初创公司尽调
尽调报告 AI infrastructure / sovereign GPU cloud late-stage private / pre-IPO 2026-06-18

Firmus

澳大利亚创立的主权 AI 基础设施建设商,正接近 ASX IPO

Firmus 的主权 AI 基础设施逻辑可信,融资势头也异常强;但预期 IPO 估值仍跑在公开披露包前面。

封面要素

最近一轮私募融资 01
505 USDm [CO028, CV001]
累计股权融资 03
1350 USDm [CO029]
Project Southgate 目标容量 05
1600 MW [CO030]

公司概况

Firmus 是一家创立于澳大利亚、总部在新加坡的 AI 基础设施公司,早期做比特币挖矿冷却系统,后来转向模块化液冷 AI 工厂和 GPU 云平台。公开证据支持其新加坡部署已上线、Project Southgate 在澳大利亚推进大规模建设,并通过 Nvidia 参投股权轮和 Blackstone 牵头债务额度获得异常强的资本通道;但收入质量、利润率和客户稳定性还不足以支撑完整承销判断。

官网
firmus.co
成立时间
2019-12-12
创始人
Oliver Curtis, Tim Rosenfield, Jonathan Levee
创立地点
Australia
总部
Singapore
产品
Firmus 建设并运营液冷 AI 工厂、接入电网的高密度算力园区,以及面向训练和推理负载的 Firmus AI Cloud GPU 容量。
客户
需要 APAC 主权 AI 算力的超大规模云厂商、政府机构、企业、教育机构和初创公司。
商业模式
资本密集型基础设施的持有与运营,叠加 GPU 云和签约 AI 算力服务。
阶段
late-stage private / pre-IPO
融资情况
2025 年 9 月至 2026 年 4 月,公司三轮股权融资约 USD 1.35 billion,又获得 Blackstone 牵头的 USD 10 billion 债务额度,并计划在 2026 年中登陆 ASX。
[CO001, CO003, CO017, CO022, CO027, CO028, CO029, CO030]

执行摘要

主要优势

  • 在澳大利亚和新加坡布局主权 AI 基础设施,契合当地真实的算力短缺和政策需求。
  • 作为澳大利亚私营初创公司,Firmus 资本获取能力异常强:Nvidia 多次参与融资,Blackstone 还提供了有条件的 USD 10B 融资安排。
  • 公司在新加坡已有上线部署,也有大型 Project Southgate 管线,不只是概念性的 AI 工厂故事。

主要风险

  • 没有招股书、经审计财报或公开签约收入排期来支撑预期 IPO 估值。
  • Oliver Curtis 的内幕交易定罪和无锁定 IPO 结构都会放大公开市场审视,创始人治理风险仍偏高。
  • 执行取决于超大规模云厂商合同质量、Nvidia 供货、建设成本纪律和债务融资扩张经济性,而这些只披露了一部分。

未决问题

  • 招股书前,经审计财报、收入确认、毛利率、烧钱速度和签约收入排期仍未披露。
  • 两个超大规模云厂商锚定客户的身份、定价、最低承诺和取消条款仍未披露。
  • PUE、每兆瓦建设成本或 GPU 利用率主张尚无独立第三方验证。

目录

Chapter 01

01公司概览

1.1 身份、产品与起源

Firmus Technologies 设计、建设并运营其所谓 AI Factories:一种模块化、液冷、高密度算力平台,专为 AI 训练和推理负载打造。公司运营总部在新加坡,主要注册实体在澳大利亚:Firmus Grid Pty Ltd(2019 年 12 月注册)和 Firmus Technologies Pty Ltd(2021 年 11 月注册)。tracxn 文件中,这家初创公司的当前注册地址位于新南威尔士州 Saint Leonards;另一篇文章称其「创立于悉尼」,早期报道则把创始运营地点指向塔斯马尼亚——本章将这一冲突保留为未解决的地理细节。 公司由 Oliver Curtis、Tim Rosenfield 和 Jonathan Levee 于 2019 年共同创立。三位创始人此前都没有数据中心或工程背景:Rosenfield 曾任内衣公司 Simone Perele CEO;Levee 在资源行业有中层管理经历;Curtis 出狱后从投资银行家转为加密货币投资者。最初业务是为比特币挖矿设备开发液冷技术,借助塔斯马尼亚凉爽气候和水电资源降低运营成本。Nick Curtis(Oliver 的父亲、前矿业高管)担任首任董事会主席。比特币价格下跌后,Telstra 前高管 Ted Pretty 接替 Nick Curtis 出任主席,并帮助公司转向 AI 数据中心应用。Ted Pretty 随后离任,2025 年末公司任命 Grant Dempsey 为候任董事会主席。 Firmus 将其 AI Factory 描述为「一种模块化、垂直化系统,在技术栈每一层优化成本和电力」,核心包括 Firmus Hypercubes(多 petascale 模块化算力抽象)、Synert(与 Eaton 在塔斯马尼亚部署的电网集成软件)和直达芯片液冷。公司还运营 Firmus AI Cloud——一个 GPU 驱动的 neocloud(H200、L40S 和 GB300 Blackwell 实例)——因此商业模式同时包括资本密集型基础设施持有和云服务收入。其旗舰外部项目是 Project Southgate:与 CDC Data Centres 和 Nvidia 合作,在澳大利亚建设全国 AI 工厂网络,目标到 2028 年提供 1.6 GW AI 算力。 [CO001, CO003, CO004, CO005, CO006, CO007]

Firmus 快照 KPI(截至 2026-06-18)
指标数值 / 状态日期 / 时点置信度缺口 / 注意事项
股权融资(累计,六个月冲刺)~US$1.35 billionApr 20262024 年前轮次未公开披露
投后估值(最新股权轮)US$5.5 billion(A$7.95 billion,估值)Apr 2026与 2025 年 11 月 A$6B 数字冲突;AUD/USD 汇率解释了大部分差异
债务工具US$10 billion(Blackstone 牵头)Feb 2026已承诺,但为 Project Southgate 建设按时间提款
IPO 目标估值区间A$8–12 billion(上调 40–100%)Apr 2026媒体引用投资者通信;Firmus 官方未确认
IPO 募资目标在 ASX 募资 A$2–3 billionApr 2026报道不一;Startup Daily / Wilson AM 报道称 A$3B 为上限
收入 / ARR未公开披露2026私营公司;新加坡业务正在产生收入(金额未知)
员工数(当前员工)未公开披露2026公司目标到 2028 年在 Project Southgate 创造 20,000+ 个建设 / 运营岗位;当前 FTE 未知
GPU 数量(新加坡已上线)两个 STT GDC 站点约 4,000 块 Nvidia GPU2024来自 NAOS/Livewire Markets 投资者笔记;公司新闻稿未确认
Project Southgate 规划容量到 2028 年,澳大利亚五个站点合计 1.6 GW2025-11取决于审批和融资提款;相对澳大利亚当前 1.35 GW 总量较为激进
电力使用效率(PUE)主张1.03(点测量)2021自报;已确认并非连续验证;专家共识认为,可比设施的一流 PUE 为 1.1–1.2

收入、员工数、客户数和 ARR 均未公开披露(私营公司)。估值数字为官方新闻稿或引用公司指引的媒体所述投后估值。PUE 数字据公司发言人为 2021 年一次点测量。

[CO023, CO026, CO027, CO028, CO029, CO036]
FO002: Firmus 业务架构与价值链

Firmus 的身份(两名联席 CEO、三个创始实体)、产品平台(AI Factory + AI Cloud + Synert)、资本结构(Nvidia 双重角色、Blackstone 债务、Coatue 股权)和供应链(CDC、Maas/Benmax)如何连接起来,为主权、企业和超大规模客户支撑 AI 基础设施价值主张。

[CO005, CO006, CO016, CO027, CO028, CO031]

1.2 领导层、治理与关键人风险

Firmus 采用联席 CEO 架构:Oliver Curtis 与 Tim Rosenfield 共同担任最高管理角色。Curtis 和 Rosenfield 是表亲;第三位联合创始人 Jonathan Levee 现任研发总监,媒体报道曾称其为 Curtis 的前内兄弟。创始层高度集中的家族与私人关系网络,带来关键人依赖,尽调必须专门处理。 Curtis 2016 年内幕交易定罪,是 Firmus 档案中最重大的治理风险。2007 年 5 月至 2008 年 6 月,Curtis 与前好友 John Hartman 合谋,利用 Hartman 对雇主 Orion Asset Management 交易意图的内幕了解,完成 45 笔盈利 CFD 交易,净利润 A$1,432,228.85。2016 年 6 月,最高法院陪审团裁定 Curtis 有罪,判处两年监禁;他在 Cooma Correctional Centre 服刑十二个月,并于 2017 年 6 月 23 日获释。2016 年 12 月 16 日,他的定罪上诉被驳回。Curtis 后来公开称该罪行为年轻时犯下的「一个错误」。 2026 年 2 月 23 日,ASX Listings Compliance 部门发布一份专门说明其「良好声誉与品格」测试的合规更新;市场观察人士普遍认为,这与即将到来的 Firmus IPO 有关。该框架会权衡过往行为的不诚实程度、时间流逝、悔意和后续监管接纳等因素。随后有报道称 Curtis 已获准以公开上市公司身份运营 Firmus;2026 年 5 月 27 日发布的 nine.com.au AFR Rich List 人物报道确认了这一批准。2026 年 5 月,Curtis 首次登上 AFR 的 The List – Australia's Richest 250,估计个人财富 A$1.25 billion。 2025–2026 年,董事会构成明显加强。Grant Dempsey 被任命为候任主席,Kirsty Godfrey-Billy 于 2025 年 9 月末出任 CFO。David Leslie(Ellerston Capital 投资总监)在 2025 年 9 月股权融资后加入董事会。2026 年 3 月起,三名新的非执行董事生效:Lee Hatton(审计与风险委员会主席;Xero 前董事会成员、Block Inc 前高管);Christine Bartlett(薪酬与提名委员会主席;现任 NSW Ports 和 CEDA 主席,曾任 IBM 和 NAB 高管);Julie Shuttleworth AM(Fortescue Future Industries 前副 CEO 和 CEO)。Bloomberg 与 Firmus 自家新闻室均在 2026 年 3 月 23 日确认这些任命。 独立机构分析师——包括 Roger Montgomery Investment Management,以及通过 Livewire Markets 发布的 NAOS Asset Management 投资者笔记——都明确指出,Curtis 的定罪与其担任上市公司董事或高管直接相关,并质疑澳大利亚监管框架是否足以校准如此规模的 IPO。 [CO005, CO006, CO007, CO008, CO009, CO011]

领导层与创始人表
人物职务背景 / 过往经验创始人—市场匹配关键人依赖
Oliver Curtis联合创始人兼联席 CEO前投行人士;2017 年联合创办 e-Nome(加密医疗初创公司);2019 年最初向 Firmus 投入 A$250K;2016 年因内幕交易被定罪,服刑 12 个月没有数据中心或工程过往经验;画像偏业务拓展和融资;在悉尼金融圈拥有投资者网络高——公司公众面孔;定罪记录是 IPO 的首要治理风险;ASX Good Fame and Character 于 2026 年 3 月通过
Tim Rosenfield联合创始人兼联席 CEOSimone Perele 内衣公司前 CEO;Oliver Curtis 的表兄弟没有 AI/DC 工程背景;具备运营管理经验;与 Curtis 共同承担高管责任以分散关键人风险高——在册联席 CEO;与 Curtis 存在治理和运营共同依赖
Jonathan Levee联合创始人兼研发总监资源行业中层管理经验;被描述为 Oliver Curtis 前姻亲未披露工程资质;研发总监角色暗示其拥有内部产品所有权;资源行业背景可能帮助工业级项目管理中——研发所有权形成技术依赖;没有独立于 Firmus 的公开形象
Grant Dempsey主席(新任,2025 年 9 月任命)公开来源未充分披露IPO 前的董事会监督与治理角色中——新任命;背景尚未进入公开领域
Kirsty Godfrey-BillyCFO(2025 年 9 月任命)公开来源未充分披露财务领导力对 IPO 进程至关重要中——新任命;背景尚未进入公开领域
David Leslie非执行董事(自 2025 年 9 月);Ellerston Capital 投资总监Ellerston Capital(Ashok Jacob 基金)投资总监;2025 年 9 月轮的基石投资者投资者阵营 NED;金融服务监督低——有机构支持的独立 NED
Lee Hatton非执行董事兼审计与风险委员会主席(自 2026 年 3 月)20+ 年金融服务经验;前 Xero 董事会成员(2014–2023);曾在 NAB、Suncorp、Afterpay、Block Inc(Global Financial Operations)任高管治理与风险监督;科技公司董事会经验低——独立 NED
Christine Bartlett非执行董事兼薪酬与提名委员会主席(自 2026 年 3 月)曾任 IBM、NAB、Jones Lang LaSalle CEO / 高管;现任 NSW Ports 和 CEDA 主席战略与复杂项目管理;上市和私营公司董事会经验低——独立 NED
Julie Shuttleworth AM非执行董事(自 2026 年 3 月)30+ 年矿业、资源和能源经验;前 Fortescue 副 CEO;Fortescue Future Industries CEO大型项目交付和能源转型专长,与 Project Southgate 建设直接相关低——独立 NED

枚举仅覆盖公开来源中具名的领导层;完整高管团队(如首席人事官 Nicole Reid、A/NZ 副总裁 Anthony McGough)并不穷尽。2025 年前董事会成员(Nick Curtis、Ted Pretty)已离任。Grant Dempsey 和 Kirsty Godfrey-Billy 的背景尚未进入公开领域。

[CO005, CO006, CO007, CO008, CO009, CO010]

1.3 融资历史与资本结构

Firmus 2024 年以前的股权历史没有公开披露。可识别的最早外部投资约为 Regal Funds Management(Phil King)的 A$10 million,据理解为约 2021 年投入的可转债,当时公司仍主要经营比特币挖矿冷却业务。2023 年 6 月与 ST Telemedia Global Data Centres(STT GDC——由新加坡主权财富基金 Temasek Holdings 持有)的合作,为 Firmus 的新加坡 AI Factory 部署奠定运营基础;安排中,Firmus 持有并运营两个 STT 数据中心站点内的 IT 网络和算力,约 4,000 块 Nvidia GPU,而 STT 保留实体设施所有权。The Witness 将 2023 年一笔 US$100 million 战略投资归因于 STT GDC;该数字没有在一手来源中独立确认,标记为未验证。 2025 年 9 月至 2026 年 4 月的三轮股权融资和一笔债务额度,把 Firmus 推成澳大利亚历史上资本最充足的私营科技公司之一: — 2025 年 9 月 16 日(SVIIC 口径为 Series B / Caplight 口径为 Private Placement):A$330 million,投后估值 A$1.85 billion。基石投资来自 Ellerston Capital(投资总监 David Leslie 加入董事会);Nvidia 战略参投。既有投资者 Regal Funds Management、Archibald Capital 和 Tectonic Investment Management 参与。Alex Waislitz 与 Pratt 家族(Heloise Pratt/ Visy)也是股东。据理解,Ellerston Capital 持股接近 7%。 — 2025 年 11 月 14 日:A$500 million,投后估值约 A$6 billion——八周内接近三倍抬升。Morgans 担任独家牵头经办人;Highbury Partnership 担任财务顾问。包括 UniSuper 和 Wilson Asset Management 在内的新增机构投资者,与既有支持方共同参与。 — 2026 年 2 月:Maas Group Holdings(ASX: MGH)按 A$6 billion 估值投资 A$100 million,取得约 1.7% 股权。Maas 子公司 JLE 将成为 Firmus AI Factory 推广的主权供应链一部分。 — 2026 年 2 月 9 日(债务):US$10 billion 债务融资额度,由 Blackstone Tactical Opportunities 和 Blackstone Credit & Insurance 牵头,Coatue 支持。Firmus 称其为澳大利亚史上规模最大的私营债务融资之一;结构为长期基础设施债。 — 2026 年 4 月 6 日(Strategic / Series F):US$505 million,投后估值 US$5.5 billion,由 Coatue Management 领投(普通合伙人 Robert Yin 被引用),Nvidia 继续参与。公告时仍受若干交割条件约束。截至 2026 年 4 月,这六个月内累计股权融资约 US$1.35 billion。 冲突说明:Firmus 公布 2025 年 11 月融资时采用 A$6 billion 估值(澳元),而 2026 年 4 月公告引用的是 US$5.5 billion 投后估值;差异大体可由 USD/AUD 汇率解释,但媒体报道中有些将公司估值写成「US$6 billion」,造成表层不一致。 截至本章生成日,Firmus 计划在 2026 年 6 月或 7 月冲刺 ASX IPO。报道称,本次上市拟募资 A$2–3 billion 新资本,IPO 顾问包括 Bank of America、JPMorgan、Morgans Financial 和 Morgan Stanley。公司已于 2026 年 4 月面向潜在投资者举行非交易路演。据报道,发给潜在股东的投资者信函显示 IPO「有望在 6 月或 7 月推进」,定价预期比当时估值高 40–100%,对应区间 A$8–12 billion。据称 James Packer 与 Nvidia 一同按约 A$145 每股买入约 A$50 million 股票。报道称,上市时不会对 IPO 前股东设置禁售期;独立基金经理已将这一点标记为重大结构性治理顾虑。 [CO023, CO024, CO025, CO026, CO027, CO028]

利益相关方或投资者图谱
利益相关方 / 投资者角色 / 关系估计持股 / 承诺控制权或经济重要性尽调问题
Nvidia (NVentures)战略股权投资者(2025 年 9 月和 2026 年 4 月轮);所有 Firmus AI Factories 的主要 GPU 供应商;架构标准制定者(DSX / Vera Rubin)未披露股权比例;Montgomery 估计 2025 年 9 月轮为 A$30–75M;2026 年 4 月继续参与关键:Nvidia 验证技术叙事,提供芯片供应优先级,并被估计会以硬件收入形式收回其大部分股权投资——形成共生但可能具腐蚀性的依赖核验实际股权比例和芯片供应合同条款;确认 Nvidia 是否拥有董事会席位或观察员权利
Ellerston Capital (Ashok Jacob)2025 年 9 月基石投资者;NED David Leslie 加入董事会~7% 持股(据 The Witness)重要治理监督;通过 David Leslie 拥有董事会席位;对 IPO 叙事具有声誉重要性确认当前持股和 IPO 前锁定意向
Coatue Management2026 年 4 月领投方(US$505M 股权)及支持贷款方(2026 年 2 月债务工具)最大股权轮领投方;未披露持股比例最大单笔股权支票;Robert Yin(GP 兼 AI Infrastructure 负责人)公开发声;增加全球机构可信度确认股权比例和董事会代表;了解与 Nvidia 的一致性
Blackstone(Tactical Opportunities & Credit & Insurance,债务融资方)2026 年 2 月 US$10 billion 债务工具牵头贷款人US$10B 工具(按时间提款)最重要的资本结构风险:契约、提款条件和违约触发器,可能在 AI Factory 建设延误或 GPU 价格下跌时限制运营灵活性获取工具条款、契约和抵押结构
Regal Funds Management(Phil King,投资方)早期投资者,约 2021 年通过 ~A$10M 可转债入场;参与 2025 年 9 月轮可转债 + 后续股权;持股已随估值上调大幅升值Phil King 的 Regal 是知名 ASX 基金;参与带来声誉合法性;早期投资者进入无托管 IPO,形成直接卖出风险确认当前持股和 IPO 卖出意向
Tectonic Investment Management投资者;约 7% 持股多家媒体报道约 7% 持股北河地区精品管理人;相对基金规模,Firmus 头寸集中核验持股,并评估集中度是否在 IPO 时造成被迫卖出风险
Maas Group Holdings(ASX: MGH;Wes Maas,投资方)战略投资者 + 供应链伙伴(JLE 电气子公司)A$100M(约 1.7% 持股);JLE 作为主权制造伙伴双重角色:财务投资者和嵌入式供应链;Maas 出售建材部门所得为 Firmus 持股提供资金评估 JLE 交付能力,以及同时作为投资者和供应商的利益冲突
Archibald Capital (Ben Madsen)早期机构投资者未披露精品私人信贷;与 Curtis 有个人关系(出席其 40 岁生日),引发关联方观感核验初始投资条款及任何顾问关系
Alex Waislitz / Pratt family(Visy / Heloise Pratt,股东)早期投资者未披露;被描述为股东墨尔本亿万富豪家族投资者网络;声誉信号确认当前持股及任何董事会 / 顾问角色
James Packer新股权投资者(2026 年 4 月)按 A$145/share 投入约 A$50M澳大利亚高知名度亿万富豪;显示 HNW 需求仍在;增加媒体叙事复杂度确认持股和任何治理权利
Tasmanian State Government战略伙伴和监管推动方;设立 AI Factory Zone(2025 年 7 月)无股权;监管 / 政策支持为塔斯马尼亚园区提供电网接入和政治合法性;水电能源是 Project Southgate 的关键差异点评估供电承诺及任何排他性或补贴安排
CDC Data Centres(Infratil / Future Fund,基础设施伙伴)Project Southgate 大陆站点基础设施 JV 伙伴不持有 Firmus 股权;设施业主和建设伙伴CDC 建设物理外壳;Firmus 安装专有 IT 内部系统——CDC 交付时间表上的相互依赖,是关键执行风险确认合同条款、排他性和 CDC 建设里程碑

股权比例来自媒体报道和投资者笔记,而不是官方股权结构表。2025 年轮次使用的 A$6 billion AUD 估值,与 2026 年 4 月官方新闻稿中的 US$5.5 billion USD 数字不同;两者按轮次语境分别使用。未披露持股反映私营公司不透明。

[CO023, CO024, CO025, CO026, CO027, CO028]
FO003: Firmus 快照 KPI

截至 2026-06-18 运行日,Firmus 的关键财务和运营指标凸显一家尚未披露收入的私营公司,在资本积累、规模野心和数据缺口上的快速扩张。

估值按近似现行 AUD/USD 汇率换算。收入和员工人数未公开披露。Project Southgate 容量和建设成本是 Firmus 的前瞻性估计,受融资、监管和建设条件约束。

[CO026, CO027, CO028, CO029, CO020, CO021]

1.4 里程碑、运营规模与不利事件

Firmus 的运营记录横跨新加坡部署、塔斯马尼亚建设,以及 Project Southgate 旗下快速扩张的澳大利亚全国 rollout。新加坡业务自 2024 Q1 通过 STT GDC 合作上线,是公司 2026 年前唯一大规模运营、能产生收入的 AI Factory。Firmus 在新加坡两个 STT 站点部署约 4,000 块 Nvidia GPU,为 SEA-LION 项目(AI Singapore 面向东南亚语言的开源 LLM)部署总算力 99 PFLOP/s。新加坡设施在 2024 年获得 DCD Asia-Pacific Data Center Project of the Year 奖,SemiAnalysis 在 2024 年和 2025 年均将 Firmus 列入其全球 Top 3 GPU Cloud providers(ClusterMAX)。 Project Southgate 是公司的旗舰增长载体,随 2025 年 9 月股权融资一同公布,并在 2025 年 10 月通过 CDC Data Centres 与 Nvidia 战略联盟扩大。初始阶段集中在塔斯马尼亚 Launceston 园区——由该岛州水电电网供电——目标 90 MW AI 算力(Stage 1a 为 44 MW,Stage 1b 为 90 MW)。此后项目扩展为全国计划,站点覆盖 Melbourne、Canberra、Sydney 和 Perth,到 2028 年规划总容量达到 1.6 GW。该计划围绕 Nvidia Vera Rubin DSX 参考设计建设,并与 CDC Data Centres 共同开发;披露总建设成本为 A$73.3 billion。Firmus 发言人在 2025 年 11 月表示,Melbourne 和 Tasmanian 中心预计分别于 2026 年 4 月和 8 月上线,员工将在 2026 年上半年进驻 Canberra 中心。 公司可被基准验证的资质包括 MLPerf V4.0 Certified Training and Power 结果,以及 2025 年 Cleantech Group APAC Top 25 奖。但 Firmus 最突出的效率主张——Power Usage Effectiveness(PUE)为 1.03——已引来实质性专家怀疑。Firmus 发言人确认,该数字是 2021 年塔斯马尼亚一次浸没式部署的「点测量」,无法独立验证,因为不存在公认的 PUE 认证机构。UNSW 高级研究员 Dr Amr Omar 称 1.03 的 PUE「已经达到当今全球任何地方可能实现水平的极端边缘」,并指出全球超大规模云厂商通常约为 1.2。Sam Maher(GovTech Australia)担心,一些投资者被叙事带动,而不是被底层基础设施逻辑说服。Roger Montgomery Investment Management 还提出额外担忧:Firmus 依赖 Nvidia(估计 Nvidia 股权投资为数千万,撬动 A$830 million 其他股权和 US$10 billion 债务,其中相当部分会以硬件收入回流 Nvidia),且没有公开披露长期有约束力的客户合同。 监管侧:2026 年 3 月,Firmus 签署一项涉及 18,400 块 Nvidia GB300 GPU 的 Melbourne AI Factory 数十亿美元协议;2025 年 7 月,塔斯马尼亚政府设立全球首个 AI Factory Zone;2025 年 7 月,Firmus 与 Singapore Port Authority 签署关于海水冷却 AI 算力的 MoU。 [CO019, CO020, CO021, CO022, CO030, CO031]

里程碑表
日期事件类型金额 / 估值 / 状态参与方含义
2019-12-12Firmus Grid Pty Ltd 在澳大利亚注册成立创立Oliver Curtis、Tim Rosenfield、Jonathan Levee;Nick Curtis(首任主席)运营实体的法律基础;公司最初聚焦于塔斯马尼亚的比特币挖矿冷却技术
2021-11-21Firmus Technologies Pty Ltd 注册成立;Hypercube 研发正式化创立同一创始团队面向 AI 算力应用的独立技术实体成立;浸没式冷却研发推进中
2021Regal Funds Management 可转债投资融资~A$10M 可转债Regal Funds Management(Phil King,投资方)已识别的最早外部机构资本;投资时公司仍是比特币挖矿冷却供应商
2022使用可再生能源的塔斯马尼亚规模化浸没式冷却数据大厅完工产品Firmus 工程团队首个已完工、具备 AI 能力的设施;1.03 的点 PUE 测量归因于该部署
2023全集成 AI 算力平台完成;e-Nome 前身活动完成产品Firmus 团队平台已准备好面向外部客户部署;从比特币冷却到 AI 的转型已完全执行
2023-06与 ST Telemedia Global Data Centres (STT GDC) 建立合作;SMC 在新加坡部署,用于主权 AI合作据称投资 ~US$100M(未核验)Firmus、STT GDC(Temasek 持有)、AI Singapore(SEA-LION LLMs)首个产生收入的国际 AI Factory;两个 STT 站点约 4,000 块 Nvidia GPU;为 SEA-LION 提供 99 PFLOP/s 算力
2024-Q3Synert 电网集成软件与 Eaton 在塔斯马尼亚部署——首个上线电网集成项目产品Firmus、Eaton证明运营软件能力;强化 Project Southgate 能源集成论据
2024-Q4新加坡 AI Factory 获 DCD 亚太年度数据中心项目奖规模奖项Data Center Dynamics、Firmus能源效率和设计获得外部验证;此后所有融资材料均引用
2024-Q4SemiAnalysis ClusterMAX Top 3 Global GPU Cloud 排名(首次获奖)规模排名SemiAnalysis与全球超大规模云厂商并列排名;在投资人材料中被重点引用,用来证明世界级性能
2025-07塔斯马尼亚政府在北塔斯马尼亚设立全球首个 AI Factory Zone监管Firmus、塔斯马尼亚州州长 Jeremy Rockliff政策背书;电网接入和规划审批已落实;支撑 Project Southgate Stage 1
2025-07与 Singapore Port Authority 签署海水冷却 AI 算力 MoU合作Firmus、Singapore Port Authority释放 APAC 扩张信号;赤道地区站点若采用海水冷却,可减少淡水消耗
2025-09-16A$330M 股权融资,投后估值 A$1.85B;Nvidia 作为战略投资方加入;Ellerston Capital 领投融资A$330M,投后估值 A$1.85BNvidia、Ellerston Capital、Regal FM、Archibald Capital、Tectonic、Waislitz、Pratt family 等投资方公司跻身独角兽;Nvidia 的战略投资增强芯片供应确定性,也带来声誉光环
2025-10宣布与 CDC Data Centres 和 Nvidia 建立战略联盟;Project Southgate 全国扩张为 A$73.3B 计划合作A$73.3B 总估算计划成本CDC Data Centres、Nvidia、Firmus 等合作方在塔斯马尼亚、墨尔本、悉尼、堪培拉、珀斯全国铺开;CDC 提供物理机房壳体;Nvidia 提供 GPU 和 DSX 架构
2025-11-14A$500M 股权融资,估值 A$6B;Project Southgate 全国铺开提速融资A$500M,投后估值约 A$6B(AUD)Morgans(独家牵头)、Highbury Partnership(顾问)、机构投资者(UniSuper、Wilson AM 等)八周内估值接近三倍;全国计划在 CDC / Nvidia 合作下获得全额资金支持
2026-02-09完成 US$10B 债务融资额度,由 Blackstone(Tactical Opportunities 和 Credit & Insurance) 牵头,Coatue 支持融资US$10B 额度Blackstone、Coatue澳大利亚史上规模最大的私募债融资之一;资金用于 Project Southgate 全国建设
2026-02Maas Group Holdings 以 A$6B 估值战略投资 A$100M;达成 JLE 供应链合作融资A$100M(约 1.7% 股权)Maas Group Holdings(ASX: MGH,Wes Maas)、Benmax 等相关方为冷却、电力和集成系统模块建立澳大利亚本土主权制造供应链
2026-02-23ASX 发布 Good Fame and Character 合规更新(近年首份专项更新)监管ASX Listings Compliance department(上市合规部门)市场普遍理解为与 Firmus 即将上市有关;框架会评估高管过往记录,包括 Oliver Curtis 2016 年的定罪
2026-03Firmus Grid Limited 董事会任命三名新的非执行董事治理Lee Hatton、Christine Bartlett、Julie Shuttleworth AM 三名董事IPO 前董事会补强;引入审计、薪酬和运营经验;直接回应治理审查
2026-03据报道 Oliver Curtis 已通过 ASX Good Fame and Character 测试监管已获批准ASX Listings Compliance、Oliver Curtis 等相关方Curtis 可在上市实体担任联席 CEO 和董事;Firmus 登陆 ASX 的主要监管障碍降低
2026-03签署数十亿澳元规模的墨尔本 AI Factory 交易;部署 18,400 块 Nvidia GB300 GPU产品数十亿 AUD 交易Firmus、Nvidia、CDC Data Centres 等合作方已宣布的最大单站点 GPU 部署;墨尔本站点预计 2026 年年中上线
2026-04-06Coatue 领投、Nvidia 参投,完成 US$505M 战略股权融资,投后估值 US$5.5B融资US$505M,投后估值 US$5.5BCoatue Management(领投)、NvidiaIPO 前最后一轮私募融资;六个月内累计股权融资 A$1.95B(US$1.35B);公司被称为 APAC 资本最充足的私营 AI 基础设施公司
2026-mid目标在 ASX IPO(2026 年 6 月或 7 月)融资目标募资 A$2–3B;IPO 隐含估值 A$8–12BBank of America、JPMorgan、Morgans Financial、Morgan Stanley(顾问)可能成为澳大利亚史上最大科技上市之一;据报道 IPO 前股东不设锁定期

里程碑日期来自 Firmus 官方新闻稿、Caplight、Bloomberg、Startup Daily、SmartCompany 和监管记录。STT GDC 投资金额(US$100M)由 The Witness 提及,但尚未获得一手来源确认。2019 年创立这一点各方口径一致;精确创立月份 / 地点仍有小冲突未解决(Tracxn 显示 2019 年 12 月注册,不同来源又写作悉尼或塔斯马尼亚)。IPO 日期和募资额仍取决于市场条件。

[CO001, CO003, CO004, CO011, CO022, CO023]
FO001: Firmus 公司里程碑时间线

这条时间线按时间梳理 Firmus 从 2019 年作为比特币挖矿冷却公司创立,到 2026 年 Series F 融资和计划 ASX IPO 的关键节点,覆盖创立、产品转向、合作、融资、治理变化和负面事件。

STT GDC 合作日期约为 2023 年中;Regal 可转债按媒体报道约在 2021 年(无一手来源确认)。IPO 日期和估值区间来自投资者通信和媒体报道,并非 Firmus 官方公告。

[CO001, CO003, CO020, CO022, CO023, CO024]

1.5 展示材料

Chapter 02

02市场分析

2.1 市场边界:AI 云、主权 AI 基础设施与 GPU 容量

Firmus 把自身放在三个相互重叠的支出类别交汇处。第一类是超大规模 GPU 算力:全球云服务商购买签约、高密度 AI 工厂容量,以满足超过自建速度的需求。第二类是主权 AI 基础设施:AI 算力必须留在本国治理之下,使用本地能源供给,并符合国家数据驻留要求。第三类是企业和政府 AI 云服务:Firmus AI Cloud 向商业和公共部门客户提供受管的前沿 GPU 容量,而客户无需自建。通用 IaaS 和传统 colocation 被明确排除:Firmus 只瞄准 AI 高密度、能源优化、GPU 原生负载。现状替代方案包括全球 hyperscaler 云(AWS、Azure、GCP)、本地自建 on-premises GPU 集群,以及 NEXTDC 和 Equinix 等传统 colocation 运营商;这些方案都无法在澳大利亚本土以规模化方式提供可再生能源供电的主权 AI 工厂经济性。澳大利亚拥有 145 个运营中的 colocation 数据中心,数据中心市场预计到 2031 年达到 USD 9.02 billion,是锚定地理市场;APAC 出口是明确目标。相邻支出类别——通用企业云、网络安全基础设施和网络运营商服务——虽相邻,但不纳入主要市场边界,因为 Firmus 商业上并不瞄准这些领域。 [CM001, CM002, CM003, CM004, CM005, CM006]

市场定义:纳入的细分、排除的支出以及与 Firmus 的关系
细分 / 类别纳入支出排除支出主要买方 / 付款方与 Firmus 的关系
超大规模 GPU 工厂容量长期合约化 AI 算力(GPU 集群、冷却、电力、网络)物流、通用 IT、非 AI 工作负载全球云 / AI 公司(capex 资金)核心收入细分;Southgate 墨尔本锚定合同
主权 AI 基础设施政府要求的境内 AI 算力、符合数据驻留要求的设施离岸主权云、非 AI 政府 IT政府部门和受监管行业(采购预算)战略差异化;澳大利亚所有权和可再生能源资质
企业 AI 云(托管)通过 Firmus AI Cloud 按需使用 GPU、托管推理和训练原始 IaaS、非 AI 企业 SaaS、传统 ERP掌握 AI 转型预算的 CIO / CTO第三个细分;单价收入较低,但客户覆盖更广
研究与教育 AI 算力大学和国家研究 GPU 资助、HPC 替代大学通用 IT 采购、网络服务研究经费,院系 / IT 联合承担早期细分;验证业务,也为政府声誉提供锚点
AI 原生初创 / 成长期公司容量用于模型训练和微调的按需 GPU 容量非 AI SaaS、支付 / 金融科技基础设施创始人或 Series A–C capex 预算利润薄但有量的细分;建设品牌和销售管线
通用 IaaS / 主机托管(排除)无——Firmus 不在此竞争不具备 GPU 密度的传统机架、VM 或算力托管企业 IT、中端市场范围外;Firmus 明确只瞄准 AI 优化工作负载
邻近领域:网络 / CDN / 边缘上线初期无;未来可能扩张边缘 CDN、运营商中立互联、WAN电信、媒体、企业邻近领域;不在当前产品路线图内

边界定义基于 Firmus 官方定位和公司公告;排除细分来自产品范围和竞争语境。收入相关性基于合同证据和客户表述作定性判断。

[CM001, CM002, CM003, CM006]
FM004: Firmus 价值链——从 GPU 供应到终端客户工作负载

这条价值链展示 NVIDIA GPU 供应、澳大利亚主权基础设施和 CDC 数据中心托管如何组合起来,把 AI 计算交付给超大规模和企业买方。

[CM001, CM004, CM020, CM039]

2.2 机会规模:相互矛盾的估算与工作边界

2026 年 AI 基础设施市场规模估算因研究者边界定义不同,出现异常宽的区间。Grand View Research 将 2026 年全球 AI 数据中心市场定为 USD 180.6 billion,并预计到 2033 年 CAGR 为 23.9%;MarketsandMarkets 使用更宽口径,得出 2025 年 USD 344 billion、2026 年升至 USD 471 billion——差距超过 2.5x,原因在于指标是否纳入与 AI 负载共同部署的全部硬件和服务,还是只统计专用 AI 算力。Gartner 收窄到主权云 IaaS,预测 2026 年全球为 USD 80 billion——同比增长 35.6%——其中 Mature Asia/Pacific 增长 87%,达到 USD 1.59 billion。Research and Markets 估计,总主权云市场(所有部署模式)2026 年为 USD 128.6 billion,并以 23.4% CAGR 增至 2030 年 USD 298 billion。Data Bridge Market Research 采用最窄定义——仅主权 AI 基础设施硬件和系统——将该细分市场 2025 年估值为 USD 14.82 billion,预测到 2033 年达到 USD 49.67 billion,CAGR 为 16.3%,APAC 是增长最快区域,CAGR 为 17.8%。澳大利亚方面,整个数据中心市场预计从 2025 年 USD 4.22 billion 增至 2031 年 USD 9.02 billion,CAGR 为 13.5%。美国前五大 hyperscaler(Microsoft、Amazon、Alphabet、Meta、Oracle)合计承诺 2026 年 capex 为 USD 660–690 billion,几乎较 2025 年翻倍,四家最大公司均自称受供给约束而非需求约束。capex 暴增形成结构性拉力,使 Firmus 工厂模式具备商业可行性:Firmus 部署的每个 GPU 集群,都可能成为 hyperscaler 在主要市场无法足够快自建时的溢出容量。衡量 Firmus 自身机会,需要叠加三层镜头:以上述全球市场数据锚定的 TAM;受 APAC 主权 AI 工厂容量约束的 SAM;以及由 Firmus 签约管线和到 2028 年建设余量推导的 SOM。 [CM008, CM009, CM010, CM011, CM012, CM013]

市场规模视角:按发布方、地域和方法论划分的 TAM / SAM 估算
发布方年份地域指标数值(USD B)CAGR方法论说明置信度局限
Grand View Research2026全球AI 数据中心市场180.623.9% (2026–2033)专用 AI DC 硬件和服务收入范围窄于 MnM;不包括邻近服务
MarketsandMarkets2026全球AI 数据中心市场471.627.5% (2026–2032)更宽口径,包含完整 AI 基础设施栈与 GVR 相差 2.5×,说明定义差异显著
Gartner2026全球主权云 IaaS 支出80.4~35%(YoY)仅 IaaS 支出;不包括 PaaS、SaaS、托管服务仅 IaaS;不计算更广义的主权 AI 硬件
Gartner2026成熟亚太主权云 IaaS 支出1.687%(YoY)包括澳大利亚、新西兰、新加坡、韩国、香港区域包含澳大利亚以外市场;未单独拆出 AU 份额
Data Bridge Market Research(市场研究机构)2025 基准 / 2033 预测全球主权 AI 基础设施市场14.8 (2025) → 49.7 (2033)16.3% (2026–2033)最窄口径:主权部署下的 AI 服务器、存储、网络、GPU 加速器定义最窄;APAC 增速最快,CAGR 17.8%
Research and Markets(市场研究机构)2026全球主权云市场(所有模式)128.623.4%(至 2030 年:$298B)包括 IaaS、PaaS、SaaS、托管主权云中(付费墙)主权定义最宽;颗粒度低于 Gartner
Mordor Intelligence2026亚太数据中心 GPU 市场22.216.3%(2026–2031:$47.1B)APAC 数据中心的 GPU 硬件和部署中(受限)页面访问受限;数字来自搜索综合
Yahoo Finance / Arizton2025 基准 / 2031 预测澳大利亚数据中心市场(总体)4.2 (2025) → 9.0 (2031)13.5% (2025–2031)澳大利亚所有主机托管和超大规模 DC 投资Firmus 容量计划约为当前已安装市场的 8–17×
Futurum Group / OfficeChai2026全球美国前五大超大规模云厂商 AI capex660–715N/A(年度 capex)Amazon、Google、Microsoft、Meta、Oracle 合计 capex 指引Capex ≠ 市场规模;包含非 AI 和物流支出

估算横跨四种不同市场定义:(1)全部 AI 数据中心收入,(2)仅主权云 IaaS,(3)窄口径主权 AI 硬件,(4)超大规模云厂商总体 capex。它们不可比,也不应相加。CAGR 适用于各来源自己的预测窗口。数值四舍五入到一位小数。Gartner 和 Futurum 数字直接来自已抓取的一手页面。

[CM008, CM009, CM010, CM011, CM012, CM014]
FM001: 市场规模金字塔——TAM、SAM 与 Firmus SOM(2026–2028 年估计)

分层展示全球 AI 数据中心 TAM、主权 AI 基础设施 SAM,以及基于 2028 年前已签约和计划容量估算的 Firmus SOM。

TAM 数字来自不同来源,定义彼此不兼容(见 TM002)。SAM 由 APAC 最高增速份额套用至 Data Bridge 全球估计得出,只是粗略代理。SOM 由作者按已签约 GPU 数量估算(约 55,000 块 GB300 GPU × 估计平均 TCO/GPU-year),高度不确定;Firmus 尚未披露收入。不要依赖金字塔比例做财务建模。

[CM008, CM012, CM040, CM041]
FM002: 市场估计区间——AI 与主权云市场口径相互矛盾(2026 年)

与 Firmus 相关的五种不同市场定义给出低、基准和高三组估计,说明定义不确定性让单一 TAM 数字容易误导。

低 / 基准 / 高边界由作者根据各来源估计区间推导(见 TM002)。全球 AI DC:GVR($180B)到 MnM($472B)。主权云 IaaS:Gartner 中点 ±10%。主权 AI infra:Data Bridge 2025 年基数按 16.3% CAGR 推至 2026 年,并上下浮动 ±20%。澳大利亚 DC:Arizton 2025 年基数按 13.5% CAGR 推算,并上下浮动 ±15%。超大规模云厂商资本开支:Futurum 和 OfficeChai 区间。各行定义不可比较;不要直接相加或横向比较。

[CM008, CM009, CM010, CM015, CM016, CM017]

2.3 买方、用户与付款方格局

三类买方驱动 Firmus 当前和近期收入。Hyperscaler 是锚定租户群体:Southgate Melbourne 一份超大规模合同,在与一家未具名全球领先科技公司的多年、数十亿美元协议下,承诺约 18,400 块 NVIDIA GB300 GPU,代表澳大利亚首个超大规模 AI 工厂合同。这类买方把容量当作基础设施债采购——长期、签约,结构常类似不动产——预算来自全球资本开支计划,而非运营 IT 预算。用户与付款方是同一实体;采购由基础设施和云投资团队牵头,不是企业 IT。Gartner 将政府列为全球主权云 IaaS 的主要买方,其后是受监管行业(金融服务、医疗、国防)和关键基础设施运营方。2025 年,政府与公共部门占全球主权 AI 基础设施市场 33%。该细分在澳大利亚是 Firmus 的战略重点:DGX Cloud(NVIDIA 的受管云平台)在 Southgate Melbourne 建立新区,让联邦和州机构首次获得认证的主权前沿 GPU 算力。在该细分中,付款方是政府 IT 或 whole-of-government 预算机构;买方是中央采购或技术办公室;用户分散(国防、科研机构、医疗、移民)。第三类——企业、教育、科研和 AI-native 初创公司——通过 Firmus AI Cloud 以受管服务获取算力。相比设施主权,这类买方更在意 token 成本和 API 可用性,预算归 CIO 或 CTO 职能所有。采用触发因素包括模型训练成本基准、延迟要求,或排除离岸云的监管数据驻留要求。全球范围内,国防与情报是增长最快的主权 AI 基础设施终端用户,CAGR 为 17.6%。 [CM019, CM020, CM021, CM022, CM023, CM024]

细分与买方地图
细分买方用户付款方工作流 / 用例预算负责人采用触发因素
超大规模云 / AI全球基础设施 / 云投资团队超大规模云平台及其终端客户全球 capex 计划LLM 训练、规模化推理、AI 实验室工作负载CFO / 基础设施资本委员会主要超大规模云数据中心供应受限;主权或 ESG 要求
政府 / 公共部门中央政府 IT 采购机构员工、情报、国防、医疗全政府或部门 IT 预算主权 AI 模型开发、涉密分析、公民服务 AI全政府 / DTA / 机构 CIO数据主权要求、安全密级、禁止使用离岸云
受监管企业银行、保险公司、电信公司、能源公司的 CIO / CTO数据和 AI 工程团队IT 转型或 AI 创新预算模型微调、风险模型、欺诈检测、客服 AICIO / CTO监管数据驻留要求;相对离岸云的成本基准
研究与教育大学 IT / 院系研究团队研究人员、博士生ARC 或 NHMRC 研究经费LLM 预训练、基因组学、气候建模科研办公室或院系院长HPC 系统算力配额耗尽;AI 项目由经费支持
AI 原生初创 / 成长期公司创始人或工程负责人ML 工程师Series A–C capex 预算模型训练、推理 API 服务、数据处理创始人 / CTO相对超大规模云 GPU 云的成本套利;上市速度
国防与情报(全球增速最快)国防采购情报分析师、自主系统团队国防能力预算监视、情报分析、自主系统国防 CIO / 能力经理国家 AI 战略;美国超大规模云厂商无法提供涉密算力

买方和付款方往往是同一法律实体,但内部职能不同。国防细分因采购路径不同且全球 CAGR 达 17.6%(Data Bridge),故与政府分列。Firmus 尚未公开澳大利亚政府买方结构。

[CM006, CM019, CM021, CM022, CM023, CM024]
FM003: 买方与细分地图——按客户类型划分采购特征

对五类买方原型进行比较,维度包括预算归属、主要采购驱动、数据敏感度、典型合同结构和当前采用阶段。

采用阶段是基于公开披露合同和 Firmus 客户沟通的定性评估。截至本报告日期,没有独立客户调研数据。

[CM006, CM019, CM022, CM023, CM024]

2.4 增长驱动、采用约束与监管摩擦

首要增长驱动是 hyperscaler capex 暴增:美国前五大基础设施公司 2026 年部署 USD 660–690 billion,并均称 AI 容量建成多快就被吸收多快。Microsoft 披露未履行 Azure 订单积压 USD 80 billion,唯一闸口是电力接入,说明约束在能源连接,不在需求。Firmus 的定位是吸收溢出:其塔斯马尼亚站点 PUE 约 1.03,借助水电 baseload,相比澳大利亚大陆偏重燃气的电网,结构上更清洁、价格也更稳定。第二个驱动是地缘政治和主权压力:已有 100 多个国家出台数据本地化或 AI 主权要求,Gartner 预测,随着 geopatriation 加速,到 2028 年 20% 主权云 IaaS 负载将从全球供应商转向本地供应商。澳大利亚数据中心支出在 2025 年 9 月季度增长超过 60%,达到 A$2.6 billion——同比增长 140%——政府则在 NSW 快速推进 15 个数据中心项目。但采用面临三项结构性约束。第一是电网容量:Transgrid(NSW 输电运营商)告诉议会调查,其没有能力为规划管线中的所有数据中心供电;AEMO 预测,到 FY2030 数据中心将消耗 12 TWh(NEM 需求的 6%),在高增长情境下到 2040 年可能升至总电力的 13%。第二是新的监管义务:AEMC 2026 年 3 月规则草案把「大型」逆变器型负载定义提高到 30 MW,并要求具备 ride-through 合规以防级联停电,这会给每个接入 National Electricity Market 的设施增加工程成本和连接时间不确定性。第三,Greenpeace 提出的暂停令呼吁得到部分州能源部长支持,形成政治风险,可能拉长规划审批,并引入附条件的 offset 要求——新设施可能需要先用新增可再生发电完全抵消其用电需求,才能获得批准。另一个不确定性是,AI 模型效率提升(类似云单位经济中已观察到的 Jevons paradox)是否会让单一负载算力强度下降快于采用扩张需求;若发生,实际市场规模可能低于当前预测。 [CM028, CM029, CM030, CM031, CM032, CM033]

增长驱动因素与采用约束
因素类型方向时间对 Firmus 的影响尽调问题
超大规模云 capex 激增(2026 年 $660–715B)驱动因素正向当前(2026 年)创造超出超大规模云自建能力的合约化 AI 工厂容量溢出需求确认包销合同期限和续约条款
主权 AI 要求和数据驻留法律(100+ 个国家)驱动因素正向当前并在加速Firmus 的澳大利亚主权资质打开政府采购,而这些采购会排除离岸云梳理哪些 AU 政府框架明确要求境内 AI 算力
地缘回流趋势(到 2028 年,20% 工作负载从全球供应商转向本地供应商)驱动因素正向2026–2028支撑受监管行业和公共部门的长期持续需求跟踪 Gartner 对 AU 主权云 IaaS 增长指标的更新
APAC AI 基础设施缺口(全球主权 AI 增速最快区域,CAGR 17.8%)驱动因素正向多年澳大利亚有条件成为面向盟友市场的 APAC 枢纽;东南亚也有扩张路径监测 ASEAN 政府 AI 基础设施采购管线
塔斯马尼亚水电电网(PUE ~1.03,接近纪录的能效)驱动因素正向当前在 token 经济性上形成结构性成本优势;也为超大规模云买方提供 ESG 资质向第三方审计方索取独立 PUE 验证
澳大利亚数据中心支出激增(2025 年 Q3 +60%,YoY +140%)驱动因素正向当前支出数据证明市场接受度;NSW 政府加速审批带来政策顺风核验 ABS 数据源,以及 Firmus capex 是否被纳入
电网容量约束(Transgrid:NSW 排队项目超过接入容量)约束负向现在–2027 年悉尼站点面临接入排队和时间不确定性;塔斯马尼亚和墨尔本站点更有利确认每个规划 Southgate 站点的电网接入状态
AEMC 2026 年 3 月电网标准草案(30 MW 门槛、穿越能力合规)约束负向2026 年–2026 年年中定稿所有接入 NEM 的设施都将新增工程合规成本和接入协议要求评估每个工厂站点为满足穿越能力要求所需的增量 capex
监管风险:AEMC / COAG 提议要求新 DC 全额可再生能源抵消约束负向2026–2027可能拉长许可周期;Firmus 的可再生能源整合可能让其获得快速通道待遇跟踪 AEMC 2026 年 7 月报告以及州部长对强制抵消规则的立场
Greenpeace 和 NGO 压力:呼吁暂停批准新数据中心约束负向当前–持续政治和声誉风险;可能放慢大陆站点(墨尔本、悉尼、珀斯)的规划审批跟踪 Kemps Creek 和大陆站点的州规划决定与社区反对
AI 模型效率风险(Jevons 悖论不确定性)约束混合2026–2030推理变便宜后,整体算力需求可能上升,也可能下降;2025 年 Gemini 服务成本下降 78%,说明效率迭代速度很快在财务模型中加入每单位工作负载 GPU-hours 下降 20% 和 50% 的敏感性

驱动 / 约束的判断基于截至 2026-06-18 的公开证据。时间估计为定性判断。「尽调问题」列列出可让评估更精确的证据缺口。AEMC 规则截至本报告日期仍是草案;最终规则预计 2026 年年中出台。

[CM017, CM018, CM028, CM029, CM030, CM031]

2.5 展示材料

Chapter 03

03竞争对手

3.1 直接 GPU 云与 Neocloud 对手

全球范围内,Firmus 最接近的结构类比是 CoreWeave:它同样从加密挖矿转型,并成为美国领先的纯 GPU 云。CoreWeave 于 2025 年 3 月以 US$40 IPO 价格上市,到 2026 年 3 月市值约 US$43 billion;FY2025 收入 US$5.13 billion,同比增长 168%。其 2026 Q1 收入 US$2.08 billion,确认超高速增长仍在继续,公司指引 2026 全年收入 US$12–13 billion。到 2026 Q1,CoreWeave 收入 backlog 达 US$99.4 billion,由与 Microsoft、OpenAI 和 Meta 的多年 take-or-pay 合同支撑。尽管规模很大,CoreWeave 仍背负约 US$25 billion 债务,GAAP 亏损扩大(FY2025 为 US$1.17B),客户集中度极高——Microsoft 占 2024 年收入 62%。CoreWeave 在 APAC 没有实质存在,主要服务美国 hyperscale AI labs。 Lambda Labs 是第二个重要 neocloud 对手,2025 年收入估计约 US$760 million(同比增长 79%),并在 2025 年 11 月完成 US$1.5 billion Series E 后估值 US$5.9 billion。Lambda 的竞争武器是激进定价:其 NVIDIA H100 实例约 US$2.49/hr,而 CoreWeave 约 US$4.25/hr;开发者导向产品吸引 AI 初创公司和研究团队。Lambda 已提交 IPO 文件,目标 2026 H1,这可能扩大其资本半径,并潜在推动 APAC 扩张。与 CoreWeave 一样,Lambda 没有既成的澳大利亚主权存在,也没有 IRAP 相关合规姿态。 Vultr 处于较低梯队,提供 GPU 实例和透明公开定价(H100 约 US$2.60/hr),拥有 32 个全球数据中心,但目标是中小企业,而非 hyperscale AI 负载。除它之外,没有其他 neocloud 展现出 Firmus 定位中的基础设施规模、APAC 主权角度或 NVIDIA 合作深度。[CP013, CP014, CP015, CP016, CP017, CP018]

竞争对手画像摘要
竞争对手类别规模 / 融资目标客群关键差异点Firmus 竞争风险
CoreWeave(美国)直接 GPU neocloudFY2025 收入 US$5.1B;市值 US$43B;债务 US$25B;2025 年 3 月 IPO大型 AI 实验室、超大规模企业规模、US$99.4B 积压订单、NVIDIA Elite 合作伙伴、SUNK/Mission Control 软件中 — 尚未进入 APAC;模式以美国为中心
Lambda Labs(美国)直接 GPU neocloud2025 年收入 US$760M(估计);估值 US$5.9B;已融资 US$2.3BAI 开发者、初创公司、研究团队激进定价(H100 ~US$2.49/hr);开发者优先 UX;目标 2026 年 IPO中 — 价格竞争;IPO 后可能进入 APAC
Vultr(美国)中端市场 GPU 云全球 32 个 DC;私有公司;收入未披露SMB、AI 初创公司、SaaS 团队透明统一定价;GPU SKU 覆盖广;托管 Kubernetes低 — 不是 APAC 主权云;面向预算较低客群
NEXTDC(澳大利亚)本土 AI 基础设施在位者2026 年 4 月融资 A$2.2B;与 OpenAI 共建 A$7B S7 园区;FY26 capex A$3B超大规模云厂商、大型企业、政府企业级认证;ASX 上市;OpenAI 主权合作;规模已验证高 — APAC 直接对手;CEO 公开质疑 Firmus 建设质量
CDC Data Centres(澳大利亚)基础设施伙伴 / 竞争者澳大利亚最大私有 DC 运营商;未上市政府、企业、超大规模云厂商主权骨干;近零用水;已是 Firmus 的 DC 基础设施供应商低-中 — 伙伴依赖;若 CDC 自建 GPU 云,可能出现冲突
Macquarie Data Centres(澳大利亚)本土基础设施在位者47 MW IC3 Super West(一期 A$350M);Dell/NVIDIA AI Factory 合作伙伴企业、政府、金融服务主权所有权;液冷;接入 Dell AI Factory;2026 年 Q3 开放低-中 — 重点是托管;尚无 GPU 云覆盖层
AUCloud / AUCyber(澳大利亚)主权云细分玩家收入约数千万澳元低位(估计);ASX 上市政府、国防、受监管行业IRAP 认证;主权 IaaS;MSSP 托管服务低 — 不具备规模化 GPU 能力;市场层级不同
AWS / Azure / GCP超大规模云在位者已承诺合计 A$20B+ 澳大利亚 capex;全球 GPU 集群企业、开发者、政府(广覆盖)覆盖面、生态、IRAP 评估服务、品牌、全球 SLA高 — 主权云快速跟进者;价格高但集成很深

规模 / 融资数据来自运行日期前的公司文件、新闻稿和分析师估计;Vultr 收入未公开披露;AUCloud/AUCyber 收入为第三方分析师估计。「Firmus 竞争风险」是基于本章证据给出的编辑严重性判断,不是量化评分。空白或未知单元格表示缺少公开披露。

[CP013, CP014, CP019, CP021, CP022, CP023]
GPU 云定价与打包对比
供应商GPU SKU公开标价(USD/hr)合同模式数据主权关键限制
FirmusNVIDIA GB300 NVL72 / GB300 cluster GPU 集群未发布企业定制;多年 take-or-pay澳大利亚主权无公开现货市场;定价不透明;仅面向企业洽谈
CoreWeaveNVIDIA H100 SXM (80 GB)~US$4.25按需 + 多年预留美国本土(无澳大利亚主权)杠杆高;债务 ~US$25B;客户集中风险
Lambda LabsNVIDIA H100 PCIe (80 GB)~US$2.49按需 + 预留美国本土(无澳大利亚主权)无 APAC 数据驻留;开发者级支持
AWS(澳大利亚区域)NVIDIA H100(通过 p5.48xlarge)~US$6.88按需 + Savings Plans + EC2 RI澳大利亚区域;部分服务 IRAP溢价定价;超大规模云捆绑;数据出站成本
Microsoft Azure(澳大利亚区域)NVIDIA H100(通过 ND H100 v5)~US$12.29按需 + Enterprise Agreement澳大利亚区域;部分服务 IRAP公开标价最高;EA 结构复杂;库存滞后
Google Cloud(澳大利亚区域)NVIDIA H100(通过 A3 系列)~US$10.98按需 + 承诺使用;TPU 替代方案澳大利亚区域;部分服务 IRAPGoogle Workspace/Analytics 捆绑;IRAP 范围受服务限制
VultrNVIDIA H100 SXM5 (80 GB)~US$2.60按需;1 年期 15-22% 折扣;3 年期 25-38% 折扣全球 32 个 DC;无澳大利亚主权仅中端市场;深度不及 neocloud 或超大规模云;无澳大利亚 DC

价格为截至约 2026 年 6 月的公开按需价或公开标价;超大规模云费率基于分析师聚合数据估算,可能不同于谈判后的企业协议。Firmus 没有公开可用定价;所有企业合作均为定制。承诺使用、预留和现货定价可让各供应商公开标价降低 15–40%。专业规模 GB300 集群定价无法直接与单 GPU H100 按需费率相比;这些行呈现可获得的最佳公开证据,不是同口径对比。

[CP020, CP035, CP007, CP009]
FP001: 竞争定位——AI 成本效率 vs 主权强度

Firmus 位于右上象限(高效率主张、高主权强度),没有单一竞争对手与其并列;NEXTDC 和 AUCloud 主权性高但 GPU 效率落后;CoreWeave 和 Lambda Labs 效率领先但没有 APAC 主权性。坐标轴是序数编辑评分(1–10),不是量化指标。

坐标轴代表基于公开证据推导的序数编辑评分(1–10 分),并非独立审计的数字指标。X 轴(AI 成本效率 / 每瓦性能)参考 SemiAnalysis ClusterMAX 数据以及各提供商定价比较。Y 轴(主权 / 本地所有权强度)反映所有权结构、IRAP 认证状态和可再生能源承诺。CoreWeave、Lambda Labs 和 Vultr 没有澳大利亚主权产品;其 y 分数只反映全球姿态。

[CP002, CP013, CP019, CP021, CP026, CP027]

3.2 澳大利亚基础设施既有玩家

在澳大利亚,NEXTDC 是最强大的本土对手。这家 ASX 上市运营商在 2026 年 4 月宣布 A$2.2 billion 融资(A$1.5B 股权 + A$700M 债务),此前获得其所称史上最大客户合同——S4 园区 250 MW 承诺,外界普遍认为来自 Microsoft。NEXTDC 将 FY26 capex 指引上调至 A$3 billion,管理层目标到 2030 年 EBITDA 达 A$1 billion,而本年度指引约 A$235 million。2025 年 12 月,NEXTDC 与 OpenAI 签署 MoU,计划在西悉尼 Eastern Creek 的 S7 站点共同开发 A$7 billion、550–650 MW AI 园区——把自身定位为 OpenAI 澳大利亚算力需求的主权载体。关键在于,NEXTDC CEO Craig Scroggie 在 2026 年 4 月公开质疑 Firmus 设施能否达到企业级质量,并反驳 Firmus 声称的建设成本约 US$6 million/MW;他引用的行业标准约为 US$12.4 million/MW。争议尚未解决:如果 Firmus 成本结构基本准确,它就是持久效率护城河;如果不准确,利润率逻辑和 IPO 估值都会面临实质风险。 CDC Data Centres 既是 Firmus 的主要基础设施伙伴,也是结构性竞争者。作为澳大利亚最大私营(非上市)数据中心运营商,CDC 为 Project Southgate 提供主权实体骨干。但 CDC 也独立服务 hyperscaler、政府和企业客户;其与 Firmus 的关系同时带来协同和依赖。如果 CDC 合作恶化——或 CDC 自建 GPU 云产品——Firmus 供给侧可能被扰动。 Macquarie Data Centres 正在悉尼建设 IC3 Super West 设施(47 MW,A$350 million Phase 1),计划 2026 Q3 开业,具备支持 AI 负载的液冷密度,可达每机架 150 kW+,并与 Dell 和 NVIDIA 合作在现场交付 Dell AI Factory 平台。其澳大利亚主权所有权和 Macquarie Group 背书,让它在受监管行业具备可信度。 AUCloud(AUCyber Limited,ASX 上市)持有 IRAP 认证,专注于面向澳大利亚政府和国防的主权 IaaS、MSSP 和受管 IT。其收入估计为数千万 AUD 低位,规模化 GPU 能力几乎可以忽略。它参与同一个主权话题,但不是 Firmus 规划规模下的 GPU 基础设施对手。[CP021, CP022, CP023, CP024, CP025, CP026]

功能 / 能力矩阵 — 关键采购标准
采购标准FirmusCoreWeaveLambda LabsNEXTDCMacquarie DCAUCloudAWS / Azure
澳大利亚 / 主权所有权是(澳大利亚多数持股)否(美国)否(美国)是(ASX 上市澳大利亚公司)是(澳大利亚)是(澳大利亚)否(美国跨国公司)
可再生能源 — 100% 承诺是(Project Southgate 承诺)部分(可持续发展目标)部分(未承诺)部分(PPA 谈判中)部分(范围未确认)Unknown部分(因地区而异)
裸金属 / 预留 GPU 云是(SLURM、RDMA、InfiniBand)是(Kubernetes 原生)是(按需 + 预留)借助 Firmus 合作 / NCP是(P5/ND H100 实例)
NVIDIA NCP 参考架构是(Elite NCP 合作伙伴)是(Elite NCP 合作伙伴)是(NCP 合作伙伴)部分(通过合作伙伴)部分(Dell AI Factory)是(AWS/Azure 符合 NCP)
IRAP 合规(澳大利亚政府)推进中 / 公司声称是(已认证)是(已认证)是(IRAP 已认证)是(部分服务)
公开发布 GPU 定价否(企业 / 定制)部分(部分标价)是(透明费率)否(定制合同)是(按需标价)
浸没式 / 芯片直触液冷是(核心架构)是(液冷)部分(部分设施)部分(较新设施)是(IC3 Super West)部分(因 DC 而异)
企业 SLA 和正常运行时间保障是(ISO 27001、SOC-2)是(企业 SLA)部分(偏初创公司)是(99.99% 正常运行时间,已认证)是(Macquarie Group 标准)是(MSSP 级 SLA)是(完整 SLA 框架)

矩阵单元格反映截至 2026 年 6 月可获得的最佳公开证据。「部分」表示能力有限、有条件或尚未确认。「未知」表示未找到公开证据。「推进中 / 公司声称」表示公司已表达意向,但仍等待独立验证。单元格不应被视为背书或正式认证。即使供应商列为已认证,IRAP 合规范围也会因服务线而异。

[CP001, CP002, CP005, CP024, CP025, CP026]
FP002: 竞争对手规模与运营成熟度——2026 年中快照

截至 2026 年中,Firmus 按运营容量看是最小的竞争对手(已确认运营约 5 MW),尽管已签约野心(155 MW)和估值都很高。CoreWeave 和 NEXTDC 的运营规模高出几个数量级,其中 NEXTDC 独特地把主权地位和澳大利亚规模的超大规模执行结合起来。

所有数字代表截至 2026 年 6 月可获得的最佳公开证据。Firmus 运营容量来自 Rask Media 2026 年 4 月报道(~5 MW vs 155 MW 已签约)。CoreWeave 数字来自 2026 年 Q1 业绩和 MLQ.ai 研究。Lambda Labs 收入来自 Sacra 估计。NEXTDC 收入按投资者数据估计;GPU 数量未披露。Macquarie DC Phase 1 按公司公告。AUCloud 收入按 AUCyber 投资者材料估计。

[CP008, CP021, CP014, CP019, CP026, CP027]

3.3 Hyperscaler、现状方案与主权云新进入者

AWS、Microsoft Azure 和 Google Cloud 合计主导全球和澳大利亚云支出。AWS 宣布在澳大利亚进行 A$20 billion、五年数据中心投资——主要围绕通用云和 AI 容量——并已在澳大利亚区域提供经 IRAP 评估的服务。Azure 和 GCP 在 Sydney 与 Melbourne 也有类似 IRAP 范围的足迹。Hyperscaler 按需 H100 定价明显高于 neocloud:AWS 约 US$6.88/hr,GCP 约 US$10.98/hr,Azure 约 US$12.29/hr,而 Lambda 为 US$2.49/hr。但 hyperscaler 把企业工具(AWS SageMaker、Azure AI、Vertex AI)、合规认证和数十年企业关系打包出售,纯 GPU 供应商很难复制。 Hyperscaler 轨迹构成 Firmus 最核心的长期商品化风险:随着 AWS、Azure 和 GCP 加速澳大利亚主权云基础设施,Firmus 用来区分自己的数据主权论点会变窄。Azure 和 GCP 均已拥有高级主权云层级;AWS 经 IRAP 评估的服务范围仍在扩张。Hyperscaler 进入不会消灭 Firmus 的能源效率主张,但会随着时间削弱主权角度。 现状替代方案——企业运行自有 on-premises GPU 集群,或通过既有 hyperscaler 协议采购——对拥有既有资本和合规框架的大型政府部门及受监管企业仍有意义。这些组织拥有对其有利的高转换成本(既有 IRAP 批准、成熟采购渠道),通常需要多年规划周期,才会迁移到新的 GPU 云供应商。 内部建设(on-prem)或与 Macquarie、CDC 做 co-location、但不叠加受管 GPU 层,对大型企业同样可行,尤其是那些不愿在缺少前期运营验证时接受五年 take-or-pay 承诺的客户。[CP035, CP036, CP037]

FP003: Firmus 竞争护城河——KPI 就绪度摘要

Firmus 的护城河在主权性、合同结构和 NVIDIA 合作上可信,但效率主张验证、运营不成熟以及供应商 / 客户集中度仍有关键未解问题——这些风险都是 IPO 尽调中的标准阻断点。

[CP002, CP009, CP028, CP037, CP008, CP006]

3.4 竞争差异化、转换动态与护城河耐久性

Firmus 的差异化论点建立在五项主张之上:(1)澳大利亚主权所有权与 100% 可再生能源;(2)同类最佳 AI Factory 效率(专有 FactoryOS、浸没 / 液冷、NVIDIA NCP 参考架构);(3)NVIDIA 同时是股权投资者和锚定 DGX Cloud 客户;(4)take-or-pay 长期合同(平均 4.8 年),总签约价值约 US$8.2 billion;(5)澳大利亚供应链带来比既有玩家更快交付和更低 capex。 独立验证还不完整。SemiAnalysis 2025 年 ClusterMAX 评级把 Firmus 的 Sustainable Metal Cloud(SMC)列入全球第一梯队——全球仅三家正确实现 InfiniBand SHARP in-network reduction 的供应商之一——这是可信的第三方技术背书。但 Firmus 标榜的 1.03 PUE 没有独立审计;一份投资者分析指出,这只是 2021 年的一次点测量,不是持续运营指标;一名 University of NSW 工程师称该数字处在全球可能性的极限边缘。NVIDIA 锚定关系是真实的,但并非排他:NVIDIA 同样是 CoreWeave 和 Lambda Labs 的股权投资者与技术伙伴,限制了这层护城河价值。 GPU 云转换成本低于传统企业软件。客户可以在多个供应商间 multi-home(Firmus AI Cloud 支持混合和多云网络),基于行业标准 Kubernetes、SLURM 和 InfiniBand 栈的负载也相对可迁移。锁定主要来自合同(take-or-pay 条款),不是技术。因此,Firmus 可持续竞争优势取决于其每 MW 成本结构和能源效率能否在生产规模反复证明——而不只是早期部署中成立。 Nomura 事件(因潜在投资者担心 GPU 出口管制,一笔面向 Firmus Metal Singapore 的 US$120 million 新加坡信贷额度难以分销)表明,Firmus 源自新加坡的业务存在地缘政治融资风险,澳大利亚主权叙事无法完全隔离。即将到来的 ASX IPO 还要求 Oliver Curtis——一名因 2007–08 年内幕交易被定罪的联合创始人——通过 ASX 的「良好声誉与品格」测试;ASX 在 2026 年 2 月正式化该测试的指引更新,被广泛解读为直接针对 Firmus 上市。[CP001, CP002, CP003, CP004, CP005, CP006]

护城河耐久性与竞争风险登记表
护城河主张主要威胁严重性缓释措施 / 尽调问题
NVIDIA Elite NCP 合作伙伴关系及 GPU 分配优先级NVIDIA 与 CoreWeave、Lambda Labs 也保持同等 NCP Elite 身份 — 并非独家;NVIDIA 持有多个竞争对手股权核验合同中 GPU 分配优先条款相对同行的差异;评估 Vera Rubin DSX 共同设计能否带来持久差异化
澳大利亚主权所有权与仅使用可再生电力超大规模云加快澳大利亚 IRAP 认证;AWS A$20B+;Azure/GCP 主权云层级扩张跟踪超大规模云 IRAP 范围扩张速度;评估 Firmus IRAP 认证时间表能否赶上政府采购截止日期
自有 AI FactoryOS 与浸没式冷却效率NEXTDC CEO 质疑 Firmus 建设质量和每 MW 成本;PUE 1.03 是 2021 年未经验证的单点测量;全球液冷技师池有限关键IPO 承销前,委托独立工程审计 Melbourne 和 Tasmania 场址的 PUE 及每 MW capex
合同价值 US$8.2B、平均期限 4.8 年的 take-or-pay 合同两个基石客户(NVIDIA 及一家未披露超大规模云)都不分散;客户集中度映射 CoreWeave 的 Microsoft 问题披露第二个基石客户身份和信用质量;要求 IPO 招股书列明合同付款瀑布和终止罚则
澳大利亚供应链(Benmax、Maas Group)以较低 capex 加快部署面向 GW 级雄心的供应链扩张尚未验证;~US$6M/MW 说法遭 NEXTDC 质疑(行业约 ~US$12.4M/MW);截至 2026 年中,仅 ~5 MW 投运,而已签约 155 MW独立核验规模化部署的每 MW capex;按 1.5 GW/yr 目标审计 Benmax 和 Maas Group 产能
SemiAnalysis ClusterMAX 顶级技术排名(2025)技术基准是时点数据;CoreWeave 和其他对手也位居前列;Blackwell/Vera Rubin 世代会重置基准GB300 部署投运后,重跑 ClusterMAX 或同等基准;跟踪竞争对手 GB300 集群性能公告

严重性评级是基于本章证据给出的编辑判断:关键 = 若被证伪,可能影响估值;高 = 重大尽调风险;中 = 重要但可管理。威胁和缓释措施仅反映公开可得信息;实际合同条款和工程规格未公开披露。

[CP002, CP003, CP006, CP008, CP009, CP024]

3.5 展示材料

Chapter 04

04财务

4.1 收入模型与商业牵引

Firmus 通过两条主要渠道销售 AI 算力容量并产生收入:一是公共云平台,提供按需 GPU 和 CPU 实例及受管存储;二是为专用 AI 基础设施定制的长期企业合同。公开定价页列出 CPU 实例价格:最小 4-vCPU 虚拟化实例为 USD $73/month,最高到 256-vCPU bare-metal 配置的 USD $4,672/month;WEKA NVMe 并行文件存储为 USD $0.11/GB/month。所有 GPU 实例层级——H200、H100、L40S、A100 和 A40——都列出规格但没有机架价,对承诺和专用访问标注「联系询价」。CPU 透明定价与 GPU 不透明定价之间的差距是商业上有意为之:GPU 合同代表经济上占主导的收入来源,且逐单谈判。 商业牵引最清晰的公开证据,是 Project Southgate 旗下两份数十亿美元 hyperscale 合同。第二份也是最近披露的一份于 2026 年 3 月 2 日签署,覆盖 Melbourne 设施约 18,400 块 Nvidia GB300 GPU,合同为与一家未具名全球领先科技公司的多年协议。该客户身份未在公司公告、Bloomberg 报道或任何已审阅公开来源中披露。因此,外部无法评估客户集中风险、交易对手信用质量或合同财务结构。 云按需与企业合同渠道之间的收入结构没有公开拆分。公司表示已经产生收入;Lennox Capital Partners 在 2025 年 10 月判断,近期 150 MW 管线支撑的 run-rate revenue 估值超过 Firmus 当时估值标记的 10x。Nvidia 的 DGX Cloud 受管 AI 平台正与 Firmus 合作,在 Southgate Melbourne 建立专用区域;这代表一个嵌入式需求渠道,混合企业与平台分发,但没有披露清晰的收入分成结构。长期 hyperscale 合同的收入质量在结构上优于现货云收入——合同代表经常性、已承诺现金流——但具体定价机制、最低承诺和取消条款不在公开记录中。[CI001, CI002, CI008, CI009, CI010, CI025]

收入流表
收入流机制单位当前价值 / 状态收入质量尽调问题
云 GPU IaaS(按需)客户从 Firmus AI Cloud 开通 GPU 实例;按小时或按月预留计费GPU-hour / instance-month 计费单位无公开标价;所有层级 GPU 定价均为「联系获取报价」低(无公开定价;现货敞口受市场费率影响)披露实际 GPU-hour 费率、利用率,以及按需与预留拆分
云 CPU 计算虚拟化和裸金属 AMD CPU 实例按月计费vCPU-month 或 metal instance-month$73/month(4 vCPU)至 $4,672/month(256 vCPU metal);公开列示中(标价透明;CPU 收入规模相对 GPU 可能较小)提供 CPU 与 GPU 渠道收入贡献
AI Cloud 存储WEKA NVMe 并行文件存储按 GB / 月计费GB-month$0.11/GB/month;公开列示中(定价透明;存储是计算合同的附属收入)披露按利用率加权后的有效费率和管理存储总量
长期企业与主权 AI Factory 合同在与超大规模云和政府客户签订的多年 take-or-pay 式协议下,部署专用 GPU 集群按合同;GPU cluster-month 或 AI-token output数十亿美元级;已确认两份合同;单价未披露高(多年承诺收入降低现货风险;合同条款不透明)披露匿名化合同期限、最低承诺和定价结构;确认是否存在取消条款

GPU 定价未发布在公开定价页;企业合同定价属于商业机密。收入质量评估基于合同结构推断,而非已披露利润率。长期合同的「当前价值」反映已披露合同规模($multi-billion),不是单价或年化费率。

[CI001, CI008, CI009, CI010, CI025]
定价与变现表
产品单位标价定价类型折扣 / 未知项来源
CPU 实例 c1.4.v(4 vCPU / 32GB / 100GB SSD)$73自助标价未列出折扣;批量定价未公开firmus.co/pricing(2026 年 6 月)
CPU 实例 c1.8.v(8 vCPU / 64GB / 200GB SSD)$146自助标价从 c1.4.v 线性扩展firmus.co/pricing(2026 年 6 月)
CPU 实例 c1.128.v(128 vCPU / 1TB)$2,336自助标价未列出折扣firmus.co/pricing(2026 年 6 月)
CPU 实例 c1.256.m metal(256 vCPU / 2TB)$4,672自助标价公开费率最高;未列出企业折扣firmus.co/pricing(2026 年 6 月)
GPU 实例(H200 SXM / H100 SXM / L40S / A100 / A40)月 / GPU-hour联系获取报价企业协商;无标价定制;可提供承诺或专用访问折扣firmus.co/pricing(2026 年 6 月)
存储 — WEKA NVMe 并行文件存储GB/month$0.11自助标价未发布分层批量定价firmus.co/pricing(2026 年 6 月)

所有价格均为 Firmus 公开定价页发布的 USD 标价,访问时间为 2026 年 6 月。GPU 定价完全未在公开页面披露。CPU 和存储标价不代表企业实际成交价;承诺或大批量部署的成交价可能更低。实际 GPU 定价属于私有商业条款,无法用于承销分析。

[CI002, CI034, CI035]
FI001: 收入模型桥——客户活动到 Firmus 收入

企业和云客户如何产生计算需求,再经 GPU 集群和 AI 工厂流入 Firmus 的合同收入。

收入确认方法(完工百分比 vs 月度交付)和实际每 GPU-hour 定价均未公开。该流程代表从公开公告和定价页面证据推断出的结构模型,不是已披露的财务机制。

[CI001, CI035, CI039]

4.2 资本结构、融资与 IPO 准备

Firmus 搭起的资本结构,相比当前经营阶段已经偏大。截至 2026 年 4 月的六个月里,公司完成三轮股权融资,合计约 USD $1.35 billion:2025 年末先以 USD $1.9 billion 估值融资 AUD $330 million,Nvidia 担任领投战略投资方;2025 年 11 月又以 USD $6 billion 估值融资 AUD $500 million;2026 年 4 月,在 Coatue Management 领投、Nvidia 继续参与下融资 USD $505 million,投后估值 USD $5.5 billion。2026 年 4 月这一轮被宣布为 IPO 前最后一轮股权融资。 结构上更关键的是 2026 年 2 月 9 日宣布的 USD $10 billion 优先担保债务融资额度,由 Blackstone Tactical Opportunities 和 Blackstone Credit & Insurance 管理的基金领投,Coatue 参与。顾问 Milbank、Allens、Kirkland & Ellis 将这笔交易描述为澳大利亚企业史上规模最大的私募债融资之一。该额度的核心结构——只有在与 hyperscaler 和蓝筹 AI 客户签署长期客户合同后才能提款——限制了投机性部署,也类似合同能源基础设施常见的项目融资结构。该设计给贷款人提供了真实的收入覆盖,但也意味着在拿到承诺前无法提款;如果合同谈判放慢或延迟,就会产生排序风险。 IPO 预计 2026 年 6 月或 7 月在 ASX 进行,目标募资约 USD $2 billion。账簿管理人为 Bank of America、JPMorgan、Morgans Financial 和 Morgan Stanley。若按目标规模完成,这将跻身澳大利亚史上最大的科技公司发行之一。Startup Daily 报道称,投资者正在按 AUD $8 billion 或更高 IPO 估值建模,较 2026 年 4 月轮次隐含的 AUD $7.95 billion 投前估值进一步抬升。到 2028 年的总 capex 计划为 AUD $73.3 billion,约为当前已融股本的 55 倍;因此,IPO 募资以及围绕新合同继续提款,是关键融资路径。月度 burn rate 和近期 runway 尚未公开披露。[CI003, CI004, CI005, CI006, CI007, CI011]

资本充足性表
项目金额 / 状态来源含义
已筹股权资金——截至 2026 年 4 月的六个月USD ~$1.35 billionFirmus 2026 年 4 月新闻稿;TNWIPO 前主要经营现金池,为运营和早期建设阶段提供近期现金续航
Blackstone 牵头债务额度(总承诺)USD $10 billionMilbank、Allens、Kirkland 顾问公告(2026 年 2 月)澳大利亚 AI 基础设施史上最大私人债务融资;建设资金的主要来源
债务提款条件与已签署的超大规模云厂商和蓝筹 AI 客户长期合同绑定Milbank 顾问公告防止投机性提款;资本部署跟合同收入对齐,降低贷款人风险
IPO 目标募资(ASX)USD 约 $2 billion(AUD 约 $2–3 billion,取决于汇率和估值上调)TNW; Startup Daily抬高总股权储备;为早期投资者释放流动性;在 AUD $8B+ 估值上测试公开市场胃口
月度现金消耗未披露任何公开来源均不可得没有该数字就无法评估现金续航;它是 IPO 前偿付能力判断的关键
到 2028 年的总资本开支计划AUD $73.3 billion(约 USD $48 billion,按 AUD/USD 0.65)Startup Daily; InnovationAus当前约 $1.35B 股权资本覆盖不到计划的 3%;剩余部分必须靠债务提款和 IPO 募资填上
下一轮触发点 / 资本充足性测试ASX IPO 目标为 2026 年 6 月或 7 月;截至 2026 年 4 月,非交易路演正在推进TNW; Startup DailyIPO 成功是近期资本充足性的核心测试;失败或延期会压紧运营现金续航

月度现金消耗和账面现金未公开披露。已筹股权资金代表公告时的总募资额;扣除费用和任何提款后的净现金未知。AUD/USD 按约 0.65 折算,仅用于指示性 USD 等值。该债务额度为已承诺额度,并非已全部提款;实际提款余额未知。

[CI003, CI004, CI005, CI006, CI007, CI011]
FI003: 财务估计区间——资本结构与估值输入

截至 2026 年 6 月,Firmus 关键资本输入、建设成本和 IPO 估值参数的有来源支撑区间。除非另有说明,所有金额均以 USD 十亿美元计。

为便于比较,AUD/USD 按约 0.65 折算。资本开支计划区间反映汇率不确定性,不代表来源材料披露了一个区间。IPO 估值区间由媒体和投资者报道推断得出,并非来自已提交招股书或正式指引。

[CI005, CI006, CI007, CI038]
FI004: 资本强度——股权和债务相对于建设成本

把资本结构中的正向流入与已承诺建设成本的流出放在一起,展示现有资本与完整 $73B 计划目标之间的缺口。

为便于比较,所有项目均以 USD 百万美元计。AUD 项目按约 AUD/USD 0.65 折算。债务融资额度为承诺总额,不是已提款余额。剩余资本开支按总计划减一期计算,数值为近似值。

[CI003, CI014, CI020, CI036]

4.3 成本结构、资本强度与供应链

Firmus 的 AI Factory 模型天生资本密集。公司称,新建项目资本强度约为每 MW $35-45 million,低于同类设施每 MW $50-55 million 的行业基准;它将这一优势归因于自研浸没式液冷技术、模块化建设和垂直供应链整合。公司还声称,相比传统风冷数据中心,建设成本可节省约 50%,同样来自冷却和模块化设计。已审阅资料中,按 MW 计算的 capex 说法和建设节省比例都没有得到独立验证。 成本结构最具体的公开证据来自 Maas Group(ASX: MGH)2026 年 5 月 18 日的 ASX 公告。文件披露,100 MW Launceston AI Factory 的 AUD $200 million 合同覆盖通过 JLE Group 交付的交钥匙模块化 PowerCube 方案和高压电气基础设施,意味着仅电气范围就达到每 MW AUD $2 million。JLE 是 Firmus 更大 3.3 GW 组合的独家电气交付伙伴,每增加一个 100 MW 模块,就对应额外 AUD $200 million 电气基础设施收入——这是一个透明成本基准,锚定了单站成本结构中的一个组成部分。 能源成本是重要 operating expense 输入。Firmus 的塔斯马尼亚设施依赖该州以水电为主的电网,可再生电力供应费率约为每 kWh USD $0.04-0.06,显著低于全球数据中心电力平均水平,是运营成本上的真实结构性优势。但塔斯马尼亚电网容量已经引发政治摩擦:当地人士担心电网承压,老牌工业客户过去也曾被拒绝追加容量。公司 1.03 的 PUE 效率说法已遭公开质疑:UNSW 的 Dr. Amr Omar 认为这处在全球可能性的极端边界,随后公司承认这只是 2021 年的单点测量,而非持续且经过验证的运营指标。领先 hyperscaler Google 和 Microsoft 通常报告 1.1 至 1.2 的 PUE。这个差异对运营成本建模很关键:真实 1.03 PUE 意味着颠覆性的能效;实际 1.15-1.20 PUE 依然优秀,但差异化会明显弱得多。[CI015, CI016, CI017, CI018, CI019, CI020]

单位经济性表
指标数值 / null置信度重要性尽调问题
毛利率null — 未披露不可用核心盈利能力指标;没有它无法评估收入质量要求提供 IPO 招股书中的经审计利润表;对标 CoreWeave 调整后 EBITDA(FY2025 为 56%)
每 MW CapEx(Firmus 声称)$35–45M / MW(公司声称)低(仅公司声称;未经独立审计)决定规模化后的资本回报和回本周期获取独立成本审计;注意 MGH 电气范围约 ~$2M/MW,说明仅电气部分就占相当比例
每 kWh 能源成本 — Tasmania 水电~$0.04–$0.06 / kWh(签约水电行业基准)中(行业基准;实际签约费率未披露)最大可变 OpEx 投入;相较全球数据中心电网平均电价 USD $0.08–$0.12 更有利向 Hydro Tasmania 供电协议或 IPO 招股书经营成本披露索取实际合同电价
GPU 机架回本周期(CoreWeave 可比)领先运营商为 24–36 个月(由 CoreWeave 推断)中(基于可比公司推断;Firmus 未披露回本周期)决定每个已部署机架何时转正自由现金流向 Firmus 索取目标利用率下的模型化回本周期;确认 GPU 代际更新周期假设
每 1,000 块 GPU 年化收入null — 未披露不可得可用于对已公布的 18,400 块和 36,800 块 GPU 集群建立收入和利润率模型向 IPO 招股书索取;可比 GPU 云 H100 市场价格为 USD $2–$4/GPU-hour(行业基准)
净利润率(调整后 EBITDA 减折旧和利息)null — CoreWeave 可比:建设扩张期接近零到为负低(仅基于可比公司推断)真正的股权层面回报取决于净利润率,不是表面 EBITDA招股书提交后,结合 Firmus 债务条款、GPU 资本开支和合同收入测算净利润率

所有 null 值都代表真实不可得的私有数据,不是研究遗漏。CoreWeave 数字只是用于划定合理区间的上市公司可比项,不是 Firmus 披露。CapEx/MW 和能源成本估算分别基于公司说法和行业基准;两者都未就 Firmus 独立核验。

[CI021, CI022, CI024, CI025, CI026, CI027]
FI002: 单位经济性桥——已知输入与关键未知

Firmus AI factory 单位经济性的关键成本和收入输入,展示公开已知、按可比公司估算以及仍不可得的信息。

所有数字均来自公司主张、行业基准和 CoreWeave 可比数据估算。Firmus 未公开披露专属毛利率、利率、折旧年限或签约收入数据。约 ~$320M 的年化 GPU 集群收入估计仅作示意,基于 CoreWeave 时代的 H100 市场费率,未必适用于 GB300 合同。

[CI040, CI041, CI042]

4.4 GTM 动作与销售效率代理指标

Firmus 采用双轨 go-to-market 模型。云渠道是直接自助式销售,公开价格页可看到透明 CPU 和存储定价,面向需要按需算力、又不想经历漫长采购周期的 AI 开发者、研究人员、初创公司和企业团队。企业与主权渠道则依赖关系和合同先行:专用 AI 基础设施容量通过多年协议销售,直接与 hyperscaler、大型企业、教育机构和政府客户谈判。Southgate Melbourne 的 DGX Cloud 区域部署引入第三层分发——Nvidia 的托管 AI 平台向 Nvidia 企业客户群转售 Firmus 基础设施访问权,从而模糊了直接销售与平台中介销售之间的边界。 传统销售效率指标——获客成本、回本周期、净收入留存、按分部划分的平均合同价值——都没有公开披露。最接近的代理指标是 Lennox Capital Partners 2025 年 10 月的分析师评估:该评估用可比 GPU-cloud 运营商的运营利润倍数作为基准,将近期 150 MW 管线估值为 Firmus 当时估值 10 倍以上的潜在规模。第二个代理指标是已披露合同规模:一个数十亿美元、跨多年、覆盖 18,400 块 GPU 的企业合同,意味着平均合同价值在数亿美元到低数十亿美元之间,符合 hyperscale AI 基础设施市场特征,但 CAC 被压缩成几乎一次性大单经济,而不是系统化销售漏斗。 考虑到 Firmus 是澳大利亚本土运营商,主权与企业焦点在战略上说得通:亚太地区需要数据驻留和低碳算力的政府与企业客户,无法轻易用美国 hyperscaler 容量替代。澳大利亚和新加坡双焦点,让 Firmus 进入该地区两个最活跃的主权 AI 基础设施市场。由于没有披露 CAC 和回本数据,外部无法判断企业销售动作是否资本高效,也无法判断除两份已宣布合同外,管线是否耐久。[CI001, CI025, CI039, CI043]

4.5 财务结论与尽调缺口

Firmus 的正面财务案例建立在三个可观察支柱上:一笔以合同收入为提款条件、结构健全的债务额度;至少两项数十亿美元级 hyperscale 承诺,在 IPO 前提供可见收入基础;以及可再生能源成本优势——若能规模化兑现,运营经济性会显著好于传统风冷竞争者。Blackstone 额度的合同提款条件类似项目融资结构,后者在能源基础设施领域过去能吸引投资级债务评级;Maas Group 的 ASX 文件也以 AUD $200 million 电气范围、35% 完工进度,独立佐证了实际建设进展。 负面案例同样证据充分。已审阅的公开来源中,Firmus 没有披露收入、毛利率、burn rate 或 runway。最接近的上市可比公司 CoreWeave 报告 FY2025 调整后 EBITDA 为 56%,但在 GPU 折旧及约 USD $21 billion 债务利息后净亏损 USD $1.17 billion——这说明 AI 云基础设施的表观 EBITDA 利润率在排除设备折旧和债务服务后,结构上会误导判断。Firmus 一旦开始规模化部署,债务结构也会带来类似动态;其 IPO 隐含 AUD $8-12 billion 估值,等于要求公开市场投资者在缺乏公开市场通常要求的信息披露深度时,支付私募市场倍数。 治理风险同样重大。Oliver Curtis 2016 年内幕交易定罪与 ASX 上市要求直接相关;ASX 2026 年 2 月关于 Good Fame and Character 测试的合规指引,也被金融媒体广泛解读为与 Firmus 上市有关。Roger Montgomery Investment Management 已公开指出,IPO 缺少 lock-up 或 escrow 安排是结构性红旗,并称早期投资者已实现 300% 估值涨幅,经济上有动力在上市首日退出。合并来看,财务结论是:从需求基本面和资本结构设计看,Firmus 论点并非不合理;但极端资本强度、未披露运营经济性、客户暴露集中以及治理悬而未决叠加在一起,意味着在无法取得管理账、IPO 招股书财务数据和合同条款清单前,本章无法支持传统财务承销结论。[CI026, CI027, CI028, CI029, CI031, CI032]

公开财务缺口表
缺失的私有指标对承销判断的影响精确尽调路径
收入(总额、按产品线、按地区)没有披露数字,就无法判断收入规模、增长率或质量;投资逻辑只能完全压在已签合同和销售管线数据上IPO 招股书(提交 ASIC)可得后从中索取;也可在尽调中要求提供管理账
按收入流拆分的毛利率没有利润率数据,就无法判断合同收入是否覆盖能源、人力和偿债等运营成本审阅 IPO 招股书财务报表;以 CoreWeave(调整后 EBITDA 56%)和 Oracle AI 基础设施利润率作参照
现金消耗和现金续航不清楚月度现金消耗,就无法评估近期偿付能力或 IPO 时间压力索取最新管理层现金流量表和 12 个月现金预测;再用员工人数和设施运营成本交叉核对
超大规模云合同定价机制和取消条款$10 billion 债务额度以已签客户合同为条件;如果合同可取消或重新定价,债务结构就会松脱尽调中索取匿名化合同条款清单;审阅 IPO 招股书重大合同披露;评估最低承诺下限和终止权
GPU 折旧计划和利用率目标AI 工厂的股权经济性关键取决于 GPU 按多快速度相对合同收入折旧、利用率跑到什么水平;两项都未披露向管理层索取资本开支计划、GPU 折旧年限假设和目标利用率;对照 CoreWeave 公开的资产层面披露

五项都是真正属于 IPO 前公司的私有指标。严重程度一律为重大且会阻断判断,因为它们合在一起,才构成对 Firmus 模型做基础财务承销判断所需的输入。预计在 2026 年 6/7 月上市前提交 ASIC 的 IPO 招股书,应能回应大部分缺口。

[CI025, CI026, CI027, CI028, CI029]

4.6 图表证据

Chapter 05

05产品与技术

5.1 产品组合与运营模型

Firmus 运营两条相互嵌套的产品线。主产品是 AI Factory——一个模块化、垂直整合的算力园区,以 HyperCube™ 硬件模块为核心。每个 HyperCube 采用液冷高密度配置,容纳 32 个 NVL72 GPU 机架,专为 AI 训练和推理设计。AI Factory 不是 colocation 服务:每个站点都是垂直工程化资产,算力、冷却、网络和电力基础设施从第一性原理共同设计;关键计算区没有 retrofit tanks,也不需要 airflow。 截至 2026 年 6 月,Firmus 在 STT GDC 合作下运营新加坡 AI Factory 站点(约 4,000 块 NVIDIA GPU、99 PFLOP/s,支持 AI Singapore 的 SEA-LION 项目),并在塔斯马尼亚 Launceston 的 Project Southgate 推进 Phase 1B 部署,目标是 36,800 块 NVIDIA GB300 Grace Blackwell GPU。早期边缘产品 Cube300 以 shipping-container 形态提供 300 kW,PUE < 1.05,占地仅为同等风冷容量的十二分之一。 第二条产品线是 Firmus AI Cloud,这是一个 GPU infrastructure-as-a-service 平台,提供 bare-metal 和 virtual-machine GPU 实例(H200、H100、A100、L40S)、RDMA 加速 InfiniBand 网络、托管 Slurm 调度以及预配置 CUDA 技术栈(PyTorch、TensorFlow)。AI Cloud 运行在自有 AI Factory 基础设施上,而非租用第三方数据中心,从而形成垂直整合 GPU 云。Firmus 还运营 Synert,这是一家电网整合子公司,向澳大利亚 National Electricity Market 提供 frequency control ancillary services(FCAS),把闲置 UPS 电池容量转化为电网稳定收入,同时将 AI Factory 能源成本降低超过 25%。 [CE001, CE002, CE003, CE031, CE032, CE033]

Firmus 产品模块和资产矩阵
模块 / 产品用户分层状态 / 成熟度差异化尽调缺口
HyperCube™(模块化硬件单元)企业 / 主权 AI 运营商生产——1B 阶段部署(2026 年 4 月);塔斯马尼亚 36,800 块 GPU每单元 32 个 NVL72 机架;液冷;无气流;可模块化部署独立 PUE 审计;浸没式维护可扩展性
AI FactoryOS™(编排软件)AI Factory 站点运营商生产——已部署于新加坡和塔斯马尼亚站点从芯片到电网统一控制;FCAS 集成;对齐 NVIDIA DSX未发布外部技术架构审计;SLA 条款未公开
Firmus AI Cloud(GPU IaaS)产品AI 开发者、研究人员、企业生产——H200、H100、A100、L40S 实例已上线裸金属 RDMA/InfiniBand;Slurm;液冷效率GPU 定价不透明(需联系询价);供应受 Nvidia 配额约束
Synert(FCAS 电网集成)电网运营商;Firmus AI Factory 站点生产——生产站点已取得 AEMO FCAS 实时认证UPS 入网收入;能源成本降低 25%+;Eaton 合作FCAS 收入规模未披露;Firmus 之外采用情况不清楚
AI Storage(WEKA NVMe PFS + VAST Data AI OS)产品云租户、AI 工作负载运营商生产——WEKA 已部署;VAST OS 于 2026 年 2 月选定RDMA 加速 NVMe;对齐 NVIDIA NCP 参考设计VAST Data 合同条款未公开;吞吐基准未披露
Project Southgate(全国 AI 园区网络)主权 AI / 企业超大规模云客户1B 阶段已运营(2026 年 4 月);到 2028 年全国扩张1.6 GW 可再生能源供电;NVIDIA DSX;与 CDC/Nvidia 共同开发A$73.3B 规模下存在时间表风险;Nvidia GB300 交付风险
Bernacchi-1 海底光纤电缆AI Factory 运营商;塔斯马尼亚经济2026 年 6 月宣布;运营目标为 2027 年 Q2(取决于许可)>60 Tbps;首条 Tasmania–NSW 直连;SUBCO SMAP 骨干海事许可延迟;SMAP 上线前单电缆集中风险

状态评估基于截至 2026-06-18 的公司公告、合作伙伴新闻稿和第三方新闻报道。Project Southgate 容量数字反映公司声明目标;只有新加坡部署的 GPU 数量(约 4,000 块)有独立核验。Bernacchi-1 的运营状态取决于许可最终落地。

[CE001, CE002, CE003, CE010, CE020, CE024]
Firmus 工作流和用例表
用户任务当前工作流(不用 Firmus)Firmus 方案可衡量收益限制
训练大语言模型从美国超大规模云厂商租用风冷 H100 集群;能源成本高Firmus AI Cloud 上带 InfiniBand 的 H200 / GB300 裸金属集群每计算单元电力效率提升约 30%(MLPerf V4.0)GPU 可得性受 Nvidia 配额限制;GPU 定价不透明
部署主权 AI 工作负载使用离岸云(数据主权、监管风险)澳大利亚 / 新加坡的 Firmus AI Factory,配本国数据控制监管合规;本地数据驻留;可再生能源资质仅限 APAC 地区;尚无欧盟 / 美国主权 AI Factory
电网频率稳定(能源公用事业)数据中心是被动、巨大的扰动型电网负荷AI Factory 站点通过 Eaton EnergyAware UPS 提供 Synert FCAS 服务已认证可对 AEMO 实时电网作快速频率响应;产生电网服务收入FCAS 收入规模未披露;仅限接入 NEM 的站点
大规模 AI 推理部署训练和推理集群分离;存储卡脖子Firmus AI Cloud 裸金属 + Slurm + RDMA NVMe 存储GPU 到存储低延迟;CUDA 预配置;混合连接没有公开推理吞吐基准

收益基于公司声称的指标,以及经 MLCommons 审阅的 MLPerf 功耗数据(仅节点层面;设施层 PUE 不在 MLCommons 范围内)。主权 AI 合规说法基于公司定位;关于数据驻留控制的独立监管评估未公开可得。

[CE003, CE009, CE011, CE015]
FE001: Firmus AI Factory 产品架构栈

七层架构从电网 / 能源一直到云交付,展示 Firmus 垂直整合的「silicon-to-grid」设计原则。

分层边界依据 Firmus 官方技术页面和合作伙伴公告中的架构描述;FactoryOS 内部子组件拆解属专有信息,尚未获得独立验证。

[CE001, CE002, CE005, CE006, CE007, CE008]

5.2 技术架构与软件栈

核心计算层建立在 NVIDIA GPU 硬件上:当前云部署使用 H200 SXM(每节点 8 × 141 GB HBM3e,NVLink 4.0 + NVSwitch 3.0,每 GPU 双向 900 GB/s);Project Southgate 规模化阶段计划采用 NVIDIA GB300 Blackwell Ultra(288 GB HBM3e,最高 8 TB/s 带宽,每 NVL72 机架 TDP >120 kW)。节点间网络使用每节点 200–800 Gb/s 双 InfiniBand,支持 RDMA 和 RoCE v2,使多节点 GPU 集群能够进行低延迟分布式训练。 冷却采用 direct-to-chip liquid cooling(CDU 供液冷板),关键计算区没有 airflow。公司称生产部署 PUE 约 1.1(工程原则页面和 MLPerf 提交中引用),同时单独营销 1.03 的 headline figure;两项数据都没有发布独立第三方年度审计。经 MLCommons 审阅的 MLPerf Training V4.0 提交显示,在 immersion-rack 层面,相比 H100 SXM 风冷系统,电力效率约高 30%;但 data-centre-level PUE 不在该轮范围内。 编排层是 FactoryOS™,Firmus 自研的全栈基础设施智能平台,把冷却、供电、热遥测和 GPU 性能作为单一系统治理。数据层是 VAST Data AI OS,2026 年 2 月被选中,用于提供统一高吞吐 NVMe 存储,并与 NVIDIA Cloud Partner(NCP)参考设计对齐;WEKA NVMe PFS 保留用于 AI Storage,提供 RDMA 加速并行文件访问。Model-to-Grid 平台把 FactoryOS 编排延伸到对澳大利亚电网价格和稳定性信号的实时响应,协调 AI 工作负载与 AEMO 及 National Electricity Market。Firmus 与 NVIDIA 基于 DSX Blueprint 合作建设这一能力,并已在 Project Southgate 初始部署。 [CE004, CE005, CE006, CE007, CE008, CE009]

Firmus 技术和运营架构
层级 / 组件作用依赖风险
NVIDIA GB300 / H200 GPU核心 AI 计算引擎(训练、推理、微调)单一供应商(NVIDIA);TSMC CoWoS-L 先进封装如果 NVIDIA 将配额重分给美国超大规模云厂商,供应会受限
NVLink 4.0 / NVSwitch 3.0节点内 GPU 互连(900 GB/s 双向)NVIDIA IP;与 HGX / NVL72 服务器板共同设计与 NVIDIA 架构路线图深度绑定;没有开放替代方案
InfiniBand 200–800 Gb/s 网络分布式训练用节点间低延迟网络NVIDIA Quantum 网络栈;ConnectX NIC成本高;APAC 的 InfiniBand 运营专家池有限
CDU 液冷(直达芯片)管理 >120 kW/rack GPU 集群的热量专有 HyperCube CDU 设计;闭环流体供应浸没式维护技能;介电流体供应链
FactoryOS™统一编排:冷却、电力、GPU 遥测、FCAS专有(Firmus IP);NVIDIA DSX API无外部审计;设施管理被单一厂商 OS 锁定
VAST Data AI OS训练、推理、检查点保存共用的统一 NVMe 数据层VAST Data 合作关系;NVIDIA NCP 参考设计依赖第三方供应商;合同条款未公开
Synert / Eaton EnergyAware UPS电网 FCAS:AI Factory 与 NEM 之间的频率缓冲Eaton UPS 硬件;AEMO 市场参与规则电池衰减;Eaton 供应链;仅限 NEM 市场
Aurora Energy / Hydro Tasmania RSA 电力协议可再生电力供应(104 MW 至最高 400 MW)三年 Aurora RSA;Hydro Tasmania 发电合同续签风险;Tasmania 工业需求竞争
SUBCO SMAP / Bernacchi-1 fiber 光纤海底连接:塔斯马尼亚至悉尼 / 墨尔本(>60 Tbps)SUBCO SMAP 电缆;海事和登陆许可许可延迟;到 2027 年 Q2 前存在运营缺口;单电缆风险

依赖和风险评估基于截至 2026-06-18 公开可得的技术文档、合作伙伴公告和新闻报道。NVIDIA 组件规格来自 NVIDIA 官方产品页。Aurora Energy 合同条款在联合新闻稿中披露;单价因商业机密未披露。FactoryOS 内部架构尚未经过独立审计。

[CE004, CE005, CE006, CE007, CE008, CE009]
FE002: Firmus 客户工作流——从 GPU 计算到 AI 输出

端到端客户工作流覆盖从 GPU 配置到模型产物交付,并纳入 Synert 电网参与闭环。

工作流基于官方产品页描述和合作伙伴公告;FactoryOS、Slurm 与 VAST Data OS 之间的具体 API 交互属专有信息,公开资料没有独立记录。

[CE005, CE009, CE011, CE013, CE014]

5.3 关键依赖、供应链与连接性

Firmus 的基础设施栈建立在三项底层依赖上,每一项都有实质集中风险。第一,NVIDIA GPU 供应:每个 HyperCube 模块、每个 AI Cloud 实例以及 Project Southgate 所有 GPU 机架,都只依赖 NVIDIA 硬件。截至 2025 年 Q4,NVIDIA 捕获了 top-5 hyperscaler AI 资本开支约 43.6%,意味着大型 hyperscaler 会与 Firmus 争夺受限芯片供应。NVIDIA 对 Firmus 的股权投资让激励更一致,但不保证优先于更大的美国客户获得分配。GB300 Blackwell Ultra 的供应链与 TSMC 封装约束,也给 Project Southgate 时间表增加交付风险。 第二,能源供应:Firmus 持有一份三年期 Aurora Energy 零售服务协议,可获得最高 104 MW 的 Hydro Tasmania 可再生电力(2026 年 8 月至 11 月 ramping)。三个计划中的塔斯马尼亚站点合计约需 400 MW——约为塔斯马尼亚能源需求的 20%——这将使 Firmus 成为该州潜在最大电力用户。政治审视已经加剧:Boyer Paper Mill 追加 45 MW Hydro 电力的请求被拒,而 Firmus 获得 104 MW 分配;Greens 议员也呼吁专门监管。塔斯马尼亚寒冷气候和水电可靠性是 Firmus 效率说法的核心,短期内该地点不可替代。 第三,数据连接:塔斯马尼亚隔 Bass Strait 的地理隔离是战略瓶颈。Firmus 2026 年 6 月与 SUBCO 宣布 Bernacchi-1 海底光缆,将通过接入 SUBCO 的 400 Tbps SMAP 系统并配置双大陆路径,向塔斯马尼亚提供超过 60 Tbps——超过现有所有 Bass Strait 光缆总和。商业运营目标是 2027 年 Q2;在此之前,Project Southgate 依靠 legacy fiber capacity 运行。全国层面,Project Southgate 目标到 2028 年在澳大利亚五座城市达到 1.6 GW,宣布建设成本为 A$73.3 billion。 [CE010, CE011, CE012, CE020, CE021, CE022]

Firmus 产品路线图和部署阶段
日期 / 阶段里程碑 / 功能状态含义来源
2019–2021Firmus 成立;为比特币挖矿开发液冷 HyperCube;早期塔斯马尼亚 20 MW 站点历史核心热管理 IP 和模块化集装箱形态在此期间成型datacenterdynamics.com(经 Wayback)
2023 年 H2STT GDC Singapore 上线(SIN01):基于 Kubernetes/OpenStack 的浸没式冷却 H100/A100 云已运营首个生产级 AI Cloud 部署;约 4,000 块 GPU;服务 AI Singapore 的 SEA-LION 项目datacenterdynamics.com(经 Wayback)
2026 年 2 月VAST Data AI OS 被选为基础数据层;确认对齐 NVIDIA NCP 参考设计已签约 / 部署中统一 PB 级数据层支撑多租户主权 AI 工作负载来源:vastdata.com / computeforecast.com
2026 年 3 月Firmus 和 NVIDIA 宣布基于 NVIDIA DSX Blueprint 的 Model-to-Grid FactoryOS 软件层已宣布 / 开发中作为 NVIDIA DSX Blueprint AI 工厂的运营软件标准;协调 AEMO来源:firmus.co/newsroom;nvidianews.nvidia.com
2026 年 4 月Project Southgate 1B 阶段已运营(首批;朗塞斯顿 36,800 块 GB300 GPU)已运营——1B 阶段部署18,400 块 GPU 的锚定超大规模云厂商合同已上线;开始产生商业收入来源:dcpulse.com;firmus.co
2026 年末Project Southgate 1b 阶段完整建成(塔斯马尼亚 90 MW)目标完成塔斯马尼亚基础底座;触发全国铺开融资提款来源:thenextweb.com;firmus.co
Q2 2027Bernacchi-1 海底光纤电缆投运(>60 Tbps,塔斯马尼亚至澳大利亚大陆)已签约 / 建设中(取决于许可)打通 Bass Strait 连接瓶颈;支持低延迟 AI token 输出来源:firmus.co/newsroom;sub.co
2028Project Southgate 全国版:墨尔本、悉尼、珀斯、堪培拉合计 1.6 GW;总成本 A$73.3B已宣布目标扩展成全国主权 AI 基础设施平台;全线采用 Vera Rubin DSX来源:firmus.co;nvidianews.nvidia.com

日期基于公司公告和合作伙伴新闻稿;2026 年 4 月之后的里程碑是公司目标,取决于监管审批、能源供应协议、建设时间表和 NVIDIA GB300 芯片交付。A$73.3B 项目成本是已声明的总投资额,不是已确认的合同支出。

[CE020, CE021, CE024, CE031, CE034]
FE003: Firmus 关键依赖图

八节点有向无环图展示 Firmus AI Factory 的关键基础设施和供应链依赖,以及主要风险向量。

依赖边代表已披露的合作和合同;每项依赖的相对关键性仅依据公开证据评估。NVIDIA 对 Firmus 与超大规模云厂商的供应分配优先级并未公开披露。

[CE010, CE022, CE024, CE027, CE028, CE029]

5.4 差异化、信任与合规

Firmus 的技术差异化建立在三根支柱上:分层专利组合、独立基准测试表现,以及云安全认证。IP 方面,专利律师事务所 FB Rice 确认,Firmus 在澳大利亚、美国及其他司法辖区持有已授权专利,覆盖核心液冷系统、电网稳定软件,以及管道设计、温控和 HyperCube 软件中的具体创新。分案申请形成分层 “patent thicket”,保护基础技术和渐进改进,在授权许可中提供杠杆,并阻止直接复制。 性能基准方面,Firmus 向 MLPerf Training V4.0 提交了数据,电力测量在 immersion-rack 层面并经 MLCommons 审阅,显示相较 H100 SXM 风冷等价系统,能效约高 30%。公司称端到端设施 PUE 估计约 1.1,并指出 data-centre-level PUE 不在该轮 MLCommons 范围内。尚无 Uptime Institute 或同等第三方年度审计发布。数据中心专家指出,持续低于 1.05 的年度 PUE 在技术上非常罕见,通常只在最佳条件下的单站点测得。 Firmus AI Cloud 服务声称具备 ISO 27001 和 SOC-2 合规,并提供传输中和静态加密。公司没有发布公开证书注册链接;覆盖范围是 AI Factory 物理基础设施还是仅限云服务并不清楚。AEMO 的一项 FCAS 认证覆盖 Synert 在现有生产站点参与电网。还存在名称空间风险:Firmus Pte Ltd(firmussec.com)是另一家新加坡网络安全公司,持有 CREST 认证——尽调必须确认哪一个实体持有哪些合规凭证。 [CE015, CE016, CE017, CE018, CE019, CE026]

Firmus 信任、质量和合规控制
控制 / 认证状态范围缺口
ISO 27001:2022已认证(公司声称)Firmus AI Cloud 服务(云计算、存储、托管服务)没有公开证书注册链接;是否覆盖 AI Factory 物理层不清楚
SOC-2合规(公司声称)云服务数据安全、可用性、保密性Type II 报告未公开披露;审计期间未说明
MLPerf Training V4.0 Power已提交并经 MLCommons 审阅LLaMA2 70B-LoRA、BERT、GPT3 等的节点级能效数据中心层 PUE 不在 MLCommons 范围内;非 MLCommons PUE 估算未经审计
FCAS / AEMO Grid Certification生产——一个站点已取得实时快速频率响应认证AEMO NEM 规则下 Firmus AI Factory(Tasmania)的 Synert FCAS 服务范围仅限现有认证站点;Bell Bay/Wesley Vale 尚未接入
独立 PUE 年度审计(Uptime Institute)未公开可得声称 AI Factory 站点 PUE 区间为 1.03–1.1重大缺口:未发布 Uptime Institute、Tier III 设计认证或同等第三方核验
CREST 渗透测试Firmus Technologies(firmus.co)未确认云安全验证CREST 资质属于 Firmus Pte Ltd(firmussec.com)——这是一家独立网络安全公司;尽调中存在混淆风险

所有合规说法都基于 Firmus 自有公开云服务页面和新闻室;截至 2026-06-18,尚未核验到独立注册机构确认。ISO 27001 和 SOC-2 声明适用于云 IaaS 产品,不一定覆盖 AI Factory 物理运营或 Synert 电网服务。

[CE015, CE016, CE017, CE026, CE030]
FE004: Firmus 产品能力成熟度矩阵

从八个维度评估 Firmus 核心技术能力的成熟度,同时给出证据质量评级和扩展至其宣称 1.6 GW 目标的可扩展性判断。

成熟度和证据质量评级是分析师基于截至 2026-06-18 的公开证据作出的评估。可扩展性判断对应公司宣称的 2028 年 1.6 GW 目标。FactoryOS 或 Synert 软件组件尚未发布独立第三方能力评估。

[CE005, CE006, CE007, CE010, CE015, CE016]

5.5 产品风险与技术风险登记

五项重大产品和技术风险需要明确尽调。第一,NVIDIA 集中:每个 AI Factory 模块和 AI Cloud 实例都是单一供应商 GPU 环境。如果 NVIDIA 将供应重新分配给更大的美国 hyperscaler、延迟 GB300 交付,或修改合作伙伴计划条款,Firmus 的扩张时间表和单位经济性都会被直接影响。NVIDIA 股权投资把 Firmus 的资本部署绑定到 NVIDIA 自身产品路线图周期。 第二,未经验证的 PUE 效率说法:Firmus 在部分材料中营销 1.03 PUE,而工程页面和 MLPerf 提交引用约 1.1。采用先进冷却的 hyperscale 运营商通常年度报告 1.1–1.15。没有独立年度审计,投资者或客户无法验证其相对竞争者的效率优势幅度。 第三,浸没冷却运营复杂度:direct-to-chip 冷板冷却已经成熟,但吉瓦级全浸没需要专门的介电液管理、泄漏检测、泵维护,以及熟悉浸没系统的劳动力。塔斯马尼亚可用且具备该规模经验的技术人员池有限。 第四,能源供应监管风险:Aurora Energy 合同是一份三年期 RSA。Firmus 在三个塔斯马尼亚站点部署 400 MW 的计划面临日益增强的政治反对,议员也在积极提出州级 AI 数据中心能源监管。 第五,建设与执行风险:A$73.3 billion 全国建设在澳大利亚没有先例;五个司法辖区同时多城市建设、供应链依赖和监管审批,会叠加执行风险。 [CE020, CE021, CE022, CE023, CE027, CE028]

5.6 图表证据

Chapter 06

06客户

6.1 客户基础分层与买方原型

截至 2026 年 6 月,Firmus 服务五个清晰客户分部,采购结构、交易规模和采用阶段差异很大。第一也是商业上占主导的分部,是全球 hyperscale 云和 AI 公司,它们以数十亿美元规模采购多年、take-or-pay 的 AI factory 容量。这些买方的动机来自本土市场的供给侧约束,也来自主权或 ESG 要求推动的本国可再生能源算力需求。截至 2026 年 6 月,Firmus 已公开披露 Project Southgate 下两份已签 hyperscale 合同,均来自未具名的 “leading global technology companies”;该分部是 Blackstone 债务额度的锚点,而该额度以与 hyperscaler 和蓝筹 AI 客户签署长期客户合同为条件。 第二个分部是政府和公共部门研究机构,代表是 AI Singapore(AISG),这是由 Singapore National Research Foundation 支持的国家 AI 计划。AISG 是迄今 Firmus 唯一公开具名的重要客户,自 2023 年以来使用 Sustainable Metal Cloud(SMC)平台进行生产级 LLM 训练。第三个分部是受监管企业——银行、保险公司、电信运营商和能源公司,它们寻求具备数据驻留保证的主权 AI 算力——Firmus 通过 Firmus AI Cloud 向其营销,但在澳大利亚尚未披露具名生产客户。第四个分部是可通过 Firmus AI Cloud 按需平台和 NVIDIA DGX Cloud Lepton marketplace 触达的 AI-native 初创公司和 scaleups。第五个分部是学术和研究机构,包括 Project Southgate 材料中明确点名的教育部门用户;除 AISG 外,尚未宣布具体部署。 这个组合反映了 Firmus 当前阶段:hyperscale 锚定合同主导已承诺收入,而企业、政府和初创分部在澳大利亚仍停留在愿景或 pilot-stage。主权 AI 定位是有意为之的差异化,瞄准公共部门和受监管工作负载的本国部署;数据驻留要求和政府指令形成结构性需求,离岸 hyperscaler 云无法满足。该分层策略与 Project Southgate 的全国雄心匹配,但短期内带来重大集中风险,也引发目标 ASX IPO 前多元化速度的问题。 [CU001, CU016, CU017, CU021, CU022, CU027]

客户分层地图——买方画像、用例和采用阶段
分层买方 / 付款方主要用例合同结构已披露收入 / 战略价值采用阶段(2026 年 6 月)
超大规模云 / AI全球云 / AI 基础设施资本委员会大规模 LLM 训练和推理;AI 工厂容量多年照付不议,数十亿美元级主导;已确认两份合同(各 $B+,客户未具名)活跃——合同已签;基础设施建设中
政府 / 主权研究国家 AI 计划(如 AISG);政府部委主权 LLM 训练;受监管 AI 工作负载MoU 加服务协议;多年期生产——AISG SEA-LION(256 块 H200 GPU,99 PFLOP/s,新加坡)生产(新加坡);愿景阶段(澳大利亚)
受监管企业银行、保险公司、电信公司、能源公司的 CIO / CTO模型微调;风险模型;欺诈检测;私有云 AIMSA 加承诺使用层级;数据驻留附录未披露;没有具名澳大利亚企业客户已投入生产愿景阶段——没有具名生产客户
AI 原生初创 / 扩张期公司创始人 / CTO(Series A–C)模型训练;推理 API 服务;数据处理通过 Firmus AI Cloud / DGX Cloud Lepton 使用按需或预留实例未披露;可通过公开定价访问(CPU 价格已发布;GPU 需“联系询价”)可访问——DGX Cloud Lepton 市场于 2025 年 6 月上线
学术和教育大学 IT 或研究负责人AI 研究;LLM 实验;基准测试研究协议或云订阅未披露;AISG 合作覆盖该分层(NRF 支持)早期——AISG 合作是唯一具名案例

分层分类基于 Firmus 官方材料、Project Southgate 公告和市场分析章节语境。收入和战略价值为定性判断。采用阶段反映截至 2026 年 6 月公开披露的证据;未公告客户可能存在。

[CU001, CU016, CU021, CU022, CU027]
FU003: 客户旅程图——细分客群、采用触点与扩张路径

绘制 Firmus 五类买方从发现、评估、承诺、部署到扩张的路径,突出已有证据和仍存在的缺口。

旅程阶段来自公开披露里程碑和 Firmus 的既定战略。公开记录没有客户旅程数据(成交周期、转化率或流失触发因素);这些阶段反映推断出的客户生命周期,而非观测指标。

[CU011, CU016, CU021, CU027, CU040]

6.2 具名客户证据——部署、结果与证据质量

相比融资规模,Firmus 的具名客户记录偏薄;但 AI Singapore 合作深度,为其运营能力提供了真实参照点。AI Singapore(AISG)是由 Singapore National Research Foundation 支持的国家 AI 计划,它使用托管在 STT GDC 新加坡设施中的 Sustainable Metal Cloud(SMC)平台,开发 SEA-LION,即首个面向东南亚语言的开源大语言模型套件。AISG 合作是 Firmus 记录最完整的客户关系:256 块 Nvidia H200 GPU 部署在 32 个节点;团队实现每设备每秒 317 GB/s NCCL 带宽和 387 TFLOP;完成超过 200 次实验;一个 27-billion-parameter 模型在 10 天内完成训练;监督微调 runs 在 4 小时内完成;checkpoint evaluations 使用 2 个节点在 1.5 小时内完成。Firmus 与 AISG 于 2025 年 3 月通过 Memorandum of Understanding 正式化合作。这是一个生产部署,且结果被量化并公开确认。 相比之下,两份 Project Southgate hyperscale 合同都来自匿名 “leading global technology companies”,客户自身没有公开披露。第二份合同于 2026 年 3 月 2 日宣布,覆盖 Melbourne facility 约 18,400 块 Nvidia GB300 GPU,是多年、数十亿美元级承诺。第一份合同覆盖 Tasmania campus 约 36,800 块 Nvidia GB300 GPU。两者客户身份都不在公开记录中。行业观察者猜测交易对手可能是 Microsoft、Google、Meta 或某个大型亚洲 hyperscaler,但没有确认。这种不透明阻止了对信用质量、合同条款、取消条款和客户集中风险的独立验证。 NVIDIA 的 DGX Cloud 平台正在 Southgate Melbourne 建立专用 region,使 Firmus 成为 NVIDIA Cloud Partners 生态成员,与 CoreWeave、Crusoe、Lambda 和 SoftBank 一同进入 DGX Cloud Lepton marketplace。这代表分发渠道和平台伙伴关系,而非单独产生收入的客户关系;但它验证了 Firmus 在 NVIDIA 全球生态中的地位,并提供了触达亚太开发者的路径。STT GDC Sustainable Metal Cloud 合作始于 2023 年 6 月,当时 Temasek 支持的 STT GDC 对其与 Firmus 的全球合资项目进行重大投资;这是最早进入生产阶段的部署,已在 STT GDC 新加坡地点上线。新加坡站点为 AISG 及其他企业和政府客户实现 99 PFLOP/s 总算力。 [CU001, CU002, CU003, CU004, CU005, CU006]

具名客户证明表
客户分部部署 / 用例生产 / 试点已披露结果证据限制
AI Singapore (AISG)政府 / 主权研究(NRF 支持)SEA-LION LLM 训练——东南亚开源 LLM 套件;256 块 H200 GPU,32 个节点生产(2023 年下半年上线;2025 年 3 月 MoU)27B 参数模型 10 天完成;200+ 次实验;NCCL 317 GB/s;387 TFLOP/设备/s;4 小时 SFT 运行Firmus 发布的案例研究;无独立第三方审计;收入未披露
超大规模客户 1(匿名)全球超大规模云 / AI塔斯马尼亚 AI 工厂——约 36,800 块 Nvidia GB300 GPU;专用高密度算力容量预生产(建设中;目标 2026 年末)未披露——客户身份、合同金额和结果均保留身份匿名;无独立确认;信用质量无法核实
超大规模客户 2(匿名)全球超大规模云 / AI墨尔本 AI 工厂——约 18,400 块 Nvidia GB300 GPU;多年期专用基础设施预生产(墨尔本预计 2026 年中上线)未披露——描述为多年期、数十亿美元级;客户身份保留身份匿名;无独立确认;集中度风险无法评估
NVIDIA DGX Cloud(Lepton 市场)平台 / 渠道伙伴(分发)位于 Southgate Melbourne 的 DGX Cloud 区域;NCP 市场挂牌,提供 APAC 开发者访问预生产 / 合作(DGX Cloud 区域 2025 年 10 月宣布;Lepton 2025 年 6 月)不是单一客户;验证 Firmus 在 NVIDIA 生态中的地位;通过市场触达开发者未披露收入分成或成交量;NVIDIA 同时是战略股权投资者——关系并非完全独立
ST Telemedia GDC (STT GDC)——经由 SMC 合作基础设施运营商(Temasek 支持;场地伙伴)Sustainable Metal Cloud (SMC) GPU IaaS 横跨 STT GDC 新加坡站点;裸金属 HPC 集群生产(2023 年下半年起上线)SMC 平台已上线;Firmus 在 STT GDC 新加坡设施内运营 H200/H100 节点集群STT GDC 是场地 / 基础设施伙伴,不是终端客户;终端客户收入未拆分

仅列出公开确认或公司披露的关系。匿名超大规模客户也纳入表中,用于体现合同证据,即便交易对手身份未公开。「生产 / 试点」依据现有证据判断;Firmus 自身披露中未使用这些术语。证据限制反映截至 2026 年 6 月的公开记录约束。

[CU001, CU002, CU003, CU004, CU005, CU007]
FU001: 客户证据质量矩阵——证据深度 vs 生产成熟度

按五个维度为每个已披露客户关系评分:证据类型(官方 / 案例研究 / 新闻)、实名与匿名、部署阶段、结果验证和来源独立性。AI Singapore 是唯一拥有已验证结果和实名身份的客户关系。

评分是分析师仅依据可得公开证据作出的定性评估。未公布客户、专有案例研究或受 NDA 保护的项目未纳入。随着 Melbourne 上线,「投产前」阶段可能转为「生产环境」。

[CU004, CU005, CU008, CU013, CU034]

6.3 采用轨迹、GPU 承诺与平台触达

衡量 Firmus 采用轨迹,最好看已承诺 GPU-equivalents,而不是已披露客户数或收入。两份 Project Southgate hyperscale 合同合计承诺约 55,200 块 Nvidia GB300 GPU,分布在塔斯马尼亚和墨尔本,隐含数十亿美元资本承诺,并预计在投运后产生合同化多年收入。新加坡 SMC 部署为 AISG 及其他企业客户,在两个 STT GDC 站点运行约 4,000 块 Nvidia GPU,生产中提供 99 PFLOP/s 总算力。AI Cloud 公共平台已上线 H200、H100、L40S、A100 和 A40 GPU 实例;但截至 2026 年 6 月,Firmus 没有公开披露活跃付费客户数、收入拆分或采用率。 独立平台验证很强。SemiAnalysis 将 Firmus 的 Sustainable Metal Cloud 评为全球仅有三家正确实现 InfiniBand SHARP in-network reduction 的 GPU 提供商之一——在首届 ClusterMAX GPU Cloud Ratings 中,Firmus 与 AWS 并列,并高于 Google Cloud。SemiAnalysis 2025 年 11 月发布的 ClusterMAX 2.0 方法论,结合 hands-on benchmarking 和超过 140 份客户调研,覆盖被评估的 84 家提供商,跟踪全集为 209 家 GPU 云运营商。根据 Hedgehog Cloud 对 ClusterMAX 2.0 分层结构的分析,Gold-tier 云(Firmus 目标层级)相较 commodity Bronze 基础设施可获得 70% 定价溢价,说明 Firmus 性能定位具有收入价值。最近一轮中,只有 CoreWeave 持有 Platinum。 Firmus 被纳入 NVIDIA DGX Cloud Lepton marketplace——2025 年 6 月 COMPUTEX 宣布——使其成为全球 NVIDIA 开发者社区可访问的 GPU 来源。Blackstone 债务额度以已签客户合同为提款条件,直接把未来基础设施部署速度与新客户 wins 绑定。Firmus 称公司正在产生收入,但没有公开披露收入数字、毛利率、burn rate 或 runway。已承诺 GPU 容量管线意味着若合同按结构交付,将产生显著收入;但这取决于建设时间表、客户履约义务和债务提款排序。 [CU002, CU003, CU011, CU012, CU014, CU015]

采用轨迹和 GPU 承诺指标
指标数值 / 状态日期 / 期间来源置信度含义缺失分母 / 缺口
超大规模 GPU 承诺(Project Southgate)两份合同合计约 55,200 块 Nvidia GB300 GPU合同签于 2025 年 7 月和 2026 年 3 月高——官方公告获行业媒体佐证澳大利亚公开宣布的最大 AI 工厂 GPU 承诺客户身份和财务条款均未披露
Singapore AI Cloud 生产 GPU(SMC / STT GDC)两个 STT GDC 站点合计约 4,000 块 Nvidia GPU;99 PFLOP/s2023 年下半年起上线高——Firmus 案例研究和 STT GDC 公告确认首个产生收入的 AI 部署;2023 年起进入生产未披露收入和活跃客户数
AISG SEA-LION 部署(H200 节点)32 个节点 / 256 块 H200 GPU;200+ 次实验;27B 参数模型 10 天完成训练2025 年 3 月正式合作高——已发布案例研究量化了结果最可信的公开生产工作负载性能证明未披露合作期限和续约状态
活跃 Firmus AI Cloud 付费客户(公开平台)未披露截至 2026 年 6 月N/A——未披露无法评估平台牵引力或收入多元化关键缺口:未发布客户数或增长率
DGX Cloud Lepton 市场收录作为 NCP 与 CoreWeave、Crusoe、Lambda、SoftBank、Foxconn 并列2025 年 6 月宣布(COMPUTEX)高——NVIDIA 官方新闻稿 + Firmus 确认为 APAC 区域增加开发者触达分发渠道未披露市场收入分成或成交量
SemiAnalysis ClusterMAX GPU Cloud Ratings(GPU 云评级)全球前三;与 AWS 并列排名,高于 Google Cloud;支持 SHARPClusterMAX v1 2025 年 3 月;v2.0 2025 年 11 月中——独立分析师基准测试;方法已发布支撑溢价定价(较 Bronze 高 70%)和企业可信度Firmus 在 v2.0 中的具体层级(Gold/Silver/Platinum)未公开说明

GPU 承诺数字来自 Firmus 官方新闻稿,并获行业媒体佐证。收入、客户数、NRR 和流失率均未公开披露。空值 / “未披露”单元格表示缺少公开数据,不代表数值为零。SMC GPU 数量来自公司来源的第一章证据;AISG 成果来自已发布案例研究(2025 年 12 月)。

[CU002, CU003, CU005, CU006, CU007, CU011]
FU002: 采用 / 部署漏斗——从发现到扩张

展示 Firmus 客户从发现到部署再到扩张的路径,标出哪些环节已有真实采用证据,哪些记录仍不透明。

漏斗数值为序数(1–6 排名,不代表收入),用来表示各阶段证据深度。各阶段的实际收入和客户数量未公开。

[CU001, CU006, CU007, CU017, CU029]

6.4 留存、合同耐久性与集中风险

Firmus 没有公开披露任何留存指标。Firmus AI Cloud 公共平台或 hyperscale 锚定合同都没有可得的 NRR、GRR、churn rate、cohort analysis 或 repeat-engagement 数据。合同结构只部分可见:hyperscale 交易被描述为 “multi-year” 和 “multi-billion-dollar”,与能源和数据中心项目常见的合同化基础设施 take-or-pay 结构一致。Blackstone 债务额度以已签长期客户合同为提款条件,为贷款人提供结构性保护;但这些合同中的最低收入承诺、取消条款和终止触发条件没有公开披露。VAST Data 被选为 AI Cloud 数据层时,明确被表述为支持多租户企业和政府工作负载,说明平台架构预期会服务需要隔离环境的受监管行业客户。 按标准基础设施承销尺度看,客户集中风险很严重。截至 2026 年 6 月,两份匿名 hyperscale 合同很可能代表 Firmus 已合同化收入管线的绝大部分。如果任一客户延迟部署时间表、终止合同或重新谈判价格,收入影响都会对 Firmus 偿付 Blackstone 债务的能力产生实质影响。缺少具名客户组合,使投资者无法独立评估交易对手信用质量、战略理由或续约概率。Gilbert and Tobin 指出,该债务额度设计为随新合同 wins 扩大提款,并点名 Meta 和 OpenAI 是这类基础设施的典型用户——但实际客户没有具名。 扩张能力被结构性嵌入模型:Project Southgate 的多站点雄心(塔斯马尼亚、墨尔本,之后是悉尼、堪培拉、珀斯、阿德莱德)为到 2028 年实现地域和客户多元化提供路径。但截至本报告日期,企业和初创分部仍属愿景。塔斯马尼亚政府 2025 年 7 月建立 AI Factory Zone 政策框架,验证了主权 AI 需求;但政策支持不等于客户合同。Endeavor Asset Management CIO、早期 Firmus 投资者 Hayden Beamish 已于 2023 年 8 月退出,他在 2026 年 5 月发出负面评价,称 ASX 上市 “will likely be priced too expensively”,且 “the easy money has already been made”,并把转型前的 Firmus 描述为 “with no legitimate business model” 的 bitcoin miner。供应链伙伴 Benmax 和 Maas Group 正为已承诺 hyperscale 合同进行工业规模 AI Factory 部件生产,表明 Firmus 运营上正在履行现有客户义务——但两份已披露 hyperscale 合同之外的客户管线仍完全不透明。 [CU004, CU017, CU018, CU019, CU022, CU026]

留存、耐久性与满意度——可用指标和缺口
指标数值 / 状态适用分部置信度尽调问题
净收入留存率(NRR)未披露全部分部N/A向管理层索取按分部和年份队列划分的 NRR;IPO 招股书可能披露
总收入留存率(GRR)未披露全部分部N/A索取 GRR,以区分扩张和缓解流失;对重合同模式至关重要
客户流失 / 合同取消无取消证据;未发布流失数据超大规模、云、企业低(没有证据 ≠ 没有流失)索取逐客户续约和取消历史;确认新加坡 SMC 无提前终止
合同期限和最低承诺多年期(超大规模);MoU(AISG);按需(公有云)超大规模、政府 / 研究、云中——合同期限类别已确认;条款未披露索取最低收入承诺、容量利用率底线和取消通知条款
客户满意度 / 基准SemiAnalysis ClusterMAX——全球前三;v2.0 方法包含 140+ 份客户调查;2024 DCD Asia Pacific 获奖者云 / 企业(基准)中——独立方法;Firmus 在 v2.0 中的具体层级未说明获取 Firmus 具体 ClusterMAX v2.0 分数层级;索取客户 NPS 或 CSAT 数据

所有留存指标均未披露。空值 / “未披露”单元格表示缺少公开可得数据,不代表确认值为零。超大规模合同的合同期限类别(多年期)已确认;财务条款(最低金额、取消条款)不在公开记录中。SemiAnalysis ClusterMAX 满意度代理指标是公开领域唯一独立的客户质量信号。

[CU015, CU025, CU026, CU029, CU031]
扩张与集中度风险
驱动因素 / 风险维度类型影响证据中的缓释尽调路径
匿名超大规模客户集中度风险关键——两个锚定客户均未具名;大部分合同收入无法按交易对手信用质量核验Blackstone 提款条件部分缓解投机性支出;这是结构性保护,不是交易对手保护IPO 招股书必须具名客户或提供集中度披露;索取前三大客户收入占比;确认续约条款
超大规模云厂商内包 / 自建风险风险高——如果主要超大规模云厂商加快自建 APAC 设施,外包 AI 工厂需求可能下降Firmus 的主权定位、澳大利亚制造 IP 和可再生能源接入抬高切换成本;共同设计理念降低可替代性跟踪超大规模云厂商 APAC 资本开支公告;检查现有合同是否包含排他性或照付不议底线
多站点地域扩张(落地后扩张)扩张驱动因素高——Project Southgate 计划到 2028 年覆盖塔斯马尼亚、墨尔本、悉尼、堪培拉、珀斯、阿德莱德站点Blackstone 融资安排按新增已执行合同提款;CDC 合作提供跨站点基础设施跟踪新站点合同公告;评估企业 / 政府分部是否能在超大规模之外填充增量容量
面向 SMB 和开发者分部的 DGX Cloud Lepton / NCP 渠道扩张驱动因素中——将可服务分部从超大规模扩展到 APAC 各地初创公司和企业 AI 开发者NVIDIA 合作和市场挂牌已确认;开发者获取率未披露索取通过 DGX Cloud Lepton 获取的开发者 / 初创客户数;跟踪按需云 GPU 利用率增长
澳大利亚企业 / 政府分部(愿景)扩张驱动因素 / 缺口中——如果公共部门主权 AI 要求转化为采购,战略价值显著Tasmania AI Factory Zone 政策于 2025 年 7 月设立;尚无澳大利亚联邦 / 州政府具名生产客户跟踪联邦 AI 战略采购;确认是否已有任何全政府 AI 基础设施框架招标在推进

影响评级是作者基于可得公开证据和可比基础设施公司风险框架作出的定性评估。缓释力度仅按公开披露机制评估;未经验证的内部控制可能存在。尽调路径是研究建议,不代表 Firmus 已确认推进。

[CU001, CU004, CU011, CU017, CU022, CU028]
FU004: 集中度与扩张风险矩阵

从五个维度评估 Firmus 关键客户集中和扩张风险的严重性与缓释力度,突出风险最高、缓释证据最弱的位置。

严重性和缓释力度是仅基于公开证据作出的定性评估。内部风险管理框架、未公布合同或专有缓释措施未反映在内,并可能实质性改变评估。

[CU004, CU022, CU029, CU037, CU038]

6.5 图表证据

Chapter 07

07风险

7.1 治理、法律与监管风险

Firmus 画像中最突出的单一治理风险,是 co-CEO Oliver Curtis 2016 年的内幕交易定罪。2007 至 2008 年间,Curtis 与一名前同事合谋,利用对一名基金经理交易意图的预先知情,执行 45 笔盈利 CFD 交易,净获利 A$1,432,228.85。他被 Supreme Court 陪审团定罪,判处两年监禁并服刑十二个月。其上诉被驳回。随后 ASX 按 Listing Rule 1.1 Condition 20 Good Fame and Character 框架让 Curtis 通过审查;这是 ASX 对任何寻求进入官方名单实体的董事和拟任董事执行的测试。ASX 可将检查延伸到高级管理人员,并可在标准未维持时拒绝或撤销上市。包括 Roger Montgomery 和 NAOS Asset Management 在内的独立基金经理公开指出,该定罪与在这一规模上市公司担任董事直接相关。 另一项结构性治理风险,是拟议上市时不对 pre-IPO 股东设置 escrow 或股份出售限制。在澳大利亚大型科技上市中,lock-up period 是标准做法,缺失该安排并不典型。独立基金经理认为,缺少 escrow 实际上把散户 IPO 买方变成早期支持者的退出流动性。极端集中进一步放大担忧:一些 boutique VC 和 PE 基金持有的 Firmus 仓位,在快速估值上调后占基金总价值超过 60%,强烈激励它们在上市时立即兑现收益。 监管方面,Firmus 计划中的数据中心运营正落在 Australia Security of Critical Infrastructure Act 2018(SOCI Act)范围内。用于向 Commonwealth、State 或 Territory 机构,或向其他关键基础设施资产责任实体提供服务的数据存储和处理资产,会被认定为 critical data storage or processing assets。一旦被纳入,责任实体必须登记资产、报告网络事件,并采用、维护和遵守书面 Critical Infrastructure Risk Management Program(CIRMP)。拟议 2026 年 SOCI Act 修正案将对高风险资产类别施加更具规定性的义务,包括供应链 mapping、供应商 FOCI 风险评估、phishing-resistant MFA、可独立运行三个月的网络隔离,以及关键工作人员 AusCheck 背景审查。拟议改革下,违规民事罚款每起事件最高可达 $3.3 million。ASIC 对 ASX 治理的调查(Final Report 于 2026 年 3 月发布)也表明,上市实体及其高管面临更趋严格的监管环境,风险管理和合规职能被要求进行转型式变革。[CR001, CR002, CR003, CR004, CR005, CR006]

监管 / 法律风险登记表
风险 / 规则 / 案件司法辖区状态可能性严重性缓释剩余暴露尽调路径
ASX Listing Rule 1.1 Condition 20——良好声誉与品格(Curtis 内幕交易定罪)澳大利亚(ASX)ASX 已于 2026 年 3 月通过 Curtis;按 Listing Rules 持续适用关键董事会刷新;与 ASIC/ASX 持续沟通机构投资者可能筛除;如再次违规,可能被迫调整董事会获取 ASX 对 IPO 持续合规的书面确认
IPO 前股东无托管 / 售股限制澳大利亚(ASX/Corporations Act)公司确认;未发现监管障碍未披露;董事会可自愿施加锁定零售投资者在上市时承受即时抛压要求主要 IPO 前持有人作出自愿托管承诺
SOCI Act 2018——关键基础设施风险管理计划(CIRMP)义务澳大利亚(Cth)范围内义务适用;拟议 2026 年修订案设置宽限期需要启动 CIRMP 项目不合规罚款最高 $3.3M;ASD 可能介入索取 CIRMP 项目文件和差距分析
Privacy Act 2026——ADM 披露要求(2026 年 12 月生效)澳大利亚(Cth)待生效;2026 年 12 月 10 日起强制需要隐私影响评估和更新政策OAIC 执法;严重违规罚款最高 AUD$50M确认 ADM 清单和披露政策时间表
Foreign Acquisitions & Takeovers Act——FIRB 审查 Nvidia 和 Blackstone 投资澳大利亚(Cth)未公开披露 FIRB 决定标准 FIRB 通知流程如果附加条件,Nvidia 供应关系或 Blackstone 契约可能受限确认 FIRB 批准状态及任何条件
拟议 2026 年 SOCI CIRMP 修订案——供应链 FOCI 供应商风险要求澳大利亚(Cth)咨询于 2026 年 5 月结束;预计将实施需要供应链映射;需要 Nvidia 供应商 FOCI 评估如果 Nvidia 或 Blackstone 被标记为 FOCI 风险,可能要求补偿控制或附加剥离条件就 FOCI 暴露咨询法律顾问;审查 GT Law CIRMP 分析

状态和可能性评估截至 2026 年 6 月。严重性按四级量表(低 / 中 / 高 / 关键)评级,依据是对 IPO 可行性、持续运营和估值的潜在影响。截至本次运行日期尚未发布招股书;监管风险披露依据分析师和法律评论以及公开可得法律。

[CR002, CR003, CR004, CR007, CR008, CR009]
FR001: 风险热力图——Firmus 主要风险的发生可能性 vs 严重性

基于截至 2026 年 6 月的公开来源定性评估,将 Firmus 十二项最高优先级风险放入发生可能性—严重性矩阵。

发生可能性和严重性是截至 2026 年 6 月来自公开分析师评论的定性专家评估。未使用量化概率或 VaR 模型。

[CR002, CR003, CR004, CR007, CR014, CR017]

7.2 运营、技术与网络安全风险

Firmus 的竞争护城河高度依赖其声称的行业领先效率,最显眼的是 1.03 的 Power Usage Effectiveness(PUE)数字。PUE 公式用总设施电力除以 IT 设备电力——满分是 1.0。包括 Google 和 Microsoft 在内的全球 hyperscaler 通常报告 1.09 至 1.2 的 PUE;多数澳大利亚数据中心在 1.4 至 1.7 之间。Firmus 后来承认,其 1.03 数据是 2021 年的一次单点测量,而不是经过独立审计的年度平均值。University of New South Wales 的 Dr. Amr Omar 认为该说法处在全球可能性的极端边界。公众没有看到任何第三方 PUE 验证。如果未来最终记录的持续能效明显不及 1.03,经济论点——即假设结构性更优的电力经济性支撑利润率——会被直接削弱。 运营模型依赖液体浸没冷却,这带来专门维护要求。全球受训并获得液体浸没冷却认证的技术人员池目前非常有限。到 2028 年从几千块 GPU 扩展到 1.6 GW,意味着需要在一个并不普及的技能上快速扩员。若该职能人员配置不足,可能表现为设施停机、保修失效或安全事故。 NEXTDC 质疑也值得单独关注,它是技术可信度信号。NEXTDC CEO Craig Scroggie 公开质疑 Firmus 声称的每 MW 约 US$6 million 建设成本,并引用每 MW 约 US$12.4 million 的行业标准。如果 Scroggie 的说法准确,而 Firmus 成本数字又没有得到招股书级别第三方评估验证,那么支撑 IPO 估值的整个利润率和资本回报框架都会受到挑战。截至本次报告日期,AI Factories 还没有任何独立建设成本审计进入公共领域。 网络安全角度看,处理匿名 hyperscale 客户 AI 工作负载的数据中心,本质上就是高价值目标。SOCI Act 要求报告网络事件并采用 CIRMP。拟议 2026 年修正案将要求达到 Essential Eight、ISO 27001、NIST CSF 2.0 或同等框架下的 Maturity Level 2——对一家 2021 年才进入该领域的运营商而言,这是显著抬升。SOCI 网络成熟度 24 个月宽限期意味着义务可能在 2028 年中到期,正好与 Project Southgate 建设高峰重合。[CR014, CR015, CR016, CR017, CR018, CR019]

运营、质量与安全风险登记表
失效模式可能性严重性缓释成熟度剩余暴露未解决缺口
PUE 1.03 未验证——实际持续效率可能大幅低于声称水平低——无第三方审计如果实际 PUE 为 1.2–1.4,利润率和估值论点会被直接削弱未发布独立年度 PUE 审计
大规模液浸冷却技术人员短缺低——全球人才池有限设施停机、安全事故、保修失效未披露扩张至 1.6 GW 所需的劳动力计划
Nvidia GPU 供应延迟——H200/GB300/Vera Rubin 瓶颈关键部分——战略 Nvidia 关系提供优先访问建设延迟期间 $10B 融资利息累积;合同罚金GPU 交付排期未公开承诺;Nvidia 路线图可能变化
电网可靠性——塔斯马尼亚和全国电网稳定性部分——水电供应合同;AEMC 已参与并网延误可能推迟投产;政治反对可能阻挡站点AEMC 数据中心并网标准尚未最终确定
AI 工厂网络安全事件——客户数据或模型外泄关键低——SOCI CIRMP 尚未完成;未确认 IRAP 认证客户终止合同;声誉受损;触发强制 SOCI 违规通知未公开披露 IRAP、ISO 27001 或 Essential Eight 认证状态
四城建设中的施工和许可延误低——项目规模新颖,过往记录有限债务契约里程碑违约;IPO 叙事恶化未对 2028 年时间表作独立项目管理验证

可能性和严重性基于截至 2026 年 6 月的公开来源和行业基准评估。缓释成熟度反映公开披露状态;内部控制尚未外部验证。

[CR014, CR015, CR016, CR017, CR018, CR019]

7.3 财务、资本与伙伴依赖风险

Blackstone 牵头的 US$10 billion 债务工具,是这里最核心的财务风险。该工具按高级担保基础设施债设计,只有在与超大规模云厂商和蓝筹 AI 客户签下长期客户合同时才可提款。设计上,它限制了投机性部署,但也引入了顺序风险:合同谈判一旦停滞,公司就拿不到提款,资本计划也会卡住。建设里程碑延误、锚定合同未能续约或留存、交易对手信用质量大幅恶化,都可能触发契约违约和贷款人权利——包括要求立即部分还款,或限制继续提款。Nomura 据称很难把一笔已出资的 Firmus $120M 四年期私人信贷工具卖下去,说明至少一家机构级银行市场参与者认为,按当时条款看,这个风险并不划算;放在 $10B 工具的语境下,这是一个重要负面信号。 截至 2028 年,总资本开支计划为 A$73.3 billion,约为截至 2026 年 4 月累计股权融资额的 55 倍。如此高的杠杆比率意味着,IPO 募资和持续债务提款不是可选项,而是计划能运转的必要条件。如果 IPO 定价显著低于目标 A$8–12 billion 区间,或簿记遇到困难,公司履行债务合同管线的能力会受损。目标定价本身——比当时私募估值高 40% 到 100%——就是一次估值跃升,且近期澳大利亚科技 IPO 历史中还没有确认可比案例。 Nvidia 依赖风险很特殊。Nvidia 既是 Firmus 的主要 GPU 供应商,也是股权投资者。Roger Montgomery 的分析把这种关系称作「共生循环」:Nvidia 的小额股权投资(未披露,可能低于 A$100 million)撬动了第三方 A$830 million 股权融资和 $10 billion 债务工具,而其中绝大多数又会以硬件收入形式流回 Nvidia。这个安排让 Nvidia 对 Firmus 的资本配置形成显著间接影响,也把 Firmus 锁进对 Nvidia 芯片交付节奏的「金手铐」依赖。H200、GB300 和即将推出的 Vera Rubin 平台的 GPU 供应都受全球结构性瓶颈约束,据称芯片交付最长可达十二个月。建设延误期间,债务工具仍会计息。Project Southgate 的两个锚定客户仍未具名,外部无法核验交易对手信用质量和合同可执行性。[CR023, CR024, CR025, CR026, CR027, CR028]

伙伴与依赖风险登记表
依赖项交易对手角色集中度失效场景严重性缓释剩余暴露
GPU 供应和股权锚定Nvidia主要 GPU 供应商和少数股权投资者关键——100% 算力硬件供应延迟、定价变化,或战略转向不再支持 Firmus关键战略 Nvidia 关系;DGX Cloud 与 Southgate Melbourne 共址对任何 Nvidia 供应中断承担完整成本和进度暴露
US$10B 债务融资Blackstone Tactical Opportunities / Blackstone Credit & Insurance(债务融资方)高级担保债务贷款方关键——主要非股权融资工具契约违约;错过里程碑;合同流失触发贷款方权利关键项目融资结构将提款与已执行合同挂钩如契约违约,可能要求提前还款;限制进一步提款
物理基础设施骨干CDC Data CentresProject Southgate 的主权物理骨干;Firmus 拥有 IT 层高——所有 Southgate 站点均位于 CDC 设施内CDC 转向自建 GPU 云;合作关系或商业条款恶化合作协议;双方在主权 AI 叙事上战略一致如果 CDC 关系终止,运营将全面受扰
锚定超大规模收入未具名全球科技公司(客户 1)多年期照付不议 GB300 合同(墨尔本)关键——首次债务提款锚点合同取消、重新谈判,或交易对手信用事件关键已执行多年期协议;假设未具名客户为投资级交易对手信用质量无法核验;合同条款不公开
锚定超大规模收入未具名全球科技公司(客户 2)首份 Project Southgate 多年期协议高——与客户 1 共同锚定债务提款与客户 1 相同;身份未披露已执行多年期协议与客户 1 相同;两个客户合计收入集中度无法量化
电网集成软件EatonSynert 电网集成软件部署伙伴Eaton 停止支持 Synert 或改变商业条款合同安排;软件可能可迁移到新平台如果 Eaton 合作失败,运营将受扰;公开细节有限
债务共同贷款方和股权共同投资者Coatue Management2026 年 4 月 Series F 股权领投方;Blackstone 债务共同参与方Coatue 下调承诺,反映市场担忧股权与债务双重关系带来利益一致性Nomura 减持出售困难,显示更广泛的市场承接风险

集中度评级仅反映公开来源分析。Blackstone 债务、两份超大规模云客户协议、CDC 合作关系和 Eaton Synert 安排的合同条款均不在公开领域。截至本报告日期,客户身份仍未披露。

[CR023, CR026, CR027, CR028, CR029, CR030]
FR002: 风险传导图——关键风险如何流向价值和融资

有向无环图展示 Firmus 主要风险源如何沿运营和财务渠道传导,影响收入、融资能力和企业估值。

[CR003, CR004, CR014, CR019, CR023, CR025]
FR003: Firmus 资本敞口——已募股权 vs 总资本开支计划

柱状图比较 Firmus 截至 2026 年 4 月的累计股权融资(约 US$1.35B)、其总资本开支目标(到 2028 年 A$73.3B)、Blackstone 债务额度上限(US$10B)以及当前企业估值(2026 年 4 月融资轮为 US$5.5B)。图表展示支撑该计划的极高杠杆比率,以及对持续债务提款和 IPO 募资的依赖。

已募股权按 AUD/USD 0.64 折算;总资本开支计划目标以 AUD 计;Blackstone 融资额度和企业估值以 USD 计。数值为截至 2026 年 4–6 月的公开披露估算,不包含未公布的额外股权或债务批次。

[CR024, CR025, CR029, CR030, CR031]

7.4 执行、团队与环境风险

Firmus 创始团队组建时,并没有数据中心或工程方面的过往专长。Oliver Curtis 创办 Firmus 之前是投行人士,后来做过加密货币投资;Tim Rosenfield 此前担任内衣公司 Simone Perele 的 CEO;Jonathan Levee 曾在资源行业担任中层管理岗位。公司引入了有经验的运营者——Ted Pretty、Grant Dempsey、CFO Kirsty Godfrey-Billy(2025 年 9 月)以及 2026 年 3 月任命的三名非执行董事——明显补强了管理层厚度。但 Godfrey-Billy 在目标 IPO 日期前任 CFO 还不到十二个月,新任董事会主席和三名新非执行董事都是过去九个月内加入;任何董事任命都不能替代公司自身十年的运营经验。创始层的个人和家族网络——Curtis 与 Rosenfield 是表兄弟,Levee 曾是 Curtis 的前姻亲——把非正式影响力集中在一个小圈子里。 按计划规模看,执行风险很高。Project Southgate 目标是在 2028 年前,于澳大利亚多个首府城市建设 1.6 GW 容量。在澳大利亚,乃至据本评审所知在全球范围内,还没有运营商尝试过以这种密度和速度建设主权 AI 基础设施。一支新近搭建的高管团队,要同时在四个城市管理电网接入、设备交付、建设许可和客户上线,这是一类最高级别的执行挑战。Firmus 承诺每 50 MW 创造 50 到 100 个全职岗位,UNSW 教授 Toby Walsh 已公开称其不现实,并指出现代自动化数据中心通常不需要这种劳动规模。如果这些数字未经限定就写入招股书,可能因前瞻性陈述招来 ASIC 审查。 环境和能源监管风险集中在塔斯马尼亚,并在全国层面开始显现。Firmus 与 TasHydro 的水电安排遭到绿党议员 Tabatha Badger 公开反对,她的理由是,传统行业此前曾被告知电网没有容量支持它们自身的绿色转型。ABC 2026 年 3 月的报道对该电力交易结构提出了正式问题。全国层面,AEMC 关于数据中心接入新电网标准的提案,以及 Greenpeace Australia 呼吁在能源转型规划完成前暂停批准新的 AI 数据中心,都显示监管逆风可能带来接入排队、电网接入费或环境许可条件,从而抬高 Firmus 成本并推迟投运时间。澳大利亚 Privacy Act 2026 改革——尤其是 2026 年 12 月生效的强制 ADM 披露要求,以及大幅提高的民事罚款(严重违规最高 AUD$50M)——也会给面向客户的产品栈顶层增加合规义务。[CR033, CR034, CR035, CR036, CR037, CR038]

人员与执行风险登记表
角色 / 职能依赖项或缺口可能性严重性缓释措施尽调路径
联席 CEO Oliver Curtis内幕交易定罪;机构基金治理筛查;上市后仍有声誉与监管敞口2026 年 3 月通过 ASX Good Fame & Character 审查;董事会独立性已加强获取 ASX 书面安慰函;核查机构投资者对治理的反馈
联席 CEO Tim Rosenfield此前没有数据中心、工程或大市值上市公司 CEO 经验经验丰富的董事会和顾问补足运营知识对照可比基础设施 CEO 资历做背景核查
CFO Kirsty Godfrey-Billy2025 年 9 月任命——到目标 IPO 日期任职不足 12 个月;首次担任此复杂度上市实体的 CFO具备财务背景;由经验丰富的顾问和账簿管理人支持尽调 CFO 对 IPO 准备度和财务控制基础设施的评估
董事会构成三名非执行董事于 2026 年 3 月获任;主席即将到任——上市时任期均不足 12 个月;对 Firmus 的集体运行记忆有限每名非执行董事都具备实质性行业资历;审计 / 风险委员会和薪酬委员会由相关任命人员担任主席评估董事会有效性流程和委员会职权范围
专业冷却工程师全球范围内,具备 1.6 GW 规模液浸冷却认证的技术人员短缺招聘管线未公开披露量化专业岗位当前人数、培训管线和流失率
项目管理——四城建设Firmus 团队或任何运营商此前都没有以这种速度和规模执行过可比项目外部顾问;CDC 关系提供设施专业能力获取对 2028 年交付计划的独立项目管理评审

可能性和严重性是基于截至 2026 年 6 月公开信息的定性评估。Firmus 没有公开内部 HR、组织架构图或董事会有效性数据。

[CR033, CR034, CR035, CR036, CR039]
缓释与否决标准表
风险可监测触发因素阈值 / 事件行动含义
Curtis 治理 / IPO 资格ASIC 或 ASX 就 Curtis 上市后继续任职适格性发起正式调查任何点名 Curtis 的 ASX 或 ASIC 不利正式通知立即复核董事会结构;若 Curtis 离任或被要求退居幕后,可能打断投资逻辑
无托管出售压力IPO 后前 90 天内部人士出售量>10% 的 IPO 前股份在上市后 30 天内售出可能造成价值毁损;尽调其余主要持有人的意向
Blackstone 契约违约公开或据报道未能达成债务提款里程碑;贷款方修订或豁免通知向 ASX 披露任何提款暂停或契约豁免投资逻辑中断——若无法提款,流动性和资本计划不可持续
Nvidia 供应中断Nvidia 芯片交付延误出现在公开沟通或业绩披露中任何主要 GPU 批次延误 >6 个月建设延误;利息成本加速;债务契约里程碑风险
能源电网监管阻断AEMC 或州政府发布正式暂停令或接入排队公告Tasmania、Victoria 或 NSW 对新数据中心接入实施任何正式监管阻断投产延误;按客户合同进行不可抗力审查
PUE / 效率偏离首次独立第三方 PUE 审计结果;工程同业评审任何主要站点经验证年度 PUE >1.15电力成本逻辑削弱;利润率被压缩;IPO 重新定价风险
客户集中——锚定合同流失未具名超大规模云客户发出任何公开的合同重谈或取消信号Project Southgate 任一锚定客户发出终止或重大修改通知投资逻辑中断——债务提款以这些合同为条件;IPO 故事受到实质性损害
IPO 撤回或推迟董事会或账簿管理人沟通 IPO 延迟或撤回IPO 推迟至 2026 年 Q3 之后或撤回资本计划缺口;烧钱加速;被迫债务重组

触发定义和阈值为指示性,基于公开分析师评论和先例 IPO 结构。截至本报告日期,实际招股说明书风险因素和债务契约尚未公开;正式否决标准必须在招股说明书发布后重新界定。

[CR004, CR023, CR025, CR026, CR028, CR037]
关键风险指标(KRI)监控仪表盘
风险类别KRI 指标监测频率黄色阈值红色阈值责任方
治理 / Curtis独立非执行董事占董事会席位的比例每月<50% 独立董事Curtis 被免职、离任,或收到不利 ASIC/ASX 通知董事会提名委员会 / ASX
无托管 IPO 结构上市后前 30 天,IPO 前股东累计减持占已发行股本比例上市后每日>10% 的已发行股本在上市后 30 天内售出>25% 的已发行股本在 30 天内售出,显示协调退出ASX/ASIC 持续披露监测方;券商
债务 / Blackstone 契约提款里程碑相对于项目计划的遵守情况每季度>1 个提款里程碑未达成且未获豁免>2 个里程碑未达成,或发出任何贷款方执行通知CFO / 资金部;Allens 法律顾问
技术 / PUE所有运营站点的年化平均 PUE每月(自报);每年(独立审计)任何主要站点年度平均 PUE >1.15年度平均 PUE >1.25,或 IPO 时无法取得第三方审计运营 / 工程;外部审计师
供应 / Nvidia GPUGPU 交付周期相对于项目计划每月任何主要 GPU 批次交付延误 >90 天延误 >180 天,或披露正式供应关系变更供应链 / Project Southgate PMO
能源 / 电网监管延迟 >6 个月的电网接入申请数量每月1 个电网接入延迟超过 6 个月目标2 个或以上接入被阻断或受到监管暂停基础设施 / 监管事务

KRI 阈值为指示性,来自公开分析师评论、行业基准和先例 IPO 结构。正式 KRI 定义必须在 ASX 上市前于风险框架文件中达成一致;招股说明书发布实际契约和里程碑规格后,应随之更新。

[CR002, CR023, CR016, CR018, CR028, CR037]

7.5 展示材料

Chapter 08

08估值

8.1 投资正方与反方

Firmus 的投资逻辑建立在四个结构性支柱上。第一,全球范围内,尤其在澳大利亚,超大规模 AI 算力需求受供给约束,IPO 前就形成了合同收入底线;美国前五大超大规模云厂商承诺的 2026 年资本开支为 USD $660–690 billion,几乎是 2025 年水平的两倍,四大厂商都自称受供给约束,而不是需求不足。第二,亚太主权 AI 基础设施具有战略稀缺价值,本土超大规模云厂商难以轻易复制;Firmus 的可再生能源定位和澳大利亚所有权结构,让它能够触达 AWS、Azure 和 Google 无法进入的主权算力预算。第三,Blackstone 的 USD $10 billion 高级担保债务工具以与超大规模云厂商签署长期客户合同为条件,提供了项目融资层面的验证,说明收入覆盖故事经过了严格机构信贷流程承销。第四,Nvidia 两轮参股及其在 Southgate Melbourne 部署 DGX Cloud 区域节点,同时锁定了芯片供应优先级和需求侧分销渠道;在 APAC,还没有其他运营商在这个规模上同时具备两者。 每个支柱都包含重大未核验假设。合同收入底线依赖两个未具名的超大规模客户,其财务条款、最低承诺和取消条款均未出现在公开记录中。随着 NEXTDC、Macquarie DC 和 CDC Data Centres 以相似或更强的治理资质建设竞争性主权容量,主权稀缺论每天都在走弱。Blackstone 工具制造了高杠杆资本结构:如果全部提款,按 7–9% 年利率计算,利息支出每年会吃掉 USD $700–900 million,超过目前任何已披露收入数字。Nvidia 的股权投资据独立分析师估计,两轮合计为 AUD $30–75 million,本质上是供应商融资工具:Nvidia 投入的每一美元,都会通过 GPU 硬件收入以数倍规模回到其资产负债表,因此它的「背书」在结构上是循环的。 IPO 前治理图景中,三个负面信号最突出。IPO 对已经实现 300–530% 账面收益的早期投资者不设托管或售股限制,上市首日的即时抛售风险由散户买家承担。Montgomery Investment Management 的 Roger Montgomery——本人也持有现有头寸——明确警告,散户 IPO 买家面临为激励充足的内部人提供「退出流动性」的风险。Wilson Asset Management 也把缺乏锁定限制称作「结构性红旗」。Firmus 的 PUE 效率 1.03 主张,已从当前测量值改口为「2021 年的点测量」,且从未经过独立审计;此前 UNSW 专家 Dr Amr Omar 称该数字处于「全球可能性的极端边缘」。早期前投资者、Endeavor Asset Management CIO Hayden Beamish 于 2023 年 8 月退出 Firmus,当时称其是「一个资产负债表拉得过长、没有合法商业模式的比特币矿工」;他现在表示「Firmus 容易赚的钱已经赚完了」,并且不会参与 IPO。[CV001, CV002, CV003, CV004, CV005, CV006]

投资逻辑与反向逻辑
维度投资逻辑论据反向逻辑反驳证据状态什么会改变判断
AI 需求拉动超大规模云厂商 2026 年资本开支 $660–690B;Firmus 已签两份数十亿级合同,在 IPO 前验证需求底部。合同对手方未具名;条款、最低承诺和取消权未披露;若 AI 效率提升(Jevons 悖论)加速,需求可能放缓。部分招股说明书披露客户身份和合同条款;超大规模云厂商 GPU 资本开支持续至 2027 年。
APAC 主权护城河澳大利亚的可再生能源、政治稳定性和贴近亚太的地理位置,形成主权算力稀缺价值;没有超大规模云厂商拥有这组资产组合。NEXTDC(OpenAI 园区)、Macquarie DC 和 CDC 正在建设主权容量,治理资质相当或更强;护城河并非独占。部分持续的 IRAP 或主权认证差异化,且竞争对手无法匹配;可证明的客户管线明确要求使用 Firmus。
Blackstone 信贷验证以合同为条件的 $10B 高级担保融资安排,说明机构贷款方相信合同收入足以支撑资本结构。全额提款按 7–9% 年利率计算,意味着 $700–900M 年利息;GAAP 盈利需要收入显著高于当前任何公开证据;CoreWeave 每季度 $740M 的 GAAP 亏损说明了这一模式。已确认(结构);未解决(财务影响)招股说明书披露 DSCR(偿债覆盖率);收入披露显示初始部署至少达到年利息覆盖 1.5x。
Nvidia 背书与供应Nvidia 参股,并在 Southgate Melbourne 部署区域 DGX Cloud,显示芯片供应优先级和需求侧分销能力,这是 APAC 竞争对手无法匹配的。Nvidia 两轮合计股权估计 <AUD $100M,而总融资为 AUD $1.35B;投资形态更像供应商融资:GPU 采购流会让资金多次回到 Nvidia,并非独立背书。部分(小额持仓已确认;供应商动态为推断)招股说明书披露 Nvidia 股权比例;证明存在非 GPU 供应商业关系;披露 Nvidia DGX Cloud 收入分成。
IPO 结构完整性全球路演已完成;四家一级账簿管理人获任;ASX Good Fame and Character 已为 Curtis 放行;董事会实质性增强。招股说明书尚未提交;没有托管;Oliver Curtis 定罪与证券语境下的治理直接相关;精品 VC 基金持有的 Firmus 超过基金价值 60%——集中抛售风险高。部分(流程推进;透明度低)招股说明书披露收入、锁定结构和经审计的效率主张;创始人和 IPO 前持有人至少 90 天托管。
能源成本优势已签约的塔斯马尼亚水电,估计价格为 USD $0.04–$0.06/kWh——处于全球数据中心电力成本最低区间——带来持久运营成本优势。PUE 1.03 的主张是 2021 年一次点测,并非独立审计;实际运营 PUE、电力合同量和定价均未公开验证;塔斯马尼亚水电交易引发政治摩擦。部分(成本区间为估计;PUE 未验证)满载独立 PUE 审计;招股说明书披露电力合同条款;电网接入持续且不受监管扰动。

证据状态(部分 / 已确认 / 未解决)由作者基于公开来源审阅得出。「部分」表示存在部分佐证证据,但关键假设仍未验证。「什么会改变判断」描述可消除不确定性的证据。

[CV005, CV006, CV011, CV019, CV020, CV021]

8.2 融资背景、入场纪律与估值立场

Firmus 六个月内的私募估值轨迹,是澳大利亚初创公司中记录最陡峭的之一:从 USD $1.9 billion(2025 年 9 月,Nvidia 牵头的 AUD $330 million 轮)到 USD $6 billion(2025 年 11 月,AUD $500 million 轮),再到投后 USD $5.5 billion(2026 年 4 月,Coatue 牵头、Nvidia 参与的最后一轮 IPO 前 USD $505 million 融资)。2026 年 4 月这一轮从 USD $6 billion 到投后 USD $5.5 billion 的表面降估值,反映的是汇率变动和交易结构,而不是平轮:按 AUD 计为 AUD $7.95 billion,略高于 2025 年 11 月隐含的 AUD $8 billion 标记。Startup Daily 和投资者通信报道称,IPO 目标最低为 AUD $8 billion,并视簿记需求冲刺 AUD $12 billion——意味着较最后一轮私募估值跃升 50–100%。 这个融资背景给公开市场投资者提出了明确的入场纪律问题。2025 年 9 月以 AUD $1.9 billion 估值入场的 IPO 前股东,在 2026 年 4 月标记下持有约 300–320% 账面收益;如果 IPO 估值为 AUD $8–12 billion,账面收益会升至 320–530%。据报道,在无托管结构下,这些持有人——包括创始人、管理层和机构——上市第一天都可以卖出。按 AUD $8 billion 投前估值募集 AUD $2 billion,意味着新增股份摊薄约 25%,但一级募集资金(进入 Firmus 资产负债表)与二级出售所得(流向售股股东)的拆分尚未公开披露;如果二级部分很大,资产负债表受益会相应更小。 Blackstone 的 USD $10 billion 工具由 Milbank(Firmus)、Allens(Blackstone)和 Kirkland & Ellis 参与设计,是高级担保、项目融资式结构:提款以已执行长期客户合同为条件,并随部署分阶段进行。该工具确实释放信用质量信号——Blackstone 的机构信贷流程很严格——但也让资本结构在全部提款后,相对于迄今已募集股权的债股比超过 7:1。按 7–9% 的私人基础设施债利率计算,全部提款意味着年度利息 USD $700–900 million,显著高于任何公开披露的收入水平,并使 GAAP 盈利路径取决于尚未披露的合同定价。CoreWeave 的平行经验很有参考价值:尽管 FY2025 调整后 EBITDA 率约为 56%,CoreWeave 在扣除债务服务和 GPU 折旧后,仅 2026 年 Q1 就录得 USD $740 million GAAP 净亏损。 建议为 TRACK。主权 APAC 市场中 AI 基础设施投资的结构性逻辑有证据支撑,也可信。问题在价格。若估值为 AUD $8 billion,且招股书不披露收入,Firmus 等于要求公开市场投资者把 CoreWeave 等价倍数(8–10x 2026E 收入)套用在一个公开记录中不存在的收入数字上。就现有证据看,入场价格偏高,但概念本身并非没有吸引力;第 5 节列出的尽调问题,精确定义了哪些证据能把 TRACK 转为 BUY。[CV001, CV002, CV005, CV007, CV023, CV024]

建议摘要
维度评估证据基础决策含义
建议观察两份已确认的超大规模云合同;Blackstone $10B 融资安排;Nvidia 背书——但没有招股说明书、未披露收入、没有托管。没有招股说明书,不应参与 IPO 投资。待收入、客户身份和锁定结构确认后再重新评估。
信心结构性逻辑证据充分;但缺少收入和招股说明书数据,限制了具体价格 / 回报判断。若招股说明书披露 AUD $800M+ 收入,且估值 ≤ AUD $8B,信心上调至高。
风险评级无托管;治理悬置问题(Curtis 定罪,已获放行);效率主张未经验证;杠杆极高;平台尚无收入。若 IPO 仍无托管,且估值超过 AUD $10B 但没有收入证明,风险降至关键级别。
估值立场偏高AUD $8–12B IPO 目标对比 CoreWeave 8–10x 收入倍数,意味着需要 AUD $800M–$1.5B 收入;招股说明书前没有证据支持。若招股说明书披露合同收入,且费率与 CoreWeave 一致,AUD $6–8B 估值可视为合理。
投资逻辑中断周期12–18 个月乐观情景取决于 GB300 部署在 2026 年 H2 上线,并且合同收入在 2027 年财务数据中兑现。监测 IPO 后前两次 ASX 业绩发布;将收入爬坡与 AUD $600M–$1.2B 基准情景第一年目标对比。

评估截至 2026 年 6 月 18 日,即招股说明书提交前。所有定性判断均由作者基于公开证据得出。估值数字均不基于未披露财务数据。

[CV001, CV002, CV007, CV005, CV013, CV025]
FV001: 推荐逻辑:从证据到 TRACK

决策链把四个结构性投资论点支柱与其关键不确定性连接到 TRACK 建议,并展示哪些问题清除后才可能转为 BUY。

[CV001, CV005, CV006, CV013, CV027]

8.3 可比公司、私募轮次与并购交易

ASX 上没有直接可比的纯 AI 工厂型 neocloud 上市公司。最相关的四个参照点是 CoreWeave(NASDAQ: CRWV)、Lambda Labs(私有,计划 2026 年 H2 IPO)、NEXTDC(ASX: NXT),以及 2025–2026 年创下交易价值纪录的 AI 和超大规模数据中心并购市场。 CoreWeave 是主要基准。CoreWeave 2025 年 3 月以每股 USD $40 IPO 时,按 2025 年 Q1 年化收入约 USD $3.9 billion 计算,隐含市值约 USD $19–27 billion。到 2026 年 5 月,CoreWeave 报告 2026 年 Q1 收入 USD $2.08 billion,重申 2026 全年指引 USD $12–13 billion,在 $2 billion Nvidia 股权投资之外又获得一笔新的 $8.5 billion 无追索、延迟提款定期贷款,并以 8.2x 到 10.6x 的隐含 EV/2026E 收入倍数交易(MarketScreener:8.26x;Multiples.vc:2026 年 6 月为 10.6x)。CoreWeave 截至 2026 年 Q1 末积压订单达到 USD $99.4 billion,提供了 Firmus 尚未公开证明的可观合同收入可见度。但 CoreWeave 同时背负约 USD $25–28 billion 的债务和股权支持工具,GAAP 亏损扩大(2026 年 Q1 尽管收入 $2.08 billion、调整后 EBITDA $1.16 billion,仍净亏损 $740 million),且交易估值仍处高溢价。CoreWeave 模型是 Firmus 最有参考意义的类比:它证明 AI 云基础设施需求真实且可资本化,但也展示了高杠杆 GPU 云运营商在调整后 EBITDA 与 GAAP 净利润之间的结构性盈利缺口。 Lambda Labs 已聘请投行为 2026 年 H2 美国 IPO 做准备。截至 2026 年 5 月,它的隐含估值约 USD $9–9.1 billion,按估计 FY2025 收入 USD $760 million 计算,约为 12x 远期收入。Lambda 面向开发者的定位不同于 Firmus 的 factory 模型,但它的收入倍数为 IPO 前私募可比分析提供了有用上限。 NEXTDC(ASX: NXT)截至 2026 年 6 月的追踪 EV/EBITDA 倍数约为 56–74x,企业价值 AUD $12–14 billion,受合同容量增长 60% 至 667 MW 以及 OpenAI 园区合作推动。NEXTDC 的高倍数反映了 AI 基础设施的溢价定价,但锚定在真实 EBITDA 上,而不是收入前承诺——这与 Firmus IPO 所需的估值基础在结构上不同。 AI 和超大规模数据中心并购市场自 2024 年初以来累计交易价值达到 USD $151 billion,其中 84% 由私募股权资金支持。已运营、已预租、可用于 AI 的资产可取得 6–10x EV/revenue 和 16–20x EV/EBITDA。鉴于 Firmus 仍处收入前状态,它应锚定在可比组低端;只有招股书披露合同收入后,才可能进入 CoreWeave 或 Lambda 的可比区间。 将 CoreWeave 的 8–10x 2026E 收入倍数套用于 Firmus 的 AUD $8–12 billion IPO 目标,意味着所需收入为 AUD $800 million 到 AUD $1.5 billion。截至运行日期,没有公开证据支持这一水平。若估值为 AUD $6 billion,只要一部分合同 GPU 管线按典型 GPU 云费率转化为收入,估值即可自洽;若估值为 AUD $8–12 billion,则必须由招股书给出证明。[CV007, CV008, CV009, CV010, CV015, CV016]

可比估值表
可比对象收入 / EBITDA 指标倍数 / 估值状态与 Firmus 的相关性局限
CoreWeave (NASDAQ: CRWV)FY2026E 收入 $12–13B;Q1 2026 收入 $2.08B;调整后 EBITDA 约 56% 利润率;$99.4B 合同积压订单EV/2026E 收入:8.2x–10.6x;EV 约 USD $95–104B;Q1 2026 净亏损 $740M公开上市(2025 年 3 月 IPO)最好的结构性类比:GPU 云 neocloud;AI 工厂模式;Nvidia 合作;高杠杆资本结构每季度收入超过 $2B;120,000+ 客户合同;相比之下 Firmus 尚无收入;美国 / 欧洲地域;$25–28B 债务
Lambda Labs(私营)估计 FY2025 收入 $760M;收入同比增长约 79%;IPO 前可转票据融资 $350M,折价 20%EV/收入约 12x,对应 $9–9.1B 隐含估值(二级市场,2026 年 5 月)后期私营;目标 2026 年 H2 IPO;已聘 Morgan Stanley、JPMorgan、Citi开发者 GPU 云纯玩家;Nvidia 背书;美国市场;IPO 前轨迹与 Firmus 类似无 APAC / 主权角度;开发者 vs 企业模式;GPU 集群更小;美国地域收入
NEXTDC (ASX: NXT)TTM EBITDA 约 AUD $222M;签约容量 667MW(2026 年增长 60%);FY2026 资本开支指引 AUD $3BEV/EBITDA 56–74x(追踪期);EV AUD $12–14B;市值 AUD $12–14B公开上市(ASX);本地市场可比对象最接近的 ASX 上市基础设施同行;澳大利亚主权 AI 数据中心运营商;OpenAI 园区合作;争夺同一批客户传统数据中心运营商(并非 GPU 原生云);AI 爬坡正在推进但商业模式不同;估值锚定 EBITDA,而 Firmus 尚无收入
CoreWeave IPO(2025 年 3 月交易)Q1 2025 收入 $981.6M(年化约 $3.9B);FY2025 收入 $5.13B(实际)IPO 价格 $40/股 → EV 约 $19–27B;上市时隐含 FY2025E 收入倍数 15–20xM&A / IPO 交易最近一次 GPU 云 IPO;为 AI 基础设施公开发行设定市场先例;直接类比 Firmus 将在 ASX 面对的环境CoreWeave IPO 时季度收入已超过 $981M,Firmus 尚无收入;美国公开市场 vs ASX;投资者基础和流动性不同
AI / 超大规模数据中心 M&A(2025–2026 年交易)EV/收入 6–10x;EV/EBITDA 16–20x+;自 2024 年 Q1 以来合计交易价值 $151B创纪录交易背景:Aligned DC $40B 收购;84% 交易价值来自 PE / 基础设施基金;PE 基础设施基金 2025 年募资 $250B+持续进行的 M&A 市场退出 / 下行情景参考;为已运营、AI 就绪基础设施资产设定底部倍数这些是已有收入和 EBITDA 的运营资产;Firmus 尚无收入;M&A 倍数可能没有反映无收入风险溢价
Firmus(IPO 前私募标记)收入:未披露;Blackstone $10B 融资安排以合同为条件;2 份超大规模云合同已确认投后估值:USD $5.5B(AUD $7.95B)2026 年 4 月;IPO 目标 AUD $8–12B;隐含 EV/收入:未定义(未披露收入)IPO 前私营公司;目标 2026 年 6–7 月 IPO标的公司;列示用于与可比组对照自我参照;尚无收入;无经审计财务;若无招股说明书收入披露,无法计算隐含倍数

除另有说明外,收入倍数均按 2026E 口径列示。所有 USD/AUD 数字在适用处按约 0.65 汇率换算。NEXTDC EV/EBITDA 反映追踪期数字,以及截至 2026 年 6 月的 AI 基础设施需求溢价环境。CoreWeave IPO 倍数基于 IPO 定价,而非当前交易价格。AI / 超大规模数据中心 M&A 区间反映 EV 可估算的交易;并非所有交易都披露财务数据。

[CV008, CV009, CV010, CV015, CV016, CV017]
FV002: 估值敏感性:按收入倍数和情景收入推算的 Firmus EV(AUD B)

按四档 EV / 收入倍数和三组收入情景测算 Firmus 企业价值,显示在 CoreWeave 可比倍数下,AUD $8–12B 的 IPO 目标隐含 AUD $800M–$1.5B 收入。

所有数字仅作示意,由作者测算,并以 AUD 十亿计。收入情景($400M、$800M、$1.2B、$1.5B)覆盖 18,400–55,000 块 GB300 GPU 在典型 neocloud GPU 小时费率下可对应的区间。EV / 收入倍数以 CoreWeave(2026E 8–10x)和 Lambda Labs(约 12x)为基准。Firmus 尚未披露任何收入数字。

[CV007, CV008, CV028]

8.4 乐观、基准与悲观情景分析

三个情景框定了 2026–2029 年 Firmus 投资案例的区间。 乐观情景需要五件事按顺序发生:ASX IPO 在 2026 年 H2 以 AUD $10–12 billion 定价并超额认购完成;招股书披露至少一个超大规模客户身份和收入规模;墨尔本和塔斯马尼亚的 GB300 GPU 部署按目标在 2026 年 H2 上线;Blackstone 工具开始基于合同收入提款;AI 训练和推理需求到 2027 年仍受供给约束,阻止 GPU 价格正常化。在这个情景中,Firmus 2027–2028 年收入可能达到 AUD $1.5–3 billion,股价可能反映 10–15x 远期收入倍数,意味着到 2028 年企业价值达到 AUD $15–20 billion。该情景概率低到中等:前三个条件单独看都可行,但五个条件在 2026–2027 年窗口内顺序发生,取决于 AI 需求延续、合同披露意愿和无瑕疵运营执行。 基准情景假设 IPO 在 2026 年 H2 以 AUD $7–9 billion 定价,招股书部分披露合同,收入到 2027 年底爬坡至 AUD $600 million–$1.2 billion。在基准假设下,随着 IPO 后增长阶段成熟、债务服务吸收大量经营现金流,倍数压缩至 6–8x;2028–2029 年企业价值 AUD $7–12 billion,意味着 IPO 投资者回报持平到温和。鉴于当前证据基础,这是最可能的情景,也反映了资本密集型、收入前基础设施运营商的市场一致假设。 悲观情景由以下任何一项触发:IPO 因市场条件不利、治理挑战或 ASIC 审查推迟至 2027 年 Q1 之后;一个或两个超大规模客户行使取消或重谈条款(条款未公开);AI 算力需求因效率提升、GPU 供给过剩或超大规模云厂商资本配置回撤而走弱;Blackstone $10 billion 工具的利息负担在确认足够收入前变得不可持续;或独立审计证实建设成本效率主张被重大夸大。在悲观情景中,Firmus 股权价值可能降至 AUD $3–6 billion(相对 AUD $8 billion IPO 价格下跌 25–65%),并可能出现债务契约压力。鉴于未解决假设数量很多,悲观情景具备中等概率。[CV032, CV033, CV034, CV035]

乐观 / 基准 / 悲观情景分析
情景关键假设估值逻辑关键风险概率信号
乐观2026 年 H2 以 AUD $10–12B 估值 IPO;合同身份在招股说明书中披露;2026 年 H2 有 200MW+ 上线;2027–2028 年收入达到 AUD $1.5–3B;AI 需求仍受供给约束;Blackstone 提款顺利推进。AUD $1.5–3B 的远期收入按 10–15x 计算,意味着 2028–2029 年 EV 为 AUD $15–20B;IPO 投资者以 AUD $10B 入场可获得 50–100% 回报。CoreWeave $99.4B 积压订单作为结构性可比,验证需求轨迹。需要五个条件同时成立:IPO 定价在区间高端、客户披露、运营爬坡、债务提款、需求持续——复合概率低。低到中等。单项条件各自有中到高概率,但联合概率更低;AI 效率提升或 GPU 供应正常化是主要下行催化剂。
基准2026 年 H2 以 AUD $7–9B 估值 IPO;合同部分披露;到 2027 年底收入爬升至 AUD $600M–$1.2B;增长正常化后倍数压缩至 6–8x;偿债吃掉大量自由现金流。AUD $600M–$1.2B 收入按 6–8x 计算,意味着 2027 年 EV 为 AUD $3.6–9.6B;IPO 投资者以 AUD $8B 入场,回报持平到小幅正收益。全额提款后利息费用(估计 $700–900M/年)吸收现金流,净回报被压缩。市场成熟后倍数压缩;治理悬置问题延续;客户披露延迟;GPU 部署慢于计划。以现有证据看,这是最可能情景;也符合资本密集、IPO 时尚无收入的基础设施运营商市场先例。
悲观IPO 延迟至 2027 年 Q1+,或定价低于 AUD $6B;主要合同取消或重谈;AI 需求走弱;建设成本效率主张被证实夸大;尚无收入时偿债已不可持续。AUD $300–$600M 实际收入按压缩后的 3–5x 收入倍数计算,EV 为 AUD $1.5–3B,较 AUD $8B IPO 入场估值回撤 62–81%。Blackstone 融资安排可能承受契约压力。下轮估值下调风险实质存在。无托管造成首日抛压;治理事件可能触发机构抵制;GPU 过剩事件;监管干预塔斯马尼亚水电分配。中等概率;任何单一触发因素(无托管暴跌、合同取消、AI 需求走弱)本身都合理可能。尚无收入 + 极高杠杆 + 无托管共同造成复合下行凸性。

所有情景估值均为作者基于可比分析(CoreWeave、Lambda、NEXTDC、AI 数据中心 M&A 先例)和已披露资本结构得出的示意区间。它们不是财务建议,也不反映任何招股说明书指引;截至 2026 年 6 月 18 日,招股说明书尚未发布。

[CV032, CV033, CV034, CV035, CV009, CV007]
FV003: 估值 / 回报区间:Firmus 与可比公司的退出价值(低 / 基准 / 高)

对 Firmus IPO 和 2028 年退出给出低、基准、高三档估值,并列示可比公司参照点,说明收入披露和市场条件不同会把结果推向多大区间。

Firmus IPO 区间基于投资者往来材料和媒体报道(低位 = 最近一轮私人市场估值;基准 = 目标底线;高位 = 报道中的目标)。Firmus 2028 年退出估值由作者按牛市 / 基准 / 熊市情景推导,不是招股书指引。CoreWeave EV 来自 MarketScreener 和 Multiples.vc 2026 年 6 月数据。Lambda Labs 估值来自 Forge Global 2026 年 5 月二级市场数据。NEXTDC 区间来自分析师估计。所有 AUD / USD 换算约按 0.65。

[CV002, CV008, CV015, CV016, CV033, CV034]

8.5 退出准备度、论点破裂触发点与最终尽调问题

Firmus 在结构性流程维度上的退出准备度较高,在透明度维度上较低。结构上,ASX 上市基础设施已经到位:已任命簿记行(Bank of America、JPMorgan、Morgans Financial 和 Morgan Stanley);已开展全球路演;ASX 在 2026 年 2 月澄清框架后,于 2026 年 3 月通过了对 Oliver Curtis 的 Good Fame and Character 评估。Blackstone 工具提供了机构可信度。董事会构成也有所加强,2026 年 3 月任命了三名新的独立非执行董事,包括 Lee Hatton(审计与风险委员会主席)和 Christine Bartlett(薪酬委员会主席)。透明度方面,截至运行日期(2026 年 6 月 18 日),尚无招股书提交 ASIC;收入、毛利率和现金消耗率均未公开披露;两个基石超大规模客户身份仍未披露;运营效率主张也没有独立审计。 缺乏 IPO 锁定或托管,是最可立即行动的结构性风险。按这个规模的科技上市惯例,创始人、管理层和主要 IPO 前机构股东通常应有 90–180 天托管。报道称,无托管结构让新进入的 IPO 投资者暴露在已实现 300–530% 账面收益的持有人抛压之下。Ellerston Capital(约 7%)、Regal Funds、Wilson Asset Management、Paradice、UniSuper 等关键 IPO 前机构,都可能从上市第一天开始卖出。Oliver Curtis 本人按最后一轮标记估计持有价值 AUD $1.25 billion 的股权。Maas Group 于 2026 年 2 月取得 AUD $100 million 股份,资金来自出售其建材部门所得——这笔头寸上市时也可立即变现。 TV005 中列出的论点破裂触发点,设计上用于在簿记和上市后监测。它们直接对应 TV006 中的六项最终尽调问题。其中两项是结构性阻断项——招股书收入披露和托管结构——应在任何承诺前解决。其余四项(合同条款披露、独立效率审计、建设成本验证和债务服务覆盖率)也很重要,但有可能在 IPO 结束前 30 天内解决。 基于当前证据,合适的投资者动作是 TRACK:跟踪 IPO 进程,招股书提交后审阅,并在收入、客户身份和托管结构得到确认后,重新评估买入 / 回避决策。若投资者短期内需要 AI 基础设施敞口,按当前 ASX 倍数看,NEXTDC 可能是风险较低的主权 APAC 代理标的。CoreWeave 以 8–10x 2026 年远期收入提供了流动性公开基准;按同样倍数,Firmus 招股书必须披露 AUD $800 million 到 AUD $1.5 billion 年收入,BUY 才站得住。[CV037, CV038, CV039, CV040, CV041, CV042]

投资逻辑中断与否决触发因素
触发因素阈值 / 信号对投资逻辑的传导行动含义
招股说明书确认无 IPO 托管招股说明书提交后,对 IPO 前持有人没有锁定或托管;主要股东没有承诺上市后继续持股。持有 300–530% 账面收益的股东带来首日抛压;散户投资者承接卖盘;IPO 价格发现受损;观察降级为回避。不参与 IPO 建簿;等待上市后 90 天价格稳定再重新评估。监测前 30 天每日成交量。
招股说明书未披露收入招股说明书提交的经审计财务显示追踪收入 <AUD $200M,或没有收入披露。无法验证任何收入倍数;定价完全依赖信念;可比框架失效;若估值 AUD $8B+,建议降级为回避。放弃 IPO 配售;在 AUD $5B 设置价格提醒(按 $2B+ 合同最低披露价值对应 CoreWeave 等效水平)。
重大合同取消或重谈Project Southgate 任一超大规模云合同取消或重大重谈;Blackstone 提款暂停;Firmus 在 IPO 后公布修订收入或资本开支指引。收入底部坍塌;触发 Blackstone 融资契约;下轮估值下调风险急剧上升;悲观情景启动。卖出或回避;每季度监测 ASX 披露和 Blackstone 提款状态。标记进入减值复核。
AI 需求走弱 / GPU 过剩事件前五大超大规模云厂商 2027 年资本开支指引同比下调 >20%;GPU 现货价格下跌 >30%;CoreWeave 或 Lambda 指引下调 >25%。整个 GPU 云板块倍数压缩;Firmus 失去合同授予顺风;$10B 债务融资安排的契约条件可能收紧。降低仓位;比较 Firmus 合同费率与市场 GPU 费率;评估支持 Blackstone 融资安排的 GPU 抵押品减值风险。
建设成本效率主张被证实严重夸大独立审计发现 PUE >1.15,或每 MW 建设成本 >USD $9M(Firmus $6M 与 NEXTDC CEO 引用 $12.4M 的中点);由 Uptime Institute 或同等机构发布。能源成本优势收窄;每兆瓦资本效率边际收窄;EBITDA 预测假设需要下修;护城河实质削弱。目标估值下调 15–25%;要求管理层提供修订版运营模型;从基础设施溢价重新归类为商品化建设。

触发因素定义为可观察、公开市场可获取的信号。阈值由作者基于可比分析和商业模式逻辑得出。行动含义是示意性的组合管理指引,不构成财务建议。

[CV005, CV006, CV013, CV032, CV035, CV039]
最终尽调问题清单
主题缺失证据重要性负责人 / 尽调路径
收入和财务报表未公开披露经审计或管理层编制的利润表、毛利率、运营烧钱速度或收入确认政策。没有收入,就无法做可比倍数分析;整个估值框架都建立在代理证据上。这是最重要的单项尽调事项。IPO 招股说明书提交 ASIC 后,审阅其中的财务报表;对标 CoreWeave 56% 调整后 EBITDA 和可比 neocloud 运营商。负责人:ASIC 提交文件(公开)。
IPO 锁定和托管结构IPO 前股东锁定、托管或股份出售限制条款没有正式招股说明书披露。没有托管,持有 300–530% 账面收益的股东可从第一天起出售,压制上市后价格;这是散户 IPO 投资者特有的结构性风险。审阅招股说明书中关于股份出售限制的章节和托管契约(如有)。负责人:IPO 法律顾问;预计在招股说明书中确认。
超大规模客户身份和合同条款Project Southgate 两个基石客户均未具名;合同财务条款、最低承诺和取消条款未进入公开记录。缺少客户身份和信用质量,Blackstone 融资覆盖率和收入耐久度无法独立评估。客户集中度风险无从判断。要求提供招股书合同附表或脱敏条款清单;也可在 IPO 时通过 ASX 披露确认客户身份。负责人:Firmus 投资者关系。
独立 PUE 和效率审计现有设施在代表性负载下没有经过独立第三方审计的 PUE 测量;2021 年单点测量已因不具代表性被撤回。如果实际运营 PUE 是 1.15–1.25 而非 1.03,能源成本优势估计会在运营成本线上收窄 15–25%,并实质影响 EBITDA 预测。要求 Uptime Institute 或 TÜV Rheinland 对满载 Melbourne AI Factory 出具符合 ASHRAE 的 PUE 审计。负责人:Firmus 技术团队;预计在 IPO 或上市后提供。
债务偿付覆盖和融资提款条件Blackstone 融资的利率、摊还安排、契约包以及首次提款对应的已签约收入门槛均未公开披露。按 7–9% 的私募基础设施债利率测算,$10B 债务每年利息支出可达 $700–900M,可能超过早期运营现金流;触发契约违约是实质下行情景。审阅招股书借款附注;向簿记行索取融资摘要。将首次提款门槛与已披露合同价值对标。
IPO 融资中新股与老股出售拆分AUD $2 billion IPO 融资尚未拆分为主要募集资金(进入 Firmus)和次级出售资金(出售股东套现)。如果次级出售占比较大,资产负债表受益会相应变小;也意味着上市首日来自出售股东的抛压更高。在招股书封面页和募集资金用途章节确认。预计 ASIC 备案招股书会披露。

截至 2026 年 6 月 18 日(招股书发布前),所有尽调事项仍未关闭。第 1、2 项是结构性阻断点,任何投资承诺前都应解决。第 3–6 项属于重大事项,预计会在招股书中部分解决。

[CV003, CV005, CV006, CV013, CV018, CV019]
FV004: 投资 KPI 记分卡:八个维度的投委会就绪度评估

按八个投委会维度结构化打分;加权结果支持继续跟踪,并明确了招股书发布后上调至买入的条件。

[CV001, CV005, CV009, CV019, CV025, CV027]

8.6 展示材料

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验。

证据索引

结论
编号陈述可信度来源
CO001 Firmus was founded in 2019. SO006, SO007, SO011, SO017
CO002 Firmus was originally focused on developing cooling technology for bitcoin mining operations. SO006, SO007, SO011, SO020
CO003 Firmus Grid Pty Ltd was incorporated in Australia on December 12, 2019 (CIN: 638040534). SO017
CO004 Firmus Technologies Pty Ltd was incorporated in Australia on November 21, 2021 (CIN: 655498105). SO017
CO005 Oliver Curtis is a co-founder and co-CEO of Firmus Technologies. SO006, SO007, SO015, SO020
CO006 Tim Rosenfield is a co-founder and co-CEO of Firmus Technologies. SO006, SO007, SO015, SO020
CO007 Jonathan Levee is a co-founder and Director of Research and Development at Firmus Technologies. SO006, SO007, SO020
CO008 Oliver Curtis and Tim Rosenfield are cousins. SO015, SO020, SO027
CO009 Jonathan Levee is described in reporting as a former brother-in-law of Oliver Curtis. SO027
CO010 Tim Rosenfield was previously CEO of lingerie company Simone Perele before co-founding Firmus. SO020
CO011 Oliver Curtis was convicted by a Supreme Court jury of conspiring to commit insider trading and sentenced to two years' imprisonment on 21 June 2016. SO014, SO020
CO012 Curtis was ordered to be released after serving one year; he walked out of Cooma Correctional Centre on 23 June 2017. SO014, SO015
CO013 The insider trading conspiracy involved 45 CFD trades between May 2007 and June 2008 generating a total net profit of A$1,432,228.85. SO014, SO015
CO014 Oliver Curtis initially invested A$250,000 in Firmus at founding. SO007, SO015, SO020
CO015 Curtis's initial A$250,000 stake in Firmus was valued at approximately A$81 million by 2024. SO007, SO015
CO016 Firmus pivoted from bitcoin-mining cooling technology to AI data-centre (AI Factory) infrastructure. SO001, SO006, SO011
CO017 Firmus's operational headquarters is in Singapore; Oliver Curtis relocated to Singapore in 2023. SO007, SO015, SO020
CO018 Firmus's registered entity Firmus Grid Pty Ltd has an address in Saint Leonards, New South Wales, Australia. SO017
CO019 Firmus claims its AI Factories can reduce energy consumption by up to 60% versus conventional air-cooled data centres. SO001, SO006, SO013
CO020 Firmus reported a PUE (Power Usage Effectiveness) of 1.03 from a 2021 immersion-cooling deployment point measurement in Tasmania. SO020, SO023
CO021 A Firmus spokesperson confirmed the 1.03 PUE figure is a point measurement from 2021 and cannot be independently verified, as no recognised certification body for PUE exists. SO020, SO022
CO022 Firmus deployed approximately 4,000 Nvidia GPUs across two ST Telemedia Global Data Centres sites in Singapore from mid-2023, generating revenue from enterprise and government customers. SO021
CO023 Firmus raised A$330 million in a private placement on 16 September 2025 at a post-money valuation of A$1.85 billion. SO007, SO011, SO018
CO024 Nvidia participated as a strategic equity investor in Firmus's September 2025 A$330 million funding round. SO007, SO011, SO013
CO025 Ellerston Capital provided a cornerstone investment in the September 2025 round; investment director David Leslie joined the Firmus board. SO007, SO011, SO020
CO026 Firmus raised A$500 million at a post-money valuation of approximately A$6 billion on 14 November 2025. SO004, SO006, SO023, SO024
CO027 Firmus closed a US$10 billion debt financing facility led by Blackstone Tactical Opportunities and Blackstone Credit & Insurance on 9 February 2026. SO003, SO013
CO028 Firmus expects to raise US$505 million led by Coatue with Nvidia participation at a post-money valuation of US$5.5 billion (announced 6 April 2026). SO002, SO013
CO029 Total equity raised by Firmus in the six months through April 2026 is approximately US$1.35 billion across three rounds. SO002, SO013
CO030 Project Southgate targets 1.6 GW of AI Factory capacity across five Australian sites by 2028. SO004, SO008, SO013
CO031 The total estimated investment in Project Southgate announced in October 2025 is A$73.3 billion in partnership with CDC Data Centres and Nvidia. SO008, SO012, SO013
CO032 Grant Dempsey was appointed incoming board chair and Kirsty Godfrey-Billy was appointed CFO of Firmus in late September 2025. SO024
CO033 Lee Hatton, Christine Bartlett, and Julie Shuttleworth AM were appointed as Non-Executive Directors of Firmus Grid Limited, effective March 2026. SO005, SO025
CO034 The ASX Listings Compliance department published a compliance update dedicated to the Good Fame and Character test on 23 February 2026, widely understood to be connected to the planned Firmus IPO. SO026
CO035 Oliver Curtis was reported to have been cleared to serve as co-CEO of a publicly listed company under the ASX Good Fame and Character test, as of March 2026. SO015, SO019
CO036 Firmus is targeting an ASX IPO in June or July 2026, seeking to raise approximately US$2 billion (A$2–3 billion) in new capital. SO013, SO027
CO037 IPO advisers for Firmus include Bank of America, JPMorgan, Morgans Financial, and Morgan Stanley. SO013, SO023
CO038 Reports indicate that Firmus's ASX IPO has been pitched with no escrow or share-sale restrictions for pre-IPO shareholders. SO022, SO027
CO039 Firmus received the DCD Asia-Pacific Data Center Project of the Year award in 2024 for its Singapore AI Factory. SO001, SO021
CO040 SemiAnalysis ranked Firmus in its ClusterMAX Top 3 Global GPU Cloud in both 2024 and 2025. SO001, SO016
CO041 Firmus achieved MLPerf V4.0 Certified Training and Power benchmark results, claiming industry-leading performance. SO001, SO016
CO042 Maas Group Holdings (ASX: MGH) invested A$100 million in Firmus in approximately February 2026, acquiring approximately 1.7% equity; Maas subsidiary JLE will form part of the Firmus supply chain. SO009
CO043 Nick Curtis (Oliver Curtis's father, a former mining executive) served as inaugural chair of the Firmus board. SO020
CO044 Former Telstra executive Ted Pretty succeeded Nick Curtis as Firmus board chair and helped pivot the company to AI data centres; Pretty subsequently stepped down from the board prior to 2025. SO020
CO045 Roger Montgomery Investment Management argued that Nvidia's small equity investment (estimated A$30–75M) unlocked approximately A$830 million in follow-on equity and US$10 billion in debt, most of which returns to Nvidia as hardware revenue. SO022
CO046 UNSW senior research associate Dr Amr Omar described Firmus's claimed PUE of 1.03 as 'at the extreme end of what is now possible anywhere in the world'; global hyperscalers typically achieve approximately 1.2 PUE. SO020, SO022
CO047 Firmus's November 2025 raise was announced at an AUD post-money valuation of A$6 billion, while the April 2026 Coatue-led round was announced at a USD post-money valuation of US$5.5 billion; media coverage citing both produced an apparent conflict (US$6B vs US$5.5B) largely explained by currency conversion. SO004, SO002, SO013
CO048 James Packer is reported to have purchased approximately A$50 million in Firmus shares alongside Nvidia at approximately A$145 per share. SO027
CO049 Firmus has two Australian legal entities: Firmus Grid Pty Ltd (incorporated Dec 2019, CIN 638040534) and Firmus Technologies Pty Ltd (incorporated Nov 2021, CIN 655498105). SO017
CO050 Jonathan Levee had middle management experience in the resources sector prior to co-founding Firmus. SO020
CO051 Regal Funds Management (Phil King) invested approximately A$10 million in Firmus via a convertible note circa 2021, when the company was still primarily a bitcoin-mining cooling provider. SO027
CO052 Oliver Curtis's initial A$250,000 Firmus stake was worth an estimated A$1.25–1.29 billion as of May 2026, as reported by the AFR Rich List. SO015, SO027
CO053 Ellerston Capital's stake in Firmus is understood to approach 7 per cent of the equity. SO027
CO054 Synert, Firmus's grid-integration software, was deployed with Eaton in Tasmania in 2024 Q3 as the first live grid-integration project for the AI Factory platform. SO016
CO055 The Tasmanian state government established the world-first AI Factory Zone in Northern Tasmania in July 2025, with Firmus as the anchor tenant. SO007, SO008
CO056 Firmus signed a multi-billion-dollar deal for a Melbourne AI Factory involving 18,400 Nvidia GB300 GPUs in March 2026. SO016, SO019
CM001 Firmus's market boundary spans three overlapping categories: hyperscale GPU compute for contracted AI factory tenants, sovereign AI infrastructure for government and regulated industries requiring on-shore data residency, and enterprise AI cloud services delivered through the Firmus AI Cloud. SM018, SM022
CM002 Firmus explicitly excludes general-purpose IaaS, traditional colocation, and non-AI compute workloads from its target market; its addressable opportunity is defined solely by AI-optimised, GPU-dense, energy-efficient infrastructure. SM018, SM021
CM003 Status-quo substitutes for Firmus customers include global hyperscaler cloud platforms (AWS, Azure, GCP), on-premises GPU cluster builds, and conventional colocation operators such as NEXTDC, Equinix, and AirTrunk—none of which offer sovereign-Australian-soil, renewable-powered AI factory economics at the scale Firmus is targeting. SM005, SM021
CM004 Project Southgate's total planned investment is up to A$73.3 billion across five Australian sites (Tasmania, Melbourne, Sydney, Perth, and Canberra), targeting 1.6 GW of AI compute capacity by 2028. SM018, SM006
CM005 Firmus's first two sites under construction—Tasmania and Melbourne—represent a combined 150 MW of AI compute capacity, delivering approximately 54,000 NVIDIA GB300 GPUs by mid-2026. SM018, SM006
CM006 Firmus targets three distinct buyer segments: hyperscale cloud and AI companies requiring contracted off-take capacity; government, enterprise, education, and regulated industries requiring sovereign access; and AI-native startups and scaleups needing on-demand GPU access. SM018, SM007
CM007 Firmus had raised approximately USD 1.35 billion in equity across three rounds in a six-month period to February-April 2026, at a post-money valuation of USD 5.5 billion, and secured a USD 10 billion debt facility led by Blackstone. SM003, SM005
CM008 Grand View Research estimates the global AI data center market at USD 180.6 billion in 2026, growing at a 23.9 percent CAGR to USD 810.6 billion by 2033. SM019
CM009 MarketsandMarkets estimates the global AI data center market at USD 344.24 billion in 2025 and USD 471.59 billion in 2026—more than 2.5 times larger than Grand View Research's 2026 estimate—reflecting a broader boundary definition that includes the full AI infrastructure stack co-deployed with AI workloads. SM019
CM010 Gartner forecasts worldwide sovereign cloud IaaS spending will total USD 80.4 billion in 2026, a 35.6 percent year-on-year increase from USD 59.3 billion in 2025. SM002, SM017
CM011 Gartner projects Mature Asia/Pacific sovereign cloud IaaS spending at USD 1.593 billion in 2026, an 87 percent year-on-year increase, making it the second-fastest growing region globally after Middle East and Africa (89 percent). SM002, SM015
CM012 Data Bridge Market Research estimates the global sovereign AI infrastructure market at USD 14.82 billion in 2025, projecting it to reach USD 49.67 billion by 2033 at a 16.3 percent CAGR—a significantly narrower estimate than broader sovereign cloud figures, reflecting a focus on dedicated AI hardware under sovereign governance. SM015
CM013 Asia-Pacific is the fastest-growing region for sovereign AI infrastructure at a 17.8 percent CAGR from 2026 to 2033, driven by rising government AI initiatives and domestic semiconductor and compute investment. SM015
CM014 The broader sovereign cloud market—covering all deployment modes including IaaS, PaaS, SaaS, and managed services—is valued at USD 128.6 billion in 2026 and projected to reach USD 298.5 billion by 2030 at a 23.4 percent CAGR (Research and Markets), representing a figure approximately eight times larger than the sovereign AI infrastructure-specific estimate. SM017
CM015 The Asia-Pacific data center GPU market is projected at USD 22.15 billion in 2026, growing at a 16.3 percent CAGR to USD 47.14 billion by 2031, with the AI inference segment growing at 17.24 percent CAGR. SM012
CM016 The Australia data center market is projected to grow from USD 4.22 billion in 2025 to USD 9.02 billion by 2031 at a 13.5 percent CAGR; 145 operational colocation data centers exist in Australia as of 2026. SM016
CM017 The five largest US cloud and AI infrastructure providers (Amazon, Alphabet, Microsoft, Meta, Oracle) collectively plan to spend USD 660–690 billion on capital expenditure in 2026—nearly doubling 2025 levels—and all self-report being supply-constrained rather than demand-constrained. SM011, SM023
CM018 The top four hyperscalers (Amazon, Google, Microsoft, Meta) have guided 2026 capex to approximately USD 715 billion; in Q1 2026 alone, Alphabet, Amazon, and Microsoft collectively spent USD 112 billion on infrastructure—more than three times the inflation-adjusted cost of the Manhattan Project in 90 days. SM023, SM011
CM019 Firmus signed a multi-year, multi-billion dollar contract with an undisclosed leading global technology company for approximately 18,400 NVIDIA GB300 GPUs at its Melbourne facility, described as the second large-scale customer for Project Southgate. SM007, SM014
CM020 The Melbourne hyperscale contract represents the first hyperscale AI factory in Australia, positioning Australia on a path to become a producer and exporter of AI tokens globally. SM007, SM006
CM021 The Firmus AI Cloud is available to enterprise, education, government, startup, and scaleup customers in Australia, providing managed access to AI compute from the Melbourne site with 18,500 NVIDIA GB300 GPUs expected online by April 2026. SM018, SM006
CM022 Gartner identifies governments as the primary buyers of sovereign cloud IaaS globally, followed by regulated industries and critical infrastructure operators such as energy, utilities, and telecommunications. SM002, SM015
CM023 Government and public sector accounted for 33.26 percent of the global sovereign AI infrastructure market in 2025, the largest single end-user segment. SM015
CM024 Defence and intelligence is the fastest-growing end-user segment in the global sovereign AI infrastructure market, with a 17.6 percent CAGR from 2026 to 2033, driven by AI-powered surveillance, intelligence analysis, and autonomous systems. SM015
CM025 DGX Cloud (NVIDIA's managed cloud platform) established a new region at Southgate Melbourne in partnership with Firmus, giving enterprises sovereign access to NVIDIA AI compute on Australian soil for the first time. SM018, SM006
CM026 Australia's combination of renewable energy resources, political stability, English-speaking workforce, and geographic proximity to the Asia-Pacific market has attracted significant AI infrastructure capital from international investors. SM005, SM004
CM027 The Tasmanian site operates on a grid overwhelmingly powered by hydroelectric generation, enabling Firmus to claim a PUE of approximately 1.03 and a low-carbon compute footprint; liquid-cooling technology is claimed to reduce energy consumption by up to 60 percent versus conventional air-cooled facilities. SM005, SM018
CM028 Australian data centre spending grew more than 60 percent in the September 2025 quarter, reaching A$2.6 billion—an increase of more than 140 percent on the previous year, according to Australian Bureau of Statistics data cited by Greenpeace. SM008
CM029 The AEMC proposed in March 2026 that data centres of 30 MW or greater connecting to the National Electricity Market must comply with voltage and frequency ride-through standards, adding engineering compliance costs and connection-timeline obligations to new facilities. SM009, SM013
CM030 AEMO estimates Australian data centres consumed 3.9 TWh of electricity in FY2025 (approximately 2 percent of NEM demand) and forecasts this growing to 12.0 TWh by FY2030 under a Step Change scenario (approximately 6 percent of NEM demand). SM013, SM008
CM031 Under a high-growth scenario, AEMO projects data centres could account for 13 percent of Australia's total national electricity demand by 2040—comparable to the combined electricity demand for electrification of all Australian cars and homes. SM008, SM010
CM032 Transgrid, NSW's transmission network provider, told a state parliamentary inquiry in May 2026 that existing regulatory arrangements were not designed for the current level of large, clustered data centre load growth, and that it does not have enough capacity to power every data centre currently in the planning pipeline. SM008, SM009
CM033 Greenpeace Australia Pacific called for an urgent moratorium on new data centre development until governments legislate transparency requirements and safeguards around electricity, water, and emissions use; a majority of federal-state energy ministers in a May 2026 meeting supported requiring data centres to fully offset their power demand with new renewable energy. SM010, SM008
CM034 AEMO's projected data centre demand in 2040 increased more than fivefold between its 2024 and 2025 forecast reports, reflecting a systemic failure of planning assumption models to anticipate the AI-driven buildout—characterised by analysts as 'none of this was forecast'. SM008
CM035 Alphabet reported reducing Gemini serving costs by 78 percent during 2025 through model optimisation, illustrating that AI inference efficiency can improve faster than demand grows—a dynamic that could reduce the aggregate GPU capacity required per unit of AI output if sustained. SM011
CM036 Microsoft disclosed an USD 80 billion backlog of Azure orders that cannot be fulfilled due to power availability constraints, establishing that energy access rather than demand softness is the primary limiting factor in AI infrastructure deployment globally. SM011, SM023
CM037 Industry observers cited by Ventureburn have questioned the scale of Firmus's efficiency claims (PUE 1.03, 60% energy reduction versus air-cooled), noting these have not been independently verified at operational scale. SM021
CM038 Most state energy ministers at the May 2026 federal-state meeting agreed with proposals to require data centres to fully offset their electricity demand and to report energy use and emissions, signalling that mandatory renewable offset requirements could become a planning condition that extends approval timelines for mainland Firmus sites. SM008, SM009
CM039 Firmus committed over A$300 million to its domestic supply chain, creating advanced manufacturing capacity for up to 1.5 GW of AI factory delivery per year with manufacturing partners Benmax and Maas Group. SM007, SM018
CM040 A working SAM estimate for Firmus can be constructed from the APAC sovereign AI factory market: applying the 17.8% CAGR from the Data Bridge $14.82B 2025 global base yields a 2026 figure of approximately $17.4B globally; APAC is estimated at roughly 15–20% of global sovereign AI infrastructure, implying an APAC SAM of approximately $2.6–3.5 billion—with Australia capturing a modest but growing share of this. SM015, SM002
CM041 The 2.5× discrepancy between Grand View Research ($180.6B) and MarketsandMarkets ($471.6B) for the same 2026 AI data center market reflects incompatible scope definitions: GVR counts purpose-built AI compute hardware and colocation revenue, while MnM includes the full IT, networking, and services stack co-deployed in facilities with AI workloads—making the figures non-comparable for sizing Firmus's directly addressable revenue opportunity. SM019, SM011
CM042 Project Southgate is expected to support the development of up to 5.1 GW of new wind, solar, storage, and hydro generation projects across Tasmania, Victoria, NSW, and the ACT through 2028, providing long-term renewable energy demand certainty for new clean energy projects. SM018, SM006
CM043 Gartner estimates that sovereign cloud IaaS spending will shift 20 percent of workloads from global to local cloud providers by 2028, driven by geopatriation projects seeking digital and technological independence from US- and China-based hyperscalers. SM002, SM016
CP001 Firmus AI Cloud platform meets ISO 27001 and SOC-2 requirements, includes encrypted InfiniBand networking and built-in observability. SP002
CP002 SemiAnalysis ClusterMAX rated Firmus's Sustainable Metal Cloud (SMC) as one of the top global GPU cloud platforms in 2025, ranking alongside AWS and above Google Cloud. SP005
CP003 SMC was one of only three GPU cloud providers globally to correctly implement InfiniBand SHARP in-network reduction, per SemiAnalysis ClusterMAX. SP005
CP004 Firmus operates AI FactoryOS, a proprietary operating system integrating data intelligence and control across the physical AI Factory stack. SP001
CP005 Firmus AI Factories use NVIDIA Cloud Partner (NCP) reference design with advanced immersion and liquid cooling as core architecture. SP002, SP011
CP006 Firmus claims its AI Factory platform can be built at approximately US$6 million per megawatt, roughly half the ~US$12.4 million per megawatt that NEXTDC CEO Craig Scroggie cited as the industry standard. SP018, SP007
CP007 Firmus has two confirmed cornerstone customers at Project Southgate: NVIDIA (DGX Cloud anchor tenant) and a second undisclosed leading global technology company. SP004, SP007, SP008
CP008 As at mid-2026, Firmus's contracted portfolio stood at approximately 155 MW backed by 55,000 GPUs, with only approximately 5 MW operational. SP018
CP009 Firmus's contracts carry an average term of 4.8 years on a take-or-pay basis, with a total contracted value of approximately US$8.2 billion from two cornerstone customers. SP018
CP010 Project Southgate targets delivery of 1.6 GW of AI compute capacity across five Australian sites by 2028, at a total planned cost of up to AUD$73.3 billion. SP004, SP012, SP014
CP011 Firmus raised US$505 million in equity led by Coatue, with NVIDIA participation, at a post-money valuation of US$5.5 billion (April 2026), bringing total equity raised over six months to US$1.35 billion. SP006, SP019
CP012 Firmus secured a US$10 billion debt financing facility led by Blackstone Tactical Opportunities and Blackstone Credit & Insurance, supported by Coatue, announced February 9, 2026. SP003, SP012
CP013 CoreWeave reported FY2025 revenue of US$5.13 billion, a 168% year-on-year increase, making it the fastest cloud platform to reach US$5 billion in annual revenue. SP024, SP026
CP014 CoreWeave went public in March 2025 at US$40 per share and traded at approximately a US$43 billion market capitalisation by March 2026. SP024, SP026
CP015 CoreWeave's revenue backlog reached US$66.8 billion at year-end 2025 and grew to US$99.4 billion by Q1 2026, driven by multi-year contracts with Microsoft, OpenAI, and Meta. SP024, SP020
CP016 CoreWeave guided full-year 2026 revenue of US$12–13 billion and 2026 capital expenditure of US$31–35 billion; Q1 2026 revenue was US$2.08 billion. SP020, SP024
CP017 CoreWeave posted a GAAP net loss of US$1.17 billion for FY2025, widening from US$863 million in FY2024, despite adjusted EBITDA of US$3.09 billion. SP024
CP018 Microsoft accounted for approximately 62% of CoreWeave's revenue in FY2024; the company is diversifying toward Meta and OpenAI but remains highly customer-concentrated. SP024, SP020
CP019 Lambda Labs achieved approximately US$760 million in annualised revenue in 2025, up 79% year-on-year, and raised US$1.5 billion in a Series E (November 2025) at a US$5.9 billion valuation. SP025
CP020 Lambda Labs lists NVIDIA H100 instances at approximately US$2.49/hr compared to CoreWeave's approximately US$4.25/hr, competing directly on price in the developer and startup segment. SP025
CP021 NEXTDC announced a A$2.2 billion capital raise (A$1.5B equity + A$700M debt) in April 2026, lifted FY26 capex guidance to A$3 billion, and targets A$1 billion EBITDA by 2030. SP018
CP022 NEXTDC secured its largest-ever customer contract — a 250 MW commitment at its S4 Western Sydney campus, widely attributed to Microsoft — as the trigger for its April 2026 capital raise. SP018
CP023 OpenAI and NEXTDC signed an MoU (December 2025) for a A$7 billion, 550–650 MW hyperscale AI campus at the S7 site in Eastern Creek, Sydney, under OpenAI's first APAC sovereign AI initiative. SP009, SP021, SP022
CP024 NEXTDC CEO Craig Scroggie publicly disputed Firmus's claimed build cost of approximately US$6 million per megawatt during an April 2026 analyst call, citing an industry standard of approximately US$12.4 million per megawatt and questioning whether Firmus facilities can meet enterprise-grade quality. SP018
CP025 CDC Data Centres provides the physical sovereign data centre backbone for Project Southgate, with near-zero water consumption and 100% renewable power as operational benchmarks. SP008, SP014
CP026 Macquarie Data Centres' IC3 Super West facility in Sydney (47 MW, A$350 million Phase 1) is due to open Q3 2026, supporting 150 kW+ per-rack density and integrating Dell AI Factory with NVIDIA GPUs. SP023
CP027 AUCloud (AUCyber Limited) is an Australian-owned sovereign cloud provider focused on IaaS, MSSP, and managed IT for government and defence, with IRAP certification but no GPU-at-scale compute offering. SP010
CP028 Firmus's headline PUE claim of 1.03 was a 2021 point measurement and has not been independently verified; a University of NSW engineer described the figure as at the extreme end of global possibility. SP017
CP029 Roger Montgomery (Montgomery Investment Management) and other Australian fund managers have publicly raised concerns about Firmus's unverified efficiency claims, absence of IPO escrow/lock-up, and rapid valuation markup from A$81M (2024) to A$6B+ in months. SP017
CP030 Nomura's attempt to syndicate a US$120 million, four-year private credit facility for Firmus Metal Singapore stalled in February 2026, with institutional investors citing concerns over US GPU export controls and potential re-export-to-China enforcement risk. SP027
CP031 Oliver Curtis, Firmus co-founder and co-CEO, was convicted of insider trading in 2007–08 and served approximately twelve months in prison; the ASX issued detailed Good Fame and Character guidance in February 2026, widely understood to address his planned directorship on a Firmus IPO entity. SP016, SP017
CP032 Tasmanian Greens MP Tabatha Badger raised questions in state parliament (March 2026) about the opacity of Firmus's power allocation deal with Hydro Tasmania, noting that rival industries including the Boyer Paper Mill were previously denied equivalent grid capacity. SP015
CP033 VAST Data was selected as the foundational AI Operating System data layer for Firmus AI Factories (February 2026), integrated with the NVIDIA NCP reference design for disaggregated high-throughput storage. SP011
CP034 Firmus committed over AUD$300 million to an Australian domestic supply chain, including manufacturing partners Benmax and Maas Group (ASX: MGH), targeting production capacity of up to 1.5 GW of AI facility delivery per year. SP007
CP035 Hyperscaler NVIDIA H100 on-demand pricing in Australian regions: AWS approximately US$6.88/hr, GCP approximately US$10.98/hr, Azure approximately US$12.29/hr — materially higher than specialist neocloud providers at US$1–4/hr. SP026, SP021
CP036 Amazon Web Services announced a A$20 billion, five-year investment in Australian data centre infrastructure; AWS and Azure both have IRAP-assessed services in Australian regions. SP021
CP037 NVIDIA is both an equity investor in Firmus (two confirmed rounds) and the anchor DGX Cloud tenant for Firmus's Southgate Melbourne deployment, though NVIDIA holds similar equity and partnership positions at CoreWeave and Lambda Labs. SP004, SP006, SP008
CP038 Firmus's immersion-cooling technology originated from its founding-era Bitcoin mining operations in Tasmania; the company pivoted from crypto cooling to AI infrastructure between 2019 and 2023. SP027, SP019
CP039 NEXTDC's current-year EBITDA guidance is approximately A$235 million; its management targets A$1 billion EBITDA by 2030, citing contracted demand from Microsoft and OpenAI as the driver. SP018
CP040 Firmus's Hypercube product is deployed in Singapore, India, and Australia; US GPU export-control enforcement concerns focus on the risk that offshore leasing of NVIDIA H100 processing capacity could be used to indirectly serve Chinese entities. SP027
CI001 Firmus generates revenue by selling AI compute capacity via a public GPU/CPU cloud platform with transparent CPU list pricing and via bespoke multi-year enterprise contracts for dedicated AI infrastructure, serving enterprise, hyperscale, government, education, and research customers. SI001, SI022
CI002 Firmus CPU instance list pricing ranges from USD $73/month for the smallest 4-vCPU virtualised instance to USD $4,672/month for the 256-vCPU bare-metal configuration. SI001
CI003 Firmus secured a USD $10 billion senior-secured debt financing facility on February 9, 2026, led by funds managed by Blackstone Tactical Opportunities and Blackstone Credit & Insurance with participation from Coatue. SI006, SI007, SI008
CI004 The Blackstone debt facility is structured to be drawn against executed long-term customer contracts with leading hyperscalers and other blue-chip AI customers, limiting speculative drawdown. SI006, SI009
CI005 Firmus raised USD $505 million in a strategic equity investment led by Coatue Management with participation from Nvidia, announced April 2026. SI002, SI003, SI015
CI006 The April 2026 USD $505 million raise values Firmus at USD $5.5 billion post-money. SI002, SI003
CI007 Total equity raised by Firmus in the six months to April 2026 reached approximately USD $1.35 billion across three equity rounds. SI002, SI015
CI008 Firmus signed a multi-billion dollar, multi-year agreement for approximately 18,400 Nvidia GB300 GPUs at the Melbourne Project Southgate facility on March 2, 2026, with an unnamed leading global technology company. SI010, SI011, SI018, SI023
CI009 The identity of the hyperscale customer in the March 2026 Melbourne contract has not been publicly disclosed in any reviewed source. SI010, SI011
CI010 The Melbourne hyperscale contract is described as the second large-scale customer secured for Project Southgate, implying a first large-scale contract pre-dates March 2026. SI010, SI018
CI011 Firmus plans a USD approximately $2 billion IPO on the Australian Securities Exchange targeting June or July 2026. SI004, SI015, SI025
CI012 IPO bookrunners are Bank of America, JPMorgan, Morgans Financial, and Morgan Stanley, with a non-deal roadshow conducted in April 2026. SI004, SI025
CI013 Total planned Project Southgate capex is AUD $73.3 billion through 2028 for 1.6 GW of AI compute capacity across five Australian sites. SI005, SI019
CI014 The initial Project Southgate build programme — covering the Tasmania campus and the Melbourne facility — has a disclosed value of AUD $4.5 billion. SI005, SI025
CI015 The Tasmania AI factory campus is targeted to host approximately 36,800 Nvidia GB300 GPUs and be completed in late 2026. SI004, SI010
CI016 Firmus claims its liquid-cooling technology reduces energy consumption by up to 60% compared with conventional air-cooled data centers. SI004, SI005
CI017 Firmus claims its modular construction and cooling approach reduces construction costs by approximately 50% compared with traditional data center builds. SI004, SI025
CI018 Firmus's previously stated PUE claim of 1.03 was a point measurement from 2021 and has not been independently verified; the company has since acknowledged this. SI012
CI019 UNSW Professor Amr Omar described Firmus's 1.03 PUE claim as being "at the extreme end of global possibility," noting global hyperscalers typically operate between 1.2 and 1.4 PUE. SI012
CI020 Firmus has committed more than USD $300 million to its Australian domestic supply chain for AI factory components, spanning advanced manufacturing and sovereign electrification. SI010, SI021
CI021 Maas Group Holdings (ASX: MGH) holds a AUD $200 million contract with Firmus for electrical delivery at the 100 MW Launceston AI Factory. SI020, SI021
CI022 As of May 18, 2026, the Launceston contract is approximately 35% complete by value with delivery and commissioning on track for calendar year 2026. SI020
CI023 JLE Group, a Maas Group subsidiary, is Firmus's exclusive electrical delivery partner for the broader 3.3 GW Australian AI factory portfolio. SI020
CI024 Each additional 100 MW AI factory module is expected to represent approximately AUD $200 million of electrical infrastructure revenue for JLE Group. SI020
CI025 Firmus describes itself as generating revenue but has not publicly disclosed specific revenue figures, annual recurring revenue, or gross margin in any reviewed public source. SI015, SI016
CI026 No gross margin, burn rate, cash runway, or income statement data have been disclosed for Firmus in any public filing, press release, or investor presentation reviewed. SI012, SI016, SI025
CI027 Firmus is not yet publicly profitable and operates in a build-out phase analogous to CoreWeave before its IPO — where headline adjusted margins overstate actual net profitability. SI016, SI024
CI028 CoreWeave reported FY2025 revenue of USD $5.1 billion and adjusted EBITDA of USD $3.1 billion (56% margin), but a net loss of USD $1.17 billion after GPU depreciation and interest on approximately USD $21 billion in debt. SI016
CI029 CoreWeave carries approximately USD $21 billion in total debt, creating a structurally analogous risk profile to Firmus's USD $10 billion debt facility combined with its planned capex obligations. SI016
CI030 Nvidia is both a strategic equity investor in Firmus across multiple rounds and the primary chip supplier for every AI factory deployment under Project Southgate, creating a capital-recycling dynamic. SI003, SI004
CI031 Roger Montgomery Investment Management reported the Firmus IPO is structured with no escrow or share-sale restriction, enabling early investors — who have realized a near-300% valuation markup — to exit at listing. SI012
CI032 Hayden Beamish, CIO of Endeavor Asset Management, exited his Firmus investment in August 2023, having described the company at that time as "a bitcoin miner with a stretched balance sheet that was making no money and had no legitimate business model." SI013
CI033 Some VC and PE funds are reported to hold Firmus at over 60% of total fund value due to rapid valuation markups, creating concentration risk for fund-level returns tied to the IPO outcome. SI012
CI034 Firmus publicly lists WEKA NVMe parallel file storage at USD $0.11 per GB per month on its pricing page. SI001
CI035 All GPU instance tiers on the Firmus public pricing page — H200, H100, L40S, A100, and A40 — show specifications but no list rate; pricing requires contacting sales for committed or dedicated access. SI001
CI036 The Blackstone debt facility proceeds will fund acquisition and deployment of AI compute, associated networking, storage, and data center infrastructure for the national Project Southgate rollout. SI006, SI009
CI037 Firmus's domestic supply chain involves Benmax for mechanical systems (cooling and power modules) and JLE Group for electrical systems, with combined capacity targeting up to 1.5 GW of AI facility delivery per year. SI021, SI020
CI038 Startup Daily reported investors expect Firmus's IPO valuation to reach AUD $8 billion or above, with some reports citing an ambition of up to AUD $12 billion. SI015
CI039 Nvidia's DGX Cloud managed AI platform is establishing a dedicated region at Project Southgate Melbourne in partnership with Firmus, routing Nvidia enterprise customers to Firmus infrastructure. SI005
CI040 Industry benchmarks for contracted renewable hydroelectric power supply in Tasmania are approximately USD $0.04-$0.06 per kWh, well below the global data center average. SI012, SI024
CI041 Industry benchmarks for AI data center capital intensity are approximately $50-55 million per MW; Firmus claims $35-45 million per MW, representing a claimed 30-35% cost reduction that has not been independently audited. SI012, SI024
CI042 GPU rack payback periods at leading AI cloud operators such as CoreWeave are typically 24-36 months at target utilization, setting the benchmark Firmus would need to match to generate positive free cash flow per deployed rack. SI016
CI043 Lennox Capital Partners assessed the near-term 150 MW Project Southgate pipeline at a potential value over 10x Firmus's then-current valuation, using comparable GPU-cloud operator profit multiples. SI014
CI044 Oliver Curtis, co-founder and co-CEO of Firmus, was convicted of insider trading in 2016 for conduct in 2007-2008, sentenced to two years' imprisonment, and served approximately twelve months. SI017, SI024
CI045 The ASX issued its first Listings Compliance update of 2026 in February 2026 dedicated entirely to the Good Fame and Character test; the update was widely understood in the financial press to be connected to the planned Firmus IPO. SI017
CE001 Firmus AI Factory is a modular, vertically integrated compute platform in which compute, cooling, power, and networking are co-designed from first principles rather than assembled from third-party colocation components. SE001, SE022
CE002 Each HyperCube™ is the physical building block of the AI Factory and houses 32 NVL72 racks with two NVIDIA Scale Units per module, in a primarily liquid-cooled, high-density configuration purpose-built for AI compute. SE001, SE003
CE003 Firmus AI Cloud offers GPU IaaS instances using H200 SXM (8 × 141 GB HBM3e each), H100, A100, and L40S configurations for bare-metal and virtual-machine workloads. SE002, SE004
CE004 H200 instance nodes include NVLink 4.0 + NVSwitch 3.0 (900 GB/s bi-directional per GPU) for intra-node GPU interconnect, and dual 200–800 Gb/s InfiniBand or 400 Gb/s Ethernet with RDMA/RoCE v2 support for inter-node distributed training. SE002
CE005 FactoryOS™ is Firmus's proprietary full-stack infrastructure intelligence platform governing cooling, power delivery, thermal telemetry, and GPU performance as a single unified system — not bolted-on observability — natively integrating the Synert FCAS grid-participation loop. SE001, SE006
CE006 Firmus selected VAST Data AI OS as the foundational data layer for its AI factory deployments in February 2026, aligned with the NVIDIA Cloud Partner (NCP) reference design and built for high-throughput disaggregated NVMe storage at petabyte-to-exabyte scale. SE010, SE024
CE007 Firmus is an NVIDIA Cloud Partner (NCP) building on the NVIDIA Vera Rubin DSX AI Factory reference design, as confirmed by NVIDIA's announcement of Firmus as a DSX Blueprint deployment partner in March 2026. SE005, SE011
CE008 The NVIDIA Vera Rubin DSX AI Factory reference design provides software libraries DSX Max-Q (maximize tokens per watt), DSX Flex (connect AI factories to power-grid services), and DSX Exchange (scalable IT/OT/energy signal integration) enabling co-designed compute, networking, and power orchestration. SE011
CE009 Firmus's Model-to-Grid architecture integrates model workload behavior, GPU thermal signatures, and real-time grid conditions into a single optimization framework coordinated with AEMO and Australia's National Electricity Market — with patented technologies in grid-aware orchestration developed since 2020. SE005, SE024
CE010 Synert, a Firmus subsidiary, provides Frequency Control Ancillary Services (FCAS) by transforming Eaton EnergyAware UPS battery systems into real-time grid frequency buffers, absorbing load shocks between the AI Factory and the grid under AEMO NEM rules. SE006, SE012
CE011 Synert's FCAS system is deployed in production at a Firmus AI Factory site in Australia, delivering certified fast-frequency response services to the live AEMO grid and generating grid-services revenue from otherwise latent UPS capacity. SE006, SE023
CE012 Synert's grid participation reduces total AI infrastructure energy costs at Firmus AI Factory sites by over 25%, as stated by Firmus in its Eaton partnership announcement. SE006
CE013 Firmus AI Cloud services include managed Slurm job scheduling for distributed training, pre-configured CUDA stacks (PyTorch, TensorFlow), GPU Fabric high-bandwidth interconnect, and CLI/Terraform/GitOps automation toolkits. SE002, SE004
CE014 Firmus AI Storage uses WEKA NVMe parallel file system with RDMA acceleration for multi-node training and checkpoint workloads, scaled alongside the VAST Data AI OS for petabyte-scale AI data pipelines. SE008, SE024
CE015 Firmus's MLPerf Training V4.0 submission, with power measurements reviewed by MLCommons at the immersion-rack level, showed approximately 30% better power efficiency compared to H100 SXM-based systems using conventional air cooling. SE009
CE016 Firmus's engineering principles page and FB Rice patent analysis cite a PUE of approximately 1.1 in production AI Factory deployments, with roughly 40% less power loss than air-cooled peers in the same jurisdictions. SE003, SE015
CE017 Firmus's marketing materials cite a PUE of 1.03 while the engineering principles page and MLPerf submission reference approximately 1.1; the two figures conflict and no independent third-party annual energy audit has been published to resolve them. SE001, SE003, SE009, SE015
CE018 Firmus claims its data centers use approximately 99% less water compared to conventional data centers that rely on evaporative cooling, enabled by closed-loop liquid cooling systems. SE001, SE014
CE019 Firmus's greenfield AI Factory campus in Tasmania operates without water consumption for more than 350 days per year under its closed-loop CDU cooling system. SE001
CE020 Project Southgate in Launceston, Tasmania, is targeting 36,800 NVIDIA GB300 Grace Blackwell GPUs in its full buildout, with Phase 1a at 44 MW and Phase 1b at 90 MW; Phase 1B initial deployment was declared operational in April 2026 with full buildout targeted for late 2026. SE005, SE014, SE020
CE021 Project Southgate expanded to a national programme targeting 1.6 GW of AI compute across Melbourne, Sydney, Perth, and Canberra by 2028, with a declared total build cost of A$73.3 billion co-developed with CDC Data Centres and NVIDIA. SE025, SE011
CE022 Firmus signed a three-year retail service agreement with Aurora Energy for up to 104 MW of Hydro Tasmania renewable power, ramping from August 2026 to maximum capacity in November 2026. SE017, SE014
CE023 Firmus plans three Tasmanian AI factory sites (St Leonards, Bell Bay, Wesley Vale) requiring approximately 400 MW total, which would make it Tasmania's largest power user at approximately 20% of total state energy demand. SE017, SE019
CE024 Firmus and SUBCO announced the Bernacchi-1 submarine fiber optic cable on 2 June 2026, delivering more than 60 Tbps of additional capacity to Tasmania — more than all existing Bass Strait fiber cables combined — with commercial operation targeted for Q2 2027. SE007, SE013
CE025 Bernacchi-1 will provide dual mainland pathways (northwest to Melbourne, northeast to Sydney via SMAP cable system), the first-ever direct Tasmania–New South Wales fiber connection; the SMAP system carries 400 Tbps total capacity. SE007, SE013
CE026 Firmus holds a multi-jurisdiction patent portfolio confirmed by FB Rice patent attorneys covering core liquid-cooling systems, grid-stabilisation software, and improvements to pipework, temperature controls, and HyperCube software — with granted applications in Australia and the US and divisional filings extending scope. SE015
CE027 NVIDIA captured approximately 43.6% of top-5 hyperscaler capital expenditure on AI infrastructure in Q4 2025 (against $130.7 billion total hyperscaler capex that quarter), representing a floor figure excluding sovereign buyers, second-tier cloud providers, and enterprise direct buyers. SE018
CE028 Firmus faces GPU supply allocation risk because large hyperscalers absorb the majority of NVIDIA's available GPU production, leaving secondary operators facing potential allocation compression if NVIDIA prioritises larger U.S. customers during periods of constrained supply. SE018, SE020
CE029 Firmus's dependency on NVIDIA is structural and self-reinforcing: every AI Factory module uses NVIDIA GPUs exclusively, NVIDIA is a strategic equity investor, and NVIDIA's DSX reference architecture underpins every Project Southgate deployment — creating incentive alignment but also single-vendor lock-in. SE005, SE020
CE030 Firmus AI Cloud Services carries company-claimed ISO 27001 and SOC-2 compliance with encryption in-flight and at rest, as stated on the official cloud services product page. SE004
CE031 Firmus's early AI Factory deployments in Singapore used OpenStack and Kubernetes-based cloud software from Canonical, running H100 and A100 GPUs inside HyperCube immersion-tank containers at STT GDC facilities. SE021
CE032 The Cube300 edge product holds a company-claimed PUE of less than 1.05 and TUE of less than 1.15 in a 300 kW shipping-container form factor with 145 rack units of capacity and N+1 reliability. SE021
CE033 Firmus HyperCube containers require no airflow in critical compute zones and occupy approximately one-twelfth the floor space of comparable air-cooled systems, enabling high-density deployment in standard industrial footprints. SE001, SE021
CE034 Firmus deployed approximately 4,000 NVIDIA GPUs across two STT GDC Singapore sites (SIN01) with 99 PFLOP/s total compute, supporting AI Singapore's SEA-LION open-source LLM project, constituting the company's first production-grade AI Cloud deployment. SE021
CE035 NVIDIA GB300 Blackwell Ultra GPUs require more than 120 kW per NVL72 rack (72 GPUs, 288 GB HBM3e per GPU, up to 8 TB/s memory bandwidth), necessitating liquid cooling as a technical prerequisite rather than an efficiency option. SE002, SE011
CE036 Firmus claims its liquid-cooling technology reduces energy consumption by up to 60% versus conventional air-cooled facilities and cuts construction costs by approximately half, as stated to media in relation to Project Southgate. SE019, SE016
CE037 Greens MP Tabatha Badger raised concerns in Tasmania's parliament about the "enormous" power requirements for Firmus's three planned factory sites and called for state-specific AI data centre legislation governing energy, water, and local employment obligations. SE017, SE016
CE038 Questions arose about preferential energy allocation when Firmus received a 104 MW Hydro Tasmania contract while the Boyer Paper Mill's request for an additional 45 MW to electrify its operations was refused, and Andrew Forrest's 250 MW green hydrogen project was abandoned partly due to insufficient power supply. SE017, SE016
CE039 Expert data-centre practitioners are sceptical that consistent annual PUEs of 1.03 are achievable at scale, noting that hyperscale facilities with advanced cooling typically report 1.1–1.15 annually, and that vendor PUE claims often exclude pump overhead, heat-exchanger losses, and auxiliary systems; no widely-accepted peer-reviewed facility has demonstrated 1.03 under independent annual audit. SE016
CE040 Firmus co-CEO Tim Rosenfield described the Bernacchi-1 cable investment as evidence of Firmus's leadership as a "full-stack AI Factory Platform company," signalling an intent to own connectivity layer as well as compute and cooling infrastructure. SE007, SE013
CU001 Firmus has publicly disclosed two signed multi-year, multi-billion-dollar hyperscale AI factory contracts under Project Southgate, both with unnamed "leading global technology companies" as of June 2026. SU014, SU017, SU018
CU002 The second Project Southgate hyperscale contract, announced March 2, 2026, covers approximately 18,400 Nvidia GB300 GPUs at the Melbourne facility in a multi-year, multi-billion-dollar agreement with an unnamed leading global technology company. SU014, SU017, SU018
CU003 The first Project Southgate hyperscale contract covers approximately 36,800 Nvidia GB300 GPUs at the Tasmania greenfield campus, targeted for completion in late 2026. SU014, SU013
CU004 Both Project Southgate hyperscale customer identities have not been publicly disclosed; Firmus refers to both only as "a leading global technology company" and "one of the world's leading technology platforms." SU014, SU019
CU005 AI Singapore (AISG) deployed 256 Nvidia H200 GPUs across 32 nodes on the Firmus Sustainable Metal Cloud platform for SEA-LION LLM training, a production deployment confirmed by a published case study. SU009, SU007
CU006 AISG researchers completed more than 200 experiments and produced more than 100 candidate models on the Firmus platform during the SEA-LION development engagement. SU009
CU007 A 27-billion-parameter SEA-LION model was trained in 10 days using 32 Firmus H200 nodes, and a 4-billion-parameter model was trained in 3.5 days using 16 nodes. SU009
CU008 Firmus and AI Singapore formalised their collaboration through a Memorandum of Understanding in March 2025, establishing a long-term partnership for AI research and infrastructure access in Southeast Asia. SU010, SU009
CU009 ST Telemedia Global Data Centres (STT GDC), a Temasek-backed Singapore data centre provider, announced a significant investment into a global venture with Firmus Technologies in June 2023 to launch the Sustainable Metal Cloud (SMC). SU001, SU005, SU024
CU010 The STT GDC-Firmus Sustainable Metal Cloud (SMC) was launched in Singapore (SIN01) in H2 2023 with planned expansion to India and Australia, offering bare-metal GPU IaaS for deep-learning AI and visual computing workloads. SU001, SU024
CU011 Firmus joined NVIDIA's DGX Cloud Lepton marketplace as a listed NVIDIA Cloud Partner (NCP), contributing Singapore and Australia-based GPU infrastructure to the unified NVIDIA developer platform announced at COMPUTEX June 2025. SU006, SU012
CU012 DGX Cloud Lepton NCP participants include CoreWeave, Crusoe, Lambda, Foxconn, GMI Cloud, Nebius, Nscale, SoftBank, Yotta Data Services, and Firmus; Firmus is the primary APAC player with sovereign GPU capacity. SU006, SU012
CU013 NVIDIA's DGX Cloud platform is establishing a dedicated region at Southgate Melbourne in partnership with Firmus, described as a distribution channel and platform relationship rather than a discrete end-customer revenue agreement. SU011, SU020
CU014 SemiAnalysis named Firmus's Sustainable Metal Cloud as one of only three GPU providers globally with correctly implemented InfiniBand SHARP in-network reduction in the inaugural ClusterMAX GPU Cloud Ratings. SU015, SU004
CU015 Firmus's SMC ranked alongside AWS and above Google Cloud in SemiAnalysis's ClusterMAX GPU Cloud Ratings (March 2025), the first independent GPU cloud performance benchmark series. SU015, SU004
CU016 Firmus's AI Cloud platform and Project Southgate target five customer segments: enterprise, education, government, startup, and scaleup users across Australia. SU011, SU020
CU017 The Blackstone USD 10 billion debt facility is structured so that drawdown is conditioned on executed long-term customer contracts with hyperscalers and blue-chip AI customers, preventing speculative infrastructure spend ahead of demand. SU002, SU011
CU018 Gilbert and Tobin described the Firmus debt facility as "the largest GPU financing for an Australian AI infrastructure company" and framed it as positioning Firmus at the centre of Australia's next phase of AI growth (February 2026). SU002, SU011
CU019 The Gilbert and Tobin firm note named Meta and OpenAI as typical users of AI infrastructure like Firmus's without identifying Firmus's actual hyperscale anchor customers, reinforcing how little independent verification of counterparties is possible. SU002
CU020 AI Singapore's Senior Director of AI Products, Leslie Teo, stated in March 2025 that the Firmus collaboration "underscores our commitment to pushing AI boundaries in Southeast Asia while ensuring accessibility and sustainability for the broader global community." SU010
CU021 Firmus AI Cloud provides H200, H100, L40S, A100, and A40 GPU instances for AI training and inference; all GPU tiers show "contact for pricing" rather than rack rates on the public pricing page. SU016, SU023
CU022 No named Australian enterprise or government customer in production is publicly identified by Firmus as of June 2026; enterprise and startup segments remain aspirational in Australia. SU016, SU011
CU023 AISG researchers described the Firmus platform as enabling "rapid experimentation and large-scale training without compromise, even under intensive training and evaluation cycles," citing proactive and responsive engineering support. SU009
CU024 The AISG-Firmus deployment achieved NCCL bandwidth of 317 GB/s and throughput of 387 TFLOP per device per second, with checkpoint evaluations completed in 1.5 hours using 2 nodes. SU009
CU025 SemiAnalysis ClusterMAX 2.0, released November 2025, evaluated 84 GPU cloud providers from a tracked universe of 209 operators and incorporated 140+ customer surveys; it uses hands-on benchmarking across ten categories to assign Bronze through Platinum tiers. SU008
CU026 Both Project Southgate hyperscale contracts are described by Firmus as multi-year, multi-billion-dollar agreements; no minimum revenue commitment, cancellation clause, or pricing floor has been disclosed. SU014, SU017
CU027 Firmus's sovereign AI positioning targets in-country deployments for public-sector and regulated workloads, ensuring data remains under local control and aligning with regional energy and sustainability priorities; this positioning is intended to create structural demand differentiated from offshore hyperscaler clouds. SU022, SU003
CU028 The Tasmanian Government established a world-first AI Factory Zone in Northern Tasmania in July 2025 with Project Southgate as its anchor tenant, providing policy legitimacy and grid access for sovereign AI infrastructure, but this is a policy framework rather than a customer contract. SU013, SU021
CU029 Firmus has not publicly disclosed NRR, GRR, churn rate, cohort data, active cloud customer count, or any retention metric for either the Firmus AI Cloud public platform or the hyperscale anchor contracts as of June 2026. SU016, SU014
CU030 Firmus AI Cloud does not display rack rates for any GPU instance tier (H200, H100, L40S, A100, A40) on its public pricing page; all GPU tiers show "contact for pricing." SU016
CU031 According to Hedgehog Cloud's analysis of SemiAnalysis ClusterMAX 2.0, Gold-tier GPU clouds can command a 70% pricing premium over Bronze commodity infrastructure on equivalent hardware; only CoreWeave holds Platinum tier as of the most recent cycle. SU008
CU032 Firmus stated it was generating revenue as of late 2025 but has not disclosed revenue figures, gross margin, burn rate, runway, or revenue breakdown by customer segment. SU023, SU011
CU033 Firmus's supply chain partners Benmax and Maas Group are producing AI Factory components at industrial scale to meet committed hyperscale customer contracts under Project Southgate. SU014, SU017
CU034 Firmus deployed approximately 4,000 Nvidia GPUs across two STT GDC sites in Singapore from mid-2023, delivering 99 PFLOP/s total compute for AISG and enterprise and government customers, representing the first revenue-generating international AI Factory. SU001, SU009, SU024
CU035 Hayden Beamish, CIO of Endeavor Asset Management and an early Firmus investor who exited in August 2023, warned in May 2026 that the ASX listing "will likely be priced too expensively" and that "the easy money has already been made," describing pre-pivot Firmus as having "no legitimate business model." SU025
CU036 The Melbourne AI Factory is expected to go online by mid-2026 for enterprise, education, government, startup, and scaleup customers via Firmus AI Cloud, following the March 2026 hyperscale contract announcement. SU014, SU020
CU037 No active customer count, customer acquisition rate, customer lifetime value metrics, or utilisation figures for Firmus AI Cloud or the SMC have been publicly disclosed as of June 2026. SU016, SU004
CU038 Independent assessment of Firmus customer concentration is impossible because both hyperscale anchor customers are anonymous, preventing verification of counterparty credit quality, revenue share, or renewal probability. SU002, SU014
CU039 The Blackstone debt facility structure creates potential sequencing risk: if hyperscale contract negotiations slow or a customer delays deployment, infrastructure buildout pace could compress the 2028 roadmap and reduce the evidence needed to attract further customers. SU002, SU014
CU040 Firmus selected VAST Data's AI Operating System as the foundational data layer for its AI Cloud platform to support secure multi-tenant enterprise and government workloads at petabyte scale across AI factory deployments. SU022, SU003
CU041 Firmus's SMC platform is independently corroborated as live in Singapore by both STT GDC's announcement and AISG's case study publication, confirming production operations for institutional and government research clients since H2 2023. SU001, SU007, SU009
CR001 Oliver Curtis was convicted by a Supreme Court jury of insider-trading in June 2016 and sentenced to two years' imprisonment, serving twelve months. SR014, SR015
CR002 The ASX passed Oliver Curtis under its Good Fame and Character framework, as confirmed by AFR coverage in May 2026, clearing him to lead Firmus as a publicly listed company. SR011, SR024
CR003 ASX Listing Rule 1.1 Condition 20 requires that every director of an applicant entity satisfy ASX of their good fame and character before admission to the official list. SR008, SR011
CR004 The Firmus IPO has been pitched with no escrow or share-sale restrictions for pre-IPO shareholders, described by independent fund managers as enabling early backers to provide exit liquidity at retail investors' expense. SR001, SR012, SR013
CR005 Some boutique VC and PE funds hold Firmus positions representing more than 60 percent of total fund value following rapid valuation markups. SR001, SR012
CR006 Independent fund managers including Roger Montgomery and Wilson Asset Management publicly characterised the escrow-free Firmus IPO structure as converting retail investors into exit liquidity for early backers. SR012, SR013
CR007 Australia's Security of Critical Infrastructure Act 2018 applies to data storage and processing assets used to provide services on a commercial basis to Commonwealth, state/territory, or critical-infrastructure responsible entities. SR002, SR004
CR008 SOCI Act requires responsible entities to register assets, report cyber incidents, and adopt and maintain a written Critical Infrastructure Risk Management Program (CIRMP). SR002, SR004
CR009 Proposed 2026 SOCI CIRMP amendments would require supply chain mapping, FOCI vendor risk assessments, phishing-resistant MFA, network segregation, and AusCheck background checks with civil penalties up to $3.3M for non-compliance. SR004, SR010
CR010 ASIC's ASX Inquiry Panel Final Report (March 2026) found that ASX's risk management and compliance practices need to mature, calling for transformational change across governance and accountability. SR008, SR011
CR011 From 10 December 2026, Australian APP entities using automated decision-making systems with significant impact on individuals must disclose the systems used, the decisions made, and the personal information processed. SR006, SR004
CR012 The OAIC is proactively auditing high-risk, data-heavy sectors as of 2026, with penalties for serious Privacy Act breaches reaching AUD$50 million or 30 percent of adjusted domestic turnover. SR006
CR013 Proposed 2026 SOCI Act CIRMP amendments impose civil penalties of up to $3.3 million per incident for corporations failing to comply with Ministerial Direction obligations. SR004, SR010
CR014 Firmus has acknowledged its PUE 1.03 figure was a point measurement from 2021, not an independently audited annual average, and no third-party PUE verification has been published. SR001, SR012
CR015 Dr. Amr Omar of the University of New South Wales described Firmus's claimed PUE of 1.03 as sitting at the extreme end of global possibility; Google and Microsoft typically report PUEs of 1.09–1.2. SR001, SR012
CR016 Most Australian data centres operate at PUEs between 1.4 and 1.7; industry-leading global facilities operate at 1.09–1.2, making Firmus's 1.03 claim more than 10 percentage points below its nearest verified comparable. SR001, SR012
CR017 Liquid immersion cooling requires a globally limited pool of specialised technicians, creating a recruitment bottleneck at the scale Firmus is planning. SR001
CR018 A 100 MW AI load remains a massive unrelenting strain on the electrical grid regardless of PUE, as noted by University of NSW engineering experts. SR001, SR017
CR019 Firmus's capital programme targets 1.6 GW of AI compute by 2028, representing an execution challenge at a speed and scale no operator has previously attempted in Australia. SR022, SR025
CR020 NEXTDC CEO Craig Scroggie publicly disputed Firmus's claimed build cost of approximately US$6 million per megawatt, citing an industry standard of approximately US$12.4 million per megawatt. SR028
CR021 Firmus's pivot from bitcoin mining to AI infrastructure has led independent analysts to question whether its facilities are truly purpose-built for enterprise AI or adapted from more speculative use cases. SR001
CR022 SOCI Act obligations require mandatory cyber incident reporting for critical infrastructure assets; proposed 2026 amendments extend requirements to Maturity Level 2 under Essential Eight, ISO 27001, or equivalent frameworks. SR002, SR004
CR023 The Blackstone $10B debt facility is structured as senior-secured infrastructure debt with drawdown conditioned on executed long-term customer contracts with hyperscalers and blue-chip AI customers. SR003, SR018
CR024 Firmus's total capex plan of A$73.3 billion through 2028 is approximately 55 times the total equity raised of approximately US$1.35 billion through April 2026, making continued debt drawdown and IPO proceeds structurally necessary. SR003, SR019
CR025 Missed build milestones, contract cancellations, or debt service coverage ratio deterioration could trigger covenant breaches and lender rights including demands for partial accelerated repayment or drawdown restrictions. SR003, SR019
CR026 Nvidia is both Firmus's primary GPU supplier and an equity investor, creating a symbiotic financing loop in which Nvidia's equity investment catalyses capital that returns to Nvidia as hardware revenue. SR001, SR029
CR027 Independent analysis estimates Nvidia's undisclosed equity investment may be less than A$100 million, yet it effectively unlocked A$830 million in third-party equity and the US$10 billion debt facility — most of which returns to Nvidia as chip purchases. SR001
CR028 Both Project Southgate hyperscale anchor customers remain unnamed in all public disclosures, making counterparty creditworthiness and contract enforceability unverifiable from outside the company. SR021, SR026
CR029 CDC Data Centres is both Firmus's primary physical infrastructure partner for Project Southgate and an independent data-centre operator that separately serves hyperscalers and enterprise clients. SR029
CR030 Nomura reportedly struggled to sell down a funded $120 million four-year private credit facility for Firmus, signalling institutional caution about the credit risk profile at the bank-market level. SR020
CR031 Ellerston Capital's stake in Firmus is understood to approach 7 percent following its cornerstone investment in the September 2025 Series B at approximately A$1.85 billion valuation. SR021, SR025
CR032 Firmus's total equity raised through April 2026 is approximately US$1.35 billion; the A$73.3 billion capex plan implies a leverage ratio of approximately 55:1 capital programme to equity raised. SR022, SR025
CR033 Co-CEO Oliver Curtis served twelve months at Cooma Correctional Centre following his 2016 insider-trading conviction, creating a governance exposure with no direct comparable at an Australian technology IPO of this scale. SR014, SR015
CR034 Firmus's founding team of Curtis, Rosenfield, and Levee had no prior data-centre or engineering background at founding; prior experience spanned investment banking, apparel retail, and mid-level resources management. SR015, SR021
CR035 CFO Kirsty Godfrey-Billy was appointed in September 2025, meaning she will have been in post for less than twelve months at the targeted ASX IPO date in mid-2026. SR025
CR036 Three non-executive directors — Lee Hatton, Christine Bartlett, and Julie Shuttleworth — were appointed in March 2026, meaning all three will have served less than six months before a targeted IPO. SR021, SR024
CR037 Greens MP Tabatha Badger publicly opposed Firmus's hydro power deal in Tasmania, noting that legacy industries had previously been told the grid lacked capacity for their own energy transitions. SR001, SR016
CR038 The AEMC has proposed new grid standards for data-centre connections and Greenpeace Australia has called for a moratorium on new AI data-centre approvals pending energy transition planning. SR016, SR017
CR039 Professor Toby Walsh of UNSW publicly doubted Firmus's job-creation promises of 50–100 full-time staff per 50 MW, noting that modern automated data centres rarely require such levels of labour. SR001
CR040 Hayden Beamish, CIO at Endeavor Asset Management and an early Firmus investor who exited in August 2023, described pre-pivot Firmus as 'terminal' and forecast the IPO will likely be priced too expensively. SR007
CR041 Firmus achieved a near-300% private valuation increase in under eight weeks between the September 2025 (A$1.85B) and November 2025 (A$6B) equity rounds, a speed of uplift that independent analysts flagged as unusually rapid. SR021, SR025
CR042 NAOS Asset Management's Q3 FY26 report noted Firmus as an unlisted investment with significant milestones in a quarter shaped by heightened market uncertainty and AI-driven sector volatility. SR005, SR009
CR043 Australian Privacy Act 2026 reforms impose mandatory ADM disclosure obligations from December 2026 and penalties for serious breaches reaching up to AUD$50 million or 30 percent of adjusted domestic turnover. SR006, SR004
CV001 Firmus Technologies raised USD $505 million in a round led by Coatue Management at a post-money valuation of USD $5.5 billion (AUD $7.95 billion), announced April 6, 2026, with Nvidia also participating. SV005, SV020
CV002 Investor correspondence and press reporting indicate Firmus is targeting an IPO valuation of AUD $8 billion or above, with some investor reports referencing an ambition of AUD $12 billion depending on bookbuild demand. SV007, SV008
CV003 As of June 18, 2026, Firmus had not lodged a prospectus with ASIC, published audited financial statements, or disclosed a formal IPO price range; the company is in a pre-prospectus roadshow phase. SV006, SV008
CV004 The Firmus ASX IPO, targeting a raise of approximately AUD $2 billion, is led by bookrunners Bank of America, JPMorgan, Morgans Financial, and Morgan Stanley, with listing targeted for June or July 2026. SV006, SV007
CV005 The USD $10 billion Firmus debt facility, announced February 9, 2026, is senior-secured and structured so drawdown is conditioned on executed long-term customer contracts with hyperscalers, staged against infrastructure deployment milestones. SV013, SV014, SV025, SV026
CV006 Public evidence supports at least two multi-billion dollar hyperscale contracts under Project Southgate, with neither customer identity, pricing, minimum commitments, nor cancellation provisions disclosed in any public source as of June 2026. SV005, SV006, SV014
CV007 Applying CoreWeave's 2026 EV/revenue multiple of 8.2–10.6x to Firmus's IPO valuation target of AUD $8–12 billion implies Firmus would need to demonstrate a revenue run rate of AUD $800 million to $1.5 billion to justify comparable pricing. SV010, SV022
CV008 CoreWeave (NASDAQ: CRWV) trades at an enterprise value / 2026 consensus revenue multiple of approximately 8.2x (MarketScreener) to 10.6x (Multiples.vc) as of June 2026, based on full-year 2026 guidance of USD $12–13 billion. SV010, SV022, SV009
CV009 CoreWeave reported Q1 2026 revenue of USD $2.08 billion (111.6% year-over-year growth), reaffirmed full-year 2026 guidance of USD $12–13 billion, and reported a contracted backlog of USD $99.4 billion as of March 31, 2026. SV012, SV009, SV030
CV010 The most analytically relevant comparables for Firmus valuation are CoreWeave (public GPU cloud neocloud), Lambda Labs (pre-IPO private GPU cloud), and NEXTDC (ASX-listed sovereign AI infrastructure), supplemented by AI data center M&A precedent transactions. SV009, SV010, SV016
CV011 Roger Montgomery of Montgomery Investment Management identified the no-escrow IPO structure, the rapid $1.9B-to-$6B valuation uplift, the unverified PUE claim, and the Oliver Curtis governance history as key overvaluation risks, stating publicly he would not invest at IPO. SV001, SV004
CV012 Hayden Beamish, CIO of Endeavor Asset Management and a former Firmus investor who exited in August 2023, stated that 'the easy money for Firmus has already been made' and that the IPO will be priced too expensively, declining to participate. SV002, SV018
CV013 The Firmus IPO has been pitched with no escrow or share-sale restrictions for early investors, meaning pre-IPO shareholders including founders, management, and institutional investors can sell from day one of listing. SV003, SV001
CV014 Wilson Asset Management — itself a pre-IPO Firmus shareholder — publicly identified the absence of IPO lock-up restrictions as a 'structural red flag,' noting that early investors who have already realised a 300% valuation markup are incentivised to sell at IPO. SV003
CV015 NEXTDC (ASX: NXT) trades at a trailing EV/EBITDA multiple of approximately 56–74x as of June 2026, with enterprise value of AUD $12–14 billion, driven by 60% contracted capacity growth to 667 MW by March 2026. SV023, SV016
CV016 Lambda Labs carries an implied valuation of approximately USD $9–9.1 billion as of May 2026 based on secondary market trading, at approximately 12x forward revenue on estimated FY2025 revenue of USD $760 million (79% year-on-year growth). SV024, SV028
CV017 AI and hyperscale data center M&A in 2025–2026 generated EV/revenue multiples of 6–10x and EV/EBITDA of 16–20x+ for operational, pre-leased AI-ready assets, with USD $151 billion in aggregate deal value since early 2024, 84% funded by private equity. SV019, SV017, SV027
CV018 Both Project Southgate hyperscale customer identities, contract pricing, minimum commitments, and cancellation provisions remain undisclosed as of June 2026, preventing independent assessment of contract quality, customer creditworthiness, or revenue durability. SV005, SV001, SV006
CV019 Firmus's PUE efficiency claim of 1.03 was disclosed by the company to be a 'point measurement from 2021' that was not independently verified, after UNSW expert Dr Amr Omar described the figure as being at 'the extreme end of global possibility'; global hyperscalers typically operate at PUE 1.1–1.2 or above. SV001, SV004
CV020 Typical Australian data centre facilities operate at PUE between 1.4 and 1.7 while global hyperscalers like Google and Microsoft struggle to achieve PUE below 1.2, providing context for how aggressive the Firmus 1.03 claim is relative to industry norms. SV001, SV004
CV021 Nvidia's equity investment in Firmus is estimated at AUD $30–75 million across both rounds (10–22% of each round), consistent with Nvidia's CoreWeave and Lambda playbook; this investment is structured to unlock GPU hardware revenue that flows back to Nvidia, creating a vendor-financing dynamic. SV001, SV004
CV022 Disclosed pre-IPO institutional shareholders include Ellerston Capital (approximately 7%), Regal Funds, Paradice Investment Management, Wilson Asset Management, Argo Investments, UniSuper, Lennox Capital Partners, Wingate, and Archibald Capital; none are reported to be subject to IPO lock-up. SV003, SV004
CV023 The exit pathway for Firmus is an ASX IPO (not a trade sale or SPAC), with secondary liquidity for pre-IPO shareholders available from day one of listing under the currently reported no-escrow structure; no secondary market trading program or lock-up alternatives have been publicly disclosed. SV003, SV006
CV024 Pre-IPO shareholders who backed the September 2025 AUD $1.9 billion round hold paper gains of approximately 300–320% at the April 2026 AUD $7.95 billion mark; at an IPO at AUD $8–12 billion, those gains would reach 320–530%. SV007, SV008
CV025 The Oliver Curtis insider-trading conviction (2016, two-year sentence, 12 months served) was assessed by the ASX under its Good Fame and Character framework following a February 2026 compliance clarification; Curtis was reported cleared to lead a publicly listed company, though the conviction remains directly relevant to governance scrutiny in a securities context. SV015, SV003
CV026 Wes Maas's Maas Group acquired a AUD $100 million stake in Firmus in February 2026 using proceeds from selling its building materials division, and former Bingo Industries boss Daniel Tartak acquired approximately AUD $3 million; neither stake is reported subject to IPO lock-up. SV003
CV027 At the Blackstone facility's private infrastructure debt rates of 7–9 percent, full drawdown of USD $10 billion would imply annual interest expense of USD $700–900 million; GPU asset depreciation (4–5 year useful life for GB300 GPUs) adds a further substantial non-cash charge, making GAAP profitability path-dependent on contracted pricing that has not been publicly disclosed. SV013, SV014, SV009
CV028 Applying CoreWeave's 2026 EV/revenue multiple of 8–10x to Firmus's IPO valuation target of AUD $8–12 billion implies required annual revenue of AUD $800 million to AUD $1.5 billion; no public evidence supports this level of revenue as of June 18, 2026. SV010, SV022, SV006
CV029 As of June 18, 2026, no IPO delay announcement has been made; bookbuild preparations are reported as ongoing and the IPO is described in market reporting as targeted for mid-2026, with Morgan Stanley, JPMorgan, Bank of America, and Morgans as bookrunners. SV006, SV007
CV030 The USD $10 billion Blackstone facility represents approximately 6.9 times the total equity raised to date (USD $1.35 billion); at full drawdown the company's debt-to-equity ratio would exceed 7:1, concentrating downside risk in equity and creating covenant risk if contracted revenue is delayed. SV013, SV014
CV031 Global AI and hyperscale data center M&A deal value reached USD $151 billion since early 2024, with 84 percent of deal value funded by private equity; private infrastructure funds raised over USD $250 billion in 2025, approximately 150 percent year-on-year growth. SV017, SV019, SV027
CV032 The risk of post-IPO multiple compression or a down-round is material: Firmus has no public revenue track record, no escrow mechanism, and is priced at a significant premium to comparable operational assets; any adverse governance event, AI demand softening, or GPU supply glut could compress multiples toward 3–5x revenue. SV001, SV004, SV010
CV033 The bull scenario for Firmus involves an IPO at AUD $10–12 billion in H2 2026, GB300 deployments online by H2 2026, and revenue reaching AUD $1.5–3 billion by 2027–2028, implying a potential 2028 enterprise value of AUD $15–20 billion at 10–15x forward revenue multiples consistent with the CoreWeave high-growth comparable. SV009, SV010, SV008
CV034 The base scenario involves Firmus IPO at AUD $7–9 billion in H2 2026 with partial contract disclosure, revenue ramp to AUD $600 million–$1.2 billion by end 2027, and multiple compression to 6–8x; net IPO investor returns would be flat-to-modest after debt service absorption. SV009, SV010
CV035 The bear scenario is triggered by IPO delay, hyperscale contract cancellation, AI demand softening, unsustainable Blackstone facility interest burden, or substantiated build-cost overstatement; in this scenario Firmus equity value could decline to AUD $3–6 billion, representing a 25–65% drawdown from an AUD $8 billion IPO price. SV001, SV004, SV010
CV036 CoreWeave's Q1 2026 GAAP net loss was USD $740 million despite revenue of USD $2.08 billion and adjusted EBITDA of USD $1.16 billion, driven by debt service and GPU depreciation, demonstrating that adjusted EBITDA metrics for highly leveraged AI cloud operators systematically overstate economic profitability. SV012, SV009, SV030
CV037 Standard IPO practice for a technology listing of this size would include a 90–180 day escrow on founders, management, and major pre-IPO institutional shareholders; the absence of any escrow for the Firmus IPO is structurally anomalous and increases immediate selling risk post-listing. SV003, SV001
CV038 Firmus's IPO structural readiness is advanced: bookrunners appointed (Bank of America, JPMorgan, Morgans Financial, Morgan Stanley); global roadshows completed in April 2026 across US, UK, Asia, and Australia; ASX Good Fame and Character clearance for Oliver Curtis obtained. SV006, SV003
CV039 No independent third-party audit or certification of Firmus's PUE efficiency claim, build cost per megawatt, or GPU utilisation rate has been published in any reviewed source as of June 18, 2026. SV001, SV004
CV040 NEXTDC (ASX: NXT) represents a lower-risk sovereign APAC AI infrastructure proxy available in the public market at current valuations; its AUD $12–14 billion enterprise value and 60% contracted capacity growth to 667 MW provide a visible public-market AI infrastructure investment alternative while Firmus's prospectus is pending. SV016, SV006
CV041 Lennox Capital Partners assessed Firmus's near-term 150 MW pipeline in October 2025 as supporting 'a run-rate revenue valuation over 10x Firmus's then-current mark' (implying AUD $190 million+ revenue at AUD $1.9 billion valuation), a proxy revenue figure that would need to grow by at least 4–8x to justify AUD $8–12 billion comparable multiples. SV001, SV004
CV042 Oliver Curtis's personal equity stake in Firmus is estimated at approximately AUD $1.25 billion at the April 2026 mark (AFR Rich List 2026); under the no-escrow structure this position is immediately liquid from day one of listing, representing one of the largest individual-holder selling risk positions in the register. SV003, SV004
CV043 NEXTDC's contracted AI capacity grew 60 percent to 667 MW by March 2026, supported by an AUD $2.2 billion capital raise in April 2026 and an OpenAI campus partnership; this progression demonstrates that competing sovereign Australian AI infrastructure is actively scaling to serve the same hyperscale demand that Firmus targets. SV016, SV006
CV044 Firmus's investor base includes Temasek (Singapore sovereign fund) alongside Blackstone and Nvidia, according to Sharecafe reporting, underscoring cross-border sovereign infrastructure investment interest but not reducing the pre-revenue valuation risk. SV002
CV045 The CoreWeave IPO in March 2025 at USD $40 per share implied a market capitalisation of approximately USD $19–27 billion on then-estimated annualised revenue of approximately USD $3.9 billion, representing approximately 15–20x FY2025 forward revenue at listing; this is the most relevant IPO transaction precedent for Firmus. SV011, SV009
来源
编号出版方标题引文
SO001 Firmus Creating the most efficient AI infrastructure - Firmus We design and operate Firmus AI Factories - modular, verticalized systems that optimise cost and power at every layer of the stack.
SO002 Firmus Firmus Raises USD$505 Million in Strategic Equity Investment Led by Coatue This transaction is expected to bring Firmus' total equity raised to USD$1.35 billion in equity over the past six months, at a post-money valuation of USD$5.5 billion.
SO003 Firmus Firmus Secures US$10 Billion Financing led by Blackstone and Coatue to Scale Energy-Efficient AI Infrastructure The transaction marks one of the largest private debt financings in Australian history.
SO004 Firmus Capital Raise to Expand Project Southgate Nationally - Firmus Firmus announces it has received commitments for a further A$500 million equity raise to accelerate the national rollout of Project Southgate... priced at a post-money valuation of approximately A$6 billion.
SO005 Firmus Firmus Welcomes Lee Hatton, Christine Bartlett and Julie Shuttleworth to the Board Firmus Technologies today announced the appointment of Lee Hatton, Christine Bartlett and Julie Shuttleworth AM as Non-Executive Directors of Firmus Grid Limited.
SO006 Startup Daily AI data centre startup Firmus valuation trebles to $6 billion in just 8 weeks after raising another $500 million Firmus was founded in 2019 by Oliver Curtis, Tim Rosenfield and Jonathan Levee, and originally focused on bitcoin mining.
SO007 Startup Daily Roxy Jacenko's husband now has a unicorn data centre startup after Nvidia backs its $330 million raise The investment gives Firmus a post-money valuation of $1.85 billion.
SO008 Startup Daily Green-powered AI data centre plan goes national with $73.3 billion project across 4 Australian capitals The bigger picture now includes new sites in Perth, Adelaide, Canberra and Sydney as well, at a total cost of up to $73.3 billion, delivering 1.6GW of compute at five Firmus sites by 2028.
SO009 Startup Daily AI data centre startup Firmus Technologies banks another $100 million The $100m investment in Firmus gives the business a 1.7% stake.
SO010 Startup Daily AI data centre startup Firmus eyes off another $725 million raise as it readies to IPO on the ASX Investors continue to pour astonishing levels of capital into Sydney-founded Firmus Technologies, which has confirmed it's set to raise another US$505 million (A$725m) at an A$8 billion valuation.
SO011 SmartCompany Newly minted unicorn Firmus raises $330 million to build 'AI factory' in Tasmania The investment gives Firmus a valuation of $1.9 billion, according to The Australian Financial Review.
SO012 SmartCompany Firmus expands Australian 'AI factory' plan with $73.3 billion from Nvidia, CDC
SO013 The Next Web Nvidia-backed Firmus targets $2bn ASX IPO after locking in $505m equity and $10bn Blackstone debt for its AI factory network Australian AI data centre company Firmus has raised $505m at a $5.5bn valuation in what it says is its final pre-IPO round, and is now targeting a $2bn listing on the ASX in June or July.
SO014 Commonwealth Director of Public Prosecutions (CDPP) Insider trader sentenced to full time imprisonment Former banker, 30-year-old Oliver Curtis was today sentenced to two years imprisonment after being found guilty by a Supreme Court jury earlier this month of conspiring to commit insider trading with his former best friend, John Hartman.
SO015 nine.com.au Oliver Curtis was jailed for insider trading. He's now worth $1.25bn Curtis began Firmus alongside his cousin Tim Rosenfield in 2019.
SO016 Silicon Valley Investclub Firmus Technologies | Silicon Valley Investclub
SO017 Tracxn Firmus - 2026 Company Profile, Team, Funding & Competitors Firmus is a series F company based in Saint Leonards (Australia), founded in 2019.
SO018 Caplight Firmus Technologies | Valuation, Funding Rounds & Stock Price | Caplight
SO019 The Daily Perspective Firmus Technologies: The $10B AI Bet With a Convicted Co-Founder The sheer pace of capital accumulation — from a $1.85 billion valuation in September 2025 to $6 billion just two months later — has prompted some investors to wonder whether the numbers reflect genuine enterprise value or the inflated enthusiasm of a global AI infrastructure boom.
SO020 The Witness (Sydney Morning Herald / The Age) Oliver Curtis and Firmus, the AI infrastructure provider heading towards a $6 billion listing on the ASX "None of the three co-founders have a technical or engineering background. Rosenfield was previously chief executive of lingerie company Simone Perele, while Levee, now Firmus' director of research and development, had some middle management experience in the resources sector."
SO021 Livewire Markets (NAOS Asset Management) AI factories – are they built on Firm(us) ground? Firmus have deployed ~4,000 NVIDIA GPUs across two of these STT data centres sites.
SO022 Montgomery Investment Management (Roger Montgomery) $6 billion hype or structural risk? Inside the Firmus ASX listing By taking small equity investments from Nvidia to secure priority chip allocations, Firmus gained a 'halo effect' that attracted other big-name investors and facilitated massive debt.
SO023 InnovationAus Data centre cooling startup soars to $6b valuation
SO024 ARN (Australian Reseller News) Firmus picks up $500M in additional equity raise The new hires come more than a month after the company hired Grant Dempsey as its incoming chair of the board and Kirsty Godfrey-Billy as its chief financial officer back in late September.
SO025 Bloomberg Nvidia-Backed Firmus Names Three Directors Ahead of Mooted IPO Australian artificial-intelligence start up Firmus Technologies Pty. appointed three new company directors ahead of an expected initial public offering later this year.
SO026 Stocks Down Under The ASX's Good Fame and Character Test: The Potential Listing of Oliver Curtis' $6bn AI Company Firmus Meant A Clarification Was Needed The ASX's timing of the February update is widely understood to be connected to the planned ASX listing of Firmus Technologies.
SO027 Wilson Asset Management / The Australian (Licensed reprint) Oliver Curtis's Firmus eyes $12bn float with no escrow restrictions for early investors investors in Firmus... have been told the initial public offering is 'on track for June or July' and expected to be priced '40 per cent to 100 per cent' above the current value.
SM001 Firmus Technologies Firmus Raises USD$505 Million in Strategic Equity Investment Led by Coatue The investment supports the rapid deployment of Firmus' AI infrastructure platform, based on the NVIDIA Vera Rubin DSX reference design, across the Asia-Pacific region.
SM002 Gartner Gartner Says Worldwide Sovereign Cloud IaaS Spending Will Total $80 Billion in 2026 Governments will remain the main buyers to meet digital sovereignty needs, followed by regulated industries and critical infrastructure organizations.
SM003 Firmus Technologies Firmus Secures US$10 Billion Financing led by Blackstone and Coatue to Scale Energy-Efficient AI Infrastructure AI is driving one of the most significant infrastructure build-outs in decades, and we believe Australia can play a central role in that transformation.
SM004 TechCrunch Firmus, the 'Southgate' AI data center builder backed by Nvidia, hits $5.5B valuation Firmus originally provided cooling technologies for Bitcoin mining and has become yet another crypto-roots-turned-AI provider company that investors love.
SM005 The Next Web Nvidia-backed Firmus targets $2bn ASX IPO after locking in $505m equity and $10bn Blackstone debt for its AI factory network The question of how governed and sovereign AI infrastructure gets built, who owns it, who controls access to it, and what regulatory frameworks apply, is one that Firmus's listing will force Australian policymakers and investors to confront in real terms for the first time.
SM006 Startup Daily Green-powered AI data centre plan goes national with $73.3 billion project across 4 Australian capitals
SM007 Firmus Technologies Firmus Signs Multi-Year Agreement with Global Hyperscale Customer at Project Southgate The contract marks the arrival of hyperscale AI factories in Australia for the first time, positioning the country on a path to become a major producer and exporter of AI tokens globally.
SM008 Australian Broadcasting Corporation 'None of this was forecast': Energy demand from AI data centres soars Data centres could account for 13 per cent of total national electricity demand in 2040 under a high-growth scenario.
SM009 Australian Energy Market Commission AEMC proposes new grid standards for data centre connections Data centres aren't passive loads anymore; they're active grid participants. When they fail to ride through faults, it has the potential to trigger cascading failures and blackouts.
SM010 Greenpeace Australia Pacific Energy Vampires: the AI data centres draining Australia The frenzied rollout of AI data centres in Australia is rushing through massive new projects, which will derail the renewable energy transition unless the government urgently intervenes.
SM011 Futurum Group AI Capex 2026: The $690B Infrastructure Sprint The five largest US cloud and AI infrastructure providers have collectively committed to spending between $660 billion and $690 billion on capital expenditure in 2026, nearly doubling 2025 levels.
SM012 Mordor Intelligence Asia-Pacific Data Center GPU Market Size & Share Analysis – Growth Trends & Forecast 2026–2031
SM013 pv magazine Australia AI is driving a data centre boom but can the grid keep up? According to AEMO, Australian data centres consumed an estimated 3.9 TWh of electricity in FY25... forecast to reach 12.0 TWh by FY30.
SM014 InnovationAus Firmus locks in hyperscale customer for $73bn AI play
SM015 Data Bridge Market Research Sovereign AI Infrastructure Market – Global Market Size, Share, and Trends Analysis Report The global sovereign AI infrastructure market was valued at USD 14.82 billion in 2025 and is projected to reach USD 49.67 billion by 2033, growing at a CAGR of 16.3% from 2026 to 2033.
SM016 Yahoo Finance / Research and Markets (Arizton) Australia Data Center Market Investment Analysis Report 2026–2031 The Australia Data Center Market is projected to surge from USD 4.22 Billion in 2025 to USD 9.02 Billion by 2031, growing at a 13.50% CAGR.
SM017 Research and Markets Sovereign Cloud Market Report 2026 The Sovereign Cloud Market, valued at USD 128.62B in 2026, is projected to reach USD 298.45B by 2030, growing at a 23.4% CAGR.
SM018 Firmus Technologies Firmus Expands Project Southgate Across Australia Investment: A$4.5 billion initial investment, with scale up to A$73.3 billion and 1.6GW of operational AI Factories through 2028.
SM019 MarketsandMarkets AI Data Center Market Report 2026–2032
SM020 ARN (Australian Reseller News) Firmus secures $330M to build green, sovereign AI factory with NVIDIA
SM021 Ventureburn Firmus Raises $505M to Expand AI Data Centre Infrastructure Some industry observers have questioned the scale of Firmus's efficiency claims.
SM022 VAST Data Firmus Technologies Group Selects VAST AI Operating System to Power Sovereign, Energy-Efficient AI Factories in Asia Pacific Firmus was founded on the belief that AI infrastructure must be engineered as a single, vertically integrated system.
SM023 OfficeChai AI Capex Spend At Top 4 Hyperscalers To Touch $715 Billion In 2026 Power and land constraints are increasingly the binding limits on deployment speed, not money or demand. Modern AI racks exceed 100 kW per rack, straining electrical infrastructure designed for a different era.
SM024 Intelligent CIO APAC Firmus Technologies Group selects VAST AI Operating System to power sovereign, energy-efficient AI factories in APAC
SM025 Tech Insider Firmus Tech's $505M Raise Powers $5.5B AI Data Center IPO [2026]
SP001 Firmus Technologies Creating the most efficient AI infrastructure — Firmus home page We design and operate Firmus AI Factories - modular, verticalized systems that optimise cost and power at every layer of the stack.
SP002 Firmus Technologies AI Cloud — Firmus Firmus AI Cloud meets ISO 27001 and SOC-2 requirements. Combined with encrypted InfiniBand networking, built-in observability, and flexible automation across hybrid and multi-cloud networks.
SP003 Firmus Technologies Firmus Secures US$10 Billion Financing led by Blackstone and Coatue The transaction marks one of the largest private debt financings in Australian history.
SP004 Firmus Technologies Firmus Expands Project Southgate Across Australia Project Southgate is a blueprint for how Australia can lead the world in scalable, sovereign AI infrastructure.
SP005 Firmus Technologies Firmus' SMC ranks among world's top GPU clouds, outperforming global giants Firmus' Sustainable Metal Cloud was named one of only three GPU providers worldwide to have correctly implemented InfiniBand SHARP in-network reduction.
SP006 Firmus Technologies Firmus Raises USD$505 Million in Strategic Equity Investment Led by Coatue This transaction is expected to bring Firmus' total equity raised to USD$1.35 billion in equity over the past six months, at a post-money valuation of USD$5.5 billion.
SP007 Firmus Technologies Firmus Signs Multi-Year Agreement with Global Hyperscale Customer at Project Southgate The contract is a multi-billion dollar commitment for approximately 18,400 NVIDIA GB300 GPUs to be deployed at the Melbourne facility.
SP008 CDC Data Centres Driving Australia's AI Future with CDC, Firmus, and NVIDIA CDC is proud to provide the data centre infrastructure for Project Southgate and play a critical role in positioning Australia as a leader in sustainable AI.
SP009 OpenAI Introducing OpenAI for Australia OpenAI today signed a Memorandum of Understanding with NEXTDC to develop a sovereign AI infrastructure partnership under the OpenAI for Australia program.
SP010 AUCyber Limited Investor Relations — AUCloud AUCyber delivers cutting-edge solutions to support enterprise customers of all sizes, as well as Australian Governments, ensuring robust protection and reliable IT infrastructure.
SP011 VAST Data Firmus Tech picks VAST AI OS for eco-friendly AI factories in Asia Pacific Firmus, an NVIDIA Cloud Partner (NCP), is building large-scale AI infrastructure designed to maximise performance per watt, supporting advanced AI training and inference workloads while minimising environmental impact.
SP012 Startup Daily Green-powered AI data centre plan goes national with $73.3 billion project across 4 Australian capitals The bigger picture now includes new sites in Perth, Adelaide, Canberra and Sydney as well, at a total cost of up to $73.3 billion, delivering 1.6GW of compute at five Firmus sites by 2028.
SP013 Smart Company Firmus expands Australian 'AI factory' plan with $73.3 billion from Nvidia, CDC The Melbourne site involves 18,500 Nvidia GB300 GPUs, which are expected to come online by April 2026, for enterprise, education, government, startup and scaleup customers across Australia.
SP014 W.Media Firmus partners CDC on Southgate to scale to 1.6 GW According to Firmus, Project Southgate will scale up to AUD 73.3 billion in investment through 2028.
SP015 ABC News (Australia) Questions raised over deal between AI company and state power generator Given other major customers have been told there is not enough energy for their needs, what can you tell us about the power deal with Firmus.
SP016 Stocks Down Under The ASX's Good Fame and Character Test: The Potential Listing of Oliver Curtis' $6bn AI Company Firmus Meant A Clarification Was Needed In February 2026, the ASX made clear exactly how it intends to answer that question. And even though we cannot prove it was a co-incidence, the forthcoming listing of Firmus would mean this issue could be coming to a head if left unaddressed.
SP017 Roger Montgomery (Montgomery Investment Management) $6 billion hype or structural risk? Inside the Firmus ASX listing While global hyperscalers like Google and Microsoft struggle to dip below 1.2, and typical Australian facilities operate between 1.4 and 1.7, Firmus's claims of 1.03 have been described by Dr. Amr Omar of the University of NSW as being at the extreme end of global possibility.
SP018 Rask Media Why NEXTDC Raised $2.2 Billion as the AI Data Centre Race Heats Up He pushed back on Firmus' claim that it can build capacity for roughly US$6 million per megawatt, versus an industry standard cited near US$12.4 million.
SP019 TechCrunch Firmus, the 'Southgate' AI data center builder backed by Nvidia, hits $5.5B valuation Asia AI data center provider Firmus on Monday announced a fresh $505 million raise led by Coatue at a $5.5 billion post-money valuation.
SP020 CNBC CoreWeave stock sinks 10% on weak revenue guidance, increased spending forecast The company ended the quarter with about 3.5 gigawatts of total contracted power, along with a $99.4 billion revenue backlog.
SP021 Introl Australia's AI Infrastructure Moment Australia has become ground zero for Asia-Pacific AI infrastructure investment. OpenAI launched its 'OpenAI for Countries' program with a $7 billion Sydney campus.
SP022 Forbes Australia NextDC, OpenAI to develop $4.6 Billion data centre in Sydney The A$7 billion ($4.6 billion) data center to be built on a 258,000-square-meter site the of NextDC's S7 facility, will serve OpenAI clients such as Commonwealth Bank of Australia and retail giant Wesfarmers.
SP023 Data Centre Dynamics Macquarie Data Centres tops out IC3 Super West facility in Sydney IC3 Super West is the next data center in our pipeline of sites planned to add circa 200MW of AI and cloud capacity in Sydney.
SP024 MLQ.ai CoreWeave — AI for investors analysis CoreWeave has built scale faster than any GPU cloud competitor, with $66.8 billion in contracted backlog, an NVIDIA strategic partnership, and the software platform to orchestrate it.
SP025 Sacra Lambda Labs revenue, valuation and funding Lambda's pricing emphasizes cost efficiency—Nvidia H100 PCIe instances at roughly $2.49 per hour compared to $4.25 at CoreWeave—helping drive utilization and expand its customer base.
SP026 S&P Global Market Intelligence (Visible Alpha) CoreWeave's revenue set to soar as demand for AI infrastructure booms Analysts expect CoreWeave's revenue to more than double in 2025 to $5 billion, before climbing to $11.6 billion in 2026.
SP027 Octus Intelligence Nomura Struggles to Selldown Funded $120M 4Y Private Credit Facility for Singapore's Firmus Tech On Concerns Over U.S.'s GPUs Export Controls Nomura is struggling to sell down any of its fully funded $120 million, four-year private credit debt facility for Singapore-based AI-focused data center company Firmus Technologies because of concerns by targeted private credit funds over the impact of US export restrictions on high-end GPUs.
SI001 Firmus Technologies Pricing — Firmus For committed and dedicated access with long term pricing, contact sales.
SI002 Firmus Technologies Firmus Raises USD$505 Million in Strategic Equity Investment Led by Coatue — Firmus This transaction is expected to bring Firmus' total equity raised to USD$1.35 billion in equity over the past six months, at a post-money valuation of USD$5.5 billion.
SI003 TechCrunch Firmus, the 'Southgate' AI data center builder backed by Nvidia, hits $5.5B valuation Asia AI data center provider Firmus on Monday announced a fresh $505 million raise led by Coatue at a $5.5 billion post-money valuation.
SI004 The Next Web Nvidia-backed Firmus targets $2bn ASX IPO after locking in $505m equity and $10bn Blackstone debt for its AI factory network The IPO, expected in June or July, would seek to raise an additional $2bn and, if completed at that scale, would rank among the largest technology listings in Australian history.
SI005 Startup Daily Green-powered AI data centre plan goes national with $73.3 billion project across 4 Australian capitals
SI006 Milbank LLP Milbank Advises Firmus in Landmark US$10B Financing Led by Blackstone and Coatue to Accelerate Energy Efficient AI Infrastructure This transaction is one of the largest financings for an Australian AI infrastructure company and positions Firmus to rapidly scale high-performance AI compute capacity to meet surging demand from hyperscalers and other blue-chip AI customers.
SI007 Allens Allens advises Blackstone on US$10bn financing of Firmus Group
SI008 Kirkland & Ellis LLP Kirkland Represents Blackstone on AI Infrastructure Company Firmus' US$10 Billion Debt Financing Facility
SI009 Data Center Dynamics Firmus secures $10bn debt financing facility for Project Southgate AI factory roll out AI is driving one of the most significant infrastructure build-outs in decades, and we believe Australia can play a central role in that transformation.
SI010 Firmus Technologies Firmus Signs Multi-Year Agreement with Global Hyperscale Customer at Project Southgate — Firmus The contract is a multi-billion dollar commitment for approximately 18,400 NVIDIA GB300 GPUs to be deployed at the Melbourne facility.
SI011 Bloomberg Nvidia-Backed AI Startup Firmus Signs New Contract Ahead of IPO The multi-billion contract covers roughly 18,400 Nvidia Corp. GB300 graphics processing units to be deployed at the facility, Firmus said in a statement Monday, without naming the customer.
SI012 Roger Montgomery Investment Management $6 billion hype or structural risk? Inside the Firmus ASX listing And given that early backers have already seen a nearly 300 per cent valuation uplift in just a few months, don't you think the lack of share dealing restrictions (escrow) suggests they will look to realise those gains by selling to IPO buyers?
SI013 Sharecafe Australian CIO Cautions on Firmus Valuation Firmus at the time was a bitcoin miner with a stretched balance sheet that was making no money and had no legitimate business model.
SI014 Lennox Capital Partners Stock in Focus — Firmus Utilising similar operating profit multiples sees our valuation for Firmus at well over 10x its most recent valuation. This is solely based on its near-term pipeline for 150MW of capacity.
SI015 Startup Daily AI data centre startup Firmus eyes off another $725 million raise as it readies to IPO on the ASX
SI016 CTOL Digital Solutions Firmus Reportedly Raises $505M — And Its $5.5B Valuation Is Already Stale Pricing mechanics, power cost passthroughs, and cancellation rights are undisclosed. The rumored IPO range of A$8–12 billion would be a dramatic markup from even the corrected valuation — before a single quarter of public financials.
SI017 Stocks Down Under The ASX's Good Fame and Character Test: The Potential Listing of Oliver Curtis' $6bn AI Company Firmus Meant A Clarification Was Needed The ASX's timing of the February update is widely understood to be connected to the planned ASX listing of Firmus Technologies.
SI018 W.media Firmus secures "multi-billion-dollar" AI contract for Melbourne
SI019 InnovationAus Firmus, CDC and Nvidia reveal $73bn 'AI factory' plan
SI020 Maas Group Holdings Limited Corporate Update — ASX Announcement 18 May 2026 Previously announced A$200 million Firmus contract (100MW Launceston AI Factory) is now approximately 35% complete with delivery and commissioning on track within calendar year 2026.
SI021 Firmus Technologies Firmus Extends Investment in Australian AI Factory Supply Chain — Firmus Over the past six months, Firmus has raised more than $800 million in equity and committed more than $300 million to its Australian AI factory supply chain.
SI022 Firmus Technologies Creating the most efficient AI infrastructure — Firmus
SI023 IT Brief Australia Firmus inks multi-year Nvidia AI factory deal in Melbourne
SI024 Tech Insider Firmus Tech's $505M Raise Powers $5.5B AI Data Center IPO [2026]
SI025 Digitrendz Firmus Plans $2B ASX IPO After $505M Raise and Blackstone Debt
SE001 Firmus Technologies AI Factories — Firmus "We design, build and operate the entire platform, from the chip to the grid. Every element in our stack was developed because existing infrastructure was too expensive, too slow or fundamentally incapable of supporting what modern AI requires."
SE002 Firmus Technologies AI Cloud Compute — Firmus "GPU: 8x NVIDIA H200 Tensor Core GPUs (Hopper architecture, HBM3e); Interconnect: NVLink 4.0 + NVSwitch 3.0 (900 GB/s bi-directional per GPU); Network: Dual 200–800 Gb/s InfiniBand or 400 Gb/s Ethernet (RDMA / RoCE v2 support)"
SE003 Firmus Technologies Engineering Principles — Firmus "32 NVL72 RACKS — 2 NVIDIA Scale Units per HyperCube module"
SE004 Firmus Technologies Cloud Services — Firmus "Firmus Cloud Services is ISO 27001 and SOC-2 compliant, with encryption in-flight and at rest."
SE005 Firmus Technologies Firmus to Build Grid-Integrated AI Factory Software with NVIDIA — Firmus "The AI factory of the future must be as intelligent about energy as it is about compute. Our architecture is designed to understand those shifts and apply that intelligence along with NVIDIA DSX Blueprint deployments to operate infrastructure at scale while working seamlessly with grid systems."
SE006 Firmus Technologies Firmus and Eaton Showcase How AI Factories Can Support the Grid — Not Strain It "One of our Firmus AI Factory sites is: Delivering certified fast frequency response services to the live grid; Generating grid services revenue from otherwise latent UPS capacity; Cutting total AI infrastructure energy costs by over 25%"
SE007 Firmus Technologies Firmus and SUBCO to Build Bernacchi-1 — Firmus "Bernacchi-1 will be the first new fibre connection between Tasmania and mainland Australia in more than 20 years and will deliver more than 60 terabits per second of additional capacity to the state on day one — more than all existing Bass Strait fibre cables combined."
SE008 Firmus Technologies AI Storage — Firmus "WEKA NVMe PFS — Distributed RDMA Storage — Best for training datasets, AI pipelines, and large-scale checkpoints"
SE009 Firmus Technologies MLPerf V4.0 — Firmus "At this level, our results show approximately 30% better performance compared to H100 SXM-based systems cooled with air, the most common method for building GPU clusters today."
SE010 Compute Forecast Firmus Technologies Group Selects VAST AI Operating System to Power Sovereign Energy-Efficient AI Factories in Asia Pacific
SE011 NVIDIA Investor Relations NVIDIA Releases Vera Rubin DSX AI Factory Reference Design and Omniverse DSX Digital Twin Blueprint With Broad Industry Support "The new NVIDIA Vera Rubin DSX AI Factory reference design provides a guide for building codesigned AI infrastructure that delivers maximum token per watt and accelerated time to first production."
SE012 Synert About — Synert "Synert's core technology — the PowerCube — is a next-generation FCAS or frequency control machine provided in modular battery units, or through upgrading existing global UPS infrastructure."
SE013 SUBCO Firmus and SUBCO to Build Bernacchi-1, Tasmania's First New Sub-Sea Fibre Connection in Over 20 Years "Bernacchi-1 provides dual pathways to the mainland — North-West to Melbourne, and North-East to Sydney. The path to Sydney is significant as the first time Tasmania has had a direct connection to New South Wales."
SE014 DC Pulse Firmus Southgate: 36,800 GPU Renewable AI Factory
SE015 FB Rice Patent & Trade Mark Attorneys How Firmus Technologies is Delivering a Cool Solution in a Hot Market "Firmus has created a layered defensive position which increases the cost and complexity for competitors seeking to emulate its approach. This also provides Firmus with potential leverage in licensing discussions, partnerships, and future market expansion."
SE016 Stockhead / The Australian Firmus Bets Big on 'AI Factories' as Power and Cooling Limits Bite
SE017 ABC News Australia AI Company Set to Become Tasmania's Largest Power User "Mr Duigan revealed Firmus would require about 400 megawatts in total to power the three sites, which would make it the largest power user in the state. An energy consultant said 400 megawatts represented about 20 per cent of the state's energy use, or more than the combined demand of all Tasmanian households."
SE018 Lambda Finance GPU Supply 2026: NVIDIA $57B Quarter, HBM Bottleneck, Hyperscaler $131B Capex "NVIDIA captures roughly 44 cents of every hyperscaler capex dollar going into AI infrastructure — a concentration unprecedented in the history of enterprise compute."
SE019 Pulse Tasmania Tech Company to Build $2.1 Billion AI Factory in Tasmania
SE020 SiliconAngle Nvidia-Backed Firmus Raises $505M at $5.5B Valuation Ahead of ASX IPO
SE021 Data Center Dynamics Australia's Firmus Launches Immersion-Cooled Bare Metal AI Cloud with STT GDC "The SMC is made up of immersion tanks from Firmus, installed in 1MW shipping containers known as 'Hypercubes' within STT GDC facilities in Singapore, India, and Australia."
SE022 Firmus Technologies About — Firmus
SE023 SiliconAngle AI Infrastructure Startup Firmus Raises $327M to Build Data Centers in Australia
SE024 VAST Data Firmus Tech Picks VAST AI OS for Eco-Friendly AI Factories in Asia Pacific "Firmus was founded on the belief that AI infrastructure must be engineered as a single, vertically integrated system. Our Model-to-Grid architecture integrates model behaviour, GPU performance, thermal dynamics, and grid conditions into one optimisation framework."
SE025 Firmus Technologies Firmus Secures US$10 Billion Financing Led by Blackstone and Coatue to Scale Energy-Efficient AI Infrastructure
SU001 ST Telemedia Global Data Centres ST Telemedia Global Data Centres and Firmus Technologies Forge Partnership to Build a Global Network of Sustainable AI Factories "STT GDC, in partnership with Firmus, is committed to providing sustainable and foundational compute infrastructure. Through our new bare-metal service, customers will have early access to cost-effective and large-scale GPU compute through Firmus' embedded AI compute platform."
SU002 Gilbert + Tobin Gilbert + Tobin drives landmark USD10 billion AI infrastructure financing for Firmus "The debt facility will be drawn against executed long-term customer contracts with leading hyperscalers and other blue-chip AI customers. They monetise this by renting that capacity to customers such as Meta and OpenAI for critical AI workloads."
SU003 Intelligent Data Centres Firmus Technologies Group selects VAST AI Operating System to power sovereign, energy-efficient AI factories in APAC "Firmus' infrastructure strategy is built on the NVIDIA Cloud Partner reference design, enabling both anchor tenant and government-backed workloads as the company scales AI capacity across Australia, Singapore, and the broader APAC region."
SU004 Firmus Technologies (SMC) Sustainable Metal Cloud ranks among world's top GPU clouds by SemiAnalysis "The combination of industry-leading performance benchmarks and global customer adoption shows that the AI Factory model delivers measurable value — at speed and scale."
SU005 Data Centre and Network News (DCNN) STT GDC and Firmus to build sustainable AI factories
SU006 NVIDIA NVIDIA Announces DGX Cloud Lepton to Connect Developers to NVIDIA's Global Compute Ecosystem "NVIDIA Cloud Partners (NCPs) 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."
SU007 AI Singapore NLP Hub — AI Singapore Programme Overview
SU008 Hedgehog Cloud The Sell Math: How ClusterMAX 2.0 Ratings Translate Into Revenue "The tier distribution is skewed. In ClusterMAX 2.0, only CoreWeave holds Platinum. Gold is occupied by Crusoe, Nebius, Oracle, Azure, Together AI, and LeptonAI. The difference between Bronze and Silver on a 1,024-GPU B200 cluster is $15.1 million in additional annual revenue on the same hardware footprint."
SU009 Firmus Technologies AI Singapore x Firmus — Case Study "32 nodes / 256 NVIDIA H200 GPUs deployed. NCCL bandwidth of 317 GB/s and throughput of 387 TFLOP per device per second. Over 200 experiments completed. A 27B parameter model trained in 10 days using 32 nodes."
SU010 Firmus Technologies AI Singapore and Firmus Technologies partner to advance sustainable AI infrastructure "AISG researchers and collaborators will gain access to Singapore-based H200 GPUs within SMC's Sustainable AI Factory. These NVIDIA-powered systems feature next-generation immersion cooling technology, reducing operational costs and energy consumption by up to 50%."
SU011 Firmus Technologies Firmus Expands Project Southgate Across Australia
SU012 Firmus Technologies Firmus Technologies expands regional AI access through NVIDIA DGX Cloud Lepton
SU013 Firmus Technologies Tasmanian world-first AI Factory Zone clears path for Firmus' Project Southgate
SU014 Firmus Technologies Firmus Signs Multi-Year Agreement with Global Hyperscale Customer at Project Southgate "This agreement is a defining milestone for Firmus and for Australia's role in the global AI infrastructure landscape. As the second major contract secured under Project Southgate, it reflects growing market conviction in our approach."
SU015 Firmus Technologies Firmus' SMC ranks among world's top GPU clouds, outperforming global giants
SU016 Firmus Technologies AI Cloud — Firmus
SU017 W.Media Firmus secures "multi-billion-dollar" AI contract for Melbourne
SU018 IT Brief Australia Firmus inks multi-year Nvidia AI factory deal in Melbourne
SU019 InnovationAus Firmus locks in hyperscale customer for $73bn AI play
SU020 SmartCompany Firmus expands Australian 'AI factory' plan with $73.3 billion from Nvidia, CDC
SU021 Startup Daily Green-powered AI data centre plan goes national with $73.3 billion project across 4 Australian capitals
SU022 VAST Data Firmus Technologies Group Selects VAST AI Operating System to Power Sovereign, Energy-Efficient AI Factories
SU023 Silicon Valley Investclub Firmus Technologies — Enhanced Profile
SU024 Data Centre Dynamics Australia's Firmus launches immersion-cooled bare metal AI cloud with STT GDC
SU025 ShareCafe Australian CIO Cautions on Firmus Valuation "Hayden Beamish, CIO at Endeavor Asset Management, warns the eventual ASX listing will likely be priced too expensively, that the easy money has already been made, and described the pre-pivot Firmus as a bitcoin miner with no legitimate business model."
SR001 Montgomery Investment Management $6 billion hype or structural risk? Inside the Firmus ASX listing "Combine that with concerns about the absence of publicly transparent, long-term, binding customer contracts and the widely reported personal history of the company's co-founder, whose insider trading conviction presents a significant governance hurdle for institutional funds, and then add the absence of any escrow arrangements."
SR002 Cyber and Infrastructure Security Centre (CISC) SOCI Act 2018 obligations for data storage and processing sector
SR003 Mingtiandi Blackstone Leads $10B in Debt Financing for Australia's Firmus
SR004 Gilbert + Tobin Strengthening critical infrastructure resilience: proposed amendments to SOCI Act CIRMP Rules "The Government proposes a new directions power to enable coordinated action where a specific vendor (or its products, equipment, services or technologies) presents a material risk to national security."
SR005 NAOS Asset Management NAOS Quarterly Investment Report Q3 FY26
SR006 Corrs Chambers Westgarth Australian privacy compliance: four key developments in 2026
SR007 Finance News Network Australian CIO Cautions on Firmus Valuation "Beamish now fears the 'easy money' for Firmus has already been made, and its eventual ASX listing... will likely be priced too expensively."
SR008 Australian Securities and Investments Commission (ASIC) 26-059MR ASIC publishes ASX Inquiry Panel Final Report and acknowledges observations
SR009 Market Index NCC:ASX Announcement — NAOS Quarterly Investment Report Q3 FY26
SR010 Protectera Australia's New Critical Infrastructure Security Rules Explained (2026)
SR011 Stocks Down Under The ASX's Good Fame and Character Test: Firmus and Oliver Curtis
SR012 Roger Montgomery Investment Management $6 billion hype or structural risk? Inside the Firmus ASX listing
SR013 Wilson Asset Management Oliver Curtis's Firmus eyes $12bn float with no escrow restrictions for early investors
SR014 The Daily Perspective The Ex-Convict and the $10 Billion AI Bet Shaking Australian Markets
SR015 The Witness Oliver Curtis and Firmus, the AI infrastructure provider heading towards a $6 billion listing
SR016 ABC News Australia Questions over electricity deal between Tas Hydro and Firmus AI
SR017 ABC News Australia AI data centres pressuring energy transition, Greenpeace says
SR018 Allens Allens advises Blackstone on US$10B financing of Firmus Group
SR019 Milbank LLP Milbank advises Firmus in landmark US$10B financing led by Blackstone and Coatue
SR020 Bloomberg Data Center Operator Firmus Seeks $120 Million Private Loan Firmus seeks $120 million private loan from Nomura amid institutional caution about US export-control risks on Nvidia GPUs
SR021 TechCrunch Firmus, the 'Southgate' AI data center builder backed by Nvidia, hits $5.5B valuation
SR022 The Next Web Nvidia-backed Firmus targets $2bn ASX IPO after $505m raise and Blackstone debt
SR023 Firmus Technologies Firmus Technologies – AI Cloud and Infrastructure
SR024 Nine / Australian Financial Review AFR Rich List 2026: Oliver Curtis debuts with A$1.25B fortune
SR025 Startup Daily AI data centre startup Firmus eyes off another $725 million raise as it readies to IPO on the ASX
SR026 InnovationAus Firmus, CDC and Nvidia reveal $73bn AI factory plan — locks in hyperscale customer
SR027 Livewire Markets NAOS investor note: Firmus governance and market risk assessment
SR028 Startup Daily AI data centre startup Firmus valuation trebles to $6 billion in just 8 weeks after raising another $500 million
SR029 CDC Data Centres Driving Australia's AI Future with CDC, Firmus and Nvidia
SR030 Maas Group Holdings (ASX: MGH) Maas Group Holdings — Strategic Investment in Firmus Technologies
SV001 Montgomery Investment Management (Roger Montgomery) $6 billion hype or structural risk? Inside the Firmus ASX listing the gap between hype and reality has rarely looked wider for an ASX listing
SV002 Sharecafe Australian CIO Cautions on Firmus Valuation the easy money for Firmus has already been made
SV003 Wilson Asset Management Oliver Curtis's Firmus eyes $12bn float with no escrow restrictions for early investors early investors who have already realized a 300% valuation markup are management accounts, IPO prospectus financials, and contract term sheets
SV004 Montinvest $6 billion hype or structural risk? Inside the Firmus ASX listing the lack of share dealing restrictions (escrow) suggests they will look to realise those gains by selling to IPO buyers
SV005 Bloomberg Nvidia-Backed Data Center Builder Firmus Raises $505 Million
SV006 The Next Web Nvidia-backed Firmus targets $2bn ASX IPO after $505m raise and Blackstone debt
SV007 SmartCompany / StartupSmart AI data centre startup Firmus looks to raise another $725 million at $8 billion valuation
SV008 Startup Daily AI data centre startup Firmus eyes off another $725 million raise as it readies to IPO on the ASX
SV009 Sacra CoreWeave revenue, valuation & funding
SV010 Multiples.vc CoreWeave – Public Comps and Valuation Multiples
SV011 Multiples.vc CoreWeave IPO Valuation Deep Dive
SV012 Business Wire / CoreWeave CoreWeave Reports Strong First Quarter 2026 Results Q1 2026 revenue of $2.08 billion, up 111.6% year-over-year; revenue backlog reached $99.4 billion
SV013 Milbank LLP Milbank Advises Firmus in Landmark US$10B Financing Led by Blackstone and Coatue Structured to support the staged acquisition and deployment of high-performance AI compute…the debt facility will be available to be drawn against executed long-term customer contracts
SV014 Data Center Dynamics Firmus secures $10bn debt financing facility for Project Southgate AI factory roll-out This transaction is one of the largest private debt financings in Australian history
SV015 Stocks Down Under The ASX's Good Fame and Character Test: The Potential Listing of Oliver Curtis's 6bn AI Company Firmus
SV016 MF & Co Asset Management (MFAM) NEXTDC (ASX: NXT) Contracted Capacity Jumps 60% to 667MW — Investor Presentation
SV017 S&P Global Market Intelligence Rising private infrastructure fundraising; data center investment boom
SV018 Finance News Network Australian CIO Cautions on Firmus Valuation
SV019 Data Center Knowledge Data Center M&A Outlook Robust in 2026 Despite Power, AI Risks
SV020 Silicon Valley Invest Club Firmus Technologies Raises $505 Million at a $5.5 Billion Valuation
SV021 The AI World Organisation Firmus Raises $505M, Eyes $2B ASX IPO in 2026
SV022 MarketScreener Valuation CoreWeave, Inc.
SV023 ValueInvesting.io NXT.AX EV/EBITDA — NEXTDC Ltd
SV024 Forge Global Lambda Upcoming IPO & Private Stock Price
SV025 Allens (law firm) Allens advises Blackstone on US$10bn financing of Firmus Group
SV026 Kirkland & Ellis LLP Kirkland Represents Blackstone on AI Infrastructure Company Firmus' US$10 Billion Debt Financing Facility
SV027 Data Center Dynamics The biggest data center M&A deals of 2025 — DCD 2025 M&A in Review
SV028 Sacra Lambda Labs revenue, valuation & funding
SV029 Unite.AI Nvidia-Backed Firmus Hits $5.5B Valuation With $505M Raise
SV030 AI Reporter News CoreWeave (NASDAQ:CRWV) Reports Strong Q1 2026 Financial Results; Revenue Backlog Reaches $100 Billion