Nscale
垂直整合的 AI 基础设施平台,融资渠道强,但公开运营披露有限。
Nscale 已搭建起欧洲最具野心的 AI 基础设施平台之一,拥有主权 AI 定位、顶级投资人和 Microsoft 锚定的需求;但公开资料仍未披露收入,客户高度集中,杠杆持续上升,英国规划审批的阻力也让当前 146 亿美元估值处于偏高区间。
覆盖范围与披露说明
截至 2026-10-18,公开来源确认公司获得大额融资、拥有大量已签约需求,并采取全栈基础设施战略;但公开资料仍未充分披露经审计收入、利润率、客户集中度细节或合同经济性,单靠公开证据无法完成投资判断。
封面要素
公司概况
Nscale 是一家总部位于伦敦的 AI 基础设施公司,围绕电力、数据中心、GPU 算力、网络和编排软件搭建垂直整合的全栈。公司以主权 AI 为定位,同时拥有大额战略融资、长期客户合同,以及不断扩大的产品组合,覆盖专用基础设施和托管云服务。
- 创立地点
- London, United Kingdom
- 总部
- London, United Kingdom
- 产品
- 面向训练、微调和推理工作负载的专用 GPU 基础设施、AI 数据中心容量、托管 AI 云服务、存储和编排软件。
- 客户
- 企业、政府、超大规模云厂商和前沿 AI 实验室。
- 商业模式
- 长期 take-or-pay 基础设施合同,叠加按 GPU 小时计费的云服务和存储定价。
- 阶段
- Series C
- 融资情况
- 2026 年 3 月完成 Series C,估值 $14.6B;2026 年另有 GPU 抵押债务和可转换融资披露。
执行摘要
主要优势
- 大量已签约需求和融资渠道支撑大规模建设。
- 从电力、数据中心、算力到软件的垂直整合,构成了差异化的战略能力。
- Microsoft、NVIDIA 等头部合作方验证了该平台的重要性。
主要风险
- 执行和并网投运延迟,可能让积压订单无法转化为实际收入。
- 客户集中、合同经济性不透明,削弱了对收入质量的判断信心。
- 相对于已披露的经营指标,杠杆和资本开支强度仍然偏高。
未决问题
- 公开资料没有审计收入、ARR、利润率、EBITDA、现金或现金消耗披露,因此估值只能靠推断支撑,尚不足以纳入投资判断。
- Microsoft 合同条款、利用率保证、终止权和定价结构均未公开,因而无法准确衡量客户集中度和收入质量。
- $1.4B DDTL 和 $790M 挪威融资的债务契约、抵押品触发条件和还款安排均未披露。
- Loughton 最终的规划审批和并网投运路径仍未解决,英国主权 AI 的时间表和资本开支假设因此暴露在风险之下。
- 除 Microsoft、BT 以及项目层面的 OpenAI 参与外,公开资料中关于客户多元化的证据仍然有限。
目录
01公司概览
1.1 身份、阶段与运营模式
Nscale 将自己定位为全栈 AI 云平台,而不是单纯出租 GPU 的房东。公司当前材料称,其在专用 GPU 基础设施、面向 AI 的数据中心以及专用能源和电力基础设施之上,提供用于推理、Kubernetes 和 Slurm 的托管云层;独立资料则称公司总部在伦敦、成立于 2024 年。外部证据最有力的阶段标志是 2026 年 3 月的 Series C:这说明 Nscale 已是 Series C 后的成长公司,而非仍处于试验阶段的基础设施初创企业。公开足迹也不止一个站点:审阅来源将其运营或容量分布在英国、美国、挪威、葡萄牙和冰岛,公司对外叙事覆盖欧洲、北美和亚洲。即便如此,审阅到的公开记录仍未披露经审计收入、收入 收入运行率、客户数量或员工人数。[CO001, CO002, CO003, CO004, CO009, CO013, CO016]
| 指标 | 数值 / 状态 | 截至 | 置信度 | 缺口 / 下一步 |
|---|---|---|---|---|
| 总部 | 英国伦敦 | 2026-03-09 至 2026-10-18 的公开记录 | 中 | |
| 成立年份 | 2024 | 2026-03-09 的第三方资料 | 中 | |
| 当前阶段 | Series C | 2026-03-09 | 高 | |
| 最近披露的估值 | $14.6B | 2026-03-09 | 高 | |
| 最近披露的股权融资 | $2.0B Series C | 2026-03-09 | 高 | |
| 后续债务之前的已披露股权融资 | >$3.7B | 2026-06-14 公司画像综合 | 中 | |
| 额外披露的债务 / 信贷 | $1.4B GPU 抵押延迟提款定期贷款 | 2026-02 / 2026-03 报道 | 中 | 获取融资工具文件、到期日和契约包。 |
| 公开足迹 | 英国、美国、挪威、葡萄牙和冰岛 | 2026 年公开资料 | 中 | 按地点核实当前运营站点与规划站点,以及已上线 MW。 |
| 收入 / 运行率 | 截至 2026-10-18 审阅的公开来源 | 低 | 索取经审计收入、运行率、利润率和现金消耗披露。 | |
| 客户 / 员工人数 | 截至 2026-10-18 审阅的公开来源 | 低 | 索取披露的客户数量、头部客户组合和当前员工基数。 |
Null 表示截至 2026-10-18,审阅的公开记录未披露可支持的数值;资料支持融资时间线,但不支持运营指标。
[CO003, CO004, CO009, CO011, CO012, CO013, CO016]1.2 领导层、治理与利益相关者结构
公司仍由创始人主导:Josh Payne 继续担任 CEO,也是融资和战略的公开代言人。当前领导层页面显示,财务、运营、基础设施、法律、安全和商业岗位已经明显扩充,这部分缓解了创始人集中度风险;毕竟 Nscale 正在多个司法辖区、多个技术领域推进资本密集型建设。2026 年治理也趋于成熟。Nscale 宣布 Sheryl Sandberg、Susan Decker 和 Nick Clegg 加入董事会;原董事会已有 Payne、Rael Nurick、Jacob Leschly 和 Aker CEO Oyvind Eriksen。同时,Aker Nscale 合资企业并入 Nscale,Aker 仍为重要股东。整体看,公司仍围绕创始人运转,但已逐步纳入资深运营者、政策人士和投资人的关系网络。[CO005, CO006, CO007, CO008, CO014]
| 人员 | 职务 | 背景 / 覆盖范围 | 创始人—市场匹配或职能覆盖 | 关键人物依赖 |
|---|---|---|---|---|
| Josh Payne | CEO、创始人 | 曾在招聘、资本市场、超级计算基础设施、能源、数据中心和硬件供应链领域连续创业和运营。 | 与 Nscale 资本密集型 AI 基础设施论点直接匹配。 | 高:仍是具名的战略和融资发言人。 |
| Alice Takhtajan | 首席财务官 | 25+ 年 JP Morgan 经历,参与 150+ 笔科技股权交易。 | 为快速融资和潜在 IPO 准备增加公开资本市场与融资深度。 | 中:让财务班底不再只依靠创始人。 |
| Justin Osofsky | 首席运营官 | 曾在 Meta 和 Instagram 负责合作伙伴、政策及全球运营。 | 支持规模化运营节奏、业务开发和跨职能执行。 | 中:执行班底更广,但叙事仍由创始人锚定。 |
| Nidhi Chappell | AI 基础设施总裁 | 前 Microsoft AI 基础设施负责人;公开简介称其管理过 2m 个 GPU。 | 与超大规模 AI 部署、GPU 运营和企业交付可信度直接相关。 | 中:对冲创始人集中度的关键技术运营力量。 |
| Sam Huckaby | 数据中心总裁 | 前 Oracle Cloud Infrastructure 数据中心负责人,拥有超大规模建设和运营经验。 | 为多站点扩张补充建设和园区交付经验。 | 中:交付风险仍由电力、建设和融资共同承担。 |
本表优先列出公开领导层页面中与融资和交付最直接相关的创始人及高级运营者;并非完整组织架构。
[CO005, CO006, CO007, CO008]| 利益相关者 | 角色 | 控制权或经济重要性 | 公开证据 | 尽调要求 |
|---|---|---|---|---|
| Josh Payne | 创始人兼 CEO | 兼具战略叙事、融资可见度和执行问责。 | 官方页面和融资公告都将 Payne 置于身份与融资主张的中心。 | 澄清创始人投票控制权、董事会影响力及 CEO 以下的继任深度。 |
| Aker ASA | 领投方兼战略合作伙伴 | 领投 Series B、联合领投 Series C,并在合资企业并入后仍是重要股东。 | 官方和新闻报道将 Aker 与融资及挪威交付治理同时联系起来。 | 获取持股比例、治理权,以及项目层面的保护或义务。 |
| 8090 Industries | Series C 联合领投方 | 作为已披露最大轮次的新联合领投方,可能影响后期治理。 | 官方和新闻报道均点名其为 Series C 联合领投方。 | 确认董事会权利、信息权和预期持有期。 |
| Microsoft | 锚定客户 / 交易对手 | 代表公开释放的需求,也是未来收入质量可能的主要集中节点。 | 审阅报道将 Nscale 的 Loughton 项目与 Microsoft 锚定租户联系起来,资料还提到更广泛的 Microsoft 相关部署。 | 索取合同经济性、利用率承诺、续约权和集中度敞口。 |
| NVIDIA | 战略投资者与生态依赖 | 出现在 Series C 银团中,也是 GPU 中心基础设施论点的支柱。 | 官方轮次披露将 NVIDIA 列为投资人;公开定位高度以 GPU 为中心。 | 测试配额安全性、定价条款,以及对产品路线或供应冲击的敏感度。 |
本图谱根据当前融资、治理、客户或执行证据中明确点名的主体选择性构建;并非完整股权结构表或交易对手登记册。
[CO005, CO009, CO010, CO014, CO015]1.3 融资轨迹、里程碑与当前执行信号
Nscale 的融资节奏是公开市场支持最清晰的证据。官方和第三方报道均指向 2026 年 3 月由 Aker ASA 与 8090 Industries 领投的 $2.0B Series C,估值 $14.6B;广泛的银团还包括 Dell、Nokia 和 NVIDIA 等战略投资者。资料显示此前融资也快速阶梯式上升:2023 年 12 月种子轮、2024 年 12 月 Series A、2025 年 9 月 Series B,以及 2025 年 10 月 Series C 前 SAFE;后续债务之前,已披露股权融资超过 $3.7B。但融资势头并未消除执行风险。审阅报道称,Loughton 数据中心项目已有规划许可和约定的 90MW 电网接入,可电力无法及时到位,迫使 Nscale 寻求现场发电,尽管公司已有融资和客户支持。因此,后续章节应把融资实力视为真实优势,但不能据此认定部署时间、利用率或经济性已经去风险。[CO009, CO010, CO011, CO012, CO014, CO015, CO016]
公开披露的融资、治理和执行里程碑显示,资本规模迅速攀升后,基础设施交付约束开始显现。
种子轮已披露 [CO011]
A 轮已披露 [CO011]
Aker 在挪威启动合资企业 [CO014]
B 轮已披露 [CO011]
C 轮前 SAFE 已完成 [CO011]
新增 GPU 抵押债务 [CO012]
C 轮重置规模 [CO009]
董事会扩张,Aker 合资企业并入 [CO007, CO014]
Loughton 供电延迟浮现 [CO015]
1.4 展示项
02市场分析
2.1 市场边界:什么才算 Nscale 的市场
公开证据显示,Nscale 参与的是比泛称“云”更窄的 AI 基础设施市场。公司材料一贯描述其覆盖能源、数据中心、GPU 算力和软件的垂直整合方案,同时点名 AI 原生公司、企业和政府为目标客户。由此,Nscale 所处市场应定义为主权或合规敏感型 AI 基础设施、大规模训练和推理容量,以及运行这些容量所需的编排层,而不是广义企业 SaaS、消费级 AI 应用或普通托管。公开记录能支持按细分市场判断买方角色,却无法判断合同细节,因此市场边界基本成立,但具体预算决策权仍只有部分证据。[CM001, CM002, CM003, CM015]
| 细分市场 / 类别 | 纳入支出 | 排除支出 | 买方 / 付款方 | 相关性 |
|---|---|---|---|---|
| 主权 AI 云与受监管企业基础设施 | AI 训练与推理容量、私有 AI 云编排,以及为数据驻留或合规要求而本地托管的算力 | 消费级 AI 应用、通用 SaaS,以及客户自有但并非从第三方基础设施提供商购买的本地硬件 | 政府、受监管企业、科研机构 | 与 Nscale 在英国和欧洲明确强调的主权与安全私有算力定位一致。 |
| 超大规模云厂商溢出容量 | 为基础模型或超大规模云厂商工作负载签约部署的多站点 GPU,以及 AI 园区容量 | 专用 AI 容量承诺之外的通用公有云用量 | 超大规模云厂商与前沿模型运营商 | 已披露信息中,Microsoft 最清楚地证明 Nscale 正在承接超大型买家的溢出需求。 |
| 企业托管 AI 平台 | 编排、托管推理与训练服务,以及与基础设施捆绑、可满足合规要求的软件 | 不含捆绑基础设施容量的独立开发者工具 | 企业 AI 平台团队 | Anyscale 与 VAST 的证据表明,差异化正在向上层软件迁移,而不是停留在单纯的 GPU 租赁。 |
| 排除在当前受证据约束的 SAM 之外 | n/a | 通用托管、消费级 AI、自助式 SMB 开发者云,以及未向第三方供应商开放的内部专用 AI 基础设施资本开支 | n/a | 公开来源尚未显示 Nscale 已大规模拿下这些预算,因此不能将其视为当前近期可争取的需求。 |
市场边界综合了公司定位、已披露部署和合作伙伴证据;排除的行反映公开证明仍不足,并非主张 Nscale 永远无法参与。
[CM001, CM002, CM003, CM004, CM011, CM015]2.2 多种测算口径显示市场很大,但无法形成干净的公开 SAM 或 SOM
可服务市场显然很大,但公开证据无法为 Nscale 提供一套口径一致的 TAM、SAM 和 SOM。独立来源给出了多个重叠视角:全球 GPU 云预测、更广义的 AI 基础设施预测、主权云增长率,以及欧洲电力需求等基础设施瓶颈。这些口径都指向结构性强增长,但分母不同,不能相加,也不能直接换算为 Nscale 近期可变现的份额。最具体的需求证据不是市场报告,而是 Microsoft 主导的部署足迹,以及 Nscale 自己披露的英国和 pipeline 容量;它们显示公司当前正从哪里争夺份额。[CM004, CM005, CM006, CM007, CM008, CM009, CM012, CM013, CM014]
| 发布方 | 年份 / 期间 | 地域 | 数值 | CAGR | 方法 | 置信度 | 局限 |
|---|---|---|---|---|---|---|---|
| Next Move Strategy Consulting | 2026-2035 | 全球 | 2026 年 USD 132.0B -> 2035 年 USD 779.0B | 21.8% | GPU 云市场预测 | 中 | 广义新型云类别,并非 Nscale 近期的精确细分市场。 |
| Ainvest 引用外部市场研究 | 2025-2030 | 全球 | 2025 年 USD 158.3B -> 2030 年 USD 418.8B | 21.5% | AI 基础设施硬件与系统预测 | 中 | 比 GPU 云更宽的相邻市场视角,可能与其他支出重叠。 |
| Next Move Strategy Consulting | 2026-2035 | 全球 | 托管 AI 平台细分市场 | 25.9% | GPU 云内部的细分增长视角 | 中 | 已披露增长率,但未披露起始收入基数。 |
| Next Move Strategy Consulting | 2026-2035 | 全球 | 主权云部署模式 | 29.4% | 部署模式增长视角 | 中 | 反映需求方向,不代表 Nscale 可变现支出的绝对规模。 |
| Acquinox / 引用的行业分析 | 2024-2030 | 欧洲 | 96 TWh -> 168 TWh 数据中心电力需求 | n/a(总增长 75%) | 基础设施瓶颈视角 | 中 | 电力需求不是收入 TAM,但它是交付能力的关键约束。 |
这些行是规模测算视角,而不是可相加的 TAM/SAM/SOM 体系;各来源使用了不同的分母、单位和类别边界。
[CM005, CM007, CM008, CM009, CM012, CM013, CM014]按序排列的市场视角显示,上层是广泛需求,下层是 Nscale 特有的交付约束;各层用于指示方向,不是可相加的 SAM / SOM 切片。
各层宽度仅表示层级,不代表数值比例。
金字塔顺序呈现了从广泛需求到 Nscale 受限交付能力的分析收窄;各层单位不同(USD、CAGR、TWh、MW),因此不应将数值相加或视为嵌套市场份额。
[CM005, CM007, CM008, CM009, CM010, CM012, CM013, CM014]2.3 主权需求和超大规模云溢出拉动采用,但电力与验证形成门槛
公开记录中最有力的采用逻辑,是主权算力需求、超大规模云溢出需求,以及市场转向全栈托管 AI 平台的共同作用。买方若需要本地控制、大规模 GPU 部署,以及减少基础设施与编排整合工作,Nscale 的定位最有优势。但限制因素不是 headline TAM,而是交付摩擦。独立报道显示,电网通电已成为英国数据中心项目的门槛,Nscale 的 Loughton 园区已经暴露在这一约束下。同样重要的是,公开证据仍未披露定价、利用率保证或非 Microsoft 业务组合,无法把战略需求转成自下而上的市场份额论证。因此,采用时间和估值相关性比广泛市场统计所暗示的更依赖具体路径。[CM004, CM009, CM010, CM011, CM012, CM013, CM015]
| 细分市场 | 买方 | 用户 | 付款方 | 工作流 | 预算负责人 | 采用触发因素 |
|---|---|---|---|---|---|---|
| 超大规模云厂商溢出部署 | 云基础设施或容量团队 | 模型训练与推理运营 | AI 基础设施预算 | 跨多个站点签约的 GPU 容量 | 基础设施或 AI 平台负责人(推断) | 内部建设速度跟不上需求 |
| 受监管企业主权 AI | CIO、CDO 或安全负责人 | 企业 AI 与数据科学团队 | 数字化转型或基础设施预算 | 私有或合规的训练与推理工作负载 | CIO 或 AI 平台负责人(推断) | 数据驻留、安全与运营控制要求 |
| 公共部门或科研容量项目 | 公共数字化或科研项目发起方 | 科研与公共服务模型团队 | 公共 IT 或科研资金 | 国家或机构级 AI 容量建设 | 项目发起方或 CIO(推断) | 对本地算力容量及符合主权要求的基础设施的需求 |
| 企业平台与编排买方 | 平台工程或 MLOps 负责人 | 应用、数据与模型团队 | AI 平台预算 | 托管编排与捆绑 GPU 供应 | CTO 或平台负责人(推断) | 希望避免拼接不同的软件与基础设施供应商 |
买方、付款方和预算负责人的标签,是根据已披露部署、公司定位和合作伙伴软件主张作出的角色层面推断;公开来源未披露账户层面的合同签署方或按细分市场划分的定价。
[CM002, CM003, CM004, CM011, CM013, CM015]2.4 展示项
03竞争者
3.1 竞争格局的广度比一对一 GPU 比较更重要
相关竞争者不只是知名新云厂商。证据显示,Nscale 面对超大规模云厂商、专业 GPU 云、软件驱动的基础设施对手、可移植软件平台,以及把第三方基础设施与开源编排结合起来的自建替代方案。买方可以用多种方式解决同一个问题:购买打包容量、自建全栈,或保持编排可移植而替换底层基础设施。收购 Anyscale 让 Nscale 在软件驱动的评估中更具可信度,但并未消除重视可移植性或运营控制、而非单一供应商便利性的客户的替代路径。[CP002, CP003, CP004, CP011, CP014]
| 竞争对手 / 替代方案 | 类别 | 规模 / 融资信号 | 目标细分市场 | 差异化 | 局限 |
|---|---|---|---|---|---|
| Nscale | 垂直整合型新型云 | 2026 年 3 月 $2.0B C 轮融资;多个已锁定的 GPU 和数据中心设施;合作伙伴称计划在全球部署 300,000 个 Grace Blackwell GPU | 需要主权或专用 AI 容量的企业、政府与 AI 原生团队 | 自有电力、数据中心、算力及不断扩大的软件层 | 执行仍受通电时点、杠杆和锚定合同集中度影响 |
| AWS / Google Cloud | 现有超大规模云厂商 | 独立报道将其列为 Nscale 捆绑方案的直接挑战者 | 已在大型公有云生态中运营的广泛企业买方 | 既有企业关系与广泛的云服务覆盖 | 已分配证据未显示其具备同样的自有物理栈叙事或主权 AI 聚焦 |
| CoreWeave / Lambda Labs | 专业 GPU 云 | 独立报道将其列为专业 GPU 云竞争者 | 优先追求快速获取 GPU 和熟悉开发者工作流的 AI 团队 | 聚焦 GPU 云方案,而非完整的工业化建设 | 已分配证据未显示其具备同等的电力与软件整合 |
| Nebius 及其他收购软件的新型云厂商 | 向上层软件迁移的基础设施同业 | 独立报道称 Nebius 以 $643M 收购了工作负载效率公司 Eigen AI | 寻求整合算力与软件层的 AI 团队 | 以收购驱动的软件整合类似 Nscale 的 Anyscale 策略 | 已分配证据未提供可比的容量、定价或合同披露 |
| 第三方基础设施上的 Anyscale | 可移植软件替代方案 | Anyscale 表示,现有客户仍可自由选择基础设施 | 优先考虑编排可移植性的机器学习团队 | 基于 Ray 的托管服务,不要求承诺使用 Nscale 基础设施 | 未披露商业定价和运营对比基准 |
| 内部建设 + 可移植编排 | 现状 / 替代方案 | 使用第三方基础设施及 Ray 等工具,而非单一供应商的技术栈 | 希望最大程度控制工作负载部署位置的成熟买方 | 最高的可移植性与治理控制 | 客户需自行组装运营、容量采购与集成 |
各行仅综合已分配的公开证据;当来源描述的是竞争对手类别而非信息完整披露的一对一同业时,采用分组替代方案。
[CP001, CP004, CP005, CP006, CP010, CP011, CP014]3.2 差异化增强了,但并非排他
Nscale 最清晰的公开差异化,是对电力、数据中心、GPU 算力和软件的垂直掌控,并得到合作伙伴叙事和 Anyscale 收购目的的强化。这使 Nscale 不同于主要转售容量,或只强调能源、主权托管或芯片可得性的供应商。但它加入的软件层并不像应用套件那样封闭:Ray 仍由 PyTorch Foundation 治理并保持开源,Anyscale 也表示客户仍可在其他基础设施上运行工作负载。结果是更强的打包方案,而不是完全锁定的生态。[CP001, CP002, CP005, CP009, CP012, CP013, CP015]
| 购买标准 | Nscale | 现有超大规模云厂商 | 专业 GPU 云 | 内部建设 / 可移植技术栈 |
|---|---|---|---|---|
| 物理栈所有权 | 公开定位为自有电力、数据中心、算力和软件 | 已分配来源中未知 | 已分配来源中未知 | 买方直接控制技术栈,但必须自行组装 |
| 软件 / 编排深度 | 由 Anyscale 收购和基于 Ray 的编排层扩展 | 已分配来源中未知 | 已分配来源中未知 | 取决于买方选择的编排工具和工程团队 |
| 可移植性 / 多归属 | Anyscale 的随处运行定位提升了可移植性,但仍是捆绑式全栈方案 | 已分配来源中未知 | 已分配来源中未知 | 若由买方自行管理,则对工作负载部署位置拥有最高控制权 |
| 目标客户 / 信任定位 | 官方定位面向从业者、企业与政府 | 已分配来源中未知 | 已分配来源中未知 | 取决于买方选择的站点、运营商和合规设计 |
| 交付依赖 | 竞争上行空间取决于大型站点能否按期通电并扩容 | 已分配来源中未知 | 已分配来源中未知 | 执行负担从供应商转移给买方 |
“未知”单元格表示已分配资料池没有该竞争类别的直接可比公开证据;应将矩阵理解为已记录的证据边界,而不是隐含的负面评分。
[CP001, CP002, CP005, CP009, CP010, CP012, CP014, CP015]有证据支撑的序数定位图,展示主要替代类别在基础设施栈所有权和买方可迁移性上的位置。
- X: 垂直整合深度
- low: 轻资产 / 组装式
- high: 拥有更多从电力到软件的栈
- description: Ordinal scoring from assigned public evidence only.
- scale: ordinal
- Y: 买方可迁移性
- low: 更受供应商绑定
- high: 更由买方控制 / 可多云部署
- description: Ordinal scoring from assigned public evidence only.
- scale: ordinal
| 条目 | 垂直整合深度 (1–5) | 买方可迁移性 (1–5) | 说明 |
|---|---|---|---|
| 1. Nscale | 5 | 3 | 拥有电力、数据中心、算力和软件,但捆绑式交付仍会造成一定锁定。 |
| 2. 现有超大规模云厂商 | 2 | 2 | 被列为直接挑战者,但指定证据没有保留可比的所有权或可迁移性细节。 |
| 3. 专业 GPU 云 | 2 | 3 | 专注算力的替代方案被列为竞争者,但指定证据显示其基础设施栈整合程度较低。 |
| 4. 购买软件的新型云厂商 | 3 | 2 | 据报道,Nebius 收购 Eigen AI,显示另一家基础设施提供商正向工作负载软件靠拢。 |
| 5. 在第三方云上的 Anyscale | 1 | 5 | Anyscale 表示,客户仍可自由选择底层基础设施。 |
| 6. 内部构建 + Ray | 1 | 5 | 可迁移的编排和买方自行运行的基础设施最大化控制力,但也把整合负担转移到内部。 |
点位坐标是以保留来源为锚的序数判断,应理解为相对定位,而非测量得出的基准分数。
[CP001, CP003, CP004, CP005, CP009, CP011, CP013, CP014]3.3 价格透明度集中在栈的边缘,切换权力则在企业合同里
整个竞争市场的公开价格透明度都很低,但证据提供了一个有用信号:Nscale 有目录级的按需 GPU 租赁和 spot 容量挂牌价,而更大的私有云和项目融资关系依赖企业合同,公开资料无法按净价比较。这带来两个同时成立的事实。第一,Nscale 能为试验客户展示易于进入的起点。第二,最关键的切换成本不在小时费率,而在多年基础设施承诺、锚定客户关系,以及平台完成整合后迁移工作负载的运营负担。[CP006, CP007, CP008, CP013, CP014, CP016]
| 选项 | 公开价格 / 合同信号 | 包含的能力 | 未知项 / 注意事项 | 含义 |
|---|---|---|---|---|
| Nscale 按需租赁 | AMD MI300X 价格为 $1.10/hr,NVIDIA H100 价格为 $3.85/hr;提供竞价实例 | 按小时计费的裸金属 GPU 租赁,含企业级 SLA | 目录层面的信号,并非协商后的企业合同 | 可见的入门价格有助于漏斗顶部比较,但不能证明实际经济性 |
| Nscale 竞价租赁 | 提供竞价实例;该目录声称可为可中断工作负载节省 30–70% | 同一项列出的 GPU 租赁服务,使用可中断容量 | 第三方目录主张;没有留存的交易记录或基准来验证实际节省 | 竞价容量为能容忍中断的工作负载提供了低成本替代方案 |
| Nscale 大型项目融资部署 | Ward County 融资由与 Microsoft 签订的 6 年期、按可用性计价的固定价格 GPU 即服务 take-or-pay 合同支持 | 规模化容量、融资杠杆和长期客户承诺 | 客户单位经济性和删节后的合同定价未公开 | Nscale 可以把锚定合同变成分销武器,但集中度也会随之上升 |
| 第三方基础设施上的 Anyscale 软件 | 已分配证据中未留存公开定价 | 可在非 Nscale 基础设施上运行的可移植编排、训练、推理和数据处理层 | 留存来源未披露商业打包方式 | 对抗基础设施锁定的买方确实可以选择可移植路径作为替代 |
| 内部建设 / 托管路线 | 没有可比的公开单位价格;经济性体现为资本开支、电力和融资决策,而不是小时费率 | 买方对站点、容量和工具拥有最大控制权 | 电网接入时点和建设复杂度可能主导成本与进度 | 替代方案存在,但切换成本从供应商依赖转为运营负担 |
本表保留已分配证据中的不对称性:可见的目录价格主要存在于 Nscale 方案的边缘,而战略性企业方案和替代方案的经济性仍部分不透明。
[CP003, CP007, CP008, CP010, CP013, CP014, CP016]3.4 护城河能否持久,既取决于定位,也取决于执行
最不利的解读是,Nscale 的护城河可能先在演示文稿里变宽,再在运营中变宽。独立报道和融资披露显示公司确有规模,但也凸显其依赖 Microsoft 支持的部署、债务融资建设,以及可能在场地已获融资和许可后仍拖慢交付的通电时间线。这并不抹去掌握更多栈的战略价值;它说明护城河仍取决于按时上线容量,并从少数高影响合同之外实现多元化。[CP006, CP007, CP010, CP013, CP014]
3.5 展示项
04财务
4.1 收入模式与公开牵引力
公开披露描述的是两层变现模式,而不是单一的批发托管业务。Nscale Infrastructure 是当前核心产品,向超大规模云厂商和前沿模型客户销售专用 GPU 集群及数据中心容量,采用长期 take-or-pay 合同。Nscale Cloud 位于其上,按 GPU 小时向客户收费,存储则按 GB 每月单独计费。如果容量确实交付,这种结构能带来较强的前向可见性;但公开记录仍缺少分部收入拆分、实现的每 GPU 小时价格,或足以证明云层软件和服务相对基础设施层已成为重要贡献者的证据。 已披露的牵引力指标进一步区分了已售需求和实际收入。截至 2026-08-31,Nscale 报告活跃及已签约 TCV 为 $103.4B,活跃或已签约 GPU 为 461,000 个;但其中只有约 $2.6B TCV 和约 25,000 个 GPU 已经活跃。因此,2026 年上半年 $140.6M 的收入,更适合看作大型已预订 pipeline 的早期转化点,而不是完整合同积压已产生现金收益的证明。投资判断的关键不是有没有需求,而是融资到位且完成通电的容量能多快转成确认收入。[CI001, CI002, CI003, CI004, CI014]
| 收入流 | 机制 | 单位 / 合同 | 当前公开数值 / 状态 | 质量判断 | 尽调要求 |
|---|---|---|---|---|---|
| Nscale 基础设施 | 面向超大规模云厂商和前沿模型客户的专用 GPU 集群及 AI 数据中心容量 | 长期 take-or-pay 合同 | BBAE 称这是已签约业务的绝大部分 | 可见度最高的收入流,但仍取决于交付 | 按设施拆分已确认收入、积压订单转化和客户集中度 |
| Nscale 云 | 基于自有 / 合作伙伴容量提供托管 AI 算力、训练、微调和推理服务 | GPU 小时定价 | 公开来源描述了产品,但未量化收入结构 | 可能是更高价值的层,但当前贡献未披露 | 按工作负载类型提供云收入结构、利用率和利润率 |
| 存储 / 数据服务 | 随算力一并销售的辅助存储和数据层服务 | 每 GB 每月 | 已披露公开定价公式,但未披露目录价或实际费率 | 更可能是配套收入,而非独立的核心论点驱动因素 | 披露附加率、实际定价和贡献利润率 |
公开来源披露了变现机制,但未披露细分收入结构;标为状态的单元格总结的是已披露内容,而非实际经济性。
[CI001, CI002, CI014]| 商业层 | 价格 / 单位 / 合同 | 目录价与实际定价 | 公开依据 | 含义 |
|---|---|---|---|---|
| 专用基础设施容量 | 对指定容量作出承诺,加权平均合同期限约 5.7 年 | 实际 $/GPU-hour 和阶梯降价未披露 | BBAE 对申报文件的总结 | 交付后可见度较高,但 headline TCV 不等于当前收入 |
| 云算力 | GPU 小时用量定价 | 未披露公开价目表或实际收益率 | BBAE 对申报文件的总结 | 云业务存在上行空间,但公开来源未显示其是否实质性改善单位经济性 |
| 存储 | 每 GB 每月 | 费率表未公开 | BBAE 对申报文件的总结 | 辅助变现存在,但无法根据公开证据测算其收入贡献 |
在已分配资料池中未找到权威公开价目表;本表记录已披露的收费机制,以及实际定价证据的缺失。
[CI002, CI014]公开披露展示了从签约需求到活跃容量、确认收入,再到当前已披露的负交付经济性的收入桥接。
| 节点 / 连接 | 节点 / 起点 | 终点 | 说明 |
|---|---|---|---|
| 节点 1 | 签约需求 | $103.4bn of active and contracted TCV and 461,000 active or contracted GPUs as of 2026-08-31. | |
| 节点 2 | 活跃容量 | Only about $2.6bn of TCV and roughly 25,000 GPUs were active. | |
| 节点 3 | 客户预付款 | Take-or-pay contracts can generate cash before the related services are recognized as revenue. | |
| 节点 4 | 确认收入 | H1 2026 revenue was $140.6m. | |
| 节点 5 | 服务交付成本 | Cost of revenue excluding depreciation and amortization was $189.6m in H1 2026. | |
| 节点 6 | 当前毛利润门槛 | Public data implies negative gross margin on the disclosed base until more capacity activates or economics improve. | |
| 连接 1 | 签约需求 | 活跃容量 | 融资、建设、通电和验收集群 |
| 连接 2 | 活跃容量 | 客户预付款 | 长期照付不议承诺 |
| 连接 3 | 活跃容量 | 确认收入 | 交付服务随时间确认 |
| 连接 4 | 确认收入 | 服务交付成本 | 电力、基础设施和支持成本 |
| 连接 5 | 服务交付成本 | 当前毛利润门槛 | 当前公开成本基数高于收入 |
这是基于 2026 年公开快照的概念桥接图;不推断未披露的转换率或未来利润率。
[CI002, CI003, CI004, CI005, CI007]4.2 成本结构、营运资金与单位经济性边界
公开数字显示了规模野心,却尚未提供健康单位经济性的公开证据。BBAE 对 Nscale 文件的审阅称,2026 年上半年扣除折旧摊销前的收入成本为 $189.6M,而收入为 $140.6M;所以即使不计折旧,披露基础上的毛利率也为负。上半年净亏损为 $1.02B,调整后 EBITDA 亏损为 $199.2M。这些数字符合公司仍在稳态利用率形成前,提前大量购买 GPU、建设数据中心并承担交付成本的情况。 现金流也容易被误读。Nscale 报告 2026 年上半年经营现金流约为 $1.69B,但同一来源将其归因于长期 take-or-pay 合同的客户预付款,而非经营盈利。同时,投资现金流出约为 $3.29B,6 月现金约为 $1.48B。因此,公开数据支持“客户预付和融资渠道强”的判断,但尚不能支持自我造血的运营模式。没有公开的 CAC、回收期、销售周期或渠道经济性披露,因此 GTM 效率的投资判断仍依赖管理层尽调,而非外部证据。[CI005, CI006, CI007, CI010, CI015]
| 指标 | 公开数值 / 状态 | 置信度 | 重要性 | 尽调要求 |
|---|---|---|---|---|
| 2026 年上半年收入 | $140.6m | 中 | 当前确认收入基数相对于签约需求和估值仍很小 | 确认 2026 年 Q3/Q4 退出年化收入及收入确认政策 |
| 2025 财年收入 | $33.0m | 中 | 显示业务已初步爬坡,但起点很低 | 提供 2025 至 2026 年增长的季度桥接 |
| 2026 年上半年不含 D&A 的收入成本 | $189.6m | 中 | 公开披露暗示折旧前毛利率为负 | 拆分电力、托管、人员和第三方服务交付成本 |
| 2026 年上半年净亏损 | $1.02bn | 中 | 即使已获取需求,扩张期亏损仍然很大 | 将经常性经营亏损与一次性融资和公允价值项目分开 |
| 2026 年上半年调整后 EBITDA 亏损 | $199.2m | 中 | 即使剔除部分非现金项目,经营亏损仍然显著 | 提供管理层通往设施层面盈利的桥接 |
| 经营现金流 | 2026 年上半年约 $1.69bn,主要来自客户预付款 | 中 | OCF 为正本身不能证明经营盈利 | 详细说明预付款时间表、退款触发条件和收入确认时点 |
| 投资现金流 / 资本开支代理指标 | 2026 年上半年流出约 $3.29bn | 中 | 突出签约容量投入运营前的资本密集度 | 拆分 GPU、土地 / 电力、装修和网络资本开支 |
| 现金余额 | 截至 2026-06-30 约 $1.48bn | 中 | 重要但若没有债务契约和已承诺融资细节仍不足 | 提供非受限现金、最低流动性契约,以及签署承诺后的备考流动性 |
| 最大客户集中度 | 2026 年上半年收入的 52%;2025 年收入的 73% | 中 | 少数客户可能同时主导收入质量和重新谈判风险 | 披露头部客户结构、续约安排和终止保护 |
| 积压订单激活 | $103.4bn 已激活及已签约 TCV,对比约 $2.6bn 已激活;461,000 个已激活 / 已签约 GPU,对比约 25,000 个已激活 GPU | 中 | 大多数标题所述需求仍需在完全变现前完成建设、融资和通电 | 按站点和上线日期提供活跃容量与已签约容量的队列视图 |
数值来自Nscale的公开报告;对于作为代理指标的数字,表格明确说明了解读方式,并要求披露缺失的底层数据。
[CI003, CI004, CI005, CI006, CI007, CI010, CI015]4.3 资本充足性与交付依赖
Nscale 显然具备融资能力。CNBC 和 TNW 均报道 2026 年 3 月 Series C 估值 $14.6B,公开报道也确认了 2026 年 2 月的 GPU 抵押延迟提款定期贷款。这些融资很关键,因为 Nscale 的模式天生资本密集:购买 GPU、建设或改造液冷园区、 확보电力,都发生在公司确认已签合同的大部分未来合同价值之前。TechFundingNews 明确将该债务工具描述为用于已签合同和客户 pipeline 相关的资本支出,而非纯粹投机扩张。 剩下的问题是充足性,而非获取渠道。BBAE 报告称,截至文件日期,Nscale 尚未取得履行 Anthropic 协议所需的具有约束力融资承诺;Bricks & Bytes 则报道,Loughton 园区电网交付延迟严重,迫使 Nscale 转向现场供电选项。两项信息共同说明,收入转化仍依赖额外融资、电力交付和项目执行。历史融资时间线属于公司概览;本节的财务结论是,新增股权和债务减轻了近期资金压力,却没有消除与交付挂钩的融资和基础设施风险。[CI008, CI009, CI011, CI012, CI013]
4.4 财务结论与尽调阻塞点
公开记录支持一个有条件的结论。如果长期 take-or-pay 合同按预期经济性转成活跃容量,且客户预付款部分抵消建设带来的营运资金负担,收入质量可能很强。但当前披露基础仍显示:相对已预订需求,实际收入不高;可由公开资料推断的毛利率为负;客户集中度高;资本支出需求大;融资和电力带来重大执行依赖。 多个投资判断阻塞点仍未解决。公开来源未披露按产品划分的实际价格、基础设施与云收入组合、产品级利润率、CAC 或回收期、客户特定合同保护、债务契约、抵押触发条件,或履行 Anthropic 承诺所需的确切融资方案。在尽调补齐这些数据前,合理结论是:Nscale 已证明需求获取和融资能力,但尚无足够公开证据自信地判断利润率路径、runway 持久性或估值支撑。[CI003, CI005, CI007, CI010, CI013, CI015]
4.5 展示项
05产品与技术
5.1 产品方案与产品线
从客户工作流看,Nscale 推销的不是单一软件功能,而是把 AI 系统从容量规划推到实际训练、微调、推理和持续运营所需的基础设施。公司和投资人材料一贯将方案描述为垂直整合,覆盖土地与电力决策、数据中心建设、GPU 供应、网络、数据服务和编排。独立市场评论也描述了两条商业界面:Nscale Infrastructure 提供专用集群和 AI 数据中心容量,Nscale Cloud 提供托管算力和推理服务。因此,尽调必须分别判断物理容量、软件控制平面和服务交付的成熟度,不能假设这是一个经济性均质的成品云产品。[CE001, CE002, CE004, CE014]
| 模块/资产/产品线 | 主要用户 | 状态/成熟度 | 差异化 | 尽调缺口 |
|---|---|---|---|---|
| Nscale基础设施 | 超大规模云厂商和前沿AI实验室 | 已商业销售且部分上线 | 专用GPU集群及AI数据中心容量,而非纯按需云 | 需要按站点展示已售、已验收和已产生收入的容量队列 |
| Nscale云 | 企业AI团队和模型运营方 | 公开描述为商业产品,但软件深度仍主要由外部叙述,文件记录较少 | 在原始基础设施之上增加托管训练、微调和推理服务 | 需要公开产品文档或SLA证据,以核实功能深度和支持范围 |
| Glomfjord/挪威已部署基地 | 早期生产客户和参考部署 | 已建成的足迹和旗舰站点均有公开提及 | 使Nscale拥有实际运营的基础设施,而非仅停留在纸面上的园区管线 | 需要按挪威站点提供当前容量利用率和客户构成 |
| Loughton AI园区 | Microsoft Azure和英国主权AI工作负载 | 已宣布并签约,但尚未通电 | 以50MW、可扩展至90MW的规模和计划中的23,040块GB300 GPU支撑英国主权定位 | 需要供电时间表和验收里程碑 |
| Stargate UK/英国主权平台 | 与OpenAI相关的主权及公共利益工作负载 | 已宣布的平台,具有分阶段承购意向 | 将业务延伸至主权AI定位,而非通用新云容量 | 需要具有约束力的商业条款和按站点列出的启用时间表 |
由于公开记录同时描述产品线和指定资产,行项目混合了二者;成熟度仅反映截至2026-10-18保留的公开证据。
[CE003, CE004, CE006, CE007, CE014]| 用户任务 | 当前工作流问题 | Nscale解决方案 | 可量化收益 | 局限 |
|---|---|---|---|---|
| 建立专用训练容量 | 企业和实验室需要有保障的GPU供应以及就绪的数据中心 | Nscale基础设施销售专用集群和AI数据中心容量 | 公开报道显示有25,000块活跃GPU以及规模大得多的已签约基地 | 公开证据未显示按客户划分的验收时间或实现的利用率 |
| 运行托管训练、微调和推理 | 买方希望使用托管计算,而不是自行构建每一层 | Nscale云被描述为位于基础设施之上的托管层 | 公开叙述明确包括训练、微调和推理 | 指定证据中没有公开定价页面或SLA细节 |
| 建设英国主权AI容量 | 政府和受监管买方希望获得英国托管的、由重要合作伙伴参与的计算资源 | Stargate UK与Loughton园区结合OpenAI、Microsoft和NVIDIA,提供主权计算 | 公开方案包括明确的GPU承诺和英国增长区定位 | 主权工作负载的可用性仍取决于站点通电和分阶段建设 |
| 在全球AI区域之间高效移动数据 | 大型模型运营方需要存储和数据路径,以保持GPU高负载 | VAST合作增加数据层,旨在优化效率和跨区域访问 | 合作伙伴证据声称具有统一数据层和由可观测性支持的运营 | 保留证据中没有对持续性能的独立验证 |
仅在来源提供具体运营信号时陈述收益;缺失的定价、SLA和验收数据仍是开放的尽调要求,而非假定的产品事实。
[CE002, CE004, CE005, CE007, CE008, CE009]5.2 架构与部署
公开架构是分层的。底层是已供电或规划中的园区,包括挪威容量以及已宣布的 Loughton 和 Stargate UK 项目;其上是 GPU 集群和高性能网络;再上是数据与编排层;最上层是托管训练、微调和推理服务。公开报道对规模成熟度的信号强于对深层技术内核的信号:截至 2026 年 8 月底,第三方报道称 Nscale 有 25,000 个活跃 GPU,对比活跃或已签约 GPU 的 461,000 个;活跃站点 5 个,对比已签约站点 12 个。这一差距说明运营栈真实存在,但 headline 平台相当一部分仍依赖将已签约资产转成活跃服务。VAST 合作和计划中的 Anyscale 收购表明,公司有意加厚软件和数据层,而不只是出租原始 GPU 容量。[CE005, CE006, CE007, CE008, CE009, CE012]
| 层/流程/组件 | 作用 | 依赖 | 风险 |
|---|---|---|---|
| 电力和土地控制 | 确保可通电站点和长期园区 | 电网接入、拥有或控制的站点以及表后发电能力 | 如果电力晚于已签约的计算需求到达,园区交付将延迟 |
| 数据中心园区 | 容纳液冷计算和实体运营 | 建设、通电和当地审批 | 活跃足迹远小于已签约管线 |
| GPU和网络织构 | 提供专用训练和推理容量 | NVIDIA供应、合作伙伴经济性和站点调试 | 在上线GPU获验收前,已售容量无法帮助客户 |
| 数据层 | 保持GPU获得数据并让各区域都能访问数据 | VAST AI操作系统和参考架构集成 | 公开证明来自合作伙伴,而非独立运营方的基准测试 |
| 编排和可观测性 | 调度工作负载并监控多层基础设施 | 内部软件以及操作系统/流程成熟度 | 公开文档较少,因此运营可靠性仍难以核实 |
| 云和推理服务 | 在裸基础设施之上获取更高价值的托管使用 | 产品软件广度及客户支持运营 | 公开证据显示了雄心,但未显示与成熟超大规模云工具的功能级对等 |
| 安全与治理控制平面 | 管理网络、实体和供应链暴露 | 专门的高管负责制和正式运营控制 | 保留证据中没有SOC、ISO、可用性报告或公开信任中心披露 |
架构依据保留的官方、合作伙伴和独立报道重建;它展示运营栈的形态,而非经核实的内部实施细节。
[CE001, CE005, CE008, CE009, CE010, CE011, CE012, CE013]公开证据将 Nscale 描述为分层运营栈:电力和园区启用支撑 GPU 集群、数据 / 编排层以及托管式 AI 服务。
- 电力与能源
- 容量建设始于受控站点和可用电力。
- 模块
- 1.37 GW 活跃和签约电力容量
- 自有或受控站点合计 10 GW 的可见储备
- Loughton 现场发电计划
- 输出
- 园区就绪
- 交付时间表
- 数据中心园区
- 物理站点是主权云和企业 AI 服务的底座。
- 条目
- Glomfjord 挪威园区
- Loughton 园区
- Stargate 英国站点
- 输出
- 活跃和签约站点
- GPU 与网络底座
- 大部分专用 AI 容量尚未启用,就已提前售出。
- 模块
- 25,000 个活跃 GPU
- 461,000 个活跃或签约 GPU
- Loughton 计划部署 23,040 个 GB300 GPU
- 输出
- 训练和推理容量
- 数据层
- 合作数据服务旨在维持各地区的高利用率。
- 模块
- VAST AI Operating System
- 统一数据层
- 输出
- 跨区域数据访问
- GPU 效率
- 编排与可观测性
- 公开材料称,软件层能够控制硬件之上部分。
- 模块
- 编排软件
- 可观测性工具
- 输出
- 工作负载调度
- 监控
- 托管式云服务
- 顶层是面向客户的服务界面。
- 条目
- 托管式训练
- 微调
- 推理
- 输出
- 企业 AI 工作负载
- 开发者访问
栈层次综合自保留的公开描述;排序反映运营依赖关系,而不是声称存在某张参考架构图。
[CE001, CE003, CE004, CE005, CE006, CE007, CE008, CE009, CE011, CE015]5.3 信任、控制与依赖
主要产品风险不是抽象的模型性能主张,而是交付依赖。公开证据显示,Nscale 已配备电力、基础设施、产品、法律和安全领导岗位,包括负责网络、物理及供应链安全的 CISO;这比没有运营责任归属的普通初创企业页面更好。但保留的证据没有披露信任中心、正常运行时间承诺、认证或事件报告,因此无法核实这些控制如何在生产环境中运转。更大的即时依赖是物理交付:Loughton 对外被描述为已融资、获许可并由 Microsoft 锚定,但外部报道称电网接入延迟到 Nscale 正在寻求现场发电。结合少数锚定客户带来的需求集中,这个产品在战略上有差异化,但园区启用和客户爬坡阶段仍高度依赖执行。[CE003, CE010, CE011, CE012, CE013]
5.4 展示项
06客户
6.1 锚定需求真实存在,但知名客户集合很窄
Nscale 的定位很广:公司称服务 AI 原生公司、企业和政府,并将主权属性作为英国和欧洲客户的重要购买标准。但审阅来源中的知名客户证据实际窄得多。独立报道反复把 Nscale 最显眼的园区与 Microsoft 需求联系起来,而 OpenAI 更多出现在项目级基础设施叙事中,并非完整描述的经常性客户账户。这在目标市场广度和有证据支持的买方集中度之间形成落差。公司显然在向大型买方销售高接触、多园区基础设施产品,但公开记录还没有显示企业、政府和研究机构中广泛的知名生产客户。[CU001, CU002, CU008, CU009, CU012]
| 分层 | 买方/用户/付款方 | 用例 | 公开证明 | 收入/战略价值 | 缺口 |
|---|---|---|---|---|---|
| 超大规模锚定租户 | Microsoft作为买方/承购锚点;Nscale交付园区容量和托管AI基础设施 | 为大规模AI工作负载预留多园区GPU和电力容量 | 在Loughton、得克萨斯州和Monarch叙述中均有提及 | 可见度最高的需求锚点,可能也是融资催化剂 | 合同价值、利用率承诺和终止权未披露 |
| 前沿模型生态项目 | 参与与OpenAI关联的项目;确切付款方结构未披露 | 挪威的主权AI和高密度训练容量 | Stargate Norway报道将OpenAI与其联系起来 | 验证其与旗舰AI工作负载的相关性 | 直接经常性客户经济性与项目品牌效应尚不清楚 |
| 企业AI建设者 | 运行训练、微调和推理工作负载的企业 | 提供计算、数据、网络和编排的生产AI基础设施 | Nscale和VAST将其描述为面向企业的多区域基础设施 | 将总潜在市场拓展到单一超大规模账户之外 | 保留来源中没有具名企业客户名单或案例研究结果 |
| 主权/受监管买方 | 政府、研究机构和重视主权的组织 | 欧洲/英国境内的私有和主权AI容量 | Nscale英国公告和关于页面强调主权 | 支持差异化的区域定位 | 此处未引用具名政府部署或采购记录 |
混合了披露的目标分层和具名客户证明;经济数据为空,表明公开来源对需求叙述的证明比对实际客户构成的证明更清楚。
[CU001, CU002, CU008, CU009, CU012]公开证据显示,Nscale 通过主权定位、锚定租户承诺和多园区扩张赢得需求,而不是依靠广泛的自助式采用。
各阶段是对多个来源所披露客户动作的分析综合,不是来源作者绘制且带有报告转换率的漏斗。
[CU001, CU002, CU005, CU006, CU008, CU009, CU012, CU013]6.2 采用证据主要是园区承诺和容量预留
最强的公开采用信号是具体站点承诺,而不是广泛的使用群组。Nscale 披露英国 Loughton 设施初始规模 50MW、最高 90MW;独立报道将后续园区与德州的 Microsoft 支持需求以及 Monarch 的 Microsoft 意向书联系起来。Datacenter Frontier 还将该组合与葡萄牙和德州分阶段 GPU 及兆瓦数字联系起来,Data Center Knowledge 则将 Nscale 与 OpenAI 的挪威项目联系起来。这些事实证明超大型 AI 基础设施买方对需求有可信度,但大多仍是已预留或规划容量的前瞻性指标,不能说明园区运营后的实际利用率、工作负载留存或扩张经济性。[CU003, CU004, CU005, CU006, CU007, CU008, CU013]
| 指标 | 数值 | 日期 | 来源 | 置信度 | 含义 | 缺失分母 |
|---|---|---|---|---|---|---|
| 英国站点容量 | 初始50MW,可扩展至90MW;目标于2026年第四季度上线 | 2025-01-13 | Nscale英国新闻稿,经Bricks & Bytes佐证 | 高 | 显示了一个披露实体规模、与Microsoft关联的具体园区 | 未披露预留利用率、定价或爬坡曲线 |
| Loughton GPU容量 | 最多45,000块NVIDIA GB200 GPU | 2025-01-13 | Nscale英国新闻稿,经Bricks & Bytes佐证 | 高 | 表明面向客户的大规模交付雄心 | 未披露实际签约或部署的GPU数量 |
| 得克萨斯州Microsoft园区 | 初始约104,000块GB300 GPU、约240MW,可扩展至1.2GW | 2026-04-03 | Datacenter Frontier | 中 | 暗示围绕锚定需求的土地扩张模式 | 未披露投产日期、利用率或收入 |
| Monarch Microsoft意向书 | 最多1.35GW的AI计算容量,交付自2027年末开始 | 2026-04-03 | Datacenter Frontier | 中 | 为美国能源园区战略提供强劲的未来需求信号 | 意向书条款、转化风险和分阶段经济性未披露 |
| 葡萄牙部署 | 计划在Sines的Start Campus部署12,600块GB300 GPU | 2026-04-03 | Datacenter Frontier | 中 | 显示超越单一园区的多区域客户足迹 | 除更广泛的Microsoft关系外,没有具名最终客户 |
| 挪威旗舰项目 | 计划中的Stargate Norway项目拥有100,000块GPU,初始计划230MW | 2026-03-09 | Data Center Knowledge | 中 | 展示其与前沿规模AI需求的项目相关性 | Nscale的直接收入份额、利用率和续约路径未披露 |
所有行项目都是披露的容量或项目动向,而非已报告的活跃客户数量;缺失分母突出显示采用证明止步于收入质量证明之前。
[CU003, CU004, CU005, CU006, CU007, CU008, CU013]| 客户 | 分层 | 部署/用例 | 生产还是试点 | 结果 | 局限 |
|---|---|---|---|---|---|
| Microsoft | 超大规模锚定租户 | Loughton英国园区,初始容量50MW,可扩展至90MW | 已锚定的未来生产站点,未证明存在实时使用 | 为英国主权AI定位提供知名租户证明 | 条款、实时利用率和收入贡献未披露 |
| Microsoft | 多区域超大规模账户 | 得克萨斯州园区(约104,000块GB300 GPU)加上最多1.35GW的Monarch意向书 | 容量预留/扩张路径,而非公开工作负载指标 | 暗示Nscale能够从同一买方获得重复园区需求 | 强化对单一具名客户的集中 |
| OpenAI | 前沿模型生态参与者 | 挪威的Stargate品牌AI中心 | 项目级基础设施关联 | 支持其在领先AI工作负载中的可信度 | 公开来源未显示OpenAI是否是Nscale直接的经常性客户 |
仅为具名公开证明的部分样本;表格记录披露的部署和关系,而非完整客户名单或经常性生产使用的证明。
[CU003, CU005, CU006, CU008, CU012, CU013]6.3 持久性、多元化和时间仍是客户侧关键未知数
核心尽调问题不是 Nscale 能否描述需求,而是需求是否多元、持久并能按时变现。一份审阅来源明确称 pitch 材料遗漏了具体客户牵引力或案例,保留的公开来源也都未披露客户数量、NRR、GRR、流失、合同条款或续约历史。因此,Microsoft 权重很高的足迹是集中度风险,而不是完整的 go-to-market 证明。不利之处在于执行时间:Bricks & Bytes 报道,已融资、获许可并由 Microsoft 锚定的 Loughton 项目仍缺少及时电力交付。若园区通电延迟,Nscale 可能拥有 marquee 客户需求,却没有投资者通常希望从持久性案例中看到的近期收入爬坡或可供引用的生产结果。[CU010, CU011, CU012, CU013, CU014]
6.4 展示项
07风险
7.1 风险排序
最重要的风险不是创造需求,而是在基础设施约束下完成交付。多份独立报道称,Loughton 需要 90MW 接入,当前电网时间线已推迟到 2030 年代初至中期,尽管 Microsoft 已被点名为锚定租户,Nscale 也在探索现场供电替代方案。这使旗舰英国项目成为检验公司能否把资本和客户承诺转成通电容量的真实案例。第二类风险是集中度:公开报道把重大部署和 backlog 主张与 Microsoft 及少数战略交易对手联系起来。第三类是融资和执行负担,大额股权融资、GPU 抵押债务和激进扩张目标会放大任何进度延误的成本。[CR001, CR002, CR003, CR004, CR005, CR007, CR013]
尽管治理层面的缓释措施可见,剩余风险严重度仍主要由供电交付、交易对手集中度和资本结构风险主导。
定性分档反映引用的公开证据,而不是经审计的损失历史、客户层面的流失数据或已披露的契约阈值。
[CR001, CR004, CR005, CR010, CR012, CR013, CR014]7.2 法律、监管与治理风险
这组公开证据显示治理建设快于独立监管验证。Nscale 领导层页面列出 Chief Legal Officer、General Counsel 和 CISO;2026 年 3 月报道称公司还让 Sheryl Sandberg、Susan Decker 和 Nick Clegg 加入董事会,并将 Aker-Nscale 合资企业并入母公司以简化治理。对于在多个司法辖区销售主权 AI 基础设施的企业,这些是切实的缓释信号。但证据集不包含监管申报、诉讼案卷、隐私证明或安全认证,无法独立确认公司的法律、隐私、出口管制或事件处理状况。因此,公开层面对法律和监管风险的消除仍不完整。[CR011, CR012, CR014]
| 风险/问题 | 司法管辖区 | 公开状态 | 可能性 | 严重性 | 缓解措施 | 剩余暴露 | 尽调路径 |
|---|---|---|---|---|---|---|---|
| Loughton通电及相关电力审批 | 英国 | 规划许可和租户承诺已公开,但据报道电网可用时间已推迟至2030年代初/中期。 | 高 | 严重 | Nscale正在评估现场发电替代方案,并持续加强治理建设。 | 高 | 核实电网接入里程碑、任何表后许可路径,以及通电再次延迟时客户可获得的救济。 |
| 主权AI、隐私和公共部门合规准备度 | 英国/欧盟/美国 | 公开治理人员配置清晰,但指定资料池不包含主要合规或认证记录。 | 中 | 高 | 首席法务官、总法律顾问和CISO职位已到位;董事会于2026年3月扩大。 | 中-高 | 要求提供每个运营地理区域的隐私、出口管制、制裁、公共部门和安全认证资料包。 |
| 挪威项目治理及关联方监督 | 挪威 | TNW和Tech Funding News称,Aker-Nscale合资企业已并入Nscale以简化治理,同时Aker仍留在董事会。 | 中 | 高 | 合资企业简化和增加独立董事应减少协调摩擦。 | 中 | 审查挪威项目和融资中的股东权利、董事会委员会结构及关联方控制。 |
严重性排序仅反映公开来源证据;在缺少监管机构、法院或认证记录的情况下,未经核实的法律/合规事项仍是尽调问题,而非已清除的风险。
[CR001, CR011, CR012, CR014]7.3 交付与运营风险
交付风险偏高,因为垂直整合模式把电力、房地产、GPU 供应、云运营和客户交付串成一条链。独立报道称 Loughton 已融资、获许可且有客户支持,但仍无法按原计划通电。合作伙伴材料还称 Nscale 正尝试在全球扩展到 300,000 个 NVIDIA Grace Blackwell GPU,并在方案中依赖外部数据层合作伙伴,因此执行风险不局限于单一站点。Sacra 明确指出 GPU 配额风险和北欧电力依赖,这与电网延迟叙事方向一致:Nscale 的瓶颈是容量交付、硬件可得性和韧性证明,而不是市场缺乏需求。[CR001, CR003, CR009, CR010, CR013, CR014]
| 失效模式 | 可能性 | 严重性 | 缓解成熟度 | 剩余暴露 | 未解决缺口 |
|---|---|---|---|---|---|
| Loughton供电交付延迟 | 高 | 严重 | 早期 | 高 | 现场发电仅是已探索的应对方案;指定证据未显示已获批准的替代时间表。 |
| GPU分配或平台转型延误 | 中-高 | 高 | 早期 | 高 | 公开材料显示了激进的Blackwell规模目标,但没有具有约束力的分配细节或下行缓冲。 |
| 北欧电力和站点集中 | 中 | 高 | 早期 | 中-高 | 公开证据未量化水力供电站点的冗余、停电历史或保险覆盖。 |
| 安全、正常运行时间和事故证明缺口 | 中 | 高 | 发展中 | 中-高 | 资料池显示了安全领导层,但没有公开SLA历史、审计鉴证或事故披露资料集。 |
本登记表区分已披露的缓解意图与经核实的运营证明;当公开证据未显示经过测试的冗余、审计历史或针对风险的交付时,“缓解成熟度”保持较低。
[CR001, CR003, CR009, CR010, CR011, CR013, CR014]7.4 交易对手集中度与资本结构风险
公开报道显示,Nscale 正在以集中的商业敞口和显著杠杆建设。TNW 和 Tech Funding News 均报道称,$2B Series C 之前有 $1.4B GPU 抵押延迟提款定期贷款;TNW 和 Sacra 则将德州、挪威和葡萄牙的重大部署与 Microsoft 联系起来。低置信度研究进一步认为,backlog 和 pipeline 高度偏向 Microsoft、NVIDIA、ByteDance 和 OpenAI。即便精确 backlog 数字仍需尽调,方向性风险已经清楚:一两个超大规模交易对手重新定价、延迟或终止,可能同时打击利用率、债务提款假设和股权价值。因此,否决标准应聚焦合同持久性、契约余量和通电里程碑,而不是 headline 融资规模。[CR004, CR005, CR006, CR007, CR008, CR012]
| 依赖项 | 交易对手 | 作用 | 集中度 | 失效情景 | 严重性 | 缓解措施 | 剩余暴露 |
|---|---|---|---|---|---|---|---|
| 锚定客户需求 | Microsoft | Loughton的锚定租户,以及据报道在得克萨斯州、挪威和葡萄牙的部署合作伙伴 | 高 | 推迟采购、重新定价或终止将削弱多个站点的利用率和积压订单可信度。 | 严重 | 董事会/治理建设和多区域足迹可能扩大未来客户选择。 | 高 |
| GPU和核心平台供应 | NVIDIA和VAST Data | 战略投资者/GPU供应商及关键数据层合作伙伴 | 高 | 分配短缺或平台集成问题会拖慢容量交付和产品性能。 | 高 | 垂直整合和深度合作改善协调,但无法消除对供应商的依赖。 | 高 |
| 资本、项目和治理交易对手 | Aker、债务融资方和其他大型投资者 | 董事会影响力、挪威项目赞助及资本支出融资 | 中-高 | 契约压力、提款延迟或治理摩擦可能在资产产生现金前拖慢建设。 | 高 | 合资企业简化和近期大额股权融资提供了跑道,但没有提供契约透明度。 | 中-高 |
集中度标签具有方向性,因为指定证据集中未公开合同经济性、终止权和按交易对手划分的收入份额。
[CR004, CR005, CR006, CR007, CR008, CR009, CR012, CR013]7.5 展示项
08估值
8.1 融资势头真实,但价格发现仍薄弱
Nscale 的估值背景比投资判断更容易描述。独立报道将 2026 年 3 月 Series C 定为 $2B、估值 $14.6B;Nscale 自己在 2026 年 9 月宣布了 $3.36B 的 pre-IPO 可转换票据结构,其中交割时已融资 $2.36B,另有 $1B 由 NVIDIA 承诺。这些融资证明公司有异常强的资本获取能力和真实战略支持,但没有披露收入、利润率或现金转化数据。证据池中唯一的升值标记来自一个市场数据页面:估值 $21.4B,状态为 public/active-IPO;它可作为情绪信号,但交叉验证太弱,不足以支撑高于最近透明私募估值的确定性入场价。[CV003, CV004, CV005, CV007, CV013, CV014]
| 指标 | 当前判断 | 决策含义 |
|---|---|---|
| 建议 | 跟踪 | 继续尽调,但不要仅凭公开证据就押注新一轮估值上调。 |
| 信心 | 中 | 融资和合作伙伴信号确实存在,但经营披露仍然有限。 |
| 风险评级 | 高 | 交付时间、客户集中度和融资条款的不确定性,可能迅速改变估值。 |
| 估值立场 | 偏高 | 2026 年 3 月 $14.6B 的估值有公开依据;高于该水平的支撑仍是间接的。 |
| 入场纪律 | 等待披露或更好价格 | 在支付更高价格前,必须看到经审计的经营指标、可转债条款和电力交付证明。 |
汇总指定证据池中的融资、执行风险和披露事实;“偏高”反映基本面支撑薄弱,而不是模型计算出的公允价值。
[CV003, CV004, CV007, CV008, CV013, CV014]建议维持跟踪,因为需求和资本获取的证据强于已披露经济性的证据。
| 节点 / 连接 | 节点 / 起点 | 终点 | 说明 |
|---|---|---|---|
| 节点 1 | 战略需求与支持者证据 | C 轮和 2026 年可转债融资显示,投资者仍希望获得敞口。 | |
| 节点 2 | 全栈平台叙事 | 公司和合作伙伴来源都描述了其对电力、数据中心、算力和软件层的控制。 | |
| 节点 3 | 规模信号 | 项目储备和 TCV 显示出野心,但不代表已实现的收入质量。 | |
| 节点 4 | 执行与集中度风险 | Loughton 通电和对标杆交易对手的依赖仍会影响估值。 | |
| 节点 5 | 价格支撑测试 | 指定证据池更能支撑其相关性和融资获取能力,而不是精确的公允价值上调。 | |
| 节点 6 | 建议 | 在经济性和融资条款披露前,以偏高立场继续跟踪。 | |
| 连接 1 | 战略需求与支持者证据 | 全栈平台叙事 | 融资证明了兴趣,但尚未证明经济性。 |
| 连接 2 | 全栈平台叙事 | 规模信号 | 平台广度支撑扩张主张。 |
| 连接 3 | 规模信号 | 执行与集中度风险 | 更大的项目放大交付和客户结构风险。 |
| 连接 4 | 执行与集中度风险 | 价格支撑测试 | 风险调整后的投资判断需要更多披露。 |
| 连接 5 | 价格支撑测试 | 建议 | 没有经营证明,结论仍是跟踪。 |
8.2 特许经营故事强于经济性故事
可投资的部分很直白:Nscale 已搭建垂直整合的 AI 基础设施叙事,从战略和财务支持者处筹得大额资金,并引入主权 AI 和企业 AI 中重要的交易对手。公司及合作伙伴材料大体一致地描述了对电力、数据中心、算力、存储、网络和编排的控制;VAST 合作则提供第三方证据,说明平台面向跨洲企业工作负载建设。反论点是,故事的证据强于经济性。TCV 不是收入,旗舰交易对手不等于多元化需求;一旦通电、利用率或客户集中度停滞,估值叙事会迅速减弱。[CV001, CV002, CV006, CV008, CV010, CV013, CV015]
| 视角 | 当前判断 | 哪些事实会改变判断 |
|---|---|---|
| 多头论点 | Nscale 已搭建起垂直整合的 AI 基础设施栈,拥有战略投资者、标杆客户和全球扩张计划。 | 披露确认收入、利用率和客户多元化情况,证明这套基础设施不只是靠标杆项目变现。 |
| 反论点 | 公开证据最能支撑融资和叙事定位,却不足以支撑单位经济性或现金创造质量。 | 提供经审计的收入、毛利率、现金消耗和资本开支披露,让投资者检验经营杠杆。 |
| 上调触发条件 | 经确认的经营指标加上清晰的可转债条款,会让当前的规模故事更值得纳入投资判断。 | 发布官方文件或投资者材料,说明收入质量、债务 / 抵押品条款和现实可行的通电时间表。 |
| 下调触发条件 | 进一步延迟、客户集中或结构化融资压力,可能让当前估值成为下一轮下调的前兆。 | 任何电力交付延期、利用率疲弱或偏向投资者的转换条款证据,都会让结论转向放弃。 |
本表将保留的来源记录转化为投资判断框架,而非评分模型;没有公开的 EV 或收入倍数可由指定证据推导。
[CV001, CV002, CV006, CV008, CV013, CV014, CV015]8.3 正确结论是跟踪,不是买入或放弃
我的基准情景是跟踪,而不是买入或放弃。直接放弃会忽略 Nscale 多次吸引大额投资人和战略交易对手;买入则需要证据池中的公开记录尚未提供的运营披露。因此,情景分析应使用已观察到的融资估值作为参考点,而不是假装知道干净的 EV / 收入倍数。悲观情景集中于 Loughton 通电延迟、marquee 项目和交易对手集中,以及可转换票据带来的优先权或稀释悬置。乐观情景要求把 TCV 和旗舰项目转成多元且已披露的收入,同时保持项目交付和融资条款足够清晰,避免新股权为他人的结构化上行买单。[CV003, CV004, CV006, CV008, CV009, CV013, CV014, CV015]
| 情景 | 假设 | 估值 / 回报逻辑 | 主要风险 | 概率信号 |
|---|---|---|---|---|
| 多头 | 2026 年的融资组合转化为多元化客户采用,Loughton 按计划解决供电问题,TCV 转化为已披露的高质量收入。 | 估值有理由高于 2026 年 3 月的私募估值,最终也可能验证其后期公开市场定位。 | 执行必须追上叙事;集中度和融资条款也必须保持可控。 | 需要官方收入披露、顺利通电和清晰的转换机制。 |
| 基准 | Nscale 仍具战略相关性且资金充足,但投资者仍看不到经审计的收入、利润率和票据条款披露。 | 2026 年 3 月的 $14.6B 估值仍是最后一个透明锚点;之后的价值信号应因不确定性打折。 | 交付时间、客户结构和可转债持有人经济利益仍未解决。 | 最符合当前指定证据池。 |
| 空头 | 电网延迟、标杆交易对手集中,或偏向投资者的转换条款,会拖慢商业化或迫使估值重定价。 | 在基本面得到验证前,市场先重估执行和融资风险,估值因此跌破最后一个透明的私募锚点。 | 通电延迟、多元化不足、债务 / 优先权结构不透明。 | 不利项目进展或显示新股权处于次位的披露会触发该情景。 |
情景锚定已观察到的融资估值和已披露风险,而不是模型 DCF 或公开市场倍数推演,因为指定证据没有披露收入、利润率或现金流。
[CV003, CV004, CV005, CV008, CV013, CV014, CV015]8.4 关键门槛是披露,不是更多叙事
下一步尽调不应是另一项品牌建设公告,而应是运营与融资披露。在接受甚至追平 2026 年 3 月估值的明显上调之前,投资者应索取收入或 ARR 定义、毛利率和利用率披露、票据转换机制、债务和抵押条款、按合同收入与确认收入划分的客户集中度,以及 Loughton 通电的最新时间表。由于董事会和高管队伍越来越像上市公司,退出准备度在方向上改善;但当前证据池仍缺少 IPO 状态故事的官方申报证据。电力交付延迟、无法超越 marquee 交易对手实现多元化,或融资条款把过多价值从新股权转给 pre-IPO 可转换票据持有人,都可能破坏论点。[CV005, CV007, CV008, CV011, CV012, CV013, CV015]
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,任何投资决策前都应直接向管理层和一手文件核实。
证据索引
| 编号 | 陈述 | 可信度 | 来源 |
|---|---|---|---|
| CO001 | Nscale describes itself as a full-stack AI cloud platform for AI teams. | 中 | SO002 |
| CO002 | Nscale says its vertically integrated offering spans cloud software, GPU compute, networking, data services, data centers, and dedicated power infrastructure. | 高 | SO001, SO002 |
| CO003 | Independent sources describe Nscale as London-based or UK-based. | 中 | SO005, SO006, SO008 |
| CO004 | Third-party profiles and reporting place Nscale's founding in 2024. | 中 | SO005, SO006 |
| CO005 | Nscale's official materials identify Josh Payne as CEO and founder. | 高 | SO001, SO003 |
| CO006 | Nscale's current leadership page publicly lists named executives across finance, operations, AI infrastructure, data centers, legal, revenue, power, accounting, product, people, engineering, and security. | 中 | SO003 |
| CO007 | On 2026-03-09, Nscale announced that Sheryl Sandberg, Susan Decker, and Nick Clegg would join its board. | 高 | SO001, SO005, SO008 |
| CO008 | Nscale said its pre-existing board included Josh Payne, Rael Nurick, Jacob Leschly, and Oyvind Eriksen. | 高 | SO001, SO008 |
| CO009 | On 2026-03-09, Nscale announced a $2.0 billion Series C at a $14.6 billion valuation led by Aker ASA and 8090 Industries. | 高 | SO001, SO005, SO008 |
| CO010 | Official and news reporting name Astra Capital Management, Citadel, Dell, Jane Street, Lenovo, Linden Advisors, Nokia, NVIDIA, and Point72 among the Series C participants. | 高 | SO001, SO005, SO008 |
| CO011 | Profile sources report a $30 million seed in December 2023, a $155 million Series A in December 2024, a $1.1 billion Series B in September 2025, and a $433 million pre-Series-C SAFE in October 2025, taking disclosed equity funding above $3.7 billion. | 中 | SO004, SO006 |
| CO012 | News reporting says the Series C followed a $1.4 billion GPU-backed delayed-draw term loan announced in February 2026. | 中 | SO005, SO008 |
| CO013 | Independent profile sources place Nscale's public footprint across the UK, US, Norway, Portugal, and Iceland. | 中 | SO004, SO006 |
| CO014 | Nscale announced that the Aker Nscale joint venture launched in July 2025 would be rolled fully into Nscale while Aker remained a leading shareholder. | 高 | SO001, SO005, SO008 |
| CO015 | Bricks & Bytes reported that Nscale's consented GBP 2 billion Loughton project had Microsoft as anchor tenant and an agreed 90MW grid connection, but not power delivered in time for opening. | 中 | SO007 |
| CO016 | None of the reviewed public sources disclosed audited revenue, revenue run-rate, public customer count, or headcount for Nscale. | 低 | SO001, SO002, SO003, SO004, SO005, SO006, SO007, SO008 |
| CM001 | Nscale positions itself as a vertically integrated AI infrastructure provider spanning energy, data centers, GPU compute, and software. | 中 | SM003, SM006, SM008 |
| CM002 | Public materials target AI natives, enterprises, and governments, implying a multi-segment buyer set rather than a single developer niche. | 中 | SM003, SM006 |
| CM003 | Sovereignty is a named purchase driver, with Nscale framing UK and European local compute as secure private infrastructure for enterprises, research institutions, and government ambitions. | 中 | SM003, SM005 |
| CM004 | Microsoft's contracted deployment of approximately 200,000 GB300 GPUs across four sites shows that Nscale's clearest disclosed near-term demand is hyperscaler overflow capacity rather than diffuse self-serve usage. | 中 | SM001 |
| CM005 | Nscale's first UK site at Loughton is designed for 50MW of AI and HPC capacity, scalable to 90MW, with up to 45,000 GB200 GPUs and a planned Q4 2026 go-live. | 中 | SM003, SM007 |
| CM006 | Nscale says recent funding will accelerate a 1.3GW greenfield data-center pipeline, including 120MW planned for development in 2025. | 中 | SM002, SM003 |
| CM007 | Next Move Strategy Consulting says the global GPU cloud market is USD 132.0B in 2026 and forecast to reach USD 779.0B by 2035, a 21.8% CAGR. | 中 | SM005 |
| CM008 | Ainvest cites external research projecting the global AI infrastructure market from USD 158.3B in 2025 to USD 418.8B by 2030, a 21.5% CAGR. | 中 | SM004 |
| CM009 | Acquinox says European data-center electricity demand was about 96 TWh in 2024 and is projected to reach 168 TWh by 2030, while grid interconnection queues can run up to seven years. | 中 | SM002 |
| CM010 | Bricks & Bytes reports that Loughton is funded, consented, and Microsoft-anchored but delayed by grid power timing, with more than 100 UK projects considering on-site generation and more than a quarter of planned UK capacity slipping in 2025. | 中 | SM007 |
| CM011 | Nscale's Anyscale acquisition positions the company to bundle orchestration software with raw compute, reflecting broader neocloud consolidation around full-stack AI platforms. | 中 | SM005 |
| CM012 | Next Move Strategy Consulting says the managed AI platform segment should grow at a 25.9% CAGR from 2026 to 2035, faster than the broader GPU cloud market. | 中 | SM005 |
| CM013 | Next Move Strategy Consulting says sovereign cloud deployments are the fastest-growing deployment model at a 29.4% CAGR through 2035 and that 52.6% of enterprises cite data privacy and sovereignty as a production challenge. | 中 | SM005 |
| CM014 | Public market evidence is broad and multi-lensed, but it does not isolate a clean Nscale-specific SAM or SOM because the available proxies mix global spend forecasts, deployment-model growth, and infrastructure bottleneck indicators. | 中 | SM002, SM004, SM005 |
| CM015 | The buyer map is directionally visible through company positioning, Microsoft anchor demand, and partner software claims, but public sources still stop short of disclosing account-level pricing, utilization guarantees, or budget authority by segment. | 中 | SM001, SM006, SM008 |
| CP001 | Nscale publicly positions itself as a vertically integrated AI infrastructure stack spanning power, data centers, GPU compute, and software. | 中 | SP004, SP006 |
| CP002 | The Anyscale acquisition is intended to move Nscale beyond raw compute and into a fuller software-plus-infrastructure platform position. | 中 | SP001, SP003, SP004 |
| CP003 | Anyscale is expected to keep operating under its own brand and let customers run workloads on infrastructure other than Nscale after the deal closes. | 中 | SP003, SP004 |
| CP004 | The relevant competitive field includes hyperscalers, specialist GPU clouds, neocloud peers moving into software, portable software platforms, and internal build as a substitute path. | 中 | SP001, SP003, SP004 |
| CP005 | Nscale's clearest public differentiation versus many GPU-cloud alternatives is ownership of physical infrastructure and power plus a growing orchestration layer. | 中 | SP002, SP003, SP006 |
| CP006 | Nscale competes with unusual financial and deployment scale for a young private company, including a $2.0 billion Series C and multiple large secured facilities tied to GPU and data-center projects. | 中 | SP007, SP008 |
| CP007 | Nscale's anchor-contract model creates distribution strength but also customer concentration, including a six-year Microsoft-backed take-or-pay structure behind the Ward County facility. | 中 | SP007, SP008 |
| CP008 | Assigned public evidence shows Nscale on-demand list pricing at $1.10 per hour for AMD MI300X and $3.85 per hour for NVIDIA H100, with spot instances also available. | 中 | SP009 |
| CP009 | Open-source portability limits moat inflation because Ray remains community-governed and Anyscale customers will still be able to run on non-Nscale infrastructure. | 中 | SP003, SP004 |
| CP010 | Nscale's competitive delivery story is vulnerable to physical execution risk because the Loughton project is funded and consented but delayed by grid-power availability. | 中 | SP005, SP007 |
| CP011 | Independent coverage reports that Nebius acquired workload-efficiency company Eigen AI and Qualcomm acquired compiler startup Modular as other moves up the AI infrastructure software stack. | 中 | SP003 |
| CP012 | Nscale's partner ecosystem with VAST and NVIDIA suggests the stack is vertically integrated but still relies on external ecosystem components rather than a fully closed architecture. | 中 | SP002, SP006 |
| CP013 | The acquisition strengthens Nscale's software story, but the core buyer trade-off remains efficiency versus lock-in when one vendor owns more of the stack. | 中 | SP003, SP004 |
| CP014 | Internal build remains a credible substitute for sophisticated buyers because portable orchestration can run elsewhere while Nscale's largest site advantages still depend on successful physical delivery. | 中 | SP003, SP005 |
| CP015 | Nscale's official positioning addresses practitioners, enterprises, and governments rather than only self-service GPU renters. | 中 | SP002 |
| CP016 | ComputeStacker lists spot instances for Nscale and claims they can reduce cost by 30–70% when workloads are interruptible. | 低 | SP009 |
| CI001 | Public reporting describes Nscale's monetization stack as two products: Nscale Infrastructure for dedicated GPU clusters and AI data-center capacity, and Nscale Cloud for managed AI compute services. | 中 | SI004, SI007, SI008 |
| CI002 | Customers generally pay by GPU-hour, with storage priced separately per gigabyte per month, and most infrastructure revenue comes from long-term take-or-pay contracts with a weighted-average life of about 5.7 years as of 2026-08-31. | 中 | SI004 |
| CI003 | As of 2026-08-31, Nscale reported approximately $103.4 billion of active and contracted TCV and 461,000 active or contracted GPUs, but only about $2.6 billion of TCV and roughly 25,000 GPUs were active. | 中 | SI004 |
| CI004 | Nscale reported H1 2026 revenue of $140.6 million and full-year 2025 revenue of $33.0 million. | 中 | SI004 |
| CI005 | Publicly disclosed H1 2026 cost of revenue excluding depreciation and amortization was $189.6 million, above the period's $140.6 million of revenue, implying negative gross margin on the disclosed base. | 中 | SI004 |
| CI006 | Nscale reported an H1 2026 net loss of $1.02 billion and an adjusted EBITDA loss of $199.2 million. | 中 | SI004, SI005 |
| CI007 | Nscale reported about $1.69 billion of operating cash flow in H1 2026 mainly because of upfront customer prepayments, while investing cash outflow was about $3.29 billion and June cash was about $1.48 billion. | 中 | SI004 |
| CI008 | In March 2026, Nscale raised a $2 billion Series C at a $14.6 billion valuation. | 高 | SI001, SI005 |
| CI009 | Nscale's February 2026 delayed-draw term loan provided $1.4 billion of GPU-backed debt intended to fund capital expenditure tied to signed contracts and customer pipeline. | 高 | SI001, SI002 |
| CI010 | Nscale's largest customer represented 52% of H1 2026 revenue and 73% of 2025 revenue, and the company expects Microsoft and Anthropic to remain significant customers. | 中 | SI004 |
| CI011 | Nscale's January 2025 UK press release said the Loughton site was equipped for 50MW of AI and HPC capacity, could scale to 90MW, and was scheduled to be live in Q4 2026. | 中 | SI003 |
| CI012 | Bricks & Bytes reported that National Grid could not deliver the Loughton connection in time for the planned opening, leaving Nscale to pursue on-site power options. | 中 | SI006 |
| CI013 | As of the filing date discussed by BBAE, Nscale had not yet obtained binding financing commitments needed to perform under the Anthropic agreements. | 中 | SI004 |
| CI014 | Public and partner-facing messaging positions Nscale as a vertically integrated AI infrastructure platform spanning data centers, networking, data services, and orchestration software rather than pure colocation. | 中 | SI001, SI007, SI008 |
| CI015 | Public sources still do not disclose realized pricing, revenue mix by product, CAC or payback, debt covenants, product-level margins, or customer-specific contract protections, so core underwriting inputs remain unavailable. | 中 | SI004, SI005, SI006 |
| CE001 | Nscale publicly positions itself as a vertically integrated AI infrastructure platform spanning power, data centers, GPU compute, networking, data services, and orchestration software. | 高 | SE001, SE006, SE008 |
| CE002 | Nscale frames its customer workflow around helping practitioners, enterprises, and governments create, deploy, and scale AI systems rather than around a single point product. | 高 | SE001, SE006 |
| CE003 | Independent reporting says the bulk of Nscale's built data-center infrastructure is in Norway, while the UK product story still centers on sovereign-AI expansion and newly announced campuses. | 中 | SE002 |
| CE004 | BBAE reports that Nscale's commercial offer splits into Nscale Infrastructure for dedicated GPU clusters and AI data-center capacity and Nscale Cloud for managed compute, training, fine-tuning, and inference. | 中 | SE004 |
| CE005 | As of Aug. 31, 2026, CNBC and BBAE report Nscale had 25,000 active GPUs, 461,000 active or contracted GPUs, and five active plus 12 contracted data-center sites. | 高 | SE004, SE005 |
| CE006 | Nscale's Loughton campus commitment with Microsoft is designed for 50MW of AI capacity, scalable to 90MW, with 23,040 NVIDIA GB300 GPUs planned for initial delivery in Q1 2027. | 高 | SE002, SE003 |
| CE007 | Nscale says Stargate UK is meant for sovereign workloads, with OpenAI exploring offtake of up to 8,000 NVIDIA GPUs in Q1 2026 and potential to scale to 31,000 GPUs over time. | 中 | SE003 |
| CE008 | VAST says Nscale placed the VAST AI Operating System and NVIDIA Cloud Partner reference architecture at the core of its offerings to optimize GPU efficiency, unify data access, and support orchestration and observability across regions. | 中 | SE008 |
| CE009 | CNBC and BBAE report that Nscale is extending the stack upward into software and inference, including a planned acquisition of Anyscale and support for inference requests across multiple models. | 高 | SE004, SE005 |
| CE010 | Nscale's public leadership disclosures show dedicated product, infrastructure, power, legal, and information-security ownership, including a CISO responsible for cyber, physical, and supply-chain security. | 中 | SE006 |
| CE011 | Bricks & Bytes reports that the Loughton project has planning permission, a Microsoft anchor tenant, and committed capital, but the grid connection will not arrive in time, pushing Nscale toward on-site generation. | 中 | SE007 |
| CE012 | Public 2026 reporting indicates commercialization is running ahead of live infrastructure because only about $2.6 billion of Nscale's $103.4 billion active-and-contracted TCV was active and only about 25,000 of 461,000 GPUs were active by late August 2026. | 高 | SE004, SE005 |
| CE013 | Customer and platform concentration remain material product dependencies because Microsoft and Anthropic anchor much of contracted demand and CNBC says one unnamed customer accounted for over half of first-half 2026 revenue. | 高 | SE004, SE005 |
| CE014 | Computer Weekly reports that Nscale acquired Kontena in July 2024 to add high-density modular generative-AI data-center capability to its proposition. | 中 | SE002 |
| CE015 | BBAE reports that as of Aug. 31, 2026, Nscale had approximately 1.37GW of active and contracted power capacity and line of sight to roughly 10GW of potential capacity. | 中 | SE004 |
| CU001 | Nscale publicly positions its target customer base as AI natives, enterprises, governments, and sovereignty-sensitive institutions that want controlled AI infrastructure. | 高 | SU003, SU006 |
| CU002 | The strongest named customer proof in the reviewed public record is Microsoft-linked anchor demand rather than a broad roster of disclosed production customers. | 中 | SU001, SU005, SU007 |
| CU003 | Nscale said its Loughton site can support 50MW of AI and HPC capacity, scale to 90MW, and is scheduled to go live in Q4 2026. | 高 | SU003, SU007 |
| CU004 | Nscale said the Loughton facility can house up to 45,000 NVIDIA GB200 GPUs. | 高 | SU003, SU007 |
| CU005 | Datacenter Frontier reported Nscale announced a letter of intent with Microsoft for up to 1.35GW of AI compute capacity at the Monarch campus, with deliveries expected to begin in late 2027. | 中 | SU001 |
| CU006 | Datacenter Frontier reported Nscale said it would provide Microsoft with roughly 104,000 NVIDIA GB300 GPUs at a Texas campus initially sized at approximately 240MW and potentially scaling to 1.2GW. | 中 | SU001 |
| CU007 | Datacenter Frontier reported the same October 2025 announcement included plans to deploy 12,600 GB300 GPUs at Start Campus in Sines, Portugal. | 中 | SU001 |
| CU008 | Data Center Knowledge reported Nscale partnered with OpenAI to launch a Stargate-branded AI center in Norway during 2025. | 中 | SU005 |
| CU009 | VAST Data described Nscale's enterprise customer target as organizations training, fine-tuning, and deploying AI models at global scale. | 中 | SU008 |
| CU010 | The reviewed public sources do not disclose customer count, NRR, GRR, churn, contract length, renewal history, or live utilization. | 中 | SU002, SU006 |
| CU011 | Startup Fundraising explicitly said Nscale's pitch materials omitted specific customer traction or case studies. | 中 | SU002 |
| CU012 | Customer concentration risk appears high because Microsoft is the recurring named buyer across the UK, Texas, and Monarch narratives while diversification beyond OpenAI-linked activity and broad enterprise targeting remains thin in the public record. | 中 | SU001, SU005, SU007, SU008 |
| CU013 | Bricks & Bytes reported that the funded and consented Loughton project is delayed by grid power timing, so marquee customer demand does not yet guarantee near-term revenue realization. | 中 | SU007 |
| CU014 | Nscale's public customer story is stronger on forward-looking campus commitments and strategic positioning than on demonstrated retention or diversified production usage. | 中 | SU001, SU002, SU003, SU005, SU008 |
| CR001 | Multiple independent reports say the Loughton site needs a 90MW connection and the grid timeline has slipped into the early-to-mid-2030s. | 中 | SR001, SR002, SR005 |
| CR002 | Independent reports describe Microsoft as Loughton's anchor tenant and say the site was planned to scale from 50MW to 90MW, so the delay affects a flagship customer deployment rather than an idle speculative build. | 中 | SR002, SR005 |
| CR003 | Public reporting says Nscale is evaluating on-site power options, including Bloom Energy fuel cells, rather than relying only on the delayed grid connection. | 中 | SR002, SR005 |
| CR004 | By March 2026 Nscale had announced a $2 billion Series C after previously disclosing a $1.4 billion GPU-backed delayed-draw term loan. | 高 | SR007, SR010 |
| CR005 | Public reporting ties Nscale to large Microsoft deployments, including a Texas facility targeting 104,000 NVIDIA GB300 GPUs. | 高 | SR007, SR009 |
| CR006 | Sacra says Microsoft's relationship with Nscale spans Norway, Texas, Portugal, and an additional $812 million AI investment partnership, implying multi-geography exposure to one counterparty. | 中 | SR009 |
| CR007 | Gracia Group's March 2026 research note says Nscale has signed take-or-pay contracts from Microsoft, NVIDIA, and ByteDance and an advanced pipeline weighted toward OpenAI, which would leave backlog quality concentrated in a few names if accurate. | 中 | SR008 |
| CR008 | Tech Startups, citing Bloomberg, reported in April 2025 that Nscale was seeking $2.7 billion tied to a pending ByteDance partnership, showing that some expansion plans were publicly linked to a small number of very large counterparties. | 低 | SR004 |
| CR009 | VAST's October 2025 partner release says Nscale planned to scale to 300,000 NVIDIA Grace Blackwell GPUs globally, increasing supplier and execution dependence as buildout accelerates. | 中 | SR006 |
| CR010 | Nscale's about page and VAST's partner release both describe Nscale as controlling multiple layers of the stack, so power, infrastructure, and software failures can propagate across one vertically integrated offering. | 高 | SR003, SR006 |
| CR011 | Nscale's public leadership roster includes a Chief Legal Officer, General Counsel, and CISO, indicating that governance and security mitigation capacity is being built internally. | 中 | SR003 |
| CR012 | TNW and Tech Funding News both report that Nscale added Sheryl Sandberg, Susan Decker, and Nick Clegg to the board and folded the Aker-Nscale joint venture into Nscale to simplify governance. | 高 | SR007, SR010 |
| CR013 | Sacra explicitly lists GPU allocation risk and Nordic power dependence as company-level risks that could delay capacity expansion and revenue recognition from committed customer contracts. | 中 | SR009 |
| CR014 | The assigned public evidence set does not independently verify Nscale's licenses, certifications, litigation status, or incident history. | 低 | |
| CV001 | Nscale consistently describes itself as a vertically integrated AI infrastructure platform spanning power, data centers, compute, and software. | 高 | SV001, SV006, SV008 |
| CV002 | Partner and company materials describe expansion across Europe and North America, including plans to scale to 300,000 NVIDIA Grace Blackwell GPUs globally. | 高 | SV003, SV008 |
| CV003 | Independent reporting said Nscale raised a $2 billion Series C in March 2026 at a $14.6 billion valuation. | 中 | SV002, SV005 |
| CV004 | Nscale announced a $3.36 billion pre-IPO convertible financing in September 2026, with $2.36 billion funded at closing and another $1 billion committed by NVIDIA. | 中 | SV001 |
| CV005 | Nscale said the September 2026 notes convert automatically on an IPO, so the eventual dilution and preference impact depend on undisclosed term-sheet details. | 中 | SV001 |
| CV006 | Nscale's September 2026 financing announcement cited more than $103 billion of total contracted value, which is a company-defined contract metric rather than disclosed revenue or ARR. | 中 | SV001 |
| CV007 | A market-data page listed Nscale at a $21.4 billion valuation and active IPO status in October 2026, but the assigned pool does not include an official filing that corroborates those assertions. | 低 | SV004 |
| CV008 | Bricks & Bytes reported that Nscale's funded, consented Loughton project had Microsoft as anchor tenant and a 90MW grid connection agreed, but power would not arrive in time for opening. | 中 | SV007 |
| CV009 | The same adverse report said Nscale grew headcount 121% in the first half of 2026 and had raised a $900 million credit facility before the delay became public, framing the near-term constraint as energization rather than immediate liquidity. | 中 | SV007 |
| CV010 | Nscale's Series B announcement tied its rollout to Microsoft, NVIDIA, and OpenAI initiatives, including Stargate Norway and a multi-billion-dollar Microsoft agreement in Narvik. | 中 | SV003 |
| CV011 | Nscale's about page shows it has recruited senior operators with public-capital, infrastructure, and hyperscale operating backgrounds, including a CFO from JPMorgan and operating leaders from Meta, Oracle, Palantir, and Microsoft. | 中 | SV006 |
| CV012 | Finance Monthly reported that Sheryl Sandberg, Susan Decker, and Nick Clegg joined Nscale's board alongside the March 2026 raise. | 中 | SV005 |
| CV013 | In the retained valuation-context sources, no publisher discloses audited revenue, ARR, gross margin, EBITDA, cash balance, or burn, so public support for Nscale's price remains narrative- and contract-based rather than financial-multiple-based. | 中 | SV001, SV002, SV005, SV006, SV007, SV008 |
| CV014 | Because the only visible step-up marker above the March 2026 $14.6 billion valuation is a low-transparency market-data page, the assigned public record supports strategic relevance and capital access more strongly than it supports a precise present-day fair value above that private mark. | 中 | SV001, SV003, SV004, SV005 |
| CV015 | The assigned public record still centers on marquee counterparties and flagship infrastructure projects rather than a clearly diversified customer base, leaving concentration as a valuation-sensitive risk. | 中 | SV003, SV007, SV008 |
| 编号 | 出版方 | 标题 | 引文 |
|---|---|---|---|
| SO001 | Nscale | Nscale Raises $2 Billion in Series C - the Largest in European History | Nscale | |
| SO002 | Nscale | The engine of superintelligence | Nscale | |
| SO003 | Nscale | About | Nscale | |
| SO004 | Sacra | Nscale valuation, funding & news | |
| SO005 | The Next Web | Nscale raises $2bn Series C at $14.6bn valuation | |
| SO006 | Silicon Valley Invest Club | Nscale - Funding, Valuation & Leadership | |
| SO007 | Bricks & Bytes | Nscale Has GBP 2bn and No Power - The Grid Connection Now Runs UK Data Centers | |
| SO008 | Tech Funding News | Nscale's $2B Series C makes it Europe's most valuable AI infrastructure startup - TFN | |
| SM001 | TechCrunch | Nscale inks massive AI infrastructure deal with Microsoft | |
| SM002 | Acquinox Capital | NScale: Europe's Billion Dollar Bet on AI Sovereignty | |
| SM003 | Nscale | AI hyperscaler Nscale to invest 2 billion GBP in the UK data centre industry | |
| SM004 | Ainvest | Nscale's Vertical AI Stack: Building the Rail for the AI S-Curve's Next Phase | |
| SM005 | Next Move Strategy Consulting | GPU Cloud Industry Update: Nscale Acquires Anyscale | |
| SM006 | Nscale | About | Nscale | |
| SM007 | Bricks & Bytes | Nscale Has 2bn GBP and No Power - The Grid Connection Now Runs UK Data Centers | |
| SM008 | VAST Data | Nscale, VAST Data, and NVIDIA Unite to Build the Global Fabric for AI | |
| SP001 | BriefGlance | Nscale Buys Anyscale, Forging a Vertically Integrated AI Hyperscaler | Nscale is now more directly challenging not only established cloud giants like Amazon Web Services and Google Cloud but also other specialized GPU cloud providers like CoreWeave and Lambda Labs. By moving beyond raw compute to offer a comprehensive software platform, Nscale is aiming to capture a larger share of enterprise AI budgets. |
| SP002 | Nscale | About | Nscale | Nscale is building the advanced infrastructure, systems and solutions that enables practitioners, enterprises, and governments across the globe to create, deploy, and scale their most transformative AI systems. |
| SP003 | LiveAIWire | Nscale Just Paid $1.65 Billion to Stop Being Just a GPU Landlord | The counterargument, implicit in why Ray remained open source even as its commercial arm changed hands, is that customers value the ability to run the same orchestration layer on infrastructure from any vendor, and lock-in risk is exactly what a fully vertically integrated provider is asking customers to accept in exchange for that efficiency. |
| SP004 | Anyscale | Nscale Acquires Anyscale, Enhancing its Full Stack AI Cloud Platform | Anyscale will continue to operate under its brand and serve its existing customers as it does today. Customers will remain free to choose the infrastructure on which they run their AI workloads, and over time will gain the additional option of running the Anyscale software layer on Nscale’s full stack AI platform. |
| SP005 | Bricks & Bytes | Nscale Has £2bn and No Power - The Grid Connection Now Runs UK Data Centers | A consented, funded, Microsoft-anchored data center still cannot switch on because the grid cannot deliver in time. |
| SP006 | VAST Data | Nscale, VAST Data, and NVIDIA Unite to Build the Global Fabric for AI | With recent expansion announcements across Europe, and with plans to scale to 300,000 NVIDIA Grace Blackwell GPUs globally, Nscale is building AI infrastructure that spans Europe and North America to support enterprises training, fine-tuning, and deploying models at unprecedented scale. |
| SP007 | The Next Web | Nscale raises $2bn Series C at $14.6bn valuation | The question is no longer whether Nscale can raise money. It is whether the infrastructure it is racing to build will be ready before the market moves on. |
| SP008 | CCIR | Nscale credit issuer page | 6-year availability-based fixed-price take-or-pay GPU-as-a-service contract with Microsoft Corporation, named as the Customer in the filed credit agreement |
| SP009 | ComputeStacker | NScale provider page | AMD MI300X On-Demand $1.10/hr ~$803 /mo NVIDIA H100 On-Demand $3.85/hr ~$2,811 /mo |
| SI001 | CNBC | Nvidia backs AI data center startup Nscale as it hits $14.6 billion valuation | AI data center startup Nscale has raised $2 billion at a $14.6 billion valuation, the company announced on Monday, as the AI infrastructure boom continues. |
| SI002 | Tech Funding News | How Nscale's $1.4B debt deal powers renewable AI clusters without equity dilution | The GPU-backed Delayed Draw Term Loan enables Nscale to use structured debt to finance a portion of capital expenditure associated with large-scale GPU clusters. |
| SI003 | Nscale | AI hyperscaler Nscale to invest £2 billion in the UK data centre industry | The site is equipped to support 50MW of AI and HPC capacity, with the ability to scale the power allocation up to 90MW... The site is scheduled to be live in Q4 2026. |
| SI004 | BBAE | Nscale ($NSCL) IPO: An Investor's Guide to the AI Hyperscaler | Customers generally pay by GPU-hour, with storage priced separately per gigabyte per month. Most infrastructure revenue comes from long-term take-or-pay contracts, meaning the customer commits to paying for specified capacity once it has been built and accepted. |
| SI005 | Sector Spotlight UK | Nscale AI Infrastructure: The UK's AI Power Play | The growth comes with a heavy price. Nscale reported a $1.02 billion net loss for the first half of 2026... The headline contract value also needs careful reading. Nscale has not disclosed how much of its $103 billion total has been recognised as revenue, or when the rest will land. |
| SI006 | Bricks & Bytes | Nscale Has £2bn and No Power - The Grid Connection Now Runs UK Data Centers | A consented, funded, Microsoft-anchored data center still cannot switch on because the grid cannot deliver in time. |
| SI007 | VAST Data | Nscale, VAST Data, and NVIDIA Unite to Build the Global Fabric for AI | The vision is bold: a fully vertically integrated, globally distributed AI cloud. Nscale controls every layer of the stack, from renewable-powered data centers to orchestration software and observability tools... |
| SI008 | Nscale | About | Nscale | Nscale is building the advanced infrastructure, systems and solutions that enables practitioners, enterprises, and governments across the globe to create, deploy, and scale their most transformative AI systems. |
| SE001 | Redalpine | Nscale: building the foundational layer for the AI era | The UK-headquartered AI infrastructure hyperscaler is building the foundational layer for global AI deployments, delivering GPU compute, networking, data services, and orchestration software through a fully vertically integrated platform. From data center design to inference workloads, Nscale handles the full stack, enabling enterprises to run AI training, fine-tuning, and inference at genuine production scale. |
| SE002 | Computer Weekly | Nscale explained: Everything you need to know | |
| SE003 | Nscale | Nscale announces UK AI infrastructure commitment in partnership with Microsoft, NVIDIA and OpenAI | Nscale | The site will deliver 50MW of AI capacity, scalable to 90MW, and will initially house 23,040 NVIDIA GB300 GPUs delivered in Q1 of 2027. |
| SE004 | BBAE | Nscale ($NSCL) IPO: An Investor’s Guide to the AI Hyperscaler | The business splits into two products. Nscale Infrastructure, which accounts for the vast majority of contracted business, provides dedicated GPU clusters and AI data-center capacity to large customers such as hyperscalers and frontier AI labs. |
| SE005 | CNBC | AI cloud provider Nscale files to go public | As of Aug. 31, Nscale held 25,000 active GPUs and 461,000 that were either active or contracted, with five active and 12 contracted data center sites. |
| SE006 | Nscale | About | Nscale | Latha serves as CISO for Nscale where she oversees cyber, physical and supply chain security. |
| SE007 | Bricks & Bytes | Nscale Has £2bn and No Power - The Grid Connection Now Runs UK Data Centers - Bricks & Bytes | A consented, funded, Microsoft-anchored data center still cannot switch on because the grid cannot deliver in time. |
| SE008 | VAST Data | Nscale, VAST Data, and NVIDIA Unite to Build the Global Fabric for AI | Nscale controls every layer of the stack, from renewable-powered data centers to orchestration software and observability tools, while VAST is a key partner in Nscale’s offerings with a unified data layer that sustains GPU efficiency and ensures seamless data access across regions. |
| SU001 | Data Center Frontier | Nscale Expands AI Factory Strategy With Power, Platform, and Scale | Anchor customers are a critical component of AI campuses at this scale, and Nscale’s letter of intent with Microsoft provides the clearest window into how the company is structuring its development strategy. |
| SU002 | Startup Fundraising | Nscale Pitch Deck: All 9 Slides + Teardown | The deck omits all financial performance data, current burn rates, and specific customer traction or case studies. |
| SU003 | Nscale | AI hyperscaler Nscale to invest £2 billion in the UK data centre industry | The new data centres will provide essential UK sovereign, secure and private AI computing infrastructure for industries, research institutions, and enterprises, ensuring data remains securely within Europe while driving innovation and stimulating economic growth. |
| SU004 | Tech Funding News | Nscale's $2B Series C makes it Europe's most valuable AI infrastructure startup | Enterprises running generative AI training, fine-tuning, and inferencing workloads gain access to tightly optimised compute environments rather than fragmented infrastructure stacks. |
| SU005 | Data Center Knowledge | Nvidia-backed Nscale Soars to $14.6B Valuation | In October, Nscale announced an expanded partnership with Microsoft that secured a reported $14 billion. Over the summer, the company partnered with OpenAI to launch a Stargate-branded AI center in Norway. |
| SU006 | Nscale | About | Nscale | Nscale is building the advanced infrastructure, systems and solutions that enables practitioners, enterprises, and governments across the globe to create, deploy, and scale their most transformative AI systems. |
| SU007 | Bricks & Bytes | Nscale Has £2bn and No Power - The Grid Connection Now Runs UK Data Centers | A consented, funded, Microsoft-anchored data center still cannot switch on because the grid cannot deliver in time. |
| SU008 | VAST Data | Nscale, VAST Data, and NVIDIA Unite to Build the Global Fabric for AI | With recent expansion announcements across Europe, and with plans to scale to 300,000 NVIDIA Grace Blackwell GPUs globally, Nscale is building AI infrastructure that spans Europe and North America to support enterprises training, fine-tuning, and deploying models at unprecedented scale. |
| SR001 | TechRadar | Nvidia-backed Neocloud Nscale faces years of delay because UK's largest AI supercomputer can't get enough power from the grid to feed its data centers | UK Power Networks has told Nscale that the electricity grid will not be able to produce enough power for the site until the early 2030s |
| SR002 | Crypto Briefing | Nscale's UK supercomputer project delayed until mid-2030s due to power issues | The grid connection delays were first flagged in July 2026, with revised availability now pushed out to the early-to-mid-2030s. |
| SR003 | Nscale | About | Nscale | Adam Rosman is Chief Legal Officer at Nscale, where he leads the company's legal, regulatory, and governance functions. |
| SR004 | Tech Startups | London AI startup Nscale seeks $2.7B in funding to build global data centers in partnership with ByteDance | According to an offering document seen by Bloomberg, the company is looking to raise a massive $2.7 billion to build out artificial intelligence infrastructure. The plan is tied to a pending partnership with ByteDance Ltd., the Chinese tech giant behind TikTok. |
| SR005 | Bricks & Bytes | Nscale Has £2bn and No Power - The Grid Connection Now Runs the Job | A consented, funded, Microsoft-anchored data center still cannot switch on because the grid cannot deliver in time. |
| SR006 | VAST Data | Nscale, VAST Data, and NVIDIA Unite to Build the Global Fabric for AI | With recent expansion announcements across Europe, and with plans to scale to 300,000 NVIDIA Grace Blackwell GPUs globally, Nscale is building AI infrastructure that spans Europe and North America to support enterprises training, fine-tuning, and deploying models at unprecedented scale. |
| SR007 | The Next Web | Nscale raises $2bn Series C at $14.6bn valuation | Since its Series B in September 2025, it has signed a $1.4 billion delayed-draw term loan backed by GPUs, which it announced in February 2026, and has secured large-scale contracts with Microsoft, including plans for a facility in Texas targeting 104,000 NVIDIA GB300 GPUs. |
| SR008 | Gracia Group | NScale — Pre-IPO Research | The company has $37 billion in signed take-or-pay contracts from Microsoft, NVIDIA, and ByteDance. When including their $18.5 billion advanced pipeline—which is largely weighted toward OpenAI—the total backlog reaches $56 billion. |
| SR009 | Sacra | Nscale valuation, funding & news | GPU supply constraints: Nscale's growth depends on securing sufficient NVIDIA H100, B200, and future Blackwell GPU allocations in a supply-constrained market where larger competitors like CoreWeave have secured massive take-or-pay agreements. |
| SR010 | Tech Funding News | Nscale's $2B Series C makes it Europe's most valuable AI infrastructure startup | The announcement comes a few weeks after raising $1.4 billion in a Delayed Draw Term Loan backed by GPUs. |
| SV001 | Nscale via PR Newswire | NSCALE RAISES $3.36B IN PRE-IPO CONVERTIBLE FINANCING | The financing consists of an initial $2.36bn tranche at closing and an additional $1bn commitment from NVIDIA with funding expected in mid-November, 2026. The loan notes are convertible into ordinary shares (or non-voting shares in the case of NVIDIA) automatically upon completion of Nscale's initial public offering. |
| SV002 | Built In London | Nscale Raises $2B Series C at $14.6B Valuation | Nscale, a London-based company focused on hyperscale AI infrastructure, announced $2 billion in Series C funding led by Aker ASA and 8090 Industries. The company, now valued at $14.6 billion, offers a vertically integrated compute foundation to help businesses train and refine AI models at scale. |
| SV003 | Nscale | Nscale Raises the Largest Series B in European History, at $1.1 billion | LONDON, UK, 08:00 BST 25th September 2025 – Nscale, the hyperscaler engineered for AI, today announced it has raised the largest Series B in UK and European history, at USD $1.1 billion. |
| SV004 | TechStackIPO | Nscale — Funding, Valuation & IPO Status | Nscale: $21.4B valuation, IPO Readiness 67, Public. Filed S-1 with the SEC on September 2026 — IPO actively in progress. |
| SV005 | Finance Monthly | Nscale Founder Josh Payne Raises $2bn as UK Expands AI Infrastructure | London-based AI infrastructure company Nscale has raised $2bn (£1.6bn) in a Series C funding round, valuing the business at $14.6bn as demand for the computing power needed to support artificial intelligence continues to surge. |
| SV006 | Nscale | About | Nscale | Nscale is building the advanced infrastructure, systems and solutions that enables practitioners, enterprises, and governments across the globe to create, deploy, and scale their most transformative AI systems. |
| SV007 | Bricks & Bytes | Nscale Has £2bn and No Power - The Grid Connection Now Runs UK Data Centers | Nscale has the cash, the site, and the customer. What it does not have is a plug. |
| SV008 | VAST Data | Nscale, VAST Data, and NVIDIA Unite to Build the Global Fabric for AI | With recent expansion announcements across Europe, and with plans to scale to 300,000 NVIDIA Grace Blackwell GPUs globally, Nscale is building AI infrastructure that spans Europe and North America to support enterprises training, fine-tuning, and deploying models at unprecedented scale. |