Preferred Networks, Inc.
日本旗舰级 AI 独角兽,垂直整合逻辑可信,但公开披露的收入、盈利能力和 2017 年后估值都很薄
Preferred Networks 仍是日本最可信的纵向一体化 AI 平台,但收入基础披露很薄,第三方估值又相互冲突($1.0B vs $2B+); 只靠公开证据,很难承销其独角兽标价。
封面要素
公司概况
Preferred Networks (PFN) 是一家总部位于东京的深度学习公司,2014 年 3 月由 Toru Nishikawa 和 Daisuke Okanohara 从 Preferred Infrastructure 拆分创立。PFN 将自己定位为垂直整合的 AI 平台:与 Kobe University 设计自有 MN-Core 深度学习加速器,运营大规模 GPU 与 MN-Core 计算基础设施,在日本 NEDO GENIAC 计划下训练 PLaMo 基础模型系列,并把这些能力落到工业机器人(Fanuc)、汽车感知(Toyota)、材料发现(与 ENEOS 合作的 Matlantis)、药物发现(Chugai Pharmaceutical)、农业(与 Oisix 合作的 CraftyFarm)和消费机器人(Preferred Robotics 旗下 Kachaka)。它被广泛视为日本估值最高的私营 AI 初创公司。
- 成立时间
- 2014-03-26
- 创始人
- Toru Nishikawa, Daisuke Okanohara
- 创立地点
- Tokyo, Japan
- 总部
- Otemachi, Tokyo, Japan
- 产品
- PFN 销售:(a)MN-Core 深度学习加速器芯片和系统(包括 L1000 LLM 推理部件);(b)计算基础设施访问(PFN cloud、KDDI 合作);(c)PLaMo 基础模型家族;(d)面向材料和化学的 Matlantis 云端原子级仿真(与 ENEOS 合资);(e)面向 Fanuc、Toyota 及其他大型企业的工业 AI / 机器人解决方案;以及(f)通过 Preferred Robotics 的 Kachaka 切入消费机器人。收入据称由解决方案 / 授权、芯片和系统销售、云与 SaaS 订阅,以及研究补助或共同开发收入混合构成。
- 客户
- 大型日本工业集团(汽车、工厂自动化、能源、材料、制药)、日本政府和学术研究项目,以及不断增长的日语 LLM 与原子级仿真企业客户;消费者触达仅限 Kachaka 家用机器人试点。
- 商业模式
- 混合 IP / 硬件 / 软件模式——与战略锚点(Toyota、Fanuc、ENEOS、Chugai)开展研发服务和共同开发;销售 MN-Core 芯片及 MN-3 / MN-Core 2 系统;获取 Matlantis 和 PFN-cloud 订阅收入;PLaMo 授权和政府资助计算(NEDO GENIAC);再加上主要投资方同时也是主要客户的股权式合作。
- 阶段
- Late-stage private (Japanese kabushiki kaisha; widely classified as a unicorn)
- 融资情况
- 2024 年 12 月首关 ¥19B(SBI 领投股权,MUFG、Resona、Shoko Chukin 和 SMBC 提供债务),到 2025 年 4 月扩展为 ¥24B 系列;叠加历史上 Toyota(2015 年 ¥1B、2017 年 ¥10.5B)、Fanuc、NTT、Mitsui、Mizuho 和 Hitachi 轮次,累计已披露融资远高于 ¥40B。The Bridge 和 Latka 将当前估值锚定在约 ¥300B / $2B,而 PremierAlts 的二级市场标记显示截至 2025 年 6 月为明显更低的 $1.0B。
执行摘要
主要优势
- 从自研 MN-Core 芯片、大规模算力、PLaMo 基础模型到应用解决方案,PFN 确实打通了纵向一体化;这在全球私营 AI 公司中罕见,在日本更独特。
- 与 Toyota(汽车感知)和 Fanuc(工业机器人)建立了多年深度战略锚定关系;两家公司既是客户,也是长期投资人。
- 应用版图分散:Matlantis(ENEOS)、药物发现(Chugai)、Kachaka 消费机器人(Preferred Robotics)、CraftyFarm 农业(Oisix),可对冲单一垂直行业风险。
- 技术可信度强(MN-3
- 到 2024 年末 / 2025 年,仍能从日本大型机构、银行和政府相关投资人(SBI、MUFG、Resona、SMBC、DBJ、Mitsubishi Corp、Sekisui House、Wacom)持续拿到资本。
主要风险
- 经审计收入、毛利率、分部经济性,以及 2017 年后的一级市场估值均未公开披露;当前 $2B 叙事依赖 Latka / The Bridge / CB Insights 追踪数据, 与 PremierAlts 的 $1.0B 二级市场标记冲突。
- 商业上重度依赖少数战略股东(Toyota、Fanuc、NTT、Mitsui),而这些股东同时也是客户,带来治理和收入集中风险。
- MN-Core 面对近乎垄断的 NVIDIA 体系、超大规模云厂商自研芯片(TPU、Trainium、MAIA、MTIA)抬头,以及资金充足的商用芯片对手; PFN 在 2024 年出售 MN-Core 2 芯片业务,也让外售芯片策略受到质疑。
- 日本特有风险包括:日元走弱压缩美元估值、美国先进 AI 芯片出口管制、相对全球超大规模云厂商更小的日本本土 LLM TAM,以及长期高级 AI 工程师短缺。
- AI 估值环境降温时,Tokyo Stock Exchange Growth / Prime 市场的近期 IPO 路径有限;后续融资可能向 PremierAlts 的 $1.0B 标记下修。
未决问题
- 经审计合并收入、毛利率和分部经济性(芯片 / 云 / Matlantis / PLaMo / 机器人 / 服务拆分)。
- 2017 年后的一级市场估值需要对账,以解释 PremierAlts 与 Latka / The Bridge / CB Insights 之间 $1.0B vs $2B+ 的差距。
- 2024 年 12 月 / 2025 年 4 月融资系列附近的完整股权结构表、清算优先权和任何二级市场活动。
- MN-Core L1000 未来商业管线、部署台数,以及 PFN 自有云之外的客户突破。
- Kachaka 家用机器人装机量、CraftyFarm 田间试验,以及 Matlantis 企业净收入留存,需要独立的部署后证据。
目录
01公司概览
1.1 身份、范围与运营模式
Preferred Networks, Inc. (PFN) 是一家总部位于东京的私营 AI 公司,成立于 2014 年 3 月 26 日,总部位于千代田区大手町大厦。公司自己的使命表述——“Make the real world computable and create the future together”——异常宽泛,但公开材料把运营范围讲得很具体:PFN 从 AI 芯片和自有超算一路做到基础模型、工业解决方案和应用。这使公司不像纯软件初创公司,更像一个整合 AI 基础设施与应用的实验室,并带有战略商业化路径。公开证据覆盖制造、出行、能源、材料、生命科学、娱乐、金融、公共服务和教育。经审阅来源没有公开收入、ARR、利润率和客户数量数据,因此后续章节应把 PFN 视为战略验证强、但财务透明度不足的后期私营公司。[CO001, CO002, CO007, CO008, CO022, CO037]
| 指标 | 数值 / 状态 | 日期 | 置信度 | 证据缺口 |
|---|---|---|---|---|
| 法律主体 | Preferred Networks, Inc. | 2026-06-14 | 高 | 名称 / 日期 / 总部无缺口;已审阅官方公司页面 |
| 成立 | 2014 年 3 月 26 日 | 2014-03-26 | 高 | 成立日期无缺口 |
| 总部 | Otemachi Building, 1-6-1 Otemachi, Chiyoda-ku, Tokyo,总部地址 | 2026-06-14 | 高 | 总部地点无缺口 |
| 阶段 | 后期私有 AI 独角兽 / 战略型创业公司 | 2026-06-14 | 中 | 未审阅经审计股权结构表或 IPO 申报 |
| 最新披露轮次 | 2024 年 12 月 / 2025 年 4 月轮次迄今合计 24B yen | 2025-04-30 | 高 | 轮次总额已披露;PFN 公告未披露投后估值 |
| 收入 / ARR | 未公开披露 | 2026-06-14 | 中 | 需要数据室、客户合同或投资人材料 |
| 客户数量 | 未公开披露 | 2026-06-14 | 中 | 有具名伙伴,但活跃客户数量不可得 |
| 员工人数 | 所审阅官方页面未披露 | 2026-06-14 | 中 | 第三方画像不一;需用薪资名册或 LinkedIn 导出核验 |
| 核心技术栈 | AI 芯片、算力基础设施、基础模型、AI 解决方案 | 2026-06-14 | 高 | 按收入拆分的商业组合未披露 |
所审阅公开来源中,私营公司的财务和运营指标未经审计,也未披露;本表将已验证身份事实与尽调缺口分开。
[CO001, CO007, CO016, CO017, CO022, CO023]PFN 公开模型把战略投资者和合作伙伴接到一体化计算栈,再延伸到工业 AI 部署和基础模型产品。
流程图是概念图,基于公开定位,不是收入分配模型。
[CO007, CO015, CO016, CO031, CO030, CO026]1.2 创始人、领导层与治理
PFN 仍由创始人主导。经审阅的公司页面列出 Toru Nishikawa 为联合创始人兼董事长,Daisuke Okanohara 为联合创始人兼首席执行官;联合创始人寄语则强调,公司由“love computer science and technology”的人打造,并希望掌握计算的每一层。这种创始人延续性有利于技术一致性和伙伴信任,但也形成关键人物集中:两位联合创始人仍比外部可见的广泛职业化管理团队更深地锚定战略、公司叙事和技术方向。PFN 披露了治理层:董事包括 Hiroshi Maruyama 和外部审计及监督委员会董事,高管包括 COO Naoto Ono、CFO Yotaro Katayama 和工程副总裁 Masaaki Fukuda。经审阅的公开来源没有披露薪酬、持股比例、接班计划或完整董事会投票安排。[CO003, CO004, CO005, CO006, CO038, CO040]
| 人物 | 职务 | 背景 / 证据 | 职能覆盖 | 关键人物依赖 |
|---|---|---|---|---|
| Toru Nishikawa | 联合创始人、董事长 | PFN 公司页面和联合创始人寄语具名 | 创始人战略、伙伴叙事、计算技术栈愿景 | 高 — 联合创始人仍是公司身份核心 |
| Daisuke Okanohara | 联合创始人、首席执行官 | PFN 公司页面和联合创始人寄语具名 | CEO;AI / 基础模型和技术领导力信号 | 高 — 联合创始人 CEO 集中执行权 |
| Naoto Ono | 首席运营官;企业规划部门负责人 | PFN 公司页面列名高管 | 企业规划和运营 | 中 — 该职务有助于执行专业化 |
| Yotaro Katayama | 首席财务官 | PFN 公司页面列名高管 | 财务、资本规划和投资人接口 | 中 — 融资复杂度使 CFO 职务重要 |
| Masaaki Fukuda | 工程副总裁;技术规划部门负责人 | PFN 公司页面列名高管 | 工程和技术规划 | 中到高 — 核心技术栈横跨芯片、算力和模型 |
| Hiroshi Maruyama | 董事;审计与监督委员会主席 | PFN 公司页面列名董事 | 审计委员会治理 | 直接运营依赖低;监督角色重要 |
列表基于 PFN 公司页面可见的具名领导者和董事;完整薪酬、继任和所有权数据为非公开信息。
[CO003, CO004, CO005, CO006, CO038, CO040]1.3 融资历史、估值信号与利益相关方
PFN 的资本历史主要由日本产业和金融战略支持者主导。Toyota 在 2015 年投资 10 亿日元,2017 年又追加约 105 亿日元,使 Toyota 成为当时最大的外部股东。FANUC 在 2015 年投资 9 亿日元,PFN 的里程碑年表还列出 2017 年与 Hakuhodo DY、Mitsui、Mizuho 和 Hitachi 的资本合作。近期融资把故事从出行 / 机器人研发转向日本制造 AI 半导体和计算基础设施:SBI 在 2024 年同意最高投资 100 亿日元;PFN 随后在 2024 年 12 月宣布 190 亿日元首关,并在 2025 年 4 月追加 50 亿日元,使该轮迄今达到 240 亿日元。独立媒体支持独角兽表述,但准确投后估值和股权结构仍未公开,因此估值应被视为媒体 / 市场数据支持,而非经审计事实。[CO011, CO012, CO013, CO015, CO016, CO017]
| 利益相关方 | 角色 / 证据 | 经济或战略重要性 | 尽调要求 |
|---|---|---|---|
| Toyota Motor | 2015 年投资 1.0B yen;2017 年追加约 10.5B yen;2026 年实体 AI 研究 | 2017 年配售后最大外部股东;验证移动出行和机器人方向 | 确认当前持股、商业排他性和知识产权权利 |
| FANUC | 2015 年 900M yen 资本联盟;后续按里程碑追加投资 | 工业机器人渠道和工厂自动化验证 | 确认当前持股和联合产品收入 |
| SBI Group | 2024 年最高 10B yen 联盟;领投 19B yen 首关 | 半导体融资和日本 AI 生态赞助方 | 核验轮次经济条款、治理权和债务条款 |
| Mitsubishi Corporation / IIJ | Preferred Computing Infrastructure 合资公司 | 用 PFN 技术栈商业化 AI 云算力基础设施 | 确认股权拆分、客户管线和资本开支义务 |
| Development Bank of Japan 政策金融机构 | 2024 年 12 月首关投资者 | 面向本土 AI 基础设施的政策一致性融资支持 | 尽调任何契约或战略限制 |
| ENEOS Innovation Partners / ENEOS | 股东;炼厂自主运营伙伴 | 能源行业部署证明和工业 AI 标杆 | 衡量收入、部署范围和安全 / 监管批准 |
| 媒体 / 内容投资者 | Kodansha、TBS、Toei Animation、Wacom 等媒体 / 创意投资方 | 表明 PLaMo / 生成式 AI 在内容工作流中的用例 | 确认商业合同与战略期权投资的区别 |
| Mizuho / MUFG / SMBC / Sumitomo Mitsui Trust 银行团 | 近期轮次中的债务或股权融资方 | 增加非稀释性资本和银行背书 | 审阅债务到期、抵押品、契约和对跑道的影响 |
持股比例和投票权未公开;图谱强调已披露的战略相关性和尽调问题,而不是股权结构权重。
[CO011, CO012, CO015, CO016, CO017, CO018]压缩展示 PFN 证据最充分的成熟度信号,以及私有公司主要披露缺口。
KPI 数值来自公开来源事实或披露状态标签,不是经审计经营指标。
[CO017, CO001, CO009, CO036, CO022, CO023]1.4 平台、子公司与商业验证点
PFN 的公司故事如今是一组相互连接的技术押注。MN-Core 和 MN-3 提供硬件与计算验证:PFN 称 MN-3 三次登顶 Green500,TOP500 和 Supermicro 则独立印证了 MN-3 系统和能效成绩。Preferred Elements 把 PFN 延伸到多模态基础模型,Matlantis 承接计算化学和材料仿真;与 Mitsubishi Corporation、IIJ、Rapidus、Sakura Internet 和 GMO 的合作公告显示,计算栈正在走向合资商业化,而不只是内部研究资产。应用侧,ENEOS 公开确认与 PFN 实现炼油厂自主运行,2026 年 6 月发布又新增与 Mitsubishi Heavy Industries 的关键任务 AI,以及与 Toyota 的 physical-AI 研究。这些证据支撑真实的工业伙伴基础,但还不能证明透明的收入基础。[CO009, CO010, CO026, CO027, CO028, CO029]
1.5 里程碑、不利事件与尽调含义
年表显示,PFN 多次从研究框架推进到工业部署,同时砍掉不再具备战略性的业务线。早期 Chainer 框架曾是 PFN 的重要资产,但 2019 年 PFN 将 Chainer 转入维护并把研究迁移到 PyTorch;对 Chainer 的独立护城河而言,这是一次不利的平台产品事件,即便公司将其解释为生态选择。官方里程碑页面还记录了 Crypko 和 Petalica Paint 在 2025 年终止消费者服务,进一步说明 PFN 愿意关闭非核心消费产品。因此,尽调应聚焦哪些业务具备持久商业拉力:近期资本用途明确绑定 MN-Core、PLaMo、AI 解决方案、人才和基础设施,而准确估值、客户数量、收入规模、毛利率和客户集中度并未公开。后续章节应检验战略投资者网络到底能转化为可重复收入,还是主要补贴国家冠军式研发。[CO023, CO025, CO039, CO041, CO022, CO037]
| 日期 | 事件 | 类型 | 金额 / 估值 / 状态 | 参与方 | 含义 |
|---|---|---|---|---|---|
| 2014-03 | PFN 成立 | 成立 | 2014 年 3 月 26 日成立 | Toru Nishikawa; Daisuke Okanohara | 从 PFI 根基出发,推出深度学习 / IoT 商业化载体 |
| 2015-08 | FANUC 资本联盟 | 融资 | 900M yen;6.0% 已发行股份 | FANUC; PFN | 工业机器人学习和工厂自动化验证 |
| 2015-12 | Toyota 资本合作 | 融资 | 1.0B yen | Toyota; PFN | 强化移动出行 AI 关系 |
| 2017-08 | Toyota 追加投资 | 融资 | 约 10.5B yen;Toyota 为最大外部股东 | Toyota; PFN | 自动驾驶 AI 的重大投资者验证 |
| 2017-12 | 战略资本合作 | 融资 | 未披露 | Hakuhodo DY、Mitsui、Mizuho、Hitachi、FANUC 等合作方 | 扩大日本产业和金融赞助方基础 |
| 2019-12 | Chainer 转入维护;迁移到 PyTorch | 不利 | 框架转换 | PFN; Facebook/PyTorch 生态 | 对 Chainer 护城河不利;对生态对齐有利 |
| 2020-05 | MN-3 开始运行 | 产品 | 后续在 2020/2021 年赢得 Green500 | PFN; Kobe University/Supermicro 生态 | 证明自有算力能效策略 |
| 2023-11 | Preferred Elements 成立 | 产品 | 基础模型子公司 | PFN; Preferred Elements | 分拆 PLaMo 商业化路径 |
| 2024-05 | ENEOS 原油装置自主运营 | 规模化 | 公告称全球首例 | ENEOS; PFN | 实验室研发之外的工业 AI 证明点 |
| 2024-12 | 19B yen 首关 | 融资 | 19B yen 股权 / 债务 | SBI; DBJ; Mitsubishi; Wacom; 贷款方 | 为 MN-Core、PLaMo 和算力基础设施提供资金 |
| 2025-04 | 扩展轮 | 融资 | 追加 5B yen;本轮迄今 24B yen | Kodansha、MUFG Trust、SMTB、TBS、Toei、Mizuho 等投资方 | 增加媒体 / 内容和银行利益相关方 |
| 2026-03 | GMO Preferred Security 合资公司 | 合作 | 新合资公司 | PFN、GMO Internet、GMO Cybersecurity by Ierae 等参与方 | 聚焦安全的日本制造 AI 环境 |
| 2026-06 | Toyota 实体 AI 研究 | 合作 | MN-Core L 系列测试 | Toyota Frontier Research Center、PFN 合作方 | 将 Toyota 战略关系延伸至机器人推理 |
| 2026-06 | Mitsubishi Heavy Industries 联盟 | 合作 | 任务关键型日本制造 AI | MHI; PFN | 将 AI 推向韧性社会基础设施应用 |
时间线优先纳入有公开来源支持的事件;个人持股比例、未披露早期融资和精确投后估值仍是缺口。
[CO001, CO011, CO012, CO013, CO023, CO025]从创立到 2026 年 6 月,PFN 的带日期里程碑涵盖融资、平台迁移、工业部署和当前战略联盟。
时间线只纳入部分公开事件;未纳入私有融资和未披露商业里程碑。
[CO001, CO011, CO012, CO023, CO009, CO030]1.6 图表
02市场分析
2.1 市场边界与现状替代方案
Preferred Networks (PFN) 不应按一家泛 AI 公司来测算规模。它的市场边界是一组 physical-AI 与 scientific-AI 视角,并由伙伴商业化锚定:与 Fanuc 和 MHI 相关的工业 AI 与智能制造、工业机器人智能、Toyota 相关 physical AI 与自动驾驶软件、MN-Core AI 基础设施和加速器、作为较窄期权价值的农业机器人,以及通过 Chugai 式实验自动化切入的 AI 驱动药物发现。纳入的支出是 PFN 可通过共同创造或授权合理影响的软件、模型、加速器和集成价值。排除的支出包括消费者 AI 应用、没有 PFN 计算角色的一般云服务、整车硬件、与自主化无关的广义农机,以及非计算或非自动化的制药湿实验室支出。结果不是一个标题式 TAM,而是买方高度不同的多市场边界。[CM001, CM002, CM003, CM004, CM005, CM022]
| 细分 / 类别 | 纳入支出 | 排除支出 | 主要买方 / 付款方 | 与 PFN 的相关性 |
|---|---|---|---|---|
| 工业 AI / 智能制造 | AI 模型、部署软件、机器人智能、数字孪生,以及面向工厂 / 基础设施的集成 | 通用企业 AI、ERP、非工业分析 | 制造商、机器人 OEM、类似 MHI 的基础设施总包方 | 通过 Fanuc、MHI 和制造业 AI 需求走向商业化的核心路径 |
| 工业机器人 | 机器人智能、自优化、感知,以及接入机器人 OEM 生态的自动化软件 | 作为商品硬件的机械臂,PFN 无经济权益 | 机器人 OEM、工厂自动化团队 | Fanuc 关系和日本机器人密度使其成为直接观察镜头 |
| 自动驾驶 / 实体 AI | 感知、推理加速、仿真和实体 AI 研究软件 | 车辆硬件、网约车车队价值、与 PFN 无关的消费级 ADAS 订阅 | Toyota 研发、移动出行工程团队 | Toyota FRC 2026 研究验证了进入渠道,但未验证收入规模 |
| AI 芯片 / 加速器 | MN-Core 处理器、AI 基础设施、冷却和内部 / 伙伴算力平台 | 商品云算力转售、PFN 没有份额的 GPU | AI 基础设施运营商、模型团队、主权 AI 项目 | TAM 很大,但生态和资本强度风险最高 |
| 农业机器人 | 自主农场机器人、机器视觉、喷洒、收割和农场自动化软件 | 传统拖拉机、种子 / 化学投入品、不含机器人的农场管理 | 农场经营者、农业设备 OEM | 小型期权价值视角;PFN 相关 CraftyFarm 证据公开更新不足 |
| AI 驱动药物发现 | 计算化学、实验自动化、分子仿真和 AI 发现平台 | 不含 AI 自动化的湿实验服务、临床试验支出 | 药企研发、发现平台团队 | Chugai 关系和 MALEXA 背景验证邻近性,但临床转化风险仍在 |
| 日本 AI 软件 / 服务 / 基础设施 | 本土 AI 基础设施、AIaaS、工业 AI 服务和主权算力 | 全球消费 AI 支出和非日本服务 | 日本企业、政府支持的基础设施项目 | 作为日本本土冠军企业,PFN 的重要 SAM 过滤器 |
边界将 PFN 可影响的软件、芯片和共同打造的垂直解决方案,与 PFN 不直接变现的广泛终端市场硬件或服务分开。
[CM001, CM002, CM003, CM004, CM005, CM021]2.2 市场规模——TAM、SAM、SOM 与分部视角
规模测算答案是区间,不是点估计。最宽的全球 AI 支出视角只适合作背景:Gartner 的 2026 年 AI 支出预测超过 $2.5 trillion,Fortune Business Insights 则给出 2026 年全球 AI 市场 $375.93 billion。PFN 更近的视角更小但更相关:2026 年智能制造为 $387.14 billion,工业机器人为 $15.5 billion,AI 加速器芯片为 $154.6 billion,自动驾驶软件按 Precedence 约为 2026 年 $2.97 billion,按 MarketsandMarkets 则从 2024 年 $1.8 billion 增至 2035 年 $7.0 billion,农业机器人为 $18.0 billion,AI 药物发现约为 $2.9 billion。日本本土需求也有意义:IDC 预计日本 AI 基础设施在 2026 年超过 $5.5 billion,而日本 AIaaS 与全 AI 预测因定义不同而差异很大。因此,PFN 的可服务市场应被限制在日本工业 / physical-AI 部署、AI 基础设施和选定伙伴垂直领域,而不是整个 AI 经济。公开来源的市场规模还带来一个重要解释规则:这些市场相邻,不能相加。Toyota physical-AI 验证点不能像完整自动驾驶汽车供应商那样估值;MN-Core 部署不能像占主导地位的商用 GPU 厂商那样估值;Chugai 或农业验证点也不能自动迁移到工业机器人。因此,本章把每个视角都视为独立尽调路径,各自拥有买方、预算所有者、转化证据和失败模式。[CM006, CM007, CM008, CM009, CM010, CM011]
| 发布方 | 年份 | 地域 | 数值 | CAGR | 方法论 | 置信度 | 局限 |
|---|---|---|---|---|---|---|---|
| Gartner | 2026 | 全球 AI 支出 | $2.52T–$2.59T | 44%–47% 同比 | 按类别自上而下测算全球 AI 市场支出 | 高 | 仅为背景 TAM;远大于 PFN 可变现范围 |
| Fortune Business Insights | 2026 | 全球 AI 市场 | $375.93B | 至 2034 年 26.60% | 分析师按组件和地域建立的市场模型 | 高 | 广义 AI 市场包含 PFN 之外的消费和企业领域 |
| Mordor Intelligence | 2026 | 全球智能制造 | $387.14B | 至 2031 年 13.53% | 工厂自动化和智能制造市场模型 | 高 | 包含 PFN 之外的硬件、控制系统和软件 |
| MarketsandMarkets | 2026 | 全球工业机器人 | $15.5B | 至 2032 年 5.0% | 按机器人类型和产品供给拆分 | 高 | 机器人专属市场;排除更广泛 AI 平台价值 |
| IFR | 2024 年实际值 | 全球工业机器人 | 542,000 台安装量;4.664M 台运营存量 | 2025 年安装量预测增长 6% | 行业联合会出货 / 安装统计 | 高 | 统计单位是台数而非收入;且为 2024 年实际值,不是 2026 年市场规模 |
| GMI | 2026 | 全球 AI 加速芯片 | $154.6B | 到 2035 年为 23.6% | AI 加速芯片市场模型 | 高 | PFN MN-Core 未披露外部份额 |
| Precedence Research | 2026 | 自动驾驶软件 | $2.97B | 到 2035 年为 13.33% | ADAS / 自动驾驶软件细分 | 中 | 口径只覆盖软件,小于整车级 AV 市场 |
| Mordor Intelligence | 2026 | 农业机器人 | $18.0B | 到 2031 年为 18.07% | 农业机器人设备和软件模型 | 高 | PFN 在农业商业化上的证据有限 |
| Grand View / R&M | 2026 | AI 药物发现 | $2.9B–$2.93B | 24.8%–26.2% | 两份独立 AI 药物发现市场报告 | 高 | 药物发现收入取决于药企验证和管线成败 |
| IDC | 2026 | 日本 AI 基础设施 | >$5.5B | 18% YoY | IDC AI 基础设施跟踪 / 支出指南 | 高 | 只覆盖基础设施;不包括日本 AI 软件 / 服务 |
| IMARC / VMR | 2025–2034 | 日本 AI 服务 / 全 AI 市场 | 2025 年 AIaaS 为 $1.25B;2025 年全 AI 为 $19.83B | AIaaS 为 31.75%;全 AI 为 34.72% | 口径不同的日本国家市场报告 | 中 | 定义分叉;应作为区间而非点估计使用 |
各行刻意混用不同测算口径;PFN 的 SAM 必须再按市场参与度、地理范围和伙伴商业化状态收窄。
[CM008, CM009, CM010, CM011, CM012, CM013]从广义 AI 背景逐层收窄到 PFN 受约束的可服务市场,以及公开 SOM 缺口。
金字塔组合了不可相加的市场视角;定义存在重叠,数值不应相加。
[CM010, CM012, CM013, CM016, CM023, CM036]低 / 基准 / 高估算用同一单位显示各分部规模差异:十亿美元。
高值在有披露时采用预测端点;自动驾驶高值是 2035 年软件预测,不是 2026 年。
[CM008, CM009, CM016, CM018, CM019, CM021]2.3 买方分层与采用路径
买方分层是理解采用的关键视角。在工业机器人中,Fanuc 式机器人 OEM 和工厂自动化团队是买方,产线工程师和机器人程序员是用户;预算所有者通常是制造工程或工厂自动化资本开支。在基础设施和智能制造中,MHI 及类似重工业主承包商购买关键任务 AI 能力,并嵌入基础设施项目。Toyota 的 Frontier Research Center 指向汽车研发和 physical-AI 工程买方,而不是零售购车者。MN-Core 和 AImod 买方包括内部 AI 模型团队、数据中心运营商、主权 AI 计划和 IIJ/JAIST 等伙伴。制药买方是 Chugai 式发现平台和研发自动化团队;农业买方会是设备 OEM 或大型农场运营商,但公开的 PFN 专属商业化证据更薄。跨分部看,PFN 的采用路径通常从联合研究开始,随后进入伙伴内部使用,再走向联合商业化或授权。[CM027, CM028, CM029, CM030, CM031, CM032]
| 细分市场 | 买方 | 用户 | 付款方 | 预算负责人 | 采用触发因素 |
|---|---|---|---|---|---|
| 工业机器人 / Fanuc | 机器人 OEM 和工厂自动化客户 | 机器人程序员、产线工程师 | OEM 研发预算或工厂自动化资本开支 | 制造工程 / 自动化 VP | 需要自优化机器人和劳动生产率提升 |
| 智能制造 / MHI | 重工业总包或基础设施运营方 | 运营工程师、基础设施维护人员 | 关键基础设施项目预算 | 业务单元 GM、CTO、基础设施项目负责人 | 日本造 AI 自主能力,用于韧性基础设施 |
| Toyota 物理 AI | Toyota FRC 和移动出行研发 | 自动驾驶和机器人研究员 | 公司研发和先进工程 | 研发高管 / 移动出行平台负责人 | 更快推理和物理 AI 模型部署 |
| MN-Core / AImod | AI 基础设施运营方、模型团队、主权算力项目 | ML 工程师和数据中心运营方 | AI 基础设施资本开支 / 研发补助 | CTO、数据中心业主、国家研发资助方 | GPU 供应压力、能效、本土算力 |
| 农业机器人 | 农场运营方或农业设备 OEM | 农场工人、农艺师、田间技术员 | 设备资本开支或服务合同 | 农场主 / OEM 产品负责人 | 劳动力短缺和精准农业 ROI |
| AI 药物发现 | 药企研发和发现平台团队 | 药物化学家、实验室自动化科学家 | 研发预算 / 平台授权 | 药物发现负责人 / 数字化转型负责人 | 缩短周期和实验自动化 |
买方角色基于合作公告和市场结构推断;PFN 未按细分市场披露详细采购流程或合同经济性。
[CM029, CM030, CM031, CM032, CM033, CM034]PFN 各分部对应不同买方、用户、付款方和触发模式。
矩阵为定性判断,基于公开合作伙伴公告和市场结构。
[CM029, CM030, CM031, CM032, CM033, CM034]2.4 增长驱动、约束与尽调缺口
增长驱动强但不均衡。AI 基础设施支出、日本主权计算优先级、机器人劳动力约束、智能制造数字化、自动驾驶 physical AI,以及制药缩短发现周期的压力,都支撑 PFN 能力的需求。MHI 和 Toyota 2026 年公告提升了证据新鲜度,也显示 PFN 仍在把研究深度转化为战略伙伴入口。约束同样重要:工业 AI 需要漫长验证周期,安全关键基础设施买方要求可靠性和可审计性,AI 加速器面临 NVIDIA 级生态锁定和晶圆代工产能限制,农业机器人单位经济性差且受季节性影响,药物发现 AI 则有临床转化风险。Chugai 管线的不利信号并非 PFN 专属,但说明计算发现应被视为高上行、高验证风险敞口。公开证据无法计算 PFN SOM;收入结构、伙伴合同经济性和外部 MN-Core 销售仍是数据室问题。[CM025, CM026, CM027, CM038, CM039, CM040]
| 驱动 / 约束 | 方向 | 时间 | 对 PFN 的影响 | 尽调问题 |
|---|---|---|---|---|
| 日本主权 AI 基础设施,以及 NEDO 支持的后 5G 研发 | 驱动 | 2026–2030 | 支撑 MN-Core / AImod 和本土算力差异化 | 确认补助经济性、AImod 利用率,以及外部客户是否为 MN-Core 产能付费 |
| 工业 AI 从试点转向嵌入式制造工作流 | 驱动 | 2026–2031 | 支撑 MHI / Fanuc 式共创,以及智能制造平台需求 | 量化合作伙伴管线从联合研究转为付费部署的情况 |
| 日本和全球工业机器人装机基数 | 驱动 | 当前且具周期性 | 为机器人智能软件提供庞大装机基础 | 判断 PFN 是否以及如何参与 Fanuc 部署收入分成 |
| Toyota 物理 AI 推理研究 | 驱动 | 2026–2029 | 验证 MN-Core L 在汽车和机器人推理中的用例 | 询问该研究是否有商业里程碑,还是只属探索性研发 |
| AI 加速器需求和 GPU 供应压力 | 驱动 | 2026–2035 | 巨大的芯片 TAM 支撑 MN-Core 叙事 | 评估晶圆代工获取、软件生态和外部销售牵引力 |
| 药企周期压力 | 驱动 | 2026–2033 | 支撑 Chugai 式计算化学和实验自动化 | 要求披露 AI 建议化合物转为已验证候选物的转化指标 |
| 依赖伙伴主导商业化 | 约束 | 持续 | PFN 可能依赖伙伴完成市场进入和收入获取 | 审查合同条款、排他性、IP 归属和毛利分成 |
| 安全关键验证和可审计性 | 约束 | 近期到中期 | 基础设施、汽车和药企买方都要求漫长验证周期 | 要求按垂直行业给出从 PoC 到生产部署的时间线 |
| 相比 NVIDIA 级平台的芯片生态壁垒 | 约束 | 持续 | MN-Core TAM 可以很大,但实际份额可能仍局限于内部或日本市场 | 对标编译器、模型支持、客户迁移成本和总拥有成本 |
| 农业机器人领域的 PFN 证据缺口 | 约束 | 当前 | 没有新的 PFN 商业化证明前,农业不应驱动估值 | 要求披露 CraftyFarm 状态、付费客户、单位经济性和部署地理 |
| Chugai AI 辅助抗体项目终止的不利信号 | 约束 | 2026 信号 | 说明药物发现 AI 即便平台有想象力,也可能无法转化 | 将平台收入与治疗里程碑假设分开 |
时间判断为定性;这些约束是尽调重点,因为 PFN 披露合作和技术比披露收入转化指标更清晰。
[CM004, CM005, CM006, CM007, CM016, CM017]PFN 的采用路径由合作伙伴主导,在可重复商业化之前会经历流失。
百分比为示意性漏斗估算;PFN 未披露转化率或销售周期指标。
[CM029, CM038, CM039]2.5 图表
03竞争对手
3.1 PFN 多重任务场景下的竞争格局
Preferred Networks 没有一个清晰的同业组;它有多个重叠战场,因为公司横跨 AI 芯片、深度学习软件、机器人感知、日本基础模型、AI 药物发现和农业机器人。在 AI 基础设施中,MN-Core 要对抗 NVIDIA H100、H200 和 Blackwell 的引力,也要面对 AMD MI300、Intel Gaudi、Google TPU、Cerebras、Graphcore 和 SambaNova。在机器人与感知中,相关替代方案包括 NVIDIA Isaac、Boston Dynamics Spot、Covariant、Skild AI、Physical Intelligence、Figure AI 和 Sanctuary AI。汽车业务要与 Waymo、Wayve、Mobileye、NVIDIA DRIVE 以及 Toyota 自有 Woven 组织比较。PLaMo 在日本本土的心智份额与 Sakana AI、rinna、ABEJA 和 ELYZA 竞争。PFN Bio 面对 Recursion、Isomorphic Labs、Insilico、BenevolentAI 和 Schrödinger,而 CraftyFarm 暴露于 Plenty、FarmWise 和 Carbon Robotics 等专业农业机器人。现状替代也很关键:许多客户可以内部自建、租用云 GPU、购买现成机器人平台,或保留领域专属团队在内部。[CP001, CP002, CP004, CP007, CP008, CP011]
| 竞争对手 | 类别 | 规模 / 融资信号 | 目标细分市场 | 主要差异化 | 相对 PFN 的关键限制 |
|---|---|---|---|---|---|
| Preferred Networks | 参照公司 | 日本 AI / 机器人独角兽;保留章节来源未披露私有经营指标 | AI 芯片、LLM、机器人、药物发现、农业 | 罕见的跨领域研发宽度,以及 MN-Core / PLaMo / CraftyFarm 期权组合 | 业务过宽稀释焦点,各垂直领域的公开产品证明强弱不一 |
| NVIDIA | AI 加速器 + 机器人平台 | 全球上市 AI 基础设施龙头,拥有 H100 / H200 / Blackwell 路线图 | 数据中心训练 / 推理、机器人、自动驾驶 | GPU 生态、CUDA / 软件、企业 AI、Isaac 和 DRIVE | 只有在定制芯片或日本本土集成压过生态引力的细分场景里,PFN 才能差异化 |
| AMD / Intel / Google TPU | AI 加速器替代方案 | 大型既有厂商或 hyperscaler 基础设施提供方 | AI 训练和推理买方 | NVIDIA 之外的采购选择;TPU 具备云集成 | 即便 MN-Core 技术有差异化,它们也会压低 PFN 定价和采用率 |
| Cerebras / Graphcore / SambaNova | 定制 AI 芯片 / 平台专家 | 专用 AI 架构供应商 | 大模型训练 / 推理和企业 AI 平台 | 非 GPU 架构和垂直整合 AI 系统 | 说明定制芯片定位已经拥挤且资本密集 |
| NVIDIA Isaac / Boston Dynamics | 工业机器人平台 | 大生态或成熟机器人平台品牌 | 机器人仿真、感知、检测、移动机器人 | 开发者生态和硬件平台可得性 | 不是类似 PFN 的日本跨领域 AI 研究栈 |
| Covariant / Skild / Physical Intelligence | 机器人基础模型专家 | 风投支持的机器人 AI 专家;Covariant 部分团队通过 Amazon 人才交易被吸收 | 仓储和通用机器人智能 | 聚焦机器人基础模型叙事 | 整合和融资竞赛可能跑在 PFN 感知技术变现之前 |
| Waymo / Wayve / Mobileye | 自动驾驶 AI | 成熟 AV / ADAS 组织,背靠大型资金方或有公开市场可见度 | 自动驾驶软件、robotaxi、ADAS / OEM 技术栈 | 部署证明、车规级技术栈、数据优势 | PFN 的汽车业务必须证明 OEM 为什么需要外部日本 AI 实验室 |
| Woven by Toyota 内部替代 | OEM 内部自建 | Toyota 控制的内部软件 / 移动出行组织 | Toyota 及其盟友的移动出行软件 | 绑定 OEM 入口并控制内部路线图 | 更像替代方案,而不是第三方供应商竞争 |
| Sakana AI | 日本 AI 基础模型 / 研究同业 | 高曝光日本 AI 初创公司 | 基础模型和 AI 研究 | 研究品牌和聚焦日本的 AI 叙事 | 较少看到 PFN 式芯片 / 机器人 / 药物发现宽度的证据 |
| rinna / ABEJA / ELYZA 日本 AI 同行 | 日本 AI 企业 | 聚焦消费者、企业或 LLM 的日本 AI 供应商 | 日语 AI 和企业部署 | 本地客户入口,以及更清晰的 AI 服务定位 | 全栈硬件 / 机器人范围窄于 PFN |
| Recursion / Isomorphic Labs | AI 药物发现龙头 | 专注 AI 药物发现的品牌,具备药企可信度 | 生物学、化学、药物发现 | 单一垂直深度和公开平台身份 | 无法匹配 PFN 的芯片 / 机器人宽度,但 PFN Bio 可能被其规模压制 |
| Insilico / BenevolentAI / Schrödinger 药物发现平台 | AI / 计算药物发现 | 专业发现和计算化学平台 | 药企研发和分子设计 | 药物发现工作流专精 | PFN Bio 必须证明差异化生物数据或伙伴牵引力 |
| Plenty / FarmWise / Carbon Robotics | 农业自动化 | 专业农业自动化供应商 | 垂直农场、除草、作物自动化 | 明确的作物 / 工作流专项 ROI 主张 | 与 PFN 的 AI 宽度可比性较低,但在狭窄农场工作流上更强 |
| 内部自建 / 现状维持 | 替代方案 | 大客户已经拥有工程师、数据、采购或传统运营 | 汽车、药企、制造、农场 | 控制力、定制化,并避免供应商锁定 | 速度较慢、新意较弱,但往往是采购阻力最小的路径 |
画像行是局部且受证据约束;若保留来源未提供当前可比数值,则省略融资和收入。
[CP001, CP004, CP007, CP008, CP011, CP012]按两个尽调维度给 PFN 与主要替代方案排序:垂直领域广度、产品 / 生态深度。 分数来自分析师基于留存来源的判断,不是经审计的市场份额数据。
序数评分依据公开产品范围和生态证据;芯片、机器人、LLM、药物发现和农业之间 没有来源支持的统一量化基准。
[CP001, CP004, CP011, CP016, CP019, CP020]3.2 竞争对手画像深挖
最强芯片竞争者更多是在生态上击败 PFN,而不是单点硅片指标。NVIDIA 已安装的软件栈、路线图和企业级封装,使 H100/H200/Blackwell 成为默认比较对象;Google TPU 有 hyperscaler 集成;AMD 和 Intel 提供采购替代;Cerebras、Graphcore 和 SambaNova 则说明非 GPU AI 硅片并非 PFN 独有。机器人竞争同样由生态主导。NVIDIA Isaac 用仿真和部署工具包围感知,Skild AI 和 Physical Intelligence 追求通用机器人基础模型。Covariant 尤其有启发:其与 Amazon 的人才和授权交易是不利证据,说明战略买方可以不收购整家公司,也能获取稀缺机器人 AI 能力。汽车领域,Waymo 和 Mobileye 是更成熟的自动驾驶参照,Wayve 在具身 AI 上更接近,Toyota 旗下 Woven 则是日本 OEM 内部自建威胁。日本 AI 群体分散但本土相关,Sakana 在研究叙事上更强,ABEJA/ELYZA 在企业部署姿态上更强。[CP004, CP005, CP006, CP009, CP010, CP011]
| 能力 | PFN | NVIDIA | Google / AMD / Intel | 机器人 AI 初创公司 | 自动驾驶专家 | 日本 AI 同业 | 药物 / 农业专家 |
|---|---|---|---|---|---|---|---|
| 数据中心 AI 训练加速器 | MN-Core 专用芯片线 | 极强:H100 / H200 / Blackwell | 强:TPU、MI300、Gaudi | Unknown | 否 | 否 | 否 |
| AI 软件生态 | 深度学习平台和 PLaMo | 企业 AI、Isaac、DRIVE 极强 | 部分覆盖,主要是基础设施 | 部分覆盖,偏机器人专用 | 自动驾驶赛道较强 | 日本 AI 领域从部分到较强 | 垂直专用 |
| 工业机器人感知 | PFN 机器人 / 感知积累 | 通过 Isaac 机器人较强 | 无直接产品焦点 | 强且聚焦 | 车辆感知部分覆盖 | 有限 | 有限 |
| 自动驾驶 AI | 汽车感知积累 | 通过 DRIVE 较强 | 直接技术栈有限 | 有限 | 极强:Waymo、Wayve、Mobileye | 有限 | 否 |
| 日语 LLM | PLaMo | 保留来源未显示其专攻日本 | 保留证据不清晰 | 否 | 否 | 同业组合较强 | 否 |
| AI 药物发现 | PFN Bio 活动 | 间接算力供应商 | 间接算力供应商 | 否 | 否 | 否 | 专业平台较强 |
| 农业机器人 | CraftyFarm 活动 | 间接机器人工具 | 否 | 通用机器人部分覆盖 | 否 | 否 | 作物 / 工作流聚焦度强 |
| 分销生态 | 日本研发与合作伙伴网络 | 全球生态极强 | 云与既有厂商渠道强 | 风险投资 / 创业公司渠道 | OEM / 运营商渠道强 | 日本企业渠道 | 制药 / 农场垂直渠道 |
| 公开定价透明度 | 低 | 低到中,常按报价 | 云与硬件定价因渠道而异 | 低 | 低 | 低 | 低 |
单元格只给方向性判断。未知和部分条目说明留存公开证据缺失,不代表不存在。
[CP002, CP004, CP007, CP008, CP011, CP013]热力图概括 PFN 覆盖更广的地方,以及留存公开证据显示聚焦型对手更深的地方。 正向表示证据强,中性表示证据不完整,警告表示证据弱或没有留存证据。
能力色调是定性判断且受证据约束;未知的私有部署不计入。
[CP001, CP011, CP016, CP020, CP023, CP026]3.3 能力、定价与分发矩阵
公开证据支持方向性比较,而非经审计的基准同等性。PFN 的主要优势是不寻常的广度:它可以从同一个研究组织出发,可信地谈芯片设计、日本 LLM、机器人、生命科学和农业。对寻找长期 AI 研发伙伴的客户来说,这种广度有价值,但不会自动在每个分部都生成 best-of-breed 产品证明。NVIDIA 和 Google 在加速器生态分发上更强,Waymo/Mobileye 在已部署自动驾驶验证上更强,Recursion/Isomorphic/Schrödinger 在药物发现品牌深度上更强,FarmWise/Carbon Robotics 在窄农业 ROI 主张上更强。定价大多不透明,因此买方会比较封装而非标价:云或服务器加速器可用性、机器人平台购买或试点条款、企业 AI 服务、制药合作经济性,以及项目专属农业自动化。未知单元格被有意标为未知或部分,因为公开页面很少披露模型质量基准、客户胜率、实际价格或部署单位经济性。[CP023, CP024, CP025, CP026, CP027, CP028]
| 竞争对手组 | 包装模式 | 已知公开定价信号 | 透明度 | 买方含义 |
|---|---|---|---|---|
| PFN MN-Core / 平台 | 定制芯片、软件、研究合作 | 未留存可比公开标价 | 低 | 尽调必须拿到芯片 / 软件实际经济性和支持承诺 |
| NVIDIA H100/H200/Blackwell | GPU / 服务器 / 云生态加企业软件 | 产品规格公开;服务器 / 云实际定价因渠道而异 | 中 | 软件生态降低采用风险,即使有溢价,默认选项也可能胜出 |
| AMD MI300 / Intel Gaudi | 加速器硬件与合作伙伴系统 | 有公开产品页,但留存来源没有标准化企业实际 TCO | 低 | 可作为采购筹码,对冲 NVIDIA 与 PFN 定制芯片 |
| Google TPU | 云加速器消耗 | 云平台开放 TPU 访问;具体工作负载经济性仍需建模 | 中 | 云可用性可能压过定制硬件的采用摩擦 |
| NVIDIA Isaac / DRIVE | 开发者平台、SDK、车辆 / 机器人栈 | 公开文档和平台定位可见;商业条款未完全公开 | 低 | 捆绑生态可能挤掉 PFN 定制感知工作 |
| Boston Dynamics Spot | 机器人平台销售与生态载荷 | 未留存标价;产品定位公开 | 低 | 硬件采购动作比采用 PFN 跨领域栈更简单 |
| 机器人基础模型初创公司 | 企业试点、授权、机器人部署 | 大多未披露 | 低 | 战略收购方可能更看重人才 / 模型访问,而非收入 |
| 日本 AI 同行 | 企业 AI 服务、LLM API、模型项目 | 留存来源大多未披露 | 低 | PFN 必须证明 PLaMo 容易购买和部署 |
| AI 药物发现平台 | 合作、平台授权、内部管线经济性 | 大多随合作而定 | 低 | PFN Bio 需要交易证据,而不是泛泛的 AI 主张 |
| 农业机器人 | 设备、服务、作物工作流自动化 | FarmWise 和 Carbon Robotics 发布 ROI 式主张,但没有完整价目表 | 低 | CraftyFarm 必须拿出相对专用设备的作物级 ROI |
该定价表刻意采用包装和透明度口径,因为多数私营和企业 AI 厂商不发布可比标价。
[CP028, CP029, CP030, CP032, CP035, CP036]按竞争赛道给 PFN 护城河就绪度做序数评分。就绪度分数越高,防御性越强; 风险分数越高,替代压力越大。
分数是基于来源支持的竞争者证据得出的定性尽调评级,不是经审计的 KPI。
[CP030, CP031, CP034, CP035, CP036, CP037]3.4 护城河持久性、切换成本与替代风险
PFN 的护城河最有防御力的地方,是其研究广度变成复利能力,而不是一组互不相关的押注。如果客户围绕 PFN 硬件和软件优化工作负载,MN-Core 可以形成切换成本,但这种优势容易受 CUDA、TPU 可用性和采购习惯冲击。如果 PFN 掌握生产数据和部署工具,机器人感知会变黏;但机器人基础模型初创公司和 NVIDIA Isaac 可以把部分技术栈商品化。PLaMo 受益于日语专长,但 GTM 更清晰的日本 AI 厂商可以赢得账户。PFN Bio 和 CraftyFarm 有期权价值,但专注药物发现或农业的竞争对手能展示更干净的垂直聚焦。最大风险是多供应商并用:客户可以用 NVIDIA GPU、Toyota 或 Mobileye 自动驾驶栈、日本 LLM 厂商、制药 AI 专家和作物专属机器人,而不在 PFN 上标准化。因此,尽调不应只问 PFN 是否有聪明技术,而要测试 PFN 在哪里拥有产品化证明、分发和聚焦对手难以中和的切换成本。[CP015, CP029, CP030, CP031, CP034, CP035]
| 护城河主张 | 威胁向量 | 严重性 | 缓释措施 / 尽调要求 | 主张 |
|---|---|---|---|---|
| MN-Core 定制硅 | NVIDIA CUDA 生态、Google TPU 云访问,以及既有加速器路线图 | 高 | 获取相对 H100/H200/Blackwell/MI300/Gaudi/TPU 的每美元基准性能、功耗、可用性和客户留存证据 | CP004; CP007; CP035 |
| 跨领域 AI 研究广度 | 聚焦型对手在各垂直领域执行胜过 PFN | 高 | 拆分平台协同与无关期权价值,并按细分市场要求产品证据 | CP001; CP031; CP040 |
| 机器人感知专长 | 机器人基础模型与 NVIDIA Isaac 将感知层商品化 | 中 | 要求量产部署指标和客户数据优势 | CP011; CP013; CP034 |
| 汽车 AI 关系 | Waymo、Mobileye、NVIDIA DRIVE、Wayve 和 Toyota Woven 降低 OEM 对 PFN 的需求 | 高 | 核验具名 OEM 管线,以及 PFN 在哪些环节已嵌入、哪些环节可替换 | CP016; CP018; CP019 |
| PLaMo 日语护城河 | Sakana、rinna、ABEJA 和 ELYZA 争走日本 AI 注意力或企业预算 | 中 | 对比模型质量、服务成本、客户背书和集成支持 | CP020; CP021; CP036 |
| PFN Bio 期权价值 | 专注 AI 药物发现的公司拥有更强垂直品牌和制药工作流 | 中 | 要求合作伙伴管线、里程碑和湿实验验证证据 | CP023; CP024; CP037 |
| CraftyFarm / 农业机器人 | 面向特定作物的机器人厂商展示更直接的 ROI 主张 | 中 | 验证日本作物 / 劳动力使用场景和单位经济性 | CP026; CP027; CP038 |
| 客户切换成本 | 客户同时使用多家 GPU、LLM、机器人平台和领域厂商 | 高 | 衡量工作负载锁定、数据可携带性、再训练成本和合同续约行为 | CP030; CP039; CP040 |
风险严重性为定性判断,基于留存公开证据;尽调要求列出验证或推翻各项风险所需的私有证据。
[CP001, CP004, CP007, CP011, CP013, CP016]3.5 图表
04财务
4.1 融资历史与资本结构
对一家日本私营 AI 公司而言,PFN 的公开融资记录异常丰富,但它仍应被读作资本结构地图,而不是完整财务报表。最新保留的一级证据是 2024 年 12 月首关 190 亿日元,随后 2025 年 4 月扩展,使融资系列达到 240 亿日元。结构是混合型:战略和财务投资者提供股权,银行提供债务,具名贷款方包括 MUFG Bank、SMBC、Resona、Shoko Chukin,之后还有 Mizuho。这种组合很重要,因为它说明 PFN 不只依赖风险资本,也能接触蓝筹资产负债表和银行;但它也带来关于担保、契约、债务到期日,以及计算基础设施是在 PFN 资产负债表上融资还是通过伙伴融资的尽调问题。Toyota 和 FANUC 的历史战略投资显示同一种模式:PFN 长期靠也想获得其技术的工业伙伴来资助深度 AI 研发。[CI001, CI002, CI003, CI005, CI006, CI007]
| 日期 | 事件 | 金额 | 工具 / 结构 | 已披露参与方 | 财务解读 |
|---|---|---|---|---|---|
| 2015-08 | FANUC 资本联盟 | ¥0.9B | 第三方定向增发 / 战略股权 | FANUC | 工业战略背书;规模小,但机器人合作伙伴信号强 |
| 2017-08 | Toyota 追加投资 | ~¥10.5B | 第三方定向发行新股 | Toyota Motor Corporation | 大额战略融资,确立 PFN 作为日本重要 AI 资产的地位 |
| 2024-08 | SBI 资本与业务联盟协议 | 最高 ¥10B | 计划于 2024 年 9 月底前完成第三方定向增发 | SBI Group / SBI Holdings | 聚焦半导体的资本联盟;最新融资系列的前奏 |
| 2024-12 | 最新融资首关 | ¥19B | 股权加债务融资 | SBI Group、DBJ、Mitsubishi Corporation、Sekisui House、Wacom;MUFG、SMBC、Resona、Shoko Chukin 等资本方 | 面向芯片、PLaMo、招聘和算力基础设施的大额多工具融资 |
| 2024-12 | The Bridge 累计融资数据点 | 已披露累计约 ¥36B | 媒体聚合 | The Bridge | 可作交叉核验,但不是公司股权表 |
| 2025-04 | 延展轮 | ¥5B | 股权加 Mizuho 债务 | Kodansha、Mitsubishi UFJ Trust、Sumitomo Mitsui Trust、TBS、Toei Animation、Sekisui House、Mizuho Bank 等资本方 | 将该系列融资总额推至 ¥24B,并扩大了战略 / 财务投资者基础 |
| 2025-06 | PremierAlts 融资数据点 | 累计融资 $315.4M | 市场数据显示最近一轮 $165.9M | PremierAlts | 独立市场数据估算;有参考价值,但与日元口径披露时间线冲突 |
| 2026-06 | 公开资金续航状态 | 未披露 | N/A | 未找到现金或烧钱披露 | 缺少资金余额和月度烧钱数据,无法把轮次规模换算成资金续航期 |
列举并不完整:仅覆盖留存公开融资事件和市场数据估算,不包括未披露股东转让或保密债务条款。
[CI001, CI005, CI006, CI009, CI010, CI011]| 类别 | 具名方 | 轮次 / 日期 | 可能的战略价值 | 未决尽调要求 |
|---|---|---|---|---|
| 战略产业股权 | Toyota, FANUC | 2015–2017 | 汽车与工厂自动化背书 | 当前持股、权利和商业承诺 |
| 财务 / 战略股权首关 | SBI Group、DBJ、Mitsubishi Corporation、Sekisui House、Wacom 等投资方 | Dec 2024 | 资本加分销、半导体和产业网络支持 | 精确股类、清算优先权和董事会权利 |
| 首关债务提供方 | MUFG Bank、SMBC、Resona、Shoko Chukin 银行方 | Dec 2024 | 与股权并行的银行授信通道 | 各贷款方债务规模、契约、期限和抵押品 |
| 延展轮投资人 | Kodansha、Mitsubishi UFJ Trust、Sumitomo Mitsui Trust、TBS、Toei Animation、Sekisui House 等投资方 | Apr 2025 | 内容、金融和影响力股权战略支持 | 权利、战略义务,以及投资人是否同时是客户 |
| 延展轮贷款方 | Mizuho Bank | Apr 2025 | 额外银行信贷能力 | 债务条款,以及授信是否由 IP 或应收账款担保 |
| 未核实种子投资人 | ENEOS, Chugai Pharmaceutical | 未在留存最新轮文件中核实 | 最新轮没有确认 | 缺少直接证据时,不纳入最新轮股权表 |
投资人地图基于具名公开披露;它不是完整资本结构表,也不包括未披露个人股东。
[CI002, CI003, CI007, CI008, CI009, CI010]PFN 的融资历史显示,公司多次引入战略产业资本,随后在 2024–2025 年完成一轮大额 股权 / 债务混合融资。
时间线只纳入留存公开事件,排除未披露转让或更早小额轮次。
[CI001, CI005, CI006, CI009, CI010, CI011]4.2 估值轨迹与市场数据冲突
估值故事有吸引力,但不干净。The Bridge 的日文报道显示,2024 年 12 月轮后 PFN 市值超过 3000 亿日元;Latka 列出 2024 年 $2 billion 估值。这些数据点与 PFN 是日本 AI 独角兽这一基础事实方向一致。然而,PremierAlts 列出截至 2025 年 6 月明显更低的 $1.0 billion 估值;任何依赖简单 $2 billion 标题的投资者都应把它视为不利数据点。承销答案是展示区间,而不是选择最漂亮的数字。如果估计收入约为 $42 million 至 $56 million,$2 billion 估值意味着约 36x 至 48x 收入;较低的 $1.0 billion 估计仍意味着约 18x 至 24x 收入。任一倍数区间都要求投资者相信 PFN 的芯片、PFCP cloud、PLaMo 模型和工业解决方案能复利成高质量毛利。[CI014, CI015, CI016, CI017, CI024, CI025]
| 估值来源 / 情景 | 估值 | 收入分母 | 隐含收入倍数 | 立场 | 含义 |
|---|---|---|---|---|---|
| The Bridge / 日本市场叙事 | >¥300B (~$2B+) | $42M Latka | ~48x | 证实 | 需要极强的芯片 / 云 / 模型上行空间 |
| Latka 2024 估值 | $2.0B | $56M AI Market Watch 上沿估算 | ~36x | 证实 | 相比多数非纯 SaaS 收入画像仍有溢价 |
| PremierAlts 反向数据点 | $1.0B | $42M Latka | ~24x | 反向 | 较低估值将头部数字砍半,但要求仍高 |
| PremierAlts 反向数据点 | $1.0B | $56M 上沿估算 | ~18x | 反向 | 只有毛利率和增长强劲,才更可承保 |
| 未选择估值 | 需要区间 | 仅收入估算 | 18x–48x | 中性 | 在提供股权表和经审计收入前,使用估值区间 |
倍数只是估值除以公开估算收入;未考虑净现金、债务、优先股条款、收入结构和时间差。
[CI014, CI015, CI016, CI017, CI024, CI025]公开估值区间横跨 3000 亿日元以上的标题口径,以及较低的 $1.0B 市场数据点。
日元数值四舍五入为等值美元区间;汇率只是示意,不是经审计换算。
[CI014, CI015, CI016, CI017, CI026]4.3 收入牵引与披露缺口
PFN 没有公布传统私营公司承销模型所需的收入、ARR、毛利率、现金余额或烧钱速度包。因此,保留证据只能支撑估计区间和披露结论。Craft 列出 FY2023 收入 77 亿日元;Latka 列出 2024 年收入 $42 million;Growjo 估计 $49.5 million;AI Market Watch 指向历史 84.86 亿日元指标和 280–340 名员工区间;RocketReach 给出低得多的 2026 年 $15.3 million,像是离群值。稳妥做法是把中间簇作为方向性证据,并明确排除经审计财务处理。收入模式本身看起来多元:伙伴共同创造、AI 解决方案、PFCP 云计算、自有 AI 芯片、PLaMo 相关产品,以及潜在合资计算服务。但分部收入、经常性占比、客户集中度、积压订单转化和收入确认仍只属于私有证据。[CI018, CI019, CI020, CI021, CI022, CI023]
| 来源 | 报告指标 | 数值 | 期间 / 日期 | 置信度 | 模型用途 |
|---|---|---|---|---|---|
| Craft | 收入 | ¥7.7B | FY2023 | 中低 | 历史锚点;仅为聚合器 |
| Latka | 收入 / ARR 表述 | $42M | 2024 / 2025 更新 | 中低 | 中间估算簇的下沿;未经审计 |
| Growjo | 估算年收入 | $49.5M | 运行日当前页面 | 中低 | 中位估算;可用于三角校验 |
| AI Market Watch 数据源 | 历史收入 | ¥8.486B (~$56M) | 当前资料页引用截至 2021 年 1 月的财年 | 低 | 上沿估算,但日期 / 陈旧度不清 |
| RocketReach | 年收入 | $15.3M | 2026 页面 | 低 | 离群值;作为警示,不作基准情形 |
| PFN 官方发布 | 收入 / ARR / 毛利率 | 未披露 | 2024-2026 | 缺失置信度高 | 确认需要管理层 P&L |
所有收入数字均为第三方估算或聚合器资料;PFN 在留存官方来源中未披露经审计收入、ARR、毛利率或分部收入。
[CI018, CI019, CI020, CI021, CI022, CI023]| KPI | 公开数值 / 估算 | 来源视角 | 状态 | 投资判断处理 |
|---|---|---|---|---|
| 收入运行率 | ~$42M–$56M 公开估算区间 | Craft / Latka / Growjo / AI Market Watch 数据源 | 估算 | 仅作为敏感性输入 |
| ARR | 官方未披露 | 官方来源沉默 | 缺失 | 要求提供 ARR 和经常性收入组合 |
| 毛利率 | 未披露 | 官方来源沉默 | 缺失 | 要求按芯片、云、模型和解决方案拆分分部 COGS |
| 员工数 | ~275 到 340 的公开估算区间 | Growjo / AI Market Watch / Latka 数据源 | 估算 | 仅用于粗略计算人均收入 |
| 估值 | 从 ~$1B 的不利口径到 ~$2B+ 的头条口径 | PremierAlts 对比 The Bridge / Latka | 冲突 | 使用估值区间 |
| 累计融资 | 已披露累计约 ¥36B,到市场数据估算约 ~$315M | The Bridge / PremierAlts / Growjo | 估算 | 与 cap table 对账 |
KPI 表有意把公开估算和公司披露事实分开;null 代表私有指标缺失,不是零。
[CI012, CI013, CI014, CI015, CI016, CI017]公开收入证据大致集中在 $42M 至 $56M,但包含一个更低的离群点。
区间只使用第三方公开估算;PFN 未披露经审计收入或 ARR。
[CI019, CI020, CI021, CI022, CI023, CI024]可投资的财务论证依赖估算和缺失的私有指标,而不是经审计的公开财务数据。
KPI 卡混合了已确认融资事实、第三方估算,以及不可得私有指标的显式空值。
[CI006, CI017, CI018, CI021, CI024, CI042]4.4 资本强度、Runway 与基础设施杠杆
PFN 最强的财务信号也是核心风险:它在垂直整合一个资本密集型 AI 栈。官方来源显示,公司正在开发 MN-Core 处理器、PFCP 计算基础设施、PLaMo 基础模型,以及面向 2026 年商业化的 MN-Core L1000 处理器。这比纯应用软件业务更贵,因为硅片开发、系统工程、高密度计算和 AI 模型训练,都需要在产品市场经济性完全可见前持续投入技术开支。2024 年 12 月和 2025 年 4 月融资减轻了短期融资压力,Preferred Computing Infrastructure 等伙伴载体也可能把一部分云商业化转入与 Mitsubishi Corporation 和 IIJ 共享的结构。METI/NEDO 计算资源项目也改善生态支持。但这些来源都没有披露在手现金、月度烧钱或 runway 月数,因此公开尽调无法证明业务可自我融资或运营利润具备持久性。[CI030, CI031, CI032, CI033, CI034, CI035]
| 项目 | 公开数值 / 状态 | 置信度 | 重要性 | 尽调要求 |
|---|---|---|---|---|
| 最新融资系列 | ¥24B Dec 2024-Apr 2025 | 高 | 改善近期流动性,也显示可获得战略 / 银行资本 | 扣除费用后的到账现金,以及债务 / 股权拆分 |
| 手头现金 | 未披露 | 缺失置信度高 | 缺少资金余额,无法计算资金续航期 | 银行对账单和董事会现金报告 |
| 月度烧钱 | 未披露 | 缺失置信度高 | AI 芯片、云和模型训练会消耗大量现金 | 过去 18 个月现金流和预测 |
| 债务义务 | 具名银行贷款方,但条款未披露 | 中 | 债务可能改变下行保护和资金续航期 | 全部信贷协议和契约时间表 |
| 算力资本开支敞口 | 部分通过 PFCI 借合作伙伴杠杆 | 中 | JV 可能降低 PFN 独立资本开支负担 | PFCI 出资、担保和服务协议 |
| 政府算力支持 | GENIAC 算力资源支持存在于生态层面 | 中 | 可能抵消模型开发成本,但不是现金收入 | PFN 专属的授予、额度和限制 |
| 下一轮融资触发条件 | 未公开说明 | 缺失严重性高 | 决定公司对融资的依赖程度 | 按里程碑和烧钱情景拆出的管理层计划 |
这里有意不计算 runway,因为现金、烧钱、债务条款和项目融资义务均未公开。
[CI004, CI006, CI033, CI035, CI036, CI037]近期融资提供了充足燃料,但在盈利能力可见之前,芯片、云和模型开发会先吃掉资本。
资金用途条目是根据已披露战略重点做出的示意性分配,不是公司预算披露。
[CI004, CI006, CI012, CI030, CI033, CI035]4.5 财务结论与尽调要求
PFN 应按一家由战略资本融资、技术差异化明显但私有经济性不透明的 AI 基础设施公司来承销。正面案例是,PFN 长期获得 Toyota、FANUC、Mitsubishi Corporation、SBI、主要信托银行和政府邻近计算计划的工业验证;近期融资系列规模大;并且拥有一套可能把受约束的日本 AI 基础设施需求转化为云和产品收入的技术栈。反面案例是,同一技术栈成本高,公开收入估计互相矛盾;如果收入只落在公开 $42 million 至 $56 million 估计区间,估值可能偏紧。因此,眼前尽调清单不是可选项:经审计或管理层编制的 P&L、分部收入、芯片 / 云 / 解决方案毛利率、债务条款、现金余额、月度烧钱、积压订单、头部客户集中度、PFCP 单位经济性,以及 PLaMo 和 MN-Core 商业化能在不无限依赖外部融资的情况下扩张的证据。[CI018, CI024, CI025, CI026, CI040, CI041]
| 缺失指标 | 严重性 | 重要性 | 具体尽调路径 |
|---|---|---|---|
| 经审计收入和收入确认 | 阻断性 | 估值敏感性取决于实际收入,不取决于新闻稿里的融资金额 | 要求提供经审计或管理层编制的财务报表,以及收入确认政策 |
| 按芯片、PFCP、PLaMo 和服务拆分的分部收入 | 重要 | 收入组合决定毛利率路径和收入质量 | 要求提供产品线 P&L 和头部客户合同 |
| 按分部拆分的毛利率 | 阻断性 | 硬件、云、服务和软件的利润率差异很大 | 要求提供 COGS 桥接表和完全负担口径毛利率 |
| 现金余额和月度烧钱 | 阻断性 | 公开来源无法计算 runway | 要求提供银行现金报告和月度现金流预测 |
| 债务条款和担保 | 重要 | 银行融资可能让新投资人处于次级地位 | 审查所有贷款协议、财务约束、抵押品和到期时间表 |
| 客户集中度和 backlog | 重要 | 战略合作不一定等于经常性收入 | 要求提供前 10 大客户收入、backlog、续约和已承诺最低额 |
| PFCP 单位经济模型 | 重要 | 云算力可能高增长,但资本开支重 | 要求提供利用率、每算力单位价格、电力成本和折旧政策 |
在把 PFN 当作已完成投资判断的标的、而不是来源交叉验证出的标的之前,这些是最低限度的财务底稿。
[CI018, CI024, CI025, CI026, CI041, CI042]4.6 图表
05产品与技术
5.1 架构与平台:垂直整合,而非单一产品
PFN 当前产品面最好理解为垂直整合的 AI 栈,而不是传统单一产品初创公司。公开首页称公司整合 AI 芯片、计算基础设施、生成式 AI、解决方案和产品;产品页面随后把该战略映射到四层。底层是 MN-Core 处理器和 PFCP 计算容量。其上是软件框架、编译器、运行时工具和开源库,帮助 PFN 团队把 PyTorch/JAX 时代的工作负载映射到 PFN 自有计算上。模型层是 PLaMo,这是一个面向日语的基础模型家族,提供 API、聊天、翻译、开放模型和企业定制界面。应用层包括用于原子级材料仿真的 Matlantis、Kachaka 和通过集团公司 Preferred Robotics 推进的工业机器人,以及面向制造、材料、生命科学、公共部门和企业客户的定制 AI 解决方案。这种架构形成真实差异化,因为 PFN 可以一起调优硅片、编译器、模型和应用;它也带来执行风险,因为每一层都有不同的商业化时钟和买方。[CE001, CE002, CE027, CE028, CE034, CE038]
| 产品 / 资产 | 主要用户 | 成熟度 / 状态 | 差异化 | 尽调缺口 |
|---|---|---|---|---|
| MN-Core / MN-Core 2 / MN-Core L1000 | PFN 算力用户、AI 基础设施买家 | 第一代已在 MN-3 跑通;MN-Core 2 列为可销售;L1000 仍在开发 | 为矩阵工作负载和 LLM 推理优化的定制 AI 芯片 | PFN / 关联方之外的外部采用,公开证据仍然薄 |
| PFCP 算力基础设施 | PFN 团队、需要 MN-Core / GPU 算力的合作伙伴 | 有公开产品页;云服务细节有限 | 把 PFN 自有加速器接到模型和仿真工作负载 | SLA、区域、安全控制和客户数未披露 |
| PLaMo 基础模型 | 日本企业、政府、开发者 | PLaMo Prime API / chat 已上线;Hugging Face 上有开放模型 | 聚焦日语,并与全栈算力打通 | 独立基准测试和安全审计覆盖不完整 |
| Matlantis / PFP | 材料和化学研发团队 | 商业化云仿真器;已进入美国 | 神经网络势能和 AI 原子级仿真工作流 | 收入规模和续约指标未披露 |
| Kachaka / 机器人技术栈 | 家庭、办公室、物流和工业机器人团队 | 商业化 Kachaka 产品;与 Toyota / FANUC 有关系 | 具身 AI 工作负载的本地推理验证场 | 单位经济模型和国际扩张不清楚 |
| 开源软件:CuPy、Optuna、pfio | ML 工程师和研究人员 | 文档 / repo 活跃;Optuna v4.0 和 CuPy 仍在维护 | 开发者信誉和生态招聘渠道 | Chainer 停止发展,暴露生态依赖风险 |
产品组合行只是截至 2026-06-14 公开可见产品和技术资产的部分枚举;客户数和收入贡献未公开。
[CE001, CE002, CE008, CE009, CE011, CE027]| 日期 | 里程碑 | 技术领域 | 含义 | 来源 |
|---|---|---|---|---|
| 2014-10 | Toyota 自动驾驶联合 R&D | 汽车 AI | 早期真实世界感知锚点 | SE006 |
| 2015-06 | Chainer 发布 | 深度学习框架 | 开源研究周期提速 | SE008 |
| 2018-12 | ChainerX 和 MN-Core 披露 | 框架 + 芯片 | PFN 追求软硬件协同设计 | SE009 / SE012 |
| 2019-12 | Chainer 维护和 PyTorch 迁移 | 框架战略 | 务实切换生态 | SE010 / SE037 |
| 2020-06 | MN-3 登顶 Green500 | 算力 | 能效获得独立验证 | SE028 |
| 2021-07 | Matlantis 云上线 | 材料仿真 | 研究转为商业云服务 | SE025 |
| 2022-12 | MN-Core 2 发布 | AI 芯片 | 从第一代基准测试转向可销售硬件路线图 | SE017 |
| 2023-11 | Preferred Elements 成立 | 基础模型 | 组建专门的 PLaMo 组织 | SE022 |
| 2024-12 | PLaMo Prime 发布 | 基础模型 | API / chat 商业化界面 | SE021 |
| 2026-06 | Toyota 使用 MN-Core L Series 开展 physical-AI 研究 | 机器人 + 芯片 | 具身 AI 路线图的最新证据 | SE024 |
时间线有选择性,强调产品技术里程碑,而不是融资或公司历史。
[CE016, CE018, CE019, CE006, CE034, CE008]PFN 把自研硅、算力基础设施、模型 / 软件资产和行业应用叠在一起,拼成一体化 AI 平台。
分层根据 PFN 产品页面和新闻稿推断;内部组件边界可能不同。
[CE001, CE002, CE008, CE027, CE034, CE038]5.2 AI 芯片与计算:MN-Core 赢得能效,商业化仍需证明
最强的硬科技证据来自 MN-Core 产品线。PFN 披露第一代 MN-Core 是一款 TSMC 12nm 矩阵运算加速器,目标是在半精度下达到 1.0 TFLOPS/W,随后在 MN-3 中使用 160 颗芯片。TOP500 独立印证了核心主张:MN-3 在 2020 年 6 月 Green500 以 21.1 GFLOPS/W 夺冠,并在 2021 年 11 月榜单以 39.38 GFLOPS/W 再次登顶。PFN 此后从基准演示走向产品化:芯片页面列出 MN-Core 2 板卡、八板 MN-Server 2,以及带有日本定价的 Devkit。同一页面声称在 Kachaka 优化和 Matlantis 仿真的真实工作负载中取得优势。但 GPU 替代尚未被证明。KDDI 公开 GPU Cloud 页面说明,主流企业 AI 基础设施仍以 NVIDIA GPUaaS 为中心,而 PFN 公开证据强调内部或关联工作负载。因此,尽调应把 MN-Core 视为可信的差异化硅片,但生态广度仍未被证明。[CE003, CE004, CE005, CE006, CE007, CE008]
| 代际 | 主要角色 | 已发布规格 / 声称能力 | 成熟度信号 | 关键风险 |
|---|---|---|---|---|
| MN-Core(gen 1) | AI 训练和 HPC 加速器 | TSMC 12nm;500W;524 TFLOPS 半精度;估算半精度效率 1.0 TFLOPS/W | 支撑 MN-3;Green500 领先地位由 TOP500 佐证 | 上一代产品;不能证明广泛市场采用 |
| MN-3 supercomputer | PFN 内部深度学习超算 | 160 颗 MN-Core 处理器,配专用互连 | 2020 和 2021 年 Green500 第 1 | 基准测试系统本身不是商业芯片业务 |
| MN-Core 2 | AI 训练 / HPC 板卡和服务器产品 | PFN 列示每块板卡 TF16 393 TFLOPS,MN-Server 2 为 3.1 PFLOPS TF16 | 入选 Hot Chips 2024;服务器 / 开发套件已列价 | 外部客户量和软件生态未知 |
| MN-Core L1000 | 生成式 AI 推理处理器 | 3D 堆叠内存 / 逻辑;声称 token 处理最高快 10x | 截至 2024-2026 路线图仍在开发 | 没有独立 token benchmark 或量产证据 |
| PFCP / MN-Core cloud 云服务 | 访问 PFN 算力的云入口 | MN-Core 2 已试验性用于 Matlantis 工作负载 | 官方计算 / 芯片页面描述了服务方向 | SLA、区域、合规以及 KDDI / PFN 托管细节未验证 |
除 Green500 结果由 TOP500 佐证外,规格均来自公司发布。除非另有说明,声称能力不是独立产品 benchmark。
[CE003, CE004, CE005, CE006, CE007, CE008]PFN 在芯片和材料上有强研究验证;安全控制和外部广泛采用 MN-Core 的公开证据更弱。
分数是 0–10 序数估算,只基于本章复核的公开证据。
[CE006, CE007, CE022, CE027, CE034, CE036]5.3 模型与软件:Chainer 遗产、PyTorch 迁移与 PLaMo 商业化推进
对一家日本工业 AI 公司而言,PFN 的软件可信度异常深。Chainer 于 2015 年发布,帮助普及 define-by-run 动态计算图;ChainerX 后来试图把 ndarray 和自动微分的性能关键路径迁入 C++。PFN 在 2019 年决定将 Chainer 转入维护并把研究迁移到 PyTorch,这本身是正向治理信号,而非失败:公司识别到框架生态正在整合,于是把工程能量转向 PyTorch 社区贡献、CuPy、Optuna 和应用专属工具。Optuna 和 CuPy 继续提供可见的开发者信号,PFN 的 Hugging Face 组织也给 PLaMo 外部模型分发界面。模型线已经从子公司实验转向核心战略:Preferred Elements 成立于 2023 年,PLaMo Prime 于 2024 年通过 API 和聊天发布,PFN 又在 2025 年宣布将吸收 Preferred Elements,以加快 PLaMo 的社会落地。[CE016, CE017, CE018, CE019, CE020, CE021]
| 资产 | 角色 | 截至 2026 年的状态 | 开发者信号 | 尽调读数 |
|---|---|---|---|---|
| Chainer | PFN 最初的深度学习框架 | 2019 年 12 月后仅维护 | GitHub repo 和 Chainer 公告仍公开 | 早期创新成立,但生态输给 PyTorch |
| ChainerX | C++ ndarray / autograd 组件 | Chainer stable 文档将其列为早期阶段功能 | 有技术文档 | 显示 PFN 的系统能力;但已不是当前生态锚点 |
| PyTorch 贡献 | 替代研究平台 | PFN 宣布迁移并合作 | PFN 官方发布 | 务实对齐主导框架 |
| CuPy | GPU NumPy / SciPy 数组库 | 项目和文档活跃 | GitHub、cupy.dev 和文档 | 持续的开源信誉 |
| Optuna | 超参数优化框架 | 文档活跃;PFN 报告 v4.0 采用 | GitHub、ReadTheDocs、PFN v4.0 发布 | 最强的广谱开发者足迹 |
| pfio | 统一文件系统 IO 库 | PFN GitHub 公开仓库 | GitHub 仓库 | 有用,但信号窄于 Optuna / CuPy |
| PLaMo models | 基础模型产品和开放模型渠道 | PLaMo Prime 加 Hugging Face org | PFN 网站和 Hugging Face | 模型层正在商业化,但基准测试覆盖还需尽调 |
开发者信号基于公开仓库、文档和模型分发页面,不基于私有使用遥测。
[CE016, CE017, CE018, CE019, CE020, CE021]PFN 的软件路径从 Chainer 创新转向 PyTorch 对齐,同时保留面向开发者的库和 PLaMo 分发。
流程把重叠工程工作流简化为里程碑顺序。
[CE016, CE017, CE018, CE019, CE020, CE021]5.4 机器人、材料与客户应用:关联工作负载验证技术栈
PFN 最具体的非模型应用落在机器人和材料。Toyota 关系始于 2014 年自动驾驶研发,2019 年扩展到服务机器人,并在 2026 年以 physical-AI 研究重新出现,使用 MN-Core L Series 处理器支持需要高速本地推理的机器人。FANUC 2015 年资本联盟给了 PFN 第二个工业机器人锚点。Preferred Robotics 的自主移动机器人产品 Kachaka 为集团提供商业机器人界面,也充当 MN-Core 工作负载案例。材料领域,PFCC 的 Matlantis 是更清晰的产品:它 2021 年作为云端原子级模拟器推出,2023 年扩展到美国,并由发表在 Nature Communications 的 PFP 神经网络势产品线支撑。这些应用很重要,因为它们为 MN-Core 和 PLaMo 创造了 captive workloads;风险在于,在披露更多非关联客户证据之前,关联验证可能高估第三方需求。[CE030, CE031, CE032, CE033, CE034, CE035]
| 客户 / 平台语境 | PFN 技术 | 工作流 | 证据强度 | 待解问题 |
|---|---|---|---|---|
| Toyota 自动驾驶和 physical AI | 感知 AI、服务机器人、MN-Core L Series | 汽车感知和机器人本地推理 | PFN 2014、2019 和 2026 年官方发布 | 商业部署规模未公开 |
| FANUC 工业机器人 | AI 机器人功能和工业自动化合作 | 工厂自动化和机器人智能 | 官方资本联盟来源 | 当前联合路线图未在公开英文来源中详述 |
| Kachaka / Preferred Robotics | 自主移动机器人加图像识别优化 | 家庭 / 办公室 / 物流机器人移动和感知 | Kachaka 网站加 PFN MN-Core 工作负载声明 | 销量和盈利能力未公开 |
| Matlantis / PFCC / ENEOS | PFP 神经网络势能和云端原子级仿真 | 材料发现和化学仿真 | PFN 发布、Matlantis 网站、Nature 论文 | ARR、留存和企业渗透未披露 |
| KDDI GPU Cloud 语境 | GPUaaS,而不是 MN-Core | 通用企业 AI 训练和开发基础设施 | KDDI 服务页确认 GPUaaS 可用 | 2024 年 KDDI 投资是否由 PFN 托管,公开资料未验证 |
| PLaMo API / Chat / Hugging Face 分发渠道 | 日语 LLM 和开放模型 | 企业生成、翻译和开发者实验 | PFN PLaMo 页面和 Hugging Face 主页 | 安全、红队和数据治理文档缺失 |
映射包含已确认关系和一个明确未确认的 KDDI 尽调项;不应把它读成完整客户名单。
[CE027, CE028, CE029, CE030, CE031, CE032]PFN 的差异化产品依赖半导体供应、算力运营、模型治理和关联应用渠道。
公开来源披露了第一代 TSMC 代工,但未披露后续芯片的完整供应链或控制证据。
[CE004, CE011, CE015, CE034, CE038, CE041]5.5 IP、研究速度、信任与尽调缺口
技术护城河是一组研究、开源、硅片 know-how 和领域专属数据,而不是公开来源中可见的单一专利墙。PFN 已发布或维护 Chainer、CuPy、Optuna 和 pfio;它用已发布的能效结果打造自研硅片;运营日语基础模型线;并把材料仿真研究转化成 Matlantis。这种广度稀缺,但也让尽调更复杂。公开来源还没有回答几个部署关键问题:MN-Core 2/L1000 的准确制程和供应承诺、PFCP 服务级架构、模型安全控制、出口管制姿态、数据驻留和客户安全认证。反面案例不是 PFN 缺技术,而是其技术可能在关联或日本特定场景中最强;与此同时,全球企业 AI 围绕 NVIDIA 基础设施、hyperscaler 云和拥有更大开发者基础的开源模型生态标准化。[CE021, CE022, CE023, CE036, CE040, CE041]
5.6 图表
06客户
6.1 具名客户组合
Preferred Networks 拥有很深的具名客户和战略伙伴组合,但这不是传统 SaaS 账户列表。最强模式是与日本大型工业既有企业共同创造,而这些企业同时也是投资者:Toyota Motor、FANUC、Hitachi、ENEOS、Chugai Pharmaceutical、Mitsui & Co.、NTT、Hakuhodo DY、Mizuho Bank、Mitsubishi Corporation 和 Mitsubishi Heavy Industries 都出现在抓取的官方或伙伴来源中。Toyota 是最清晰的汽车锚点,2017 年追加投资与自动驾驶 AI 相关,2026 年 Frontier Research Center 合作则使用 MN-Core L 处理器推进 physical AI。FANUC 是最深的工业机器人关系,从 2015 年研发和资本联盟开始,延伸到产品化 AI 功能,以及 FANUC-Hitachi-PFN Intelligent Edge System JV。组合可信,但客户状态各不相同:有些名称是生产 / 产品证明,有些是研发伙伴,有些是战略投资者。[CU001, CU002, CU003, CU004, CU005, CU006]
| 名称 | 分部 | 关系证据 | 阶段 | 关键限制 |
|---|---|---|---|---|
| Toyota Motor | 汽车 / physical AI | 2017 年投资;2026 年 FRC 联合研究 | 战略 R&D 伙伴 | 没有公开收入或生产合同金额 |
| FANUC | 工厂自动化 / 机器人 | 2015 年 R&D + 资本联盟;AI 功能;JV | 产品化伙伴 | 收入贡献未披露 |
| Hitachi | 工业 / 社会基础设施 | 2018 年与 FANUC 和 PFN 成立 Intelligent Edge System JV | JV 伙伴 / 投资人 | JV 经济性未披露 |
| ENEOS / Matlantis | 材料模拟 | PFP 共同开发;Matlantis 发布;v7 版本发布 | 商业产品 / JV | 客户数和 ARR 未披露 |
| Chugai Pharmaceutical | 药物发现 | 全面合作 + 投资 | 战略药企伙伴 | 未披露药物发现收入 |
| NTT 集团 | 算力基础设施 | NTT Com/NTTPC 案例研究与超算支持 | 基础设施供应商 / 合作伙伴 | 供应商支出与 PFN 收入边界不清 |
| JR East | 铁路维护机器人 | 2026 年自主轨道巡检机器人公告 | 试点 / 部署伙伴 | Preferred Robotics 子公司,而非 PFN 母公司直接关系 |
| SoftBank / KDDI | GPU 云生态 | 2026 年 SoftBank GPU 云;KDDI GPU Cloud 服务 | 基础设施生态 | 面向 PFN 的商业条款未公开 |
| Hakuhodo DY | 广告 / 创意 AI | 资本联盟;PaintsChainer 漫画产品 | 投资方 / 产品伙伴 | 历史创意场景已验证,当前收入不明 |
| MHI / Mitsubishi Corp. | 关键任务工业 AI | 2026 年 MHI 联盟;2024 年 Mitsubishi Corp. 资本 / 业务联盟 | 战略伙伴 | 时间太短,尚不足以证明留存 |
| Oisix ra daichi 农业伙伴 | 食品 / 农业 | 已获取 Oisix 官方页面,但没有 PFN 侧佐证 | 未验证 | 需管理层提供证据 |
行项目反映截至 2026-06-14 保留的公开来源;null / 未披露单元格表示未找到公开指标,并不表示不存在关系。
[CU001, CU002, CU003, CU008, CU011, CU014]PFN 往往从联合研究或资本联盟切入,再把其中一部分转化为产品、JV 或基础设施服务。
[CU030, CU033, CU038, CU039]6.2 分部与使用场景
对一家私营 AI 公司而言,PFN 的分部组合异常广。汽车和 physical AI 围绕 Toyota 的机器人与自动驾驶研究议程。工厂自动化围绕 FANUC 机床、机器人和 ROBO-MACHINE 功能。工业边缘和社会基础设施围绕 FANUC-Hitachi JV 以及 2026 年 MHI 关键任务 AI 联盟。材料仿真由 ENEOS 和 Matlantis 锚定,PFP 模型开发已经变成商业模拟器。医疗包括 Chugai 药物发现工作,以及 Mitsui 支持的 Preferred Medicine 癌症检测研究。通信基础设施由 NTT 数据中心 / GPU / 超算支持体现,KDDI 和 SoftBank 则显示较新的 GPU-cloud 生态路线。机器人包括 Preferred Robotics 的 JR East 铁路巡检工作,以及 Kachaka/Kachaka Pro 产品。广告和创意 AI 由 Hakuhodo DY 投资和 PaintsChainer 商业化体现。[CU011, CU012, CU013, CU014, CU015, CU016]
| 细分市场 | 代表客户 | 买方 / 用户 / 付款方 | 用例 | 战略价值 | 缺口 |
|---|---|---|---|---|---|
| 汽车 / 物理 AI | Toyota | OEM 研发 / 机器人研究人员 | 物理 AI 推理、自动驾驶 AI | 战略锚点和投资方 | 量产部署经济性未知 |
| 工厂自动化 | FANUC | 机器人 / 机床 OEM | 面向 FA/ROBOT/ROBO-MACHINE 的 AI 功能 | 产品化证据 | 终端客户采用情况未披露 |
| 材料模拟 | ENEOS / Matlantis | 材料研发团队 | 原子级模拟器与 PFP 模型 | 商业产品剥离 | 缺少 ARR / 客户数 |
| 医疗健康 / 制药 | Chugai; Preferred Medicine/Mitsui | 药企研发 / 诊断研究人员 | 药物发现;早期癌症检测 | 高价值受监管领域 | 临床商业化不明 |
| 算力基础设施 | NTT; KDDI; SoftBank | AI 基础设施买方 / 供应方 | GPU 云、数据中心、超算支持 | 支撑 PFN AI 栈扩展 | 供应商与客户角色因项目而异 |
| 机器人 / 基础设施 | JR East;Kachaka 用户 | 铁路运营商 / 设施用户 | 轨道巡检;AMR 运输 | 直接机器人部署路径 | 子公司经济性未单独拆分 |
| 广告 / 创意 | Hakuhodo DY; Hakusensha | 广告主 / 出版商生态 | 漫画上色 / 生成式创意 AI | 非工业用例广度 | 历史证据,并非当前收入证明 |
| 关键任务工业 AI | MHI; Mitsubishi Corp. | 工业主承包商 / 基础设施业主 | 面向关键应用的日本制造 AI | 2026 年扩张方向 | 时间太近,无法证明留存 |
行项目反映截至 2026-06-14 保留的公开来源;null / 未披露单元格表示未找到公开指标,并不表示不存在关系。
[CU003, CU009, CU011, CU014, CU016, CU018]成熟度分数反映各细分领域的公开验证质量、生产清晰度和本年度新鲜度。
分数是来自公开证据质量的序数尽调估算,不是 PFN 报告指标。
[CU003, CU011, CU014, CU026, CU038, CU039]6.3 牵引证据与采用路径
公开采用路径最好理解为一个漏斗:资助型联合研究、资本 / 业务联盟、产品化功能、合资公司,然后是商业产品或基础设施服务。FANUC 和 ENEOS 是这条路径的最佳证据。FANUC 从 2015 年研发和投资推进到 2018–2019 年发布 AI 功能和 Intelligent Edge System JV。ENEOS 从 PFP 共同开发推进到 Matlantis,这是一个专门的模拟器业务,已扩展到美国并在 2024 年发布 version-7。Toyota 2026 年 FRC 合作具有战略重要性,但仍处研究阶段。JR East/Preferred Robotics 是铁路维护机器人中的试点 / 部署证明,SoftBank 和 KDDI 则是生态基础设施路径,而非直接 PFN 客户合同。当前年份最强信号是 Toyota FRC、JR East、SoftBank GPU Cloud 和 MHI,均在 2026 年活跃。[CU030, CU033, CU037, CU038, CU039, CU040]
| 关系 | 最早证据 | 最新证据 | 状态 | 证据质量 |
|---|---|---|---|---|
| Toyota | 2017 年追加投资 | 2026 年 FRC 联合研究 | 战略研究伙伴 | 高:PFN + 独立 Toyota 投资来源 |
| FANUC | 2015 年研发 / 资本联盟 | 2019 年 AI 功能发布 | 产品化伙伴 | 高:多份 PFN 发布 + JV 报道 |
| FANUC/Hitachi JV | 2018 年 JV 协议 | 2018 年行业报道 | JV / 工业边缘 | 高:PFN + ACN + ARC |
| ENEOS / Matlantis | PFCC/Matlantis 发布 | 2024 年 PFP v7 | 商业模拟器业务 | 高:PFN + ENEOS + Business Wire |
| Chugai | 2018 年全面协议 | 2018 年投资 | 战略药物发现伙伴 | 高:Chugai + PFN |
| NTT 集团 | 2017 年超算发布 | 现有案例 / 使用页面 | 基础设施案例 | 高:NTT 官方页面 |
| JR East / Preferred Robotics | 2026 年公告 | 2026 年 PR Times | 机器人试点 / 部署 | 高:JR East PDF + PR Times |
| MHI | 2026 年联盟 | 2026 年公告 | 新战略联盟 | 高但时间短:MHI 官方 |
行项目反映截至 2026-06-14 保留的公开来源;null / 未披露单元格表示未找到公开指标,并不表示不存在关系。
[CU004, CU008, CU011, CU014, CU016, CU020]| 证据项 | 金额 / 规模 | 来源解读 | 收入相关性 | 限制 |
|---|---|---|---|---|
| Toyota 追加投资 | 10.5 billion yen | 面向移动 AI 研发的战略投资 | 验证战略重要性 | 不是 PFN 客户收入 |
| FANUC 资本联盟 | 900 million yen | 研发联盟之后的战略投资 | 验证工厂自动化投入 | 不是经常性收入 |
| 2017 年战略融资 | 超过 2 billion yen | FANUC、Hakuhodo、Hitachi、Mizuho、Mitsui 等资本方 | 多家既有巨头背书 | 投资方组合,不是客户支出 |
| Chugai 投资 | 约 700 million yen | 2018 年融资的一部分 | 药企伙伴投入 | 不是药物发现收入 |
| GENIAC | 政府支持项目入选 | METI/NEDO 基础模型项目 | 非稀释性 / 研发支持信号 | 本处未量化合同 / 补贴经济性 |
| MHI 2026 年联盟 | 未披露 | 面向关键应用的联合开发 | 潜在新增企业收入 | 尚无交易规模或部署 |
行项目反映截至 2026-06-14 保留的公开来源;null / 未披露单元格表示未找到公开指标,并不表示不存在关系。
[CU004, CU007, CU010, CU015, CU022, CU023]公开牵引力 KPI 强调关系广度和新鲜度,而不是收入指标。
计数基于留存公开来源和本章具名行。
[CU032, CU036, CU037]截至 2026 年运行日期,PFN 主要客户、合作伙伴和商业化验证点的时间线。
[CU004, CU006, CU008, CU014, CU020, CU024]6.4 留存、扩张与客户持久性
留存通过关系扩展可见,而不是通过披露的 NRR 可见。FANUC 给出最强的纵向证据:2015 年研发联盟、2015 年资本联盟、2018 年 JV,以及 2018–2019 年 AI 功能发布。ENEOS 也显示持久性,因为关系嵌入 Matlantis 和 PFP 发布,而不是单次公告。NTT 的证据偏供应商 / 客户案例,看起来对计算基础设施具备持久性,但不能证明经常性软件收入。Chugai 和 Mitsui/Preferred Medicine 展示了可信医疗合作;不过公开来源尚未证明规模化经常性临床或药物发现收入。Toyota 有大量战略证明,但当前 2026 年工作仍是联合研究。跨所有关系看,PFN 的客户成功指标是里程碑转化——从研究到产品或 JV——而不是传统留存 cohort,因此尽调应按具名账户索取 cohort 收入、续约率和付费生产状态。[CU016, CU017, CU030, CU032, CU035, CU038]
| 指标 | 数值或状态 | 细分市场 | 置信度 | 尽调要求 |
|---|---|---|---|---|
| FANUC 关系持续时间 | 2015-2019+ 多步扩张 | 工厂自动化 | 高 | 要求披露 FANUC 相关产品的付费年收入 |
| ENEOS 关系持续时间 | Matlantis/PFP 产品发布延续至 2024 年 | 材料模拟 | 高 | 要求披露 Matlantis ARR、续约和客户标识 |
| Toyota 关系持续时间 | 从 2017 年投资延续到 2026 年 FRC 研究 | 汽车 | 高 | 确认是否有 Toyota 部署进入付费生产 |
| NTT 关系类型 | 基础设施支持和案例研究使用 | 算力基础设施 | 高 | 将供应商支出与转售 / 分销收入拆开 |
| Chugai/Mitsui 医疗健康 | 有合作关系和研究产出,但没有规模化收入证据 | 医疗健康 | 中 | 要求披露临床里程碑和授权经济性 |
| NRR / GRR | 未公开披露 | 全部 | 低 | 要求披露过去三个财年的分组留存指标 |
| 流失 / 取消 | 保留来源中未发现公开客户流失信息 | 全部 | 中 | 向管理层询问丢失的试点、流失客户和未续约情况 |
行项目反映截至 2026-06-14 保留的公开来源;null / 未披露单元格表示未找到公开指标,并不表示不存在关系。
[CU032, CU035, CU038, CU039, CU040, CU041]6.5 集中度与验证风险
主要客户尽调风险不是缺少名称,而是经济性含糊。PFN 拥有许多顶级 logo,但公开来源很少区分投资者、伙伴、供应商、研究合作者和付费客户角色。Toyota、FANUC、ENEOS、Chugai、Mitsui、NTT、Hitachi、Hakuhodo DY、Mizuho 和 Mitsubishi Corporation 都验证了战略入口,但没有任何一方披露对 PFN 的收入贡献。CNBC 引用的三到五年商业化周期强化了这一风险:一些亮眼合作可能是长周期研发,而非近期经常性收入。地域集中度也很重要:大部分证明以日本为中心,Matlantis 美国发布是最清晰的国际商业化信号。最后,请求核实的 Oisix/CraftyFarm 关系无法从已抓取公开来源中印证;除非管理层提供一级证据,否则应作为未解决尽调项处理。[CU031, CU032, CU033, CU034, CU035, CU041]
| 风险 | 当前证据 | 潜在影响 | 缓释 / 尽调路径 |
|---|---|---|---|
| Toyota/FANUC 依赖 | 点名关系最成熟,战略历史也最长 | 若收入依赖少数工业锚点,影响为高 | 要求按客户披露前 10 大客户收入和管线 |
| 投资方与客户边界模糊 | 许多标识同时是投资方和合作伙伴 | 可能夸大付费客户牵引力 | 按付费生产、付费试点、供应商、投资方给每个标识分类 |
| 商业化周期长 | CNBC 引述从联合研究到实际推出需要 3-5 年 | 亮眼试点转收入可能滞后 | 询问试点到生产的转化率 |
| 组合以日本为中心 | 大多数证据来自日本既有企业 | 地理集中和采购风险敞口 | 要求披露国际收入和管线 |
| Oisix/CraftyFarm | 保留公开来源中没有佐证 | 若作为已验证标识呈现,可能虚增背书 | 要求管理层提供合同 / 试点来源,否则移除 |
| 流失 / NRR 未披露 | 没有公开留存指标 | 外部无法量化留存质量 | 要求披露 NRR、GRR、流失试点和客户背调 |
行项目反映截至 2026-06-14 保留的公开来源;null / 未披露单元格表示未找到公开指标,并不表示不存在关系。
[CU031, CU032, CU033, CU034, CU035, CU042]07风险
7.1 商业化与客户集中风险
Preferred Networks 仍是一家上行高、执行风险也高的公司,因为其既定战略横跨半导体、计算基础设施、基础模型、机器人和垂直应用。这种广度带来商业化挑战:PFN 必须把研究级技术转成可重复产品,同时继续为 Toyota、FANUC 和其他工业客户维护伙伴专属解决方案。历史上的 Chainer 向 PyTorch 迁移,是 PFN 自创平台失去独立战略重要性的最清晰公开案例;管理层自己也表示,深度学习框架作为竞争边缘的时代已经成熟。客户集中度也与普通企业集中度结构不同,因为 Toyota 和 FANUC 既是战略合作者,也是生态守门人。Toyota 内部 Woven by Toyota 能力形成替代风险,而 FANUC 专属 FIELD 集成同时带来锁定和依赖风险。公开记录没有披露收入结构、合同最低额、积压订单、总留存,或 Toyota/FANUC 在 2026 年是否仍是重要收入来源,因此在私有尽调完成前,合适的风险姿态是高严重性、中等可能性。[CR001, CR002, CR007, CR008, CR009, CR010]
| 类别 | 风险 | 严重性 | 可能性 | 缓释措施 | 状态 | 证据 |
|---|---|---|---|---|---|---|
| 商业化 | 偏研究的垂直整合未能转化为可复制产品 | 严重 | 中 | 要求提供产品线 P&L、客户试点和付费转化里程碑 | 未关闭;没有公开收入结构 | SR001, SR024; CR001, CR031 |
| 集中度 | Toyota 和 FANUC 仍是战略依赖方或路线图守门人 | 严重 | 中 | 分散已披露客户,并要求非排他路线图治理 | 未关闭;伙伴条款未公开 | SR004, SR005, SR006; CR009-CR012 |
| 竞争 | NVIDIA/CUDA 和超大规模云厂商芯片挤压 MN-Core 采用 | 严重 | 高 | 证明特定工作负载的 TCO、编译器成熟度和生态支持 | 活跃市场威胁 | SR010-SR012, SR034; CR015-CR017 |
| 地缘政治 | 出口管制限制 AI 芯片供应链、客户群或制造伙伴 | 高 | 中 | 维护分类矩阵、最终用途控制和许可法律顾问 | 活跃监管制度 | SR014-SR016; CR019, CR038 |
| 融资 | AI 泡沫回调导致降估值融资或 IPO 延后 | 高 | 中 | 延长现金续航,披露单位经济性,并按里程碑分阶段融资 | 未关闭;财务数据未公开 | SR024, SR031; CR027, CR030 |
| 人才 | AI / 半导体 / 机器人人才稀缺拖慢执行 | 高 | 中 | 留任授予、继任计划、全球招聘、大学管线 | 未解决;流失情况未披露 | SR022; CR024, CR040 |
| 监管安全 | 机器人安全或 AI Act 义务拖慢部署 | 高 | 中低 | 安全论证、ISO 映射、EU AI Act 角色分析 | 未解决;认证情况未披露 | SR017-SR020; CR021, CR022 |
| 知识产权 | 专利权属、开源或合作方 IP 冲突浮现 | 中 | 中 | 专利 FTO、转让审计、开源合规审查 | 未解决;未发现公开诉讼 | SR021, SR003; CR023, CR045 |
| 宏观汇率 | 日元波动扭曲美元估值和进口算力成本 | 中 | 中 | 汇率对冲和货币标准化 KPI 披露 | 未解决;宏观环境波动 | SR023, SR035; CR026 |
| 治理 | Nishikawa / Okanohara 相关的关键人物或治理弱点 | 高 | 中低 | 继任计划、关键人物保险、董事会控制 | 未解决;未发现离职 | SR001, SR033; CR025, CR032 |
| 声誉 | 估值跑在产品验证前面,市场质疑升温 | 高 | 中 | 发布客户牵引和量产案例 | 未解决;分析师信号偏负面 | SR009, SR024, SR031; CR043 |
| MN-Core 交易 | 传闻中的芯片业务收缩或出售仍未验证 | 中 | unknown | 索取公司交易文件以及 Sakura / PFN 确认 | 公开证据未解决 | CR033 |
各行按严重性和投资影响排序;可能性反映截至 2026-06-14 的公开证据,而不是公司内部风险评分。
[CR001, CR009, CR015, CR019, CR024, CR027]| 交易对手 | 角色 | 风险动态 | 严重性 | 缓释 | 证据 |
|---|---|---|---|---|---|
| Toyota | 投资方、合作方、潜在客户 | 对路线图的影响力叠加 Woven 内部能力,可能压低独立需求 | 极高 | 非排他协议、独立治理和客户多元化 | SR005-SR007 |
| FANUC | 工业合作方和工厂自动化渠道 | FIELD 集成可能让 PFN 依赖 FANUC 的战略优先级 | 高 | 扩大工厂客户群,并证明产品模块可迁移 | SR004 |
| Kobe University | MN-Core 共同开发方 | 学术合作可能让 IP 和路线图控制变复杂 | 中 | 审计 MN-Core 专利和 know-how 的转让与授权 | SR002 |
| Facebook / PyTorch 社区 | 框架生态依赖 | Chainer 转为维护后,PFN 依赖外部 PyTorch 路线图 | 中 | 开源贡献策略和内部 fork 政策 | SR003 |
| 公共部门出口监管机构 | 市场准入守门人 | 许可和最终用途限制可能挡住客户或组件 | 高 | 合规体系和外部律师审计 | SR014-SR016 |
客户和投资方角色来自公开合作与融资公告;合同经济性未披露。
[CR007, CR009, CR010, CR011, CR012, CR019]商业化、集中度和竞争替代集中在矩阵的高严重性半区。
x=严重性,y=可能性,采用 1–5 序数尺度,来自公开证据综合。
[CR041, CR042]PFN 的负面时间线主要由转向和市场压力主导,并非公开丑闻。
日期反映已抓取公开来源的发布日期或公告日期。
[CR007, CR011, CR014, CR019, CR027, CR030]7.2 竞争与技术替代风险
PFN 的 MN-Core 策略面对一组异常强的竞争对手。NVIDIA 不只是芯片供应商;它把先进加速器、CUDA、库、开发者心智和采购生态绑在一起,切换成本很高。CSIS 对 CUDA 生态效应的讨论说明了一个问题:即便利基加速器技术效率不错,商业化仍可能卡住,因为客户必须把软件、工具和运营经验从既有技术栈迁出来。超大云厂商自研芯片进一步加压。AWS Trainium 和 Google TPU 不是孤立出售的芯片,而是嵌在云采购、支持、定价和模型工作流里。开放基础模型也压缩了芯片之上软件层的差异化空间。PFN 自有 MN-Core 页面提供了真实产品和价格证据,但公开资料没有证明它有广泛第三方部署、单位经济、利用率,或可与 CUDA、Trainium、TPU 相比的开发者生态。因此,竞争替代是前三大投资论点风险之一。[CR004, CR005, CR006, CR013, CR014, CR015]
| 情景 | 竞争者路径 | 机制 | 可能性 | 严重性 | 所需缓释证据 |
|---|---|---|---|---|---|
| CUDA 锁定阻挡 MN-Core 采用 | NVIDIA | 开发者不愿把模型、内核和运维迁到更小的生态 | 高 | 极高 | 生产负载从 CUDA 迁到 MN-Core 的基准迁移证据 |
| 云厂商自研芯片赢下 AI 训练 / 推理 | AWS Trainium / Google TPU | 客户购买与云服务打包的加速器容量 | 高 | 高 | 与 Trainium / TPU 对比的 TCO 证据,并纳入支持和可用性 |
| 开放模型商品化基础层 | Meta Llama 和开源模型 | 模型差异化转向渠道、数据和成本 | 中 | 中 | 专有模型基准和客户付费意愿 |
| 日本主权 AI 采购停留在小众 | 国内 AI 芯片计划 | 政策支持未转化为全球规模 | 中 | 高 | 超出补贴或试点的已签多年规模合同 |
| MN-Core 仍只是 HPC 展示样板 | 内部 / 专用负载 | Green500 / SC 性能未带来广泛开发者采用 | 中 | 高 | 外部付费部署和利用率证据 |
情景描述可能的竞争路径,不代表已经观察到的客户流失。
[CR014, CR015, CR016, CR017, CR018, CR034]竞争和监管类别有最广泛的负面来源支持;财务不透明很关键,但证据来自私有领域。
计数代表映射到各类别的本地论点,不是统计意义上的事件频率。
[CR043, CR045]按主要风险类别展示缓释成熟度矩阵,既满足计划中的风险热力图展项,也保留象限评分图。
定性矩阵来自风险登记表和公开证据;不是内部控制评估。
[CR041, CR043, CR045]展示技术、监管和融资风险如何传导到收入质量和估值。
因果链由公开风险证据推断,未按概率加权。
[CR030, CR038, CR041, CR042]7.3 地缘政治、出口管制、安全和 AI 监管风险
PFN 的半导体雄心落在不断收紧的美国、日本和欧盟监管边界内。BIS 和 CSIS 资料显示,先进 AI 芯片、EDA 软件、半导体制造设备和高端算力供应链都是受监管的卡点。日本 METI 管制又叠加了本国出口管制考量;任何 PFN 硬件或软件若销售给涉华、涉军或受限终端用途,都可能触发许可、最终用途或客户筛查义务。实体机器人和移动应用还会带来工业机器人标准下的安全敞口,例如 ISO 10218;PFN 技术一旦从实验室进入工厂、物流、自动移动或服务机器人场景,风险更明显。欧盟 AI Act 则给投放或使用在欧洲的 PFN 系统再加一层风险。公开证据没有显示出口管制分类矩阵、AI Act 合规映射、安全事故记录,或第三方机器人安全认证。因此,剩余敞口为中到高严重度,尽调应聚焦法律意见、许可证、安全文件和客户最终用途控制。[CR019, CR020, CR021, CR022, CR023, CR038]
| 制度 | 司法辖区 | 暴露路径 | 严重性 | 证据 | 尽调要求 |
|---|---|---|---|---|---|
| BIS 先进计算和半导体管制 | 美国 / 域外适用 | AI 芯片、EDA、晶圆代工或客户最终用途筛查 | 高 | BIS 和 CSIS 出口管制资料 | 索取 ECCN 分类、许可证、最终用途筛查政策 |
| METI 半导体出口管制 | 日本 | 国内出口许可和受控工具 | 高 | METI 出口管制资料 | 索取 METI 律师备忘录和受限国家销售清单 |
| EU AI Act | 欧盟 | 在欧盟市场部署或投放的 AI 系统 | 中高 | AI Act 和 EUR-Lex 法规 | 将 PFN 角色映射为提供者、部署者、进口商,并确定风险类别 |
| ISO 10218 机器人安全 | 全球 / 客户合同 | 工业机器人和集成机器人系统部署 | 中高 | ISO 10218-1 和 10218-2 | 索取安全文件、合规评估、事故日志 |
| JPO / AI 专利审查 | 日本 | AI 发明可专利性、权属和 FTO | 中 | JPO AI 专利材料 | 索取专利转让和自由实施意见 |
监管暴露基于公开法律 / 监管资料;未发现针对 PFN 的具体执法行动。
[CR019, CR020, CR021, CR022, CR023, CR038]7.4 财务、融资、人才和宏观风险
PFN 是一家私有、重资本 AI 基础设施公司。公开市场数据来源确认了融资和独角兽定位,但没有披露收入、ARR、毛利率、经营亏损、现金消耗、跑道、硬件毛利率、资本开支需求或债务义务。定制半导体、编译器工作、基础模型和工业部署叠在一起,意味着成本底座很重;一旦 AI 基础设施情绪走弱,融资可能变难。Reuters 关于 AI 泡沫担忧的报道不是专门针对 PFN 的指控,但对一家公开财务披露有限的私有 AI 公司而言,直接关系到下一轮融资周期。日元走弱利弊都有:日元计价成本折成美元可能更有利,但进口算力、EDA、设备和云成本会上升,美元估值比较也会更复杂。人才风险同样重要,因为 PFN 需要 AI 研究员、编译器工程师、机器人专家和半导体产品人才的稀缺组合,而竞争对手是超大云厂商和国家级冠军企业。[CR024, CR025, CR026, CR027, CR028, CR029]
| 依赖项 | 风险 | 可能性 | 严重性 | 缓释 | 尽调路径 |
|---|---|---|---|---|---|
| Toru Nishikawa | CEO / 公开技术领袖离任,或客户触达能力下降 | 中低 | 高 | 继任计划和董事会关系图谱 | 索取关键人物保险和留任方案 |
| Daisuke Okanohara | 研究领导力和技术公信力集中 | 中低 | 高 | 扩大技术领导层梯队 | 索取组织架构图和关键岗位留任情况 |
| 半导体 / 编译器工程师 | 稀缺人才拖慢 MN-Core 软件成熟 | 中 | 高 | 大学人才管线和全球招聘 | 索取流失率、offer 接受率和薪酬基准 |
| 机器人和安全工程师 | 工业部署需要实体安全专业能力 | 中 | 中高 | 安全团队和认证流程 | 索取事故日志和安全论证负责人 |
| 未来投资者 | 若烧钱仍高,需要私人资本接续 | 中 | 高 | 里程碑融资和收入披露 | 索取现金 runway、burn 和下一轮计划 |
公开证据支持这些风险类别,但不支持员工流失或薪酬判断;这些仍是私下尽调事项。
[CR024, CR025, CR027, CR028, CR030, CR031]优先级最高的三项风险是产品化、集中度和竞争替代。
严重性 / 可能性来自风险登记表的序数判断。
[CR041, CR042]7.5 治理、声誉、知识产权和不利事件风险
截至运行日期,所审阅的公开记录给出了多项不利或反证信号,但没有确认的丑闻、执法行动、创始人离职、会计问题或裁员事件。最强的不利数据点并不戏剧化:Chainer 转入维护模式,独立分析师把 MN-Core 作为利基加速器审视,CB Insights 显示 Mosaic Score 下行,宏观和市场来源警示 AI 泡沫与日元风险。这个模式很关键,因为 PFN 的估值取决于一个信念:深厚技术资产能变成耐久产品。知识产权风险重要但大多潜伏:JPO 指引确认 AI 发明是活跃审查领域;PFN 与 Toyota、FANUC、Kobe University、Facebook/PyTorch 等伙伴长期合作,所有权、贡献权、专利许可和开源义务因此成为尽调重点。没有公开争议,不应被误读为治理风险低;这只说明决定性证据藏在非公开董事会纪要、客户合同、股权结构表、专利转让文件和管理层留任协议里。[CR023, CR027, CR028, CR030, CR032, CR033]
| 风险 | 可监控触发项 | 阈值 / 事件 | 行动含义 |
|---|---|---|---|
| 商业化失败 | 付费客户和收入披露 | 下一轮融资前仍没有重要的非 Toyota / FANUC 生产客户 | 重定价估值或暂停投资 |
| 集中度 / 替代 | Toyota 或 FANUC 范围变化 | 战略项目流失、不续约或转为内部自研 | 重新评估收入质量和战略独立性 |
| 竞争性芯片失败 | MN-Core 部署指标 | 相比 NVIDIA / Trainium / TPU 的 TCO 差,或没有外部规模订单 | 把 MN-Core 视为研发期权,而非核心估值支撑 |
| 出口管制问题 | 监管许可或客户筛查事件 | 许可证被拒、发现受限客户或合规违规 | 暂停芯片市场扩张论点 |
| 融资 / down round | 新融资 term sheet | 持平 / down round,或估值低于上一轮且结构苛刻 | 重设股权和下行情景 |
| 关键人物事件 | 创始人 / 研究负责人变化 | Nishikawa 或 Okanohara 离任且没有可信继任者 | 重新核保客户触达和产品路线图 |
否决标准是投资者尽调触发项,不是对实际事件的预测。
[CR025, CR030, CR031, CR038, CR041, CR042]7.6 附录
08估值
8.1 融资历史和隐含估值
Preferred Networks 的估值档案里有一个异常久远但仍重要的锚点:2017 年 8 月 Toyota 融资。PFN 自己的新闻稿确认 Toyota 投资约 ¥10.5 billion;独立报道则称该轮约 $95 million,并将其与数十亿美元隐含价值联系起来。由于 2024 年和 2025 年公告披露了资本规模,却没有披露投后估值,这仍是最后一个清晰的外部估值标记。2024 年 12 月首关很重要——SBI Group 领投股权,加上银行债务,合计 ¥19 billion——随后 2025 年 4 月又有 ¥5 billion,2025 年 6 月还有一笔未披露金额的扩展轮。这些融资在战略上是正面信号,因为它们在 Toyota 的历史支持基础上加入了 SBI、Development Bank of Japan、Mitsubishi Corporation 和 Wacom。但它们不是估值证明。公开证据支持一个保守结论:PFN 仍是独角兽质量资产,但 2024–2026 窗口内任何 $2.5–3.0 billion 估值说法都未获确认,应作为尽调假设,而不是事实。[CV001, CV002, CV003, CV004, CV005, CV006]
| 日期 | 轮次 / 事件 | 披露资本 | 披露估值 | 估值读数 | 主要限制 |
|---|---|---|---|---|---|
| 2017-08 | Toyota 投资 | ~¥10.5B / ~$95M | PFN 公告未披露;报道暗示数十亿美元估值 | 最后一笔明确外部锚点;常被引用在 ~$2B 附近 | 没有当前投后估值,且汇率随日期不同 |
| 2024-08 | SBI 资本 / 业务联盟 | 公告未披露金额 | 未披露 | 日本大型金融集团对芯片的战略验证 | 公告未披露轮次规模或投后估值 |
| 2024-12 | SBI Group 领投的首关,外加银行债务 | ¥19B 股权 / 债务合计 | 未披露 | MN-Core、PLaMo、云和产品的重大融资 | 债务和股权混合,遮蔽估值 |
| 2025-04 | 延展轮 | ¥5B 股权 | 未披露 | 2024 年首关后继续获得资本 | 未披露价格 / 股份或优先权条款 |
| 2025-06 | 追加延展轮 | 未披露股权 | 未披露 | 进一步延长资本 runway 的信号 | 未披露金额或投资者经济条款 |
列举并不完整,因为排除了更早的小轮融资和未披露的延展细节;披露资本不等于投后估值,不同来源日期的日元兑美元换算也未标准化。
[CV001, CV002, CV003, CV004, CV006, CV007]| 维度 | 评估 | 置信度 | 决策含义 |
|---|---|---|---|
| 建议 | 继续研究 | 高 | 没有 data room 证明时,不要按传闻中的 $2.5–3.0B 买入 |
| 估值立场 | 偏高 | 中 | 历史 ~$2B 锚点叠加无公开 ARR,让上行空间脆弱 |
| 风险评级 | 高 | 中 | 收入不透明、硬件毛利风险和潜在估值重置 |
| 目标回报数学 | $2.5–3.0B 入场价要做到 3x,需要 $7.5–9.0B 退出 | 中 | 入场价格纪律是 IC 的核心问题 |
| 退出路径 | IPO 前战略 M&A | 中 | 审计准备度和 TSE 流程限制近期 IPO 信心 |
建议只反映公开证据;经审计收入、ARR、毛利率和优先权结构条款都可能实质改变立场。
[CV036, CV037, CV039, CV041, CV042, CV043]PFN 披露的资本轨迹为正,但历史 Toyota 锚点之后,投后估值披露消失。
时间线包括已披露融资和战略事件;不意味着所有事件都是有定价的股权轮。
[CV001, CV003, CV004, CV006, CV007, CV013]8.2 可比公司倍数
可比组刻意拆开看,而不是混成一个头部倍数。NVIDIA 和 AMD 能提示 AI 芯片上行空间,但它们是有规模的上市半导体公司,具备供应链、毛利率和平台优势;PFN 尚未公开证明这些能力。Palantir、C3.ai 和 UiPath 可用于框定企业 AI 软件倍数,但它们也比 PFN 更纯软件,PFN 则混合了芯片、云、机器人和材料仿真服务。Fanuc、CYBERDYNE 和 SenseTime 给出反方向参照:当增长、利润率或监管敞口不达预期时,机器人和 AI 实施业务可以用更低倍数交易。私有可比公司把区间拉宽:OpenAI 和 Anthropic 显示前沿模型稀缺性可以有多贵;Cohere、Mistral、Figure、Wayve 和 Sakana AI 更接近,但仍不完美。Sakana AI 这个数据点尤其相关:据报道其估值为 $1.5 billion,这压缩了 PFN 的日本 AI 稀缺性溢价,也削弱了 PFN 理所当然应享有本土最高 AI 倍数的简单说法。[CV015, CV016, CV017, CV018, CV019, CV020]
| 公司 | 可比类别 | 证据来源 | 估值用途 | 关键限制 |
|---|---|---|---|---|
| NVIDIA | AI 芯片 | SEC 10-K + Yahoo Finance | AI 基础设施倍数上沿 | 已规模化的上市龙头,不是初创公司 |
| AMD | AI 芯片 | SEC 10-K + Yahoo Finance | 较低芯片倍数交叉检查 | 更宽的半导体组合 |
| Palantir | 企业 AI 软件 | SEC 10-K + Yahoo Finance | 高软件倍数参照 | 政府 / 数据平台经济性不同 |
| C3.ai | 企业 AI 软件 | Yahoo Finance | 较低企业 AI 软件参照 | 增长和盈利画像不同 |
| UiPath | 自动化软件 | Yahoo Finance | 自动化 / AI 工作流参照 | 纯软件且经历炒作退潮的倍数 |
| Fanuc | 工业机器人 | Yahoo Finance | 机器人估值底部参照 | 成熟工业自动化 |
| CYBERDYNE | 机器人 | Yahoo Finance | 日本机器人风险参照 | 规模小,公开市场波动大 |
| SenseTime | AI 软件 | Yahoo Finance | AI 落地 / 监管参照 | 中国市场治理和监管不同 |
倍数是市场数据快照,可能大幅波动;本表用于方向性对标,不是机械套用同业中位数。
[CV015, CV016, CV017, CV018, CV019, CV020]| 公司 | 披露估值 / 融资 | 类别 | 与 PFN 的相关性 | 局限 |
|---|---|---|---|---|
| OpenAI | 融资 $6.6B,对应估值 $157B | 前沿 AI | 显示前沿模型估值天花板 | 规模和生态远超 PFN |
| Anthropic | Series E 轮投后估值 $61.5B | 前沿 AI | 模型稀缺性的上限可比项 | 未横跨硬件 / 机器人 |
| Mistral | 融资 €600M | 基础模型 | 欧洲基础模型基准 | 估值来源可比性较弱 |
| Cohere | 估值 $5.5B | 企业 AI | 更接近的企业 AI 私有可比项 | 仍比 PFN 更偏纯软件 |
| Sakana AI | 披露估值 $1.5B | 日本 AI | 直接的日本 AI 稀缺性可比项 | 公司更年轻,产品重心不同 |
| Figure AI | 估值 $2.6B | 物理 AI 机器人 | 可比的物理 AI 融资规模 | 人形机器人赛道不同 |
| Wayve | >$1B Series C 轮 | 具身自主 | 显示战略资本对物理 AI 的兴趣 | 自动驾驶商业模式不同 |
| Covariant / Amazon | 战略 AI 机器人交易 | 机器人 M&A | 支撑战略退出路径 | 交易经济条款未充分披露 |
私有轮估值只是名义数字,证券条款、清算优先权和投资者权利各不相同;不应把它们当作干净的普通股可比项。
[CV021, CV022, CV023, CV024, CV025, CV026]方向性公开可比区间把成熟机器人、企业 AI 和前沿 AI 稀缺性分开。
示意性收入倍数等价值基于公开市场和私募轮可比性,不是经审计的 PFN 收入。
[CV015, CV016, CV017, CV018, CV019, CV020]PFN 处在日本战略 AI、物理机器人和 AI 芯片基础设施之间,并不能清晰映射到单一同业组。
X 轴代表 AI / 软件稀缺性;Y 轴代表物理世界 / 硬件敞口。评分为定性判断。
[CV017, CV021, CV022, CV025, CV026, CV027]8.3 情景和分部加总估值
最能自洽的方法是分部加总叠加情景分析。单一收入倍数会制造虚假精确度,因为 PFN 没有公开披露收入、ARR、分部利润率、客户集中度或单位经济。因此,模型分别给 AI 芯片、PLaMo 与企业 AI、机器人、材料 / 药物发现软件、云基础设施和战略期权价值设定估值区间。基准情景约为 $2.0–2.8 billion,大致符合高质量日本 AI 独角兽,但不足以保证在传闻区间高端进入后获得风险投资回报。熊市情景为 $1.0–1.6 billion,前提是 AI 估值重置压力、硬件利润率拖累和薄弱经常性收入占主导。牛市情景为 $4.0–6.0 billion,但需要证明 MN-Core 或 PLaMo 的表现更像稀缺 AI 基础设施,而不是定制化日本工业研发。若以 $2.5 billion 进入,3x 目标回报需要 $7.5 billion 退出;若以 $3.0 billion 进入,则需要 $9.0 billion,且还未计入稀释或清算优先权。[CV029, CV032, CV033, CV034, CV035, CV036]
| 分部 | 熊市 $M | 基准 $M | 牛市 $M | 理由 |
|---|---|---|---|---|
| AI 芯片 / MN-Core | 300 | 800 | 1800 | 战略资本瞄准 MN-Core,但芯片收入未公开 |
| PLaMo / 企业 AI | 250 | 650 | 1600 | 基础模型有上行空间,但未公开 ARR,需要折价 |
| 机器人 / 物理 AI | 200 | 450 | 900 | CNBC 机器人信号,以及 Figure / Wayve 可比项 |
| 材料 / 药物发现 / PFP | 150 | 300 | 600 | ENEOS / PFP 有验证,但商业化规模不清楚 |
| 云 / 基础设施 | 100 | 250 | 700 | 计算基础设施支撑内部和外部 AI 工作负载 |
| 战略期权溢价 | 0 | 350 | 400 | 日本主权 AI 稀缺性和产业联盟 |
数值是分析师估算,单位为百万美元;由于分部收入和利润率未披露,估算有意避免叠加所有前沿 AI 可比倍数。
[CV012, CV013, CV021, CV025, CV026, CV029]| 场景 | 估值区间 | 概率信号 | 以 $2.5B 进入的回报 | 论点触发条件 |
|---|---|---|---|---|
| 熊市 | $1.0–1.6B | 估值重置得到佐证;无 ARR;利润率弱 | 0.4–0.6x | 标记为回避,或等待 down round |
| 基准 | $2.0–2.8B | 有战略资本,但商业披露有限 | 0.8–1.1x | 仅跟踪 / 继续研究 |
| 牛市 | $4.0–6.0B | MN-Core / PLaMo 有收入,且高利润率经常性 AI 需求成立 | 1.6–2.4x | 若退出估值未超过高端情景,以 $2.5B 进入仍不到 3x |
| 风险投资目标 | $7.5B+ | 以 $2.5B 进入实现 3x 所需 | 3.0x+ | 需要公开市场或战略稀缺性溢价 |
场景区间并非公司指引;它们结合了公开可比项、私有可比项和定性概率信号,同时排除了未知优先权结构的影响。
[CV033, CV034, CV035, CV036, CV037, CV042]情景估值从熊市的 $1.0B 到牛市的 $6.0B;若按传闻入场价计算,风险投资回报目标需要高得多的退出估值。
单位为百万美元;目标回报线不计未来稀释和清算优先权。
[CV033, CV034, CV035, CV036, CV037]基准情景价值分布在芯片、PLaMo、机器人、材料、云和战略期权价值上。
基准情景单位为百万美元;分部价值是分析师基于公开证据和可比公司作出的估计。
[CV012, CV013, CV032, CV033]估值取决于未披露 KPI,而不只是 PFN 战略支持方的质量。
KPI 摘要混合已披露事实和模型输出;未披露指在已审阅公开来源中未找到。
[CV004, CV006, CV007, CV036, CV039, CV042]8.4 退出路径和流动性
PFN 有可行但并非临近的流动性路径。短期看,Tokyo Stock Exchange Growth Market IPO 比 Prime 更可信,因为 PFN 仍是一家私有、由风险资本支持的科技公司;但 JPX 指引说得很清楚,上市是经过审计的多步骤流程,不是靠公告驱动的事件。本文审阅的公开来源没有显示已审计财务报表、上市公司治理准备度或近期 IPO 申报。因此,战略并购可能是更可信的早期路径,尤其是在 Toyota、SBI、Mitsubishi 相关投资者和 MHI 都释放工业兴趣之后。潜在买家不会按通用 AI 实验室估值:他们承销的是芯片、云基础设施、机器人、材料仿真和日本主权 AI 定位。建议是在估值偏紧的前提下继续研究。只有尽调证明高毛利经常性收入、芯片需求、可信的 PLaMo 商业化和干净的优先权经济,才应上调;若所谓估值重置得到印证,或 ARR 与毛利率仍不披露,则下调为回避。[CV013, CV014, CV028, CV030, CV040, CV041]
| 退出路径 | 可能时间 | 估值支撑 | 风险 | 所需尽调证据 |
|---|---|---|---|---|
| TSE Growth IPO | 若审计就绪,2–4 年 | 日本 AI 稀缺性叠加战略背书 | 审计、治理和收入披露缺口 | 经审计报表、治理准备度、申报计划 |
| TSE Prime IPO | 更晚 / 概率更低 | 需要更大规模和流动性 | 公开市场标准更高 | 多年收入规模和盈利路径 |
| 战略 M&A | 若买方有主权 AI 动机,1–3 年 | Toyota / SBI / MHI / Mitsubishi 生态兴趣 | 买方可能因收入不透明而打折 | 分部收入、IP 所有权、客户管线 |
| 二级转让 | 任何融资窗口 | 可在未退出时提供流动性 | 因优先权和不透明度产生折价 | 409A、股权结构表、转让限制 |
退出时间是分析师估算;JPX 指引提供流程约束,但不代表 PFN 特定上市意图。
[CV013, CV014, CV028, CV030, CV040, CV041]| 要求 / 触发器 | 门槛 | 重要性 | 动作 |
|---|---|---|---|
| 经审计收入和 ARR | FY2023–FY2026,按分部 | 支撑收入倍数和 SOTP 校准 | 买入前必须拿到 |
| 分部毛利率 | 芯片、云、机器人、PLaMo、PFP | 区分硬件服务和软件经济性 | 若毛利率低于 40%,重新定价 |
| 股权结构表和优先权 | 全部优先股条款和债务义务 | 决定普通股回报瀑布 | 进入前先建模稀释 |
| MN-Core 管线 | 已签订单、ASP、毛利率 | 验证芯片分部上行空间 | 若偏弱,下调牛市情景 |
| PLaMo 商业化 | ARR、流失率和客户名单 | 验证基础模型倍数 | 没有 ARR 就不给溢价倍数 |
| 不利估值佐证 | 可靠的 down round 或 409A 证据 | 可确认估值过高风险 | 下调至回避 |
该清单并不完整,按价值权重排序;完整数据室还应加入客户合同、IP 所有权、安全合规和招聘计划。
[CV008, CV009, CV032, CV038, CV039, CV042]8.5 附录
免责声明
本报告汇总截至 2026-06-14 可获得的公开信息,仅用于尽调研究,不构成投资建议。Preferred Networks 的重大财务、治理和产品事实仍未公开;所有第三方估值、收入和客户数量估计都应视为指示性信息,待一手披露出现后可能修订。
证据索引
| 编号 | 陈述 | 可信度 | 来源 |
|---|---|---|---|
| CO001 | Preferred Networks, Inc. was established on March 26, 2014 and is located at Otemachi Building, 1-6-1 Otemachi, Chiyoda-ku, Tokyo. | 高 | SO001, SO031 |
| CO002 | PFN states its mission as “Make the real world computable and create the future together.” | 高 | SO001, SO002 |
| CO003 | PFN’s co-founders are Toru Nishikawa and Daisuke Okanohara. | 高 | SO001, SO002, SO031 |
| CO004 | As of the company page reviewed on 2026-06-14, Toru Nishikawa is Co-Founder and Chairman while Daisuke Okanohara is Co-Founder and Chief Executive Officer. | 中 | SO001 |
| CO005 | PFN lists directors including Hiroshi Maruyama, Kaname Masuda, Shinichi Koizumi and Hiroyuki Morikawa, with Maruyama chairing the Audit and Supervisory Committee. | 中 | SO001 |
| CO006 | PFN names Naoto Ono as COO, Yotaro Katayama as CFO and Masaaki Fukuda as VP of Engineering and Division President of Technology Planning. | 中 | SO001 |
| CO007 | PFN’s official business positioning is vertical integration across AI semiconductors, computing infrastructure, generative AI foundation models, solutions and applications. | 高 | SO001, SO004, SO013 |
| CO008 | PFN applies its technologies across manufacturing, materials and chemicals, life sciences, entertainment, retail and distribution, finance, public services, education and enterprise domains. | 高 | SO001, SO004, SO013 |
| CO009 | PFN’s MN-3 supercomputer, powered by MN-Core, topped the Green500 ranking three times in 2020 and 2021. | 高 | SO001, SO005, SO025, SO028 |
| CO010 | PFN’s official materials state that it has subsidiaries for materials discovery, robotics and foundation models. | 高 | SO013, SO018, SO032, SO033 |
| CO011 | Toyota invested 1.0 billion yen in PFN in December 2015 and agreed to invest an additional approximately 10.5 billion yen in August 2017. | 高 | SO008, SO029 |
| CO012 | FANUC announced a 900 million yen capital alliance with PFN in 2015, acquiring 6.0% of PFN’s issued stock. | 高 | SO009, SO003 |
| CO013 | PFN’s milestone page records December 2017 capital tie-ups with Hakuhodo DY Holdings, Mitsui & Co., Mizuho Bank and Hitachi and an additional investment from FANUC. | 高 | SO003, SO001 |
| CO014 | PFN’s milestone page records a June 2019 capital and business tie-up with JXTG Holdings, the predecessor context for ENEOS-related industrial AI activity. | 中 | SO003, SO012 |
| CO015 | SBI Holdings and PFN agreed in August 2024 to form a capital and business alliance for next-generation AI semiconductors, with SBI investing up to 10 billion yen. | 高 | SO011, SO021 |
| CO016 | PFN announced on December 23, 2024 that it raised 19 billion yen in the first close of its latest round, combining equity led by SBI Group with debt financing. | 高 | SO013, SO020 |
| CO017 | PFN announced on April 30, 2025 that an extension round added 5 billion yen and brought total amount raised in that round to date to 24 billion yen. | 高 | SO014, SO003 |
| CO018 | PFN’s December 2024 first-close investors included Development Bank of Japan, Mitsubishi Corporation, SBI Group, Sekisui House Investment Limited Partnership and Wacom. | 高 | SO013, SO020 |
| CO019 | PFN’s April 2025 extension-round investors included Kodansha, Mitsubishi UFJ Trust and Banking, Sekisui House, Sumitomo Mitsui Trust Bank, TBS Innovation Partners and Toei Animation, with Mizuho Bank as a lending institution. | 高 | SO014, SO003 |
| CO020 | CNBC described PFN in March 2025 as a Japanese AI unicorn using deep learning for real-world problems including trucking and robots. | 中 | SO019 |
| CO021 | The Bridge reported PFN’s December 2024 financing as a 19 billion yen AI-development-unicorn round including debt financing. | 高 | SO020, SO013 |
| CO022 | The reviewed official and investor-facing sources do not disclose PFN revenue, ARR, gross margin, net revenue retention or customer count. | 中 | SO001, SO013, SO014, SO030 |
| CO023 | PFN’s careers page was active as of 2026-06-14, but the reviewed official pages did not disclose total headcount. | 中 | SO007, SO001 |
| CO024 | CB Insights profiles PFN with products, competitors, financials, employees and headquarters, but its accessible text is a market-data profile rather than audited company disclosure. | 中 | SO030 |
| CO025 | PFN announced in 2019 that Chainer would move into maintenance phase as PFN migrated its deep learning research platform to PyTorch. | 中 | SO010 |
| CO026 | PFN and ENEOS announced the world’s first continuous AI-based autonomous operation of a crude oil processing unit at ENEOS Kawasaki Refinery. | 中 | SO012 |
| CO027 | PFN and Mitsubishi Heavy Industries formed a June 2026 business alliance to jointly develop Japan-made AI technologies for mission-critical applications. | 中 | SO015 |
| CO028 | PFN and Toyota’s Frontier Research Center started June 2026 joint research to accelerate physical AI using MN-Core L series processors. | 中 | SO016 |
| CO029 | PFN, GMO Internet Group and GMO Cybersecurity by Ierae announced a March 2026 joint venture, GMO Preferred Security, to deliver secure Japan-built AI environments. | 中 | SO017 |
| CO030 | PFN established Preferred Elements in November 2023 for development and sales of multimodal foundation models. | 中 | SO018 |
| CO031 | Mitsubishi Corporation and IIJ corroborate the December 2024 establishment plan for Preferred Computing Infrastructure, a joint venture for AI cloud computing using PFN technology. | 高 | SO022, SO023 |
| CO032 | Rapidus reported an agreement with PFN and SAKURA internet toward Japan-made green AI cloud infrastructure. | 中 | SO026 |
| CO033 | Matlantis is positioned as an AI simulator for predicting atomic-level phenomena and originated from PFN’s computational chemistry group company. | 高 | SO032, SO033 |
| CO034 | TOP500 lists MN-3 as a Preferred Networks MN-Core Server system using MN-Core and MN-Core DirectConnect. | 高 | SO024, SO005 |
| CO035 | Supermicro’s case study independently corroborates MN-3’s Green500 #1 achievement. | 高 | SO028, SO025 |
| CO036 | PFN’s shareholder list includes Chugai Pharmaceutical, Development Bank of Japan, ENEOS Innovation Partners, FANUC, Hakuhodo DY Holdings, Hitachi, Mitsubishi Corporation, Mitsubishi UFJ Trust and Banking, Kodansha, Mitsui & Co., Mizuho Bank, NTT, SBI Group, Sekisui House, Shin-Etsu Chemical, Sumitomo Mitsui Trust Bank, TBS, TEL Venture Capital, Toei Animation, Toyota Motor and Wacom. | 高 | SO001, SO003 |
| CO037 | PFN remains a private late-stage company in the reviewed sources; no IPO, S-1/prospectus, audited revenue filing or public listing was found. | 中 | SO001, SO019, SO030 |
| CO038 | PFN’s governance and strategy remain founder-centered because both co-founders occupy the chairman and CEO roles as of the reviewed company page. | 中 | SO001, SO002 |
| CO039 | PFN’s official milestones record consumer-service terminations for Crypko in June 2025 and Petalica Paint in July 2025, indicating selective pruning of non-core consumer products. | 中 | SO003 |
| CO040 | PFN’s AI Products and Solutions Division states ISO 27001 certification for development, commissioned work and provision of products and services. | 中 | SO001 |
| CO041 | The December 2024 and April 2025 financing announcements specify use of proceeds for talent acquisition, MN-Core processors, PLaMo, AI solutions and large-scale computing infrastructure. | 高 | SO013, SO014 |
| CO042 | PFN’s business-model evidence points to R&D partnerships, AI solutions, cloud/computing infrastructure and hardware-enabled strategic alliances rather than a single packaged SaaS line. | 中 | SO004, SO013, SO022, SO023 |
| CM001 | PFN presents itself as vertically integrated across AI chips, computing infrastructure, generative AI foundation models and applications. | 高 | SM001, SM002 |
| CM002 | PFN business co-creation examples include Fanuc for industrial robots, ENEOS for plant automation and Chugai for experiment automation. | 高 | SM002, SM003 |
| CM003 | PFN and Toyota Frontier Research Center started 2026 joint research to accelerate physical AI using MN-Core L-series processors. | 高 | SM004, SM003 |
| CM004 | PFN, IIJ and JAIST launched AImod full-scale operations in April 2026 using direct liquid-cooled high-density AI servers. | 高 | SM005, SM006 |
| CM005 | The AImod project is tied to NEDO-supported post-5G infrastructure R&D and Japan domestic AI compute capacity. | 高 | SM005, SM006, SM012 |
| CM006 | IFR reported 542,000 new industrial robots installed globally in 2024 and 4.664 million in operating stock. | 高 | SM007, SM008 |
| CM007 | IFR reported Japan installed 44,500 industrial robots in 2024 and had 450,500 in operational stock. | 高 | SM007, SM008 |
| CM008 | MarketsandMarkets valued the industrial robotics market at $15.5 billion in 2026 and forecast $20.8 billion by 2032. | 高 | SM019, SM007 |
| CM009 | Mordor valued the smart manufacturing market at $387.14 billion in 2026 with 13.53% CAGR to 2031. | 高 | SM020, SM032 |
| CM010 | Gartner forecast worldwide AI spending at roughly $2.5 trillion to $2.6 trillion in 2026. | 高 | SM009, SM010 |
| CM011 | Gartner listed 2026 AI software spending of about $452.5 billion and AI services spending of about $588.6 billion. | 高 | SM009, SM010 |
| CM012 | Fortune Business Insights projected the global AI market at $375.93 billion in 2026. | 高 | SM021, SM010 |
| CM013 | IDC forecast Japan AI infrastructure spending above $5.5 billion in 2026 after rapid 2022-2025 expansion. | 高 | SM011, SM012 |
| CM014 | Value Market Research projected Japan AI market growth from $19.83 billion in 2025 to $289.88 billion by 2034. | 中 | SM018 |
| CM015 | IMARC forecast Japan AIaaS growth from $1.2545 billion in 2025 to $15.0048 billion by 2034. | 中 | SM017 |
| CM016 | GMI valued the AI accelerator chips market at $154.6 billion in 2026. | 高 | SM013, SM009 |
| CM017 | The AI accelerator chip lens is broader than PFN’s realistic serviceable market because PFN competes with hyperscaler and NVIDIA-class ecosystems. | 中 | SM013, SM005, SM006 |
| CM018 | MarketsandMarkets projected autonomous driving software from $1.8 billion in 2024 to $7.0 billion by 2035. | 高 | SM023, SM025 |
| CM019 | Precedence Research projected autonomous driving software at $2.97 billion in 2026 and $8.04 billion by 2035. | 中 | SM025, SM023 |
| CM020 | Mordor estimated the autonomous-car market at $220.58 billion in 2026, a broader vehicle-level lens than PFN software. | 高 | SM024, SM023 |
| CM021 | Mordor projected agricultural robots at $18.0 billion in 2026 and $41.3 billion by 2031. | 高 | SM014, SM026 |
| CM022 | Public 2026 searches did not find a fresh PFN CraftyFarm announcement, making agriculture an option-value segment rather than a proven core revenue market. | 中 | SM003, SM014, SM026 |
| CM023 | Grand View projected AI drug discovery at $2.9 billion in 2026 and $13.8 billion by 2033. | 高 | SM015, SM016 |
| CM024 | Research and Markets valued AI in drug discovery at $2.93 billion in 2026 with 26.2% CAGR to 2030. | 高 | SM016, SM015 |
| CM025 | Chugai describes AI-leveraging drug discovery and MALEXA under its digital transformation program. | 高 | SM029, SM022 |
| CM026 | Fierce Biotech reported Chugai discontinued an AI-assisted antibody candidate, an adverse signal for drug-discovery conversion risk. | 中 | SM030, SM029 |
| CM027 | MHI and PFN formed a 2026 business alliance for Japan-made AI technologies in mission-critical applications. | 高 | SM027, SM028 |
| CM028 | The MHI alliance extends PFN’s industrial AI market beyond factory robots into social infrastructure autonomy. | 中 | SM027, SM028, SM020 |
| CM029 | Chugai, Fanuc, Toyota, MHI, IIJ and JAIST evidence show PFN reaches buyers through partner-led co-creation rather than a single horizontal SaaS motion. | 中 | SM002, SM004, SM005, SM027, SM029 |
| CM030 | Industrial robotics buyers are factory automation teams and robot OEMs, while PFN’s user is usually a partner R&D or automation group. | 中 | SM002, SM007, SM019 |
| CM031 | Automotive physical-AI buyers are OEM research centers and mobility engineering teams rather than consumer end users. | 中 | SM004, SM023, SM025 |
| CM032 | AI chip buyers are AI infrastructure operators and internal model teams; PFN’s proof point is AImod rather than merchant-chip share. | 中 | SM005, SM006, SM013 |
| CM033 | Drug discovery buyers are pharma R&D and platform teams, with Chugai evidence supporting experiment automation and computational chemistry adjacency. | 中 | SM002, SM022, SM029 |
| CM034 | Agriculture robotics buyers would be farm operators or equipment vendors, but PFN-specific commercialization evidence remains sparse. | 中 | SM003, SM014, SM026 |
| CM035 | PFN’s broad TAM is best expressed as multiple lenses rather than one blended number because each segment has different buyers and adoption constraints. | 中 | SM009, SM013, SM014, SM016, SM019, SM023 |
| CM036 | A defensible PFN SAM should include Japan industrial AI, physical-AI partnerships, AI infrastructure and selected vertical solutions, not the entire $2.5 trillion AI-spending pool. | 中 | SM009, SM010, SM011, SM020, SM027 |
| CM037 | Public data cannot support a precise PFN SOM because PFN does not disclose segment revenue, customer count, utilization or MN-Core external sales. | 低 | SM001, SM002, SM005 |
| CM038 | PFN’s market timing is strongest where 2026 partner announcements align with large markets: Toyota physical AI, MHI infrastructure AI and AImod compute. | 中 | SM004, SM005, SM027, SM009 |
| CM039 | Adoption constraints include long industrial qualification cycles, partner commercialization dependency, chip ecosystem barriers and regulated pharma validation risk. | 中 | SM013, SM023, SM030, SM027 |
| CM040 | Smart manufacturing and industrial robotics estimates conflict in scope: $387.14 billion smart manufacturing includes broad factory software and equipment, while $15.5 billion industrial robotics is robot-specific. | 中 | SM020, SM019 |
| CM041 | Autonomous-car market estimates overstate PFN’s addressable opportunity because vehicle hardware and fleet value are broader than perception or physical-AI software. | 中 | SM024, SM023, SM025 |
| CM042 | AI drug-discovery market estimates from Grand View and Research and Markets are tightly aligned around $2.9 billion for 2026. | 高 | SM015, SM016 |
| CM043 | Japan AI market estimates vary materially by scope, with infrastructure, AIaaS and all-AI market definitions producing different 2026 baselines. | 中 | SM011, SM017, SM018 |
| CM044 | Matlantis shows PFN can commercialize scientific AI as SaaS-like vertical software, although materials discovery is adjacent to this chapter’s core seven lenses. | 中 | SM031, SM002 |
| CP001 | Preferred Networks publicly positions its business across AI chips, deep-learning software, robotics, foundation models, drug discovery, and agriculture-related initiatives. | 高 | SP001, SP002, SP003 |
| CP002 | PFN frames MN-Core as a specialized AI chip line intended to improve training and inference speed, efficiency, cost, power use, and availability relative to general-purpose GPUs. | 中 | SP001 |
| CP003 | PLaMo-13B was released by PFN as an open-source large language model supporting Japanese and English. | 中 | SP003 |
| CP004 | NVIDIA H100, H200, and Blackwell create the highest-pressure accelerator comparison because NVIDIA offers successive datacenter GPUs with large-model training and inference positioning. | 高 | SP004, SP005, SP006 |
| CP005 | H100 public specifications include Hopper architecture, Transformer Engine, NVLink, HBM memory, and enterprise AI software positioning that make it a mature alternative to custom accelerators. | 中 | SP004 |
| CP006 | H200 and Blackwell extend NVIDIA competition beyond PFN chip hardware into a full roadmap and datacenter ecosystem that customers can standardize on. | 中 | SP005, SP006 |
| CP007 | AMD MI300, Intel Gaudi, and Google TPU are material accelerator substitutes because they target AI training or inference buyers through merchant or cloud infrastructure channels. | 中 | SP009, SP010, SP011 |
| CP008 | Cerebras, Graphcore, and SambaNova compete with PFN-like custom silicon narratives by emphasizing non-GPU AI architectures or vertically integrated AI platforms. | 中 | SP012, SP013, SP014 |
| CP009 | SemiAnalysis coverage of Google Gemini infrastructure underscores that hyperscaler TPU stacks can be strategically differentiated rather than commodity compute. | 中 | SP038 |
| CP010 | IEEE Spectrum coverage of Intel Gaudi 3 shows that Intel is explicitly challenging NVIDIA in the AI accelerator market. | 中 | SP039, SP010 |
| CP011 | NVIDIA robotics and Isaac-related product surfaces compete against PFN in robotics AI by bundling simulation, perception, and deployment tooling around NVIDIA hardware. | 中 | SP007 |
| CP012 | Boston Dynamics Spot is a credible industrial robotics alternative for mobile inspection, but it is a general robot platform rather than a PFN-style multi-vertical AI software and chip stack. | 中 | SP015 |
| CP013 | Skild AI, Physical Intelligence, Figure AI, and Sanctuary AI show that robotics foundation models and humanoid embodiments are attracting specialized full-stack AI robotics entrants. | 中 | SP016, SP017, SP018, SP019 |
| CP014 | Covariant remains a named robotics-AI competitor in warehouse automation and robot foundation models. | 中 | SP020 |
| CP015 | Amazon hired Covariant founders and about a quarter of Covariant employees while licensing Covariant models, an adverse signal that Big Tech can absorb robotics-AI talent without a full acquisition. | 中 | SP021, SP040 |
| CP016 | Waymo, Wayve, Mobileye, NVIDIA DRIVE, and Toyota Woven represent autonomous-driving AI alternatives to PFN automotive perception work. | 中 | SP008, SP022, SP023, SP024, SP025 |
| CP017 | Waymo competes as a deployed autonomous-vehicle operator, while Wayve competes through embodied-AI autonomous-driving software. | 中 | SP022, SP023 |
| CP018 | Mobileye SuperVision and NVIDIA DRIVE compete through vehicle-grade ADAS/autonomy stacks that can be bought or adopted by OEMs instead of custom PFN perception work. | 中 | SP008, SP024 |
| CP019 | Woven by Toyota is a direct internal-build threat in Japan because Toyota can develop software, autonomy, and mobility infrastructure in-house rather than buying PFN modules. | 中 | SP025 |
| CP020 | Sakana AI, rinna, ABEJA, and ELYZA form the Japan-focused AI competitor set most relevant to PLaMo mindshare and domestic enterprise AI budgets. | 中 | SP026, SP027, SP028, SP029 |
| CP021 | Sakana AI competes most directly with PLaMo on AI research visibility and Japanese foundation-model narrative rather than on PFN chips or robotics hardware. | 中 | SP026, SP003 |
| CP022 | ABEJA and ELYZA are more enterprise-AI and LLM deployment threats, while rinna adds a consumer and conversational-AI heritage in Japan. | 中 | SP027, SP028, SP029 |
| CP023 | Recursion, Isomorphic Labs, Insilico Medicine, BenevolentAI, and Schrödinger are the most visible AI-enabled drug-discovery competitors to PFN Bio. | 中 | SP030, SP031, SP032, SP033, SP034 |
| CP024 | Recursion and Isomorphic Labs appear more directly scaled around AI-first drug discovery platforms than PFN Bio based on their public company/product surfaces. | 中 | SP030, SP031, SP002 |
| CP025 | Insilico, BenevolentAI, and Schrödinger pressure PFN Bio through discovery platforms, pharma workflows, and computational chemistry tooling. | 中 | SP032, SP033, SP034 |
| CP026 | Plenty, FarmWise, and Carbon Robotics demonstrate that agricultural automation competition includes controlled-environment farming, AI computer-vision weeding, and laser-based field robotics. | 中 | SP035, SP036, SP037 |
| CP027 | CraftyFarm faces status-quo substitution from human farm labor, equipment dealers, and crop-specific automation because public evidence for PFN agriculture deployments is thinner than for global ag-robotics specialists. | 中 | SP002, SP035, SP036, SP037 |
| CP028 | Public pricing is opaque across most accelerator, robotics, Japanese AI, drug-discovery, and agriculture competitors, so packaging and ecosystem leverage are more observable than list prices. | 中 | SP004, SP009, SP015, SP030, SP036 |
| CP029 | NVIDIA, Google, Toyota, Amazon, and Mobileye have stronger distribution or ecosystem leverage than PFN in their respective chip, autonomous-driving, robotics, and OEM lanes. | 中 | SP006, SP011, SP021, SP024, SP025 |
| CP030 | PFN switching costs are strongest when customers adopt proprietary chips, deep-learning frameworks, or trained models as infrastructure, but weaker when buyers can substitute GPU cloud capacity or commodity robotics platforms. | 中 | SP001, SP004, SP011, SP015 |
| CP031 | PFN has breadth across more verticals than most competitors, but this breadth also exposes it to focused rivals with deeper ecosystems in each lane. | 中 | SP001, SP002, SP004, SP022, SP030, SP036 |
| CP032 | Unsupported capability cells in this chapter are marked unknown or partial because public pages rarely disclose benchmark-equivalent model quality, customer prices, or deployment metrics. | 中 | SP004, SP009, SP015, SP030 |
| CP033 | A capability matrix that scores PFN high on breadth but below NVIDIA on accelerator ecosystem and below Waymo/Mobileye on deployed autonomy is consistent with retained public evidence. | 中 | SP001, SP004, SP006, SP022, SP024 |
| CP034 | The principal adverse robotics-AI risk is that foundation-model capability becomes concentrated inside large platforms or well-funded specialists faster than PFN can monetize its own robotics perception stack. | 中 | SP007, SP016, SP019, SP021, SP040 |
| CP035 | The principal accelerator risk is that PFN must compete not only on chip performance but also against CUDA, cloud TPU availability, NVIDIA enterprise software, and hyperscaler procurement habits. | 中 | SP004, SP006, SP011, SP038 |
| CP036 | The principal Japanese-LLM risk is that PLaMo competes for attention and deployments against domestic AI companies with clearer pure-play enterprise AI positioning. | 中 | SP003, SP026, SP028, SP029 |
| CP037 | The principal drug-discovery risk is that PFN Bio may be outscaled by companies whose public brands and partner narratives are dedicated to AI drug discovery. | 中 | SP030, SP031, SP032, SP034 |
| CP038 | The principal agriculture risk is that CraftyFarm must prove crop-specific ROI against specialized ag-robotics companies with direct weeding or controlled-environment automation claims. | 中 | SP035, SP036, SP037 |
| CP039 | Internal build remains a serious substitute because automakers, manufacturers, pharma companies, and farms can assemble models, GPUs, software teams, and robotics partners without buying a PFN-branded stack. | 中 | SP004, SP011, SP025, SP030, SP036 |
| CP040 | PFN moat readiness is therefore highest in cross-domain research capability and Japan ecosystem credibility, and weakest where customers demand productized pricing, global cloud ecosystem depth, or vertical-specific deployment proof. | 中 | SP001, SP002, SP003, SP004, SP021, SP030 |
| CI001 | PFN announced a December 2024 first close totaling 19 billion yen, combining SBI-led equity financing with debt financing from Japanese financial institutions. | 高 | SI001, SI003 |
| CI002 | The December 2024 company-disclosed investors were Development Bank of Japan, Mitsubishi Corporation, SBI Group, Sekisui House Investment Limited Partnership, and Wacom. | 中 | SI001 |
| CI003 | The December 2024 company-disclosed lenders were MUFG Bank, Resona Bank, Shoko Chukin Bank, and Sumitomo Mitsui Banking Corporation. | 中 | SI001 |
| CI004 | PFN said the December 2024 proceeds would fund talent acquisition, MN-Core processor development and sales, PLaMo enhancement, AI solutions, and large-scale computing infrastructure. | 中 | SI001 |
| CI005 | PFN announced an April 2025 extension round of 5 billion yen through third-party share allotment and debt financing. | 中 | SI002 |
| CI006 | The April 2025 extension brought the December 2024 to April 2025 financing series to 24 billion yen. | 中 | SI002 |
| CI007 | The April 2025 extension investors included Kodansha, Mitsubishi UFJ Trust and Banking, Sekisui House Investment Limited Partnership, Sumitomo Mitsui Trust Bank, TBS Innovation Partners Fund III, and Toei Animation, with Mizuho Bank as lender. | 高 | SI002, SI024 |
| CI008 | Sumitomo Mitsui Trust Bank described its April 2025 PFN investment as an impact-equity investment supporting PFN's vertically integrated AI value chain. | 中 | SI024 |
| CI009 | The August 2024 SBI-PFN agreement contemplated SBI Group investing up to 10 billion yen through a third-party allocation by the end of September 2024. | 高 | SI006, SI007, SI005 |
| CI010 | Toyota agreed in August 2017 to invest approximately 10.5 billion yen in PFN through a third-party allocation of new shares. | 高 | SI008, SI009 |
| CI011 | FANUC and PFN announced a 2015 capital alliance under which FANUC would finance 900 million yen and acquire 6.0% of PFN's issued stock. | 高 | SI010, SI011 |
| CI012 | The Bridge reported that PFN's disclosed cumulative funding reached approximately 36 billion yen after the December 2024 first close. | 中 | SI004 |
| CI013 | Public market-data estimates for PFN total funding vary materially, including Growjo at $314 million and PremierAlts at $315.4 million. | 中 | SI014, SI015 |
| CI014 | The Bridge reported PFN's post-round valuation exceeded 300 billion yen, positioning it at the top of Japan's unicorn rankings. | 中 | SI003, SI004 |
| CI015 | Latka listed PFN at a $2 billion valuation in 2024 while estimating 2024 revenue at $42 million. | 中 | SI013 |
| CI016 | AI Market Watch described PFN as valued above 300 billion yen and estimated 2025-2026 headcount at roughly 280 to 340 employees. | 中 | SI016 |
| CI017 | PremierAlts listed a materially lower $1.0 billion PFN valuation as of June 30, 2025, creating a valuation conflict against the 300 billion yen and $2 billion narrative. | 中 | SI015 |
| CI018 | PFN is a private company and its official website and financing releases do not provide audited public revenue, ARR, gross-margin, cash-balance, or burn-rate disclosures. | 高 | SI001, SI002, SI027, SI028 |
| CI019 | Craft lists PFN as private and active with FY2023 revenue of 7.7 billion yen, but the page is an aggregator profile rather than a company financial statement. | 中 | SI012 |
| CI020 | Latka estimated PFN's 2024 revenue at $42 million and described that figure as revenue rather than company-disclosed audited ARR. | 中 | SI013 |
| CI021 | Growjo estimated PFN's annual revenue at $49.5 million and employee count at 275. | 中 | SI014 |
| CI022 | AI Market Watch cited historical PFN revenue of 8.486 billion yen for a fiscal year ending January 2021 and a 2025-2026 headcount range of 280 to 340 employees. | 中 | SI016 |
| CI023 | RocketReach gave a much lower 2026 annual-revenue figure of $15.3 million, underscoring that public PFN revenue estimates are inconsistent and should not be treated as audited data. | 中 | SI030 |
| CI024 | A reasonable public revenue range for PFN is roughly $42 million to $56 million, excluding the RocketReach low outlier and relying on Craft, Latka, Growjo, and AI Market Watch estimates. | 中 | SI012, SI013, SI014, SI016 |
| CI025 | At a $2 billion valuation and a $42 million to $56 million revenue-estimate range, PFN would trade at roughly 36x to 48x estimated revenue before any adjustment for cash, debt, or low-margin hardware and services mix. | 中 | SI013, SI016, SI003, SI004 |
| CI026 | At PremierAlts' $1.0 billion valuation and the same $42 million to $56 million revenue-estimate range, PFN would trade at roughly 18x to 24x estimated revenue. | 中 | SI015, SI013, SI016 |
| CI027 | PFN describes its business as vertically integrated across semiconductors, computing infrastructure, solutions, and applications, rather than as a pure software company. | 中 | SI027 |
| CI028 | PFN's official business page says it serves diverse industries through business co-creation and supports partners through commercialization of jointly developed technologies. | 中 | SI027 |
| CI029 | PFN's computing-infrastructure page states that since 2024 it has offered PFCP, a cloud-based service using PFN computing infrastructure. | 中 | SI018 |
| CI030 | PFN's AI-chips page says generative AI is pushing general-purpose GPUs to limits in performance, cost, power efficiency, and availability, explaining why proprietary AI chips are central to its strategy. | 中 | SI017 |
| CI031 | PFN announced that MN-Core 2 began operating in 2023 and that it planned to provide MN-Core 2 computing power to external parties in 2024. | 中 | SI019 |
| CI032 | ServeTheHome independently described MN-Core 2 as focused on HPC and AI cluster tasks and power-efficient compute. | 中 | SI020 |
| CI033 | PFN announced MN-Core L1000 as a generative-AI processor under development for 2026 commercialization, targeting up to a ten-fold speed increase versus conventional processors for inference. | 中 | SI021 |
| CI034 | Mitsubishi Corporation said its PFN investment supports a strategic AI alliance and promotion of PFN's MN-Core processor series. | 中 | SI022 |
| CI035 | Mitsubishi Corporation, PFN, and IIJ announced Preferred Computing Infrastructure, scheduled to begin operations in early 2026 to provide and support PFCP customers. | 中 | SI023 |
| CI036 | The PFCI joint venture shifts at least part of PFN's compute go-to-market and operating burden into a partner-backed cloud infrastructure vehicle rather than leaving all commercialization on PFN alone. | 中 | SI023, SI022 |
| CI037 | METI and NEDO selected 16 Cycle 4 GENIAC projects in June 2026 to receive computing-resource support for AI model development. | 中 | SI025 |
| CI038 | Government compute-resource programs may offset some AI model development cost for qualifying participants, but they are not equivalent to PFN cash revenue or unconstrained runway. | 中 | SI025 |
| CI039 | SoftBank's 2026 Telco AI Cloud announcement is relevant as an infrastructure comparable, but the retained source does not make SoftBank a direct PFN investor or customer. | 中 | SI026 |
| CI040 | Neither the December 2024 nor April 2025 retained PFN financing releases list ENEOS or Chugai Pharmaceutical among the named investors. | 高 | SI001, SI002 |
| CI041 | The retained official PFN and partner sources did not verify the claim that KDDI purchased 1,000 MN-Core chips or GPUs from PFN. | 高 | SI018, SI019, SI020 |
| CI042 | PFN's public sources do not disclose monthly burn, cash on hand, net debt, or runway months, so runway cannot be calculated from public evidence. | 高 | SI001, SI002, SI027 |
| CI043 | PFN's financial diligence should request audited or management-prepared P&L, revenue by line, gross margin by segment, cash balance, monthly burn, debt schedule, backlog, and customer concentration. | 低 | |
| CI044 | PFN's recent capital base supports near-term investment in chips, cloud, and PLaMo, but the same proceeds signal continuing capital intensity rather than proof of self-funding profitability. | 中 | SI001, SI002, SI017, SI021, SI023 |
| CE001 | PFN positions itself as a vertically integrated AI company spanning AI chips, computing infrastructure, generative AI, solutions and products. | 高 | SE001, SE002 |
| CE002 | PFN’s product architecture links proprietary accelerators, PFCP compute infrastructure, PLaMo foundation models, and applied solutions rather than a single SaaS product. | 高 | SE001, SE002, SE003, SE004 |
| CE003 | PFN began developing the MN-Core processor series with Kobe University in 2016. | 高 | SE002, SE012 |
| CE004 | The first-generation MN-Core was described in 2018 as a TSMC 12nm processor with 500W estimated power, 524 TFLOPS half-precision peak performance, and 1.0 TFLOPS/W estimated half-precision efficiency. | 高 | SE012, SE002 |
| CE005 | MN-3 was built around 160 MN-Core processors connected by a specialized interconnect and began operation in 2020. | 高 | SE002, SE013 |
| CE006 | TOP500 reported MN-3 as the most energy-efficient Green500 system in June 2020 at 21.1 gigaflops per watt. | 高 | SE028, SE013 |
| CE007 | TOP500 reported MN-3 as the No. 1 Green500 system in November 2021 at 39.38 gigaflops per watt. | 高 | SE029, SE016 |
| CE008 | PFN says MN-Core 2 provides FP64 12 TFLOPS, FP32 49 TFLOPS, TF32 98 TFLOPS, and TF16 393 TFLOPS per board. | 高 | SE002, SE017 |
| CE009 | PFN lists commercial MN-Core 2 products including an eight-board MN-Server 2 with 3.1 PFLOPS TF16 theoretical performance and a Japan-only Devkit package. | 高 | SE002, SE017 |
| CE010 | PFN says MN-Core 2 was accepted for presentation at Hot Chips 2024, a technical credibility signal for the chip architecture. | 高 | SE019, SE002 |
| CE011 | PFN began developing MN-Core L1000 in 2024 as a generative-AI inference processor using 3D-stacked memory and logic. | 高 | SE020, SE002 |
| CE012 | PFN claims MN-Core L1000 can deliver up to tenfold faster token processing than existing GPUs and processors, but this remains a company claim without independent benchmark publication. | 中 | SE020, SE002 |
| CE013 | PFN’s chips page claims first-generation MN-Core accelerated Kachaka image-recognition-model optimization sevenfold versus GPU. | 高 | SE002, SE035 |
| CE014 | PFN’s chips page claims first-generation MN-Core ran Matlantis neural-network atomistic simulation more than five times faster than GPU. | 高 | SE002, SE018 |
| CE015 | PFN’s chips page states the second-generation MN-Core 2 was experimentally used through PFCP for Matlantis and outperformed GPUs on low-atom-number simulations. | 高 | SE002, SE018 |
| CE016 | PFN’s original Chainer framework was released in June 2015 as an open-source deep-learning framework. | 高 | SE008, SE030 |
| CE017 | ChainerX was released as a C++ ndarray/autograd implementation integrated into Chainer v6 beta to improve performance. | 高 | SE009, SE038 |
| CE018 | The Chainer project announced in December 2019 that Chainer would shift to a maintenance phase with development limited to bug fixes and maintenance. | 高 | SE037, SE010 |
| CE019 | PFN announced in December 2019 that it migrated its deep-learning research platform to PyTorch. | 高 | SE010, SE037 |
| CE020 | PFN announced in May 2020 that it deepened collaboration with the PyTorch community after the migration. | 高 | SE011, SE010 |
| CE021 | CuPy is maintained as a NumPy/SciPy-compatible array library for GPU-accelerated computing and originated with PFN/Preferred Infrastructure copyright. | 中 | SE031, SE040 |
| CE022 | Optuna remains an active hyperparameter-optimization framework with documentation describing define-by-run search spaces, pruning, visualization, and integrations. | 中 | SE032, SE039 |
| CE023 | PFN reported Optuna v4.0 in 2024 with over 10,000 GitHub stars and use in over 16,000 software applications. | 中 | SE032, SE039 |
| CE024 | PFIO is a PFN open-source IO library for unified access to various filesystems. | 中 | SE033 |
| CE025 | PFN established Preferred Elements in 2023 for development and sales of multimodal foundation models. | 中 | SE022 |
| CE026 | PFN announced in 2025 that it would absorb Preferred Elements to bolster development and social implementation of PLaMo. | 高 | SE023, SE022 |
| CE027 | PFN’s PLaMo business page describes PLaMo as a family of Japanese-focused foundation models developed from scratch and includes open models developed through GENIAC. | 高 | SE004, SE036 |
| CE028 | PFN launched PLaMo Prime in December 2024 through PLaMo API and PLaMo Chat. | 高 | SE021, SE004 |
| CE029 | PFN’s Hugging Face organization page shows an external developer distribution channel for PFN models. | 中 | SE036 |
| CE030 | PFN and Toyota began joint R&D on self-driving cars in 2014. | 中 | SE006 |
| CE031 | Toyota and PFN began joint development of service robots in 2019. | 中 | SE030, SE024 |
| CE032 | In June 2026 PFN and Toyota’s Frontier Research Center started joint research to accelerate physical AI using MN-Core L Series processors for high-speed on-premise robot inference. | 高 | SE024, SE030 |
| CE033 | FANUC and PFN announced a capital alliance in 2015, anchoring PFN’s industrial robot channel. | 中 | SE007 |
| CE034 | PFCC launched Matlantis as a cloud-based atomistic simulator in 2021. | 高 | SE025, SE034 |
| CE035 | PFCC launched Matlantis in the United States in 2023 and described it as a high-speed universal atomistic simulator for AI-driven materials discovery. | 高 | SE026, SE034 |
| CE036 | The PFP neural network potential underlying Matlantis was published in Nature Communications as applicable to arbitrary combinations of 45 elements. | 高 | SE041, SE027 |
| CE037 | PFN and ENEOS announced an updated PFP neural network potential for Matlantis, with later product materials stating expanded chemistry coverage. | 高 | SE027, SE034 |
| CE038 | Kachaka is a commercial Preferred Robotics autonomous mobile robot product, and PFN uses it as a workload example for MN-Core acceleration. | 高 | SE035, SE002 |
| CE039 | KDDI’s GPU Cloud page confirms a carrier-grade NVIDIA GPUaaS offering, but public sources reviewed here do not verify a named 2024 KDDI investment hosted by PFN. | 中 | SE042 |
| CE040 | The principal adverse product-technology risk is that MN-Core remains much less broadly adopted than NVIDIA GPUs despite PFN’s efficiency wins and internal workload results. | 中 | SE002, SE028, SE029, SE042 |
| CE041 | PFN’s stack exposes a supplier concentration risk because MN-Core generation-one disclosures specify TSMC 12nm fabrication but public materials do not give equivalent manufacturer, packaging, yield, or volume data for MN-Core 2 and L1000. | 中 | SE012, SE002, SE020 |
| CE042 | Public PFN materials do not provide SOC 2, ISO 27001, model safety audit, export-control, or customer data-residency documentation for PFCP or PLaMo. | 低 | |
| CU001 | PFN’s customer and partner evidence spans automotive, factory automation, industrial edge systems, materials simulation, pharmaceuticals, communications infrastructure, robotics, generative AI, advertising, and mission-critical industrial AI. | 高 | SU001, SU002, SU003, SU009, SU014, SU018, SU020, SU032 |
| CU002 | PFN’s current shareholder roster includes Toyota Motor, Fanuc, Hitachi, Mitsui & Co., Mizuho Bank, NTT, Chugai Pharmaceutical, ENEOS Innovation Partners, Hakuhodo DY Holdings, Mitsubishi Corporation, and others. | 中 | SU001 |
| CU003 | PFN and Toyota’s Frontier Research Center began 2026 joint research to accelerate physical-AI inference for robots using MN-Core L series processors. | 中 | SU003 |
| CU004 | Toyota’s 2017 additional investment in PFN was 10.5 billion yen and targeted AI R&D in mobility fields such as automated driving. | 高 | SU004, SU005 |
| CU005 | Toyota has been both a long-running strategic investor and an active R&D partner, making it PFN’s clearest automotive anchor relationship. | 中 | SU003, SU004, SU005 |
| CU006 | FANUC and PFN agreed in June 2015 to an R&D alliance applying machine learning and deep learning to machine tools and robotics. | 中 | SU006 |
| CU007 | FANUC invested 900 million yen in PFN under an August 2015 capital alliance. | 中 | SU007 |
| CU008 | FANUC, Hitachi, and PFN agreed in 2018 to establish Intelligent Edge System, LLC for AI edge devices in industrial and social infrastructure fields. | 中 | SU009, SU010, SU011 |
| CU009 | FANUC AI functions developed with PFN moved beyond research into productized factory automation and robot functions by 2018-2019. | 中 | SU012, SU013 |
| CU010 | PFN raised over 2 billion yen from FANUC, Hakuhodo DY Holdings, Hitachi, Mizuho Bank, and Mitsui & Co. in December 2017. | 中 | SU008 |
| CU011 | ENEOS and PFN co-developed the PFP technology powering Matlantis and released version 7 in 2024. | 高 | SU014, SU015 |
| CU012 | Matlantis operates as a dedicated simulator business with Japan and U.S. offices, suggesting the ENEOS/PFN collaboration has become a customer-facing product company rather than a one-off project. | 中 | SU016, SU017 |
| CU013 | Business Wire described Preferred Computational Chemistry as a joint venture between PFN and ENEOS that launched Matlantis in the United States for AI-driven materials discovery. | 中 | SU017 |
| CU014 | Chugai and PFN entered a comprehensive partnership agreement in 2018 to apply deep learning and AI to innovative drug discovery. | 中 | SU018 |
| CU015 | Chugai invested about 700 million yen in PFN as part of the July 2018 financing round. | 高 | SU019, SU018 |
| CU016 | NTT Communications and NTT PC Communications supported PFN’s private-sector supercomputer through data-center housing, networks, operations, and technical support. | 高 | SU020, SU021, SU022 |
| CU017 | NTT DOCOMO Business and NTTPC publish PFN customer/use-case pages, corroborating an infrastructure supplier relationship rather than a pure investor logo. | 高 | SU021, SU022 |
| CU018 | KDDI’s GPU Cloud page positions KDDI as a GPU cloud provider and partner-services channel for AI learning, big-data analysis, and R&D workloads. | 中 | SU023 |
| CU019 | SoftBank announced a 2026 AI Data Center GPU Cloud powered by Infrinia AI Cloud OS as part of its Neocloud business. | 中 | SU024 |
| CU020 | JR East announced 2026 autonomous track-inspection robot work, and Preferred Robotics announced development of railway-infrastructure maintenance robots with JR East. | 高 | SU025, SU026 |
| CU021 | Kachaka Pro is sold as a compact AMR for transport automation, giving the PFN/Preferred Robotics group a direct robot-product commercialization path outside enterprise co-development. | 中 | SU027 |
| CU022 | PFN and Preferred Elements were selected for GENIAC Cycle 2, a METI/NEDO-supported project to improve Japan’s generative-AI foundation-model development capabilities. | 高 | SU028, SU029 |
| CU023 | METI describes PFN/PFE as GENIAC awardees that built a 100B-parameter multimodal foundation model in Cycle 1 and targeted efficient 8B-scale models in Cycle 2. | 中 | SU029 |
| CU024 | MHI and PFN entered a June 2026 business alliance to jointly develop Japan-made AI technologies for mission-critical applications. | 中 | SU032 |
| CU025 | Mitsubishi Corporation subscribed to PFN shares and entered a capital and business alliance in December 2024. | 中 | SU033 |
| CU026 | Preferred Medicine, a joint venture between PFN and Mitsui & Co., presented machine-learning-based early cancer-detection research using circulating miRNA profiles. | 中 | SU034 |
| CU027 | MN-Core processors have been developed with Kobe University since 2016 and are now tied to PFN’s AI-chip customer and infrastructure story. | 中 | SU035, SU036 |
| CU028 | Hakuhodo DY Holdings agreed to invest in and strategically partner with PFN for AI business development and implementation. | 高 | SU038, SU008 |
| CU029 | Hakuhodo DY Digital launched colorized manga products with PFN cooperation using PaintsChainer in 2018, providing a creative/advertising use-case proof point. | 中 | SU037 |
| CU030 | PFN’s go-to-market pattern is co-creation first: joint research or capital/business alliances precede commercialization in Toyota, FANUC, Chugai, ENEOS, MHI, and Hakuhodo examples. | 中 | SU002, SU003, SU006, SU014, SU018, SU032 |
| CU031 | Several relationships are simultaneously investor, partner, and customer-proof relationships, which strengthens strategic depth but increases concentration exposure to Japanese incumbents. | 中 | SU001, SU003, SU008, SU014, SU018 |
| CU032 | Public evidence does not disclose PFN revenue by customer, ARR, NRR, GRR, churn, or customer-count metrics. | 中 | SU001, SU002, SU030 |
| CU033 | CNBC quoted PFN’s CEO saying commercialization can take three to five years from joint research to practical launch, highlighting long pilot-to-product cycles. | 中 | SU030 |
| CU034 | No fetched source corroborated the suggested Oisix ra daichi/CraftyFarm relationship; the retained Oisix official page only establishes Oisix’s food-business context. | 中 | SU039 |
| CU035 | No public source reviewed quantified revenue from Toyota, FANUC, or any other single customer, so top-customer concentration cannot be calculated externally. | 中 | SU003, SU005, SU007, SU008 |
| CU036 | Named executive/customer-side sources exist for Toyota, Chugai, NTT, JR East, ENEOS, SoftBank, MHI, Mitsubishi Corporation, and Hakuhodo, meeting the chapter’s customer-proof requirement. | 高 | SU003, SU015, SU018, SU020, SU024, SU025, SU032, SU033, SU038 |
| CU037 | PFN’s 2026 customer evidence is unusually active: Toyota FRC, JR East/Preferred Robotics, SoftBank GPU Cloud, and MHI alliance were all announced or active in 2026. | 高 | SU003, SU024, SU025, SU026, SU032 |
| CU038 | The FANUC relationship shows repeat depth from 2015 R&D and capital alliance to 2018-2019 AI functions and the FANUC-Hitachi-PFN JV. | 中 | SU006, SU007, SU009, SU012, SU013 |
| CU039 | The ENEOS relationship shows product durability from PFP/Matlantis co-development to a U.S. launch and 2024 version-7 release. | 高 | SU014, SU015, SU017 |
| CU040 | The NTT relationship appears infrastructure-oriented rather than end-customer SaaS revenue: public proof centers on data-center, network, GPU, and supercomputer support. | 高 | SU020, SU021, SU022 |
| CU041 | The medical AI segment contains Chugai drug discovery and Mitsui/Preferred Medicine cancer-detection work, but public sources do not show scaled recurring clinical revenue. | 中 | SU018, SU019, SU034 |
| CU042 | PFN’s customer base is primarily Japan-centered; public production proof outside Japan is clearest for Matlantis U.S. launch rather than broad multinational customer deployments. | 中 | SU001, SU016, SU017, SU032 |
| CR001 | PFN describes itself as vertically integrated across semiconductors, computing infrastructure, solutions and applications, which expands execution scope beyond a pure software startup. | 中 | SR001 |
| CR002 | PFN’s company page says it has engaged in joint R&D with industry leaders since 2014, indicating a long research-collaboration operating model. | 中 | SR001 |
| CR003 | PFN has publicly established AI governance, but public materials do not disclose incident history, model-risk metrics, or external audit outcomes. | 中 | SR001 |
| CR004 | The MN-Core series is PFN’s proprietary AI-chip line developed with Kobe University since 2016 and positioned against general-purpose GPUs. | 中 | SR002 |
| CR005 | PFN lists commercial MN-Core 2 products including MN-Server 2 at 20 million yen and a MN-Core 2 devkit at 2.0–2.5 million yen, proving productization but not broad market adoption. | 中 | SR002 |
| CR006 | PFN claims MN-3 topped the Green500 list multiple times, supporting technical efficiency but not necessarily customer-scale commercial demand. | 中 | SR002 |
| CR007 | PFN moved Chainer into maintenance mode in 2019 and migrated its deep learning R&D platform to PyTorch, a material platform pivot from an internally controlled framework to an external ecosystem. | 高 | SR003, SR028 |
| CR008 | PFN’s Chainer announcement explicitly said the era when the deep-learning framework itself was a competitive edge had matured, reducing the strategic value of PFN’s original framework differentiation. | 中 | SR003 |
| CR009 | PFN’s FANUC collaboration placed PFN inside the FIELD system and named FANUC, Cisco and PFN as providers of middleware platform software. | 中 | SR004 |
| CR010 | The FANUC FIELD system announcement tied PFN to factory analytics, robots, CNCs, sensors and Chainer-based middleware, creating partner-specific integration and switching-cost risk. | 中 | SR004 |
| CR011 | Toyota’s third-party allocation to PFN and Toyota service-robot collaboration make Toyota a financial and strategic counterparty rather than an ordinary customer. | 中 | SR005, SR006 |
| CR012 | Woven by Toyota’s public mobility-and-technology mandate creates a plausible in-house Toyota alternative for AI mobility capabilities that historically overlapped PFN’s Toyota work. | 中 | SR007, SR006 |
| CR013 | Reuters reported PFN’s domestic AI-chip development in 2023, confirming that MN-Core strategy remains visible to independent technology media. | 中 | SR008 |
| CR014 | SemiAnalysis coverage of MN-Core 2 places PFN in a specialized accelerator market where product assessment depends on performance, software ecosystem and deployment depth. | 中 | SR009 |
| CR015 | NVIDIA’s public 10-K and developer materials demonstrate the scale, software ecosystem and pace of incumbent AI accelerator competition confronting niche chips such as MN-Core. | 高 | SR010, SR034 |
| CR016 | CSIS identifies NVIDIA CUDA and customer ecosystem effects as a key barrier for customers leaving NVIDIA chips, a direct adoption obstacle for PFN accelerators. | 高 | SR015, SR010 |
| CR017 | AWS Trainium and Google Cloud TPU provide hyperscaler-backed alternatives for AI training and inference, reducing the addressable market for independent AI accelerators. | 高 | SR011, SR012 |
| CR018 | Open-source ecosystems and broadly available frameworks reduce PFN software differentiation unless PFN proves proprietary deployment, data or chip-integration advantages. | 中 | SR003, SR028 |
| CR019 | BIS and CSIS sources corroborate that advanced AI chips and semiconductor equipment are exposed to U.S. export-control chokepoints. | 高 | SR014, SR015 |
| CR020 | Japan’s METI export-control posture adds a domestic regulatory layer for semiconductor manufacturing equipment and dual-use technology. | 中 | SR016 |
| CR021 | The EU AI Act and EUR-Lex regulation create risk-tiered obligations that can attach to AI systems placed on or used in the EU market. | 高 | SR017, SR018 |
| CR022 | ISO 10218-1 and ISO 10218-2 are the relevant industrial robot and robot-system safety standards for deployments involving physical robot systems. | 高 | SR019, SR020 |
| CR023 | JPO AI-patent materials show that AI-related inventions remain an active legal-examination area in Japan, making patent freedom-to-operate and ownership diligence material. | 中 | SR021 |
| CR024 | IPA talent materials indicate Japan tracks IT and digital talent as a policy issue, supporting the risk that PFN competes in a constrained domestic AI-engineering labor market. | 中 | SR022 |
| CR025 | PFN’s most visible named leaders in public materials are Toru Nishikawa and Daisuke Okanohara, creating key-person diligence requirements around succession, retention and investor/customer access. | 中 | SR001, SR033 |
| CR026 | IMF commentary that yen weakness has limited benefits supports a macro risk: USD investors may see PFN valuation volatility and imported compute components may become more expensive in yen terms. | 中 | SR023, SR024 |
| CR027 | CB Insights lists PFN as having raised $308.23M and being in unicorn collections, while also showing a -70 Mosaic Score movement in the past 30 days. | 中 | SR024 |
| CR028 | Crunchbase and PitchBook profiles confirm PFN remains a private-market company with funding-history opacity from public sources. | 中 | SR025, SR026 |
| CR029 | JPX listing materials confirm a public-market route exists in Japan, but they do not demonstrate that PFN meets growth-market timing, profitability, governance or liquidity expectations. | 中 | SR027 |
| CR030 | Reuters coverage of AI bubble concerns is an adverse market signal for PFN’s future financing terms because PFN is a capital-intensive AI infrastructure company. | 中 | SR031 |
| CR031 | No public source reviewed disclosed PFN revenue, ARR, gross margin, operating loss, burn rate or cash runway as of 2026-06-14. | 中 | |
| CR032 | No public evidence of PFN layoffs, accounting scandal, enforcement action, or founder departure was found in reviewed sources through 2026-06-14. | 中 | |
| CR033 | The sale or transfer thesis around MN-Core 2 could not be confirmed from accessible Sakura URLs; the risk remains an unresolved diligence question rather than a validated adverse event. | 低 | |
| CR034 | PFN’s SC23 presence provides technical proof of ongoing MN-Core promotion but does not by itself validate customer traction or revenue scale. | 中 | SR032 |
| CR035 | The combination of Toyota funding, Toyota collaboration and Woven by Toyota creates a customer-investor overlap that can produce conflicts over roadmap priorities and independence. | 中 | SR005, SR006, SR007 |
| CR036 | FANUC dependency is partly mitigated by PFN’s broader vertical strategy, but historical FIELD integration still creates concentration risk if FANUC reduces strategic emphasis. | 中 | SR004, SR001 |
| CR037 | NVIDIA, AWS and Google together represent a three-front competitive threat: merchant GPU platforms, hyperscaler custom silicon and cloud-integrated TPU/Trainium services. | 高 | SR010, SR011, SR012, SR034 |
| CR038 | Export-control risk is high severity because PFN’s AI-chip work depends on global semiconductor tooling, foundry supply chains, and access to restricted customers and components. | 高 | SR014, SR015, SR016 |
| CR039 | Robot-safety and AI-regulation exposure is moderate because PFN sells into real-world industrial and mobility contexts where physical harms and regulated AI use cases can arise. | 高 | SR017, SR018, SR019, SR020 |
| CR040 | Talent risk is high because PFN needs scarce semiconductor, compiler, robotics and foundation-model engineers while competing against Japanese champions and global hyperscalers. | 中 | SR022, SR010, SR011, SR012 |
| CR041 | The top three thesis-break risks are commercial productization failure, Toyota/FANUC concentration or displacement, and inability to compete with NVIDIA/hyperscaler AI infrastructure. | 高 | SR001, SR004, SR005, SR010, SR011, SR012 |
| CR042 | Investor kill criteria should include evidence of customer churn from Toyota or FANUC, new financing at a down round, export-control licensing denial, and MN-Core unit economics below plan. | 中 | SR004, SR005, SR014, SR024 |
| CR043 | Multiple independent adverse sources exist for this chapter: SemiAnalysis, NVIDIA SEC filing, CSIS, BIS/METI regulatory materials, Reuters AI-bubble reporting, IMF yen analysis and CB Insights score movement. | 高 | SR009, SR010, SR014, SR015, SR016, SR023, SR024, SR031 |
| CR044 | PFN’s own statement that Chainer’s framework differentiation era had matured is a rare company-issued adverse datapoint on historical moat erosion. | 中 | SR003 |
| CR045 | The risk register requires private diligence on revenue mix, Toyota/FANUC contract terms, MN-Core customer pipeline, export-control classification, patent ownership, and management retention to resolve material gaps. | 中 | SR001, SR002, SR004, SR021, SR024 |
| CV001 | PFN announced that Toyota invested about 10.5 billion yen in August 2017, making the Toyota round the cleanest primary-source valuation anchor. | 高 | SV001, SV002 |
| CV002 | Third-party coverage of Toyota’s investment reported a roughly $95 million amount and an implied valuation near the multi-billion-dollar range, but the exact post-money is not in PFN’s press release. | 中 | SV002 |
| CV003 | SBI Holdings and PFN announced a capital and business alliance for next-generation AI semiconductors in August 2024. | 中 | SV003 |
| CV004 | PFN announced a first close of 19 billion yen in December 2024, combining equity financing led by SBI Group with debt financing from financial institutions. | 高 | SV004, SV007 |
| CV005 | The December 2024 financing named Development Bank of Japan, Mitsubishi Corporation, SBI Group, Sekisui House Investment Limited Partnership and Wacom as investors. | 中 | SV004 |
| CV006 | PFN announced an additional 5 billion yen extension financing in April 2025. | 中 | SV005 |
| CV007 | PFN announced a further undisclosed extension financing in June 2025, so the latest public capital total is not enough to compute a full post-money valuation. | 中 | SV006 |
| CV008 | No PFN source reviewed for the 2024 and 2025 financing rounds disclosed an explicit post-money valuation. | 高 | SV003, SV004, SV005, SV006 |
| CV009 | A low-reputation adverse article alleged a 50% drop in PFN valuation, which is insufficient to override primary financing releases but is a useful down-round risk flag. | 低 | SV008 |
| CV010 | CNBC described PFN as a Japanese AI unicorn pursuing deep-learning applications in real-world robotics and trucking contexts. | 中 | SV009 |
| CV011 | J-Startup lists Preferred Networks as a selected Japanese startup, reinforcing government-recognition but not valuation. | 中 | SV010 |
| CV012 | ENEOS and PFN released version 7 of the PFP neural network potential, supporting the materials-simulation line in the sum-of-parts model. | 中 | SV011 |
| CV013 | PFN and Mitsubishi Heavy Industries announced a 2026 business alliance, a fresh strategic-proof signal relevant to exit-premium assumptions. | 中 | SV012 |
| CV014 | JPX states that an IPO process commonly takes about one year from kick-off to listing and requires audited financial statements, so PFN is not IPO-ready without public-quality audits. | 中 | SV013 |
| CV015 | NVIDIA’s SEC filing and Yahoo Finance market data make it a high-growth AI-chip public comparable rather than a direct startup peer. | 高 | SV014, SV017 |
| CV016 | AMD’s SEC filing and Yahoo Finance market data provide a lower-multiple AI-chip comparator than NVIDIA for chip exposure. | 高 | SV015, SV018 |
| CV017 | Palantir’s SEC filing and market data make it the most relevant public enterprise-AI software multiple for PFN’s PLaMo and solutions exposure. | 高 | SV016, SV019 |
| CV018 | C3.ai and UiPath provide public enterprise-AI and automation references, but their business models remain more software-pure than PFN’s hardware-and-services mix. | 中 | SV020, SV021 |
| CV019 | Fanuc and CYBERDYNE provide Japanese robotics comparables that anchor a lower multiple range than frontier-AI software. | 中 | SV022, SV023 |
| CV020 | SenseTime provides an AI-software public comparable with China-market and regulatory differences that limit direct applicability. | 中 | SV024 |
| CV021 | Anthropic announced a $61.5 billion post-money Series E, setting an upper-bound frontier-model comp far above PFN’s current evidence base. | 中 | SV025 |
| CV022 | Reuters and TechCrunch reported OpenAI’s 2024 financing at a $157 billion valuation, an extreme upper-bound comp not directly transferable to PFN. | 高 | SV026, SV027 |
| CV023 | Reuters reported Mistral AI raised 600 million euros, supporting the European foundation-model comp set. | 中 | SV028 |
| CV024 | Crunchbase News reported Cohere raised $500 million at a $5.5 billion valuation, a more relevant enterprise-AI private comp than OpenAI. | 中 | SV029 |
| CV025 | Reuters and Figure’s announcement indicate Figure raised $675 million at a $2.6 billion valuation, a physical-AI robotics comp close to PFN’s unconfirmed range. | 高 | SV030, SV031 |
| CV026 | Wayve announced and TechCrunch reported a more than $1 billion Series C led by SoftBank, validating large physical-AI funding rounds for embodied autonomy. | 高 | SV032, SV033 |
| CV027 | Forbes and Nikkei Asia reported Sakana AI’s large 2024 financing, with Nikkei describing a $1.5 billion valuation that challenges the “most valuable Japanese AI startup” narrative. | 高 | SV034, SV035 |
| CV028 | Amazon’s Covariant announcement supports the view that robotics-AI exits may occur through strategic acquisitions or acqui-hires rather than near-term IPOs. | 中 | SV036 |
| CV029 | Damodaran’s sector price-to-sales data supports using revenue multiples as a valuation cross-check when company revenue can be estimated. | 中 | SV037 |
| CV030 | PwC’s 2026 M&A outlook supports modeling strategic M&A as a realistic liquidity route when IPO readiness is not established. | 中 | SV038 |
| CV031 | CB Insights’ AI 100 provides a broad AI-startup benchmark set, but it does not substitute for PFN-specific revenue or margin evidence. | 中 | SV039 |
| CV032 | A sum-of-parts approach is more appropriate than a single revenue multiple because PFN spans AI chips, PLaMo/foundation models, robotics, materials simulation, and cloud infrastructure. | 中 | SV004, SV009, SV011, SV012 |
| CV033 | The base-case valuation range of $2.0 billion to $2.8 billion assumes current strategic financing converts to commercial chip and LLM revenue but no OpenAI-style frontier-model multiple. | 中 | SV004, SV015, SV016, SV019, SV037 |
| CV034 | The bear-case valuation range of $1.0 billion to $1.6 billion assumes a Japan AI valuation reset, hardware margin drag, and no disclosed ARR to support premium software multiples. | 中 | SV008, SV018, SV022, SV023, SV037 |
| CV035 | The bull-case valuation range of $4.0 billion to $6.0 billion requires credible evidence that MN-Core or PLaMo can command venture-scale AI-chip or foundation-model economics. | 中 | SV014, SV017, SV021, SV025, SV026 |
| CV036 | At an unconfirmed $2.5 billion entry valuation, a 3x target return requires a $7.5 billion exit before dilution and preference effects. | 中 | SV004, SV037 |
| CV037 | At an unconfirmed $3.0 billion entry valuation, a 3x target return requires a $9.0 billion exit, which public evidence does not yet support. | 中 | SV004, SV037 |
| CV038 | PFN’s late-2024 and 2025 strategic financings increase preference-stack and dilution complexity even though public filings do not reveal liquidation preferences. | 中 | SV004, SV005, SV006 |
| CV039 | The strongest diligence ask is audited revenue, ARR, gross margin, segment contribution and cap-table preference data because public valuation support is otherwise indirect. | 高 | SV004, SV013, SV037 |
| CV040 | A Tokyo Stock Exchange Growth Market listing is more plausible than Prime if PFN pursued a near-term IPO, but audited statements and scale disclosures remain prerequisites. | 中 | SV013 |
| CV041 | Strategic M&A by an industrial, cloud, semiconductor or robotics acquirer is likely more realistic than a near-term IPO if PFN seeks liquidity before multi-year audit readiness. | 中 | SV012, SV030, SV036, SV038 |
| CV042 | The preferred investment stance is research-more rather than buy because the last confirmed valuation is historical and current post-money valuation is not publicly disclosed. | 高 | SV001, SV004, SV005, SV006, SV008 |
| CV043 | The valuation stance is stretched at any assumed $2.5 billion to $3.0 billion price unless diligence proves high-margin recurring revenue or semiconductor gross margins. | 中 | SV014, SV017, SV019, SV025, SV037 |
| CV044 | The high-confidence positive case is strategic validation: Toyota historically, SBI and DBJ in 2024, and MHI in 2026 each point to Japanese industrial support for PFN. | 高 | SV001, SV003, SV004, SV012 |
| CV045 | The high-confidence negative case is evidence quality: valuation, revenue, ARR, gross margin and IPO timing are not publicly disclosed with enough precision to underwrite a primary investment. | 高 | SV004, SV005, SV006, SV013 |
| 编号 | 出版方 | 标题 | 引文 |
|---|---|---|---|
| SO001 | Preferred Networks | Company - Preferred Networks, Inc. | Company name Preferred Networks, Inc.; Established March 26, 2014; Location Otemachi Building, 1-6-1 Otemachi, Chiyoda-ku, Tokyo. |
| SO002 | Preferred Networks | Co-Founders' Message - Company - Preferred Networks, Inc. | Preferred Networks is committed to mastering every aspect of computing, advancing our business daily. |
| SO003 | Preferred Networks | Milestones and Awards - Company - Preferred Networks, Inc. | Founded in March 2014, PFN has engaged in joint research and development with industry leaders. |
| SO004 | Preferred Networks | Business - Preferred Networks, Inc. | Vertical integration of computer science—from semiconductors and computing infrastructure to solutions and applications. |
| SO005 | Preferred Networks | PFN’s Supercomputers - Preferred Networks | MN-3 topped Green500 ranking 3 times as world’s most energy-efficient. |
| SO006 | Preferred Networks | MN-Core Series - Preferred Networks | MN-Core Series is PFN’s proprietary processor series for AI workloads. |
| SO007 | Preferred Networks | Careers - Preferred Networks, Inc. | Join the PFN team. |
| SO008 | Preferred Networks | Preferred Networks received about 10.5 billion yen in investments from Toyota Motor Corporation | PFN will receive an additional investment of approximately 10.5 billion yen from Toyota Motor Corporation. |
| SO009 | Preferred Networks | FANUC and Preferred Networks announce capital alliance | Amount of finance: 900 million JPY. |
| SO010 | Preferred Networks | Preferred Networks Migrates its Deep Learning Research Platform to PyTorch | Chainer will move into a maintenance phase. |
| SO011 | Preferred Networks | SBI Holdings and PFN Agree to Form Capital and Business Alliance for Next-Generation AI Semiconductors | SBI Holdings agreed to invest up to 10 billion yen through third-party allocation of new shares. |
| SO012 | Preferred Networks | ENEOS and PFN Begin World’s First AI-Based Autonomous Operation of Crude Oil Processing Unit | World’s first continuous AI-based autonomous operation of a crude oil processing unit. |
| SO013 | Preferred Networks | PFN Raises Total of 19 Billion Yen in Latest Round | PFN announced that it has raised a total of 19 billion yen in the first close. |
| SO014 | Preferred Networks | PFN Raises Additional 5 Billion Yen in Extension Round | The extension round follows the initial funding in December 2024, bringing the total amount raised to date to 24 billion yen. |
| SO015 | Preferred Networks | Mitsubishi Heavy Industries and Preferred Networks Form Business Alliance | Jointly develop Japan-made AI technologies for mission-critical applications. |
| SO016 | Preferred Networks | PFN Starts Joint Research with Toyota’s Frontier Research Center | Companies to test ultra-high-bandwidth MN-Core L series for robots requiring high-speed on-premise inference. |
| SO017 | Preferred Networks | Preferred Networks, GMO Internet Group and GMO Cybersecurity by Ierae to Establish Joint Venture GMO Preferred Security | New joint venture to deliver secure, Japan-built AI environment from hardware to software. |
| SO018 | Preferred Networks | PFN Establishes New Subsidiary Preferred Elements for Development and Sales of Multimodal Foundation Model | PFN established Preferred Elements for development and sales of multimodal foundation models. |
| SO019 | CNBC | This Japanese AI unicorn has big plans to use deep learning to fix real-world problems | This Japanese AI unicorn has big plans to use deep learning to fix real-world problems. |
| SO020 | The Bridge | Preferred Networks, AI Development Unicorn, Raises 19B Yen Including Debt Financing—SBI and Mitsubishi Corporation Among Investors | Preferred Networks, AI Development Unicorn, Raises 19B Yen Including Debt Financing. |
| SO021 | JAKOTA Index | SBI Holdings to Invest ¥10 Billion in Preferred Networks for AI Chip Development | SBI Holdings to invest ¥10 billion in Preferred Networks for AI chip development. |
| SO022 | Mitsubishi Corporation | PFN, Mitsubishi Corporation and IIJ to Establish Joint Venture Preferred Computing Infrastructure | PFN, Mitsubishi Corporation and IIJ to establish a joint venture for AI cloud computing. |
| SO023 | Internet Initiative Japan | PFN, Mitsubishi Corporation and IIJ to Establish Joint Venture Preferred Computing Infrastructure for AI Cloud Computing | Preferred Computing Infrastructure will provide AI cloud computing services. |
| SO024 | TOP500 | MN-3 - MN-Core Server, Xeon Platinum 8260M 24C 2.4GHz | MN-3 system profile lists Preferred Networks MN-Core and MN-Core DirectConnect. |
| SO025 | TOP500 | November 2021 Green500 List | Green500 list ranks supercomputers by energy efficiency. |
| SO026 | Rapidus | PFN, Rapidus and SAKURA internet Reach Basic Agreement toward Japan-Made Green AI Cloud Infrastructure | PFN, Rapidus and SAKURA internet reach basic agreement toward realization of Japan-made green AI cloud infrastructure. |
| SO027 | Qualcomm AI Hub | PLaMo-1B | PLaMo-1B model is listed on Qualcomm AI Hub. |
| SO028 | Supermicro | Supermicro Contributes to the MN-3 Supercomputer Earning #1 on Green500 | MN-3 supercomputer earned #1 on Green500. |
| SO029 | Global Venturing | Toyota provides $95m to its Preferred Networks | Toyota provides $95m to its Preferred Networks. |
| SO030 | CB Insights | Preferred Networks - Products, Competitors, Financials, Employees, Headquarters Locations | CB Insights profiles Preferred Networks products, competitors, financials, employees and headquarters. |
| SO031 | Wikipedia | Preferred Networks | Preferred Networks is a Japanese technology company focused on artificial intelligence. |
| SO032 | Matlantis | Matlantis | AI simulator for predicting atomic-level phenomena | Matlantis is an AI simulator for predicting atomic-level phenomena. |
| SO033 | Preferred Networks | PFCC Launches Matlantis Atomistic Simulator as Cloud-Based Service | PFCC launches Matlantis atomistic simulator as a cloud-based service. |
| SM001 | Preferred Networks, Inc. | Preferred Networks, Inc. | PFN says it develops AI chips, computing infrastructure, generative AI foundation models and applications in-house. |
| SM002 | Preferred Networks, Inc. | Business - Preferred Networks, Inc. | PFN lists Fanuc industrial robots, ENEOS plant automation and Chugai experiment automation as business co-creation examples. |
| SM003 | Preferred Networks, Inc. | News - Preferred Networks, Inc. | PFN newsroom lists 2026 releases including Mitsubishi Heavy Industries and Toyota Frontier Research Center announcements. |
| SM004 | Preferred Networks, Inc. | PFN Starts Joint Research with Toyota’s Frontier Research Center | PFN announced joint research with Toyota Frontier Research Center to accelerate physical AI using MN-Core L series processors. |
| SM005 | Preferred Networks, Inc. | PFN, IIJ and JAIST Deploy Direct Liquid-Cooled AI Servers | PFN, IIJ and JAIST said full-scale operation of AImod would begin in April 2026. |
| SM006 | Internet Initiative Japan Inc. | PFN, IIJ and JAIST Deploy Direct Liquid-Cooled, High-Density AI Servers | The IIJ release describes direct liquid-cooled high-density AI servers using PFN MN-Core series semiconductors. |
| SM007 | International Federation of Robotics | World Robotics 2025 report – INDUSTRIAL ROBOTS – released by IFR | IFR reported 542,000 industrial robots installed globally in 2024 and 4,664,000 in operational stock. |
| SM008 | International Federation of Robotics | International Federation of Robotics Shares Top Five Global Robotics Trends for 2026 | IFR identified AI-driven robotics and automation trends as core 2026 themes. |
| SM009 | Gartner | Gartner Forecasts Worldwide AI Spending to Grow 47% in 2026 | Gartner forecast worldwide AI spending of $2.59 trillion in 2026. |
| SM010 | Gartner | Gartner Says Worldwide AI Spending Will Total $2.5 Trillion in 2026 | Gartner table listed 2026 AI software spending of $452.458 billion and AI services spending of $588.645 billion. |
| SM011 | IDC | Japan’s AI Infrastructure Will Surge Past $5.5 Billion in 2026 | IDC said Japan domestic AI infrastructure spending will exceed $5.5 billion in 2026. |
| SM012 | NEDO | New Energy and Industrial Technology Development Organization | NEDO describes itself as Japan’s national R&D agency promoting technological development for a sustainable society. |
| SM013 | Global Market Insights | AI Accelerator Chips Market Size & Share | Industry Report, 2035 | GMI valued the global AI accelerator chips market at $120.2 billion in 2025 and $154.6 billion in 2026. |
| SM014 | Mordor Intelligence | Agricultural Robots Market Size, Share & Report 2031 | Mordor projects agricultural robots at $18.0 billion in 2026, growing to $41.3 billion by 2031. |
| SM015 | Grand View Research | Artificial Intelligence In Drug Discovery Market Report, 2033 | Grand View projects AI in drug discovery from $2.9 billion in 2026 to $13.8 billion by 2033. |
| SM016 | Research and Markets | AI in Drug Discovery Market Report 2026 | Research and Markets valued AI in drug discovery at $2.93 billion in 2026 with 26.2% CAGR to 2030. |
| SM017 | IMARC Group | Japan Artificial Intelligence-as-a-Service Market Statistics | IMARC forecast Japan AIaaS to grow from $1.2545 billion in 2025 to $15.0048 billion by 2034. |
| SM018 | Value Market Research | Japan Artificial Intelligence Market Size, Share, Growth, Demand, 2034 | VMR projected Japan AI market from $19.83 billion in 2025 to $289.88 billion by 2034. |
| SM019 | MarketsandMarkets | Industrial Robotics Market Size, Share and Growth | MarketsandMarkets valued industrial robotics at $15.5 billion in 2026 and $20.8 billion by 2032. |
| SM020 | Mordor Intelligence | Smart Manufacturing Market Size, Share, Forecast Report 2025–2031 | Mordor valued smart manufacturing at $387.14 billion in 2026 and names FANUC among major players. |
| SM021 | Fortune Business Insights | Artificial Intelligence (AI) Market | Global Report 2034 | Fortune Business Insights projects the global AI market at $375.93 billion in 2026. |
| SM022 | Chugai Pharmaceutical | Platforms & Technologies | R&D | Innovation | Chugai describes research platforms and technologies for drug discovery and pharmaceutical R&D. |
| SM023 | MarketsandMarkets | Autonomous Driving Software Market Report 2024-2035 | MarketsandMarkets projects autonomous driving software from $1.8 billion in 2024 to $7.0 billion by 2035. |
| SM024 | Mordor Intelligence | Autonomous Car Market Size, Share, Trends Report Analysis 2025-2031 | Mordor estimates autonomous cars at $220.58 billion in 2026. |
| SM025 | Precedence Research | Autonomous Driving Software Market Size to Attain USD 8.04 Bn by 2035 | Precedence Research projects autonomous driving software at $2.97 billion in 2026 and $8.04 billion by 2035. |
| SM026 | Folio3 AgTech | Role of Robotics in Agriculture in Farming in 2026 | Folio3 describes farm robotics use cases including monitoring, spraying, harvesting and labor substitution. |
| SM027 | Mitsubishi Heavy Industries | Mitsubishi Heavy Industries and Preferred Networks Form Business Alliance | MHI and PFN announced a business alliance to jointly develop Japan-made AI technologies for mission-critical applications. |
| SM028 | Preferred Networks, Inc. | Mitsubishi Heavy Industries and Preferred Networks Form Business Alliance | PFN’s release says the MHI alliance will accelerate intelligence and autonomy of social infrastructure. |
| SM029 | Chugai Pharmaceutical | AI-leveraging drug discovery | Chugai describes AI-leveraging drug discovery and the MALEXA platform under its digital transformation program. |
| SM030 | Fierce Biotech | Chugai drops only AI-assisted antibody from pipeline but still holds high hopes for tech | Fierce Biotech reported Chugai discontinued an AI-assisted antibody while still expressing confidence in the technology. |
| SM031 | Matlantis | Matlantis | AI simulator for predicting atomic-level phenomena | Matlantis markets an AI simulator for predicting atomic-level phenomena. |
| SM032 | IDC | Charting the AI-driven future of manufacturing | IDC describes manufacturing AI adoption across automation, asset optimization, quality and supply-chain use cases. |
| SP001 | Preferred Networks | AI Chips - Business | PFN says its chips are optimized for faster, more efficient AI training and inference. |
| SP002 | Preferred Networks | Preferred Networks, Inc. | |
| SP003 | Preferred Networks | PFN Releases PLaMo-13B Open-Source Large Language Model in Japanese and English | |
| SP004 | NVIDIA | NVIDIA H100 GPU | H100 uses Hopper architecture, Transformer Engine, HBM, NVLink and confidential-computing features. |
| SP005 | NVIDIA | NVIDIA H200 GPU | |
| SP006 | NVIDIA | NVIDIA Blackwell Architecture | |
| SP007 | NVIDIA | NVIDIA Robotics Platform | |
| SP008 | NVIDIA | NVIDIA DRIVE AI Solutions | |
| SP009 | AMD | AMD Instinct MI300 Series Accelerators | |
| SP010 | Intel | Intel Gaudi AI Accelerator Products | |
| SP011 | Google Cloud | Tensor Processing Units | |
| SP012 | Cerebras | Product - Chip - Cerebras | |
| SP013 | Graphcore | IPU Processors | |
| SP014 | SambaNova | SambaStack Full-Stack Enterprise AI Platform | |
| SP015 | Boston Dynamics | Spot | Boston Dynamics | |
| SP016 | Skild AI | Skild AI | |
| SP017 | Figure AI | Figure | |
| SP018 | Sanctuary AI | Sanctuary AI | |
| SP019 | Physical Intelligence | Physical Intelligence | |
| SP020 | Covariant | Covariant | |
| SP021 | TechCrunch | Amazon hires the founders of AI robotics startup Covariant | Amazon hired Covariant founders and about a quarter of employees while licensing Covariant robotic foundation models. |
| SP022 | Waymo | Waymo - Self-Driving Cars - Autonomous Vehicles - Ride-Hail | |
| SP023 | Wayve | Wayve: Reimagining Autonomous Driving with Embodied AI Technology | |
| SP024 | Mobileye | Mobileye SuperVision | |
| SP025 | Woven by Toyota | Woven by Toyota | |
| SP026 | Sakana AI | Sakana AI | |
| SP027 | rinna | AIりんな | |
| SP028 | ABEJA | ABEJA | |
| SP029 | ELYZA | ELYZA | |
| SP030 | Recursion | Pioneering AI Drug Discovery | Recursion | |
| SP031 | Isomorphic Labs | Reimagining Drug Discovery Process with AI | |
| SP032 | Insilico Medicine | Main | Insilico Medicine | |
| SP033 | BenevolentAI | BenevolentAI | AI Drug Discovery | AI Pharma | |
| SP034 | Schrödinger | Computational Platform for Molecular Discovery & Design | |
| SP035 | Plenty | Indoor Vertical Farming | Plenty | |
| SP036 | FarmWise | Feeding Our World and Our Future | |
| SP037 | Carbon Robotics | Carbon Robotics: First & Only Commercial LaserWeeder | |
| SP038 | SemiAnalysis | Google Gemini Eats The World | SemiAnalysis discusses Google TPU/Gemini infrastructure as a differentiated hyperscaler AI compute stack. |
| SP039 | IEEE Spectrum | Intel’s Gaudi 3 Goes After Nvidia | |
| SP040 | GeekWire | Amazon hires Covariant founders, inks licensing deal with AI startup in latest reverse acquihire | |
| SI001 | Preferred Networks | PFN Raises Total of 19 Billion Yen in Latest Round | PFN today announced that it has raised a total of 19 billion yen in the first close of the latest equity financing round led by SBI Group combined with debt financing from financial institutions. |
| SI002 | Preferred Networks | PFN Raises Additional 5 Billion Yen in Extension Round | The extension round follows the initial funding in December 2024, bringing the total amount raised to date to 24 billion yen. |
| SI003 | The Bridge | Preferred Networks, AI Development Unicorn, Raises 19B Yen Including Debt Financing—SBI and Mitsubishi Corporation Among Investors | Preferred Networks announced on December 23 that it has raised 19 billion yen in its latest funding round. |
| SI004 | The Bridge | AI開発ユニコーンのPreferred Networks、デット含め190億円を調達——SBIや三菱商事ら参加 | 今回の調達を受けて、同社のこれまでの累積調達額は明らかになっている範囲で約360億円に達した。同社の時価総額は3,000億円を超えており |
| SI005 | MarketScreener | Preferred Networks, Inc. announced that it expects to receive ¥10 billion in funding from SBI Holdings, Inc. | Preferred Networks, Inc. announced that it expects to receive ¥10 billion in funding from SBI Holdings, Inc. |
| SI006 | Preferred Networks | SBI Holdings and PFN Agree to Form Capital and Business Alliance for Next-Generation AI Semiconductors | SBI Group plans to invest a maximum of 10 billion yen in PFN through SBI Holdings through a third-party allocation of new shares by the end of September 2024. |
| SI007 | SBI Holdings | 次世代AI半導体開発等に向けた資本業務提携に関する基本合意のお知らせ | |
| SI008 | Toyota Motor Corporation | Toyota to Make Additional Investment in Preferred Networks, Inc. | The investment will amount to 10.5 billion yen, and Toyota will acquire stock in PFN through the allocation of new shares to a third party. |
| SI009 | Preferred Networks | Preferred Networks received about 10.5 billion yen in investments from Toyota Motor Corporation | PFN agreed to receive an additional investment of approximately 10.5 billion yen from Toyota Motor Corporation. |
| SI010 | FANUC | Announcement for capital tie-up between FANUC CORPORATION and Preferred Networks Inc. | FANUC CORPORATION and Preferred Networks Inc. came to an agreement on capital tie-up. |
| SI011 | Preferred Networks | FANUC and Preferred Networks announce capital alliance | Amount of finance: 900 million JPY. |
| SI012 | Craft | Preferred Networks Company Profile - Office Locations, Competitors, Revenue, Financials, Employees, Key People, Subsidiaries | Total Funding $129.9 M Revenue ¥7.7 B FY, 2023. |
| SI013 | Latka | Preferred Networks Revenue 2024: $42M ARR, $2B Valuation | In 2024, Preferred Networks's revenue reached $42M. |
| SI014 | Growjo | Preferred Networks: Revenue, Competitors, Alternatives | Preferred Networks's estimated annual revenue is currently $49.5M per year. |
| SI015 | PremierAlts | Preferred Networks Valuation: $1.0B (2026) | Preferred Networks is currently valued at $1.0B as of June 30, 2025. |
| SI016 | AI Market Watch | Preferred Networks - AI Startup Profile | Revenue: ¥8,486 million (~$56M) as of FY ending Jan 2021; ~280-340 employees as of 2025-2026; valuation >300 billion yen (~$2B+). |
| SI017 | Preferred Networks | AI Chips - Business | AI development and usage currently depend heavily on general-purpose GPUs, but the rapid rise of generative AI is pushing these chips to their limits in performance, cost, power efficiency and availability. |
| SI018 | Preferred Networks | Computing Infrastructure - Business | Since 2024 PFN has also offered the Preferred Computing Platform (PFCP), a cloud-based service. |
| SI019 | Preferred Networks | PFN’s AI Processor MN-Core 2 Accepted to Hot Chips 2024 | PFN plans to provide MN-Core 2’s computing power to external parties in 2024. |
| SI020 | ServeTheHome | Preferred Networks MN-Core 2 for HPC and AI | The MN-Core 2 is focused on HPC and AI cluster tasks, and specifically power efficient compute. |
| SI021 | Preferred Networks | PFN Begins Development of Generative AI Processor MN-Core L1000 | PFN plans to market L1000 as the latest product in its proprietary MN-Core series of AI processors. |
| SI022 | Mitsubishi Corporation | MC Invests in Preferred Networks to Establish Strategic Alliance | MC is pleased to announce its investment in Preferred Networks, Inc. to establish a strategic alliance focused on AI. |
| SI023 | Mitsubishi Corporation | PFN, Mitsubishi Corporation and IIJ to Establish Joint Venture Preferred Computing Infrastructure for AI Cloud Computing | Scheduled to begin operations in early 2026, PFCI will primarily provide, operate and support customers of Preferred Computing Platform. |
| SI024 | Sumitomo Mitsui Trust Bank | 株式会社 Preferred Networks への出資について | PFN は現在、低消費電力の AI プロセッサー MN-Core シリーズや、国産生成 AI 基盤モデル PLaMo および幅広い領域の AI ソリューション・プロダクトの開発・販売を進めており |
| SI025 | METI / NEDO | Selection of 16 New Projects to support the development of AI Models under the GENIAC Computing Resource Provision Support Project (Cycle 4) | METI and NEDO will provide support to the 16 projects selected this time for the computing resources necessary for the development of AI models. |
| SI026 | SoftBank Corp. | SoftBank Corp. Announces Telco AI Cloud Vision to Build Social Infrastructure for the AI Era | Telecommunications operator is integrating GPU cloud, AI-RAN and software for AI data centers to evolve into an AI infrastructure provider. |
| SI027 | Preferred Networks | Company - Preferred Networks, Inc. | We are a company committed to vertical integration of computer science—from semiconductors and computing infrastructure to solutions and applications. |
| SI028 | Preferred Networks | Milestones and Awards - Company | April 2025 Forms a capital tie-up with companies including Kodansha, Sekisui House, TBS, Toei Animation, Sumitomo-Mitsui Trust Bank, Mitsubishi UFJ Trust Bank. |
| SI029 | Sacra | Preferred Networks funding, news & analysis | This report is for information purposes only and is not to be used or considered as an offer. |
| SI030 | RocketReach | Preferred Networks, Inc. Information | The Preferred Networks, Inc. annual revenue was $15.3 million in 2026. |
| SE001 | Preferred Networks | Preferred Networks corporate homepage | PFN vertically integrates the AI value chain from AI chips, computing infrastructure, generative AI, solutions and products. |
| SE002 | Preferred Networks | AI Chips - Business | Since 2016, PFN has been developing the MN-Core processor series with Kobe University. |
| SE003 | Preferred Networks | Computing Infrastructure - Business | |
| SE004 | Preferred Networks | Generative AI foundation models - Business | |
| SE005 | Preferred Networks | AI Products and Solutions - Business | |
| SE006 | Preferred Networks | Joint R&D with Toyota on Self-driving Cars | |
| SE007 | Preferred Networks | FANUC and Preferred Networks announce capital alliance | |
| SE008 | Preferred Networks | PFN is at Cisco Live! at San Diego | |
| SE009 | Preferred Networks | Preferred Networks releases ChainerX | |
| SE010 | Preferred Networks | Preferred Networks Migrates its Deep Learning Research Platform to PyTorch | |
| SE011 | Preferred Networks | Preferred Networks Deepens Collaboration with PyTorch Community | |
| SE012 | Preferred Networks | Preferred Networks develops a custom deep learning processor MN-Core | |
| SE013 | Preferred Networks | Preferred Networks’ MN-3 Tops Green500 List | |
| SE014 | Preferred Networks | Preferred Networks’ MN-3 Supercomputer Breaks Previous Record by 23.3% | |
| SE015 | Preferred Networks | PFN’s MN-3 Tops Green500 List for Second Time | |
| SE016 | Preferred Networks | PFN’s MN-3 Achieves 39.38 GFlops/W, Tops Green500 for Third Time | |
| SE017 | Preferred Networks | PFN Unveils Deep Learning Accelerator MN-Core 2 | |
| SE018 | Preferred Networks | PFN’s MN-Core Processor Now Powers Matlantis | |
| SE019 | Preferred Networks | PFN’s AI Processor MN-Core 2 Accepted to Hot Chips 2024 | |
| SE020 | Preferred Networks | PFN Begins Development of Generative AI Processor MN-Core L1000 | |
| SE021 | Preferred Networks | PFN Launches Flagship Japan-Made Large Language Model PLaMo Prime | |
| SE022 | Preferred Networks | PFN Establishes New Subsidiary Preferred Elements | |
| SE023 | Preferred Networks | PFN to Absorb and Merge with Subsidiary PFE | |
| SE024 | Preferred Networks | PFN Starts Joint Research with Toyota’s Frontier Research Center | |
| SE025 | Preferred Networks | PFCC Launches Matlantis Atomistic Simulator as Cloud-Based Service | |
| SE026 | Preferred Networks | PFCC Launches Matlantis in United States | |
| SE027 | Preferred Networks | PFN and ENEOS Release v7 of PFP for Matlantis | |
| SE028 | TOP500 | June 2020 Green500 list commentary | The most energy-efficient system on the Green500 is the MN-3, based on a new server from Preferred Networks. |
| SE029 | TOP500 | November 2021 Green500 list commentary | The system to claim the No. 1 spot for the Green500 was MN-3 from Preferred Networks in Japan. |
| SE030 | GitHub | chainer/chainer repository | |
| SE031 | GitHub | cupy/cupy repository | |
| SE032 | GitHub | optuna/optuna repository | |
| SE033 | GitHub | pfnet/pfio repository | |
| SE034 | Matlantis | Matlantis product site | |
| SE035 | Preferred Robotics | Kachaka product site | |
| SE036 | Hugging Face | pfnet model hub profile | |
| SE037 | Chainer Project | Chainer/CuPy v7 release and Future of Chainer | |
| SE038 | Chainer Documentation | ChainerX Documentation | |
| SE039 | Optuna Documentation | A hyperparameter optimization framework | |
| SE040 | CuPy Project | CuPy home page | |
| SE041 | Nature Communications | Towards universal neural network potential for material discovery | |
| SE042 | KDDI | KDDI GPU Cloud service page | |
| SU001 | Preferred Networks | Company - Preferred Networks, Inc. | PFN lists shareholders including Toyota Motor, Fanuc, Hitachi, Mitsui & Co., Mizuho Bank, NTT, Chugai Pharmaceutical, ENEOS Innovation Partners, and Hakuhodo DY Holdings. |
| SU002 | Preferred Networks | Business - Preferred Networks, Inc. | PFN says it co-creates with partners and provides AI solutions and products to diverse industries. |
| SU003 | Preferred Networks | PFN Starts Joint Research with Toyota’s Frontier Research Center | PFN and Toyota FRC started joint research to accelerate inference processing for physical AI research and development using MN-Core L series processors. |
| SU004 | Preferred Networks | Regarding additional investment by Toyota Motor Corporation | PFN clarified Toyota Motor shareholding after Toyota announced an additional investment. |
| SU005 | JCN Newswire via ADVFN | Toyota to Make Additional Investment in Preferred Networks, Inc. | Toyota agreed to invest 10.5 billion yen in PFN to accelerate AI R&D in mobility fields including automated driving. |
| SU006 | Preferred Networks | R&D alliance with FANUC Corporation | FANUC and PFN agreed on an R&D alliance using machine learning and deep learning for machine tools and robotics. |
| SU007 | Preferred Networks | FANUC and Preferred Networks announce capital alliance | FANUC and PFN reached a capital alliance agreement, with FANUC investing 900 million yen. |
| SU008 | Preferred Networks | PFN raises over 2 billion yen from FANUC, Hakuhodo DYHD, Hitachi, Mizuho Bank, and Mitsui & Co. | PFN allocated new shares to FANUC, Hakuhodo DY Holdings, Hitachi, Mizuho Bank, and Mitsui & Co. for over 2 billion yen. |
| SU009 | Preferred Networks | FANUC, Hitachi, and PFN establish Intelligent Edge Systems JV | FANUC, Hitachi, and PFN agreed to establish a JV to develop Intelligent Edge Systems using AI in edge devices. |
| SU010 | JCN Newswire via ACN Newswire | FANUC, Hitachi, and Preferred Networks to Establish a Joint Venture Company | The JV planned for April 2, 2018 would develop Intelligent Edge Systems for industrial and social infrastructure fields. |
| SU011 | ARC Advisory Group | Fanuc, Hitachi, and Preferred Networks to Establish a JV | ARC summarized the FANUC-Hitachi-PFN JV for intelligent edge systems. |
| SU012 | Preferred Networks | FANUC’s new AI functions utilizing machine learning and deep learning | FANUC, in collaboration with PFN, developed new AI functions for FA, ROBOT, and ROBO-MACHINE products. |
| SU013 | Preferred Networks | FANUC’s new AI functions that utilize machine learning and deep learning | FANUC developed and would release AI Servo Monitor and other AI functions in collaboration with PFN. |
| SU014 | Preferred Networks | PFN and ENEOS Release v7 of PFP for Matlantis | PFN and ENEOS released PFP version 7, the core technology powering Matlantis. |
| SU015 | ENEOS Corporation | PFN and ENEOS Release v7 of PFP Neural Network Potential | ENEOS and PFN state Matlantis is powered by PFP co-developed by the companies and now supports all naturally occurring elements. |
| SU016 | Matlantis Corporation | Company Profile | Matlantis | Matlantis Corporation lists its establishment, offices, and corporate identity for the simulator business. |
| SU017 | Business Wire | PFCC Launches Matlantis in United States | PFCC is described as a joint venture between Preferred Networks and ENEOS that provides Matlantis for AI-driven materials discovery. |
| SU018 | Chugai Pharmaceutical | Chugai Enters into Comprehensive Partnership Agreement with Preferred Networks | Chugai and PFN entered a comprehensive partnership agreement to apply deep learning and AI to innovative drug discovery. |
| SU019 | Preferred Networks | PFN raises capital from Chugai Pharmaceutical and Tokyo Electron | PFN agreed to receive about 700 million yen from Chugai Pharmaceutical as part of about 900 million yen in investments. |
| SU020 | NTT Communications | PFN Launches Private Sector Supercomputer | NTT Communications and NTT PC Communications supported PFN’s private-sector supercomputer with housing, networks, operations and technology. |
| SU021 | NTT DOCOMO Business | 導入事例 株式会社 Preferred Networks | NTT DOCOMO Business presents PFN as a customer case study for AI R&D infrastructure with high computing requirements. |
| SU022 | NTTPC Communications | Preferred Networks × NTTドコモビジネス × NTTPC use case | NTTPC describes a PFN GPU/supercomputer use case involving NTT DOCOMO Business and NTTPC Communications. |
| SU023 | KDDI Corporation | KDDI GPU Cloud | KDDI presents KDDI GPU Cloud and partner services for AI learning, big-data analysis, and R&D workloads. |
| SU024 | SoftBank Corp. | SoftBank to Launch AI Data Center GPU Cloud | SoftBank announced an AI Data Center GPU Cloud powered by Infrinia AI Cloud OS as part of its Neocloud business in October 2026. |
| SU025 | JR East | 線路内自律走行型ロボットによる線路点検を推進します | JR East announced promotion of track inspection using autonomous robots, naming Preferred Robotics in the joint development context. |
| SU026 | PR Times | Preferred Robotics、JR東日本と鉄道インフラの維持管理ロボットを開発 | Preferred Robotics announced development of railway infrastructure maintenance robots with JR East. |
| SU027 | Preferred Robotics | Kachaka Pro | Kachaka Pro is marketed as a compact AMR for efficient transport automation. |
| SU028 | Preferred Networks | GENIAC第2サイクルに継続採択 | PFN and Preferred Elements were selected for GENIAC Cycle 2, implemented with METI and NEDO cooperation. |
| SU029 | METI | Preferred Networks key people discuss generative AI development | METI describes PFN and PFE as GENIAC awardees building a 100B-parameter multimodal foundation model in Cycle 1 and smaller efficient models in Cycle 2. |
| SU030 | CNBC | Japan AI unicorn Preferred Networks has big plans in trucking, robots | PFN’s CEO said commercialization and practical launch can take three to five years after joint research begins. |
| SU031 | J-Startup / METI | Preferred Networks, Inc. | J-Startup | J-Startup describes PFN as founded in March 2014 with AI and AI control technology. |
| SU032 | Mitsubishi Heavy Industries | MHI and PFN Form Business Alliance | MHI and PFN entered a business alliance to jointly develop Japan-made AI technologies for mission-critical applications. |
| SU033 | Mitsubishi Corporation | Preferred Networksとの資本業務提携について | Mitsubishi Corporation subscribed to PFN third-party allocation and entered a capital and business alliance. |
| SU034 | Preferred Medicine | Preferred Medicine Announces ASCO 2021 Presentation | Preferred Medicine is described as a joint venture between PFN and Mitsui & Co. presenting joint research on machine-learning-based early cancer detections. |
| SU035 | Preferred Networks | MN-Core | PFN says MN-Core processors have been developed with Kobe University since 2016. |
| SU036 | Preferred Networks | PFN’s MN-Core Processor Powers Matlantis | PFN states MN-Core, co-developed with Kobe University, began powering PFP for Matlantis. |
| SU037 | Preferred Networks | Hakusensha and Hakuhodo DY Digital launch colorized manga products using PaintsChainer | Hakuhodo DY Digital launched colorized manga products with cooperation from PFN using PaintsChainer. |
| SU038 | Hakuhodo DY Holdings | Hakuhodo DY Holdings agrees capital and business alliance with PFN | Hakuhodo DY Holdings agreed to invest in and strategically partner with PFN for AI business development and implementation. |
| SU039 | Oisix ra daichi | Oisix ra daichi official website | Oisix ra daichi describes its food-oriented business, but the fetched official page did not corroborate a PFN relationship. |
| SR001 | Preferred Networks | Company — mission, strengths, leadership and AI governance | PFN says it is committed to vertical integration from semiconductors and computing infrastructure to solutions and applications. |
| SR002 | Preferred Networks | AI Chips — MN-Core series | MN-Core 2 products include MN-Server 2 with a standard price of 20 million yen excluding tax. |
| SR003 | Preferred Networks | PFN Announces Transition from Chainer to PyTorch | Chainer will move into a maintenance phase, and PFN will migrate its deep learning research platform to PyTorch. |
| SR004 | Preferred Networks | Industrial Automation Leaders Collaborate: Optimizing Industrial Production through Analytics | FANUC, Cisco and Preferred Networks provide enabling middleware platform software for the FIELD system. |
| SR005 | Business Wire | Preferred Networks Raises 6 Billion Yen through Third-Party Allocation to Toyota Motor Corporation | |
| SR006 | Business Wire | Toyota Motor Corporation and Preferred Networks to Collaborate on Service Robots | |
| SR007 | Woven by Toyota | About Us — mobility and technology | Woven by Toyota describes itself around mobility, technology, news and careers. |
| SR008 | Reuters | Japan AI startup Preferred Networks developing domestic AI chips | |
| SR009 | SemiAnalysis | Preferred Networks MN-Core 2 — A Japanese AI Accelerator | SemiAnalysis places MN-Core 2 in the AI accelerator market, a market dominated by larger accelerator ecosystems. |
| SR010 | SEC EDGAR / NVIDIA | NVIDIA Corporation Form 10-K for fiscal year ended January 26, 2025 | NVIDIA files public risk factors for a business with dominant AI accelerator supply, software ecosystem, and export-control exposure. |
| SR011 | Amazon Web Services | AWS Trainium — machine learning accelerator | AWS states customers including Anthropic, Databricks, Ricoh and Uber are realizing performance and cost benefits of Trainium. |
| SR012 | Google Cloud | Cloud TPU documentation | Google Cloud provides TPU resources, creation overview, pricing and support. |
| SR014 | U.S. Bureau of Industry and Security | Commerce strengthens export controls to restrict China advanced chip capabilities | BIS press releases document strengthened export controls on advanced computing and semiconductor manufacturing items. |
| SR015 | Center for Strategic and International Studies | Choking off China’s Access to the Future of AI | CSIS describes U.S. policy as exploiting chokepoints in AI chip design, EDA software and semiconductor manufacturing equipment. |
| SR016 | Japan Ministry of Economy, Trade and Industry | Update of Japan export-control measures for semiconductor manufacturing equipment | |
| SR017 | Artificial Intelligence Act EU | The EU Artificial Intelligence Act | The EU AI Act classifies AI systems by risk and creates obligations for providers and deployers. |
| SR018 | EUR-Lex | Regulation (EU) 2024/1689 laying down harmonised rules on artificial intelligence | |
| SR019 | International Organization for Standardization | ISO 10218-1: Robots and robotic devices — Safety requirements for industrial robots | |
| SR020 | International Organization for Standardization | ISO 10218-2: Safety requirements for industrial robot systems and integration | |
| SR021 | Japan Patent Office | AI-related inventions and patent examination information | |
| SR022 | Information-technology Promotion Agency, Japan | IT human resources and digital talent information | |
| SR023 | International Monetary Fund | Japan’s Weak Yen Has Limited Benefits | IMF analysis frames yen weakness as having limited benefits for Japan. |
| SR024 | CB Insights | Preferred Networks — products, competitors, financials, employees, headquarters | CB Insights lists Preferred Networks as having raised $308.23M and shows a Mosaic Score movement of -70 points in the past 30 days. |
| SR025 | Crunchbase | Preferred Networks company profile | |
| SR026 | PitchBook | Preferred Networks company profile | |
| SR027 | Japan Exchange Group | Listing on TSE — basic information for growth companies | |
| SR028 | GitHub | chainer/chainer repository | |
| SR029 | Wikipedia | Preferred Networks | |
| SR031 | Reuters | Nvidia forecast tests AI boom as investors worry about bubble | |
| SR032 | Preferred Networks | PFN at SC23 International Conference for High Performance Computing | PFN exhibited achievements and performance of the MN-Core series at SC23. |
| SR033 | Preferred Networks | PFN CEO Toru Nishikawa keynote at CEATEC 2018 | |
| SR034 | NVIDIA Developer | NVIDIA Hopper architecture in depth | |
| SV001 | Preferred Networks | Preferred Networks received about 10.5 billion yen in investments from Toyota Motor Corporation | PFN said it received about 10.5 billion yen in investments from Toyota Motor Corporation. |
| SV002 | Global Venturing | Toyota provides $95m to its Preferred Networks | Global Venturing reported Toyota provided $95m to Preferred Networks. |
| SV003 | Preferred Networks | SBI Holdings and PFN Agree to Form Capital and Business Alliance for Next-Generation AI Semiconductors | SBI Holdings and PFN agreed to form a capital and business alliance for next-generation AI semiconductors. |
| SV004 | Preferred Networks | PFN Raises Total of 19 Billion Yen in Latest Round | PFN announced it raised a total of 19 billion yen in the first close of the latest equity financing round led by SBI Group combined with debt financing. |
| SV005 | Preferred Networks | PFN Raises Additional 5 Billion Yen in Extension Round | PFN said it raised an additional 5 billion yen in equity financing through third-party allocation of shares. |
| SV006 | Preferred Networks | PFN Raises Additional Fund in Latest Extension Round | PFN announced an undisclosed additional equity financing in its latest extension round. |
| SV007 | The Bridge | Preferred Networks, AI Development Unicorn, Raises 19B Yen Including Debt Financing | The Bridge characterized Preferred Networks as an AI development unicorn and reported the 19 billion yen financing. |
| SV008 | UNDERCODE News | AI Development Firm Preferred Networks Faces 50% Drop in Valuation | The article framed Preferred Networks as facing a valuation decline, providing an adverse signal to test against primary evidence. |
| SV009 | CNBC | This Japanese AI unicorn has big plans to use deep learning to fix real-world problems | CNBC called Preferred Networks a Japanese AI unicorn and described trucking and robotics ambitions. |
| SV010 | J-Startup | Preferred Networks, Inc.|J-Startup | J-Startup lists Preferred Networks among selected Japanese startups. |
| SV011 | ENEOS | PFN and ENEOS Release v7 of PFP Neural Network Potential for Universal Atomistic Simulator | ENEOS and PFN released version 7 of the PFP neural network potential for universal atomistic simulation. |
| SV012 | Preferred Networks | Mitsubishi Heavy Industries and Preferred Networks Form Business Alliance | Mitsubishi Heavy Industries and Preferred Networks formed a business alliance in 2026. |
| SV013 | Japan Exchange Group | Overview of IPO | JPX says it takes about one year from kick-off to listing and the company needs audited financial statements for the two most recent years. |
| SV014 | U.S. Securities and Exchange Commission | NVIDIA Form 10-K fiscal 2025 | NVIDIA filed its fiscal 2025 Form 10-K with SEC financial data. |
| SV015 | U.S. Securities and Exchange Commission | AMD Form 10-K fiscal 2024 | AMD filed its fiscal 2024 Form 10-K with SEC financial data. |
| SV016 | U.S. Securities and Exchange Commission | Palantir Form 10-K fiscal 2024 | Palantir filed its fiscal 2024 Form 10-K with SEC financial data. |
| SV017 | Yahoo Finance | NVIDIA Corporation Valuation Measures & Financial Statistics | Yahoo Finance provides current valuation measures for NVIDIA. |
| SV018 | Yahoo Finance | Advanced Micro Devices Valuation Measures & Financial Statistics | Yahoo Finance provides current valuation measures for AMD. |
| SV019 | Yahoo Finance | Palantir Technologies Valuation Measures & Financial Statistics | Yahoo Finance provides current valuation measures for Palantir. |
| SV020 | Yahoo Finance | C3.ai Valuation Measures & Financial Statistics | Yahoo Finance provides current valuation measures for C3.ai. |
| SV021 | Yahoo Finance | UiPath Valuation Measures & Financial Statistics | Yahoo Finance provides current valuation measures for UiPath. |
| SV022 | Yahoo Finance | Fanuc Corporation Valuation Measures & Financial Statistics | Yahoo Finance provides current valuation measures for Fanuc. |
| SV023 | Yahoo Finance | CYBERDYNE Valuation Measures & Financial Statistics | Yahoo Finance provides current valuation measures for CYBERDYNE. |
| SV024 | Yahoo Finance | SenseTime Group Valuation Measures & Financial Statistics | Yahoo Finance provides current valuation measures for SenseTime. |
| SV025 | Anthropic | Anthropic raises Series E at $61.5B post-money valuation | Anthropic announced a Series E at a $61.5 billion post-money valuation. |
| SV026 | Reuters | OpenAI closes $6.6 billion funding haul with investment from Microsoft, Nvidia | Reuters reported OpenAI closed a $6.6 billion funding round at a $157 billion valuation. |
| SV027 | TechCrunch | OpenAI raises $6.6B and is now valued at $157B | TechCrunch reported OpenAI raised $6.6 billion and was valued at $157 billion. |
| SV028 | Reuters | Mistral AI raises 600 mln euros in latest funding round | Reuters reported Mistral AI raised 600 million euros in its latest funding round. |
| SV029 | Crunchbase News | Cohere Raises $500M At $5.5B Valuation | Crunchbase News reported Cohere raised $500 million at a $5.5 billion valuation. |
| SV030 | Reuters | Robotics startup Figure raises $675 mln from Microsoft, Nvidia, OpenAI | Reuters reported Figure raised $675 million from Microsoft, Nvidia, OpenAI and other investors. |
| SV031 | PR Newswire | Figure Raises $675M at $2.6B Valuation | Figure announced a $675 million Series B at a $2.6 billion valuation. |
| SV032 | Wayve | Wayve Raises Over $1 Billion Led by SoftBank to Develop Embodied AI | Wayve announced it raised over $1 billion in Series C financing led by SoftBank. |
| SV033 | TechCrunch | Wayve raises $1B led by SoftBank | TechCrunch reported Wayve raised $1 billion led by SoftBank. |
| SV034 | Forbes | Nvidia Joins Japanese Startup Sakana AI’s $100 Million Series A Round | Forbes reported Nvidia joined Sakana AI’s $100 million Series A round. |
| SV035 | Nikkei Asia | Japan’s Sakana AI worth $1.5bn in latest megabank fund raise | Nikkei Asia reported Sakana AI was worth $1.5 billion in its latest megabank fund raise. |
| SV036 | Amazon | An update on how we are accelerating the use of AI in robotics at scale | Amazon announced an update on using Covariant AI technology and talent to accelerate robotics at scale. |
| SV037 | Aswath Damodaran | Price to Sales Ratios by Sector | Damodaran publishes sector-level price-to-sales multiples used as valuation anchors. |
| SV038 | PwC | Global M&A industry trends: 2026 outlook | PwC provides a 2026 global M&A industry trends outlook. |
| SV039 | CB Insights | AI 100 2024 | CB Insights publishes an AI 100 list for private AI company benchmarking. |