloud 和 SK Telecom 的直接生产证据。第二层是第三方报道的其他韩国云部门客户名单,以及与沙特相关的需求。第三层是覆盖日本、美国和泰国的更广但更薄的扩张证据。 这一区分很重要,因为公开客户深度并不均衡。KT 和 SKT 有公司直接声明或客户邻近证据支撑;Kakao、Naver、Saudi Aramco 以及美国 / 日本 / 泰国地域,主要由 Forbes 和公司融资公告支撑。这足以判断 Rebellions 已经不是单一客户故事,但还不足以高置信度承保其广泛商业覆盖。[CU001, CU002, CU003, CU004, CU006, CU012]

客户分群表
分群买方 / 用户 / 付款方公开具名证明阶段战略价值缺口
韩国电信云锚定客户KT/kt cloud 与 SK Telecom 的云和 AI 基础设施团队Series B 首批客户表述、KT 数据中心交付报道、SKT A. 部署引述生产公开证据质量最高,且大概率构成早期核心收入基础未披露收入拆分、合同规模和续约条款
韩国云与互联网平台SK、Kakao 和 Naver 的云部门与平台运营方Forbes 蓝筹客户名单第三方报道的名单证明显示其覆盖范围不止两家电信公司没有客户方表述或工作负载细节
沙特 / 中东主权 AI 买方与沙特生态相关的能源、主权 AI 和数据中心运营方Forbes 报道的首个 Saudi Aramco 交易,加上 Wa'ed 桥头堡叙事接近生产 / 部分部署来自报道验证 GCC 滩头阵地和主权 AI 相关性无客户方案例、合同金额或利用率数据
日本验证路径日本电信和创新团队DOCOMO Innovations MOU、日本电信会议和 Series C 地域扩张说法从验证到早期商业信号韩国以外的重要扩张向量未披露日本具名付费生产客户
美国企业和 AI 实验室管线企业基础设施买家、超大规模云厂商和大语言模型实验室Series C 轮美国部署表述、Forbes 所称美国订单,以及美国扩张推进管线 / 部分未具名部署一旦转化,TAM 大、战略可信度高尚未披露已投产的具名美国客户
渠道牵引的企业采购平台团队借助生态和合作伙伴路径采购Red Hat OpenShift AI 发布和合作伙伴覆盖渠道赋能降低韩国以外的采购摩擦已转化终端客户数未公开

公开客群按买家工作流和证据质量分组,而不是按已披露收入占比分组;阶段标签用于区分生产证据、第三方名单提及和伙伴牵引的管线证据。

[CU001, CU002, CU003, CU006, CU012, CU014]
客户增长 / 采用轨迹表
里程碑公开证据日期分类含义限制
KT 和 kt cloud 被列为领投方及首批客户Rebellions Series B 轮新闻稿2024-05-01生产证据ATOM 在数据中心场景最早的清晰商业锚点首批客户表述没有披露出货量或商业条款
KT 累计投资和数据中心交付见诸报道Korea JoongAng Daily 称 KT 累计投资 66.5B 韩元,并收到用于云服务的 ATOM 芯片2024-05-07生产证据证实投资方与客户重合,也证实上线部署意图独立交易经济性仍不清楚
Wa'ed 支持的中东桥头堡公布Series B 延伸轮新闻稿把沙特关系定位为商业扩张路径2024-07-01扩张助推在后续商业表述前,先打开沙特生态入口投资本身不等同于经常性客户收入
Forbes 商业快照扩充客户名单Forbes 点名 SK、Kakao、Naver 云部门、首个 Saudi Aramco 交易,以及美国 / 日本 / 泰国订单2025-04-14第三方商业快照显示客户故事不只发生在韩国多个账户仍未具名,或缺少客户侧证言支撑
SKT 和 DOCOMO Innovations 公告Rebellions-SKT-DOCOMO 新闻稿聚焦下一代 AI 基础设施合作2025-04-30验证证据通过知名电信创新机构,显示面向日本的评估路径公告尚未点名付费生产账户
Series C 轮地域披露Rebellions 称 ATOM 已部署给日本、沙特和美国客户2025-09-30生产与验证混合证据按地域证实国际牵引力客户名称和账户规模大多未披露
Red Hat 企业渠道发布关于搭载 Rebellions NPU 的 OpenShift AI 的联合与独立报道日期:2025-12-10 to 2025-12-11渠道证据改善直销之外的企业采购路径未披露来自该渠道的已转化客户数
Pre-IPO 发布和 2026 年主权 AI 公告Pre-IPO 新闻稿加上 SKT/Arm/Rebellions 报道日期:2026-03-30 to 2026-04-10扩张证据支撑企业与政府客户扩张叙事仍看不到 RebelRack 或 RebelPOD 的具名客户交付

该轨迹表使用有日期的公开里程碑,而不是真正的客户漏斗,因为 Rebellions 没有发布活跃客户数、部署数或经常性收入队列。

[CU003, CU004, CU006, CU020, CU008, CU014]
FU001: 客户旅程图

Rebellions 通常先靠基础设施验证拿下机会,再把证明转化为伙伴带动的企业扩张。

[CU003, CU007, CU008, CU015, CU024, CU039]

6.2 具名客户证据与证据分层

最强的具名客户证据仍在韩国。Rebellions 在 Series B 新闻稿中称 KT 和 kt cloud 是领投方和首批客户,Korea JoongAng Daily 另行报道 Rebellions 已向 KT 数据中心交付 ATOM 芯片,用于运行云服务。Rebellions 自己的 about page 又给出了最清晰的 SK Telecom 生产表述,引用 SKT 的 Tony Ha 称 SKT 正在 A. 中部署 Rebellions NPU;该页面称 A. 是韩国最大的 LLM 服务。 离开这些锚点后,客户名单就没那么直接,必须谨慎分层。Forbes 报道了 SK、Kakao 和 Naver 的蓝筹云部门客户、首个 Saudi Aramco 交易,以及截至 2024 年底来自美国、日本和泰国的订单。Rebellions 2025 年和 2026 年公告随后加入了地域级部署表述,但多数非韩国客户名称仍未披露。另一个层面,Rebellions/SKT/DOCOMO 和 SKT/Arm/Rebellions 公告是有价值的验证证据,但不能把它们误读成广泛、具名、付费客户名单。[CU003, CU004, CU006, CU007, CU008, CU011]

具名客户证据表
客户细分领域公开证据生产还是试点结果或商业信号局限
KT / kt cloud 客户韩国电信云Series B 轮称其为首批客户;Korea JoongAng 称 ATOM 芯片已提供给 KT 数据中心生产ATOM 在数据中心场景最早的具名锚定客户证据未披露出货量、收入贡献或续约状态
SK Telecom韩国电信 / AI 服务商Rebellions 关于页面引用 SKT:SKT 在 A. 部署该 NPU生产韩国线上大语言模型服务中的具名工作负载未公开合同金额、期限或扩张节奏
Saudi Aramco 相关沙特部署中东主权 / 能源生态Forbes 称首个 Saudi Aramco 交易;Wa'ed 和 Wamda 强化沙特桥头堡与部署叙事接近生产 / 可能已部署公开材料中最强的中东商业证据仍缺客户侧案例或使用指标
Kakao 和 Naver 云部门韩国互联网 / 云平台Forbes 将 SK、Kakao、Naver 的云部门列为客户第三方报道显示国内客户名单不止 KT 和 SKT无客户侧声明、无具体工作负载,也无运营方直接背书
DOCOMO Innovations / 日本电信验证日本创新机构Rebellions-SKT-DOCOMO 新闻稿,加上 Korea JoongAng 关于日本电信会谈的报道验证通过知名电信创新实体,显示可信的日本切入口未披露为付费生产客户

这是截至 2026-05-20 的部分公开枚举,仅限客户侧证据中讨论过的具名账户或具名实体;各行有意区分生产证据、验证证据和第三方名单提及。

[CU003, CU004, CU006, CU007, CU012, CU013]
生产 vs 验证 vs 管线表
证据类别账户或地域来源依据商业解读仍缺什么
生产证据KT / kt cloud 客户Series B 轮新闻稿加 Korea JoongAng Daily早期付费且已部署锚定账户的最佳公开证据单位数量、收入贡献和合同期限
生产证据SK TelecomRebellions 关于页面推荐语韩国的具名线上工作负载定价、扩张节奏和续约细节
第三方名单证据SK、Kakao 和 Naver 云部门Forbes有助于证明客户广度,但弱于客户侧案例运营方引述和工作负载细节
验证 / 联合设计DOCOMO Innovations、SKT 和 Arm2025 和 2026 年合作公告评估信号强,也带来生态入口采购订单、具名上线或经常性收入披露
地域管线美国、日本和泰国Forbes 加 Series C 轮显示韩国以外存在需求,也有部分部署表述具名客户、日期和生产状态
渠道证据Red Hat 企业生态联合发布和合作伙伴报道改善采购路径,但本身不是终端客户订单已转化部署和渠道归因收入

这张分类表用于防止过度解读新闻稿;它把直接生产证据同合作伙伴验证、名单提及和仍未具名的管线证据拆开。

[CU003, CU006, CU007, CU012, CU014, CU015]
FU003: 客户证据矩阵

证据质量在韩国电信锚定客户处最高;国际覆盖、续约和新产品交付披露最弱。

[CU012, CU013, CU014, CU015, CU017, CU024]

6.3 渠道与扩张回路

Rebellions 并不只依赖韩国企业直销。公司围绕获客搭建了渠道和生态层,最显眼的是 Red Hat、Marshall Choy 的商务负责人招聘,以及美国实体设立。Red Hat OpenShift AI 发布很重要,因为它给 Rebellions 一条更适合采购的企业路径:潜在客户不必单独审查一整套硬件栈,而是可以通过成熟企业 AI 平台和渠道关系接触产品。独立行业媒体也报道了该公告,提高了外界对发布不只是单家公司新闻稿的信心。 扩张回路也经过主权和电信基础设施合作。Wa'ed 支持的中东桥头堡、SKT/Arm 主权 AI 合作,以及面向日本的 DOCOMO Innovations 验证,都会创造客户邻近入口。不过,这些机制大多是打开路径的证据,不是已披露的重复收入证明。商业上行空间真实存在,但投资人需要把伙伴杠杆和已验证终端客户韧性分开看。[CU008, CU009, CU010, CU011, CU020, CU022]

扩张与集中度风险表
驱动因素或风险公开证据影响当前判断尽调路径
韩国电信客户集中KT/kt cloud 和 SKT 是最清晰的生产锚点即便确切收入结构未披露,重大集中度风险也可能存在要求提供前五大客户收入、毛利率和账户敞口
国际订单名称仍很少Forbes 和 Series C 轮提到美国、日本和泰国需求,但名称有限削弱对韩国以外客户广度的信心要求提供含生产、试点和管线状态的具名客户清单
验证容易被误读成收入DOCOMO/SKT/Arm 等公告证明生态,不自动等同于终端客户订单必须保守划分合作伙伴公告要求提供每项合作从管线转生产的数据
Red Hat 渠道杠杆Red Hat 打开更适合企业采购的路径正向上行空间可信,但仍处早期要求提供经 Red Hat 生态导入的来源管线和已成交订单
RebelRack 和 RebelPOD 交付不透明Pre-IPO 发布点名新产品,但没有点名客户交付新基础设施产品尚未获得公开客户验证要求提供首批装机客户背书和发货日期
沙特桥头堡可信度Wa'ed、Forbes 和 Wamda 支撑沙特商业路径正向桥头堡看起来可信,但终端用途细节仍薄要求沙特账户提供运营方引述、站点背书或部署 KPI

该表同时列出上行驱动和下行风险,因为二者共同决定今天的客户证据能否沉淀为多元化装机基础。

[CU020, CU022, CU024, CU025, CU029, CU030]
FU002: 采用与部署流程

公开记录显示,路径从基础设施需求出发,经过验证、渠道支持、生产落地,再走向韩国之外仍待证明的覆盖广度。

[CU001, CU020, CU022, CU023, CU024, CU025]

6.4 留存韧性与集中度风险

到留存韧性这里,公开记录明显变弱。本章没有找到公开披露的客户数、NRR 或 GRR、流失率、合同期限披露,也没有账户级扩张数据。这不意味着客户基础弱;而是说承保负担从公开指标转向定性判断。公开来源足以证明真实采用,但不足以证明这种采用有多粘、多分散。 最重要的风险是集中度。KT/kt cloud 和 SK Telecom 是最清晰的生产锚点;国际需求则常通过未具名订单、地域级部署主张或伙伴主导的 go-to-market 表达。Forbes 和 Korea JoongAng Daily 都强化了反向情景:从 Nvidia 手中赢走数据中心买方很难,Rebellions 仍未披露证明广度所需的分母数据。实际判断是:客户故事有可信旗舰账户,也有可信扩张向量,但不透明度仍然太高,无法排除集中度和续约风险。[CU017, CU018, CU033, CU034, CU035, CU036]

留存 / 重复使用 / 满意度表
指标公开数值细分置信度说明了什么尽调问题
付费客户总数全公司客户基数没有公开分母要求按地域和细分披露活跃付费客户数
净留存率(NRR)全公司未公开披露收入留存要求提供过去八个季度 NRR 及定义
总留存率(GRR)/ 客户流失全公司未公开披露流失要求按企业客户队列提供年度 GRR 和客户流失
合同期限 / 续约节奏具名企业账户看不到公开的合同期限或续约时点要求锚定账户的条款清单或匿名合同期限分布
重复工作负载证据2024–2026 年资料反复提到 KT/SKT,显示一定持续性韩国电信锚定客户暗示客户在最初公告后并未消失要求按账户提供历年收入和续约里程碑
国际耐久性已披露日本、沙特和美国地域,但名称仍稀疏国际账户显示确有扩张,但无法判断是否重复要求提供韩国以外的具名上线账户和上线日期

空值表示所审阅来源未公开披露该指标;唯一可见的耐久性证据是锚定账户被反复提及,其证明力弱于队列或合同数据。

[CU017, CU018, CU033, CU034, CU035, CU036]
Chapter 07

07风险

7.1 按严重度排序的风险画像与证据缺口

Rebellions 的风险栈首要来自证据、收入和商业化缺口,而不是技术野心不足。公开技术证据显示 REBEL-Quad 能运行前沿规模模型,官方发布也描述了机架级系统,但公司仍缺少公募市场投资人在对标 Nvidia 或 AMD 承保时会期待的独立基准包。MLPerf Datacenter 没有 Rebellions 提交结果,Forbes 和 Korea JoongAng Daily 也都把非 Nvidia 采购描述为异常困难。这个证据缺口很重要,因为估值野心已经跑在公开收入披露前面:审阅来源中唯一硬收入基线是 FY2023 收入约 2.7 billion KRW,而 Park 的 2025 年 100 billion KRW 收入目标和约 $2.34 billion 的 pre-IPO 估值,并没有由经审计 FY2025 数据连接起来。更广的行业背景也不宽容。Graphcore 最终被 SoftBank 收购,SambaNova 据报道曾探索出售,FuriosaAI 仍在韩国境内融资。合起来看,基准不透明、收入执行和生存压力,是前三大剩余风险。[CR001, CR002, CR003, CR004, CR005, CR006]

运营 / 质量 / 安全风险登记表
失效模式可能性严重性缓释成熟度剩余敞口未解决缺口
没有与 Nvidia / AMD 的独立同口径基准测试关键IPO 叙事依赖性能主张时,敞口极高需要第三方基准测试或客户侧部署规模结果
相比上次披露的 FY2023 基线,收入执行存在缺口关键中低高,因为公开估值已经跑在披露收入前面需要经审计的 FY2024-FY2025 桥接和在手订单转化证据
晶圆代工或 HBM 分配不足中高高,因为路线图和封装仍以 Samsung 为中心需要产能承诺、缓冲库存和第二供应源路线图
未公开披露出货量或生产量高,因为投资者无法检验经营杠杆或产品发布后的牵引力需要披露出货量、积压订单和利用率
RebelRack / RebelPOD 从发布到可供货的时间不清中高中高,直到能看到独立部署需要装机基数证明、客户背书和交付节奏
非 Nvidia 采购摩擦拖慢国际转化高,因为公司仍在对抗根深蒂固的采购流程需要客户销售周期数据和独立的切换成本证据

可能性和严重程度基于已审阅的公开证据排序,而不是内部经营数据;公开记录仍缺少生产量、利用率和毛利率可见度。

[CR001, CR002, CR003, CR004, CR005, CR006]
FR001: 风险热力图

基于公开证据而非内部风险评分,对本章主要残余风险做可能性—影响矩阵。

[CR001, CR004, CR009, CR025, CR033, CR043]

7.2 监管、上市与政策暴露

这里的法律和监管风险主要是暴露与披露风险,不是已知的主动诉讼。Rebellions 当前扩张叙事横跨韩国、沙特阿拉伯、日本和美国;本章审阅的 2025–2026 年法律与监管材料都指向同一方向:先进计算和 AI 模型出口规则仍在生效、演进,并且对芯片供应商很重要。律师事务所警报把近期 BIS 动作描述为范围很广,CSIS 则强调盟友执行不均,可能让通过伙伴分销更复杂。即使 Rebellions 本身没有被公开指控违规,缺少公开分类函或筛查披露,也意味着投资人无法验证合规准备度。IPO 风险也在同一类。Seoul Economic Daily 报道目标是在 2026 年 8 月进行 KOSPI 初步审查,但审阅来源没有找到初步批准、招股说明书或经审计公开申报文件。由于 National Growth Fund 支持和 K-Nvidia 叙事把融资与政策叙事绑定,披露缺口可能同时变成监管和声誉问题。[CR012, CR013, CR014, CR015, CR016, CR017]

监管 / 法律风险登记表
规则 / 案件 / 敞口司法辖区状态可能性严重性关键缓释剩余敞口尽调路径
先进计算 / AI 扩散出口分类美国加跨境销售规则已生效;产品分类未公开披露中高关键出口律师加目的地筛查分类和客户路径得到证明前,敞口仍高获取 ECCN 或分类备忘录、目的地矩阵和经销商控制
最终用途和主权 AI 目的地筛查美国加合作伙伴司法辖区全球扩张已公开;筛查控制未公开使用可信渠道和最终用途认证中高,因为主权部署提高敏感度审查经销商 KYC、终端用户限制和拒绝方工作流
KRX / KOSPI 上市披露准备韩国IPO 目标见报道;未审阅批准文件或招股书Pre-IPO 资本和顾问争取时间初步审查通过、审计报表出现前,敞口仍高要求提供初步审查提交材料、审计师时间表和申报清单
并购后披露与控制整合韩国SAPEON 合并完成;控制框架未公开单一 CEO 和新资金可支持整合工作中高,因为控制缺口常在上市时暴露审查整合 PMO、内控负责人图谱和董事会汇报节奏
知识产权 / 专利自由实施悬而未决美国 / 全球未发现公开案件中低假设已有标准 FTO 和许可流程中,因为没有公开 FTO 披露要求提供专利律师备忘录和关键第三方知识产权许可
政府项目附加条件 / K-Nvidia 叙事韩国National Growth Fund 直接支持已确认中高国内政策背书可能支持融资渠道中,因为政策预期会压缩战略自由度审查基金契约、报告义务和本土上市预期

覆盖范围有限,因为该登记表只纳入公开来源可见的监管和法律敞口;非公开律师备忘录、私下合同和内控材料无法审阅。

[CR013, CR014, CR015, CR016, CR017, CR018]
FR002: 风险传导图

主要风险事件及其向收入、融资和估值结果级联的路径。

[CR004, CR010, CR013, CR027, CR033, CR044]

7.3 供应链与生态依赖

Rebellions 正试图扩张并走向公开上市,但运营模式的集中度异常高。Samsung 不只是供应商,也是路线图锚点:下一代 REBEL 项目宣布基于 Samsung 4nm 和 Samsung HBM3E,独立 Hot Chips 报道也仍显示 REBEL-Quad 集中在 Samsung 制程和 HBM 上。Park 对 foundry 产能和 HBM 短缺的表述,把这种依赖变成一阶风险,而不是泛泛的半导体 caveat。需求和渠道侧也有同样集中度。SK Telecom 和 Arm 提供了醒目的主权 AI 验证,但也把商业化绑定到一小组政策对齐项目和交易对手。公开市场时点又是另一项依赖,而不只是一个里程碑,因为 KRX 和韩国 IPO 窗口现在夹在 Rebellions 与下一次估值重置之间。没有披露第二供应来源计划、产能预留或出货量时,内存分配、封装或主权项目时点的任何滑坡,都可能传导为交付延迟、基准变弱,以及更脆弱的 IPO 故事。[CR013, CR022, CR025, CR026, CR027, CR028]

合作伙伴 / 依赖风险清单
依赖项对手方角色集中度失效情景严重程度缓释措施残余敞口
晶圆制造和先进存储Samsung Foundry / Samsung HBM3EREBEL 系列产品的路线图锚点制程、封装或 HBM 配额延误,拖慢出货关键除战略合作外,未公开披露其他缓释措施
出口管制把关BIS 及盟友出口主管机关敏感目的地的市场准入许可分类或筛查失误会阻断主权部署关键法律审查和渠道纪律只是隐含存在,尚无证据
旗舰主权 AI 设计定点SK Telecom / Arm 生态样板案例、渠道可信度和部署背书方象征意义高度集中试点或路线图延误会同时削弱收入叙事和可信度战略伙伴能改善准入,但也收窄依赖面
公开市场时点KRX / KOSPI 审查流程IPO 执行的外部门槛单一上市地集中审查延误或估值重置会压缩融资可选性IPO 前的大额现金缓冲能争取时间中高
政策资本Korea National Growth Fund本土资本支持和国家冠军信号重要但不是唯一资本来源政策优先级转向,或报告负担上升中高共同投资者基础较广
技术生态杠杆Arm 及其他战略栈伙伴参考架构和市场准入可信度适中路线图或生态支持落地慢于预期Series C 轮扩大了投资者基础

本清单聚焦能改变收入、供应、融资或市场准入的依赖项;其中几项是战略关系,不是传统供应商,所以集中度用定性方式描述。

[CR013, CR022, CR025, CR026, CR027, CR028]
FR003: 依赖图谱

关键外部依赖横跨供应商、监管方、市场窗口和旗舰商业赞助方。

[CR013, CR022, CR025, CR036, CR047, CR048]

7.4 整合、领导力与否决条件

SAPEON 合并通过打造更大的韩国 AI 芯片冠军,解决了一个战略问题,但也增加了整合、治理和领导力集中度。合并后的公司在公开层面与 Sunghyun Park 深度绑定;他仍是具名 CEO,也是主要技术和外部代表。公开资料没有找到继任计划、协同评分卡或合并后治理框架。这很重要,因为未来十二个月挤满了执行事件:证明基准可信度,把 pre-IPO 资金转化为可审计增长,处理出口管制暴露,并管理 IPO 流程。正确的投资人姿态不是假设失败,而是坚持可衡量的投资逻辑失效触发条件。如果经审计 FY2025 结果远低于目标,如果 KRX 审查显著滑坡,如果 Samsung 或 HBM 分配约束发布量,或者 Park 或合并整合让领导团队不稳,剩余风险画像会很快从高风险变成投资逻辑失效。资本实力存在,但证据和控制仍落后于野心。[CR030, CR031, CR032, CR033, CR034, CR035]

人员 / 执行风险清单
角色 / 职能依赖或缺口可能性严重程度缓释措施尽调路径
CEO / 公开技术代表(Sunghyun Park)Park 是合并后公司公开具名的领导者,也是主要对外发言人关键董事会监督和更厚的管理梯队可能存在,但公开材料没有描述审查继任计划、高管梯队和授权经营权限
合并后整合领导力SAPEON 整合评分卡和重叠岗位削减情况未公开最近融资可为整合工作提供资金索取协同跟踪表、组织架构图和核心高管留任数据
财务 / 披露准备度虽有 IPO 传闻,但审计后的 FY2024-FY2025 桥接数据尚未公开IPO 前融资降低了短期紧迫性索取审计状态、关账日程和披露控制情况
国际企业客户拓展韩国以外赢单仍会撞上 Nvidia 采购锁定效应战略伙伴和主权试点有帮助索取销售周期数据、韩国以外管线和可背书客户
基准测试和发布计划管理独立证明、出货量和产品可得性仍然薄弱中高Hot Chips 和 ISSCC 提供了技术可见度索取外部基准测试计划、客户试点和发布期出货量证据

这里的人员风险,与其说是创始人魅力,不如说是在合并叠加 IPO 周期中,执行、披露和技术可信度都集中在少数可见领导者身上。

[CR030, CR031, CR032, CR033, CR034, CR035]
缓释与否决标准表
风险可监控触发项阈值 / 事件行动含义
基准测试证明缺口独立基准测试或参与 MLPerf首次公开申报时仍没有第三方基准测试包将其视为估值支撑的投资逻辑破坏点;追加资本前要求客户侧证明
收入 / 审计披露缺口经审计的 FY2025 收入和毛利率披露结果仍不披露,或明显低于管理层叙事重定承销假设,推迟任何公开市场式估值
出口管制 / 市场准入风险分类、许可或筛查事件新规、许可被拒或出货暂停影响目标市场暂停扩张假设,并要求法律顾问背书的补救方案
Samsung / HBM 集中度产能分配和交付节奏重大发布延误或供应约束让系统交付推迟超过一个季度假设收入转化更慢、营运资本压力更高
IPO 流程风险KRX 初步审查和估值口径审查明显延期,或预期估值下调超过 25%从 IPO 可选性框架切换到融资风险框架
领导层 / 整合风险高管流失和合并里程碑审计桥接数据公开前,Park 离职或核心高管反复流失升级为治理红旗,重新评估执行信心

阈值是可监控的代理指标,不是合同契约;设计目的,是把今天的叙事风险变成未来投资者或刷新研究时的明确尽调关口。

[CR009, CR014, CR029, CR044, CR046, CR048]

7.5 证据项

Chapter 08

08估值

8.1 投资逻辑、反向逻辑与建议

Rebellions 值得进入严肃投资讨论,不是因为它又是一家同质化 AI 芯片初创公司。作为韩国半导体创业公司,它融资规模异常大,与 Samsung 的制造体系深度绑定,也站在主权 AI 叙事里,因此本土需求路径比许多西方推理芯片同行更受保护。2026 年 3 月官方 Pre-IPO 公告还显示,公司正从单卡硅片走向 RebelRack、RebelPOD 这类系统级产品;这一点有战略意义,因为客户买的是可部署的推理容量,不是独立裸片。 但在当前价格下,反向逻辑更关键。公开披露的最近收入基线仍只是 FY2023 的 2.7 billion KRW,且净亏损 13.7 billion KRW;本章审阅的公开记录仍未披露 FY2024 或 FY2025 实际收入、毛利率、积压订单、客户数或复购指标。也就是说,当前估值不是由可见商业规模支撑,而是由可选性支撑。估值已经计入成功商业化和最终 IPO 路径,因此正确立场是继续研究,不是买入。审计级或招股书级财务披露追上融资故事之前,置信度应保持低、风险保持高,入场纪律也要严格。 [CV001, CV002, CV005, CV011, CV012, CV018]

投资建议摘要表
维度评估置信度决策含义
投资建议继续研究不要在没有新审计财务披露的情况下,把 2026 年 3 月估值视为清晰买点。
风险评级如有仓位,仓位规模也应假设 IPO 前披露风险和商业化风险都带有二元性。
估值立场偏高当前价格已经计入重大执行进展,而这些进展尚未公开披露。
入场纪律更偏好更低入场价或更强证据只有在经审计收入、积压订单和毛利率证据改善,或价格重置后再做投资测算。
目标回报 / 持有逻辑当前估值下,若没有乐观情景证据,很难证明 >2x 回报合理现在要拿到风投式回报,大概率需要更低价格,或明显更强的商业化证明组合。
最重要的上调触发项经审计的 FY2024-FY2025 规模,加上可信的 IPO 申报只有财务披露开始验证战略叙事时,才上调判断。

建议基于截至 2026-05-20 的公开证据;当前融资支持很强,但经审计的经营支撑仍然缺位。

[CV043, CV044, CV045, CV046, CV047]
投资逻辑 / 反向逻辑表
投资逻辑支撑程度反向逻辑改变判断的证据
主权 AI 定位让 Rebellions 拿到一个受保护的韩国需求切入口,许多西方同行没有这个位置。主权 AI 需求具备战略意义,但不等于已经披露且可重复的收入。披露与主权或电信部署绑定的经审计国内合同收入和积压订单。
接入 Samsung 生态改善可制造性和战略可信度。制造资源本身不能证明商业需求,也不能证明健康的单位经济性。按产品代际提供出货规模、重复订单和毛利率的第三方证据。
RebelRack、RebelPOD 等系统级发布把变现面从单颗芯片拓宽到系统。发布公告没有量化客户转化、利用率或定价权。具名生产部署,并披露规模、性能和商业条款。
非美国 AI 芯片资产稀缺,可能支撑 IPO 前估值溢价。一旦披露不及预期或退出时点延后,稀缺性溢价会迅速反转。招股书级披露显示收入规模足以支撑溢价,而不只是靠稀缺性。
私有 AI 芯片同行显示,投资者仍愿为可选性付费。SambaNova 和 Graphcore 表明,商业化卡住后,可选性可能坍塌。证据显示 Rebellions 正走向 Cerebras 式收入披露,而不是 SambaNova/Graphcore 式不透明。

投资逻辑真实存在,但主要是战略性的;反向逻辑是,公开财务证据仍大幅落后于估值。

[CV018, CV019, CV020, CV021, CV029, CV030]
FV001: 建议逻辑

从战略利好和支撑估值的叙事输入,走到相反方向的披露缺口,最终给出低置信度的继续研究结论。

[CV011, CV012, CV018, CV019, CV021, CV044]

8.2 当前价格支撑、融资背景与可比估值锚

当前价格支撑首先来自融资势能,而不是经营披露。Rebellions 从 2025 年 9 月 $1.4 billion 的 Series C 估值,升至 2026 年 3 月约 $2.34 billion 的投后估值,约六个月抬升约 67%。2026 年 4 月媒体将公司描述为正准备 KOSPI IPO,但本章审阅的来源中,没有可见的公开申报、价格区间或经审计招股书。换句话说,通常由公开市场尽调包强制披露的信息尚未打开,投资人已经在为 Pre-IPO 期权付费。 可比分析同时强化了上行叙事和谨慎理由。Cerebras 说明,只要财务披露真实、增长可见,有收入支撑的 AI 芯片公司可以获得很高的公开市场估值。Groq、Tenstorrent 和 FuriosaAI 说明,私人投资人仍愿意为 AI 硅片可选性支付数十亿美元估值。但 SambaNova 据报道探索出售、Graphcore 被 SoftBank 吸收,同样提醒我们:一旦资本强度、商业化或战略定位打破故事,独立 AI 芯片公司的结局会突然压缩。相对于这组同行,Rebellions 仅看私人估值并不荒谬;真正紧绷的是,它披露的收入基础远薄于价格已经隐含的规模。 [CV006, CV007, CV008, CV009, CV016, CV024]

可比估值表
可比公司状态 / 最新估值结果最近披露收入倍数 / 信号与 Rebellions 的相关性局限
Cerebras已上市;申报并上市后,2026 年 5 月市值大约在 $8B 以上收入:FY2025 $510M(SEC S-1/A)约十几倍中段的收入倍数(估计)最佳收入支撑型 AI 芯片锚点,显示披露能在公开市场解锁什么价值公开可比公司,披露远好于 Rebellions,已证明规模也明显更高
Groq私有;完成 $750M 融资后估值 $6.9B(据报道)未公开披露可选性锚点;没有干净的收入倍数显示投资者仍会为推理芯片可选性支付高额溢价收入不透明,无法逐项对比
SambaNova私有;据报道融资困难后探索出售未公开披露困境信号,不是干净倍数资本密集型 AI 芯片初创公司的重要下行先例结果由交易流程驱动,不是已确定的估值标记
Graphcore2024 年被 SoftBank 收购所审阅来源未公开披露退出时收入战略结果信号,不是干净倍数显示独立 AI 芯片公司可能以战略吸收告终,而不是 IPO退出条款不是透明的公开市场估值基准
FuriosaAI私有;2025 年完成 $125M 融资,据报道还计划更大规模融资未公开披露韩国同行估值信号;无收入倍数韩国推理芯片在地域和品类上最接近的同行收入不透明度仍高
Tenstorrent私有;据报道 2024 年底融资 $693M 并成为独角兽未公开披露平台可选性估值信号;无收入倍数说明投资者对非 Nvidia AI 计算平台仍有胃口产品宽度不同,且没有公开收入披露
Rebellions私有;2026 年 3 月 IPO 前融资后投后估值 $2.34B最后披露为 FY2023 收入 $2.7B KRW(约 $2.1M)按最后披露收入基线隐含 >1,100x(估计)本报告标的;估值落在私有可比公司区间内,但缺少收入支撑FY2024-FY2025 实际收入仍未披露,基线已经陈旧

表中混合了公开、私有和结果型可比对象;私有公司估值来自第三方报道,且多数缺少收入披露,所以更适合方向性使用,而不是机械套用。

[CV016, CV024, CV025, CV026, CV027, CV028]
FV002: 估值敏感性

隐含估值对披露和商业化里程碑的敏感度,而不是对单一干净收入倍数的敏感度;当前信息缺口决定了这一点。

数值为分析师按百万美元估计,展示里程碑敏感度,并非公司指引。

[CV039, CV040, CV041, CV042, CV043, CV053]

8.3 乐观、基准、悲观情景与投资逻辑破裂触发点

公司尚未向公开市场投资人提供足以搭建清晰收入或 EBITDA 框架的经营披露,因此情景分析比点估式倍数更站得住。在悲观情景下,Rebellions 仍有战略吸引力,但商业证据不足:FY2024-FY2025 收入仍未披露,IPO 延后,公开财务显示业务规模明显低于当前价格要求。该情景支撑约 $0.8 billion-$1.2 billion 区间。基准情景下,公司拿出足够的经审计规模,客户集中度看起来可控,IPO 路径仍可行,对应约 $1.5 billion-$2.5 billion。乐观情景下,主权 AI 部署转化为可见商业化、全球设计赢单和强 IPO 需求,可支撑约 $3 billion-$5 billion。 新投资人的问题在于,当前 $2.34 billion 标记已经接近基准情景区间的上半段。因此,下一个重大披露节点分量很重。最清晰的投资逻辑破裂触发点,是经审计或 IPO 申报收入明显低于 $2.34 billion 价格隐含的水平。第二个触发点来自流程:如果 IPO 延迟或撤回,又没有规模化部署证据抵消,稀缺性溢价应会迅速压缩。因此,下行不是理论风险,而是集中在少数几个高度可监测的披露上。 [CV040, CV041, CV042, CV043, CV049, CV050]

乐观 / 基准 / 悲观情景表
情景关键假设估值区间概率信号关键风险或解锁项
悲观FY2024-FY2025 实际收入仍然疲弱或令人失望,IPO 延后,公开披露显示规模有限。$0.8B-$1.2B若当前不透明最终揭晓为负面结果,风险重大。下行由稀缺性溢价压缩和商业化证据薄弱驱动。
基准Rebellions 展示可信但仍有限的商业规模,推进 IPO 申报,并维持主权 AI 定位。$1.5B-$2.5B按当前证据,这是概率最高区间。需要披露质量足以维持 IPO 路径。
乐观经审计增长、强劲主权 AI 部署和旺盛 IPO 需求,共同验证 Rebellions 是稀缺 AI 芯片资产。$3.0B-$5.0B需要多项有利披露,而不只是叙事继续升温。上行取决于证明规模,而不只是保住可选性。

情景区间是分析师估计,锚定融资背景、披露质量和 AI 芯片同行结果,而不是干净的收入倍数。

[CV040, CV041, CV042, CV043]
投资逻辑破坏和否决触发项表
触发项阈值 / 事件对投资逻辑的传导行动含义
收入披露不及预期IPO 申报或审计结果显示,收入远低于 $2.34B 估值所隐含水平可选性溢价失去核心支撑,估值必须向悲观情景压缩立即重新测算;在出现新的价格发现前,假设估值偏高
IPO 延误且无补偿KOSPI 流程延期或消失,且没有合同或收入披露来抵消信息风险仍高,稀缺性和流动性溢价同步压缩下调任何建设性立场,完全转为观望
毛利率披露偏弱首次审计披露显示硬件经济性结构性偏弱削弱从增长叙事走向公开市场半导体倍数支撑的路径用更低的基准和乐观区间重做情景表
客户集中度冲击审计披露显示极度依赖一两个锚定客户收入质量变弱,可重复性假设下降将增长视为更不耐久,并上调悲观情景概率
没有重复部署证据新披露确认只是试点或一次性项目,而不是可重复的规模化铺开系统级产品故事没有转化为商业复利从投资测算中剔除 IPO 准备度溢价

触发项定义聚焦可观察披露事件,而不是叙事情绪,因为这个估值取决于少数可监控的公开证明。

[CV049, CV050, CV051]
FV003: 估值 / 回报区间

以百万美元计的悲观、基准和乐观估值区间,显示如果公开披露不及预期,当前标记估值几乎没有缓冲。

中点仅为分析师估计,反映高度披露不确定性下的情景加权判断。

[CV040, CV041, CV042, CV047]

8.4 退出准备度、证据缺口与最终尽调问题

现有公开证据支持 IPO 可能性,但不支持 IPO 已准备就绪。Seoul Economic Daily 的报道把 Rebellions 放进韩国 AI 芯片上市竞赛,KOSPI 这一上市地点也意味着,一旦最终申报,公司必须给出丰富得多的披露。但当前公开记录仍没有披露经审计的 FY2024 或 FY2025 收入、毛利率、积压订单、集中度或股权结构经济条款。公司因此处在披露真空地带:定价像后期 Pre-IPO 资产,文件透明度却更像风投支持的私营初创公司。 这也限制了退出分析。从本章审阅来源看,KOSPI 上市仍是唯一可支撑的公开退出路径。战略收购方角度理论上可能存在,尤其在主权 AI 基础设施和非 Nvidia 供应链增值的世界里,但公开证据不足以让 M&A 成为今天的估值支柱。正确做法不是在证据之外强行给出明确看多或看空结论,而是列出会改变判断的缺失文件。经审计 FY2024-FY2025 财务、优先股堆叠细节、积压订单和客户集中度披露、经验证的复购部署数据,是从继续研究转向更建设性立场前所需的最低材料包。 [CV008, CV009, CV010, CV048, CV051, CV052]

最终尽调问题表
主题缺失证据重要性负责人 / 尽调路径
经审计的 FY2024-FY2025 财务数据FY2023 之后两年的收入、毛利率、经营亏损和烧钱速度桥接这是当前估值只能按期权价值处理的最大障碍KOSPI 申报文件、经审计招股书,或公司 / 投资方数据室直接披露
优先股堆叠与股权结构条款清算优先权、反稀释条款、老股交易权利,以及任何结构化 IPO 前保护如果优先股堆叠很重,回报结果会与名义投后估值显著不同律师审阅融资文件,或投资方股权结构包
客户集中度与积压订单具名客户、收入集中度、复购节奏和已承诺积压订单要支撑 IPO 和下行保护,收入质量与收入规模同样关键招股书客户风险章节、尽调访谈,或经审计附注披露
产品层面单位经济芯片卡、服务器和机柜级系统的毛利率与定价证据即使有商业规模,硬件经济性不健康,也撑不起稳健的公开市场倍数管理层指引、经审计分部披露,或 NDA 下的投行材料
IPO 准备度与投行流程交易所时间表、承销商阵容、申报状态和治理准备度当前价格含有 IPO 前溢价,不能盲目支付持续跟踪 KOSPI 流程披露,以及公司 / 投资方直接确认
战略退出证据真实战略兴趣信号,而非理论上的并购契合没有真实交易对手意愿证据前,不应把并购计入估值支撑董事会或投行尽调,而不是公共叙事外推

各项按承销重要性排序;前三项会阻断任何高于继续研究的动作。

[CV048, CV051, CV052]
FV004: 投资 KPI

IC 风格的八维评分,强调战略定位明显好于已披露商业证据或证据质量。

[CV018, CV023, CV038, CV045, CV046, CV053]

免责声明

本报告由 AI 研究工作流根据截至 2026-05-20 的公开来源生成,仅供信息参考,不构成投资建议。私营公司数据仍部分不透明,若干结论依赖第三方报道、管理层表述和比较分析,而非经审计文件。

证据索引

结论
编号陈述可信度来源
CO001 Rebellions was co-founded in 2020 by five Korean engineers as a fabless AI chip startup focused on inference NPUs. SO001, SO009
CO002 Rebellions' headquarters is located at 3F, 6 Jeongjail-ro 156beon-gil, Bundang-gu, Seongnam-si, Gyeonggi-do, South Korea. SO001
CO003 Rebellions builds purpose-engineered AI accelerators for energy-efficient, high-performance AI inference at data-center scale. SO001, SO006
CO004 CEO Sunghyun Park holds a master's and doctorate from MIT's Computer Science and Artificial Intelligence Laboratory (CSAIL). SO005, SO012
CO005 Park Sung-hyun built his pre-Rebellions career at Intel, SpaceX, and Morgan Stanley before co-founding Rebellions. SO005, SO012
CO006 The other four co-founders of Rebellions are not individually identified by name in any reviewed public source. SO009, SO001
CO007 Forbes' April 2025 profile of Park Sung-hyun describes him as 40 years old, covering Korea's Richest 2025. SO008
CO008 In November 2025, Rebellions appointed Marshall Choy as Chief Business Officer to lead its newly established U.S. entity and global go-to-market. SO020
CO009 Post-merger, SAPEON Korea is the surviving legal entity but operates under the Rebellions name and is led by Rebellions' CEO Park Sung-hyun. SO004, SO005, SO018
CO010 The merged Rebellions-SAPEON entity is described as Korea's first AI chip unicorn, with corporate value exceeding 1 trillion Korean won. SO005, SO007
CO011 Rebellions closed a $124 million Series B on January 30, 2024, led by KT, with Pavilion Capital (Temasek), KDB, Mirae Asset Venture Investment, IMM Investment, KT Investment, and SV Investment participating. SO002, SO009
CO012 After the January 2024 Series B, Rebellions' total capital raised exceeded $200 million, making it the most-funded semiconductor startup in South Korea at the time. SO002
CO013 The Series B also included Korelya Capital (France), DGDV (Japan), KB Securities, KB Investment, Seoul Techno Holdings, and several other Korean institutional investors. SO002
CO014 Wa'ed Ventures, a $500 million venture capital fund wholly owned by Saudi Aramco, invested $15 million in Rebellions in a Series B extension on July 23, 2024. SO003
CO015 On August 18, 2024, Rebellions and SAPEON Korea signed a definitive merger agreement with an equity value ratio of 2.4:1 (Rebellions:SAPEON). SO004, SO014, SO018
CO016 The merger agreement stated that SKT, SK Square, and SK Hynix would sell 3% of their SAPEON shares prior to the merger to ensure Rebellions' management would become majority shareholders of the merged entity. SO004, SO014
CO017 The Rebellions-SAPEON merger was completed on December 2, 2024, with Park Sung-hyun confirmed as CEO of the merged entity. SO005, SO007
CO018 Rebellions raised $250 million in a Series C at a $1.4 billion post-money valuation on September 30, 2025. SO022, SO019
CO019 In November 2025, Rebellions extended the Series C with new participation from Kindred Ventures and Top Tier Capital Partners. SO019
CO020 Rebellions closed a $400 million pre-IPO round on March 30, 2026, led by Mirae Asset Financial Group and the Korea National Growth Fund. SO006, SO011, SO013
CO021 The March 2026 pre-IPO round valued Rebellions at approximately $2.34 billion post-money. SO006, SO013
CO022 As of March 30, 2026, Rebellions' total capital raised was $850 million across all rounds. SO006, SO011
CO023 In the six months preceding the March 2026 pre-IPO close, Rebellions raised $650 million—over 75% of its lifetime capital to that date. SO006, SO022, SO011
CO024 Rebellions taped out the ATOM SoC in June 2022 on a Samsung GDDR6-based process as its first-generation AI inference accelerator. SO001, SO009
CO025 Rebellions delivered the first ATOM chips to kt cloud in May 2023, making ATOM the first Korean-developed inference chip in live data-center deployment. SO001, SO009
CO026 ATOM is targeted for mass production on Samsung's 5nm technology. SO009, SO024
CO027 ATOM targets AI inference for models with up to 7 billion parameters. SO009
CO028 Rebellions taped out the REBEL SoC in November 2024, described as the world's first UCIe-Advanced AI chiplet integrating 144GB of HBM3E. SO001, SO023
CO029 Rebellions established a Japan subsidiary in February 2025 as its first overseas branch. SO001
CO030 Rebellions established a Saudi Arabia subsidiary in August 2025 based on Wa'ed Ventures' strategic investment and Saudi AI market opportunity. SO001, SO003
CO031 Following the SAPEON merger, SK Telecom, SK Square, and SK Hynix became strategic investors in Rebellions through their inherited SAPEON shareholdings. SO004, SO005, SO014
CO032 SK Telecom tested Rebellions' ATOM-based NPU servers for AI services including A.Dot call summarization, PASS spam filtering, PASS financial assistant, and X Caliber, aiming to commercialize ATOM-Max by year-end 2025. SO015, SO025
CO033 In October 2023, Samsung Electronics and Rebellions announced a strategic partnership to co-develop REBEL on Samsung's 4nm foundry process with HBM3E memory integration. SO024, SO001
CO034 In April 2026, SK Telecom, Arm, and Rebellions signed an MOU to jointly develop AI inference servers combining Arm's AGI CPU and Rebellions' RebelCard accelerator for sovereign AI applications. SO010, SO016, SO017
CO035 In December 2025, Rebellions and Red Hat introduced Red Hat OpenShift AI powered by Rebellions NPUs, providing enterprise AI inference on Rebellions hardware. SO021
CO036 In April 2025, Rebellions, SK Telecom, and DOCOMO Innovations (NTT DOCOMO subsidiary) signed an MOU to evaluate Rebellions' ATOM-based NPU servers within SK Telecom's NPU farm. SO025
CO037 Rebellions raised $650 million in the six months ending March 2026, representing over 75% of its total lifetime capital to that date. SO006, SO022, SO011
CO038 Rebellions is preparing for a future IPO as of March 2026; no specific exchange, price range, or filing date has been publicly announced. SO006, SO013
CO039 At the time of the August 2024 merger announcement, Rebellions was valued at approximately 900 billion Korean won ($664 million) and SAPEON Korea at approximately 550 billion KRW. SO007
CO040 Marshall Choy, appointed CBO in November 2025, is a Silicon Valley-based industry veteran with more than two decades of experience in enterprise systems and AI. SO020
CO041 Rebellions presented REBEL-Quad at Hot Chips Symposium 2025 on August 27, 2025, targeting Blackwell-grade performance with superior energy efficiency, as its next-generation AI accelerator after ATOM. SO023
CO042 Arm became a strategic investor in Rebellions through the Series C round announced September 30, 2025. SO022, SO019
CO043 Samsung Ventures participated in the Rebellions Series C announced September 30, 2025. SO022, SO019
CO044 Pegatron VC participated in the Rebellions Series C, strengthening strategic ties in AI hardware and AI module assembly. SO022
CO045 Korea Development Bank (KDB) made a follow-on investment in the Rebellions Series C. SO022
CO046 Nvidia held approximately 94% of the AI chip market share as of 2023, making it the dominant competitor Rebellions must displace to achieve significant market presence. SO007, SO012
CO047 The global AI chip market was valued at $34.3 billion in 2023, according to The Investor at the time of the SAPEON merger reporting. SO007
CO048 The Rebel100 platform is described as Rebellions' current production platform as of the March 2026 pre-IPO announcement. SO006, SO011
CO049 RebelRack and RebelPOD were launched alongside the March 2026 pre-IPO announcement, described as fully deployable, vertically integrated AI infrastructure for production-scale environments. SO006, SO011
CO050 Forbes featured Park Sung-hyun in its 'Korea's Richest 2025' coverage in April 2025, positioning Rebellions as South Korea's AI chip champion with global ambitions. SO008
CO051 Rebellions competes in a market where Nvidia holds approximately 94% share with a deeply entrenched software ecosystem (CUDA), creating significant barriers to customer adoption for alternative inference hardware. SO007, SO012
CO052 Rebellions has not disclosed any revenue, ARR, gross margin, customer count, or burn rate in any public communication reviewed for this chapter. SO006, SO001
CO053 No IPO exchange, filing date, price range, or lock-up structure has been publicly announced by Rebellions as of May 2026, despite the March 2026 pre-IPO round. SO006, SO013
CM001 Nvidia's Compute & Networking segment revenue in FY2026 (year ended January 25, 2026) was $193.5 billion, up 67% year-over-year, representing the best available proxy floor for the global AI accelerated compute market. SM001, SM003
CM002 Nvidia's total revenue in FY2026 was $215.9 billion, up 65% from $130.5 billion in FY2025, with gross margin declining from 75.0% to 71.1% partly due to a $4.5 billion H20 inventory charge from export control changes. SM001, SM008
CM003 The global AI chip market was approximately $34.3 billion in calendar year 2023, with Nvidia holding approximately 94% market share at that time — a figure that understates the current scale given rapid growth since. SM009, SM010
CM004 Nvidia's Compute & Networking revenue in FY2025 (year ended January 26, 2025) is estimated at approximately $115.9 billion, derived by dividing FY2026 C&N revenue ($193.5B) by the reported 67% growth rate — confirming rapid acceleration in AI compute spending. SM001, SM008
CM005 Training costs for ML systems grew at 0.49 orders of magnitude per year between 2009 and 2022 (90% CI: 0.37–0.56), with all-systems cost range spanning from $0.02 to approximately $80,000 — illustrating the exponential compute demand that is now shifting from training to inference deployment. SM002
CM006 Inference — running trained AI models at production scale for billions of end users — is the fastest-growing component of AI chip demand in 2025–2026, as generative AI applications shift from model development to deployment, driving a structural transition in accelerated compute workload mix. SM001, SM003, SM004
CM007 Training costs for large-scale ML systems grew more slowly at 0.2 OOMs/year since October 2015 versus 0.51 OOMs/year for all systems, suggesting compute efficiency gains in large frontier model training that shift the economic pressure toward inference optimization. SM002
CM008 Nvidia's FY2026 10-K discloses that the company was 'effectively foreclosed from competing in China's data center computing market' by the end of FY2026 due to US government export controls, materially benefiting alternative silicon suppliers in markets seeking AI supply-chain independence. SM001, SM008
CM009 Nvidia's data center customers span all major public and private cloud service providers, AI model makers, enterprises, startups, and public sector entities globally, as stated in its FY2026 annual report. SM001, SM003
CM010 Samsung SDS achieved 23.9% managed cloud service provider (MSP) market share in South Korea in 2023, per IDC's Domestic Managed Cloud Service Market Share Report, making it the leading domestic managed cloud provider in Korea. SM005
CM011 Samsung SDS ranked second in Korea's domestic public cloud (CSP) market with 11.0% market share per IDC 2023 data, confirming Samsung Group affiliates as the dominant cloud buyers in Korea and a strategically important inference chip customer pathway for Rebellions. SM005
CM012 SK Telecom, Arm, and Rebellions announced a formal sovereign AI chip collaboration in April 2026 targeting Korean telco data centers, establishing the first confirmed policy-linked AI chip procurement pathway for Rebellions in Korea. SM015, SM016, SM017, SM024
CM013 Arm positioned its compute platform as 'the trusted foundation for AI inference deployment across edge to cloud' and published the Arm AI Readiness Index report surveying global enterprise AI adoption, identifying inference deployment as a rapidly growing workload. SM004
CM014 Rebellions' RebelServer integrates its NPU technology for data center AI inference, marketed to cloud service providers and enterprise customers seeking efficient AI infrastructure. SM006
CM015 Samsung Electronics invested in Rebellions and established a co-development partnership for next-generation AI chips targeting the generative AI market, creating a strategic buyer-investor-foundry relationship that is structurally unique among AI chip startups. SM020, SM019
CM016 NTT Docomo innovations joined the Rebellions/SKT infrastructure partnership to accelerate next-generation AI deployment, extending the Korean sovereign AI chip model to Japan — the second-largest AI chip market in Asia. SM023
CM017 Red Hat certified Rebellions' NPU technology for OpenShift AI, providing an enterprise software integration pathway that reduces migration risk for enterprise on-premises inference buyers adopting Rebellions silicon. SM022
CM018 Hyperscaler capital expenditure on AI infrastructure drove Nvidia's FY2026 Compute & Networking revenue to $193.5 billion, confirming ongoing hyper-investment in AI compute as the dominant market driver — even after China market closure. SM001, SM003
CM019 US export controls on AI chips foreclosed Nvidia from China's data center compute market by FY2026, creating demand for alternative silicon suppliers in markets prioritizing AI supply-chain independence — directly benefiting Korea-origin AI chip companies. SM001, SM014
CM020 Epoch AI's analysis of 124 ML systems shows that training costs for large-scale models grew more slowly than average (0.2 vs. 0.51 OOMs/year), suggesting compute efficiency improvements in frontier model training that may accelerate the economics shift toward inference optimization. SM002
CM021 Nvidia's gross margin declined from 75.0% in FY2025 to 71.1% in FY2026, partly due to a $4.5 billion charge for H20 inventory excess caused by US export control changes — illustrating geopolitical supply risk embedded in the incumbent AI chip stack. SM001, SM008
CM022 Rebellions raised $250 million in a round backed by Arm and Samsung, demonstrating that industry-tier strategic investors validated the Korea-origin inference chip ecosystem as commercially investable. SM019, SM012
CM023 The AI inference chip market faces a power-efficiency adoption driver favoring purpose-built NPUs: Rebellions' REBEL Quad chiplet debuted at Hot Chips 2025 claiming breakthrough TPS/W performance, explicitly targeting the power-efficiency constraint of GPU-based inference in data centers. SM006, SM021
CM024 Korea's Ministry of Science and ICT backed a sovereign AI chip initiative formalizing domestic chip preference in government and telco AI infrastructure, creating a policy-driven demand category distinct from pure commercial chip adoption. SM014, SM015
CM025 Rebellions' REBEL Quad chiplet debuted at Hot Chips 2025 with claimed breakthrough in performance-per-watt targeting what the company described as 'AI's energy tax,' but no independent third-party benchmark corroboration of performance claims was identified in this research. SM021
CM026 The SK Telecom / Rebellions / Arm sovereign AI chip partnership, formally announced in April 2026, positions Rebellions silicon as the AI inference layer for Korean sovereign telco data centers, establishing the first concrete large-telco procurement pathway for Rebellions in Korea. SM015, SM016, SM017, SM024
CM027 The Rebellions + SKT sovereign AI initiative aligns Korean industrial policy objectives (national security, telecom infrastructure independence, AI semiconductor development) with commercial chip procurement, creating a compound demand driver not present for most AI chip challengers. SM014, SM015
CM028 NTT Docomo's partnership with Rebellions and SKT on next-generation AI infrastructure extends the Korean sovereign AI NPU model to Japan, indicating that major Asian telcos see commercial value in non-Nvidia inference silicon for regional AI infrastructure. SM023
CM029 Rebellions raised a Silicon Valley-backed Series C prior to its $250 million Arm/Samsung round, indicating cross-border US investor interest in Korean AI chip challengers as a viable investment category. SM018
CM030 Korea's vertically integrated domestic AI chip ecosystem — Samsung Electronics as foundry and investor, SK Telecom as sovereign AI chip customer, Arm as co-development partner — represents a structural support structure for Rebellions that most non-US AI chip startups lack. SM015, SM019, SM020, SM024
CM031 IDC data reported by BusinessKorea shows Samsung SDS dominates Korean cloud infrastructure with 23.9% MSP share, establishing Samsung Group affiliates as the dominant cloud buyer in Korea and a natural — though not yet confirmed — procurement pathway for Rebellions inference chips. SM005
CM032 Rebellions' inference-only NPU strategy avoids direct head-to-head competition with Nvidia in the training market, focusing on inference economics where total-cost-of-ownership advantages may be clearer to buyers evaluating power consumption and cost per inference token. SM006, SM003
CM033 Arm's AI Readiness Index report benchmarks global enterprise AI adoption and identifies inference deployment as a rapidly growing enterprise workload, supporting the thesis that enterprise demand for inference-optimized silicon will expand beyond hyperscalers. SM004
CM034 Nvidia's Rubin platform — successor to Blackwell — is expected to commence production shipments in H2 FY2027 (approximately H2 2026 calendar year), delivering claimed 10x reduction in cost per token versus Blackwell, indicating the incumbent's inference optimization roadmap that challengers must anticipate. SM001
CM035 Epoch AI's extrapolation from 2009–2022 ML training data suggests frontier model training costs may eventually exceed $233 billion annually (~1% of US GDP), reinforcing the economic imperative to deploy trained models on inference-optimized silicon to reduce per-inference cost. SM002
CM036 Samsung Electronics and Rebellions established a formal co-development partnership for next-generation AI chips targeting the generative AI market, with Samsung providing manufacturing (foundry) and strategic investment alongside the product development collaboration. SM020, SM019
CM037 Rebellions' multiple NPU product lines — ATOM, REBEL, REBELRACE, and REBEL Quad chiplet — indicate a multi-product portfolio strategy spanning different inference performance and power tiers, consistent with targeting heterogeneous buyer segments from telco to enterprise. SM006, SM007, SM021
CM038 Converge Digest's coverage of the Rebellions/Arm/SKT partnership describes the initiative as positioning Rebellions silicon as 'the inference layer for Korean sovereign AI telco data centers,' representing first-mover strategic positioning in Korea's government-backed AI infrastructure plan. SM016
CP001 Groq was founded in 2016 and pioneered the LPU (Language Processing Unit) as the first chip purpose-built for AI inference. SP001
CP002 Groq raised $750 million in Series D financing at a $6.9 billion post-money valuation in September 2025, with backers including Samsung, Cisco, Disruptive, BlackRock, and Neuberger Berman. SP002, SP004
CP003 Groq powers more than two million developers and Fortune 500 companies with fast, affordable AI inference via GroqCloud. SP001, SP002
CP004 In December 2025, Groq entered a non-exclusive licensing agreement with Nvidia; Groq's founder Jonathan Ross and President Sunny Madra left Groq to join Nvidia as part of the arrangement. SP003, SP004
CP005 Saudi Arabia committed $1.5 billion to Groq AI inference infrastructure, announced at the LEAP 2025 conference in February 2025. SP005, SP006
CP006 Groq operates data centers in North America, Europe, and the Middle East as of May 2025. SP002, SP005
CP007 Groq was named an official inference provider for HUMAIN, a Saudi AI company launched to transform economies through large-scale AI capabilities. SP005
CP008 Tenstorrent raised $693 million in Series D financing led by Samsung Securities and LG Electronics. SP009, SP010
CP009 Tenstorrent's Wormhole AI accelerator cards target both training and inference workloads. SP007, SP008
CP010 Tenstorrent takes an open-source RISC-V approach to hardware and software ecosystem development, aiming to build a broader developer community than proprietary NPU vendors. SP007, SP009
CP011 SambaNova's SN50 RDU (Reconfigurable Dataflow Unit) is the company's fifth-generation AI inference processor, designed specifically for large-scale agentic workloads. SP012
CP012 SambaNova's RDU architecture maps model execution directly onto the processor, minimizing data movement to memory, which SambaNova positions as the key inference efficiency advantage over GPU-based systems. SP011, SP012
CP013 On SambaNova Cloud (SambaCloud), DeepSeek-V3.1 achieves up to 200 tokens per second and MiniMax M2.7 achieves 435 tokens per second, both independently measured by Artificial Analysis. SP011, SP013
CP014 SambaNova's SN50 can run multiple models simultaneously using a tiered memory architecture, enabling rapid model switching with minimal latency for agentic workloads. SP012
CP015 Cerebras WSE-3 contains four trillion transistors and delivers 125 petaflops of compute in a single wafer-scale chip, with redundant compute cores enabling a fail-in-place architecture. SP014, SP015
CP016 As of May 2026, Cerebras has completed its IPO, per NextPlatform's coverage of the company's post-IPO product direction. SP024
CP017 Cerebras targets both AI training and inference with its wafer-scale approach, competing in segments distinct from Rebellions' air-cooled data-center inference focus. SP014, SP015
CP018 Hailo focuses exclusively on edge AI inference with products including Hailo-8, Hailo-10H, and Hailo-15 processors targeting edge devices such as cameras, automotive systems, and robotic platforms. SP018
CP019 Hailo's processors are designed for 1–10W power profiles and real-time deep learning inference on edge devices, not data-center AI workloads. SP018
CP020 Hailo does not compete with Rebellions in data-center AI inference; they occupy categorically different market segments (edge vs. data center) with different buyers and power profiles. SP018
CP021 FuriosaAI's RNGD chip delivers 512 TFLOPS (eight processing elements at 64 TFLOPS FP8 each), 48 GB HBM3 memory, 1.5 TB/s memory bandwidth, and a 180W TDP targeting air-cooled data centers. SP016, SP017
CP022 FuriosaAI plans to raise $500 million before an IPO, per reports from January 2026. SP017
CP023 LG CNS partnered with FuriosaAI in February 2026 to bring South Korean NPU chips to enterprise AI services. SP017
CP024 FuriosaAI began shipping RNGD chips in January 2026. SP016, SP017
CP025 FuriosaAI's CEO June Paik stated that 'the AI data centers of 2036 won't be filled with GPUs,' reflecting the company's thesis that inference-optimized NPUs will displace GPU-centric infrastructure over the coming decade. SP017
CP026 AMD Instinct MI325X offers 304 compute units, 256 GB HBM3E memory, and 6 TB/s peak theoretical memory bandwidth on CDNA3 architecture. SP019
CP027 AMD claims the Instinct MI325X delivers up to 1.3x the AI performance versus competitive accelerators; this comparison is company-originated and not independently validated. SP019
CP028 AMD Instinct MI300 series targets both AI training and inference workloads with CDNA3 architecture supporting FP64, FP32, FP16, BF16, INT8, and FP8 precision formats. SP019
CP029 Nvidia H100 delivers up to 30x higher AI inference performance for the largest LLM models compared to A100, based on Megatron chatbot inference benchmarks at 530B parameters. SP020, SP021
CP030 Nvidia GB200 NVL72 connects 36 Grace CPUs and 72 Blackwell GPUs in a rack-scale liquid-cooled design and delivers 30x faster real-time LLM inference versus H100. SP020, SP021
CP031 Nvidia GB200 NVL72 delivers 25x more performance at the same power compared to H100 air-cooled infrastructure, representing a step-change in energy efficiency as well as raw performance. SP021
CP032 Nvidia's CUDA software ecosystem is estimated to have over 4 million registered developer users and more than 3,800 GPU-accelerated applications, representing roughly 20 years of ecosystem investment since CUDA launched in 2006. SP020, SP027
CP033 Google Cloud TPU Ironwood (7th generation) features 9,216 liquid-cooled chips per pod, provides 42.5 ExaFlops of compute, and is optimized for large-scale training, reasoning, and inference including agentic AI workloads. SP022
CP034 Google Cloud TPU 8i delivers 80% better performance-per-dollar compared to prior-generation TPUs for low-latency inference of large MoE models. SP022
CP035 AWS Trainium3, built on a 3nm process, provides 2.52 petaflops of FP8 compute, 144 GB HBM3e memory, and 4x better energy efficiency compared to Trainium2 UltraServers. SP023
CP036 AWS Trainium3 UltraServers deliver 4.4x more performance versus Trainium2 UltraServers; Anthropic, Databricks, Decart, and poolside are among the early adopters of Trn3. SP023
CP037 FuriosaAI is Rebellions' most direct Korean-market competitor, sharing the same 180W air-cooled inference profile, same Samsung Foundry manufacturing access, and largely the same target customer set of Korean telcos and sovereign AI programs. SP016, SP017
CP038 Hailo (edge AI), Google Cloud TPU (captive Google-only ASIC), and AWS Trainium (captive AWS-only ASIC) do not directly compete with Rebellions for external data-center inference chip procurement. SP018, SP022, SP023
CP039 Groq differentiates primarily on its GroqCloud API layer and developer ecosystem (OpenAI-compatible API, 2M+ developer users) rather than bare-metal chip hardware sales; this positions Groq as competing with Rebellions at the cloud API layer but not in the hardware-plus-SDK enterprise sales channel. SP001, SP003, SP004
CP040 Nvidia's CUDA software ecosystem and its 20-year head start in GPU-accelerated application development constitute the primary structural barrier for all AI inference chip challengers including Rebellions, because every customer migration to an NPU requires re-validating model compatibility, operator support, and deployment tooling. SP020, SP021, SP027
CP041 The December 2025 Groq-Nvidia licensing agreement, in which Groq's founder and president joined Nvidia, signals that Nvidia can neutralize inference chip startup competition through talent and technology licensing deals rather than full acquisitions. SP003, SP004
CP042 Rebellions' Samsung Foundry manufacturing partnership and SK Telecom strategic investor status provide a Korean domestic market moat unavailable to Western inference chip competitors including Groq, SambaNova, Cerebras, and Tenstorrent. SP007, SP010, SP027
CP043 Tenstorrent's $693M raise from Samsung Securities and LG Electronics, combined with its open-source RISC-V strategy, represents the most direct Western silicon challenger to Rebellions' mid-range inference NPU positioning, with institutional investor overlap creating potential channel conflict. SP009, SP010
CP044 Cerebras' wafer-scale WSE-3 approach requires liquid cooling and targets a higher-power, training-oriented performance envelope that differs fundamentally from Rebellions' air-cooled 180W chiplet inference product, making them indirect rather than direct competitors. SP014, SP015
CP045 SambaNova's full-stack model (proprietary RDU hardware plus SambaCloud inference API) competes at the cloud API layer rather than as a direct hardware chip vendor, creating partial overlap with Rebellions in enterprise AI but through a different go-to-market motion. SP011, SP012
CI001 Rebellions' revenue model is based on hardware sales of AI accelerator chips (ATOM, REBEL-Quad/Rebel100), server systems (RebelServer), rack-scale systems (RebelRack, RebelPOD), and accelerator cards (RebelCard), with software (SDK) bundled at no separate charge. SI001, SI002, SI004
CI002 ATOM and ATOM-Max chips were in mass production and deployed with customers across Japan, Saudi Arabia, and the United States, and powered Korea's largest commercial AI service as of September 2025. SI004
CI003 RebelRack and RebelPOD rack-scale systems became available in March 2026 as part of the pre-IPO announcement, described as fully deployable vertically integrated AI infrastructure. SI001, SI002
CI004 CEO Sunghyun Park stated a 2025 revenue target of approximately 100 billion KRW (roughly $68 million), as reported by Forbes Asia in April 2025. SI008
CI005 Rebellions partners with Taiwanese assembler Pegatron to develop AI servers powered by the Rebel chip, and with Penguin Solutions for cluster deployment assistance, embedding partner channel economics into the cost structure. SI008, SI001
CI006 The Rebellions SDK is designed to be open-source-aligned, supporting vLLM, PyTorch, Triton, Hugging Face, and Red Hat OpenShift AI, and is bundled with hardware rather than sold as a standalone subscription. SI001, SI002
CI007 Rebellions has not publicly disclosed list pricing for any of its hardware products as of May 2026; pricing is negotiated on a customer-by-customer basis. SI001, SI004, SI008
CI008 CEO Park claims that in terms of total cost of ownership for AI inference, the Rebel chip is cheaper than the Nvidia H100, citing lower peak power consumption (350W vs 400W for H100) and greater HBM3E memory capacity (144GB vs 80GB for H100). SI008
CI009 Korean corporate registry filings reviewed by Forbes Asia show Rebellions posted 2.7 billion KRW (~$2.1M) in revenue for fiscal year 2023, with a net loss of 13.7 billion KRW (~$10.5M), widening from a net loss of 8.1 billion KRW (~$6.2M) in 2022. SI008
CI010 Rebellions has raised $850 million in total capital as of March 30, 2026, per the official pre-IPO press release and corroborated by CNBC and the official company newsroom. SI001, SI002, SI003
CI011 The March 30, 2026 pre-IPO round raised $400 million at a post-money valuation of approximately $2.34 billion, led by Mirae Asset Financial Group and the Korea National Growth Fund. SI001, SI002, SI003
CI012 The Series C round raised $250 million at a $1.4 billion post-money valuation on September 30, 2025, with Arm as the lead strategic investor alongside Samsung Ventures and Pegatron VC. SI004, SI005
CI013 $650 million—over 75% of Rebellions' total lifetime capital—was raised in the six months between September 2025 and March 2026, indicating compressed and highly concentrated fundraising ahead of the planned IPO. SI001, SI002
CI014 The Series C was extended on November 10, 2025, adding Kindred Ventures and Top Tier Capital Partners; Kindred's investment marked its first ever investment in a Korean startup. SI005, SI004
CI015 Sungkyue Shin serves as CFO of Rebellions, as identified in the Series C extension press release dated November 10, 2025. SI005
CI016 The Series B of $124 million closed January 30, 2024, led by KT Corporation—Korea's premier data-center operator—with Korelya Capital, Korea Development Bank, and Samsung Ventures also participating. SI007, SI009
CI017 The Korea National Growth Fund chose Rebellions as its first investment under the K-Nvidia national AI semiconductor initiative, per the pre-IPO press release. SI001
CI018 Mirae Asset Financial Group has backed Rebellions since Series A and led the pre-IPO round, with Mirae Asset Venture Investment CEO Eung-Suk Kim quoted in the pre-IPO press release. SI001, SI003
CI019 The Series B extension of $15 million closed July 23, 2024, funded by Wa'ed Ventures (Saudi Aramco's venture arm)—its first investment in a Korean startup—directly linked to the Saudi Arabia market relationship. SI010, SI008
CI020 The SAPEON Korea merger (all-stock, December 2024) brought SK Telecom, SK Square, and SK Hynix into Rebellions as strategic investors through their SAPEON shareholdings, valuing the combined entity at approximately 1.3 trillion KRW (~$1 billion at merger). SI017, SI016, SI025
CI021 CEO Park holds approximately 10% of the merged Rebellions entity, per Korean regulatory filings reviewed by Forbes Asia (April 2025). SI008
CI022 Rebellions operates as a fabless semiconductor company that eliminates manufacturing capex but incurs recurring non-recurring engineering (NRE) and mask set costs paid to Samsung Foundry for each tape-out, typically tens of millions of dollars per advanced-node design. SI009, SI013, SI011
CI023 Samsung Foundry offers advanced process technologies including 14/10/8/5/4nm FinFET, 3nm GAA with EUV from 5nm, and integrated 3D/2.5D packaging—the nodes and packaging approaches used for Rebellions' REBEL-Quad chiplet architecture. SI013, SI006
CI024 CEO Park acknowledged a shortage in foundry manufacturing capacity and a shortage of HBM3E memory as structural supply-chain risks, as quoted by Forbes Asia in April 2025. SI008
CI025 The Rebellions turnkey relationship with Samsung Electronics covers silicon manufacturing, HBM3E memory supply, and packaging, making Samsung the sole foundry and primary supply-chain partner. SI009, SI006, SI013
CI026 Rebellions' REBEL-Quad chip uses 144GB of HBM3E memory, compared to 80GB of HBM3 on the Nvidia H100, requiring access to scarce high-bandwidth memory supply that constrains production volumes. SI004, SI008
CI027 The FY2023 net loss of 13.7 billion KRW against revenue of 2.7 billion KRW implies a deeply negative operating margin at that stage, consistent with a pre-scale fabless chip startup investing heavily in R&D and initial commercialization. SI008
CI028 The pre-IPO proceeds are designated for US market expansion (led by CBO Marshall Choy), scaled production of the Rebel100 platform, and preparation for a future IPO. SI001, SI002
CI029 Rebellions announced a partnership with Pegatron (Taiwan-based electronics assembler) to develop AI servers powered by the Rebel chip, and with Penguin Solutions to assist customers with cluster deployment. SI008, SI001
CI030 Rebellions has expanded global operations through subsidiaries in Japan (established February 2025), Saudi Arabia (August 2025), and a US entity led by CBO Marshall Choy, representing growing operational expenditure ahead of revenue confirmation. SI001, SI021
CI031 No publicly disclosed cash-on-hand figure, monthly burn rate, or remaining runway has been released by Rebellions as of May 2026, preventing precise capital adequacy assessment. SI001, SI003, SI008
CI032 The qualitative runway estimate for Rebellions, based on the $400 million pre-IPO round and disclosed expansion plans, is approximately 18 to 24 months from April 2026—a highly uncertain figure that could be materially lower if post-merger burn has accelerated. SI001, SI008
CI033 Rebellions' planned IPO is the central next capital event, but no exchange, filing date, IPO price range, or underwriter mandate has been publicly disclosed as of May 2026. SI001, SI002, SI003
CI034 The Korea National Growth Fund's participation in the pre-IPO round is expected to carry governance obligations typical of public-fund investments, including reporting requirements and strategic use-of-proceeds alignment conditions, though no specific conditions are publicly disclosed. SI001
CI035 Forbes Asia describes Rebellions as "a minnow by comparison" relative to Nvidia, whose data center segment generated $35.6 billion in Q4 FY2025 revenue—more than Rebellions' total raised to date. SI008, SI003
CI036 Forbes Asia states that persuading data centers to source AI chips from companies other than Nvidia is "almost as difficult as making the chips themselves," reflecting the adverse market access challenge Rebellions faces in converting capital into sustainable revenue. SI008, SI003
CI037 Rebellions has not released FY2024 or FY2025 revenue, gross margin, operating loss, customer count, or ARR; the last confirmed public revenue figure is FY2023's 2.7 billion KRW (~$2.1M), which predates the SAPEON merger and all 2025–2026 fundraising. SI001, SI003, SI008
CI038 The $2.34 billion pre-IPO valuation cannot be validated against revenue or earnings multiples using publicly available data, as FY2024–2025 financials have not been disclosed—representing a diligence blocker for any external investor forming a conviction view. SI001, SI003, SI011
CI039 No customer count, customer concentration data (e.g., top-3 revenue share), or churn figures have been publicly disclosed by Rebellions, preventing assessment of revenue diversification or dependency risk. SI001, SI004, SI008
CI040 Forbes Asia (April 2025) noted that Rebellions was "unwilling to disclose how many Rebel chips it plans to produce," citing confidentiality agreements with customers, confirming deliberate opacity on production volume. SI008, SI001
CE001 The ATOM-Max card (RBLN-CA25) delivers 128 TFLOPS FP16, 512 TOPS INT8, and 1024 TOPS INT4 from four ATOM chips in a multi-die package, with 64 GB GDDR6 at 1024 GB/s, at 350 W TDP, in a PCIe Gen5 x16 full-height full-length dual-slot form factor. SE001, SE019
CE002 The ATOM-Max Server integrates eight ATOM-Max cards in a 4U chassis, providing 1024 TFLOPS FP16 aggregate, 512 GB GDDR6, and 8 TB/s total memory bandwidth at 3.4 kW typical / 4.3 kW maximum power. SE002
CE003 The ATOM-Max POD is an 8-server mini-cluster providing 64 NPUs and a 400 GB/s RDMA inter-node fabric for distributed tensor-parallel inference. SE003
CE004 Rebellions produces two single-chip ATOM variants: RBLN-CA21 (one ATOM chip, <75 W, no external power connector required) and RBLN-CA22 (one ATOM chip, up to 90 W, requires external power). SE019
CE005 Each ATOM SoC die delivers 32 TFLOPS FP16, 128 TOPS INT8, and 256 TOPS INT4, with 16 GB GDDR6 at 256 GB/s per chip (16 Gbps data rate on a 128-bit bus). SE019, SE001
CE006 The ATOM-Max card uses a PCIe Gen5 x16 host interface in a full-height full-length dual-slot board form factor. SE001
CE007 The ATOM-Max product family began shipping in H1 2024; the ATOM SoC was taped out in June 2022 and first shipped in May 2023 per earlier research. SE026
CE008 The REBEL-Quad uses a quad-chiplet architecture with UCIe-Advanced die-to-die interconnect operating at 16 Gbps per lane, fabricated on Samsung Foundry's 4nm SF4X process node. SE005, SE006, SE018
CE009 The REBEL-Quad integrates four HBM3E memory stacks (each 36 GB at 9.6 GT/s), providing 144 GB total HBM3E capacity and 4.8 TB/s aggregate memory bandwidth. SE019, SE006
CE010 The ISSCC 2026 peer-reviewed paper (IEEE Xplore document 11409003) reports the REBEL-Quad achieves 56.8 tokens per second on LLaMA v3.3 70B with 2,048-token input and 2,048-token output sequences. SE012, SE006
CE011 The REBEL-Quad package is built on Samsung SF4X and CoWoS-S, integrating four compute ASICs, four HBM3E sites, and four integrated silicon capacitors (ISC) on a single package. SE018, SE012
CE012 At Hot Chips 2025 (August 27, 2025), Rebellions demonstrated a live LLaMA 3.3 70B inference session on REBEL-Quad hardware and showed Qwen3 235B MoE execution capability, independently observed by trade press. SE018, SE006
CE013 The RebelServer is a 5U system integrating eight RebelCard accelerators, rated up to 2 PFLOPS FP8, with dual AMD EPYC 9355 (32-core/64-thread) host CPUs, 1.5 TB DDR5 host memory, 4× 400G networking ports, consuming 4–6 kW typical and 7 kW maximum. SE004
CE014 RebelRack and RebelPOD were announced and described as 'available now' in the March 30, 2026 press release accompanying the $400M pre-IPO funding announcement. SE007, SE011
CE015 As of May 2026, the Rebellions website navigation shows both RebelRack and RebelPOD as 'Coming Soon', contradicting the March 30, 2026 press release that described them as 'available now'. SE007, SE011
CE016 RBLN SDK version 0.10.3 was released in May 2026 and includes three packages: rebel-compiler 0.10.3, optimum-rbln 0.10.3, and vllm-rbln 0.10.3.post1. SE013, SE015
CE017 The vllm-rbln 0.10.3.post1 package was published to the public Python Package Index (pypi.org) on May 18, 2026, extending availability beyond the proprietary pypi.rbln.ai mirror. SE015
CE018 Rebellions' NPU is listed as a supported hardware backend in the official vLLM documentation (docs.vllm.ai) alongside Google Cloud TPU, Intel Gaudi, and AMD Instinct, granting it first-class integration tier status. SE014
CE019 The vllm-rbln package does not yet support three features that are available in GPU vLLM: speculative decoding, distributed KV cache (cross-node memory sharing), and prefill/decode disaggregation — all listed as in-development in the May 2026 SDK release notes. SE013, SE014
CE020 The Rebellions model zoo includes 300+ supported models spanning LLMs, vision transformers, and classical inference workloads across PyTorch and TensorFlow frameworks. SE008
CE021 Rebellions' Kubernetes NPU Operator version 0.4.0 is certified for Red Hat OpenShift AI (December 2025 partnership), with Helm chart distributed via OCI registry. SE009, SE006
CE022 The RBLN SDK runtime identifies ATOM devices as rebellions.ai/ATOM (RBLN-CA* series) and REBEL devices as rebellions.ai/REBEL (RBLN-CR* series) for Kubernetes resource allocation. SE013
CE023 The Rebellions GitHub organization (rebellions-sw) contains only archived or internal repositories; no public SDK source code, compiler binary, or model code is published as of May 2026. SE016
CE024 As of May 2026, Rebellions has not submitted results to any MLCommons MLPerf Inference Datacenter benchmark round; its hardware does not appear in the published leaderboard. SE017
CE025 All Rebellions perf/watt performance claims for ATOM deployments (Mongolia: 2.7× TPS/Watt vs GPU; UAE: 2× performance-per-watt) are either company-cited or certified by TTA (Korea's Telecommunications Technology Association) — no independent international benchmark corroboration exists. SE025, SE017
CE026 Samsung is Rebellions' sole vendor for all three critical REBEL-Quad supply inputs: logic fabrication (4nm SF4X process), HBM3E memory (4× 36 GB stacks), and advanced packaging (CoWoS-S); no disclosed alternative source exists for any of these inputs. SE010, SE023, SE018, SE028
CE027 Alphawave Semi supplies the UCIe-Advanced SerDes IP enabling 16 Gbps per-lane die-to-die connectivity within the REBEL-Quad package. SE018
CE028 Rebellions joined the Arm Total Design ecosystem in October 2025 with plans to integrate Arm's Neoverse compute subsystem (CSS) into a future AGI CPU product alongside its NPU. SE022, SE027
CE029 Red Hat and Rebellions announced an OpenShift AI partnership in December 2025, enabling enterprise Kubernetes deployments on Rebellions NPU hardware with certified operator support. SE009, SE006
CE030 In April 2026, Rebellions, SK Telecom, and DOCOMO Innovations announced a partnership for sovereign AI datacenter validation and AI infrastructure co-development in Korea and Japan. SE024
CE031 The REBEL-Quad uses a dual PCIe Gen5 x16 host interface; ServeTheHome notes this may lag NVIDIA's GB300 which is expected to support PCIe Gen6, representing a generational I/O bandwidth difference for host-bottlenecked workloads. SE018
CE032 The ATOM-Max card integrates four ATOM SoC dies in a multi-die package (MDP), each contributing 32 TFLOPS FP16, 128 TOPS INT8, 256 TOPS INT4, and 16 GB GDDR6 at 256 GB/s bandwidth. SE019, SE001
CE033 The RSD (Rebellions Scalable Design) architecture enables tensor-parallel inference scaling from a single ATOM card (128 TFLOPS FP16 / 64 GB) to a configurable rack system (512–7168 TFLOPS FP16, 256 GB–3.5 TB GDDR6) with Kubernetes and OpenStack support. SE025
CE034 Rebellions ATOM cards are deployed at ECOPEACE in the UAE for water robot vision AI, achieving 2× performance-per-watt versus a GPU baseline per TTA certification. SE025
CE035 Rebellions ATOM hardware is deployed at Mongolia's customs authority, achieving 2.7× TPS/Watt versus a GPU baseline for customs AI inference, per company-cited data. SE025
CE036 Rebellions presented the ATOM SoC at ISSCC 2024 alongside AMD and Intel papers; the chip was described as mass-production ready at the time of the conference. SE021, SE020
CE037 The RBLN SDK supports inference in FP32, FP16, FP8, FP6, and FP4 numeric precision formats, covering the full precision range used by leading LLM deployments. SE013
CE038 The REBEL chip was taped out in November 2024 on Samsung Foundry's 4nm SF4X process, making it one of the first AI inference chiplets in production on this node. SE005, SE010
CE039 The RBLN SDK identifies ATOM hardware as device type RBLN-CA* and REBEL hardware as RBLN-CR*, exposed as Kubernetes device resources under the rebellions.ai resource namespace. SE013, SE009
CE040 Arm made a strategic investment in Rebellions as part of the December 2024 $250M Series C, marking Arm's first disclosed investment in an AI chip startup, alongside Samsung Securities. SE022, SE020
CU001 Rebellions' official solutions surface targets telecom, sovereign AI, enterprise AI, and data-center buyers. SU023
CU002 Rebellions' product pages present ATOM Max Server and Rebel Server as server-level inference infrastructure products. SU024, SU025
CU003 Rebellions said KT and kt cloud were lead investors and first customers in its Series B round. SU002, SU005
CU004 Rebellions said ATOM mass production was intended for deployment into KT's data center. SU002
CU005 Korea JoongAng Daily reported that KT had invested a cumulative 66.5 billion won in Rebellions by May 2024. SU005
CU006 Korea JoongAng Daily reported that Rebellions provided ATOM chips to KT's data center to run cloud services. SU005
CU007 Rebellions' about page quotes SK Telecom's Tony Ha saying SKT is deploying Rebellions' NPU in A., Korea's largest LLM service. SU001
CU008 Rebellions' April 2025 announcement with SK Telecom and DOCOMO Innovations describes next-generation AI infrastructure collaboration rather than a broad named paying-customer roster. SU006
CU009 SK Telecom's 2026 newsroom post says SKT, Arm, and Rebellions plan to optimize the A.X K1 model on a jointly developed REBEL server for sovereign AI. SU007
CU010 Business Wire corroborated that the SKT-Arm-Rebellions collaboration targets sovereign AI and telecom infrastructure. SU018
CU011 Independent trade coverage from HPCwire and Converge Digest framed the 2026 SKT-Arm-Rebellions announcement as telco sovereign-AI infrastructure collaboration rather than a broad new customer roster. SU019, SU021
CU012 Forbes reported that Rebellions' blue-chip client list included the cloud divisions of SK, Kakao, and Naver. SU004
CU013 Forbes reported that Rebellions had signed its first Saudi Aramco deal. SU004
CU014 Forbes reported that Rebellions had customer orders from the United States, Japan, and Thailand by the end of 2024. SU004
CU015 Rebellions' Series C release says ATOM is deployed with customers across Japan, Saudi Arabia, and the United States. SU008, SU004
CU016 Rebellions' pre-IPO release says commercial deployments are already live across enterprises and governments. SU009
CU017 Neither the Series C release nor the pre-IPO release publicly names the U.S. customer or discloses total customer count. SU008, SU009
CU018 Neither the pre-IPO release nor the Rebel Server product page publicly confirms a named customer delivery for RebelRack or RebelPOD. SU009, SU025
CU019 Rebellions' Wa'ed extension release says Wa'ed Ventures is the venture capital arm of Saudi Aramco. SU003
CU020 Rebellions' Wa'ed extension release says the investment was intended to establish a strategic Middle East bridgehead. SU003
CU021 Wamda described Rebellions' Series C as supporting AI chip deployment in Saudi Arabia. SU017
CU022 PR Newswire said Marshall Choy joined as Chief Business Officer to strengthen Rebellions' customer-centric strategy and that the company established a U.S. entity. SU016
CU023 KoreaTechDesk also described a U.S. expansion push tied to Rebellions' global customer strategy. SU015
CU024 Rebellions and Red Hat announced OpenShift AI powered by Rebellions NPUs as an enterprise deployment option. SU010
CU025 Red Hat's own press release corroborates the OpenShift AI partnership as an enterprise channel and ecosystem route to market. SU011, SU010
CU026 TechPowerUp, BusinessKorea, and Seoul Economic Daily independently covered the Red Hat-Rebellions launch. SU012, SU013, SU014
CU027 Rebellions' solutions and product surfaces support multiple customer motions spanning telco data centers, sovereign AI, enterprise AI, and packaged infrastructure. SU023, SU024, SU025
CU028 The Investor reported that the SAPEON merger created Korea's first AI-chip unicorn and deepened SK Telecom ecosystem alignment. SU022
CU029 Rebellions announced a $250 million Series C in 2025 backed by Arm and Samsung. SU008
CU030 Rebellions announced a $400 million pre-IPO round in 2026 and launched RebelRack and RebelPOD for global expansion. SU009
CU031 Korea JoongAng Daily said Rebellions had been in meetings with Google and Japanese telcos. SU005
CU032 Korea JoongAng Daily said IBM was conducting a qualification test on Rebellions' chips. SU005
CU033 Public sources reviewed for this chapter do not disclose Rebellions' total paid customer count. SU004, SU008, SU009
CU034 Public sources reviewed for this chapter do not disclose Rebellions' NRR, GRR, logo churn, or renewal rate. SU001, SU004, SU009
CU035 Public sources reviewed for this chapter do not disclose contract length, renewal cadence, or expansion spend by named account. SU004, SU009, SU011
CU036 The clearest production proofs in public sources remain KT/kt cloud and SK Telecom, while international evidence is more often third-party reported or unnamed. SU001, SU002, SU004, SU008
CU037 Forbes said Rebellions' international expansion would be tricky because Nvidia remains deeply entrenched. SU004
CU038 Korea JoongAng Daily said persuading data centers to buy chips other than Nvidia was almost as difficult as making the chips. SU005
CU039 The combination of sparse named international customers and concentrated Korean telco proof implies material customer concentration risk. SU002, SU004, SU008
CU040 Public evidence supports classifying KT/kt cloud and SK Telecom as production proof, DOCOMO and SKT-Arm announcements as validation proof, Red Hat as channel proof, and U.S./Japan/Thailand orders as partly unnamed pipeline or geography-level proof. SU001, SU002, SU004, SU006, SU007, SU008, SU010, SU011
CR001 The May 2026 MLPerf Inference Datacenter results include major incumbent vendors but no Rebellions submission. SR009
CR002 The IEEE ISSCC 2026 paper reports 56.8 TPS on LLaMA v3.3 70B for Rebellions’ quad-chiplet AI SoC under the paper’s stated sequence settings. SR010
CR003 ServeTheHome reported that REBEL-Quad uses a Samsung SF4X chiplet design with 144 GB HBM3E and was shown running LLaMA 3.3 70B live at Hot Chips 2025. SR011
CR004 Reviewed public sources do not provide an apples-to-apples third-party benchmark or TCO table against Nvidia H100 or AMD MI300-class hardware. SR009, SR010, SR011
CR005 Korea JoongAng Daily wrote that persuading data centers to source chips from non-Nvidia vendors is almost as difficult as making the chips. SR002
CR006 Forbes described Rebellions’ sales challenge, especially internationally, as a tricky proposition. SR001
CR007 Forbes reported FY2023 revenue of about 2.7 billion KRW, or roughly $2.1 million. SR001
CR008 Forbes reported that Sunghyun Park targeted 100 billion KRW of revenue for 2025. SR001
CR009 No reviewed 2026 source disclosed FY2025 actual revenue, gross margin, or customer shipment figures. SR001, SR004, SR014, SR023
CR010 Rebellions’ March 2026 pre-IPO release said the company raised $400 million, reached $850 million of cumulative funding, and was valued at about $2.34 billion. SR004, SR014, SR024
CR011 CNBC independently reported that Rebellions raised $400 million ahead of its IPO. SR014
CR012 Seoul Economic Daily reported that Rebellions was targeting KOSPI preliminary review in August 2026 at a valuation of roughly 3.4 trillion won. SR023
CR013 KRX is the listing venue that administers Korea’s review process, so IPO timing depends on an external approval gate rather than a unilateral company decision. SR022, SR023
CR014 No KRX preliminary approval notice, prospectus, or audited public filing was identified in the reviewed sources as of 2026-05-20. SR022, SR023, SR024
CR015 The Federal Register published the Framework for Artificial Intelligence Diffusion on 2025-01-15, showing that advanced AI compute and model diffusion remained active U.S. policy territory in 2025. SR015
CR016 BIS issued a May 13 2025 policy statement on training AI models, indicating that model training and advanced chip access remained explicit export-control concerns. SR016
CR017 CSIS found that allied legal authority to implement AI and semiconductor export controls is uneven, which increases cross-border compliance complexity. SR017
CR018 The National Law Review described BIS as issuing four key updates on advanced computing and AI export controls in May 2026. SR018
CR019 Wiley said BIS announced a new regulatory framework for AI and controls on advanced computing technology and AI models in January 2025. SR019
CR020 WilmerHale described the December 2024 BIS package as sweeping additional restrictions on semiconductors and advanced computing. SR020
CR021 Morrison Foerster framed export-control risk management as a live issue across the AI chip ecosystem in February 2026. SR021
CR022 Rebellions’ 2025 and 2026 releases describe expansion into Japan, Saudi Arabia, and the United States, making export classification and end-use screening commercially relevant. SR004, SR005
CR023 Reviewed legal and regulatory sources did not identify a disclosed enforcement action or active litigation involving Rebellions. SR015, SR016, SR018, SR019, SR020, SR021
CR024 The legal overhang is compliance and disclosure risk rather than known active litigation. SR018, SR019, SR020, SR021
CR025 Samsung said it partnered with Rebellions on the next-generation REBEL AI chip using Samsung 4nm foundry and HBM3E memory. SR012, SR011
CR026 Independent Hot Chips coverage still ties REBEL-Quad to Samsung-process and HBM3E concentration in the current roadmap. SR011, SR012
CR027 Forbes quoted Park acknowledging shortage in foundry capacity and shortage of HBM as structural constraints. SR001
CR028 No reviewed public source disclosed a second-source foundry or alternative HBM supply plan for REBEL-family products. SR001, SR004, SR011, SR012
CR029 The March 2026 launch of RebelRack and RebelPOD was official, but reviewed sources did not independently verify production volume, commercial availability at scale, or installed-base size. SR004, SR014
CR030 Rebellions and SAPEON signed a definitive merger agreement in August 2024. SR003, SR007
CR031 Rebellions said the SAPEON merger completed in December 2024 and created Korea’s first AI-chip unicorn. SR003, SR008
CR032 The merged entity is publicly led by CEO Sunghyun Park. SR008, SR013
CR033 Public sources did not disclose a quantified synergy plan, overlap reduction target, or post-merger governance framework. SR003, SR007, SR008, SR013
CR034 Rebellions’ about page and merger-completion release make Park the named technical and corporate face of the combined company. SR008, SR013
CR035 No public succession plan was identified in the reviewed source set. SR008, SR013
CR036 Wowtale said the pre-IPO round was the first direct investment under the Korea National Growth Fund. SR024
CR037 The official pre-IPO narrative also positioned the round within the K-Nvidia industrial-policy theme. SR004, SR024
CR038 Graphcore announced that it joined SoftBank Group, showing one notable AI-chip startup ended as a strategic acquisition rather than a standalone public-scale winner. SR025, SR026
CR039 EE Times separately described Graphcore as acquired by SoftBank. SR026
CR040 DataCenterDynamics reported SambaNova explored a sale after struggling to secure further funding, with BlackRock marking shares down to about $2.4 billion from a $5 billion peak. SR027
CR041 Forbes and BusinessWire reported FuriosaAI closed a $125 million round to scale production of its next-generation inference chip. SR028, SR029
CR042 Tech Funding News reported FuriosaAI was seeking up to $500 million for next-generation AI chips, indicating active capital competition inside Korea. SR030
CR043 Graphcore’s acquisition, SambaNova’s sale exploration, and SemiAnalysis’ warning that few startups reach meaningful production support a view that AI-chip survival is concentrating among a few well-funded or strategically backed players. SR025, SR026, SR027, SR031
CR044 Burn rate, gross margin, unit shipment, and audited FY2024-FY2025 financials remain undisclosed in reviewed public sources, leaving runway assessment qualitative rather than underwritten. SR001, SR004, SR014, SR023
CR045 The reviewed sources did not describe a public export-control compliance program, reseller screening policy, or classification letter for Rebellions products. SR004, SR005, SR006, SR015, SR016, SR021
CR046 The highest-severity thesis-break triggers are continued benchmark opacity into IPO filing, a failed or delayed KRX review, a Samsung or HBM supply slip, new export-control restrictions on key markets, or leadership disruption around Park. SR009, SR022, SR023, SR027
CR047 SK Telecom announced an April 2026 MOU with Arm and Rebellions for next-generation AI infrastructure and sovereign-AI deployment. SR006
CR048 The SK Telecom and Arm relationship deepens Rebellions’ dependence on a small number of strategic ecosystem sponsors for flagship sovereign-AI wins. SR005, SR006
CR049 Public-market timing, policy capital, and flagship sovereign deployments are all external dependencies outside Rebellions’ direct product roadmap. SR022, SR023, SR024, SR006
CV001 Rebellions closed a $400 million pre-IPO financing round in March 2026. SV001, SV002, SV003
CV002 Rebellions said the March 2026 pre-IPO round valued the company at approximately $2.34 billion post-money. SV001, SV002, SV003, SV005
CV003 Mirae Asset Financial Group and the Korea National Growth Fund were identified as key participants in the March 2026 pre-IPO round. SV001, SV002, SV005
CV004 The March 2026 pre-IPO followed Rebellions' $250 million Series C round announced in September 2025. SV001, SV002, SV009
CV005 Rebellions' total disclosed equity funding reached $850 million after the March 2026 pre-IPO round. SV001, SV002, SV003
CV006 Rebellions' September 2025 Series C announcement stated a $1.4 billion valuation. SV009
CV007 The step-up from a $1.4 billion Series C valuation to a $2.34 billion pre-IPO valuation was approximately 67%. SV001, SV002, SV009
CV008 Seoul Economic Daily reported in April 2026 that Rebellions was gearing up for a KOSPI IPO as Korea's AI-chip listing race accelerated. SV004
CV009 No public IPO filing, price range, or audited prospectus was identified in the sources reviewed for this chapter as of 2026-05-20. SV004, SV005, SV006
CV010 Korea Exchange operates the KOSPI market that would ultimately require fuller public disclosure once a listing is filed. SV006
CV011 Forbes Asia reported that Rebellions posted FY2023 revenue of 2.7 billion KRW, roughly $2.1 million. SV007
CV012 Forbes Asia reported that Rebellions posted a FY2023 net loss of 13.7 billion KRW. SV007
CV013 The public sources reviewed for this chapter do not disclose Rebellions' FY2024 or FY2025 actual revenue. SV003, SV004, SV007, SV008
CV014 Forbes Asia reported that management targeted approximately 100 billion KRW of revenue for FY2025. SV007
CV015 The sources reviewed for this chapter do not confirm whether Rebellions achieved the stated FY2025 revenue target. SV003, SV004, SV007
CV016 Rebellions' $2.34 billion valuation implies an estimated revenue multiple above 1,100x on the last disclosed FY2023 revenue baseline. SV002, SV007
CV017 The disclosed FY2023 financial baseline is stale for point-multiple underwriting because it predates the SAPEON merger and later product launches. SV007, SV009
CV018 Rebellions' current price support rests mainly on sovereign-AI optionality, Samsung ecosystem access, and IPO scarcity rather than on disclosed commercial scale. SV001, SV004, SV008, SV009
CV019 Korea JoongAng Daily described Rebellions' Samsung relationship as a turnkey setup spanning memory, manufacturing, and packaging. SV008
CV020 Rebellions' September 2025 Series C announcement said ATOM and ATOM-Max were in mass production and deployed with customers in Japan, Saudi Arabia, and the United States. SV009
CV021 Rebellions' March 2026 announcement said the company launched RebelRack and RebelPOD to support global expansion. SV001
CV022 The public sources reviewed for this chapter do not disclose customer count, backlog, gross margin, or ARR for Rebellions. SV003, SV004, SV007
CV023 The disclosure gap around recent scale and economics makes bullish underwriting low-confidence at the current price. SV003, SV004, SV007, SV008
CV024 TechCrunch reported that Cerebras filed for an IPO in April 2026. SV011
CV025 Cerebras' May 2026 SEC filing disclosed FY2025 revenue of $510 million. SV013
CV026 CompaniesMarketCap indicated Cerebras traded at roughly an $8 billion-plus market cap in May 2026, implying a mid-teens multiple on FY2025 revenue. SV013, SV014
CV027 Groq announced a $750 million round, and Data Center Dynamics reported a $6.9 billion valuation for that financing. SV015, SV016, SV017, SV018
CV028 The Groq sources retained for this chapter do not disclose revenue, so Groq is an optionality anchor rather than a clean revenue-multiple comp. SV015, SV016, SV017, SV018
CV029 Data Center Dynamics and TechStartups reported that SambaNova explored a sale after struggling to secure further funding in late 2025. SV019, SV020
CV030 SambaNova's reported distress shows that private AI-chip valuations can compress sharply when commercialization and financing do not keep pace. SV019, SV020
CV031 Graphcore joined SoftBank in 2024, ending its path as an independent public-market AI-chip comparable. SV021, SV022
CV032 Graphcore's outcome reinforces the downside case for standalone AI-chip startups that fail to scale enough to remain strategically independent. SV021, SV022
CV033 Forbes and BusinessWire reported that FuriosaAI closed a $125 million funding round in July 2025. SV023, SV025
CV034 BizChosun and TechFundingNews reported that FuriosaAI was also seeking a larger raise of up to $500 million earlier in 2025. SV024, SV026
CV035 FuriosaAI is the closest Korean peer in geography and product category, but its missing public revenue disclosure limits precise multiple comparison. SV023, SV024, SV025, SV026
CV036 Ars Technica and Crunchbase News reported that Tenstorrent raised $693 million and reached unicorn status in late 2024. SV027, SV028
CV037 Tenstorrent shows investors still fund AI-chip platform optionality at multibillion-dollar levels even without public revenue disclosure. SV027, SV028
CV038 Across Groq, SambaNova, Graphcore, FuriosaAI, and Tenstorrent, private AI-chip valuations reflect optionality and strategic positioning more than transparent financial comparability. SV016, SV019, SV020, SV022, SV025, SV027, SV028
CV039 Relative to that peer set, Rebellions sits within the private valuation band but looks expensive relative to its last disclosed revenue base. SV007, SV016, SV019, SV020, SV023, SV027
CV040 A bear case with limited commercialization, no meaningful audited scale disclosure, and IPO slippage supports an estimated valuation range of $0.8 billion to $1.2 billion. SV007, SV019, SV020, SV022
CV041 A base case with real but still modest commercialization evidence and workable IPO readiness supports an estimated valuation range of $1.5 billion to $2.5 billion. SV001, SV004, SV009, SV016, SV023
CV042 A bull case with strong sovereign-AI deployment proof, cleaner disclosures, and strong IPO demand supports an estimated valuation range of $3 billion to $5 billion. SV001, SV004, SV009, SV023, SV027
CV043 The current $2.34 billion mark already sits near the upper half of the base-case band, leaving limited room for execution misses. SV001, SV002, SV007, SV016
CV044 Because the current price is supported mostly by optionality while disclosed commercial scale is thin, the appropriate recommendation is research-more rather than buy. SV007, SV013, SV016, SV020, SV022
CV045 Confidence should be low because FY2024-FY2025 revenue, gross margin, backlog, and cap-table terms remain undisclosed. SV004, SV005, SV007
CV046 The appropriate risk rating is high and the valuation stance is stretched. SV007, SV019, SV020, SV022
CV047 At the current mark, a venture-style greater-than-2x return is hard to justify without either a lower entry price or new evidence of commercial scale. SV007, SV013, SV016
CV048 Public evidence reviewed for this chapter does not reveal the pre-IPO preference stack, liquidation terms, or secondary-liquidity structure. SV001, SV002, SV004, SV005
CV049 A clean thesis-break trigger would be audited or IPO-filed revenue that lands materially below what a $2.34 billion price implies. SV007, SV013
CV050 A second thesis-break trigger would be a delayed or withdrawn IPO process without offsetting evidence of scaled commercial deployments. SV004, SV005, SV006
CV051 Final diligence priorities are audited FY2024-FY2025 financials, cap-table terms, customer concentration, backlog, and gross margin. SV004, SV005, SV007
CV052 Public evidence does not support underwriting a strategic M&A angle today, so it should remain a diligence question rather than a valuation support pillar. SV004, SV008, SV009
CV053 Because most private comparables lack disclosed revenue, scenario ranges are more defensible than point-estimate multiples for Rebellions. SV016, SV019, SV020, SV022, SV025, SV027, SV028
来源
编号出版方标题引文
SO001 Rebellions About - Rebellions "Rebellions builds purpose-engineered AI accelerators to redefine energy-efficiency and scale in the age of large-scale AI."
SO002 Rebellions Korean AI chipmaker Rebellions Closes $124M Series B Fundraise
SO003 Rebellions South Korea's foremost AI Chip company 'Rebellions' Secures $15M Series B Extension round by Wa'ed Ventures
SO004 Rebellions Rebellions and SAPEON Korea Sign Definitive Merger Agreement
SO005 Rebellions Rebellions and SAPEON Korea Complete Merger, Launching Korea's First AI Chip Unicorn "Leading the new entity is CEO Sunghyun Park, an expert in AI and system semiconductors with a Ph.D. from MIT's Computer Science and AI Lab (CSAIL). With experience spanning Intel, Space X, and Morgan Stanley."
SO006 Rebellions Rebellions Closes $400 Million Pre-IPO and Launches RebelRack™ and RebelPOD™ to Accelerate Global Expansion "Rebellions has raised $400 million in a pre-IPO funding round led by Mirae Asset Financial Group and the Korea National Growth Fund. The round follows the company's $250 million Series C in September 2025, bringing total funding to $850 million and valuing the company at approximately $2.34 billion."
SO007 The Investor Rebellions, Sapeon Korea merges to challenge Nvidia "Rebellions, a fabless chip designer valued at 900 billion won ($664 million), and Sapeon Korea, the AI chip supplier for data centers valued at 550 billion won... The global AI chip market was valued at $34.3 billion in 2023, and Nvidia took 94 percent of the market share."
SO008 Forbes South Korea's AI Chip Champion Is Poised To Carve Out Global Niche
SO009 The Korea Times Rebellions seeks to become first Korean startup to mass-produce AI chips for language models
SO010 SK Telecom Newsroom SK Telecom Signs MOU with Arm and Rebellions for Next-Generation AI Infrastructure Innovation
SO011 PR Newswire Rebellions Closes $400 Million Pre-IPO and Launches RebelRack™ and RebelPOD™ to Accelerate Global Expansion
SO012 Korea JoongAng Daily Why Rebellions' strategy against Nvidia is not so outrageous "In the Nvidia-dominated world of AI chips, a five-year-old startup from Korea, Rebellions, has some outrageous ideas to win the game."
SO013 CNBC Samsung-backed AI chip firm Rebellions raises $400 million ahead of IPO
SO014 eeNews Europe Merger terms agreed to form Korean AI unicorn
SO015 BusinessKorea SKT, Rebellions Team Up to Build Sovereign AI Ecosystem
SO016 HPCwire Rebellions Collaborates with SK Telecom and Arm Targeting Sovereign AI
SO017 Converge Digest Rebellions, Arm, and SK Telecom Partner on Sovereign AI for Telco Data Centers
SO018 SK Telecom SAPEON Korea and Rebellions Inc. Sign Definitive Merger Agreement
SO019 Rebellions Rebellions Scales Global Growth with Silicon Valley-backed Series C
SO020 Rebellions Rebellions Accelerates Global Expansion and Strengthens Customer-centric Strategy with Significant Executive Appointments
SO021 Rebellions Rebellions and Red Hat Introduce Red Hat OpenShift AI Powered by Rebellions NPUs
SO022 Rebellions Rebellions Raises $250 Million to Advance the Next Generation AI Infrastructure, Backed by Arm and Samsung "Rebellions, Asia's fastest-growing AI inference chip company, today announced it has raised $250 million in Series C round at a valuation of $1.4 billion."
SO023 Rebellions Rebellions Debuts REBEL-Quad at Hot Chips 2025, Breaking AI's Energy Tax with High-Performance Chiplet Innovation
SO024 Rebellions Rebellions and Samsung Electronics Partner to Target the Emerging Generative AI Market with Next-Generation AI Chip Co-development
SO025 Rebellions Rebellions, SK Telecom, and DOCOMO Innovations Partner to Accelerate Next-Gen AI Infrastructure
SM001 U.S. Securities and Exchange Commission (Nvidia Corporation) NVIDIA Annual Report on Form 10-K — Fiscal Year 2026 (Year Ended January 25, 2026)
SM002 Epoch AI (Jaime Sevilla) Trends in the dollar training cost of machine learning systems
SM003 NVIDIA Corporation NVIDIA Data Center — Accelerated Computing for AI and HPC
SM004 Arm Ltd. AI on Arm: Secure, Scalable Intelligence from ML to Agentic AI
SM005 BusinessKorea Samsung SDS Dominates Korean Cloud Market, According to IDC Report
SM006 Rebellions Rebellions — Power AI Inference. Efficiently. At Scale.
SM007 Rebellions Rebellions Newsroom
SM008 U.S. Securities and Exchange Commission (Nvidia Corporation) EDGAR Filing Index — Nvidia FY2026 10-K (Acc-no 0001045810-26-000021)
SM009 The Investor (Herald Corp) Rebellions, SAPEON to merge to form Korea's 1st AI chip unicorn
SM010 Forbes South Korea's AI Chip Champion Is Poised To Carve Out Global Niche
SM011 Korea Times Rebellions seeks to become first Korean startup to mass-produce AI chips for language models
SM012 CNBC AI chip startup Rebellions raises $400 million ahead of IPO
SM013 eeNews Europe Merger terms agreed to form Korean AI unicorn
SM014 BusinessKorea Korea's Sovereign AI Drive: SK Telecom, Rebellions, Arm Collaborate on AI Chip
SM015 HPCwire Rebellions Collaborates with SK Telecom and Arm Targeting Sovereign AI
SM016 Converge Digest Rebellions, Arm and SK Telecom Partner on Sovereign AI for Telco Data Centers
SM017 SK Telecom SK Telecom AI Partnership Press Release
SM018 Rebellions Rebellions Scales Global Growth with Silicon Valley-Backed Series C
SM019 Rebellions Rebellions Raises $250 Million to Advance Next-Generation AI Infrastructure, Backed by Arm and Samsung
SM020 Rebellions Rebellions and Samsung Electronics Partner to Target the Emerging Generative AI Market with Next-Generation AI Chip Co-Development
SM021 Rebellions Rebellions Debuts REBEL Quad at Hot Chips 2025, Breaking AI's Energy Tax
SM022 Rebellions Rebellions and Red Hat Introduce Red Hat OpenShift AI Powered by Rebellions NPUs
SM023 Rebellions Rebellions, SK Telecom, and DOCOMO Innovations Partner to Accelerate Next-Gen AI Infrastructure
SM024 SK Telecom SK Telecom AI and Rebellions Partnership — Official Press Release
SM025 PR Newswire Rebellions Closes $400 Million Pre-IPO and Launches RebelRack and RebelPod to Accelerate Global Expansion
SP001 Groq Groq is fast, low cost inference. Groq pioneered the LPU in 2016, the first chip purpose-built for inference.
SP002 Groq Groq Raises $750 Million as Inference Demand Surges $750 million in new financing at a post-money valuation of $6.9 billion.
SP003 Groq Groq and Nvidia Enter Non-Exclusive Inference Technology Licensing Agreement to Accelerate AI Inference at Global Scale Jonathan Ross, Groq's Founder, Sunny Madra, Groq's President, and other members of the Groq team will join Nvidia to help advance and scale the licensed technology.
SP004 Groq Groq Newsroom
SP005 Groq Groq Solidifies Status as Emerging Hyperscaler with New Global Deployment Groq has been named an official inference provider for HUMAIN, a newly launched AI company headquartered in Saudi Arabia.
SP006 Groq Saudi Arabia Announces $1.5 Billion Expansion to Fuel AI-powered Economy with AI Tech Leader Groq Silicon Valley AI pioneer Groq has secured a $1.5 billion commitment from the Kingdom of Saudi Arabia (KSA) for expanded delivery of its advanced LPU-based AI inference infrastructure.
SP007 Tenstorrent Tenstorrent
SP008 Tenstorrent Wormhole AI Accelerator Cards
SP009 Tenstorrent Newsroom | Tenstorrent
SP010 Tenstorrent Tenstorrent Raises $693 Million in Series D Led by Samsung Securities and LG Electronics
SP011 SambaNova Systems SambaNova | The Fastest AI Inference Platform SambaNova leads on speed at 435 output tokens/s, >3x faster than any other provider.
SP012 SambaNova Systems RDU | Next-Gen AI Chip for Inference at Scale The SN50 RDU (Reconfigurable Dataflow Unit) is SambaNova's fifth-generation AI inference processor, designed specifically for large-scale, agentic workloads.
SP013 SambaNova Systems SambaNova Blog
SP014 Cerebras Systems Cerebras
SP015 Cerebras Systems Product — Chip — Cerebras Four trillion transistors. 125 petaflops. One silicon wafer.
SP016 FuriosaAI Homepage — FuriosaAI RNGD (pronounced 'Renegade') delivers high-performance LLM and multimodal deployment capabilities while maintaining a radically efficient 180W power profile.
SP017 FuriosaAI Newsroom — FuriosaAI AI Chip Startup FuriosaAI Plans $500 Million Round Before IPO
SP018 Hailo Hailo AI on the Edge Processors | Edge AI Chip Solutions Hailo offers breakthrough AI processors uniquely designed to enable high performance deep learning applications on edge devices.
SP019 AMD AMD Instinct MI300 Series Accelerators Meet the AMD Instinct MI325X Accelerators... industry leading 256 GB HBM3E memory and 6 TB/s bandwidth.
SP020 Nvidia NVIDIA H100 GPU H100 extends NVIDIA's market-leading inference leadership... Inference by up to 30X.
SP021 Nvidia NVIDIA GB200 NVL72 30x faster real-time trillion-parameter large language model (LLM) inference... 25x more performance at the same power.
SP022 Google Cloud Tensor Processing Units (TPUs) Ironwood 7th-generation energy-efficient TPU engineered for large-scale training, reasoning, and inference. Features 9,216 liquid-cooled chips per pod, provides 42.5 ExaFlops.
SP023 Amazon Web Services AWS Trainium AWS Trainium3 chip provides 2x higher compute performance to 2.52 petaflops (PFLOPs) of FP8 compute... over 4x better energy efficiency compared to Trn2 UltraServers.
SP024 NextPlatform The Next Platform With Its IPO Done, Cerebras Can Get Back To Pushing The AI Envelope
SP025 HPCwire HPCwire — Home
SP026 SiliconAngle SiliconANGLE — The Voice of Enterprise and Emerging Tech
SP027 SemiAnalysis SemiAnalysis — Semiconductor Industry Analysis
SP028 TechCrunch AI News & Artificial Intelligence | TechCrunch
SI001 PRNewswire / Rebellions Rebellions Closes $400 Million Pre-IPO and Launches RebelRack™ and RebelPOD™ to Accelerate Global Expansion Rebellions has raised $400 million in a pre-IPO funding round led by Mirae Asset Financial Group and the Korea National Growth Fund. The round follows the company's $250 million Series C in September 2025, bringing total funding to $850 million and valuing the company at approximately $2.34 billion.
SI002 Rebellions Rebellions Closes $400 Million Pre-IPO and Launches RebelRack™ and RebelPOD™
SI003 CNBC AI chip startup Rebellions raises $400 million, plans IPO
SI004 Rebellions Rebellions Raises $250 Million to Advance the Next Generation AI Infrastructure, Backed by Arm and Samsung Rebellions has raised $250 million in Series C round at a valuation of $1.4 billion… ATOM and ATOM-Max are already in mass-production and deployed with customers across Japan, Saudi Arabia, and the United States, and power Korea's largest commercial AI service.
SI005 Rebellions Rebellions Scales Global Growth with Silicon Valley-backed Series C Sungkyue Shin, CFO of Rebellions: "The participation of leading Silicon Valley venture firms reflects global recognition of our innovation."
SI006 Rebellions Rebellions and Samsung Electronics Partner to Target the Emerging Generative AI Market with Next-Generation AI Chip Co-Development
SI007 Rebellions Rebellions Korean AI Chipmaker Closes $124M Series B Fundraise
SI008 Forbes Asia South Korea's AI Chip Champion Is Poised To Carve Out Global Niche Rebellions is targeting 100 billion won in revenue, roughly $68 million, in 2025. According to the latest available filings, the company posted 2.7 billion won in revenue in fiscal 2023, with a net loss that widened to 13.7 billion won from 8.1 billion won in 2022.
SI009 Korea JoongAng Daily Why Rebellions' strategy against Nvidia is not so outrageous Park is putting up a strong fight against a coalition of big players like Nvidia, SK hynix and TSMC, while Rebellions relies solely on Samsung Electronics… "It's a turnkey with Samsung since it handles memory, manufacturing, and packaging."
SI010 Rebellions South Korea's Foremost AI Chip Company Rebellions Secures $15M Series B Extension Round by Wa'ed Ventures
SI011 SemiAnalysis AI Semiconductor Startup Economics — Capital Intensity and Path to Profitability
SI012 NextPlatform Rebellions Raises $250 Million as It Navigates Toward IPO
SI013 Samsung Semiconductor Foundry Overview — Samsung Foundry Capabilities and Advanced Process Technologies Samsung Foundry offers competitive processes, design technologies, IP, and high-volume manufacturing capability… full suite of advanced process technologies includes 28FD-SOI, 14/10/8/5/4nm FinFet, and 3nm GAA with EUV technology from 5nm… integrated package solutions including 3D/2.5D.
SI014 Mirae Asset Financial Group Rebellions Pre-IPO Investment Announcement
SI015 VentureBeat Rebellions raises $250M Series C to challenge Nvidia in AI inference
SI016 EE News Europe Merger Terms Agreed to Form Korean AI Unicorn Sapeon and Rebellions have agreed to merge based on an equity value ratio of 1:2.4 between the two companies… Rebellions was founded in 2020. It has raised more than US$225 million in funding.
SI017 Rebellions Rebellions and SAPEON Korea Complete Merger — Launching Korea's First AI Chip Unicorn
SI018 Nikkei Asia Korean AI chip startup Rebellions raises $400M at $2.34 billion value
SI019 Arm Newsroom Arm Invests in Rebellions Series C
SI020 TechCrunch Rebellions raises $250M Series C with Arm and Samsung
SI021 SK Telecom Newsroom SK Telecom Signs MOU with Arm and Rebellions for Next-Generation AI Infrastructure Innovation
SI022 HPCwire Rebellions Collaborates with SK Telecom and Arm Targeting Sovereign AI
SI023 Business Korea Samsung SDS Dominates Korean Cloud Market, According to IDC Report
SI024 The Investor Rebellions, Sapeon Korea Merges to Challenge Nvidia
SI025 Rebellions Rebellions and SAPEON Korea Sign Definitive Merger Agreement
SI026 DART Korea (Financial Supervisory Service) Rebellions Inc. — Korean Corporate Interim Report Filing (DART Registry, Receipt No. 20231128000396)
SE001 Rebellions Inc. ATOM-Max — Rebellions Product Page 128 TFLOPS FP16, 512 TOPS INT8, 1024 TOPS INT4 at 350W TDP with 64 GB GDDR6 at 1024 GB/s, PCIe Gen5 x16 FHFL dual-slot.
SE002 Rebellions Inc. ATOM-Max Server — Rebellions Product Page 8x ATOM-Max cards in a 4U chassis; 1,024 TFLOPS FP16 aggregate; 512 GB GDDR6 at 8 TB/s; 3.4 kW typical, 4.3 kW max.
SE003 Rebellions Inc. ATOM-Max POD — Rebellions Product Page 8-server mini-POD; 64 NPUs; 400 GB/s RDMA fabric; turnkey rack-scale inference solution.
SE004 Rebellions Inc. Rebel Server — Rebellions Product Page 5U, 8x RebelCard; up to 2 PFLOPS FP8; 2x AMD EPYC 9355 (32C/64T); 1.5 TB DDR5; 4x 400G networking; 4–6 kW typical, 7 kW max.
SE005 Rebellions Inc. REBEL — Rebellions Product Overview Page REBEL: 4nm chiplet architecture with HBM3E memory; UCIe-Advanced die-to-die; designed for next-generation AI inference at scale.
SE006 Rebellions Inc. Rebellions Debuts REBEL-Quad at Hot Chips 2025 REBEL-Quad features four compute ASICs, 144 GB HBM3E at 4.8 TB/s, and UCIe-Advanced die-to-die at 16 Gbps, demonstrated live running LLaMA 3.3 70B and Qwen3 235B MoE at Hot Chips 2025.
SE007 Rebellions Inc. Rebellions Closes $400M Pre-IPO and Launches RebelRack and RebelPOD Rebellions launches RebelRack and RebelPOD — available now — to deliver rack-scale AI inference infrastructure for hyperscale operators.
SE008 Rebellions Inc. Rebellions Model Zoo Rebellions model zoo: 300+ supported models across LLMs, vision transformers, and classical inference workloads for PyTorch and TensorFlow.
SE009 Rebellions Inc. RBLN SDK — Rebellions Developer Page RBLN SDK comprises rebel-compiler, optimum-rbln, and vllm-rbln for end-to-end model deployment on Rebellions NPU hardware.
SE010 Rebellions Inc. Rebellions and Samsung Electronics Partner on Next-Generation AI Chip Co-Development Rebellions and Samsung Electronics will co-develop the next-generation REBEL AI chip on Samsung's 4nm foundry process with HBM3E memory integration.
SE011 PR Newswire Rebellions Closes $400 Million Pre-IPO and Launches RebelRack and RebelPOD Rebellions today announced the launch of RebelRack and RebelPOD — available now — to power hyperscale AI inferencing deployments globally.
SE012 IEEE Xplore / ISSCC 2026 A Quad-Chiplet AI SoC with Full-Chip Scalable Mesh Over 16Gb/s UCIe-Advanced Die-to-Die Interface for Large-Scale AI Inferencing A quad-chiplet AI SoC with UCIe-Advanced die-to-die at 16Gbps achieves 56.8 TPS on LLaMA v3.3 70B with 2k/2k input/output sequences at 4nm process node.
SE013 Rebellions Inc. (docs.rbln.ai) RBLN SDK v0.10.3 Release Notes RBLN SDK v0.10.3 (May 2026): rebel-compiler 0.10.3, optimum-rbln 0.10.3, vllm-rbln 0.10.3.post1; supports vLLM v0.18.0; paged attention and continuous batching implemented; speculative decoding and distributed KV cache in development.
SE014 vLLM Project vLLM AI Accelerator Installation — Supported Backends Supported AI accelerators for vLLM include: Google Cloud TPU, Intel Gaudi, AMD Instinct, Rebellions NPU, and others. Each is listed as a first-class hardware plugin backend.
SE015 Python Package Index (PyPI) vllm-rbln 0.10.3.post1 on PyPI vllm-rbln 0.10.3.post1 released on PyPI on May 18, 2026; provides vLLM-compatible inference backend for Rebellions NPU hardware.
SE016 GitHub (rebellions-sw) Rebellions GitHub Organization — rebellions-sw The rebellions-sw GitHub org contains primarily internal or archived repositories (e.g., rbcn-k8s-platform-archive-public). No public SDK source code or model repository is available as of May 2026.
SE017 MLCommons MLPerf Inference Datacenter Benchmark Results MLPerf Inference Datacenter benchmark results (May 2026 edition) include submissions from NVIDIA, AMD, Google, Intel, Qualcomm, and others. No Rebellions submission is present.
SE018 ServeTheHome Rebellions REBEL-Quad UCIe and 144GB HBM3E Accelerator at Hot Chips 2025 Rebellions REBEL-Quad is built on Samsung SF4X and CoWoS-S with four compute ASICs, four HBM3E sites, and four ISC on each package — a PCIe card using UCIe-Advanced for chiplet interconnect, running LLaMA 3.3 70B live at Hot Chips 2025.
SE019 Chips and Cheese Rebellions: From High Frequency Trading to AI Acceleration At Supercomputing 2024, Rebellions confirmed each ATOM chip has 32 TFLOPS FP16, 128 TOPS INT8, 256 TOPS INT4, and 16 GB GDDR6 at 256 GB/s; the RBLN-CA25 (ATOM-Max) packs four chips for 128 TFLOPS / 64 GB / 1024 GB/s at 350 W. The REBEL will use 4× 36 GB HBM3e stacks at 9.6 GT/s totalling 144 GB and 4.8 TB/s.
SE020 EE Times Rebellions Raises $250M Series C for AI Chip Development Rebellions closed a $250M Series C co-led by Samsung and Arm; Arm made a strategic investment, its first in an AI chip startup, alongside SK Hynix and other investors.
SE021 IEEE Spectrum AI Chip Startup Rebellions Builds Toward the Big Leagues Rebellions is building a two-generation AI chip portfolio — the ATOM inference SoC and the REBEL chiplet-based accelerator — with Samsung Foundry as its sole manufacturing partner and Arm as a strategic investor.
SE022 Arm Limited (Newsroom) Arm Invests in Rebellions Series C Arm is making a strategic investment in Rebellions as part of the company's Series C. Rebellions joins the Arm Total Design ecosystem, with plans to integrate Arm Neoverse compute subsystems into future products.
SE023 Samsung Newsroom Samsung and Rebellions Partner on Next-Generation AI Chip Samsung and Rebellions are partnering to develop the next-generation REBEL AI chip on Samsung's 4nm foundry process, integrating Samsung HBM3E memory.
SE024 Rebellions Inc. Rebellions, SK Telecom, and DOCOMO Innovations Partner on Next-Gen AI Infrastructure Rebellions, SK Telecom, and DOCOMO Innovations announce collaboration on sovereign AI infrastructure validation and co-development for AI datacenter deployments.
SE025 Rebellions Inc. Rebellions Solutions Overview Rebellions ATOM deployed at ECOPEACE (UAE) for water robot vision AI at 2× performance-per-watt vs GPU, TTA-certified; at Mongolia Customs Authority at 2.7× TPS/Watt vs GPU for customs AI inference.
SE026 Rebellions Inc. ATOM SoC — Rebellions Product Overview ATOM is Rebellions' first-generation AI inference SoC, designed for energy-efficient AI inference at datacenter scale, with GDDR6 memory and PCIe connectivity.
SE027 Rebellions Inc. Rebellions Raises $250M Series C Backed by Arm and Samsung Rebellions closes $250M Series C; Arm and Samsung co-lead. Arm joins as strategic investor and Rebellions joins the Arm Total Design ecosystem.
SE028 SemiAnalysis AI Semiconductor Startup Economics Fabless AI chip startups face a structural challenge: single-foundry dependence on TSMC or Samsung, high NRE costs for leading-edge nodes, and a winner-take-most inference market dominated by NVIDIA. Few startups have reached production at meaningful volume.
SU001 Rebellions Rebellions About We are deploying Rebellions' NPU in A., Korea's largest LLM service.
SU002 Rebellions Korean AI Chipmaker Rebellions Closes $124M Series B Fundraise KT and kt cloud join as lead investors and first customers.
SU003 Rebellions South Korea's Foremost AI Chip Company Rebellions Secures $15M Series B Extension Round by Wa'ed Ventures Wa'ed Ventures, the venture capital arm of Saudi Aramco, invested in Rebellions to support expansion into the Middle East.
SU004 Forbes South Korea's AI Chip Champion Is Poised To Carve Out Global Niche Increasing sales at home, much less international growth, is a tricky proposition.
SU005 Korea JoongAng Daily Why Rebellions' Strategy Against Nvidia Is Not So Outrageous Persuading data centers to source chips from companies other than Nvidia is almost as difficult as making the chips themselves.
SU006 Rebellions Rebellions, SK Telecom and DOCOMO Innovations Partner to Accelerate Next-Gen AI Infrastructure
SU007 SK Telecom SKT, Rebellions and Arm MOU for AI Inference Server
SU008 Rebellions Rebellions Scales Global Growth with Silicon Valley-Backed Series C ATOM is deployed with customers across Japan, Saudi Arabia, and the United States.
SU009 Rebellions Rebellions Closes $400 Million Pre-IPO and Launches RebelRack and RebelPOD to Accelerate Global Expansion Proven commercial deployments already live across enterprises and governments.
SU010 Rebellions Rebellions and Red Hat Introduce Red Hat OpenShift AI Powered by Rebellions NPUs to Fuel Choice and Flexibility in Enterprise AI
SU011 Red Hat Red Hat and Rebellions Introduce Red Hat OpenShift AI Powered by Rebellions NPUs to Fuel Choice and Flexibility in Enterprise AI
SU012 TechPowerUp Rebellions and Red Hat Introduce Red Hat OpenShift AI Powered by Rebellions NPUs
SU013 BusinessKorea Rebellions and Red Hat Introduce OpenShift AI Powered by Rebellions NPUs
SU014 Seoul Economic Daily Rebellions Unveils NPU-Based Red Hat OpenShift AI Platform
SU015 KoreaTechDesk Rebellions U.S. Expansion Korea AI Deep Tech Globalization
SU016 PR Newswire Rebellions Accelerates Global Expansion and Strengthens Customer-Centric Strategy with Significant Executive Appointments
SU017 Wamda Rebellions Closes $250 Million Series C to Power AI Chip Deployment in KSA
SU018 Business Wire Rebellions Collaborates with SK Telecom and Arm Targeting Sovereign AI and Telecom Infrastructure
SU019 HPCwire Rebellions Collaborates with SK Telecom and Arm Targeting Sovereign AI
SU020 Evertiq Rebellions Partners with SKT, Arm to Bolster AI Infrastructure
SU021 Converge Digest Rebellions, Arm and SK Telecom Partner on Sovereign AI for Telco Data Centers
SU022 The Investor Rebellions' SAPEON Merger Creates Korea's First AI Chip Unicorn
SU023 Rebellions Rebellions Solutions
SU024 Rebellions ATOM Max Server
SU025 Rebellions Rebel Server
SR001 Forbes South Korea's AI Chip Champion Is Poised To Carve Out Global Niche Increasing sales at home, much less international growth, is a tricky proposition.
SR002 Korea JoongAng Daily Why Rebellions' strategy against Nvidia is not so outrageous Persuading data centers to source chips from non-Nvidia companies is almost as difficult as making the chips.
SR003 The Investor Rebellions, Sapeon Korea merges to challenge Nvidia
SR004 Rebellions Rebellions Closes $400 Million Pre-IPO and Launches RebelRack™ and RebelPOD™ to Accelerate Global Expansion The round follows the company's $250 million Series C in September 2025, bringing total funding to $850 million and valuing the company at approximately $2.34 billion.
SR005 Rebellions Rebellions Raises $250 Million to Advance the Next Generation AI Infrastructure, Backed by Arm and Samsung
SR006 SK Telecom Newsroom SK Telecom Signs MOU with Arm and Rebellions for Next-Generation AI Infrastructure Innovation
SR007 Rebellions Rebellions and SAPEON Korea Sign Definitive Merger Agreement
SR008 Rebellions Rebellions and SAPEON Korea Complete Merger, Launching Korea's First AI Chip Unicorn Leading the new entity is CEO Sunghyun Park.
SR009 MLCommons MLPerf Inference Datacenter Benchmark Results The May 2026 MLPerf Inference Datacenter results include major incumbent vendors, but no Rebellions submission is present.
SR010 IEEE Xplore / ISSCC 2026 A Quad-Chiplet AI SoC with Full-Chip Scalable Mesh Over 16Gb/s UCIe-Advanced Die-to-Die Interface for Large-Scale AI Inferencing
SR011 ServeTheHome Rebellions REBEL-Quad UCIe and 144GB HBM3E Accelerator at Hot Chips 2025
SR012 Samsung Newsroom Samsung and Rebellions Partner on Next-Generation AI Chip Samsung and Rebellions are partnering to develop the next-generation REBEL AI chip on Samsung's 4nm foundry process, integrating Samsung HBM3E memory.
SR013 Rebellions About - Rebellions
SR014 CNBC Samsung-backed AI chip firm Rebellions raises $400 million ahead of IPO
SR015 Federal Register Framework for Artificial Intelligence Diffusion
SR016 Bureau of Industry and Security, U.S. Department of Commerce AI Policy Statement on Training AI Models
SR017 Center for Strategic and International Studies Understanding U.S. Allies’ Current Legal Authority to Implement AI and Semiconductor Export Controls
SR018 The National Law Review BIS Issues Four Key Updates on Advanced Computing and AI Export Controls
SR019 Wiley BIS Announces New Regulatory Framework for AI and Controls on Advanced Computing Technology and AI Models
SR020 WilmerHale BIS Issues Sweeping Additional Restrictions on Semiconductors and Advanced Computing, Entity List Designations
SR021 Morrison Foerster Managing Export Control Risks in the AI Chip Ecosystem
SR022 Korea Exchange Global KRX
SR023 Seoul Economic Daily Rebellions Gears Up for KOSPI IPO as AI Chip Listing Race Heats Up
SR024 Wowtale Rebellions Raises $424M in Pre-IPO Round as First Direct Investment Under National Growth Fund
SR025 Graphcore Graphcore joins SoftBank Group to build next generation of AI compute
SR026 EE Times AI chip startup Graphcore acquired by SoftBank
SR027 DataCenterDynamics SambaNova exploring sale after struggling to secure further funding SambaNova was exploring a sale after struggling to raise further funding, with BlackRock marking shares to about $2.4B from a $5B peak.
SR028 Forbes South Korean AI Chip Startup FuriosaAI Raises $125 Million
SR029 BusinessWire FuriosaAI Closes $125M Investment Round to Scale Production of Next-Gen AI Inference Chip
SR030 Tech Funding News FuriosaAI: Nvidia challenger from Seoul seeks up to $500M for next-gen AI chips
SR031 SemiAnalysis AI Semiconductor Startup Economics Few startups have reached production at meaningful volume.
SV001 Rebellions Rebellions Closes $400 Million Pre-IPO and Launches RebelRack™ and RebelPOD™ to Accelerate Global Expansion
SV002 PRNewswire / Rebellions Rebellions Closes $400 Million Pre-IPO and Launches RebelRack™ and RebelPOD™ to Accelerate Global Expansion
SV003 TechCrunch AI chip startup Rebellions raises $400 million at $2.3 billion valuation in pre-IPO round
SV004 Seoul Economic Daily Rebellions gears up for KOSPI IPO as AI chip listing race
SV005 WOWTale Rebellions raises $400 million in pre-IPO round at $2.3 billion valuation
SV006 Korea Exchange Korea Exchange Global Website
SV007 Forbes Asia South Korea's AI Chip Champion Is Poised To Carve Out Global Niche Rebellions is targeting 100 billion won in revenue, roughly $68 million, in 2025, while the latest available filings showed 2.7 billion won of 2023 revenue and a 13.7 billion won net loss.
SV008 Korea JoongAng Daily Why Rebellions' strategy against Nvidia is not so outrageous
SV009 Rebellions Rebellions Raises $250 Million to Advance the Next Generation AI Infrastructure, Backed by Arm and Samsung
SV010 Rebellions Rebellions Korean AI Chipmaker Closes $124M Series B Fundraise
SV011 TechCrunch AI chip startup Cerebras files for IPO
SV012 U.S. Securities and Exchange Commission Cerebras Systems S-1/A filing index (April 2026)
SV013 U.S. Securities and Exchange Commission Cerebras Systems S-1/A filing index (May 2026)
SV014 CompaniesMarketCap Cerebras Systems market capitalization
SV015 Groq Groq raises $750 million as inference demand surges
SV016 Data Center Dynamics Groq raises $750M for $6.9B valuation
SV017 Public.com Groq company profile
SV018 Tracxn Groq funding and investors
SV019 Data Center Dynamics SambaNova exploring sale after struggling to secure further funding report
SV020 TechStartups AI chip startup SambaNova, once valued at $4 billion, explores sale after failing to raise new funding
SV021 Graphcore Graphcore joins SoftBank Group to build next generation of AI compute
SV022 EE Times AI chip startup Graphcore acquired by SoftBank
SV023 Forbes South Korean AI chip startup FuriosaAI raises $125 million
SV024 BizChosun FuriosaAI seeks larger fundraising ahead of AI-chip scale-up
SV025 BusinessWire / FuriosaAI FuriosaAI Closes $125M Investment Round to Scale Production of Next-Gen AI Inference Chip
SV026 TechFundingNews FuriosaAI, Nvidia challenger from Seoul, seeks up to $500M for next-gen AI chips
SV027 Ars Technica Tenstorrent takes on Nvidia with open-source AI chip and $693 million in backing
SV028 Crunchbase News Tenstorrent AI chips unicorn and Jim Keller financing profile
SV029 Rebellions South Korea's Foremost AI Chip Company Rebellions Secures $15M Series B Extension Round by Wa'ed Ventures
SV030 CNBC AI chip startup Rebellions raises $400 million, plans IPO