初创公司尽调
尽调报告 Infrastructure / AI Data Centers / Climate Tech Growth / private unicorn 2026-07-02

Firmus Technologies

主权 AI 基础设施已获真正战略验证,但资本开支和披露风险沉重

作为私营 AI 基础设施平台,Firmus 拿到了少见的强战略验证;但当前估值已经预设多园区商业化会顺利跑通,而公开经济性还没成熟到可以激进承保。

封面要素

最新估值 01
1850 AUD M [CO009]
最新融资 02
330 AUD M [CO008]
旗舰园区 03
36000 GPUs [CO012]
Southgate 扩容路径 04
90MW by 2026; further 300MW planned [CO013]
成立时间 05
2019 [CO001]
总部 / 运营 06
Singapore HQ; Sydney registered office; Tasmania flagship build-out [CO002, CO003, CO006]

公司概况

Firmus Technologies 是一家创立于澳大利亚、总部位于新加坡的 AI 基础设施公司,建设并运营模块化、液冷的“AI 工厂”及配套云服务,服务 AI 训练、推理和 HPC 工作负载。公司的策略,是把主权算力需求同顾及可再生能源的选址、垂直整合的基础设施设计对齐;设计范围覆盖冷却、电力、编排和 GPU 云交付。

官网
firmus.co
成立时间
2019-01-01
创始人
Tim Rosenfield, Oliver Curtis, Jonathan Levee
创立地点
Australia
总部
Singapore
产品
面向高密度 AI 工作负载和主权部署优化的模块化 AI 工厂、GPU 云算力、裸金属集群、RDMA 存储和编排服务。
客户
亚太地区的 AI 原生初创公司、企业 AI 团队、研究人员,以及政府 / 主权算力用户。
商业模式
资本密集型基础设施平台,靠 AI 云服务、预留集群、主权 AI 工厂部署,以及伙伴牵头的算力容量协议变现。
阶段
Private unicorn / growth stage
融资情况
2025 年 9 月以 A$1.85 billion 投后估值完成 A$330 million 股权配售,Ellerston Capital 担任基石投资者,NVIDIA 参与。
[CO001, CO003, CO004, CO005, CO006, CO008, CO009, CO017]

执行摘要

主要优势

  • NVIDIA 和 Ellerston 为一家 APAC 私营 AI 基础设施初创公司提供了少见的战略和机构验证。
  • Tasmania、Singapore 和 Batam 拼出一个可信的主权算力版图,地点也贴合可再生能源敏感性或战略相关性。
  • 产品故事靠液冷、模块化 AI 工厂设计,以及从模型到电网的效率定位做差异化,而不只是泛化托管机房。

主要风险

  • 即便已获独角兽估值,收入、利用率、毛利率、客户集中度和股权结构条款仍未披露。
  • 多园区扩张取决于电力交付、许可和未来融资结构;这些安排可能让普通股经济性被次级化。
  • 存量玩家和资本更厚的新云竞争者,能凭更深资产负债表和更大装机基础进攻同一个 AI 基础设施机会。

未决问题

  • 完整股权结构条款、清算优先权,以及任何项目层面的优先融资仍不可得。
  • 当前收入、利用率和客户集中度数据未公开。
  • Southgate 阶段定义和后续扩张经济性,还需要一套标准化里程碑和 capex 材料包。

目录

Chapter 01

01公司概览

1.1 身份、产品与地理足迹

Firmus 应被理解为 AI 基础设施运营商,而不是普通托管机房提供商。从官网首页、基础设施页面到 AI 云材料,公司都反复把自己描述成 AI 工厂的垂直整合开发商和运营商,设计范围从芯片延伸到电网。它的运营主张把模块化高密度设施、液冷、编排软件,以及面向 AI 训练、推理和 HPC 工作负载的云服务放在一起。地理位置对这个论点很关键。官方材料把当前云运营和开发者访问放在新加坡,同时把公司注册和资本市场活动锚定在悉尼,把旗舰主权建设项目锚定在塔斯马尼亚。若把它读作一家创立于澳大利亚、总部位于新加坡的区域平台,这种组合是自洽的;但公开描述在不同来源之间仍有差异,因此报告应保留这种细微差别,而不是强行贴上单一司法辖区标签。塔斯马尼亚是主权算力叙事的核心,因为可再生电力、凉爽气候和政府支持在那里能够对齐;新加坡则是在线服务、参考工作负载和受监管区域需求的证明点。[CO001, CO002, CO003, CO004, CO005, CO006]

Firmus 快照 KPI 表
指标数值 / 状态日期置信度缺口 / 限制
成立时间20192019官方关于页面与 SmartCompany、DCD 报道相互印证
公司 / 总部口径总部位于新加坡的集团,在悉尼有注册办公室,塔斯马尼亚业务占比高2025-2026公开来源使用多个法律和地域标签;更适合视作多司法辖区运营足迹
最新披露融资A$330m 股权配售2025-09-16官方公告与多家独立报道一致
最新披露估值A$1.85b 投后2025-09-16源自 AFR 的独立报道与官方、ARN 报道一致
旗舰园区塔斯马尼亚北部的 Project Southgate2025-2026随着审批和供电推进,阶段细节仍在变化
一期容量信号2026 年达到 90MW,其中 1a 阶段 44MW,1b 后 90MW2025-2026官方来源同时使用 84MW 关键 IT 负载和 90MW 分阶段交付口径
Southgate 总路径两阶段 36,000 张 GPU;长期区域潜力 400MW2025-2026长期容量仍部分面向未来,且依赖审批
现有运营足迹新加坡 AI 云,加上澳大利亚 / 塔斯马尼亚建设2025-2026云和合作页面确认新加坡运营;塔斯马尼亚建设中
披露状态私营公司,未公开收入、ARR、客户数量或员工数2026必须从招聘、合作和基础设施承诺推断规模

表格保留当前最有支撑的事实,并明确区分硬披露指标与仍依赖审批、供电和客户爬坡的前瞻性建设信号。

[CO001, CO002, CO003, CO008, CO009, CO012]
FO002: Firmus 公司快照逻辑

Firmus 把主权选址、模块化基础设施、云服务和战略伙伴串成一个 AI 工厂论点。

[CO003, CO004, CO005, CO006, CO007, CO019]

1.2 创始人、领导层与治理信号

公开记录显示,Tim Rosenfield 和 Oliver Curtis 是最常露面的两位高管,材料反复称二人为联合 CEO;第三方报道还把 Jonathan Levee 列为联合创始人。SmartCompany 和 Data Center Dynamics 都把成立年份写作 2019,SmartCompany 还补充了公司早期围绕比特币挖矿散热起步、后来转向 AI 基础设施的背景。领导层在运营和技术领域释放的信号很强,但传统治理披露较弱。投资者沟通材料确认了悉尼注册办公室和股东文件流程,SmartCompany 报道称 Ellerston 投资总监 David Leslie 将在 2025 年融资后加入董事会。除此之外,董事会构成、投票控制权和保护性权利并未公开到足以绘制完整治理图的程度。尽调中还有一个声誉层面的细节:SmartCompany 提到 Curtis 曾在 2016 年被判内幕交易罪,时间早于 Firmus 成立多年。这并不否定基础设施论点,但意味着治理尽调不能停留在标准的创始人—市场匹配问题上。[CO001, CO002, CO010, CO026, CO027, CO028]

领导层和创始人表
人员职位背景 / 语境职能覆盖核心人物依赖
Tim Rosenfield联席 CEO / 联合创始人公司、政府和伙伴公告中最常见的发言人融资、政策定位、主权算力叙事、合作伙伴
Oliver Curtis联席 CEO / 联合创始人Project Southgate 和基础设施叙事的公开共同负责人基础设施建设、战略、投资者叙事、政府沟通
Jonathan Levee联合创始人独立报道将其列为创始团队成员创立背景和早期公司组建
David LeslieEllerston Capital 投资总监;据报道将加入董事会2025 年融资后被 SmartCompany 点名投资人监督和资本市场纪律
Toby Langley投资者关系总经理出现在投资者沟通和 2026 年发布中股东沟通和外部资本接口
Daniel Kearney首席技术官在 VAST 合作和产品架构材料中被引用模型到电网架构、数据层、系统设计

创始人能见度很高,但完整董事会构成、委员会结构和控制权并未公开披露。

[CO026, CO027, CO029, CO030, CO041]
利益相关方 / 投资人图谱
利益相关方角色控制权 / 经济重要性重要性尽调要求
Ellerston Capital2025 年 9 月融资的基石投资人主要机构背书,可能影响董事会锚定独角兽轮次和本地机构支持获取准确持股比例、董事席位条款和任何投资人保护
NVIDIA战略投资人和平台伙伴战略供给、生态和需求信号验证 GPU 路线图对齐和市场入口澄清排他性、配额权和未来硬件承诺
塔斯马尼亚政府项目和政策推动者非股权战略利益相关方支持分区、主权算力叙事和社区许可核验审批、土地状态和电力接入里程碑
ST Telemedia Global Data Centres 等数据中心运营商2023 年合资伙伴新加坡的平台和设施伙伴加速 SMC 启动和区域运营足迹确认当前经济性,以及 SMC 是否仍是新加坡主要运营模式
AI Singapore / 公共部门伙伴需求侧验证者参考客户和生态伙伴展示研究和主权算力可信度澄清合同期限、收入结构和重复使用经济性
现有私人支持方(Regal、Archibald、Tectonic、Waislitz/Pratt family)早期和 / 或持续股东潜在股权结构影响力显示澳大利亚资本网络深度要求完整股权结构,以及各轮次二级 / 一级融资组合

经济角色大方向清楚,但持股比例、优先权和否决权不公开。

[CO008, CO009, CO010, CO011, CO019, CO020]

1.3 资本基础、战略验证与里程碑

Firmus 记录最充分的里程碑,是 2025 年 9 月的融资。公司官方材料和多家独立媒体口径一致:公司完成了扩容后的 A$330 million 股权配售,Ellerston Capital 担任基石投资者,NVIDIA 参与,投后估值为 A$1.85 billion。资金用途不是抽象表述,而相当具体:位于塔斯马尼亚北部的 Project Southgate 被描述为一个 36,000-GPU 旗舰园区,分两个阶段建设,第一阶段目标是在 2026 年前后交付约 90MW,更大规模的后续扩建取决于审批。战略验证不止来自这轮融资。NVIDIA 不仅是投资者,还通过 DGX Cloud Lepton、基于 Spectrum-X 的架构,以及之后的 Batam 园区公告,成为云和平台伙伴;AI Singapore、HTX、MPA、STT GDC 和 VAST 则分别验证了产品栈或需求故事的不同部分。结果是,对于一家仍为私有的基础设施公司,Firmus 的伙伴阵容强于平均水平。与此同时,时间线推进很快,投资者应把已经验证的当前里程碑,同依赖执行、许可和供电交付的前瞻性园区主张分开看。[CO008, CO009, CO011, CO012, CO013, CO014]

里程碑表
日期事件类型金额 / 状态参与方含义
2019Firmus 在澳大利亚注册 / 创立创立已成立创始人包括 Tim Rosenfield、Oliver Curtis、Jonathan Levee开启 AI 基础设施平台故事
2023-06-22STT GDC 合作在新加坡推出 Sustainable Metal Cloud 合资项目合作战略合资项目已宣布STT GDC;Firmus让 Firmus 拥有实际的新加坡运营足迹和超大规模级主机伙伴
2024SemiAnalysis 和性能奖项开始出现在官方材料中规模外部验证Firmus / SMC / SemiAnalysis / DCD在重大融资前支撑技术可信度叙事
2025-03宣布与 AI Singapore 围绕 SEA-LION 和基准测试合作合作战略研究合作AI Singapore;Firmus形成研究和主权用途需求的公开证据
2025-05-27HTX 与 Firmus 签署 MoU,研究面向公共安全系统的可持续算力合作政府研究 MoUHTX;Firmus增加新加坡公共部门背书
2025-06塔斯马尼亚宣布 Green AI Factory Zone,并支持 Project Southgate监管园区已设立塔斯马尼亚政府;Firmus提升主权园区的社会许可和规划动能
2025-06-12Firmus 加入 NVIDIA DGX Cloud Lepton 市场合作Cloud Partner 身份NVIDIA;Firmus强化市场路径和区域 GPU 获取能力
2025-09-16Firmus 完成 A$330m 股权配售,投后估值 A$1.85b融资A$330m / A$1.85b 投后Ellerston;NVIDIA;其他澳大利亚投资人确认独角兽地位,并为 Southgate 建设提供资金
2025-12AI Singapore 案例研究发布 SEA-LION 部署结果规模32 个节点 / 256 张 H200 GPU;200+ 次实验AI Singapore;Firmus将合作转化为量化的工作负载证据
2026-02-24VAST 被选为主权 AI 工厂的 AI 操作系统数据层产品技术栈扩展VAST;Firmus表明产品向更大规模主权部署成熟
2026-06Firmus 宣布与 NVIDIA 合作的 170,000 张 GPU Batam 园区,规划至 2034 年规模360MW 园区;170,000 个加速器覆盖至 2027-2028 年Firmus;NVIDIA;DayOne显示塔斯马尼亚和新加坡之外的野心,也增加执行复杂度
2026-06Firmus 发布正式的澳大利亚能源和用水政策治理政策框架已发布Firmus建立可对照政府期待衡量的 ESG 和电网整合承诺
2026-07-02ABC 报道凸显 Southgate 周围的电力与就业争议反向公开质疑已被记录ABC;塔斯马尼亚政治利益相关方确认电网可用性和本地经济承诺仍是当前尽调问题

这条时间线混合了已验证的历史事件和尚在发展的扩张里程碑;后续园区阶段和公开上市时间应视作前瞻性内容,而非既定事实。

[CO001, CO008, CO009, CO012, CO013, CO019]
FO001: Firmus 公司里程碑时间线

公开里程碑显示,公司从基础 R&D 快速推进到新加坡证明点、塔斯马尼亚主权园区开发,以及独角兽融资事件。

[CO001, CO008, CO009, CO019, CO020, CO022]
FO003: Firmus 快照 KPI

最重要的公开指标强调融资额、园区规模和效率定位,而非成熟 SaaS 公司常见的业务牵引力披露。

阶段容量口径和效率主张混合了已披露里程碑与公司自定义基准;应把它们作为方向性证据,而非经审计运营 KPI。

[CO008, CO009, CO012, CO013, CO033, CO034]

1.4 披露缺口、身份摩擦,以及后续章节需要压测的事项

概览章节给下游尽调留下了几处重要缺口。公开材料没有披露收入、ARR、客户数量、利用率、毛利率、烧钱速度或债务结构;甚至员工规模也只能从当前招聘广度间接推断。总部措辞和未来资本市场计划也存在一些报道不一致,这进一步说明,投资者叙事必须和硬披露分开处理。最尖锐的身份问题并非财务,而是数字入口:用户提供的 firmus.ai 域名目前指向一个无关的施工文档 AI 产品,该产品现已归入 Bluebeam;AI 基础设施公司的活跃公开阵地则在 firmus.co 及相关 SMC 资产上。这种错配会给交易对手带来本可避免的混淆,也凸显了本报告为何必须做来源验证。最后,ABC 的报道有价值地注入了对塔斯马尼亚电力充足性和长期就业密度的怀疑。公司的能源和用水政策,方向上符合澳大利亚政府对 2026 年 AI 基础设施开发商的预期;但这些政策也已经形成一套可衡量的标准,执行表现将据此被检验。[CO003, CO018, CO026, CO030, CO031, CO032]

1.5 图表

Chapter 02

02市场分析

2.1 市场边界与现状替代品

Firmus 位于几个相邻市场的交叉处,因此宽泛的“AI 基础设施”标签不足以支撑估值工作。真正纳入的核心市场,是高密度 AI 工厂容量:为 AI 训练和推理优化的专用园区与云端交付的 GPU 容量,以及要求工作负载留在本辖区内的主权算力项目。这个边界包括实体园区、液冷和编排能力,以及与底层容量紧密绑定的 AI 云或 GPUaaS 层。它不包括普通企业托管机房、常规 SaaS 支出和商用半导体收入,因为这些资金池不能直接衡量 Firmus 正在试图销售的东西。 最重要的替代品不是小型初创公司,而是已经控制稀缺土地、公用事业接入和客户采购路径的既有超大规模云厂商和传统托管机房提供商。超大规模云厂商可以租赁、自建,或推出自有服务的主权版本;经典托管机房仍是许多不需要 AI 工厂级密度工作负载的默认外壳。因此,实际分析差异不在语义,而在架构:在 AI 工厂里,算力密度、液冷、编排和电网行为都是产品的一部分,而不是事后补丁。这个差异很重要,因为只有当买方足够重视这些特征、愿意离开现状时,Firmus 才能从更大市场扩张中受益。[CM001, CM002, CM003, CM004, CM045]

市场定义表
细分 / 类别纳入支出排除支出买方 / 付款方与 Firmus 的相关性
AI 工厂 / AI 数据中心面向训练和推理工作负载的高密度 AI 园区、预留容量、液冷与编排通用企业托管空壳容量和非 AI 算力机房主权项目、neocloud 厂商、模型开发商、受监管买方Firmus 宣称设计和电网行为有差异化,这是其核心实体市场
Neocloud / GPUaaS面向 AI 训练、尤其推理优化的云端 GPU 容量大宗化 IaaS 和无关开发者工具AI 原生初创、模型开发商、企业 AI 团队重要相邻收入层:比直接销售企业园区更快变现稀缺容量
主权算力 / 主权云绑定司法辖区的 AI 与云基础设施,本地控制数据、运营和治理缺少主权保障的跨境标准云政府、公共研究、关键基础设施、受监管行业Firmus 明确主打本土、政策对齐基础设施,战略契合度高
超大规模云和传统托管主要云和基础设施既有厂商已租用或自建的容量基础机房壳无法提供的专用 AI 工厂增值超大规模云厂商、大型业主、企业云买家更像主要替代品和对标对象,不是 Firmus 干净的可服务市场
企业本地部署 / 现状既有 IT 资产内的自建集群和渐进扩容区域共享基础设施和外部主权容量企业 IT、研究团队、业务线预算主要是买方无法证明迁移到外部 AI 工厂容量合理时的慢采用备选

边界逻辑把实体 AI 工厂容量、云变现层和主权项目,同宽泛托管或半导体池区分开,避免章节夸大可服务支出。

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

Firmus 真正相关的市场,从全球 AI 驱动基础设施增长,收窄到 APAC 供电受限、主权敏感的容量;Firmus 正是在这块容量上竞争。

各层刻意混用电力、MW 和收入视角。它们不能相加,应视为一层层更适合决策的过滤器,而不是一个算术漏斗。

[CM005, CM007, CM009, CM011, CM017, CM020]

2.2 用多重测算视角取代单一标题式 TAM

没有一个公开市场数字能干净描述 Firmus。最偏物理层的视角是用电量:IEA 分析把 2024 年数据中心需求放在约 415 至 460 TWh,到 2030 年约 945 至超过 1,000 TWh,且以 AI 为核心的设施增速快于品类平均。下一个视角是容量和 capex。JLL 预计 2026 至 2030 年全球将新增约 97 至 100 GW 数据中心容量,并把这轮建设定义为到 2030 年最高 $3 trillion 的房地产和租户合计投资;McKinsey 更宽的工业视角则达到约 $7 trillion。这些数字在规模方向上相互印证,但不能直接相互比较。 APAC 比全球总量数字更重要,因为 Firmus 是区域运营商。JLL 的亚太报告指出,到 2027 年将新增 4.8 GW 供给,其中 78% 已预租;DatacenterDynamics 报道称,2025 年区域开发管线为 19.4 GW,未来五到七年 APAC 托管机房建设 capex 约 $116 billion。Neocloud 和主权算力视角又不同。Gartner 较窄的口径意味着,到 2030 年,neocloud 将从 $267 billion 的 AI 云市场中拿到 20% 份额,对应约 $53 billion 收入;ABI 更宽的 GPUaaS 口径到 2030 年达到 $250 billion,Gartner 的主权云 IaaS 视角则在 2026 年已经达到 $80 billion。应把这些作为并行视角保留,而不是相加成一个虚假精确的 TAM。[CM005, CM006, CM007, CM008, CM009, CM010]

TAM/SAM/SOM 或规模测算视角表
发布方 / 视角年份地理范围数值CAGR / 增长方法置信度限制
IEA 电力需求视角2024-2030全球2024 年 415-460 TWh;到 2030 年约 945 至 >1,000 TWh基准情景到 2030 年年增长约 15%对数据中心总用电量建模衡量能源需求,不衡量收入或 Firmus 份额
JLL 全球容量视角2026-2030全球97-100 GW 新增容量到 2030 年 CAGR 约 14%基于 AI 和云增长的行业容量预测实体容量视角,不是客户收入
JLL / McKinsey 资本开支视角2030全球$3,000B 至 $7,000B 累计建设n/a房地产加租户配套视角,对照更广义的工业建设视角不同发布方的范围差异很大
JLL 亚太供给视角2027亚太4.8 GW 新供给;78% 已预租空置率预计大致保持在 6.5%-7.0%近期区域供给和预租展望供给视角几乎不说明终端客户付费意愿
DCD / Cushman 亚太管线视角2025亚太19.4 GW 管线(3.7 GW 在建;15.7 GW 规划中)2025 年新增 13.8 GW 运营容量区域管线和执行跟踪包含可能延期或融不到资的规划项目
DCD / Cushman 亚太托管资本开支视角2026-2031亚太12.45 GW 管线需要 $116B 建设投入5-7 年部署窗口托管专项资本需求估算仅托管;不含部分主权或自用建设
Gartner neocloud 视角2030全球隐含新云厂商收入约 $53B$267B AI 云市场的 20%新云厂商拿到的 AI 云收入份额口径比 GPUaaS 或基础设施资本开支更窄,只看服务收入
ABI 新云厂商 GPUaaS 视角2030全球$250B 收入机会到 2030 年,推理占收入 80%聚焦 GPUaaS 的新云厂商收入预测比 Gartner 的份额视角更宽,也更偏厂商口径
Gartner 主权云视角2026全球$80B 主权云 IaaS 支出较 2025 年同比增长 35.6%IaaS 支出预测仅限主权 IaaS;不等同于实体 AI 园区收入

各行刻意不相加。它们保留了能源、MW、资本开支、新云厂商收入和主权云支出之间彼此矛盾但有用的测算视角。

[CM005, CM007, CM008, CM009, CM010, CM011]
FM002: 市场估算区间

围绕 Firmus 的非可加 $B 视角区间,保留定义差异,而不是压成一个 TAM 标题数字。

所有行都用 $B 单位,但混合了资本开支、服务收入和支出池。这是有意为之,因为公开来源没有提供一个可比的 Firmus 市场总量。

[CM008, CM015, CM017, CM018, CM020, CM021]

2.3 买方、用户与付款方分层

买方地图很碎片化。AI 原生初创公司和模型开发者往往先像紧迫用户,再像纪律严明的付款方:他们要的是稀缺 GPU 访问、低摩擦部署,以及愿意比超大规模云厂商动作更快的供应商。企业 AI 团队是大算力消费者,但通常通过中央云或 IT 预算采购,这意味着像 Firmus 这样的园区运营商,往往要通过 neocloud 或基础设施伙伴间接触达。政府、公共研究机构和关键基础设施组织又是另一类买方,因为主权、司法辖区和可审计性可以和原始吞吐量同样重要。 超大规模云厂商扮演着不寻常的双重角色。它们通过吸收巨额 capex、教育客户把算力当作战略投入,验证了这个品类;但它们也会压缩可触达市场,因为它们预租容量、自建园区,并推出自有主权版本。区域运营商由此留下一个现实的早期采用走廊:主权项目、需要本地控制的受监管买方、服务 AI 原生需求的 neocloud 或基础设施伙伴,以及那些更看重 APAC 已批准容量、而不是超大规模云区域最低单位成本的工作负载。Firmus 符合这条走廊,但前提是它能把基础设施设计优势转化为合同,而不只是叙事上的相邻性。[CM023, CM024, CM025, CM026, CM027, CM038]

细分客群 / 买方图谱
客群买方用户付款方工作流预算负责人采用触发因素
AI 原生创业公司 / 模型开发者创始人、基础设施负责人或模型平台负责人ML 工程师和平台团队类资本开支的基础设施支出或已承诺云预算突发训练、推理服务、快速迭代基础设施或平台预算负责人快速拿到稀缺 GPU,并愿意试用非传统供应商
企业 AI 团队CIO、CTO 或云平台负责人数据科学、MLOps 和应用团队中心化云、IT 或转型预算Copilot、内部 LLM、数据管道和推理密集型应用中心 IT / 云 FinOps 负责人需要容量、数据本地化,或比超大规模云默认方案更低的实际单位成本
政府 / 公共研究数字部门、研究机构或项目发起方研究人员、政策实验室和公共服务团队公共预算或项目资金国家 AI 能力、公共研究和安全模型开发政府项目负责人司法辖区控制、韧性和本地能力建设
受监管行业行业 CIO、风险负责人或基础设施发起人合规、数据和 AI 应用团队IT、风险或业务线预算带主权或驻留要求的敏感数据处理行业平台负责人需要可审计的本地控制,而不是最低成本的通用云
超大规模云厂商和基础设施伙伴云平台团队或机房托管采购团队基础设施工程和部署团队大规模资本开支和长期租赁计划园区扩张、合作伙伴转售或主权变体基础设施资本开支委员会需要已获批且可规模化的土地和电力

同一份容量可能被完全不同的用户和付款方组合消耗。Firmus 的采用路径取决于谁掌握预算,以及谁最先感到电力或主权痛点。

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

买方细分的差异,不在于原始 AI 兴趣有多强,而在于预算由谁掌握、主权有多重要,以及通往 Firmus 的路径是直连还是渠道带动。

单元格是有证据支撑的定性标签,不是数值评分,因为公开来源没有披露 Firmus 特定的买方转化数据。

[CM023, CM024, CM025, CM026, CM027, CM038]

2.4 增长驱动、约束与绿色准入溢价

增长逻辑很强。AI 落地、云采用和数字化正在扩大 APAC 整个品类;以推理为主的生产工作负载正在成为主导设计点;主权云需求在希望获得更多数字独立性的地区增长最快。这些驱动因素有利于能快速把新容量推向市场、又能让公用事业公司和监管机构接受的运营商。因此,从市场方向看,Firmus 的区域论点说得通。 问题在于,采用受制于物理瓶颈,而不是兴趣不足。电力可得性是第一道筛选,报道中的电网等待时间从一些新兴市场约两年,到核心市场超过八年不等。约 100 kW 的 AI 机架迫使液冷和更重的机械设计成为必要条件;变压器、涡轮机、先进芯片和相关组件仍然紧缺。与此同时,租金上涨、空置率低,买方在锁定定制容量前仍需要对利用率有信心。因此,市场所谓的绿色溢价不应被理解成普遍涨价。新加坡和澳大利亚都明确把效率、电网行为、用水和社区匹配纳入审批或优先排序。对 Firmus 这类公司而言,溢价更可能体现为排队准入和政策兼容性,而不是立刻获得定价权;这一区别对估值很关键。[CM028, CM029, CM030, CM031, CM032, CM033]

增长驱动因素与约束表
驱动因素 / 约束方向时间影响尽调问题
亚太 AI 落地、云采用和数字化正向当前 / 中期把总机会扩展到单一国家或单一买方类别之外在每个目标地区,把真实已承诺需求与宽泛数字化转型话术拆开
推理密集型生产 AI 工作负载正向当前 / 中期能交付高密度、低延迟、常时在线容量的运营商受益,而不是只做一次性训练集群向早期客户索取真实生产工作负载结构和配套采用率
主权云和本地化要求正向当前 / 中期为辖区内基础设施和本地治理保证创造需求厘清买方需求来自法律、采购政策,还是内部风险偏好
电力可得性和并网延迟负向当前 / 结构性把价值推向已有获批电力路径的场址和运营商逐场址取得供电时间表、排队位置和备用方案
散热密度和用水审查负向当前 / 结构性抬高建设复杂度,也让效率主张具备商业实质验证实载下实测 PUE、用水量和散热表现
变压器、涡轮、芯片和设备供应链瓶颈负向当前 / 2027即便需求和许可都在,也可能拖慢交付判断部署节奏前,梳理长周期物料和供应商集中度
绿色政策和国家利益审查契合者正向 / 不契合者负向当前 / 结构性把效率和社区契合度变成审批筹码,而不是可有可无的品牌包装测试 Firmus 的承诺是写进合同,还是只停留在营销
租金上升、空置率低和资本开支强度负向当前 / 结构性买方接受定制容量经济性前,仍需要利用率信心索取客户批次级利用率、合同期限和续约数据

市场需求充沛但执行受限。驱动因素和约束同时起作用,所以已获批 MW 的获取能力和利用率证明,比宽泛的品类热度更重要。

[CM028, CM029, CM030, CM031, CM032, CM033]
FM004: 采用漏斗或价值链地图

采用路径从急迫的 AI 容量痛点出发,穿过审批和电力门槛,最终落到经常性工作负载;因此,容量可得性比抽象 TAM 更关键。

[CM029, CM031, CM032, CM034, CM035, CM039]

2.5 尽调缺口与定义敏感的矛盾

最大的未解问题不是品类是否存在,而是 Firmus 能否捕获足够份额。公开证据支持方向——电力受限的 AI 需求、APAC 外溢、主权算力兴趣,以及政策对高效建设的偏好——但没有披露自下而上测算 SOM 所需的商业细节。客户组合、合同期限、在线利用率、实际定价和扩张权仍是私有信息。缺少这些输入,市场分析只能界定机会边界,不能证明份额。 市场也没有一个被普遍接受的定义。有些来源衡量用电量或 MW 需求,有些关注云服务收入,有些看主权 IaaS,还有些看累计基础设施 capex。这不是本章的缺陷,而是市场真实的分析状态。正确的尽调反应,是保留彼此矛盾的视角,要求公司提供内部管线和利用率数据,并避免把宽泛的全球 AI 基础设施大标题当成 Firmus 可触达收入池。[CM022, CM037, CM039, CM040]

2.6 图表

Chapter 03

03竞争者

3.1 格局与替代层

Firmus 面对的不是一个干净的单一同业集合。买方可以用超大规模云厂商的 GPU 云、能托管或拼接私有 AI 环境的 AI 就绪机房业主、把 GPU 访问和软件打包的专业 neocloud,或由最大型项目自行建设,来完成同一项任务。因此,竞争压力来自任何最能缓解获得容量时间、司法辖区控制和运营确定性痛点的替代方案。客户愿意留在既有云关系里时,超大规模云厂商会赢;土地、电力和互联成为稀缺输入时,Equinix、Digital Realty、AirTrunk、NEXTDC、Keppel 和 GDS 很重要;买方想要不必等待定制园区的 AI 原生栈时,CoreWeave、Lambda 和 Crusoe 很重要。因此,评判 Firmus 时,不应把它看成独立托管机房,而应把它看成一个区域性 AI 基础设施整合商,试图在更大类别的替代品面前守住更窄的主权楔子。[CP001, CP002, CP015, CP036, CP037, CP042]

FP001: 竞争定位地图

物理主权控制相对于分销与生态力量的序位图。

坐标轴评分是基于足迹、包装、资本和生态信号综合出的有证据支撑序位判断,而非经审计的市场份额数据。

[CP001, CP002, CP015, CP019, CP025, CP031]

3.2 超大规模云厂商划定外层竞争边界

AWS、Google Cloud 和 Azure 划定了外层竞争边界,因为它们已经提供集群级 GPU 基础设施、广泛区域覆盖,以及大多数买方默认信任的采购关系。AWS 推出带 UltraClusters 和液冷效率主张的 H100、H200 P5 系列;Google 把加速器优化的 A-series 机器、全球区域和基于承诺的定价结合起来;Azure 将 ND H100 和 A100 系列同广泛地理版图和明确主权选项配套。Microsoft 自己的 AI 工厂叙事也很重要,因为它表明最大的云厂商不再只提供通用算力,而是在用巨额 capex 建设专用 AI 园区。这意味着 Firmus 的可触达市场只剩那些足够看重物理主权、本地能源姿态或定制部署、愿意离开默认选项的工作负载。没有这块楔子,超大规模云厂商捆绑的相邻价值太多,难以撼动。[CP002, CP003, CP004, CP005, CP006, CP007]

功能 / 能力矩阵
购买标准Firmus超大规模云厂商(AWS/GCP/Azure)Equinix / Digital RealtyAirTrunk / NEXTDCCoreWeave / Lambda / Crusoe
公开 GPU 云服务有限 / 伙伴主导否 / 有限
亚太主权选址叙事Tasmania/Singapore 叙事强混合:区域内云强,定制园区控制较少通过本国设施达到中等混合;取决于地区
公开披露高密度 AI 散热
互联生态中等很强中等有限到中等
公开挂牌价部分部分到强
合规与驻留覆盖宽度初现很强中等中等到强
锚定客户 / 资本信号初现很强很强强但分化
报价制定制园区方案有限部分已承诺交易

单元格仅总结公开证据。“部分”表示存在部分价格或能力证据,但不足以做同口径经济性对比;“否”通常意味着报价制或未披露。

[CP002, CP006, CP009, CP012, CP013, CP019]
FP002: 功能广度 / 能力地图

按组展示 Firmus 与主要竞争类别的能力。

单元格有意只概括公开证据;“混合”和“部分”表示披露确实存在,但不能直接对比。

[CP002, CP012, CP019, CP029, CP032, CP034]

3.3 机房业主与主权容量既有玩家

机房业主类别在结构上不同于超大规模云厂商,但对 Firmus 同样危险。Equinix 和 Digital Realty 把 AI 就绪设施同大型互联生态配对,让客户不依赖较小区域运营商,也能组装私有、混合或主权 AI。AirTrunk 和 NEXTDC 在物理论点上更接近:两者都在加码高密度 APAC 容量,也都能凭更大的资产负债表和更成熟的客户入口,销售主权或区域受控基础设施的叙事。Keppel DC REIT 和 GDS 虽然不那么以开发者为中心,但仍然重要,因为它们在本可供给区域新进入者的市场里,拥有或融资了大量既有数据中心容量。实际竞争中,这意味着 Firmus 拼的不只是冷却设计,还要争夺稀缺土地、电力和互联的控制权;而这些市场里的既有玩家已经很大,并且越来越懂 AI。[CP011, CP012, CP013, CP014, CP015, CP016]

竞争对手画像表
竞争对手类别规模 / 资本信号目标客群差异化局限
AWS超大规模云厂商39 个区域 / 123 个 AZ;H100/H200 UltraClusters全球企业、模型开发者、受监管买方深云服务捆绑,加上面向 AI 的 GPU 机群物理主权以云为中心,而不是以园区为中心
Google Cloud超大规模云厂商43 个区域 / 130 个可用区;A4/A3 加速器系列全球 AI 开发者和企业强全球网络,加上公开定价工具实体场址定制不如云产品包装显性
Microsoft Azure超大规模云厂商广泛地理覆盖,加上专用 AI 数据中心资本开支大型企业、OpenAI 邻近生态、受监管工作负载强采购能力、驻留选项和 H100 横向扩展默认落到 Azure,会压缩区域运营商空间
EquinixAI 就绪型机房业主 / 互联既有龙头280 个数据中心;10,500+ 客户;507,000+ 互联混合多云、私有 AI、全球企业互联市场和 AI 就绪密度自身不是 AI 原生云产品
Digital RealtyAI 就绪型机房业主 / 私有 AI 平台遍布 55+ 都会区的 300+ 数据中心私有、混合和主权 AI 买方PlatformDIGITAL 和经伙伴验证的 AI 方案定价和准确 AI 套餐经济性多为报价制
AirTrunk亚太超大规模机房业主Blackstone 牵头 A$24b 交易;覆盖 APME 的超大规模平台亚太 / 中东全球云和超大规模客户区域规模,加上深厚资本背书公开软件、定价和工作负载工具不透明
NEXTDC澳大利亚主权 AI 机房业主FY25 收入 A$427m;S4 350MW;S7 550+MW澳大利亚主权 AI、超大规模云厂商、企业本土主权叙事,加上高密度液冷设计仍以报价制为主,更像机房业主而非 AI 云
Keppel DC REIT区域组合持有者10 个国家 25 个数据中心;AUM 折合约 US$6.3b长期超大规模和企业需求资产负债表能力覆盖亚太 / 欧洲枢纽集成云或开发者动作证据较少
GDS Holdings中国既有运营商FY2025 收入 RMB11.43b;利用率 75.5%中国企业、超大规模、托管云买方装机基础和中国布局公开定位不是 AI 原生云
CoreWeaveAI 专用新云厂商US$60.7b RPO;OpenAI / Meta 大额承诺AI 实验室、前沿模型开发者、大型企业 AIAI 原生云,具备规模和锚定合同客户高度集中,并暴露于超大规模云捆绑
LambdaAI 专用新云厂商公开挂牌价和 16 到 2,000+ GPU 集群包装开发者、研究团队、企业 AI 开发者透明包装,加上模块化 AI 工厂设计生态和地理足迹小于超大规模云厂商
Crusoe电力优先的 AI 基础设施云1.2GW Abilene 一期;3.0GW 在建项目能源密集型 AI 建设和快速部署客户电力编排和垂直整合地理覆盖窄于全球云

各行聚焦与 Firmus 决策最相关的竞争者和替代方案,而不是列举所有数据中心业主;局限来自公开证据约束,不是完整产品拆解。

[CP011, CP013, CP015, CP017, CP021, CP023]

3.4 AI 专业 neocloud 与全栈同业

如果问题是,谁最像 Firmus 集成化抱负的放大版本,答案不是 Equinix 或 AirTrunk,而是 CoreWeave、Lambda 和 Crusoe。CoreWeave 把 AI 原生 云交付同足以重塑供给和融资决策的大客户承诺结合起来,但它的 10-K 也显示了代价:规模伴随着沉重的客户集中度,以及对超大规模云厂商捆绑的直接暴露。Lambda 规模较小,但信息量很高,因为它公开小时价格、使用模块化 AI 工厂语言,并提出明确的液冷和合规主张,比按报价销售园区的运营商更容易比较。Crusoe 又不同:它的边缘在于电力编排,以及横跨云、数据中心建设和电气制造的垂直整合。合在一起,这些公司展示了全栈 AI 基础设施竞争者在同时解决供给入口、产品包装和部署速度时会长成什么样。[CP025, CP026, CP027, CP028, CP029, CP030]

定价 / 包装对比
供应商 / 类别公开套餐公开价格信号合同模式哪些仍不透明Firmus 含义
AWS P5 / P5e / P5enEC2 和 SageMaker 内的按需 GPU 实例引用页面披露相对节省和能力,但没有简单通用的集群挂牌价按用量计费云,加上企业承诺实际折扣、预留容量经济性和主权包装没有私下报价,很难判断直接价格能否持平
Google Cloud Compute EngineGPU VM、Spot、持续使用以及 1 年或 3 年承诺公开页面给出明确定价框架和折扣机制按用量计费,可选择承诺分区域实际价格和 GPU 预留经济性为比较云替代方案的买方设定公开基准
Azure ND 系列带规模集和 InfiniBand 集群的 GPU VM现有来源没有清晰挂牌价按用量计费和企业合约按地区和期限拆分的 H100 实际经济性即使价格透明度较弱,Azure 也能靠捆绑赢单
Lambda实例、1-Click Clusters 和 SuperclustersB200 $6.69/hr;H100 $3.99/hr(小时价格)自助服务加预留容量区域可用性和企业折扣同业中最透明的 AI 云价格信号
CoreWeave / Crusoe 已承诺交易AI 原生云,加上长期容量合同公开规模信号,但没有挂牌价多年 take-or-pay 与按需组合单位价格、最低承诺和利润率结构更接近 Firmus 经济性,但仍以私下交易为主
AI 就绪型机房业主 / 主权园区私有 AI、托管和定制园区多为报价制定制合同、MW 承诺和伙伴主导包装每 MW 价格、最低期限、包含的云软件和利用率假设Firmus 竞争的定价层最不透明

本表把挂牌价证据与合同经济性拆开。公开可见价格主要是云式方案;主权园区和许多新云厂商已承诺交易仍不透明。

[CP003, CP006, CP026, CP030, CP038]

3.5 定价、锁定与护城河耐久性

竞争经济性异常不均衡。公开价格发现最清楚的是云式产品——Lambda 发布标价,Google 记录 Spot 和承诺机制,AWS 发布性能和相对成本主张——而大多数 neocloud 承诺交易,以及几乎所有主权园区或机房业主报价,都仍基于报价。这种不透明更有利于拥有成熟采购团队的既有玩家,而不是新运营商。锁定也会非对称累积:超大规模云厂商受益于账单、安全和数据重力关系;Equinix 和 Digital Realty 受益于生态;AirTrunk 受益于资本和客户覆盖;CoreWeave 展示了多年照付不议 合同如何把 AI 云嵌进去。因此,Firmus 有真实但狭窄的护城河。当买方需要 APAC 物理控制、定制能源或冷却设计,以及本地执行时,护城河最强;当买方主要想要 GPU 供给、快速签约和熟悉的采购路径时,护城河最弱。被替代的风险不是假设——它已经嵌在市场内竞争对手的资本密度和分销优势里。[CP028, CP033, CP034, CP038, CP039, CP040]

护城河耐久性 / 竞争风险登记表
护城河主张威胁向量严重性证据缓解措施 / 尽调问题
亚太主权选址超大规模云厂商已提供区域内云和广泛驻留地图AWS、Google 与 Azure 的覆盖范围远大于 Firmus 已披露布局要求列出必须物理控制的具体工作负载,而不只是区域内云服务
节能型 AI 园区更大对手也在披露液冷和能效计划AWS、NEXTDC、Lambda 与 Crusoe 都在营销液冷 AI 算力要求证明 Firmus 在能源经济性或许可结果上显著更好
从芯片到电网的一体化栈Neocloud 同行已把云产品封装和基础设施结合起来CoreWeave、Lambda 与 Crusoe 都在销售一体化 AI 基础设施叙事验证 Firmus 是否掌握足够软件控制权,避免只做机房房东
借头部背书方获取资本对手资产负债表或合同积压规模更大AirTrunk 背后有 Blackstone/CPP;CoreWeave 披露 US$60.7b RPO要求提供 Firmus 的硬件分配权和已承诺融资文件
通过合作伙伴分发既有厂商掌握采购通道和互联生态超大规模云厂商、Equinix 与 DLR 都更贴近企业既有采购路径确认 Firmus 能否借合作伙伴渠道放量,同时守住经济性
按报价定制部署不透明定价既可能藏着优势,也可能掩盖短板多数园区和承诺型交易定价不公开获取真实客户报价方案和折扣梯度
主权计算叙事各家对手的政策话术越来越趋同NEXTDC、DLR 与云厂商都在使用主权或区域内控制话术要求看到已签主权合同,而不是营销文案
与供应商及新进入者的距离NVIDIA 和超大规模云厂商可以继续下探技术栈Microsoft AI 工厂、AWS UltraClusters 与 NVIDIA 参考栈都已公开压力测试:如果供应商变成直接替代方案,Firmus 差异化还剩多少

严重程度是作者基于规模、分销和供给不对称作出的判断,并非上市公司风险评级。

[CP028, CP034, CP035, CP038, CP039, CP040]
FP003: 护城河 / 就绪度 KPI

几项公开指标显示,主要对手规模已经更大、分布也更广。

这些 KPI 不是评分,而是公开参照点,显示对手的规模、分布或定价可见度,供尽调团队与 Firmus 的非公开披露对照。

[CP004, CP011, CP015, CP017, CP026, CP030]

3.6 图表

Chapter 04

04财务

4.1 收入模型与定价不透明

公开证据足以勾勒 Firmus 的收入入口,但不足以给它们定价。AI Cloud Compute 围绕 H200 级系统提供按需实例和预留集群;Bare Metal 增加可预留的专用单租户或多机架 GPU 集群;Cloud Services 在其上叠加编排、托管 Slurm、CUDA 栈、可观测性和混合连接。这不是一个 SKU。它至少是三层商业栈:云访问、预留基础设施和托管运营。缺失的关键层是商业具体性。受审阅的 Firmus 页面都没有发布每 GPU-hour、每集群或托管服务标价,所有页面都把买方导向询价或预留表述。这意味着本章可以描述活动应如何转化为收入,但无法判断实际组合:使用量收入、最低承诺预留容量、专业服务或支持。健康的解读是,这是一种借 NVIDIA DGX Cloud Lepton 获得部分市场分销的企业基础设施收入。保守解读则是,产品包装领先于公开商业披露。承销时,正确姿态既不是“没有商业模式”,也不是“清晰的 SaaS 定价”,而是“可信的变现入口,但实际定价未解”。[CI001, CI002, CI003, CI004, CI005, CI006]

收入来源表
收入来源机制单位 / 依据当前公开状态收入质量判断尽调问题
AI Cloud Compute面向 AI/HPC 工作负载的按需实例和预留集群按实例、集群预留或工作负载消耗计费产品公开,实际成交价格未披露需求表面看起来真实,但变现率不透明提供每 GPU 小时价格、最低预留要求和实际混合 ASP
专用裸金属集群预留的单租户或多机架 GPU 集群按预留节点 / 集群期限计费预留条款表述公开,合同经济性未公开ACV 可能更高,但销售周期更长、交付负担更重提供合同期限、安装费和取消条款
托管云服务AIFactoryOS、托管 Slurm、CUDA 栈、可观测性、混合连接按托管环境、支持等级或捆绑服务计费能力公开,商业化条款未公开可能形成经常性支持 / 服务层,但附着率未知披露托管服务定价,以及其在基础设施交易中的附着率
市场平台撮合容量DGX Cloud Lepton 将买家导向 Firmus 区域 GPU 容量通过市场平台路径获取按需或长期容量分发路径公开,NVIDIA 与 Firmus 之间的经济安排未公开可以拓宽管线,但市场平台抽佣率和组合未知厘清渠道经济性、收入分成,以及客户关系归属
标杆合作牵引的主权 / 公共部门工作负载通过标杆合作获取科研、企业和政府工作负载按合同或项目计费用例公开,合同金额未公开能增强可信度;也可能需要定制,并拉长采购周期提供科研、企业、政府和合作伙伴渠道的收入拆分

各行只列出可见变现路径。「当前公开状态」指披露状态,不代表收入表现;实际价格、折扣和渠道经济性仍未公开。

[CI001, CI003, CI005, CI007, CI009, CI010]
定价 / 商业化表
产品公开价格 / 单位标价 vs 实际价格已知信息未知信息来源视角
AI Cloud Compute实际价格未知按需和预留访问已公开,规格为 H200 级,并配有可观测性工具小时价格、承诺使用折扣和最低预留期限Firmus AI Cloud Compute 页面
Bare Metal 集群实际价格未知专用集群可预留,并提供 24/7 运营支持集群日定价、安装费和支持加价Firmus Bare Metal 页面
Cloud Services独立销售还是捆绑定价未知AIFactoryOS、托管 Slurm、CUDA 栈和混合连接已公开服务是单独计费、捆绑销售,还是某些合同的必选项Firmus Cloud Services 页面
DGX Cloud Lepton 路径渠道经济性未知市场平台支持按需和长期区域容量收入分成、抽佣率、账单归属和支持义务NVIDIA DGX Cloud Lepton 页面
标杆 / 主权项目很可能采用定制定价AI Singapore 和公共部门案例证明能力及区域化交付合同金额、预付款结构、SLA,以及是否涉及补贴或拨款Firmus 案例研究和新加坡报道

价格单元格为 null 表示价格未公开。本表区分可见产品封装与未知的实际商业条款。

[CI006, CI008, CI009, CI013, CI014]
FI001: 收入模式桥接图

公开产品入口和渠道大致如何把客户活动转成确认收入。

该桥接图只展示公开材料可见的变现逻辑,不估算实际收入结构或成交价。

[CI001, CI003, CI005, CI007, CI009, CI013]

4.2 GTM 路径与销售效率代理指标

Firmus 看起来像直接企业销售和市场辅助分销的混合体。产品页面反复强调询价式采购、预留、混合部署、可观测性和运营支持。这指向高接触销售,而不是大规模自助转化。AI Singapore 以及 MPA 相关的新加坡材料也强化了这个判断:公开证明点是参考工作负载、主权或公共部门相关性,以及高性能技术交付,而不是广泛的客户标识 数量或交易式网页注册。NVIDIA 的 DGX Cloud Lepton 通过提供共同市场和区域容量发现层,改变了漏斗顶部,但没有让底层服务变成低接触。预留集群、公共部门买方和 24/7 支持仍意味着账户级资格评估和实施工作。公司没有发布任何常规效率指标——CAC、回收期、管线转化率、NRR 或支持人员数——因此最好的公开代理指标只能是间接的。新加坡在线工作负载、进入 NVIDIA 生态以及以参考为主的伙伴渠道,都说明买方有兴趣;它们还不能证明变现有效率。投资者应把销售效率视为未回答的执行问题,而不是隐藏强项。[CI008, CI009, CI010, CI011, CI012, CI014]

单位经济性表
指标数值 / 公开代理指标置信度重要性尽调问题
当前收入 / ARR缺少当前收入,增长、估值倍数和回本期都无法进入可投资分析提供过去 12 个月收入,以及按收入来源拆分的 run-rate / ARR 桥
当前客户数没有活跃账户数,就无法测算客户集中度和先落地再扩张的账户经济性提供活跃客户数、前 10 大收入占比,以及按细分市场拆分的账户数
利用率 / 已预订容量GPU 利用率决定固定成本吸收,也验证园区是否高效填满提供新加坡云、Tasmania 爬坡和预留集群的利用率
技术牵引代理指标为 AI Singapore 提供 256 块 H200 GPU、跑 200+ 次实验、10 天训练 27B 模型能说明使用强度可信,但不能说明付费意愿或留存把旗舰技术使用量换算成收入和毛利贡献
销售效率指标CAC、回本期、NRR 和支持负担决定高触点 GTM 能否规模化按工作负载类型提供 CAC、回本期、赢单率、管线转化率和支持 FTE
同行成本结构参照Equinix 收入成本主要来自折旧、租赁、公用事业、带宽、人员、维护和安保勾勒规模化 AI 基础设施运营商可能的固定成本结构将 Firmus 站点层面的电力、租赁、人工和支持成本桶映射到同行披露口径
融资强度参照CoreWeave:$12.9bn 债务承诺和 $2.6bn 经营租赁负债;AirTrunk:A$16bn 再融资可比 AI 和超大规模运营商通常使用很大的融资栈提供当前债务、信用证、供应商融资和项目融资计划

null 值表示该指标未公开。同行行只用于提示承销压力点,并非对 Firmus 业绩的估计。

[CI011, CI015, CI026, CI028, CI030, CI034]
FI002: 单位经济模型桥接图

从工作负载需求到利润的定性桥接,突出缺失的内部指标。

未知值保持明确,不用伪精确填补;同业披露只用于标出成本节点和压力点。

[CI014, CI015, CI025, CI030, CI036, CI037]

4.3 成本结构、capex 强度与运营义务

公开证据强烈显示,Firmus 的成本基础很重资本。Southgate 站点页面和公司承诺页面显示,运营模式围绕高密度 GPU 基础设施、液冷、确定性电力、网络连接、编排软件和持续支持搭建。2026 年 6 月的南澳协议把这种成本姿态从愿景变成具体义务:12 年合同电力、600MW 确定性电力、1.2GW 关联可再生能源、1.5GWh 储能,以及每年最高 220 小时需求响应义务。Firmus 还称,将支付商业电价,并自费建设输电或连接升级。这些是经济义务,不只是 ESG 修辞。外部报道让规模问题更尖锐。ABC 报道 Launceston 项目约 A$2.1 billion,并引用公司说法称第一阶段需要 90MW。CoreWeave、Equinix、NEXTDC 和 AirTrunk 的可比披露指向同一方向:AI 和超大规模基础设施经济性由固定资本、电力、租赁、融资和支持负担主导。这使 Firmus 更像数据中心或项目融资业务,而不是传统软件公司。[CI016, CI018, CI019, CI020, CI021, CI022]

FI004: 资本强度 / 现金流地图

披露股权、园区建设和电力义务如何叠加成潜在融资依赖。

该地图是方向性判断。它识别公开来源可见的现金需求节点,不代表管理层内部项目模型或正式预测。

[CI021, CI022, CI023, CI028, CI031, CI040]

4.4 公开牵引缺口与资本充足性

最强的公开牵引信号是技术性的,不是财务性的。AI Singapore 公开描述了在 Firmus 基础设施上使用 256 块 H200 GPU、超过 200 次实验,以及快速模型训练周期。NVIDIA 将 Firmus 列入 DGX Cloud Lepton,多份公开材料也把公司放在新加坡云和主权算力语境中。这些都是可信证明点,说明服务存在,也说明一些成熟买方愿意使用。它们不能替代财务 KPI。受审阅来源没有披露当前收入、ARR、已预订容量、客户数量、客户标识集中度、利用率、毛利率、现金、烧钱速度或债务栈。2025 年 9 月 A$330 million 融资真实且具有战略意义,但公开义务增长快于披露。单是 Southgate 就被描述为数十亿美元级资产,南澳平台又增加了长期电力、储能和电网责任。因此,SmartCompany 报道称 Firmus 预计会在拟议的 2026 年上市前继续融资,即便确切路径未确认,也方向上可信。公开记录更支持“资金足以继续建设”,而不是“相对已披露管线已经全额融资”。[CI010, CI011, CI016, CI017, CI024, CI032]

资本充足性表
项目公开数值 / 状态重要性证据质量融资含义尽调问题
最新披露股权融资A$330m最新硬性股权事实,也是即时资本缓冲资本有分量,但相对多园区野心仍偏小提供扣除交易费用和近期用途后的备考现金余额
最新披露估值官方 A$1.85b;基于 AFR 的独立报道取整为 ~A$1.9b设定融资背景,也显示公开记录存在取整噪声任何估值工作都应以准确融资文件为准提供已签配售文件和交割后股权结构表
官方募集资金用途加速 Project Southgate确认资金投向园区建设,而不只是泛泛叙事说明股权资金正在被 capex 消耗,而不是作为冗余现金留存提供逐站点募集资金用途计划
Tasmania 项目成本信号ABC 报道约 A$2.1bn旗舰园区 capex 的外部规模信号单个项目 capex 可能是最新股权融资的数倍提供董事会批准的 Tasmania capex 预算和提款计划
South Australia 电力承诺12 年、600MW 批发协议在收入完全披露前,就形成长期运营和融资义务意味着需要项目式承销要求和需求风险管理提供购电 / 承购条款、抵押品,以及介入 / 违约条款
可再生能源 / 储能联动到 2032 年新增 1.2GW 可再生能源,加 1.5GWh 电池储能;每年 220 小时负载灵活性说明 Firmus 承担的不只是计算硬件把增长绑定到第三方基础设施建设和电力市场条件披露每项联动资产由谁出资,以及交付延误时如何处理
电网 / 输电政策Firmus 称其出资建设接入所需输电和网络基础设施接入 capex 可能实质改变现金需求和时间表提高追加债务、项目融资或新股权融资的概率提供接入协议、资本化电网支出和付款里程碑

本表关注前瞻资本充足性,而非历史融资轮次。已披露币种为澳元的数值继续使用澳元,不做合成美元换算。

[CI016, CI017, CI021, CI022, CI023, CI024]
公开财务缺口表
缺失指标或文件重要性当前公开替代信息对承销的影响精确尽调路径
按收入来源拆分的收入 / ARR用于判断规模和结构公开信息只有产品界面和技术案例研究无法为增长或收入质量定价要求提供按 AI 云、裸金属和服务拆分的过去 12 个月收入
实际价格、折扣和渠道抽佣率用于把工作负载换算成毛利按需 / 预留 / 市场平台机制可见,但价格不可见无法判断商业化效率要求提供价目表、标准合同模板和实际净价瀑布
客户集中度和合同期限用于衡量流失风险和议价能力AI Singapore 和 NVIDIA 证明需求入口,但不能证明结构无法评估收入耐久性或集中风险要求提供前 20 大客户结构、已承诺容量和续约计划
按站点拆分的利用率 / 已预订容量用于判断固定成本吸收技术案例使用量公开,商业利用率未公开无法区分可信需求和闲置基础设施要求按园区提供月度利用率、积压订单和已预订容量看板
毛利率和电力成本转嫁用于承销高耗能 AI 基础设施的单位经济性同行披露只能提供比较参照无法判断效率主张能否转化为利润要求按工作负载提供毛利率桥,包括电力和支持成本分摊
现金、烧钱、债务和项目融资栈用于测试 runway 和融资依赖公开信息只有 2025 年股权融资和同行融资类比无法验证资本充足性要求提供当前资产负债表、债务明细、LOC,以及任何项目融资条款清单
逐站点 capex、接入成本和提款计划用于核对建设野心与资金来源ABC 和公司政策只给出方向性规模信号无法评估是否很快还要融资或举债要求提供 Tasmania、South Australia 及后续园区的董事会批准 capex 模型

本表有意梳理仍未公开的内容。每行都列出从叙事型尽调推进到承销型尽调所需的最低文件或数据集。

[CI015, CI032, CI036, CI037, CI038, CI041]
FI003: 财务估算区间

把 Firmus 已披露股权事实与可比基础设施运营商的公开基准区间做资本规模对比。

Firmus 估值沿用披露时的澳元口径。可比项只用于展示资本规模,不暗示经济性相同。

[CI016, CI024, CI033, CI034, CI044]

4.5 财务结论与尽调阻塞点

从财务上看,Firmus 比较容易相信,但很难承销。公司有可信的收入触点、真实的技术需求验证;以私有 AI 基础设施运营商标准看,电力和冷却架构也异常具体。但同一组证据也显示,业务成败取决于能否在资本需求跑赢已披露股权之前,把电力、冷却和融资承诺转化为高利用率经常性收入。因此,收入质量是未证明,不是已证伪。公开证据还无法告诉投资者实际定价是什么样、收入基础有多集中、扣除电力和支持成本后还能留下多少利润,或客户预承诺能否显著抵消 capex 曲线。正确结论是:Firmus 有可信技术牵引,资本强度高,公开披露缺口严重。核心尽调问题包括按收入流划分的当前收入、实际定价和折扣政策、合同期限和续约条款、按站点划分的利用率、按工作负载类型划分的毛利率、当前现金和烧钱速度、各站点 capex 与电网接入提款计划,以及塔斯马尼亚、南澳和任何后续园区的确切融资结构。没有这些,常规承销仍然卡住。[CI037, CI039, CI040, CI041, CI042, CI043]

Chapter 05

05产品与技术

5.1 用客户工作流定义方案

Firmus 没有把自己包装成单一的一体化产品,而是在销售一个分层 AI 基础设施工作流:先提供 GPU 算力访问,再叠加把团队推向生产所需的存储、编排和应用入口。实际用户旅程从按需或预留 GPU 访问开始,取决于工作负载是突发型还是承诺型;随后接上检查点和数据集存储,最后叠加托管运营、开发者工具包和推理端点。这个结构很重要,因为它让 Firmus 更像垂直整合的 AI 工厂运营商,而不是普通 GPU 转售商。公司也明确表示,技术栈面向多类买方——从开发者和企业平台团队,到教育和政府用户——因此这个工作流既要吸收实验,也要承接更受控的生产使用。[CE001, CE002, CE003, CE004, CE005, CE006]

工作流与用例表
用户任务当前工作流痛点Firmus 产品路径声称收益限制
训练多节点 LLM集群稀缺,互联配置复杂Cloud Compute 或 Bare Metal,加 Slurm 和 InfiniBand在 H200 支撑的集群上做分布式训练没有公开的工作负载专属价格或吞吐曲线
部署智能体 AI运行时和推理封装碎片化Cloud Applications,加 GPU Cloud 上的 NIM API从原型到生产更快按需申请的套件未完整说明
运营企业 ML 管线混合集成和运维负担AIFactoryOS、可观测性和混合连接跨工作负载治理和可视性没有公开控制映射或管理员截图
管理大型数据集和检查点训练中的存储瓶颈带 RDMA 或 NVMe 加速的 AI 存储在集群规模向 GPU 供给数据,不让存储拖慢未公布耐久性或复制目标
运行主权或公共部门 AI土地、电力和冷却受限HyperCube 概念,加上与 HTX 或 MPA 相关的设计在受限场址主打更低土地与能源占用公开证据仍集中在研究,而不是生产案例

收益来自产品文案和合作伙伴材料的明示或推断;缺口显示尽调仍需直接运营证据。

[CE002, CE003, CE004, CE008, CE021, CE034]
FE001: 客户工作流和运营流程

市场化用户旅程从计算访问开始,经编排和应用工具,进入生产级 AI 使用。

[CE002, CE003, CE004, CE005, CE006, CE008]

5.2 AI Cloud 与 AI 工厂的模块和资产地图

目前对外销售的模块集合覆盖 Cloud Compute、AI Storage、Bare Metal、Cloud Services 和 Cloud Applications,这些软件和服务入口又被绑定回实体 AI 工厂资产。Engineering Principles 页面很有用,因为它把商业接口接到公司更深的基础设施抽象上:HyperCubes 被描述为核心物理构件,而不只是品牌包装;同一页面还连接了当前新加坡资产、Southgate 项目和已宣布的 Batam 扩张。换句话说,产品地图横跨服务模块和承载这些模块的设施。尽调需要注意这一点:客户体验质量不仅取决于 API 或集群规格,也取决于 Firmus 投产高密度液冷基础设施、并让这些资产与云服务层保持同步的能力。[CE001, CE009, CE010, CE011, CE012, CE042]

产品模块与资产矩阵
模块或资产主要用户当前状态差异化尽调缺口
AI Cloud Compute模型构建者、研究人员、平台团队公开产品页已上线;提供按需和预留选项以 H200 为主的液冷节点,配 Slurm 和 InfiniBand需要公开定价、区域列表和 SLA 条款
Bare Metal企业和承诺型训练用户公开产品页已上线;以预留为主单租户 4x-8x H200 集群,配 24/7 运营支持需要当前在线容量和开通周期
AI Storage运行多节点训练和检查点的团队公开产品页已上线面向 AI 管线的 RDMA 加速 NVMe 和分布式存储需要吞吐量、耐久性和供应商范围细节
Cloud Services / AIFactoryOS平台工程和运维团队公开产品页已上线治理、遥测、托管 Slurm 和混合连接需要 API 文档、发布节奏和具名客户案例
Cloud Applications开发者和数据科学家公开产品页已上线;部分套件需申请CUDA、Jupyter、AI Workbench 和 NIM 推理界面需要 GA 矩阵、支持边界和兼容性细节
HyperCube / AI Factory 资产锚定租户、主权工作负载、云运营新加坡和澳大利亚资产,加 Batam 路线图围绕高密度 AI 基础设施共同设计的多 petascale 模块化单元需要按站点披露当前在线 MW 和已部署 GPU 数

各行概括公开模块界面及其下方物理资产层;状态指公开网页证据,不一定代表广泛商业可用。

[CE001, CE005, CE006, CE008, CE010, CE011]
FE002: 产品架构地图

公开技术栈把工厂基础设施、计算、存储、编排和开发者界面分层,拼成一个 AI 交付系统。

[CE001, CE010, CE011, CE012, CE020, CE021]

5.3 架构:冷却、网络、GPU、编排和数据层

以私有基础设施公司的标准看,公开架构故事异常具体。Compute 页面披露了以 H200 为主的节点设计:每节点八块 GPU、大容量 HBM3e 内存池、节点内 NVLink 和 NVSwitch,以及用于横向扩展的 InfiniBand 或高速 Ethernet。Bare Metal 随后把同一架构延伸到预留单租户集群,AI Storage 则描述了 RDMA 加速的 NVMe 和检查点密集型工作流。硬件之上,Slurm 被持续点名为调度器,AIFactoryOS 被定位为编排和遥测层。VAST 伙伴关系补充了一个独立的数据平面线索:Firmus 试图让存储和元数据管理随算力与能源同步扩展。净结果是,这套栈有意为分布式训练优化,而不是面向普通虚拟机托管。[CE013, CE014, CE015, CE016, CE017, CE018]

技术与运营架构表
层级或组件作用依赖关键披露细节风险
GPU 计算节点训练与推理执行NVIDIA H200 / H100 技术栈公开的 8x H200 节点、大容量 HBM3e 池、NVLink 与 NVSwitchGPU 供应商集中,价格透明度有限
横向扩展互联多节点通信InfiniBand 或高速以太网200-800 Gb/s InfiniBand 选项,或支持 RDMA / RoCE v2 的 400 Gb/s 以太网实际部署拓扑和交换机选择未公开
存储与数据层数据集与检查点吞吐RDMA NVMe 加 VAST AI OS面向流水线的 RDMA 存储,以及解耦式 AI 数据层没有端到端存储性能公开基准
调度与编排资源分配与可视性Slurm 加 AIFactoryOS托管 Slurm、治理、工作负载自动化、遥测没有公开 API 文档或管理员界面截图
冷却与电力层密度与效率管理液冷、浸没式计量、电网互动液冷节点、浸没式机架功率测量、电网感知控制主张设施级效率仍部分依赖自证
市场平台与渠道层区域分发与采购NVIDIA DGX Cloud Lepton跨提供商通用工作流,以及 Firmus 参与市场平台商业化路径仍与 NVIDIA 生态入口紧密绑定

该架构表把 Firmus 明确披露的内容,与仍决定执行质量的外部依赖拆开。

[CE013, CE014, CE015, CE017, CE019, CE020]
FE003: 关键依赖地图

Firmus 掌控集成层,但交付仍依赖 NVIDIA、VAST、基准机构、政府伙伴和站点执行对手方。

依赖地图聚焦可见技术和商业卡点,而不是未公开的供应商合同。

[CE023, CE024, CE025, CE028, CE029, CE040]

5.4 部署、可靠性、支持与信任控制

部署叙事围绕降低运营摩擦展开:Firmus 宣传托管运营、可观测性、混合连接,以及从 Jupyter 和 CLI 工作流延伸到 NIM 支持推理的工具链。这让产品比裸租容量更可用,但可靠性证据仍然参差。最强的公开证明来自 MLPerf 工作,Firmus 在其中披露了节点级功耗测量方法,并与风冷 H100 系统做效率比较;不过公司也承认,设施级 PUE 主张不在 MLCommons 验证范围内。信任控制同样可见但不完整。Cloud Services 声称具备 ISO 27001、SOC-2、静态和传输中加密;网络安全领导层招聘也显示公司正在把真正的控制职能制度化。尚未可见的是公开 SLA 材料、可供客户使用的控制包,或公开事故历史机制。[CE030, CE031, CE032, CE033, CE034, CE035]

信任、质量与合规表
控制项或证明点公开状态范围证据质量缺口
ISO 27001声称Cloud Services公司公开产品页声明未见公开证书或范围说明
SOC 2声称Cloud Services公司公开产品页声明报告类型和信任服务范围未公开
传输中和静态加密声称AI 流水线与云服务流量公司公开产品页声明没有公开 BYOK、KMS 或密钥轮换细节
基准透明度部分有证据MLPerf 节点级功率方法公司方法说明加 MLCommons 框架背景设施级 PUE 仍未经过独立验证
网络安全负责人补强公开招聘证据基础设施、身份与应用官方招聘启事未见公开事件响应、正常运行时间或审计控制资料包

公开信任控制足以支撑尽调框架,但多数证据仍是自证且范围有限。

[CE032, CE033, CE037, CE038, CE039, CE045]
FE004: 产品成熟度和证据地图

计算和编排已经面向市场,但控制平面可靠性、专利和大规模园区执行的独立证明更薄。

矩阵反映证据质量,不代表绝对工程能力;各行区分面向市场的界面与仅停留在路线图或研究阶段的元素。

[CE032, CE034, CE035, CE037, CE041, CE045]

5.5 差异化、路线图与开放风险

Firmus 的差异化主张在被描述成系统集成时最有说服力,而不是单一基准测试结果。公司把液冷、HyperCube 模块化、model-to-grid 编排、解耦的 VAST 数据层,以及深度 NVIDIA 对齐,组合进一个运营模型。如果公司真能把电力、冷却和数据层协同转化为更低的每 token 成本,这套架构就有价值。路线图也具体到足够有意义:Lepton 分销、VAST 采用和 Batam DSX 园区都已公开。但同样的事实也制造风险。最雄心勃勃的容量扩张仍是未来事项,公开路线图没有写清 AIFactoryOS 或存储层面向客户的发布日期,抓取材料也没有披露专利号,无法让投资者区分受保护的专有诀窍 与难以审计的实操能力。执行质量、伙伴依赖和证据深度仍是主要承销问题。[CE041, CE042, CE043, CE044, CE045, CE046]

路线图、发布与开发阶段表
日期或阶段功能或里程碑状态含义来源
2025 研究阶段与 HTX 合作研究液冷 AI 基础设施已宣布研究支撑主权公共安全设计叙事,不是 GA 服务证明Firmus 与 HTX 材料
2025 研究阶段MPA 海水冷却模块化 AI 工厂研究已宣布研究验证受限场址的滨水部署逻辑Firmus 与 MPA 材料
2026 公开上线产品面Cloud Applications、Cloud Services 与 AIFactoryOS 营销页面公开页面已上线产品越来越由软件牵引,不只是设施牵引Firmus 产品页
2026 基准MLPerf Training and Power v4 披露公司结果已发布为效率叙事补上基准证据Firmus MLPerf 页面与 MLCommons
2026 伙伴扩张参与 DGX Cloud Lepton 市场平台已宣布通过统一界面扩大区域触达Firmus 与 NVIDIA
2027-2028 扩容路线图Batam 360 MW DSX 园区,最高 170,000 个加速器已宣布路线图若落地,上行空间大;但建设和供应商风险也集中Firmus 与 Tech Wire Asia

路线图在合作伙伴和园区公告上最强;在客户可见的软件发布时间和服务级承诺上最弱。

[CE012, CE029, CE031, CE040, CE043, CE044]

5.6 图表

Chapter 06

06客户

6.1 分层地图与当前公开参考足迹

Firmus 公开瞄准的买方集合,比它具名参考名单显示的更宽。对外销售足迹覆盖 AI 原生初创公司、企业 AI 团队、研究机构、政府和公共部门用户、主权或受监管工作负载,以及通过 NVIDIA 和其他伙伴牵头的渠道访问。问题在于,分层地图远比具名客户名单丰富。AI Singapore 是最清楚的公开证明点,因为它给出了具体机构用户、特定模型家族和可衡量的工作负载规模。HTX 和 MPA 说明新加坡公共机构愿意围绕主权和可持续基础设施问题同 Firmus 接触,但这些参考仍偏研究或设计导向。NVIDIA、STT GDC 和 VAST 扩大了访问和可信度,却仍是伙伴表面,而非终端客户留存证明。结果是,客户故事在研究、公共部门兴趣和渠道验证上最强,在可独立验证的企业生产采用上薄得多。[CU001, CU002, CU003, CU004, CU019, CU020]

客户分群表
客群买方 / 用户 / 付款方公开用例公开证明 / 规模战略价值 / 缺口
AI 原生初创公司创始人、ML 工程师、平台负责人;常见为按用量付费的买方需要弹性云容量的训练、推理和智能体应用工作负载Batam 与 Reuters 材料明确瞄准 AI 原生客户,但没有公开具名初创客户标识上行空间大的客群,但现有证据多是面向未来的需求语言,而不是已披露的活跃账户
企业 AI 团队与 ISV企业平台团队、软件厂商和内部模型构建者从原型到生产的算力、区域部署和多云可移植性Lepton 与 Batam 材料提到企业和 ISV 用户;未找到具名 Fortune 500 或 ISV 生产参考商业 TAM 很宽,但企业证据质量明显弱于研究或政府证据
研究机构研究人员、模型开发者和公共 AI 项目SEA-LION 训练、模型评估、基准测试和托管AI Singapore / SEA-LION 是最强的具名参考,披露了 32 个节点和 256 块 H200 GPU这是高质量可信度信号;但若后续没有更多实验室,一个旗舰机构可能夸大覆盖宽度
政府 / 公共部门机构赞助方、任务运营方和公共安全负责人公共安全算力设计、可持续性牵引的基础设施研究,以及主权 AI 能力规划HTX 和 MPA 均为具名的新加坡公共机构,但两个参考仍处于研究或课题阶段能证明政府参与有效,但还不能证明重复采购或规模化实时收入
主权 / 受监管算力用户政府、受监管行业和对本地性敏感的工作负载本国托管、低延迟部署和数据主权匹配NVIDIA、VAST 和政策来源都把主权需求描绘为真实存在,但 Firmus 未公布具名受监管行业客户叙事契合度高;具体到账户层面的披露仍薄
渠道 / 合作伙伴市场平台运营商、数据中心托管方、存储 / 平台厂商提供分发、托管和生态可信度,而非直接消耗工作负载NVIDIA Lepton、STT GDC 和 VAST 显著扩大触达,但它们是合作伙伴界面,不是留存的终端客户有用的市场触达杠杆,但也提高对合作伙伴经济性和执行的依赖

各行把终端客户需求客群与伙伴牵引的触达路径拆开;公开证明在研究和公共部门客群最强,在具名企业买方最弱。

[CU001, CU003, CU019, CU020, CU023, CU024]
FU001: 客户旅程地图

公开可见的客户推进从主权或研究需求形成,进入具名研究,再走向生产托管和伙伴带动的扩张。

旅程阶段综合直接客户证明、伙伴页面和政策材料;没有公开来源披露 Firmus 端到端销售周期时长,或按客群划分的准确转化率。

[CU001, CU005, CU011, CU015, CU019, CU020]

6.2 采用轨迹与部署证明

最强的公开采用证据集中在一个参考账户:AI Singapore 在 Firmus 基础设施上的 SEA-LION 工作。Firmus 称 AISG 使用其平台完成快速实验、大规模训练和评估;案例研究披露了 32 个节点、256 块 H200 GPU、超过 200 次实验,以及具体模型训练时间线。这比客户标识墙实质性强得多,因为它显示了实际运行内容和规模。在这个参考之外,轨迹就不那么商业化,而更偏开发性。HTX 描述了围绕公共安全和主权任务算力目标的液冷 AI 基础设施联合研究;MPA 描述了对海水冷却模块化 AI 工厂的研究,且须经过规划、污染控制和环境审查。Lepton 伙伴关系随后借渠道入口拓宽访问,但它证明的是市场参与,而不是客户黏性。因此,从公开材料看,Firmus 有需求形成和部署能力的证据,但只有一个具名参考披露了详细运营指标。[CU005, CU006, CU007, CU008, CU009, CU010]

客户增长 / 采用轨迹表
指标数值日期来源置信度含义缺失分母
具名研究参考客户已上线AI Singapore / SEA-LION 合作正式化2025-03Firmus AI Singapore 合作页面显示新加坡有可引用的机构用户未披露总客户基数或赢单率
具名客户的 GPU 部署规模32 个节点 / 256 块 H200 GPU2025-12Firmus AI Singapore 案例研究至少一个账户的具体工作负载规模可见缺少利用率、支出或合同价值分母
实验量已完成 200+ 个实验2025-12Firmus AI Singapore 案例研究说明存在重复使用,不只是仪式性公告缺少与平台总实验量的对比
模型训练吞吐27B 模型 10 天;4B 模型 3.5 天2025-12Firmus AI Singapore 案例研究暗示平台能支撑严肃训练周期缺少与竞争提供商或客户支出的独立基准对比
公共部门主权算力切入HTX MoU 已宣布2025-05-27HTX 官方发布释放其与公共安全和主权工作负载相关的信号缺少采购金额、上线部署日期或转化率
滨水主权基础设施切入MPA 海水冷却研究已宣布2025-06-11MPA 官方发布显示另一家公共机构愿意测试该架构仍是研究,不是已预订的生产合同
渠道扩张路径Firmus 加入 DGX Cloud Lepton 市场平台2025-06-12Firmus 与 NVIDIA Lepton 材料拓宽 AI 原生和企业构建者的获客路径未披露 GMV、席位数或客户转化率
前瞻需求信号已承诺协议支撑的前六年承购预期 US$25-30B2026-06Firmus、Reuters 与 TechWire Asia暗示园区若交付,容量已有买方绑定交易对手名称、信用支持和收入确认假设均未公开

这张表是混合轨迹:前四行反映已披露使用证据,后三行反映机构或渠道扩张,最后一行是前瞻管线,不是已实现留存。

[CU005, CU007, CU008, CU009, CU011, CU015]
具名客户证明表
客户 / 参考客群部署 / 用例生产 vs 试点结果 / 证据质量局限
AI Singapore / SEA-LION研究机构 / 国家 AI 项目面向东南亚 LLM 的大模型训练、评估、托管和基准测试类生产研究部署最高质量的公开证明:具名用户、满意度引述、32 个节点 / 256 块 H200 GPU、200+ 个实验,以及模型训练时间线仍由公司发布;未披露合同价值、续约期限或独立采购记录
HTX政府 / 公共部门针对公共安全和应急响应用例,共同研究液冷 AI 基础设施研究 / 设计阶段HTX 官方发布确认主权任务算力意图,并点名 Firmus 为合作伙伴未披露生产工作负载指标、合同价值或上线客户服务
MPA政府 / 公共部门基础设施规划方围绕新加坡滨水区域研究并试点测试模块化海水冷却 AI 工厂研究 / 试点阶段MPA 官方发布确认机构参与,并列明监管 / 规划工作流该参考证明公共部门入口,不证明重复算力消耗或商业部署

该枚举有意只覆盖部分样本,并限于公开具名终端用户或机构参考;在已审阅公开材料中,未找到具名企业或超大规模云厂商终端客户。

[CU006, CU007, CU008, CU011, CU012, CU015]
FU002: 采用 / 部署漏斗

公开证据集很快从宽泛客群定位收窄到少数带运营细节的具名引用。

阶段计数反映截至 2026-07-02 审阅的公开来源;这是证据载体的数量,不是按收入加权的客户群组。

[CU006, CU007, CU010, CU011, CU015, CU021]

6.3 按细分市场划分的具名参考质量

参考质量在不同细分市场之间差异很大。AI Singapore 是高质量客户证明,因为合作具名、说明了工作负载、引用了客户引语,且产出与 SEA-LION 项目相连。HTX 和 MPA 仍是有意义的参考,因为二者都是新加坡官方机构,也都描述了明确用例;但它们尚未证明已签约生产消费、已披露合同金额或重复采购。商业侧公开证明弱得多。DGX Cloud Lepton 表明 Firmus 进入了经过筛选的 NVIDIA 生态,STT GDC 与 VAST 表明基础设施伙伴愿意同它共建,但这些来源都没有点名一个正在生产中使用 Firmus 的企业终端客户。这是本章的核心张力:公司作为主权和研究基础设施提供商看起来越来越正当,但公开记录仍没有显示一个广泛、具名、且有耐久商业支出的企业客户基础。[CU003, CU006, CU010, CU011, CU015, CU019]

FU003: 客户证明矩阵

公开证明质量在研究引用上最强,公共部门研究居中,企业广度最弱。

位置反映公开证据质量,而非收入规模;伙伴证明可以提升可信度,但不能证明终端客户消费具备持久性。

[CU006, CU011, CU015, CU021, CU027, CU028]

6.4 耐久性、留存与扩张信号

耐久性是公开客户记录中最弱的一环。受审阅来源没有披露客户数量、净收入留存、毛留存、流失、合同期限或续约节奏,因此投资者无法判断 Firmus 是在留住账户,还是主要制造新的试点关注。公开可见的最佳耐久性代理指标是定性的:AI Singapore 案例研究描述了一段正式化的长期伙伴关系,引用了对工程团队的满意度,并把关系框定为从研究走向生产。扩张逻辑比留存数学更可见。Lepton 降低了开发者、AI 原生团队和需要跨区域从原型走向生产工作流的企业建设者的分销摩擦;Batam 和澳大利亚叙事则暗示未来更大的 AI 原生、企业、ISV、主权和超大规模需求池。但这些仍只是渠道和路线图信号。没有具名续约、客户群组 或多账户扩张指标,公开记录能更清楚地证明可触达需求和部分部署成功,而不是耐久商业复利。[CU020, CU022, CU023, CU024, CU025, CU029]

留存 / 重复使用 / 满意度表
指标数值客群置信度尽调问题
净收入留存全部客群索取按客户群组划分的 NRR,以及过去 12 个月扩张收入与收缩收入对比
毛留存 / 流失全部客群索取客户数流失、工作负载流失,以及任何已终止的公共部门或研究合作
合同期限 / 续约周期全部客群索取云、市场平台和主权合同的标准期限,以及最早续约日期
重复使用代理指标AI Singapore 案例研究描述了 200+ 个实验和正式确立的长期合作研究核实实验量是否转化为合同化经常性支出或续约承诺
客户满意度代理指标AI Singapore 案例研究中有关于响应速度和顺畅运营的正面客户引述研究获取 Firmus 自有媒体之外的独立客户访谈或客户自撰证言
公开生产服务 SLA 披露企业 / 主权 / 市场平台按产品线索取标准 SLA、正常运行时间历史和服务补偿条款

空值表示已审阅来源未公开;两个非空行是定性代理指标,不应误作留存 KPI。

[CU010, CU029, CU030, CU047]
FU004: 持久性可见度矩阵

Firmus 披露了部分采用输入,但最能说明客户持久性和集中度的指标大多仍不透明。

矩阵跟踪披露可见度,而非业绩质量;空白或不透明单元格意味着公开记录回答不了该问题,不代表公司缺乏能力。

[CU026, CU029, CU034, CU035, CU045, CU049]

6.5 集中度、伙伴依赖与采购摩擦

主要客户风险是透明度不足和依赖,而不是需求明显不足。公开披露没有量化头部客户占比或细分组合,因此无法用开放来源给集中度风险定量。Batam 项目以已承诺承购协议和 AI 原生需求来营销,但客户名称和合同结构仍是私有信息,因此锚定租户 风险真实存在。伙伴依赖也很重,因为公开的市场路径依赖 NVIDIA 硬件和 Lepton 分销、STT GDC 托管历史、VAST 数据层,以及 DayOne 的印尼园区建设。采购和许可摩擦对主权用户尤其相关:MPA 明确把环境和海事审查放进任何滨水冷却路径,新加坡 AI 战略强调由效率约束的算力增长,澳大利亚预期文件把 AI 工厂审批同主权、能源、社区和清洁基础设施测试挂钩。澳大利亚独立报道又补充了第二重警示:电力、水和社会许可问题可能拖慢大型项目。简言之,公开需求信号有希望,但客户质量仍容易受到伙伴集中、政府转化周期,以及锚定买方真实身份披露稀疏的影响。[CU016, CU025, CU026, CU034, CU035, CU036]

扩张与集中风险表
扩张驱动集中风险影响尽调路径
AI Singapore 成功案例可带来更多研究参考若没有其他实验室具名,一个旗舰案例可能主导叙事可能夸大覆盖宽度,掩盖多账户渗透不足要求提供更多具名研究或高校用户,以及其活跃 GPU 消耗
HTX 与 MPA 可打开主权 / 公共部门需求两者仍是研究阶段关系,采购和监管周期都长公共部门转化可能比投资者预期更久要求逐项政府合作提供商业里程碑、采购状态和转化标准
DGX Cloud Lepton 扩大开发者和企业触达获客可能依赖 NVIDIA 市场平台经济性和政策触达市场的杠杆提高,但利润率可见度可能下降索取 Lepton 收入分成条款、预留容量经济性和获客组合
Batam 承购承诺暗示规模需求交易对手名称和集中度未公开锚租户或头部客户失效,可能显著拖累园区爬坡索取承诺承购背后的客户名单、合同期限、信用支持和最低承诺结构
STT GDC、VAST 和 DayOne 扩大容量和运营范围执行依赖多个外部基础设施伙伴伙伴运营或商务延误,会削弱服务交付和客户留存梳理每个伙伴对应的客户可见依赖、终止权和替代方案
澳大利亚和新加坡主权定位受益于政策顺风电力、水、社区和环境阻力可能推迟部署容量延迟可能推迟主权和超大规模用户的上线或扩张按主要园区审查电网接入状态、环境审批和社会许可计划

本表聚焦增长和集中度机制,而非单纯风险严重度;最大的盲点是公开来源没有量化客户或承购集中度。

[CU025, CU026, CU034, CU035, CU036, CU037]

6.6 图表

Chapter 07

07风险

7.1 按严重程度排序的风险栈

Firmus 的风险不主要在 AI 算力需求是否存在,而在公司能否足够快地把需求转成获批、通电、可融资的容量。今天最强的公开证据很窄:St Leonards 有 104 MW 零售服务协议并已开工,Bell Bay 有详细 FAQ 和输电叙事,澳联邦也已明确对高耗能 AI 基础设施的期待。最弱的证据,恰好落在投资人需要形成承保信心的地方:St Leonards 之外的最终审批、客户或包销披露,以及已签署的长期塔斯马尼亚能源安排。这个组合使风险排序很清楚。第一层是电力、审批和社会许可,因为它们可以直接让站点无法上线。第二层是对 Aurora、Hydro、NVIDIA 和 VAST 的伙伴与平台依赖,因为 Firmus 的产品和市场路径仍与外部交易对手高度绑定。第三层是融资、治理和利用率不透明,因为公开记录对需求质量的披露仍远薄于项目野心。[CR001, CR002, CR003, CR015, CR029, CR030]

FR001: 风险热力图

电力和审批执行处在最高影响、最高可能性角落,伙伴锁定和需求不透明紧随其后。

位置反映可见缓释后的剩余投资风险,不代表工程确定性;单元格综合本章证据基础,而不是量化评分模型。

[CR001, CR016, CR024, CR029, CR030, CR042]

7.2 监管、法律和社会许可风险

监管负担不是一张许可证,而是一层叠一层的要求。在澳联邦层面,针对数据中心和 AI 基础设施的新期待让政府有理由优先支持对齐的提案、搁置不对齐的提案,尤其是能源、韧性或社区收益偏弱时。与此同时,AI 的法律环境已经牵涉隐私、董事义务、过失和消费者法风险;SOCI 下的网络改革也已经直接指向数据存储系统和风险管理计划。这些都不能证明 Firmus 不合规,但会抬高一家服务主权和公共部门工作负载公司的尽调门槛。更直接的问题是社会许可。ABC 和 ABC Listen 的报道显示,社区在咨询、水、噪声和公共收益上反弹;Bell Bay 和 Wesley Vale 最新公开规划状态仍取决于公司无法完全控制的市政流程。新加坡 MPA 和 HTX 关系有利于战略定位,但它们是研究 MoU,不是运营审批的替代品。[CR002, CR003, CR005, CR006, CR007, CR008]

监管 / 法律风险登记表
规则 / 流程司法辖区状态可能性严重性缓释措施剩余敞口尽调路径
联邦政府预期下的电力与审批匹配澳大利亚预期已发布;定位是优先级工具,不是直接许可极高Firmus 已公开可再生能源、可调度能力和电网支撑承诺若审批方认为项目在电力或社区收益上不匹配,风险仍在要求管理层逐一把每个站点映射到联邦政府预期和各州审批要求
Bell Bay 和 Wesley Vale 规划审批塔斯马尼亚申请已提交;已审阅来源未公开证明最终结果极高既有工业用地、存量基础设施,以及新增社区沟通会市政审批节奏和本地阻力仍不受 Firmus 控制取得两个站点当前 DA 案卷、公众意见和预计决策时间表
Privacy Act 与 OAIC AI 指引澳大利亚现行法律和监管指引已经适用隐私政策披露、合同控制和工作负载治理公开记录尚未显示可交付客户的控制包或 AI 专项隐私操作流程要求提供隐私影响评估、DPA 模板和公共部门控制映射
SOCI 与 Cyber Security Act 义务澳大利亚改革已生效;是否适用取决于资产和工作负载范围风险管理计划、事件流程和受保护信息控制尚无公开证据显示 Firmus 如何把这些义务落到运营中要求提供关键基础设施法律分析、CIRMP 状态和事件治理材料
新加坡政府合作新加坡已与 MPA 和 HTX 签署 MoU把项目当作 R&D 和信誉渠道,而不是审批研究阶段合作可能被误读为商业或监管放行要求提供工作说明书、交付物,以及从研究走向生产部署的路径
公开披露依赖与诉讼可见度澳大利亚网站条款限制依赖;已审阅来源未浮现官方诉讼记录投资测算只使用已抓取的一手来源和高质量独立来源投资者仍缺少直接的法院、股权表或董事会流程文件要求律师和管理层就诉讼、股权表和董事会治理作出陈述

各行按严重性排序,只限于 2026-07-02 可见的公开监管和法律流程;抓取审查期间无法直接使用市政门户,因此状态依据已审阅的公司、媒体和 RTI 材料,而不是经核验的实时案卷导出。

[CR002, CR003, CR005, CR006, CR007, CR008]

7.3 运营、交付、用水和安全风险

运营上,Firmus 要投资人同时相信几件难事:高密度液冷 AI 工厂能在压缩时间表内爬坡,干冷假设经得起塔斯马尼亚真实条件,站点级运营复杂度能在 24 小时设施里管住。公开证据有利有弊。Firmus 给出过具体用水数字,并称 Bell Bay 每年大约只有 10 个炎热日需要冷却水;但同一轮公共讨论也说明,风险并没有因此消失——居民在质疑假设,独立研究者也指出,干冷可能用额外耗电换取节水。安全带来另一种运营负担。Cyber.gov 指引把数据完整性、加密、来源和生命周期控制视为 AI 系统可靠性的核心;商业模式一旦依赖托管主权或企业 AI 工作负载,这些期待就更重要。AEMO、IEA、JLL、WEF 和 Deloitte 的全球研究进一步说明,建设周期、电网瓶颈和电气设备短缺已经是结构性约束,不再是一次性问题。[CR018, CR019, CR020, CR021, CR022, CR023]

运营 / 质量 / 安全风险登记表
失效模式可能性严重性缓释成熟度剩余敞口未解决缺口
St Leonards 爬坡错过 2026 年 8–11 月负载时间表中等:零售供电和站点建设信息公开,但没有独立投产证明如果第一座工厂爬坡延误,收入确认节奏和可信度都会下滑需要投产里程碑、实时容量数据和客户上线时间表
Bell Bay 干冷假设低估实际电力或用水强度中等:FAQ 和管理层披露了假设,但公开第三方验证很薄弱高温表现或备用冷却可能实质改变成本和社区反应需要工程审查冷却模式、设计气象基准和最坏情形取水量
AI 数据完整性或隐私控制失效影响客户工作负载早期:政府指引清楚,但 Firmus 落实控制的公开证据有限一次控制失效可能同时打击公共部门信任和主权工作负载需求需要控制包证据、渗透测试节奏、加密细节和 AI 数据治理流程
电网、设备或建设瓶颈拖延后续塔斯马尼亚站点低至中:行业研究解释了风险,但不是 Firmus 专属缓冲收入验证前,站点排序和资本开支资金调用可能被拉长需要变压器、开关柜和主要电气包采购时间表
Bell Bay 全天候运营会拉紧招聘、维护和轮班覆盖低:岗位承诺存在,但运营模型细节稀少多个站点投运后,用工缺口可能演变成正常运行时间和安全问题需要组织架构图、轮班设计、维护人员配置计划和承包商策略
施工后,社区对噪音和振动的担忧持续被动响应:反弹后 Firmus 增加了现场咨询会和网络研讨会投诉拉长后,可能转化为审批条件、监测要求或运营限制需要噪音监测计划、升级流程和投产后社区报告

严重性反映对收入兑现时间和公共部门可信度的潜在影响,不只看工程难度;多行依据公开假设,而不是经核验的运营遥测。

[CR018, CR019, CR020, CR021, CR022, CR023]

7.4 伙伴、电网和平台依赖风险

Firmus 的商业架构仍明显依赖交易对手。塔斯马尼亚故事从 Aurora 和 Hydro 的初始电力开始,延伸到 TasNetworks 和 AEMO 的接入与输电经济性,再押注公共政策接受三处站点合计消耗超过 400 MW。Bell Bay 自己的 FAQ 仍称最终能源安排还在谈判,这提醒投资人:在更大规模铺开上,公司叙事多于合同披露。南澳大利亚州的 Gunvor 协议给出更具体的缓释样本,但也凸显对供应商执行的依赖:只有承诺的可再生发电、电池储能和限发机制按时到位,价值才会兑现。技术侧,NVIDIA 既是供应商,也通过 DGX Cloud Lepton 成为渠道;VAST 则是唯一公开披露的基础数据层伙伴。依赖图谱很清楚:Firmus 控制集成和品牌,但几个关键的吞吐、价格和可靠性杠杆仍不在它直接控制之下。[CR022, CR023, CR027, CR028, CR029, CR030]

合作伙伴 / 依赖风险登记表
依赖项交易对手角色集中度失效场景严重性缓释措施剩余敞口
初始塔斯马尼亚供电Aurora Energy / Hydro TasmaniaSt Leonards 的电力零售和发电路径旗舰站点集中度高定价、时间表或政治审查迫使爬坡放慢,或单位经济性恶化极高初始 104 MW 协议已公开,Firmus 称会按市场费率付费更长期的塔斯马尼亚条款和经济性仍只披露一部分
Bell Bay 输电与能源安排TasNetworks / 未来供应商并网、研究和协商能源路径第二站点推出依赖度高并网审批或商业谈判晚于建设就绪极高Firmus 称会自筹输电资金,并以可调度方式管理需求Bell Bay FAQ 显示最终安排仍在谈判
南澳大利亚可再生能源补位Gunvor Group长期稳固供电,加可再生能源和电池建设Firmus 负载继续爬坡时,可再生能源或储能交付滞后12 年合同,加明确的发电和储能承诺只有供应商按期执行、弃电机制真实落地,缓释才成立
GPU 与渠道生态NVIDIA硬件路线图、Lepton 市场入口和买方信任GPU 分配、定价或市场平台经济性恶化Firmus 已列为 Lepton 云合作伙伴,并宣传多代就绪能力公开可见的获客路径仍紧贴 NVIDIA 生态
基础数据层VAST DataAI 操作系统和数据平面中至高数据层路线图或经济性与 Firmus 工作负载模型错配公开选择具名平台,降低了当前架构的不确定性未见公开替代路径、迁移权利或多供应商数据平面策略
需求侧 / 承购基础未披露锚定租户负载利用率和收入转化在可持续合同需求可见前,容量先落地除一般市场需求和合作伙伴信号外,公开渠道看不到其他缓释公开来源无法压力测试客户集中度和利用率

这张表把外部卡点单独拎出,而不是重复内部执行风险;最后一行有意写成未披露承购,因为客户可见度本身就是重大依赖风险。

[CR012, CR014, CR022, CR027, CR028, CR029]
FR003: 依赖地图

Firmus 掌控集成和站点叙事,但关键电力、GPU、数据层和审批依赖仍在外部对手方手里。

本图突出外部集中节点和信息不对称,而不是逐行复述静态合作伙伴清单。

[CR022, CR027, CR032, CR042, CR043, CR044]

7.5 融资、治理和论点失效标准

最后一层风险是财务和治理质量。Treasury 的 RTI 披露显示,即使单一接入升级也可能被认定为重大资本投资,并在没有公开商业论证的情况下送达部长。公开层面,政府也以商业保密为由拒绝披露合同细节。这种不透明很重要,因为公开材料在兆瓦、水和可持续性承诺上远比需求可见度或项目级经济性丰满。所审阅来源没有指出塔斯马尼亚站点的锚定租户、已签约包销方或利用率承诺,意味着客户集中度和利润率韧性还无法用公开证据检验。治理披露同样薄:联席 CEO 在公告中很突出,但审阅材料没有出现 CFO 或独立董事会细节。正确的承保反应应聚焦可观察的否决条件:最终审批、已签署的能源和接入文件、新增发电回补证据、客户或包销披露,以及任何州政治从审视转向更严格监管或暂停审批的信号。[CR012, CR013, CR014, CR046, CR047, CR048]

人员 / 执行风险登记表
角色 / 职能依赖或缺口可能性严重性缓释措施尽调路径
高管领导 / 对外代表公开披露集中在两位联席 CEO 身上管理层明显参与社区和政府对话要求提供继任计划、授权运营权限和具名站点负责人
财务与董事会治理已审阅公开来源未出现 CFO 或独立董事会细节公开不可见要求提供董事会构成、审计监督和项目融资治理材料
社区关系能力反弹后,沟通动作才明显起来现场咨询会和网络研讨会已启动要求提供社区沟通计划、升级日志和获批后报告承诺
运营和维护人员配置Bell Bay 假设 24/7 运营,并配置 >100 名本地 FTE公司提到大型工业劳动力池和可迁移工种要求提供人员爬坡、外包组合和维护 KPI 目标
披露纪律商业保密和部分公开证据限制投资者可见度RTI 和媒体审查正在形成更多披露的外部压力要求提供已签合同摘要、项目仪表盘,以及季度建设进度与计划对比报告

这里的执行风险,是把项目雄心转化为可靠交付所需的人员和披露系统,不是泛泛的招聘挑战。

[CR013, CR046, CR051, CR052, CR053]
缓释措施与否决标准表
风险可监测触发器门槛 / 事件行动含义
审批与社会许可规划流程收紧Bell Bay 或 Wesley Vale 审批延迟、咨询窗口不断延长,或议会审查转化为州层面的限制新的关键路径重设前,不要把完整塔斯马尼亚建设纳入投资测算
电力可得性初始站点和第二站点能源路径变弱St Leonards 错过公开的 104 MW 爬坡,或 Bell Bay 建设已就绪后仍卡在能源谈判中把案例从扩张论证下调为只验证首站点
可再生能源补位可信度新发电能力跟不上负载增长负载承诺上升时,公司无法证明 Hydro 或其他塔斯马尼亚补位安排把可持续性主张视为叙事,而不是成本或政策保护
NVIDIA 依赖GPU 和市场议价能力恶化没有替代渠道时,分配、定价或商业条款恶化假设利润率被压缩、获客变慢
数据层集中VAST 路线图或经济性错配Firmus 无法说明退出路径、迁移路径或双供应商策略在运营模型中加入平台锁定折价
客户 / 利用率不透明没有锚定租户证据尽调中,管理层仍无法披露合同需求、承购质量或利用率假设不要把项目级现金流纳入投资测算
隐私与网络安全控制缺少控制包证据未提供隐私影响评估、事件治理或可交付客户的控制映射假设主权和公共部门销售周期仍受约束
治理与披露规模扩大但不透明持续项目增加后,仍没有商业案例可见度、董事会清晰度或定期建设进度与计划对比报告上调风险评级,要求更强融资契约,或回避

触发器表有意前瞻:每一行都把风险转成可监测事件,能够改变投资判断姿态,而不是只重复静态登记表。

[CR023, CR031, CR033, CR042, CR043, CR047]
FR002: 风险传导地图

最重要路径从审批和电力传导到收入时点,再进入融资、利润率和估值支撑。

该地图展示方向性因果,而不是数字概率树;若干边因公开谈判证据或未披露承购而更强。

[CR013, CR014, CR023, CR029, CR030, CR042]
Chapter 08

08估值

8.1 当前价格锚与真正已证明的内容

文件里最干净的估值事实就是本轮融资本身:Firmus 官方完成 A$330 million 股权配售,投后估值 A$1.85 billion,Ellerston Capital 和 NVIDIA 在投资团内。这个价格买到的是有形项目,而不是泛泛的 AI 故事。Project Southgate 被记录为塔斯马尼亚旗舰园区,建设计划为 36,000 张 GPU;项目页还补充了 84 MW 关键 IT 负载、低于 1.10 的 PUE 和强用水效率主张。文件也显示公司已有面向投资者关系的股东沟通页面,SmartCompany 还报道其有 2026 年上市意图。但估值文件仍缺少投资人从“欣赏醒目估值”进入承保所需的证据:收入、毛利率、利用率、客户集中度,以及本轮新资金所附的清算优先或优先级条款均未公开披露。因此,A$1.85 billion 可以作为市场出清事件成立,但作为普通股承保包仍不完整。[CV001, CV002, CV003, CV004, CV005, CV006]

建议摘要表
维度当前观点重要性置信度
建议跟踪质量和市场顺风真实存在,但公开材料仍太不完整,不足以支持激进进场判断。
置信度融资事件和合作伙伴证据可信,但经济性和条款仍不透明。
风险评级规模被证明前,资本强度、电力时间表和融资层风险可能压缩普通股价值。
估值立场合理至偏高只有商业化推进、未来资本以有利条款到位,A$1.85b 才站得住。
进场纪律分阶段进入或等待要求基于里程碑的投资测算、运营数据权利和股权表清晰度。
退出姿态监测,不预先纳入投资测算公司有 IPO 目标,但尚未看到达到公开市场标准的披露。

该表概括当前轮次价格下的投资姿态;它不能替代股权表审查或完整运营模型。

[CV002, CV039, CV046, CV049, CV050, CV055]
FV004: 投资 KPI

当前投资判断最需要的公开事实简明记分卡。

[CV002, CV004, CV007, CV039, CV055]

8.2 正反论点:市场、产品、客户和监管

正向逻辑不难讲清。Firmus 所在市场里,主权 AI 基础设施、受电力约束的容量和高密度冷却正在同时增值。官方页面显示,公司没有停在一个塔斯马尼亚园区:它与 NVIDIA 拥有 360 MW Batam 合作,声称最多 170,000 个加速器,引用收入分成和信用支持结构,并称已承诺包销 6 年可达 US$25 billion 至 US$30 billion。南澳大利亚州能源协议又叠加一层规模,把 600 MW 供电与 1.2 GW 可再生能源、电池储能和 2.7 GW 规划容量绑定。参与 DGX Cloud Lepton 以及与 AI Singapore 合作,强化了产品和伙伴证据。反向逻辑同样清楚。行业研究称,电力、资本结构和企业变现能力会把赢家和 GPU 经纪人分开。澳大利亚政策也在主权、能源、技能和社会许可上加入真实义务。独立报道还暗示,本地就业承诺可能没有宣传口径那么持久。业务方向可能是对的,但在当前价格下,对普通股而言仍可能过于资本密集。[CV007, CV008, CV009, CV010, CV011, CV012]

投资论点 / 反论点表
视角看多论点反论点改变观点的证据
市场Neocloud 和主权 AI 需求快速扩张,电力稀缺会奖励早期持有容量的人。即便市场快速增长,只要电力、政策或融资收紧,估值仍可能重校。已签客户需求和特定市场电力准入
产品高能效、液冷 AI 工厂设计贴合成本和可持续性叙事。市场可能把 Firmus 看成资本密集型建设项目,而不是差异化的软件式基础设施。经测量的成本、正常运行时间和利用率优势
客户DGX Cloud Lepton 和 AI Singapore 显示,合作伙伴牵引的需求正在形成。相较估值规模和未来资本开支负担,公开客户证据仍薄。具名创收交易对手和集中度数据
竞争在 APAC 落地主权能力,可能打开现有厂商尚未完全本地化的切口。Equinix、Digital Realty 等现有厂商已在全球销售主权和适配 AI 的基础设施。Firmus 在速度、价格和本地性上胜出的证据
资本结构新股本和合作伙伴结构能比纯资产负债表融资更快推动建设。盈利出现前,收入分成、优先资本或项目债可能让普通股经济性让位。完整条款清单、担保排序和项目级资金计划
监管澳大利亚政策明确重视主权、本地能力和能源纪律。同一套政策框架也会放慢或重定价不满足社会许可和基础设施预期的项目。各站点许可状态和监管一致性

每条反论点都与估值相关,而不只是运营问题;关键是它多快会改变普通股结果。

[CV009, CV013, CV014, CV029, CV030, CV034]

8.3 可比框架和情景区间

可比分析在这里主要是纪律工具。AirTrunk 说明,APAC 数据中心平台可以支撑很高的私募估值,但前提是已承诺容量、未来土地储备和庞大融资平台都能看见。CoreWeave 更鲜明地展示了 AI 原生基础设施的上行空间:数十亿美元收入和数百亿美元剩余履约义务,可以与十亿美元级亏损同时存在。Equinix、Digital Realty、NEXTDC、GDS 和 Keppel DC REIT 等成熟上市平台给出另一条经验:公开市场奖励透明度、持续披露和融资韧性。在这个背景下,Firmus 的估值不算荒唐,但确实偏早。因此,基准情景把当前轮次视为大致合理,前提是商业化快速跟上且资本持续可得。乐观情景要求多件事同时顺利,尤其是 Southgate 执行和 Batam 转化。悲观情景不需要需求消失;只要融资、许可或已披露变现低于预期,且发生在 Firmus 达到公开或准公开可比公司已展示的规模和透明度之前,就足够了。[CV016, CV017, CV018, CV019, CV020, CV021]

乐观 / 基准 / 悲观情景表
情景明确假设参考公允价值(A$bn)未来 12–24 个月必须成立的条件概率信号
牛市情景Southgate 按期爬坡,Batam 已承诺承购转成合约,主权需求仍然稀缺,后续资本结构足够标准化,普通股也能分享上行。2.4–3.0商业交付里程碑兑现,交易对手质量高,也没有惩罚性优先资本出现。需要多项绿灯同时亮起
基准情景Southgate 证明商业化可行,Batam 上行仍有一部分未被证实,还需要更多资本,但条款可控。1.5–2.0下一次重大融资前,运营证明先到位,披露也明显改善。按现有证据最站得住脚
熊市情景商业化落后,已披露经济性仍然偏薄,资本以成本更高或级别更优先的结构进入。0.8–1.2时间线滑坡,融资利差走阔,或客户证据仍以叙事为主。无需需求崩塌也说得通
融资压力情景需求存在,但项目债、优先股或伙伴经济性吸收的上行,超过普通股投资者预期。0.6–0.9轮次条款或项目层文件显示,普通股之上存在严重价值渗漏。最需要盯住的悬念

区间是情景锚点,不是精确目标;驱动因素是已披露证明、资本结构风险和里程碑交付,而非对未披露收入的点估计。

[CV040, CV041, CV046, CV047, CV048, CV056]
可比估值表
可比对象公开估值 / 规模信号重要性Firmus 参考价值局限
AirTrunkA$24b 收购;>800MW 已承诺;>1GW 未来增长;A$16b 再融资APAC 私有数据中心平台稀缺价值的最佳锚点说明可见规模、客户承诺和融资深度能撑起什么估值阶段远更成熟,且已获机构化融资
CoreWeave$5.1b 收入;$60.7b RPO;$1.2b 净亏损说明 AI 原生基础设施可以爆发式放大,同时仍极度吃资产负债表可类比 AI 基础设施的上行空间和资金胃口披露强得多,客户画像也不同
Equinix280 个数据中心;10,500+ 客户;$9.2b 收入透明度、可复制性和全球平台价值的公开标尺说明公开市场投资者期待的披露门槛成熟互联与托管平台,不是从零建设 AI 工厂的故事
Digital RealtyAI 专用主权方案,加上公开季报和 SEC 报告节奏说明在位者已经在销售可支撑 AI、受司法辖区约束的基础设施可参考竞争定位和买方替代方案成熟 REIT 经济性不同于 Firmus 的建设与爬坡曲线
NEXTDCA$427.2m 收入;A$2.2b 资本计划;A$2.9b 债务平台区域公开可比:有可见融资支撑面向 AI 的扩张更接近 Firmus 的 APAC 公开市场参照,可看资本开支和融资观感仍是更成熟的托管运营商
GDS / Keppel DC REITUS$1.63b 收入,或 ~$6.2b AUM,且有公开报告说明 APAC 上市平台价值通常在报告规模可见后显现可作为亚洲市场透明度和融资能力的晴雨表地域、结构和客户组合不同

本表有意混合私有交易、上市运营公司和上市平台载体,因为 Firmus 介于 neocloud 增长故事和数据中心基础设施建设之间。

[CV016, CV017, CV018, CV019, CV020, CV021]
FV002: 估值敏感性

相对于大致公允的基准情形,现阶段少数执行和资本变量最能拉动估值波动。

敏感性条是围绕基准判断的方向性调整,并非统计模型;它们对应最可能先撬动普通股价值的变量。

[CV033, CV034, CV040, CV041, CV047, CV048]
FV003: 估值 / 回报区间

当前估值接近基准区间上沿;只有下一个关键执行闸门顺利通过,才有上行空间。

区间以澳元计价,反映对执行、披露和融资质量的情景级判断,而不是对未披露收入的点估计。

[CV046, CV047, CV048]

8.4 入场纪律、摊薄悬顶、退出准备度和最后问题

真正的估值争论不是 Firmus 是否有意思,而是新投资人是否应该在资本结构和收入引擎更清楚之前接受今天的价格。审阅来源显示,公司在并行推进 Southgate、Batam 和南澳大利亚州扩张时,可能还需要更多外部资本。行业来源也说明,这些资本现在来自哪里:优先股、项目融资、ABS、CMBS、私人信贷,以及其他可能排在普通股之上或稀释普通股的结构。因此,入场纪律是核心。一笔新投资应分阶段投入、以文件为先,并明确摊薄、下行保护和信息权。退出愿景存在,因为公司已经有股东沟通基础设施,媒体也报道了上市意图;但退出准备度仍落后于成熟上市可比公司展现的披露门槛。因此,实际建议是 Track,而不是 Buy:保持接触,要求更严密的承保包,只有当商业化、交易对手和融资条款改善速度快于估值时才行动。[CV039, CV040, CV041, CV046, CV049, CV050]

论点破裂与退出触发因素表
触发因素阈值 / 事件如何传导到投资论点行动含义
Southgate 交付延误商业交付或送电错过下一个外部可见里程碑窗口推迟货币化证明,并抬高融资需求暂停,或扩大估值折扣
商业指标继续不披露下一轮融资步骤前,仍没有可信的 ARR、利用率或客户集中度披露当前估值仍靠叙事支撑,无法扎实承销不追加新资金
优先资本出现优先股、担保或项目层结构拿走大量经济利益或控制权规模尚未证实,普通股上行已先被抽走从零重做股权结构承销
Batam 承诺弱化交易对手、量或承购经济性未能转成可见合同拿掉牛市情景的主要规模驱动切换到基准或熊市情景
监管或社区摩擦上升政策匹配、电力接入或本地社会许可明显恶化拉长收入兑现周期,抬高执行风险提高折现率,下调公允价值
在位者更快完成主权本地化超大云厂商或成熟业主凭更强资产负债表,提供类似的区域内 AI 容量压缩差异化和定价权下调战略溢价假设

这几个变量最直接决定 Firmus 是有意思的战略资产,还是在当前估值下缺乏吸引力的普通股入口。

[CV034, CV035, CV037, CV041, CV051, CV056]
最终尽调清单
主题缺失证据重要性负责人或尽调路径
当前 ARR / 年化收入按业务线拆分的当前经常性收入、已确认收入和增长桥没有这组数据,当前估值无法严谨对标任何公开或私有可比组管理层材料,以及经审计或董事会口径 KPI 摘要
利用率和单位经济性园区利用率、毛利率、电力成本假设,以及每 token 成本或等效工作负载经济性决定效率主张能否真正转成股权价值运营模型审阅和现场走访
已签约承购交易对手Southgate 和 Batam 交易对手的名称、信用质量、期限和定价区分叙事需求和可融资、可银行化需求合同审阅,附交易对手集中度表
股权结构表和轮次条款A$330m 条款清单、清算优先权、担保顺位、董事会权利,以及任何老股转让部分决定普通股下行和稀释悬念法律尽调和完整股权结构表滚动更新
未来融资计划Southgate、Batam 和南澳大利亚的项目层资本开支计划与资金来源说明增长能否在不采用惩罚性结构的情况下融资按站点列出的资金用途与来源计划
客户集中度和续约头部客户占比、续约画像,以及按同期群拆分的销售管线转化检验合作伙伴标识能否转成持久经常性价值收入集中度明细和同期群留存材料
许可和电网里程碑电力、输电和本地审批的时间线、依赖项和应急方案执行滑坡是估值压缩最快路径之一监管事项追踪表和公用事业方往来函件

这些问题是从跟踪观察推进到可承销投资观点所需的最低资料包。

[CV039, CV040, CV049, CV050, CV052, CV053]
FV001: 建议逻辑

当前建议以真实融资事件和战略验证为起点,但在现有估值下,披露不足和资本结构风险要求保持谨慎。

[CV002, CV039, CV040, CV041, CV049, CV055]

8.5 图表

免责声明

本报告是基于公开证据的尽调快照,不构成投资建议。重要的财务、法律、技术和合同事实仍未公开;作出任何投资决定前,应直接向管理层和一手文件核实。

证据索引

结论
编号陈述可信度来源
CO001 Firmus traces its founding to 2019. SO002, SO020, SO024
CO002 Firmus maintains a registered Sydney office for shareholder communications at Level 14, 333 George Street, Sydney. SO008
CO003 Firmus publicly presents itself as a Singapore-headquartered or Singapore-based company in several 2025-2026 partnership and news materials. SO011, SO015, SO020, SO024
CO004 Firmus describes itself as a vertically integrated developer and operator of AI infrastructure or AI factories rather than a generic colocation provider. SO001, SO005
CO005 Firmus says it designs and operates the stack from the chip to the grid. SO005, SO012
CO006 Project Southgate in northern Tasmania is Firmus’s flagship sovereign AI infrastructure initiative. SO003, SO004, SO006
CO007 Firmus has live AI cloud operations in Singapore and uses that footprint to serve research, enterprise, and government workloads. SO003, SO007, SO011
CO008 Firmus closed an A$330 million equity placement in September 2025. SO003, SO019, SO020, SO021
CO009 The September 2025 financing closed at a A$1.85 billion post-money valuation. SO003, SO019, SO021
CO010 The September 2025 raise was described as materially upsized and attracted institutional plus high-net-worth Australian investors beyond the cornerstone backers. SO003, SO020
CO011 Morgans was sole lead manager and Highbury Partnership financial adviser on the September 2025 raise. SO003, SO019, SO021
CO012 Project Southgate is designed around 36,000 NVIDIA GPUs built over two stages. SO003, SO019, SO021
CO013 Official Tasmania-related materials frame Southgate stage delivery at 44MW in stage 1a and 90MW after stage 1b by 2026, with a further 300MW second stage planned later. SO004, SO018, SO024
CO014 Stage 1a of Tasmania’s AI Factory Zone was described as involving up to A$2.1 billion of investment over 12 months. SO004, SO018, SO024
CO015 The Tasmanian Government release projected up to 100 direct jobs from Southgate stage 1a with hundreds more supported indirectly. SO004
CO016 Firmus’s core infrastructure unit is the modular HyperCube AI Factory platform. SO005, SO016
CO017 Firmus AI Cloud offers GPU compute, bare metal clusters, RDMA storage, and managed cloud services for AI and HPC workloads. SO007
CO018 Firmus states that its AI Cloud environment meets ISO 27001 and SOC-2 requirements. SO007
CO019 AI Singapore partnered with Firmus to support SEA-LION and other sustainable regional AI research workloads. SO010, SO017, SO027
CO020 HTX signed a 2025 MoU with Firmus to research sustainable AI infrastructure for Singapore public-safety systems. SO014
CO021 MPA signed a 2025 MoU with Firmus to study seawater-cooled modular AI infrastructure around Singapore’s waterfront areas. SO015, SO024
CO022 ST Telemedia Global Data Centres announced a significant investment into a 2023 venture with Firmus to launch Sustainable Metal Cloud in Singapore. SO016
CO023 VAST Data said in February 2026 that Firmus selected the VAST AI Operating System as a foundational data layer for sovereign AI factories across Asia-Pacific. SO025
CO024 Firmus joined NVIDIA’s expanded DGX Cloud Lepton marketplace in June 2025 using Singapore- and Australia-based infrastructure. SO011
CO025 Firmus announced in June 2026 a Batam, Indonesia campus with NVIDIA covering up to 170,000 accelerators and 360MW through a longer partnership horizon. SO013
CO026 SmartCompany reported that Firmus planned to list publicly in 2026. SO020
CO027 SmartCompany reported that Ellerston investment director David Leslie was set to join the Firmus board after the 2025 raise. SO020
CO028 SmartCompany identified Regal Funds Management, Archibald Capital, Tectonic Investment Management, Alex Waislitz, and the Pratt family as part of Firmus’s shareholder base around the 2025 raise. SO020
CO029 Independent reporting names Jonathan Levee alongside Tim Rosenfield and Oliver Curtis as a co-founder of Firmus. SO020, SO024
CO030 Public materials do not disclose full board composition, voting control, or investor rights in enough detail to map Firmus governance with confidence. SO008, SO020
CO031 Revenue, ARR, customer count, and audited headcount are not publicly disclosed in the reviewed source set. SO001, SO003, SO007, SO020
CO032 Firmus’s careers page shows hiring across Australia, Singapore, and San Francisco in engineering, operations, finance, security, and corporate development roles. SO009
CO033 Firmus says its infrastructure can use up to 60% less energy and up to 99% less cooling water than traditional data-centre approaches. SO004, SO012, SO020
CO034 SMC claims up to 48% lower CO2 emissions for H100 training in Singapore versus an air-cooled H100 baseline in a 1.30 PUE data centre. SO016
CO035 Firmus publicly emphasizes MLPerf-style benchmarking and independently reviewed power measurements as part of its technical credibility narrative. SO027
CO036 The prompt-supplied firmus.ai domain currently resolves to a different construction-document AI site, while the AI-infrastructure company’s active public web presence is on firmus.co. SO001, SO026
CO037 Firmus’s 2026 energy and water policies are framed as an explicit response to the Australian Government’s expectations for data centres and AI infrastructure developers. SO012, SO022, SO023
CO038 ABC’s July 2025 coverage records live political concern that Tasmania may not have enough renewable power for Southgate’s later expansion stages. SO018
CO039 ABC’s interview with UNSW AI scientist Toby Walsh argues Southgate may create fewer long-run operating jobs than promotional materials imply. SO018
CO040 Firmus positions its sovereign-compute offering toward researchers, enterprises, governments, and other users that need in-region AI training or inference. SO003, SO007, SO014
CO041 SmartCompany reported that co-founder Oliver Curtis had been found guilty of insider trading in 2016, before Firmus was founded. SO020
CM001 Firmus positions itself as an AI-factory operator serving sovereign AI training and inference rather than as a generic colocation landlord. SM020
CM002 Southgate is described as infrastructure for both AI training and inference, placing Firmus across physical campus and compute-service layers rather than in a single narrow market bucket. SM020
CM003 The most defensible included spend for Firmus covers AI-factory capacity, AI-cloud or GPUaaS delivery, and sovereign-compute programs, while excluding commodity enterprise colocation, generic SaaS, and merchant semiconductor revenue. SM020, SM022, SM027
CM004 Hyperscalers and conventional colocation providers remain the status-quo substitutes because they already control much of the buyer relationship, pre-lease scarce capacity, and self-build when economics justify it. SM006, SM011, SM022
CM005 IEA base-case analysis puts data-center electricity demand around 415 to 460 TWh in 2024 and roughly 945 to more than 1,000 TWh by 2030. SM001, SM003
CM006 IEA says total data-center electricity demand rose 17% in 2025 and AI-focused facilities grew even faster. SM002, SM014
CM007 JLL projects roughly 97 to 100 GW of new global data-center capacity between 2026 and 2030, implying about 14% CAGR and a doubling of sector size. SM007, SM006
CM008 Published capex lenses for the AI data-center buildout diverge materially, with JLL framing up to $3 trillion by 2030 and McKinsey framing about $7 trillion of global spending by 2030. SM007, SM005
CM009 JLL expects Asia Pacific to deliver 4.8 GW of new supply by 2027 and says 78% of that near-term supply is already preleased. SM009
CM010 JLL says grid-connection waits in APAC run from about 24 months in emerging markets to more than eight years in core markets. SM009, SM008
CM011 DatacenterDynamics reported that APAC’s 2025 development pipeline reached 19.4 GW, including 3.7 GW under construction and 15.7 GW planned. SM025
CM012 Southeast Asia accounted for 31% of APAC under-construction capacity in 2025, making it the largest construction share in the regional pipeline. SM025
CM013 Johor and Mumbai are among APAC’s fastest-growing markets, while Johor and Batam gain attention because they offer more scalable land and power than tighter hubs such as Singapore. SM025, SM011
CM014 CBRE says Singapore remained one of APAC’s tightest and most expensive data-center markets in 2026 at roughly 2% vacancy and $330 to $475 per kW per month pricing. SM011, SM012
CM015 DatacenterDynamics says the APAC colocation pipeline alone requires about $116 billion of buildout capital over the next five to seven years. SM026
CM016 Neocloud providers were projected by JLL-cited analysis to grow about 82% CAGR through 2025 as buyers scrambled for AI-ready GPU capacity. SM015
CM017 Gartner expects neocloud providers to capture 20% of a $267 billion AI cloud market by 2030, implying about $53 billion of revenue on that narrower AI-cloud-share lens. SM022
CM018 ABI Research’s broader GPUaaS lens puts the 2030 neocloud opportunity around $250 billion, preserving a much larger estimate than Gartner’s narrower share-of-AI-cloud framing. SM024, SM022
CM019 ABI expects inference workloads to account for about 80% of neocloud revenue by 2030, shifting the category from training relief toward production AI operations. SM024
CM020 Gartner forecasts sovereign cloud IaaS spending to reach $80 billion in 2026, up 35.6% from 2025, with governments remaining the main buyers. SM027
CM021 TheCUBE Research says customers may direct several trillion dollars of cumulative spend toward sovereign and GPU-specialized clouds over the next decade, including more than $1 trillion of neocloud infrastructure investment and about a quarter-trillion of sovereign-cloud infrastructure investment. SM023, SM028
CM022 Firmus does not map to one clean TAM because electricity demand, physical MW buildout, cloud-service revenue, sovereign-cloud spend, and neocloud GPUaaS revenue are all relevant but non-additive lenses. SM005, SM007, SM022, SM024, SM027
CM023 AI-native startups and model builders are natural neocloud users because they value fast GPU access, flexible contracts, and willingness to adopt nontraditional infrastructure stacks. SM015, SM022, SM028
CM024 Enterprise AI teams are heavy users of AI compute but often buy through cloud, procurement, or central IT budgets rather than directly financing dedicated campuses. SM006, SM022
CM025 Governments, public research labs, and critical-infrastructure operators are the clearest sovereign-compute payers because jurisdiction, auditability, and national-interest criteria matter alongside throughput. SM027, SM018, SM019
CM026 Hyperscalers validate demand but also shrink Firmus’s directly reachable market because they self-build, pre-lease supply, and are launching their own sovereign offerings. SM006, SM007, SM022
CM027 Colocation landlords and infrastructure partners remain important channel actors because much AI demand is landing in leased capacity rather than in enterprise-owned facilities. SM006, SM011
CM028 Category growth is being pulled by AI implementation, cloud adoption, and digitalisation across APAC rather than by one standout national market alone. SM010, SM009
CM029 Inference-heavy AI workloads are becoming the main design point for new AI infrastructure, with JLL expecting inference to overtake training after 2027 and represent a major share of workloads by 2030. SM007, SM024
CM030 AI facilities are moving toward rack densities around 100 kW and specialized liquid-cooling requirements, which is far beyond standard enterprise-colocation assumptions. SM007, SM015
CM031 Power availability is now the dominant site-selection constraint, with some markets facing multi-year delivery waits and core APAC hubs pushing demand into Malaysia, Thailand, and Indonesia. SM008, SM009, SM011
CM032 Supply chains for transformers, gas turbines, advanced chips, and other IT components tightened further during 2025, so deployment timing is constrained by hardware and grid inputs as much as by customer demand. SM002, SM005
CM033 Cooling and water management are now adoption constraints because denser AI facilities invite scrutiny over water use and local grid stress. SM016, SM018, SM019
CM034 Singapore’s Green DC Roadmap makes energy efficiency, low-carbon operations, and system-level sustainability part of expansion eligibility rather than optional marketing extras. SM016, SM017
CM035 Australia’s 2026 expectations explicitly test national interest, energy transition, water, jobs, and local capability, turning sustainability and community fit into a permitting screen for AI factories. SM018, SM019
CM036 Firmus says Southgate’s flagship campus is designed for 84MW of critical IT load, PUE below 1.10, and 99% less water than traditional cooling. SM020
CM037 Firmus says its AI-factory model is designed to reduce demand when power prices spike and to match or exceed its own load with new renewable and storage commitments. SM021
CM038 Localized control over data, operations, and governance is becoming a material purchase driver for AI infrastructure outside the United States and China. SM027, SM022
CM039 Rising rents, low vacancy, and capex intensity mean buyers still need utilization confidence before committing to bespoke AI capacity. SM007, SM011, SM026
CM040 Public sources are not enough to build a bottom-up SOM for Firmus because they do not disclose customer mix, contract duration, utilization, realized pricing, or live workload mix. SM020, SM022, SM026
CM041 IEA says electricity consumption from AI-focused data centers is poised to triple by 2030 even though power use per AI task is falling quickly. SM002, SM003
CM042 IEA expects renewables to meet nearly half of additional data-center electricity demand through 2030, but gas and coal still supply a large share of incremental load. SM001
CM043 Gartner identifies mature Asia/Pacific as one of the fastest-growing sovereign-cloud regions in 2026, supporting a regional demand case for Firmus rather than a purely Western one. SM027
CM044 Singapore’s power limits are already shifting growth into neighboring markets such as Malaysia and Indonesia, which supports a regional hub-and-spoke logic for AI-factory deployments. SM008, SM011, SM025
CM045 AI factories differ from status-quo colocation because compute density, liquid cooling, orchestration, and grid behavior are part of the product rather than merely attributes of the building shell. SM015, SM020, SM021
CP001 Firmus competes across four buyer alternatives: hyperscaler GPU clouds, AI-ready landlords, AI-specialized neoclouds, and internal build for the largest buyers. SP002, SP011, SP013, SP023, SP029
CP002 The hyperscalers are direct substitutes because they bundle GPU instances, regional footprints, security controls, and adjacent cloud services into one procurement path. SP001, SP005, SP007, SP008
CP003 AWS markets P5, P5e, and P5en instances with up to eight H100 or H200 GPUs, up to 3,200 Gbps of EFA networking, and UltraClusters that scale to 20,000 H100 or H200 GPUs. SP001
CP004 AWS says its cloud spans 39 geographic regions and 123 availability zones, giving it far broader physical reach than any regional operator. SP002
CP005 AWS says its 2024 global data-center PUE was 1.15 and that it uses configurable liquid-to-chip cooling for AI processors, pairing scale with efficiency claims. SP002, SP003
CP006 Google Cloud positions A4, A3, and A2 accelerator-optimized machine families as GPU-native building blocks and pairs them with usage-based billing, Spot discounts, and one- or three-year commitments. SP004, SP006
CP007 Google Cloud says it operates across 43 regions and 130 zones, connected by 10 million kilometers of fiber across 200-plus countries and territories. SP005
CP008 Azure's ND H100 v5 series exposes eight H100 GPUs per VM, scales to thousands of GPUs, and uses 400 Gb/s InfiniBand per GPU for tightly coupled AI training. SP008
CP009 Azure's geography map shows a wide in-region and sovereign-residency footprint across Asia Pacific, Australia, Indonesia, Malaysia, and sovereign options in Germany. SP007
CP010 Microsoft now openly describes purpose-built AI datacenters as AI factories, with Fairwater alone representing tens of billions of dollars of investment and hundreds of thousands of AI chips. SP009
CP011 Equinix's 2025 annual report says it had 280 data centers, 10,500-plus customers, 77 markets, 507,000-plus interconnections, and $9.2 billion of revenue. SP010, SP011, SP012
CP012 Equinix differentiates with an interconnection marketplace linking 3,000-plus clouds, 2,000-plus networks, and 5,500-plus enterprises, plus AI-ready high-density sites with 99.9999%-plus uptime. SP011, SP012
CP013 Digital Realty sells private, hybrid, and sovereign AI infrastructure built around high-density colocation, low-latency interconnection, and in-region deployment. SP013, SP014
CP014 Digital Realty's partner-validated AI infrastructure offers let it package private AI deployments without needing to be a public neocloud brand itself. SP013, SP014
CP015 AirTrunk says it is well capitalized to fund hyperscale expansion across Asia-Pacific and the Middle East, and its Blackstone-led A$24 billion acquisition shows how much capital can back that buildout. SP015, SP016
CP016 AirTrunk's main edge over Firmus is regional hyperscale capital and relationships with global technology customers, while its public cloud software and pricing layers are much less disclosed. SP015, SP016
CP017 NEXTDC reported FY25 revenue of A$427.2 million, a net loss of A$60.5 million, and 72.2 MW of new contracted utilisation, up 42% year over year. SP017
CP018 NEXTDC says its contracted pipeline exceeds everything it has built to date and positions the company to more than double revenue and EBITDA over the next few years. SP017
CP019 NEXTDC's sovereign-AI build path now includes S4 Sydney at 350 MW, S7 Sydney at 550-plus MW, and M3 Melbourne at 225 MW. SP017, SP018, SP019
CP020 NEXTDC is pushing a high-density cooling story too: M3 targets a 1.29 average PUE, the portfolio reported 1.44 PUE in FY25, and the company says it deployed its first 40 MW direct-to-chip liquid-cooling system. SP017, SP018
CP021 Keppel DC REIT ended 2025 with 25 data centres across 10 countries, about $6.3 billion of assets under management, 95.8% occupancy, and a 6.7-year weighted average lease expiry. SP020
CP022 Keppel is closer to a landlord and portfolio allocator than to an integrated AI cloud, so it pressures Firmus mainly on regional capacity ownership and balance-sheet staying power. SP020
CP023 GDS says it offers colocation, managed hosting, and managed cloud services to large Chinese data customers, making it an incumbent capacity operator rather than an AI-native GPU cloud. SP021
CP024 GDS reported FY2025 revenue of RMB11.43 billion and a 75.5% utilization rate, evidence of a scaled China incumbent that can compete for regional enterprise and hyperscale demand. SP021, SP022
CP025 CoreWeave presents itself as an AI-native, Kubernetes-native cloud with software-defined liquid cooling, rack-scale networking, and early access to NVIDIA GPUs. SP023
CP026 CoreWeave's 2025 10-K says it had $60.7 billion of remaining performance obligations at year-end and that its committed contracts had a weighted-average term of about five years. SP024
CP027 CoreWeave's 2025 revenue increased by $3.2 billion or 168%, but 67% of revenue came from Microsoft, showing both exceptional scale and meaningful customer concentration. SP024
CP028 CoreWeave explicitly lists AWS, Google, Microsoft, and Oracle as larger rivals that can use broader portfolios, lower pricing, bundling, and data-egress friction to win business. SP024
CP029 Lambda markets modular AI factories with direct-to-chip liquid plus precision air cooling, a roadmap toward 1 MW rack-scale designs, and enterprise compliance certifications. SP025
CP030 Lambda publishes unusually transparent AI-cloud pricing, including B200 SXM6 at $6.69 per hour and H100 SXM at $3.99 per hour, alongside cluster offers from 16 to 2,000-plus GPUs. SP026
CP031 Crusoe says it designs and builds data centers, operates Crusoe Cloud, manufactures critical electrical components in-house, and has 3.0 GW of active projects under development. SP027, SP028
CP032 Crusoe says it delivered the first phase of its 1.2 GW Abilene Stargate campus in under 12 months and combines renewable-linked power, on-site backup, and direct liquid-to-chip cooling. SP027, SP028
CP033 NVIDIA now markets a full-stack AI data-center platform around Blackwell, networking, and accelerated-computing software, making supplier-led standardization and self-build more credible. SP029
CP034 Hyperscalers offer the broadest public footprint for data residency and compliance, so Firmus's sovereignty edge is strongest only where buyers require local physical control rather than merely in-region cloud. SP002, SP005, SP007
CP035 Equinix and Digital Realty attack Firmus through partner ecosystems and interconnection, letting customers stitch private AI, data, and clouds together without depending on a smaller operator. SP012, SP013, SP014
CP036 AirTrunk, NEXTDC, Keppel, and GDS compete primarily on land, power, and scarce capacity rather than on developer workflow, which compresses Firmus's opportunity when buyers split real estate from cloud experience. SP015, SP017, SP020, SP022
CP037 CoreWeave, Lambda, and Crusoe are the closest integrated peers because they combine AI-specific facilities or power design with cloud delivery, not just bare colocation. SP023, SP025, SP027, SP028
CP038 Public pricing is transparent mainly in hyperscaler-style cloud offers and Lambda's self-serve model; enterprise colo, sovereign campus, and most committed neocloud contracts remain quote-based or private. SP001, SP006, SP024, SP026
CP039 Switching costs are highest where buyers adopt adjacent cloud, interconnection, and procurement rails, not where they only rent raw megawatts or rack space. SP012, SP014, SP024
CP040 Firmus's moat is most defensible in APAC sovereign campuses that combine energy narrative, customization, and local physical control; it is weakest when buyers can accept hyperscaler regions or established AI-ready landlords. SP013, SP015, SP017, SP027
CP041 Distribution power today sits with hyperscalers' billing and compliance rails, Equinix's interconnection marketplace, Blackstone-backed AirTrunk, and neoclouds with anchor-customer contracts. SP002, SP012, SP016, SP024
CP042 Likely entrant risk is high because Microsoft already talks about AI factories, AWS sells UltraClusters plus liquid-cooled H100 and H200 fleets, and NVIDIA markets a full-stack AI data-center reference stack. SP001, SP009, SP029
CP043 Internal build remains viable only for the largest buyers that can manage NVIDIA stacks or reserve cluster capacity directly on hyperscalers, which caps Firmus's pricing power at the high end. SP002, SP008, SP029
CP044 Firmus is competing in a market where capital access and hardware allocation can outweigh clever design, so displacement risk rises if rivals secure capacity faster than Firmus converts sovereign demand into contracts. SP016, SP024, SP028
CI001 Firmus publicly offers AI Cloud Compute as on-demand instances and reserved clusters for AI and HPC workloads. SI001
CI002 The AI Cloud Compute page publicly exposes H200-based instance specifications and built-in GPU-cost observability, indicating metered infrastructure even though the commercial meter is undisclosed. SI001
CI003 Firmus publicly offers dedicated bare-metal GPU clusters by reservation, including single-tenant and multi-rack configurations. SI002
CI004 Bare Metal marketing emphasizes 24/7 operational support and predictable reserved performance, implying contractual service obligations beyond raw hardware access. SI002
CI005 Firmus Cloud Services publicly add AIFactoryOS, managed Slurm, CUDA stacks, observability, and hybrid connectivity on top of the infrastructure layer. SI003
CI006 No reviewed Firmus product page publishes per-GPU-hour, per-instance, per-cluster, or managed-service list pricing as of 2026-07-02. SI001, SI002, SI003
CI007 The public monetization surface implies at least three revenue layers: usage-based cloud compute, reserved bare-metal capacity, and managed cloud-services tooling/support. SI001, SI002, SI003
CI008 Repeated “Enquire,” “available by reservation,” and “available on request” language implies a contact-led procurement motion rather than credit-card self-serve cloud pricing. SI001, SI002, SI003
CI009 NVIDIA and Firmus materials show Firmus participates in DGX Cloud Lepton, where customers can access regional GPU capacity on either on-demand or long-term terms through a shared marketplace. SI006, SI013, SI014
CI010 The AI Singapore case study is public evidence that Firmus already delivers live Singapore-based compute rather than only future-campus promises. SI007, SI027
CI011 Firmus publicly cites 32 nodes, 256 H200 GPUs, over 200 experiments, and a 27B-parameter model trained in 10 days for the AI Singapore engagement. SI007
CI012 Public sources position Firmus against research, enterprise, government, and developer workloads rather than low-touch consumer segments. SI001, SI003, SI007, SI012
CI013 Because product packaging is visible but realized pricing is not, the public record can describe revenue surfaces without proving revenue quality. SI001, SI002, SI003
CI014 DGX Cloud Lepton may reduce top-of-funnel friction by routing developers to regional GPU supply, but it does not eliminate the implementation burden of reserved clusters and sovereign workloads. SI006, SI013, SI014
CI015 Reviewed public materials do not disclose CAC, payback, pipeline conversion, sales-cycle length, NRR, or support headcount as of 2026-07-02. SI001, SI003, SI005, SI023
CI016 Firmus officially says it closed a A$330 million equity placement in September 2025 at a A$1.85 billion post-money valuation with Ellerston Capital as cornerstone investor and NVIDIA participating. SI005, SI027
CI017 The stated use of proceeds from the 2025 raise is to accelerate Project Southgate. SI005
CI018 Project Southgate is publicly framed as a 36,000-GPU sovereign campus built in two stages and as the largest deployment of Firmus' AI Factory platform to date. SI005, SI008, SI027
CI019 The Southgate page publicly lists 84MW of critical IT load, sub-1.10 PUE, and 99% less water than traditional cooling for the Launceston flagship. SI008, SI009
CI020 Firmus says the Launceston design uses water for cooling only on the hottest days and estimates annual cooling-water use at roughly 20 Tasmanian households. SI009
CI021 Firmus' June 2026 South Australia deal is a 12-year, 600MW wholesale electricity agreement linked to 1.2GW of new renewable generation and 1.5GWh of new battery storage by 2032. SI004, SI009
CI022 The South Australia structure contractually commits Firmus to reduce electricity consumption for up to 220 hours per year when price thresholds signal grid stress. SI004, SI009
CI023 Firmus publicly says it will pay commercial energy prices, fund transmission and connection upgrades, and avoid subsidies or special deals. SI009, SI010
CI024 ABC reported the Launceston AI factory as roughly A$2.1 billion and quoted Firmus saying the first stage required 90MW of energy. SI011
CI025 Across product and policy pages, the public cost stack appears dominated by GPUs, power, cooling, networking, orchestration, facilities, and round-the-clock support rather than by software-only opex. SI001, SI002, SI003, SI009
CI026 CoreWeave's S-1 shows that AI-cloud revenue can be extremely concentrated, with 77% of 2024 revenue from its top two customers and 62% from the largest one. SI015
CI027 CoreWeave's S-1 also shows multi-year take-or-pay revenue and $15.1 billion of remaining performance obligations as of December 31, 2024. SI015
CI028 CoreWeave disclosed more than 250,000 GPUs, over 360MW of active power, about 1.3GW of contracted power, $12.9 billion of debt commitments, and $2.6 billion of operating lease liabilities, illustrating financing-heavy AI-cloud expansion. SI015
CI029 Equinix's 2025 annual report shows what mature infrastructure monetization looks like: $9.2 billion of revenue, $1.6 billion of annualized gross bookings, 10,500+ customers, and 49% adjusted EBITDA margin. SI017
CI030 Equinix's 10-K says its cost of revenues is dominated by depreciation, leased-facility rent, electricity and other utilities, bandwidth, personnel, maintenance, supplies, and security, with most of the base fixed until new capacity is opened. SI018
CI031 Equinix also disclosed $23.6 billion of property, plant and equipment, about $2.1 billion of non-capital commitments including power purchases, and 12-year operating lease duration, underscoring long-lived capital lock-in. SI018
CI032 CoreWeave, Equinix, Digital Realty, NEXTDC, and AirTrunk all maintain dedicated quarterly, annual, or report portals, highlighting how thin Firmus' public KPI disclosure remains by comparison. SI016, SI019, SI020, SI021, SI023, SI026
CI033 NEXTDC says record contracted utilisation growth and a fully funded A$2.2 billion capital plan are being used to accelerate AI-ready infrastructure at scale. SI022, SI023
CI034 AirTrunk says it closed a A$16 billion ex-Japan sustainability-linked refinancing and now has an A$18 billion-plus financing platform, showing that hyperscale expansion often depends on very large debt structures. SI024, SI026
CI035 AirTrunk's Malaysia release describes 280MW of new IT load, more than 700MW across four campuses, about US$6.8 billion of committed investment, and existing campuses that are almost 100% contracted. SI025
CI036 Firmus has public technical-traction signals—AI Singapore workloads, NVIDIA Cloud Partner distribution, and Singapore cloud/public-sector references—but no comparable disclosure of revenue or unit-economics KPIs. SI006, SI007, SI012, SI013
CI037 Reviewed public sources do not disclose Firmus' current revenue, ARR, customer count, utilization, gross margin, cash, burn, or debt stack as of 2026-07-02. SI001, SI003, SI005, SI009, SI023
CI038 SmartCompany reported that Firmus is expected to continue raising capital ahead of a proposed 2026 ASX listing. SI028
CI039 Independent Tasmanian reporting shows power availability and long-run jobs claims remain contested rather than universally accepted. SI011
CI040 The latest disclosed equity round is meaningful, but it does not fully de-risk a model that publicly contemplates multistage campuses plus long-dated power, storage, and transmission commitments. SI004, SI005, SI009, SI022, SI025
CI041 Revenue quality is presently not underwritable because realized pricing, contract duration, concentration, utilization, and margin data remain private. SI001, SI002, SI003, SI015
CI042 Firmus' public obligations and infrastructure posture make the business resemble data-center or project-finance capital structures more than an asset-light software model. SI004, SI009, SI010, SI015, SI018, SI024
CI043 The main diligence blocker is the absence of a current cash, burn, and financing bridge tied to site-specific capex, grid-connection costs, and customer pre-commitments. SI005, SI009, SI011, SI022, SI024
CI044 Independent coverage rounds the latest valuation to about A$1.9 billion while Firmus and ARN state A$1.85 billion post-money, so even basic financing facts need source control and exact-document confirmation. SI005, SI027, SI028
CI045 Near-term revenue likely depends more on Singapore cloud and partner channels than on megacampuses that are still being built or still tied to future power delivery. SI006, SI007, SI011, SI025
CE001 The public AI Cloud surface spans compute, storage, bare metal, cloud services, and cloud applications as separately marketed modules. SE001, SE002, SE003, SE004, SE005
CE002 Firmus markets a workflow that starts with GPU access and then layers storage, orchestration, and application kits rather than a single black-box SaaS product. SE001, SE002, SE003, SE004, SE005
CE003 Cloud Compute is marketed for large-model training, agentic AI, ML pipelines, and CUDA or HPC workloads. SE001
CE004 Cloud Applications adds CUDA dev environments, a data-science stack, AI Workbench, and NIM inference APIs for developer workflows. SE005
CE005 Cloud Compute offers both on-demand instances and reserved clusters. SE001
CE006 Bare Metal is reservation-led and positioned as single-tenant dedicated infrastructure. SE003
CE007 The public GPU lineup includes H200, H100, A100, and L40S options. SE001
CE008 AI Storage is positioned as RDMA-accelerated NVMe storage for datasets, checkpoints, and model artifacts. SE002
CE009 Firmus says its AI Cloud serves developers, enterprises, educational institutions, and government users. SE026
CE010 Engineering Principles describes HyperCubes as multi-petascale, highly available, modular, and thermally optimized compute-scale instruments. SE007
CE011 Each HyperCube module is described as 32 NVL72 racks and two NVIDIA Scale Units. SE007
CE012 Firmus ties cloud delivery to physical AI-factory assets in Singapore and Australia, and to a roadmap campus in Batam. SE007, SE012, SE027
CE013 H200 Cloud Compute nodes are publicly specified with eight NVIDIA H200 GPUs. SE001, SE015
CE014 The same H200 node spec advertises about 1.13 TB of total HBM3e memory. SE001, SE015
CE015 The H200 node spec uses NVLink 4.0 and NVSwitch 3.0 for intra-node GPU interconnect. SE001, SE015
CE016 The H200 node spec pairs GPUs with dual Intel Xeon Platinum 8462Y+ CPUs and 2 TB DDR5 system memory. SE001
CE017 Public compute networking options include dual 200–800 Gb/s InfiniBand or 400 Gb/s Ethernet with RDMA or RoCE v2 support. SE001
CE018 Bare Metal clusters are marketed at four to eight H200 GPUs per node. SE003
CE019 Bare Metal adds InfiniBand plus high-throughput RDMA or RoCEv2-capable storage for distributed jobs. SE003
CE020 Slurm is named across Cloud Compute, Bare Metal, and Cloud Services as the scheduler for multi-GPU environments. SE001, SE003, SE004
CE021 Cloud Services presents AIFactoryOS as a proprietary orchestration and telemetry layer for governance, workload automation, and system-wide visibility. SE004
CE022 Engineering Principles says the factory operating system integrates telemetry, cooling, GPU orchestration, and grid interaction into one layer. SE007
CE023 Firmus selected VAST AI OS as a foundational data layer for next-generation sovereign AI factories. SE018, SE025
CE024 VAST describes Firmus's model-to-grid architecture as an optimization framework spanning model behavior, GPU performance, thermal management, and grid conditions. SE018, SE025
CE025 VAST says the chosen data layer is high-throughput, disaggregated, and aligned with NVIDIA Cloud Partner reference designs. SE018, SE019, SE025
CE026 NVIDIA describes the H200 as a high-memory Hopper GPU tuned for generative AI and HPC, matching Firmus's decision to foreground H200 nodes. SE001, SE015
CE027 NVIDIA's InfiniBand platform highlights SHARP, self-healing, quality of service, adaptive routing, and hypercube-supporting topologies that fit the distributed-training profile Firmus advertises. SE016, SE003
CE028 DGX Cloud Lepton is presented by NVIDIA as a common workflow for development, training, and inference across regional cloud providers. SE017
CE029 Firmus says its Lepton participation contributes Singapore- and Australia-based infrastructure to that marketplace. SE009, SE017
CE030 Firmus's MLPerf page says it measured node power at immersion-rack power shelves. SE006
CE031 Firmus says those MLPerf measurements showed about 30% better node-level performance versus air-cooled H100 SXM systems. SE006
CE032 The same MLPerf page says datacenter-level PUE estimates were not within MLCommons's verification scope. SE006, SE020
CE033 MLCommons and its GitHub repository confirm that the benchmark suite itself is industry-run and publicly versioned rather than vendor-private. SE020, SE021
CE034 Bare Metal and Cloud Services both advertise observability, hybrid connectivity, and managed operations as part of deployment. SE003, SE004
CE035 Bare Metal explicitly promises 24/7 operational support for enterprise AI deployments. SE003
CE036 Cloud Compute and Cloud Applications say teams can work through CLI, Terraform, GitOps, Jupyter, and NIM APIs. SE001, SE005
CE037 Cloud Services claims ISO 27001 and SOC-2 compliance plus encryption in flight and at rest. SE004
CE038 Firmus publishes a corporate privacy policy, but the fetched public materials did not expose a product-specific AI data-processing addendum or named customer data-residency control pack. SE008, SE004
CE039 The Head of Corporate IT & Cyber Security role is tasked with securing infrastructure, applications, and identity systems for global hyperscale cloud growth in APAC. SE026
CE040 HTX and MPA collaboration materials show Firmus tailoring liquid-cooled and seawater-cooled concepts for public-safety and waterfront deployments with tight land and power constraints. SE010, SE011, SE023, SE024
CE041 Both government-linked collaborations are framed as studies or joint research rather than evidence of already-live public-sector production deployments. SE010, SE011, SE023, SE024
CE042 The clearest differentiation claim is the coupling of liquid cooling, HyperCube modularity, model-to-grid orchestration, and a disaggregated VAST data layer. SE007, SE018, SE019, SE025
CE043 The Batam announcement says HyperCube is co-designed to NVIDIA DSX blueprints to bring capacity online faster and improve tokens per watt and resiliency at scale. SE012
CE044 Firmus's public roadmap is concrete on Lepton access, VAST data-layer adoption, and Batam scale-out. SE009, SE012, SE018
CE045 The fetched public pages did not expose customer-visible release dates for AIFactoryOS features, storage classes, or formal SLA targets. SE004, SE005
CE046 The fetched public materials describe proprietary infrastructure and software, but they did not disclose patent numbers or granted IP assets for the cooling or orchestration stack. SE004, SE007, SE012
CE047 Firmus's NVIDIA dependency spans GPUs, NIM APIs, DSX blueprints, networking options, and DGX Cloud Lepton distribution. SE001, SE005, SE012, SE017
CE048 Tech Wire Asia describes the Batam plan as a 360 MW, 170,000-GPU future campus, underscoring that the largest scale claim is still a roadmap execution story rather than a currently shipped service. SE027, SE028, SE012
CU001 Firmus publicly markets AI infrastructure to a mix of AI-native builders, enterprise teams, research users, and sovereign or commercial workloads rather than to a single buyer archetype. SU001, SU015, SU016, SU027
CU002 Firmus's AI Cloud page frames the user journey as moving from experiment to deployment with Jupyter, CUDA stacks, and NIM-powered inference kits on one platform. SU001
CU003 Firmus says AI Singapore is a national AI programme launched by Singapore's National Research Foundation and that SEA-LION is its open-source regional LLM suite. SU002
CU004 AI Singapore's SEA-LION surfaces describe the model family as open-source, multilingual, and designed for Southeast Asian languages, cultures, and contexts. SU005, SU006
CU005 The Firmus-AI Singapore partnership is framed around three concrete workstreams: access to Singapore-based H200 GPUs, hosting SEA-LION, and benchmarking AI training and inference workloads. SU002, SU003
CU006 Firmus says AI Singapore used its AI Cloud platform for rapid experimentation, large-scale training, and efficient model evaluation for SEA-LION. SU003
CU007 The AI Singapore case study reports that the SEA-LION engagement deployed 32 nodes and 256 NVIDIA H200 GPUs on Firmus infrastructure. SU003
CU008 The same case study says the engagement completed more than 200 experiments and produced more than 100 candidate models for evaluation. SU003
CU009 Firmus says a 27B-parameter model was trained in 10 days on 32 nodes and a 4B-parameter model in 3.5 days on 16 nodes for the SEA-LION effort. SU003
CU010 Firmus frames the AI Singapore relationship as a long-term partnership and the case-study quotes describe the team as responsive, proactive, and effective during intensive experimentation. SU002, SU003
CU011 HTX describes its relationship with Firmus as a memorandum of understanding for joint research into advanced liquid-cooled AI infrastructure rather than as a production procurement award. SU008, SU010
CU012 HTX says the work is intended to uplift Singapore's sovereign capability in mission-critical compute for public safety and emergency response systems. SU008, SU009
CU013 HTX's release says Firmus has already been deployed in AI factories in Singapore and Australia that deliver enterprise-grade services for LLM training, inference, and agentic workloads. SU008
CU014 HTX's broader AI TechXplore recap places Firmus alongside Google, Microsoft, and Mistral within HTX's 2025 partnership stack. SU009
CU015 MPA says its collaboration with Firmus is a study of sustainable, modular AI factories using seawater for cooling, including research and pilot testing rather than a disclosed commercial deployment. SU011, SU012, SU013
CU016 MPA says any seawater-cooled deployment path must account for navigation safety, pollution-control rules, siting of seawater intakes, discharge management, and environmental impact analysis. SU011, SU014
CU017 MPA says the study will engage local research institutions and industry stakeholders as part of the workstream. SU011, SU014
CU018 Independent coverage of the MPA collaboration still describes it as a sovereign-grade infrastructure exploration or feasibility effort rather than a live customer deployment. SU013, SU014
CU019 NVIDIA describes DGX Cloud Lepton as a unified AI platform for AI natives, model builders, and fast-iterating teams that need one workflow across development, training, and inference. SU016, SU018
CU020 Firmus says its DGX Cloud Lepton participation brings compute to developers across Asia-Pacific and helps customers meet both sovereign and commercial AI requirements. SU015
CU021 NVIDIA independently names Firmus as one of the cloud partners contributing GPU infrastructure to DGX Cloud Lepton. SU017, SU018
CU022 NVIDIA's Lepton materials repeatedly emphasize regional placement, data sovereignty, and prototype-to-production workflows, which aligns the channel with sovereign and regulated workloads as well as with startups. SU016, SU017, SU018
CU023 Firmus's Batam announcement explicitly targets AI-native, enterprise, and ISV customers for NVIDIA-powered cloud services. SU027
CU024 Reuters says the NVIDIA partnership is intended to help smaller and emerging AI firms access infrastructure more cost-effectively, with Firmus selling NVIDIA-powered cloud services to AI Native customers among others. SU026
CU025 Tech Wire Asia says the Batam site is planned as a multi-tenant project serving AI-native customers, while Firmus's Australian projects are aimed at hyperscaler customers. SU028
CU026 The same Tech Wire Asia report says a Southgate project has secured an unnamed global hyperscaler customer, but the counterparty is not publicly identified. SU028
CU027 STT GDC's venture with Firmus launched a GPU-centric bare-metal IaaS offering intended to serve AI use cases for businesses, governments, and society through a shared channel model. SU023
CU028 VAST says its data-layer partnership with Firmus is intended to support anchor-tenant and government-backed workloads as AI capacity scales across Asia-Pacific. SU024
CU029 No reviewed public source discloses Firmus customer count, active account count, net revenue retention, gross retention, churn, or contract length. SU001, SU003, SU015, SU016, SU027
CU030 The strongest public advocacy signal is the AI Singapore case-study quote praising Firmus's responsiveness and smooth operations, but it remains company-published rather than independently issued by the customer. SU003
CU031 Public reference quality is uneven because AI Singapore provides concrete workload metrics while HTX and MPA disclose only study-stage or design-stage collaboration details. SU003, SU008, SU011
CU032 HTX and MPA are valid proof of public-sector engagement, but neither source publicly proves recurring public-sector compute revenue or deployed production usage at scale. SU008, SU011, SU013
CU033 No named enterprise AI team, Fortune 500, or ISV production customer was found in the reviewed public materials beyond AI Singapore and public-sector collaborations. SU001, SU015, SU016, SU027
CU034 No reviewed source discloses top-customer share, top-five customer share, or revenue mix by segment, leaving customer concentration publicly opaque. SU015, SU026, SU027, SU028
CU035 The Batam materials cite US$25-30 billion of first-six-year expected receipts from committed offtake agreements, but the customer identities and contract structures are not named in the public record. SU026, SU027, SU028
CU036 Firmus's public route to market is structurally partner-dependent because it relies on NVIDIA hardware and Lepton distribution, STT GDC hosting history, VAST's data layer, and DayOne's Batam campus development. SU015, SU017, SU023, SU024, SU028
CU037 Australia's expectations document says large AI factories should advance data sovereignty, clean energy, water efficiency, community engagement, and favorable compute access for startups, researchers, and not-for-profits. SU021
CU038 Singapore's AI-strategy materials say the country will secure more compute, embed AI more deeply across government, and strengthen itself as an AI hub while improving deployment efficiency. SU019, SU020
CU039 Gartner says governments will remain the main buyers of sovereign cloud IaaS, followed by regulated industries and critical-infrastructure organizations such as energy, utilities, and telecommunications. SU022
CU040 Reuters and Tech Wire both frame the Batam-NVIDIA partnership as a way to lower infrastructure barriers for smaller or emerging AI firms, which makes the AI-native segment a stated expansion vector rather than a fully evidenced current customer cohort. SU026, SU028
CU041 Singapore and Australia's AI cooperation MOU is designed to increase access to AI technologies, markets, talent, and research-industry linkages across government and business domains. SU029
CU042 AI Singapore's public SEA-LION materials show the models are open, community-oriented, and available through multiple external distribution surfaces, so Firmus's proof is strongest around hosting and training support rather than exclusive control of the model's distribution. SU005, SU006, SU007
CU043 NVIDIA's Lepton materials promise a consistent workflow from prototype to production across regions and providers, so Firmus's channel role can lower friction for enterprise teams even without named enterprise logos. SU016, SU018
CU044 Because SEA-LION is open-source and broadly accessible, the AI Singapore reference proves compute credibility and hosting relevance more clearly than it proves customer lock-in or exclusivity for Firmus. SU005, SU006, SU007
CU045 ABC's Tasmania coverage reports local concern about power availability and describes part of Firmus's larger second-stage plan as aspirational, highlighting that future sovereign or hyperscale demand depends on grid and planning execution. SU025
CU046 Tech Wire reports that Australian scrutiny of data centres includes questions about energy use, water consumption, noise, waste, and site selection, which can slow customer conversion even when demand is present. SU028
CU047 STT GDC publicly referenced industry-standard uptime SLAs for the earlier SMC bare-metal offer, but equivalent standalone SLA disclosure was not found on the Firmus AI Cloud pages reviewed for this chapter. SU001, SU023
CU048 The named institutional reference set is geographically concentrated around Singapore because AI Singapore, HTX, and MPA are all Singapore-linked bodies. SU003, SU008, SU011
CU049 Firmus's public customer evidence clusters around research, public-sector, sovereign, and channel narratives, while direct proof of commercial enterprise repeat usage remains sparse. SU003, SU008, SU011, SU015, SU016, SU027
CR001 The highest residual risks are power-and-approval execution, grid and community acceptance, partner concentration, and demand or financing opacity rather than demand for AI compute itself. SR001, SR010, SR012, SR035, SR036
CR002 The Australian Government says it will prioritise data-centre proposals that are most closely aligned with the Commonwealth expectations. SR001, SR002
CR003 The same expectations say energy-intensive proposals that are not closely aligned will not be prioritised by Commonwealth regulatory assessments. SR001, SR002
CR004 The expectations are framed to work alongside existing national, state, and territory laws rather than creating a separate approval regime. SR001, SR003
CR005 The Australian legal landscape for AI already reaches privacy, directors' duties, negligence, and consumer-law exposure for AI operators. SR003, SR023
CR006 Australia's 2024 cyber reforms clarified obligations to protect certain data storage systems that hold business-critical data. SR019, SR020
CR007 Those reforms also created powers to direct responses to all-hazards incidents and to force changes to deficient risk-management programs. SR019
CR008 The OAIC says privacy obligations apply both to personal information put into AI systems and to AI-generated outputs that contain personal information. SR023
CR009 The OAIC recommends that organisations avoid entering personal or sensitive information into publicly available generative AI tools as a best practice. SR023
CR010 Firmus's website terms say the company does not warrant the accuracy, completeness, or suitability of public site content and may change it without notice. SR007
CR011 Firmus says its website terms are governed by New South Wales law and that it follows the Privacy Act 1988 and OAIC APP guidelines. SR007
CR012 Tasmanian Treasury's RTI release says Firmus submitted three transmission connection enquiries and sought to expand St Leonards from 20 MVA to 104 MVA. SR036
CR013 The same RTI release says the St Leonards connection upgrade was treated as a major capital investment and that TasNetworks had not provided a business case. SR036
CR014 The RTI release says connection assets are not regulated services and that the load proponent pays the full cost of studies, connection assets, and AEMO assessment fees. SR036
CR015 Firmus is constructing St Leonards and has lodged development applications for Bell Bay and Wesley Vale. SR014, SR015, SR011
CR016 ABC's July 2026 coverage says some residents felt they had not been adequately consulted and remained worried about water and energy use. SR015, SR016
CR017 The Exeter community meeting and follow-on reporting show that social-licence risk is already active rather than hypothetical. SR014, SR015, SR016
CR018 ABC's national water reporting says Australia has more than 250 data centres and experts argue new facilities should avoid relying on drinking water where possible. SR013
CR019 ABC reported that the Bell Bay proposal requested 19.2 million litres of water a year from TasWater, although the company said it expected to use less than half. SR014
CR020 ABC reported that Firmus projected annual water use of roughly 3.3 million litres at St Leonards, 8.7 million at Bell Bay, and 700,000 at Wesley Vale. SR014
CR021 The Bell Bay FAQ says cooling water is expected to be needed on only about 10 days a year above 26 degrees Celsius, with dry cooling used otherwise. SR035
CR022 The Bell Bay FAQ says the site would draw about 288 MW and connect directly to three 220 kV TasNetworks transmission lines. SR035
CR023 The Bell Bay FAQ says the site is designed to be dispatchable and would reduce electricity use during system constraint. SR035
CR024 Cyber.gov says AI data security depends on controls such as encryption, signatures, provenance, and lifecycle safeguards because data-integrity failures can distort outcomes. SR020
CR025 The MPA and HTX arrangements are memoranda of understanding centered on study and research rather than operating approvals. SR017, SR034
CR026 The HTX collaboration focuses on sovereign public-safety compute use cases, raising expectations around security and reliability for government-facing workloads. SR020, SR034
CR027 Utility Magazine says Aurora Energy signed a three-year retail service agreement to supply up to 104 MW for Launceston, with operations ramping from August 2026 to full contracted load by November 2026. SR031
CR028 ABC reported that Oliver Curtis confirmed a 104 MW supply via Aurora using Hydro power for the initial St Leonards stage. SR010
CR029 ABC reported that if all three Tasmanian sites proceed, Firmus would become Tasmania's largest power user. SR011, SR015
CR030 ABC's Marinus coverage said the three Tasmanian sites would need more than 400 MW in aggregate. SR012, SR015
CR031 Climate Change Authority chair Matt Kean said large AI data-centre loads like Firmus's could undermine the Marinus Link business case. SR012
CR032 Firmus says the Gunvor agreement gives it a 12-year 600 MW wholesale supply arrangement linked to 1.2 GW of new renewable generation and 1.5 GWh of battery storage by 2032. SR005
CR033 Firmus says the Gunvor agreement includes a demand-response commitment that can reduce electricity consumption for up to 220 hours a year when power prices cross agreed thresholds. SR005
CR034 ABC's March 2026 power-deal story says another manufacturer's request for more power had been rejected less than a year earlier. SR010
CR035 The same ABC story says Boyer Paper Mill's request for an additional 45 MW was declined while the Firmus deal proceeded. SR010
CR036 AEMO now forecasts data centres as a standalone electricity-demand category and estimated they used about 4 TWh, or 2.2% of NEM demand, in FY2025. SR024
CR037 AEMO forecasts data-centre demand could rise about 25% a year to around 12 TWh by 2029-30 under its Step Change scenario. SR024
CR038 IEA said data-centre electricity use grew 17% in 2025 versus 3% overall electricity-demand growth, showing how quickly supply bottlenecks can tighten. SR025
CR039 JLL said global data-centre demand is surging despite supply and power constraints, making early power access a strategic advantage. SR026
CR040 WEF said grid connectivity is becoming the strategic bottleneck for AI because power infrastructure is expanding more slowly than data-centre investment. SR027
CR041 Deloitte said AI data-centre buildouts face rising stress from grid, land, and construction-supply constraints. SR028
CR042 NVIDIA's DGX Cloud Lepton announcement lists Firmus among the cloud partners contributing GPUs to the marketplace. SR008, SR032
CR043 VAST says Firmus selected VAST AI Operating System as a foundational data layer for its AI factories. SR033
CR044 The Bell Bay FAQ says final energy-supply arrangements for that site were still being negotiated. SR035
CR045 ABC's July 2026 reporting says Firmus had not yet outlined a Tasmania-specific plan for funding new renewables beyond current negotiations. SR014, SR015
CR046 ABC's March 2026 report said Premier Rockliff would not detail the Firmus power deal because it was commercial-in-confidence. SR010
CR047 Firmus says it will initially match its power use with renewable energy certificates and contract suppliers to build new generation and storage. SR004, SR014
CR048 Firmus says it will self-fund new transmission infrastructure, invest in firming assets such as batteries, and pay market rates for electricity. SR014, SR035
CR049 Public sources reviewed for Tasmania do not name anchor tenants, contracted offtakers, or utilisation commitments for the Tasmanian sites. SR006, SR009, SR035
CR050 That missing offtake disclosure keeps project-level customer concentration and utilisation risk opaque. SR006, SR009, SR035
CR051 The Bell Bay FAQ says the site would support more than 100 full-time local roles once operational and run around the clock across three shifts. SR035
CR052 ABC reported management's rule of thumb of about half a full-time role per megawatt across sites, implying uneven job intensity relative to electricity draw. SR014
CR053 Public disclosures center on the co-CEOs, while no CFO or independent board detail appears in the reviewed sources. SR006, SR009, SR035
CR054 Digital.gov.au's December 2025 AI policy says government AI use now requires designated accountability, strategic adoption approaches, and use-case impact assessment. SR022
CR055 ABC's July 2026 coverage says the Greens want a moratorium and parliamentary oversight or reporting for large AI data centres until state-specific regulation exists. SR015
CR056 A power-thesis break would be visible through delayed connection approvals, unfinalised Bell Bay energy arrangements, or a failure to backfill demand with new generation. SR015, SR035, SR036
CR057 The highest-value diligence asks are final DA determinations, executed Hydro or TasNetworks documents, anchor-tenant disclosure, and direct litigation or cap-table records. SR006, SR035, SR036
CR058 A social-licence breakdown would be observable through extended consultation windows, calls for moratoria or parliamentary oversight, and persistent resident complaints on water, noise, or transparency. SR014, SR015, SR016
CV001 Firmus officially said it closed a A$330 million equity placement with Ellerston Capital as cornerstone investor and NVIDIA participating. SV001, SV009
CV002 Firmus said the round closed at a A$1.85 billion post-money valuation. SV001, SV009
CV003 Firmus said the raise funds Project Southgate, a 36,000-GPU flagship campus in northern Tasmania built over two stages. SV001, SV009
CV004 The Southgate project page describes a Launceston campus with 84 MW critical IT load, under-1.10 PUE, and 99% lower water use than traditional cooling. SV003
CV005 SmartCompany reported that Firmus reached a A$1.9 billion valuation and was planning a public listing in 2026. SV008
CV006 Firmus maintains a shareholder-communications page covering annual reports, meeting notices, and an investor-relations contact. SV002
CV007 Firmus announced a dedicated 360 MW NVIDIA DSX AI Factory campus in Batam running through 2034. SV004
CV008 Firmus said the Batam agreement covers up to 170,000 NVIDIA accelerators through 2027 and 2028. SV004
CV009 Firmus said the Batam structure uses revenue sharing and credit support with NVIDIA. SV004
CV010 Firmus expects between US$25 billion and US$30 billion from committed Batam offtake during the first six years of the partnership. SV004
CV011 Firmus signed a 12-year wholesale energy agreement for 600 MW of firm electricity with Gunvor. SV005
CV012 The South Australia agreement supports 1.2 GW of new renewables, 1.5 GWh of battery storage, and 2.7 GW of planned Firmus capacity. SV005
CV013 Firmus joined NVIDIA DGX Cloud Lepton with infrastructure in Singapore and Australia. SV006
CV014 The AI Singapore partnership says Firmus provides Singapore-based H200 access and up to 50% lower operating cost and energy use via immersion cooling. SV007
CV015 ABC described the Launceston AI factory as a A$2.1 billion project and quoted skepticism that long-run operating jobs will match construction hype. SV010
CV016 Blackstone agreed to acquire AirTrunk at an implied enterprise value of more than A$24 billion. SV011, SV012
CV017 At sale announcement, AirTrunk had more than 800 MW of customer-committed capacity and land supporting over 1 GW of future growth. SV011
CV018 AirTrunk later disclosed A$16 billion of ex-Japan refinancing and more than A$18 billion of total financing platform backed by over 60 banks and financiers. SV013
CV019 CoreWeave reported $60.7 billion of remaining performance obligations at December 31, 2025 with roughly five-year weighted average committed contract duration. SV014
CV020 CoreWeave reported 2025 revenue of $5.1 billion versus $1.9 billion in 2024. SV014
CV021 CoreWeave still reported a 2025 net loss of $1.2 billion. SV014
CV022 Equinix reported 280 data centers, 10,500-plus customers, 77 markets, and more than 507,000 interconnections in 2025. SV015, SV032
CV023 Equinix said 2025 revenue reached $9.2 billion, annualized gross bookings reached $1.6 billion, and adjusted EBITDA margin was 49%. SV032
CV024 Digital Realty markets sovereign and high-density AI infrastructure and publishes annual reports, quarterly results, and SEC filings. SV016, SV017, SV018, SV019
CV025 NEXTDC reported FY25 revenue of A$427.2 million and said a new A$2.9 billion syndicated debt agreement refinanced prior facilities. SV021
CV026 NEXTDC said its April 2026 updates were backed by a fully funded A$2.2 billion capital plan to scale AI-ready infrastructure. SV020
CV027 GDS reported 2025 revenue of US$1.6348 billion, 670,106 square meters committed or pre-committed, and a 47.3% adjusted EBITDA margin. SV022
CV028 Keppel DC REIT said assets under management were about $6.2 billion excluding a February 2026 acquisition and explicitly linked future growth to the AI wave. SV023
CV029 Gartner forecasts neocloud providers will capture 20% of a US$267 billion AI cloud market by 2030. SV030
CV030 ABI Research forecasts more than US$250 billion of neocloud GPUaaS revenue by 2030 and more than 2,200 neocloud-operated data centers by 2035. SV029
CV031 CBRE said APAC data-centre investment reached US$11.6 billion in 2025, average new builds now exceed 100 MW, and power availability is a major constraint. SV024
CV032 Colliers said 2025 global data-center investment exceeded US$580 billion and build costs rose 47% year over year. SV025
CV033 Colliers said 40% to 50% of total project cost now sits in power infrastructure and that early-stage funding increasingly comes from private credit. SV025
CV034 Ropes & Gray said power availability rather than capital is now the primary development constraint and that preferred equity, project finance, GPU financings, and forward sales are common. SV026
CV035 S&P Global estimated lenders committed US$121.91 billion of data-center credit in 2025 and highlighted facilities, ABS, CMBS, and industrial revenue bonds as active financing tools. SV027
CV036 JLL’s 2026 outlook described a roughly US$3 trillion data-center supercycle and warned that power scarcity and community acceptance now determine which projects can advance. SV028
CV037 Australian government expectations require AI-factory developers to support data sovereignty, bring new clean energy or storage, cover infrastructure costs, and invest in local skills. SV031
CV038 Large infrastructure valuations are most defensible once capacity, customer commitments, financing platforms, and repeat public reporting are visible at scale. SV011, SV013, SV021, SV022, SV023, SV032
CV039 Reviewed public Firmus materials do not disclose revenue, gross margin, utilization, customer concentration, or the preference and security terms of the A$330 million round. SV001, SV002, SV003, SV004, SV005, SV006, SV007, SV008, SV009
CV040 Firmus is likely to require additional external capital beyond the recent equity raise because it is simultaneously pursuing Southgate, Batam, and South Australian expansion. SV004, SV005, SV025, SV026, SV027
CV041 The Batam revenue-sharing and credit-support structure increases the risk that future economics are split across partners or senior capital layers rather than accruing cleanly to common equity. SV004, SV026
CV042 The strongest bull thesis is that sovereign AI demand, power scarcity, and Firmus’s energy-efficient design create a rare APAC platform that can lock in scarce capacity before incumbents localize supply. SV003, SV004, SV005, SV024, SV029, SV030, SV031
CV043 The strongest anti-thesis is that Firmus remains a capital-intensive project developer with strong narrative but insufficient disclosed monetization, and later capital can reprice common equity even if demand stays real. SV025, SV026, SV027, SV031, SV008
CV044 Equinix and Digital Realty already market AI-ready, sovereignty-aware infrastructure globally, so Firmus must win on APAC-specific energy execution rather than generic AI-colocation messaging. SV015, SV019, SV032
CV045 ABI warns that neoclouds risk margin pressure and irrelevance if they remain GPU brokers rather than winning enterprise demand and broader platform control. SV029
CV046 A reasonable base-case view is that the A$1.85 billion round is roughly fair if Southgate starts commercial delivery and capital markets stay open, but it is not obviously cheap on disclosed evidence. SV001, SV003, SV025, SV026, SV027, SV031
CV047 A reasonable bull-case fair-value range is A$2.4 billion to A$3.0 billion if Southgate lands on time, Batam offtake converts, and future capital remains non-punitive. SV004, SV005, SV029, SV030, SV031
CV048 A reasonable bear-case fair-value range is A$0.8 billion to A$1.2 billion if commercialization lags, power or permitting slip, or new capital arrives senior to common. SV025, SV026, SV027, SV031
CV049 Fresh entry should be staged and price-disciplined, with hard diligence rights on unit economics, offtake, and financing terms rather than blind acceptance of the unicorn headline. SV025, SV026, SV027, SV001
CV050 The public evidence supports exit aspiration more than exit readiness because shareholder communications and reported IPO intent exist, but audited operating disclosure still trails public-market norms. SV002, SV008, SV016, SV018, SV032
CV051 The main thesis-break triggers are Southgate delivery, disclosed commercialization metrics, cap-table terms, and whether future capital comes as plain equity or more senior structures. SV003, SV004, SV025, SV026, SV027
CV052 Mandatory diligence items are current ARR or revenue, utilization, signed offtake counterparties, customer concentration, project-level capex, and exact financing terms. SV001, SV004, SV005, SV025, SV026, SV027
CV053 Public Southgate materials use multiple scale frames, including 84 MW critical load, 36,000 GPUs over two stages, and a 45 MW first-stage framing in media coverage, so milestone definitions need normalization. SV001, SV003, SV008
CV054 Sovereignty is a real buyer-side driver because Gartner and Australian policy both emphasize jurisdictional control and data localization as enterprise decision factors. SV030, SV031, SV019
CV055 On current public evidence, the investment call is Track with medium confidence and a fair-to-stretched entry at the present mark. SV001, SV025, SV026, SV027, SV031
CV056 Downside transmission is nonlinear because 2026 sector financing increasingly rewards power certainty and pre-leased capacity first, so valuation can re-rate before revenue catches up. SV025, SV026, SV028
CV057 Ellerston Capital describes itself as an investment manager serving sovereign wealth, superannuation funds, international funds, family offices, and high-net-worth investors, strengthening the institutional-quality signal around Firmus’s 2025 round. SV033
来源
编号出版方标题引文
SO001 Firmus Creating the most efficient AI infrastructure - Firmus
SO002 Firmus About - Firmus
SO003 Firmus Firmus closes $330m raise - Firmus
SO004 Firmus Tasmanian world-first AI Factory Zone clears path for Firmus' Project Southgate - Firmus
SO005 Firmus AI Factories - Firmus
SO006 Firmus Southgate - Firmus
SO007 Firmus AI Cloud - Firmus
SO008 Firmus Investor Communications - Firmus
SO009 Firmus Careers - Firmus
SO010 Firmus AI Singapore and Firmus Technologies partner to advance sustainable AI infrastructure - Firmus
SO011 Firmus Firmus Technologies expands regional AI Access through NVIDIA DGX Cloud Lepton - Firmus
SO012 Firmus Our Commitments - Firmus
SO013 Firmus Firmus to Build 170,000 GPU AI Factory Campus with NVIDIA for Global AI-Natives - Firmus
SO014 HTX Protecting Singapore and the planet: HTX signs MoU with Firmus Technologies to boost sustainable computing
SO015 Maritime and Port Authority of Singapore MPA and Firmus Technologies to Study Seawater Cooling for Sustainable AI Infrastructure
SO016 ST Telemedia Global Data Centres ST Telemedia Global Data Centres and Firmus Technologies Forge Partnership to Build a Global Network of Sustainable AI Factories
SO017 Firmus AI Singapore x Firmus - Case Study | GPU AI Cloud Infrastructure - Firmus
SO018 ABC News Tasmania enters the 'AI race' with factory in north of state
SO019 ARN Firmus secures $330M to build green, sovereign AI factory with NVIDIA
SO020 SmartCompany Newly minted unicorn Firmus raises $330 million to build ‘AI factory’ in Tasmania
SO021 techpartner.news Firmus Technologies raises $330m for renewable-powered 'AI factory'
SO022 Australian Government Department of Industry, Science and Resources Expectations of data centres and AI infrastructure developers
SO023 Australian Government Department of Industry, Science and Resources New data centre expectations help bring the benefit of AI to all Australians
SO024 Data Center Dynamics AI cloud provider Firmus signs MoU with Singapore port authority for seawater-cooled AI compute
SO025 VAST Data Firmus Tech picks VAST AI OS for eco-friendly AI factories in Asia Pacific
SO026 Firmus Firmus | Preconstruction AI Design Review & Risk Analysis
SO027 MLCommons MLCommons MLPerf Training Benchmark
SM001 web.archive.org Energy supply for AI – Energy and AI – Analysis - IEA
SM002 web.archive.org Data centre electricity use surged in 2025, even with tightening bottlenecks driving a scramble for solutions - News - IEA
SM003 web.archive.org Energy demand from AI – Energy and AI – Analysis - IEA
SM004 web.archive.org Data centers: The race to power AI
SM005 web.archive.org The $7 trillion data center build-out: How industrials can capture their share
SM006 JLL Global data center demand surges despite supply and power constraints
SM007 JLL JLL 2026 Global Data Center Outlook
SM008 JLL Power progress in your global data center expansion
SM009 JLL Asia Pacific Data Centre Report Year-end 2025
SM010 CBRE 2026 Asia Pacific Data Centre Trends & Outlook
SM011 CBRE Global Data Center Trends 2026
SM012 CBRE Global Data Center Trends 2025
SM013 web.archive.org DCD>Market Review | APAC 2025
SM014 DatacenterDynamics Global data center electricity use to double by 2026 - IEA report
SM015 Data Center Knowledge Neocloud Services Surge as AI Strains Data Centers
SM016 Infocomm Media Development Authority Green DC Roadmap | IMDA
SM017 Infocomm Media Development Authority Architects of SG Digital Future
SM018 Australian Government Department of Industry, Science and Resources Expectations of data centres and AI infrastructure developers
SM019 Australian Government Department of Industry, Science and Resources New data centre expectations help bring the benefit of AI to all Australians
SM020 Firmus Technologies Southgate - Firmus
SM021 Firmus Technologies Our Commitments - Firmus
SM022 Gartner Gartner Predicts Neocloud Providers Will Capture 20% of the $267 Billion AI Cloud Market by 2030
SM023 Cisco / theCUBE Research FINAL-theCUBE-Research-White-Paper-Neoclouds-and-Sovereign-Clouds-Feb-2026
SM024 ABI Research The State of Neocloud: Four Trends for 2026
SM025 DatacenterDynamics APAC 2025 development pipeline hits new record of 19.4GW
SM026 DatacenterDynamics More than $100bn needed for APAC colo data center pipeline - report
SM027 Gartner Gartner Says Worldwide Sovereign Cloud IaaS Spending Will Total $80 Billion in 2026
SM028 Cisco Neocloud Providers Are Making Waves—and Cisco Is Helping Them Do It
SP001 Amazon Web Services Amazon EC2 P5 Instances
SP002 Amazon Web Services Global Infrastructure
SP003 Amazon Web Services Data Centers
SP004 Google Cloud GPU machines in accelerator-optimized machine family | Compute Engine | Google Cloud Documentation
SP005 Google Cloud Global Locations - Regions & Zones | Google Cloud
SP006 Google Cloud VM instance pricing
SP007 Microsoft Azure Choose the Right Azure Region for You | Microsoft Azure
SP008 Microsoft Learn ND family virtual machine size series - Azure Virtual Machines
SP009 Microsoft Inside the world’s most powerful AI datacenter - The Official Microsoft Blog
SP010 Equinix EQIX- 2025 10-K
SP011 Equinix Data Centers
SP012 Equinix Scalable Data Center Infrastructure | Global Data Centers | Equinix
SP013 Digital Realty AI & ML Infrastructure Solutions for Growth | Digital Realty
SP014 Digital Realty PlatformDIGITAL | Digital Realty
SP015 AirTrunk Our Investors | AirTrunk
SP016 AirTrunk Completion Of AirTrunk Acquisition By Blackstone - Marking New Era Of Growth | AirTrunk
SP017 NEXTDC NEXTDC Limited FY25 Annual Report
SP018 NEXTDC M3 Melbourne
SP019 NEXTDC S4 Sydney NEXTDC Data Centre
SP020 Keppel DC REIT KDC REIT Annual Report 2025
SP021 GDS Holdings HOME - GDS
SP022 GDS Holdings GDS Holdings Limited Reports Fourth Quarter and Full Year 2025 Results | GDS Holdings Ltd
SP023 CoreWeave The Essential Cloud for AI | CoreWeave
SP024 United States Securities and Exchange Commission crwv-20251231
SP025 Lambda AI infrastructure | Lambda
SP026 Lambda AI Cloud Pricing | GPU Compute & AI Infrastructure | Lambda
SP027 Crusoe Crusoe AI Data Centers | Energy-First Acceleration
SP028 Crusoe Crusoe's 2025 Impact Report: Building sustainable intelligence
SP029 NVIDIA NVIDIA Data Centers for the Era of AI Reasoning
SI001 Firmus AI Cloud Compute - Firmus
SI002 Firmus Bare Metal - Firmus
SI003 Firmus Cloud Services - Firmus
SI004 Firmus Firmus secures 600 MW energy supply agreement in South Australia, linked to 1.2 GW of new renewable generation and battery storage Firmus today announced a landmark 12-year wholesale energy supply agreement with Gunvor Group for 600 MW of firm electricity to support the next phase of Project Southgate.
SI005 Firmus Firmus closes $330m raise - Firmus The raise closed at a post-money valuation of AUD $1.85 billion.
SI006 Firmus Firmus Technologies expands regional AI Access through NVIDIA DGX Cloud Lepton - Firmus Firmus Technologies joins NVIDIA's expanded DGX Cloud Lepton marketplace as a Cloud Partner, contributing its Singapore and Australia-based infrastructure to the unified platform.
SI007 Firmus AI Singapore x Firmus - Case Study | GPU AI Cloud Infrastructure - Firmus 32 nodes / 256 NVIDIA H200 GPUs deployed.
SI008 Firmus Southgate - Firmus Capacity 84MW Critical IT Load.
SI009 Firmus Our Commitments - Firmus We fund the transmission and network infrastructure needed to connect our AI Factories to the energy grid.
SI010 Australian Government Department of Industry, Science and Resources Expectations of data centres and AI infrastructure developers New data centres and AI infrastructure should not place upward pressure on energy prices and should make a positive contribution to Australia’s energy transition.
SI011 ABC News Tasmania enters the 'AI race' with factory in north of state "But there's a big problem here — there isn't enough power."
SI012 Data Center Dynamics AI cloud provider Firmus signs MoU with Singapore port authority for seawater-cooled AI compute The SMC site says AI factories in India and Thailand are “coming soon.”
SI013 NVIDIA NVIDIA DGX Cloud Lepton Developers can purchase GPU capacity directly from participating cloud providers through the marketplace or bring their own compute clusters.
SI014 NVIDIA NVIDIA Announces DGX Cloud Lepton to Connect Developers to NVIDIA’s Global Compute Ecosystem NVIDIA Cloud Partners including CoreWeave, Crusoe, Firmus, Foxconn, GMI Cloud, Lambda, Nebius, Nscale, Softbank Corp. and Yotta Data Services will offer NVIDIA Blackwell and other NVIDIA architecture GPUs on the DGX Cloud Lepton marketplace.
SI015 U.S. Securities and Exchange Commission / CoreWeave CoreWeave, Inc. Form S-1 We recognized an aggregate of approximately 77% of our revenue from our top two customers for the year ended December 31, 2024.
SI016 CoreWeave CoreWeave - Financials - Quarterly Results
SI017 Equinix, Inc. EQIX 2025 Annual Report PDF We grew revenue to $9.2 billion. We achieved record annualized gross bookings of $1.6 billion, up 27 percent year over year.
SI018 Equinix, Inc. 10-K - 02/11/2026 - Equinix, Inc. PDF The largest components of our cost of revenues are depreciation, rental payments related to our leased IBX data centers, utility costs including electricity...
SI019 Equinix, Inc. Annual Reports
SI020 Equinix, Inc. All SEC Filings
SI021 Digital Realty Trust Annual Reports | Digital Realty Trust
SI022 NEXTDC Record Contracted Growth and A$2.2bn Capital Plan to Scale AI-Ready Infrastructure Record contracted utilisation growth and a fully funded A$2.2bn capital plan position NEXTDC to accelerate delivery of next-generation AI-ready infrastructure at scale.
SI023 NEXTDC Financial Reports
SI024 AirTrunk AirTrunk closes A$16 billion (ex Japan) sustainable financing to accelerate APJ growth and impact | AirTrunk AirTrunk has closed a A$16 billion (ex Japan) refinancing, the region’s largest-ever sustainability linked financing.
SI025 AirTrunk AirTrunk doubles down in Malaysia with two new hyperscale campuses in Johor Bahru | AirTrunk The existing JHB1 and JHB2 campuses (totalling more than 420MW of IT load) are almost 100% contracted and tracking well ahead of investment plans.
SI026 AirTrunk Report | AirTrunk
SI027 ARN Firmus secures $330M to build green, sovereign AI factory with NVIDIA The raise closed at a post-money valuation of $1.85 billion.
SI028 SmartCompany Newly minted unicorn Firmus raises $330 million to build ‘AI factory’ in Tasmania The company is expected to continue raising capital ahead of a slated ASX listing next year.
SE001 Firmus AI Cloud Compute - Firmus
SE002 Firmus AI Storage - Firmus
SE003 Firmus Bare Metal - Firmus
SE004 Firmus Cloud Services - Firmus
SE005 Firmus Cloud Applications - Firmus
SE006 Firmus MLPerf - Firmus
SE007 Firmus Engineering Principles - Firmus
SE008 Firmus Privacy Policy - Firmus
SE009 Firmus Firmus Technologies expands regional AI Access through NVIDIA DGX Cloud Lepton - Firmus
SE010 Firmus MPA and Firmus sign MoU to advance sustainable AI Infrastructure using Seawater Cooling - Firmus
SE011 Firmus Firmus HTX MOU partnership - Firmus
SE012 Firmus Firmus to Build 170,000 GPU AI Factory Campus with NVIDIA for Global AI-Natives - Firmus
SE013 NVIDIA NVIDIA Spectrum-X Ethernet Platform for AI Networking
SE014 NVIDIA Scaling Power-Efficient AI Factories with NVIDIA Spectrum-X Ethernet Photonics
SE015 NVIDIA H200 GPU | NVIDIA
SE016 NVIDIA Accelerated InfiniBand Solutions for HPC | NVIDIA
SE017 NVIDIA Connect Developers to Global GPU Compute | NVIDIA DGX Cloud Lepton
SE018 VAST Data Firmus Tech picks VAST AI OS for eco-friendly AI factories in Asia Pacific - VAST Data
SE019 VAST Data The Operating System for Artificial Intelligence - VAST Data
SE020 MLCommons MLCommons MLPerf Training Benchmark
SE021 GitHub GitHub - mlcommons/training: Reference implementations of MLPerf training benchmarks
SE022 GitHub GitHub - SchedMD/slurm: Slurm: A Highly Scalable Workload Manager
SE023 Maritime and Port Authority of Singapore MPA and Firmus Technologies to Study Seawater Cooling for Sustainable AI Infrastructure
SE024 HTX Protecting Singapore and the planet: HTX signs MoU with Firmus Technologies to boost sustainable computing
SE025 Intelligent CIO APAC Firmus Technologies Group selects VAST AI Operating System to power sovereign, energy-efficient AI factories in APAC
SE026 Greenhouse Job Application for Head of Corporate IT & Cyber Security at Firmus Technologies
SE027 Tech Wire Asia Nvidia-backed Firmus plans 170,000-GPU Batam AI data centre
SE028 U.S. News & World Report Australia's Firmus Technologies Strikes AI Access Deal With Nvidia
SU001 Firmus AI Cloud - Firmus
SU002 Firmus AI Singapore and Firmus Technologies partner to advance sustainable AI infrastructure - Firmus
SU003 Firmus AI Singapore x Firmus - Case Study | GPU AI Cloud Infrastructure - Firmus
SU004 AI Singapore Home - AI Singapore
SU005 AI Singapore SEA-LION | Empowering Open Multilingual AI for Southeast Asia
SU006 AI Singapore SEA-LION | SEA-LION Documentation
SU007 GitHub GitHub - aisingapore/sealion: South-East Asia Large Language Models
SU008 HTX Protecting Singapore and the planet: HTX signs MoU with Firmus Technologies to boost sustainable computing
SU009 HTX Because we're stronger together
SU010 Firmus Firmus HTX MOU partnership - Firmus
SU011 Maritime and Port Authority of Singapore MPA and Firmus Technologies to Study Seawater Cooling for Sustainable AI Infrastructure
SU012 Firmus MPA and Firmus sign MoU to advance sustainable AI Infrastructure using Seawater Cooling - Firmus
SU013 Data Center Dynamics AI cloud provider Firmus signs MoU with Singapore port authority for seawater-cooled AI compute
SU014 OpenGov Asia Singapore: Seawater Cooling for Sustainable AI Infrastructure - OpenGov Asia
SU015 Firmus Firmus Technologies expands regional AI Access through NVIDIA DGX Cloud Lepton - Firmus
SU016 NVIDIA NVIDIA DGX Cloud Lepton
SU017 NVIDIA NVIDIA Announces DGX Cloud Lepton to Connect Developers to NVIDIA’s Global Compute Ecosystem
SU018 NVIDIA Introducing NVIDIA DGX Cloud Lepton: A Unified AI Platform Built for Developers | NVIDIA Technical Blog
SU019 Smart Nation Singapore National AI Strategy
SU020 Ministry of Digital Development and Information Update to Singapore's National AI Strategy: Refreshed Priorities to Harness AI for the Public Good (Factsheet)
SU021 Department of Industry, Science and Resources Expectations of data centres and AI infrastructure developers
SU022 Gartner Gartner Says Worldwide Sovereign Cloud IaaS Spending Will Total $80 Billion in 2026
SU023 ST Telemedia Global Data Centres ST Telemedia Global Data Centres and Firmus Technologies Forge Partnership to Build a Global Network of Sustainable AI Factories
SU024 VAST Data Firmus Tech picks VAST AI OS for eco-friendly AI factories in Asia Pacific
SU025 ABC News Australia Tasmania enters the 'AI race' with factory in north of state
SU026 U.S. News & World Report / Reuters Australia's Firmus Technologies Strikes AI Access Deal With Nvidia
SU027 Firmus Firmus to Build 170,000 GPU AI Factory Campus with NVIDIA for Global AI-Natives - Firmus
SU028 Tech Wire Asia Nvidia-backed Firmus plans 170,000-GPU Batam AI data centre
SU029 Ministry of Digital Development and Information Singapore and Australia Expand Cooperation on AI with New Memorandum of Understanding
SR001 Department of Industry, Science and Resources Expectations of data centres and AI infrastructure developers
SR002 Department of Industry, Science and Resources New data centre expectations help bring the benefit of AI to all Australians
SR003 Department of Industry, Science and Resources The legal landscape for AI in Australia
SR004 Firmus Our Commitments - Firmus
SR005 Firmus Firmus secures 600 MW energy supply agreement in South Australia, linked to 1.2 GW of new renewable generation and battery storage - Firmus
SR006 Firmus Tasmanian world-first AI Factory Zone clears path for Firmus' Project Southgate - Firmus
SR007 Firmus Privacy Policy - Firmus
SR008 Firmus Firmus Technologies expands regional AI Access through NVIDIA DGX Cloud Lepton - Firmus
SR009 ABC News Tasmania enters the 'AI race' with factory in north of state
SR010 ABC News Questions raised over deal between AI company and state power generator
SR011 ABC News AI company set to become Tasmania's largest power user
SR012 ABC News AI factory 'brings into doubt' Marinus Link future
SR013 ABC News What we know about water use of the over 250 data centres in Australia
SR014 ABC News 'Less water than one restaurant': AI data centre company responds to concerns
SR015 ABC News Is Tasmania ready for the AI data centre boom?
SR016 ABC Listen 'Take us seriously': Why this community has concerns about a proposed 'AI factory' - ABC listen
SR017 Maritime and Port Authority of Singapore MPA and Firmus Technologies to Study Seawater Cooling for Sustainable AI Infrastructure
SR018 Infocomm Media Development Authority Singapore launches new tools to help businesses protect data and deploy AI in a trusted ecosystem | IMDA
SR019 Cyber and Infrastructure Security Centre Cyber and Infrastructure Security Centre Website
SR020 Australian Cyber Security Centre AI data security | Cyber.gov.au
SR021 Digital Transformation Agency AI Policy Update: Strengthening responsible use across government
SR022 Digital.gov.au Policy for the responsible use of AI in government - Version 2.0
SR023 Office of the Australian Information Commissioner Guidance on privacy and the use of commercially available AI products
SR024 Australian Energy Market Operator AEMO’s updated forecasting methodology targets rapidly growing electricity loads, following industry consultation
SR025 International Energy Agency Data centre electricity use surged in 2025, even with tightening bottlenecks driving a scramble for solutions - News - IEA
SR026 JLL Global data center demand surges despite supply and power constraints
SR027 World Economic Forum Is power grid connectivity the strategic bottleneck for AI?
SR028 Deloitte Can US infrastructure keep up with the AI economy?
SR029 MSCI When AI Meets Water Scarcity: Data Centers in a Thirsty World | MSCI
SR030 Data Center Frontier JLL's 2026 Global Data Center Outlook: Navigating the AI Supercycle, Power Scarcity and Structural Market Transformation
SR031 Utility Magazine Aurora Energy signs $5B deal for 104MW AI project - Utility Magazine
SR032 NVIDIA NVIDIA Announces DGX Cloud Lepton to Connect Developers to NVIDIA’s Global Compute Ecosystem
SR033 VAST Data Firmus Tech picks VAST AI OS for eco-friendly AI factories in Asia Pacific
SR034 HTX Protecting Singapore and the planet: HTX signs MoU with Firmus Technologies to boost sustainable computing
SR035 Firmus Project Southgate Bell Bay: FAQs - Firmus
SR036 Tasmanian Department of Treasury and Finance Firmus Technologies Pty Ltd connection enquiry - St Leonards expansion (RTI release)
SV001 Firmus Technologies Firmus closes $330m raise - Firmus
SV002 Firmus Technologies Investor Communications - Firmus
SV003 Firmus Technologies Southgate - Firmus
SV004 Firmus Technologies Firmus to Build 170,000 GPU AI Factory Campus with NVIDIA for Global AI-Natives - Firmus
SV005 Firmus Technologies Firmus secures 600 MW energy supply agreement in South Australia - Firmus
SV006 Firmus Technologies Firmus Technologies expands regional AI Access through NVIDIA DGX Cloud Lepton - Firmus
SV007 Firmus Technologies AI Singapore and Firmus Technologies partner to advance sustainable AI infrastructure - Firmus
SV008 SmartCompany Newly minted unicorn Firmus raises $330 million to build ‘AI factory’ in Tasmania
SV009 ARNnet Firmus secures $330M to build green, sovereign AI factory with NVIDIA
SV010 ABC News Australia Tasmania enters the 'AI race' with factory in north of state
SV011 Blackstone Blackstone Announces Agreement to Acquire AirTrunk in a A$24B Transaction
SV012 AirTrunk Our Investors | AirTrunk
SV013 AirTrunk AirTrunk closes A$16 billion (ex Japan) sustainable financing to accelerate APJ growth and impact
SV014 Securities and Exchange Commission CoreWeave, Inc. Form 10-K for fiscal year ended December 31, 2025
SV015 Equinix Data Centers
SV016 Digital Realty Trust Annual Reports | Digital Realty Trust
SV017 Digital Realty Trust Quarterly Results | Digital Realty Trust
SV018 Digital Realty Trust SEC Filings | Digital Realty Trust
SV019 Digital Realty AI & ML Infrastructure Solutions for Growth | Digital Realty
SV020 NEXTDC Record Contracted Growth and A$2.2bn Capital Plan to Scale AI-Ready Infrastructure
SV021 NEXTDC Limited NEXTDC FY25 Annual Report
SV022 GDS Holdings Ltd GDS Holdings Limited Reports Fourth Quarter and Full Year 2025 Results
SV023 Keppel DC REIT Keppel DC REIT Annual Report 2025
SV024 CBRE 2026 Asia Pacific Data Centre Trends & Outlook
SV025 Colliers 2026 Data Center Marketplace Report
SV026 Ropes & Gray LLP Data Center Investment in 2026: AI Demand, Power Constraints, and Private Equity Trends
SV027 S&P Global Market Intelligence Banks meeting data center demand with billions in credit facilities, bonds
SV028 Data Center Frontier JLL's 2026 Global Data Center Outlook: Navigating the AI Supercycle, Power Scarcity and Structural Market Transformation
SV029 ABI Research The State of Neocloud: Four Trends for 2026
SV030 Gartner Gartner Predicts Neocloud Providers Will Capture 20% of the $267 Billion AI Cloud Market by 2030
SV031 Australian Government Department of Industry, Science and Resources Expectations of data centres and AI infrastructure developers
SV032 Equinix, Inc. Equinix 2025 Annual Report (10-K)
SV033 Ellerston Capital About | Ellerston Capital