Andromeda
GPU 流动性经纪商已有真实牵引,但底层核验披露偏薄
Andromeda 看起来确实在解决 GPU 采购痛点,早期牵引也有分量;但如果没有更扎实的毛利质量、集中度和治理证据,当前 $1.5B 价格很难承保。
封面要素
公司概况
Andromeda 是一家总部位于 San Francisco 的私营公司,为需要灵活算力、又不想承担超大规模云厂商式长期承诺的 AI 团队撮合并运营第三方 GPU 集群访问。公司源自围绕 Nat Friedman 和 Daniel Gross 网络搭建的 Andromeda Cluster,如今由 CEO Wil Moushey 运营,并把自己定位成中立市场层:对供应商做基准测试、标准化合同、整合账单,并帮助买方在碎片化供应中部署和管理大规模算力。
- 创始人
- Nat Friedman, Daniel Gross
- 创立地点
- San Francisco, CA
- 总部
- San Francisco, CA
- 产品
- 一个市场与托管运营层,用于在众多第三方供应商之间采购、认证、定价、签约、部署并支持 GPU 集群。
- 客户
- 需要突发或规模化训练与推理基础设施的 AI 原生初创公司、前沿实验室和企业 AI 团队。
- 商业模式
- 经纪式 GPU 采购与运营,围绕第三方供应提供标准化合同、整合账单和支持。
- 阶段
- Private / growth
- 融资情况
- 2026 年 3 月由 Paradigm 支持的融资,估值 $1.5B;Paradigm 至今披露出资 $60M,但公司累计资本和优先权结构仍不透明。
执行摘要
主要优势
- 100+ 家供应方、1,000+ 笔交易和据报 ~$100M 2025 收入运行率,指向真实市场活动。
- 中立的多供应方寻源、标准化合同和托管运营,切中了 GPU 采购真实摩擦。
- 创始人与资助方履历强,让 Andromeda 早期就接触到前沿 AI 需求和资本。
主要风险
- Nat Friedman 和 Daniel Gross 退居非活跃角色后,治理与继任仍不透明。
- 公开材料没有披露客户留存、集中度,以及毛收入与净收入质量。
- 依赖第三方供应方,且公开信任 / 合规细节稀疏,可能挤压毛利和持久性。
未决问题
- 已审计数据或尽调室口径,打通确认收入、抽佣率、毛利率和现金转化。
- 客户与供应方集中度、NRR/GRR、合同期限和 SLA 表现数据。
- 股权结构表、老股交易或优先权条款、债务或担保,以及清晰的治理 / 继任图谱。
目录
01公司概览
1.1 身份、产品模式与总部信号
即便部分成立细节仍模糊,公开证据已相当清楚地说明 Andromeda 想成为什么。官网主页、洞察页、客户页、供应商页和隐私页面都指向一个双边算力市场:买方说明工作负载需求,供应方提交 GPU、网络、地点和可用性细节;Andromeda 对供应做基准测试、标准化合同、路由需求,并给客户一个统一账单和支持关系。这个组合说明,Andromeda 更像 AI 算力的中立市场基础设施层,而不是拥有大部分所售硬件的传统云运营商。官方文案中的产品叙事也一致:公司试图把采购、认证、定价和运营包进一个界面,让碎片化的 GPU 容量可用起来。 公司身份字段更窄。官方页脚和隐私政策把法律实体标为 Andromeda Cluster, Inc.。Built In、公司 X 账号和 Gaebler 都指向 San Francisco 作为当前基地,Gaebler 还给出 228 Grant Ave 这个邮寄地址信号。这个信息有用,但不应被视为总部最终结论,因为官网并未公布街道地址。创立时间也需要细分:Upstarts 把 Wil Moushey 的招募放在 2023 年 11 月下旬,而 2026 年 3 月的发布文章称团队已花三年时间搭建算力市场基础设施。因此,后续章节最稳妥的事实底座是:项目起源于 2023 年末,运营基地在 San Francisco,当前商业模式以撮合并运营第三方算力为核心,而不是直接拥有超大规模云厂商级足迹。[CO001, CO002, CO003, CO004, CO005, CO006]
Andromeda 把创始人与赞助人、供应商供给、买方需求、标准化合同和运营支持连成一个算力市场论点。
[CO001, CO002, CO003, CO004, CO010, CO020]1.2 创始人、当前领导层与治理不透明
创始人故事一致,但公开高管名单并不完整。Built In 和 Gaebler 都称 Andromeda Cluster 由 Nat Friedman 和 Daniel Gross 创立。独立来源随后补上这件事为何重要:DCD 和 CB Insights 把 Friedman 定位为 GitHub 前 CEO、NFDG 联合创始人;Daniel Gross 自己的网站和 DCD 则把他与 Apple、Y Combinator、NFDG 以及如今的 Meta 算力工作联系起来。这个组合让创始人在稀缺 AI 基础设施、客户引荐和资本获取上具备异常强的创始人与市场契合度。Upstarts 补上缺失的运营桥梁:Gross 在 2023 年末招募 Wil Moushey 来运营该项目,第三方 2026 年 3 月报道则称 Moushey 为 CEO。 治理透明度远弱于创始人与市场契合度。官网没有公开董事会名单、高管团队页面,也没有清晰的所有权或控制权地图。相反,当前公开可见的运营班底主要来自活跃招聘:法务、财务、合作伙伴、采购、业务运营、解决方案工程、软件和 SRE 岗位都在招聘。这显示公司确实在搭组织,但也意味着部分重要管理职能仍在补人,或至少尚未公开点名。因此,关键人物依赖度较高。Moushey 似乎是当前运营重心,而创始人仍主导历史叙事和赞助方网络。Friedman 与 Gross 后来转向 Meta,以及 Upstarts 称他们不再与 Andromeda 保持活跃关系,是公开记录中最重要的领导层变化;在董事会、所有权和继任细节直接确认前,应一直作为尽调重点。[CO009, CO011, CO012, CO013, CO014, CO015]
| 人物 / 层级 | 角色 / 状态 | 背景 / 公开证据 | 创始人市场契合或职能覆盖 | 关键人物依赖 |
|---|---|---|---|---|
| Nat Friedman | 创始人;NFDG 联合创始人 | 第三方来源称 Friedman 是前 GitHub CEO,也是 Andromeda 背后 NFDG 赞助方之一。 | 提供创始人可信度、AI 网络访问,以及早期资本 / 客户渠道。 | 高 |
| Daniel Gross | 创始人;NFDG 联合创始人 | 个人网站和第三方来源把 Gross 与 Apple、Y Combinator、NFDG 以及后来的 Meta 算力工作联系起来。 | 提供算力市场论点、赞助方网络和早期项目启动能力。 | 高 |
| Wil Moushey | 首席执行官 | Upstarts 称 Gross 在 2023 年末招募 Moushey 负责该项目;2026 年 3 月的跟踪资料将他列为 CEO。 | 当前运营负责人、外部发言人,也是把集群项目商业化为公司的桥梁。 | 高 |
| 运营班底(通过公开招聘可见) | 正在招聘,尚未完整披露具名领导层 | Ashby 和 Built In 显示,公司当前在法务、财务、合作、采购、业务人员、解决方案、软件和 SRE 等方向招聘。 | 显示公司正在围绕法务、资本规划、供给、GTM 和可靠性搭建职能覆盖。 | 中 |
| 董事会 / 治理层 | 未公开披露 | 已审阅官方页面没有公开董事会名单、观察员地图或委员会结构。 | 治理、控制权和继任仍不透明,需要管理层室披露。 | 高 |
公开材料清楚点名了创始人和现任 CEO,但更广泛的管理层和董事会可见度主要来自岗位招聘信号,而不是正式高管名册。
[CO009, CO011, CO012, CO013, CO014, CO015]| 利益相关方 | 角色 | 控制权或经济重要性 | 尽调问题 |
|---|---|---|---|
| Paradigm | 已披露的领投资本方 | 锚定 2026 年 3 月 $1.5B 估值,以及公开的 $60M 总投资数字。 | 披露确切分笔金额、证券类型、董事会或信息权,以及是否包含任何老股转让。 |
| NFDG | 创始赞助方 / 前身资本来源 | 看起来资助了原始集群建设,并通过投资组合公司种下早期需求。 | 将拆分前资产资金与当前公司股权分开,并澄清任何保留所有权或合同权利。 |
| Nat Friedman | 创始人 / 赞助方 / 历史网络节点 | 前 GitHub CEO;在集群阶段,其网络很可能对资本、人才和客户访问很重要。 | 澄清当前持股、任何持续权利,以及转去 Meta 后是否仍保留治理或商业影响力。 |
| Daniel Gross | 创始人 / 赞助方 / 历史算力战略家 | 招募了 Moushey,并与 Andromeda 背后的算力市场论点高度绑定。 | 澄清剩余持股、顾问角色,以及转去 Meta 后是否存在任何商业关系。 |
| Wil Moushey | CEO 和运营负责人 | 当前执行、商业化和外部叙事的运营重心。 | 评估留任、股权激励、继任覆盖,以及围绕 Moushey 的客户或供应商关键人物集中度。 |
| AI Grant / NFDG 投资组合公司 | 早期客户渠道 | Upstarts 显示,早期利用和验证来自包括 ElevenLabs 和 Pika 在内的赞助投资组合公司。 | 量化当前关联方收入占比,以及该需求是否仍然集中或已经多元化。 |
| 供应商网络 | 核心供给侧利益相关方群体 | 100+ 家供应商以及标准化认证或合同,是商业模式的核心。 | 按头部供应商绘制集中度、排他条款、SLA 执行,以及经纪供给与直接供给的经济性。 |
公开来源识别了主要资本和生态利益相关方,但没有披露完整股权结构、投票权、老股组合或商业集中度。
[CO009, CO010, CO012, CO013, CO014, CO020]1.3 融资历史、估值锚点与规模指标
最强的清晰融资锚点是 2026 年 3 月的 Paradigm 交易。Upstarts 称,公司从 Paradigm 获得新融资,估值 $1.5B,且 Paradigm 对 Andromeda 的总投资现为 $60M。Raising.fi 和 Gaebler 都重复了 $60M 这一数字,Gaebler 将该事件归为风险股权融资;SiliconANGLE 则称,尽管 Paradigm 的累计投资达到这一水平,新近完成那一融资批次 的确切规模并不清楚。实务结论是,估值和赞助方身份有较好锚定,但轮次标签、证券类型和公司累计融资额没有。这个缺口重要,因为 NFDG 似乎也曾在拆分独立前为前身集群投入超过 $100M,公开来源没有清楚区分拆分独立前资产建设与当前公司的股权资本。 运营规模看起来比股权结构披露更扎实。官方和独立来源都指向 100+ 供应商和 1,000+ 笔交易,官方文案还补充了超过 10 亿 GPU-hour 活动量,以及跨众多供应商的数十个集群。Upstarts 和 SiliconANGLE 都报道,2025 年收入 run-rate 约 $100M,2024 年收入超过 $50M,且自推出以来已经盈利。但几项重要指标仍缺失或只有区间:客户数量未公开披露,员工数只能从 Built In 的 17 人和 Upstarts 的约 20 人侧面观察,也没有公开来源清楚确认二级交易、债务工具或拆分独立后完整资本结构。因此,后续章节应把估值、收入 run-rate、供应商广度和交易量作为可复用锚点,同时明确提示客户集中度、融资机制和团队规模仍有保留。[CO016, CO018, CO019, CO022, CO023, CO024]
| 指标 | 数值 / 状态 | 日期 | 置信度 | 缺口 / 备注 |
|---|---|---|---|---|
| 创立 / 项目起源 | 2023 年末启动项目;确切注册成立日期未披露 | 2023-11 | 中 | 公开来源支持 2023 年末已运营,但不支持精确的法律成立日期。 |
| 法律实体 | Andromeda Cluster, Inc. | 2026-07-01 | 高 | 官方页脚和隐私政策提供支持。 |
| 总部 / 基地 | San Francisco;Gaebler 列出 228 Grant Ave 邮寄地址 | 2026-07-01 | 高 | 街道地址信号来自第三方,因为官网没有公布地址。 |
| 当前阶段 | 私营、独立、从原项目拆分后由风投注资的算力市场 | 2026-03 to 2026-07 | 中 | 轮次标签没有清楚披露;Upstarts 称相当于 Series A,而 Gaebler 称为未披露的风险股权融资。 |
| 估值锚点 | $1.5B | 2026-03-18 | 高 | 这是公开资料中对 2026 年 3 月融资支持最充分的估值锚点。 |
| Paradigm 迄今投入资本 | 累计投入 $60M | 2026-03-18 至 2026-03-19 | 中 | 来源一致认同 Paradigm 的总投资额,但未必等同于最新一笔资金规模。 |
| 累计融资额 | 2026-07-01 | 低 | 无法清楚支撑,因为 NFDG 前身集群资本与当前公司股权没有明确拆开。 | |
| 收入运行率 | 2025 年 $100M;2024 年 >$50M | 2025 | 高 | 仅来自独立报道;没有经审计的财务披露。 |
| 盈利能力 | 自上线以来已盈利 | 2025-2026 | 高 | 由 Upstarts 和 SiliconANGLE 报道,并非来自经审计文件。 |
| 供应商网络 | 100+ 家算力供应商 | 2026-03 to 2026-07 | 高 | 官方计数器和独立报道相互印证。 |
| 交易量 | 已完成 1,000+ 笔交易 | 2026-03 to 2026-07 | 高 | 公司说法被 SiliconANGLE 重复。 |
| GPU 小时 / 集群规模 | 数十亿 GPU 小时流动性;超过 10 亿 GPU 小时;数十个集群 | 2026-07-01 | 高 | 来自官方营销计数器和洞察叙事,而不是经审计运营披露。 |
| 客户数 | 2026-07-01 | 低 | 公开来源描述了领先 AI 实验室和创业公司,但没有披露客户总数。 | |
| 员工数 | Built In 显示 17 人,Upstarts 称约 20 人 | 2026-03 to 2026-07 | 中 | 只能作为有用区间;没有官方员工数披露。 |
| 地点 / 招聘覆盖 | San Francisco 办公室,加上美国、北美和全球远程招聘 | 2026-07-01 | 高 | 招聘覆盖是公开的,尽管办公室总数没有清楚披露。 |
| 债务 / 信贷额度 | 2026-07-01 | 低 | 财务招聘和 Moushey 评论暗示融资复杂度,但没有发现公开融资额度条款。 |
公开指标混合了官方计数器、官方工作流页面和独立报道;null 表示该指标对判断很重要,但无法由公开证据支撑。
[CO005, CO006, CO008, CO018, CO019, CO022]最可用的当前公开指标把强估值锚、显著供应商和收入规模,与仍不透明的客户和资本结构细节混在一起。
客户数量、累计融资和债务仍未披露;多行 KPI 使用区间或混合来源的当前状态近似值,而不是经审计的公司报告。
[CO019, CO022, CO023, CO026, CO027, CO031]1.4 里程碑、商业化路径与当前风险信号
里程碑故事重要,因为 Andromeda 的公开身份滞后于其运营历史。Upstarts 显示,项目在 2023 年末交接给 Moushey,2024 年初为 NFDG 支持的公司扩展到 4,000 多块 GPU,随后走向更广泛的商业化。DCD 增加了 2025 年 6 月这个重要检查点:在 2026 年 3 月公开发布前,Andromeda Cluster 已按公开单 GPU 价格向非 NFDG 公司销售。2026 年 3 月 16 日的隐私政策修订是最清晰的公开法律标记,3 月 18 日的 X 帖和网站亮相则是最干净的外部发布标记。同一周也带来了 Paradigm 融资和最强公开估值锚点。截至 2026 年 7 月本报告运行日,公司仍在法务、财务、合作伙伴、采购、解决方案、软件和可靠性岗位活跃招聘,说明它正围绕资本与需求故事搭建运营底盘。 风险信号比典型困境公司红旗更隐蔽,但确实存在。DCD 2025 年报道称,Meta 围绕 NFDG 的部分收购谈判会如何影响 Andromeda 尚不清楚;Upstarts 后来称 Friedman 和 Gross 已不再与公司保持活跃关系。问题恰恰在于,创始人曾是早期供应、资本和客户叙事的中心,因此这带来连续性疑问。信任和隐私页面显示管理层重视法律与安全基础设施,但相对于收入和估值叙事,公开信任页面仍很稀疏。再加上董事会披露不透明、资本结构细节未解、客户数量未公开,本章应把 Andromeda 描述为一家高速扩张、似乎已盈利的算力市场;但其可投资性仍取决于管理层会议中对所有权、治理和集中度风险的确认。[CO007, CO011, CO016, CO021, CO025, CO028]
| 日期 | 事件 | 类型 | 金额 / 估值 / 状态 | 参与方 | 含义 |
|---|---|---|---|---|---|
| 2023-11 | Daniel Gross 招募 Wil Moushey 负责 Andromeda Cluster 项目。 | 治理 | 运营负责人到位 | Daniel Gross;Wil Moushey | 标志项目从创始人侧项目转向专职运营建设。 |
| 2024-01 | NFDG 支持的集群扩张到超过 4,000 张 GPU,并服务 ElevenLabs 和 Pika 等早期投资组合公司。 | 合作 | 2024 年初超过 4,000 张 GPU | NFDG;AI Grant / 投资组合公司 | 显示早期需求来自赞助方生态,并验证初始产品需求。 |
| 2025-06-21 | DCD 报道 Andromeda 正在以列明的单 GPU 定价向非 NFDG 公司销售算力。 | 产品 | $2.40-$3.00 / GPU-hour 报价 | Andromeda;DCD;外部买方 | 确认公司在 2026 年 3 月公开发布前已经商业化。 |
| 2025-07 | Nat Friedman 和 Daniel Gross 转去 Meta,之后被描述为不再活跃参与 Andromeda。 | 负面 | 创始人连续性风险上升 | 相关方:Meta;Nat Friedman;Daniel Gross | 原始赞助方退居幕后,制造了治理和连续性问题。 |
| 2026-03-16 | 隐私政策为当前商业用途市场平台更新。 | 监管 | 政策更新 | Andromeda Cluster, Inc. | 提供了已审阅来源集中最清楚的公开法律 / 合规里程碑。 |
| 2026-03-18 | X 帖文和网站发布公开宣布 Andromeda,此前公司已用三年建设算力市场基础设施。 | 创立 | 公开品牌发布 | Andromeda | 锚定当前公司身份的外部发布日期。 |
| 2026-03-18 | Paradigm 支持的融资披露,估值为 $1.5B。 | 融资 | $1.5B 估值;Paradigm 总投资后来被报道为 $60M | Paradigm;Andromeda | 为后续章节建立最清楚的资本和估值锚点。 |
| 2026-03-19 | 公司社交帖公开感谢 Nat Friedman、Daniel Gross 和 Hersh Desai。 | 治理 | 赞助方致谢 | 相关方:Andromeda;Nat Friedman;Daniel Gross;Hersh Desai | 显示拆分后仍与创始赞助方网络保持叙事连接。 |
| 2026-07-01 | 官网和报道显示有 100+ 家供应商、1,000+ 笔交易和超过 10 亿 GPU 小时活动。 | 规模 | 当前运营规模主张 | Andromeda;供应商网络 | 即便没有公开客户数披露,也展示了有意义的市场活动。 |
| 2026-07-01 | 活跃招聘覆盖法务、财务、采购、合作、算力市场、解决方案、软件和 SRE 岗位。 | 规模 | 公开可见 10+ 个职能族群 | Andromeda 招聘团队 | 表明公司正在围绕供给、治理、GTM 和可靠性继续搭组织。 |
这条时间线是在 2026 年运行日期前,公开材料中关于创立、商业化、法律、融资、治理和规模里程碑的最佳序列;方向性很强,但并不穷尽。
[CO007, CO011, CO016, CO021, CO022, CO023]从 2023 年底招募运营者,到 2026 年 3 月发布以及当前连续性风险,梳理关键公开里程碑。
[CO007, CO011, CO016, CO021, CO022, CO023]1.5 图表
02市场分析
2.1 市场边界、纳入支出与现状替代方案
最干净的市场定义,应先界定 GPU 即服务(GPUaaS)包含什么,再收窄到 Andromeda 的实际赛道。The Business Research Company 将 GPUaaS 定义为按需虚拟化 GPU 访问,用于机器学习、高性能计算和高级可视化,客户无需拥有硬件。其市场定义包括 GPU 托管、工作负载编排、资源配置,以及训练或优化服务。Andromeda 自己的表述更窄也更具体:它不只是出租自有集群,而是作为市场基础设施,对第三方供应做基准测试、标准化合同、路由工作负载,并在 AI 构建者与基础设施供应商之间运营单一支持关系。这意味着,最相关的支出池是外包 GPU 容量,加上让异构供应可用起来所需的协调层。 重要相邻市场大于 Andromeda 的直接品类。AWS、Google Cloud 和 Azure 的超大规模云厂商加速器实例是核心替代品,因为买方可以直接预留或突发使用 H100 和 H200 级集群。RunPod、Lambda、CoreWeave 和 Vast 等专业自有 GPU 云是另一组替代品;它们在没有中立经纪商居中的情况下销售按需、spot、可中断或预留容量。与电信运营商、加密矿工、MSP,以及主权云或初创云直接双边交易,也属于现状采购路径,因为这些正是 Andromeda 称其聚合的原始供应。 排除项或仅相邻支出同样重要,可约束估值纪律。购买 GPU、融资服务器、建设数据中心、租赁电力,或为内部自有集群签长期容量,不等于购买经纪式外包算力。模型 API 支出也不是直接等价物:API 推理可以替代一部分推理用例,但买不到专用集群控制权。实务结论是,应把 Andromeda 放在外包、多供应商训练与推理基础设施市场和采购流程中分析,而不是整个 AI 资本开支浪潮。[CM001, CM002, CM003, CM004, CM005, CM006]
| 细分 / 类别 | 纳入支出 | 排除支出 | 买方 / 付款方 | 相关性 |
|---|---|---|---|---|
| 经纪式多供应商 GPU 算力 | 跨第三方供应商的外包训练或推理集群、工作负载路由、基准测试、标准化合同、账单、SLA 和运营支持 | 自有集群 capex、芯片采购、数据中心建设和纯模型 API 支出 | AI 实验室、创业公司、平台团队,以及集中式云 / 采购预算 | Andromeda 试图占据的核心类别 |
| 专业第一方 GPU 云 | RunPod、Lambda、CoreWeave、Vast 等运营商直接销售的按需、竞价、可中断或预留容量 | 中立经纪撮合经济性和跨供应商标准化 | ML 工程师、基础设施负责人,以及直接支付云账单的训练团队 | 主要替代品集合,也是最直接的清算价基准 |
| 超大规模云加速器实例 | AWS P5、Google A3、Azure ND H100/H200 及相关集群工具 | 跨供应商经纪、第三方裸金属供给和中立市场路由 | 企业云预算、受监管 HPC 团队和战略账户 | 现状下的高端替代品和信任基准 |
| 直接双边供应商合同 | 与电信运营商、加密矿工、MSP、主权云或创业云签订的专用协议 | 市场聚合、单一发票整合和通用基准 | 愿意谈判定制交易的基础设施买方和采购团队 | Andromeda 声称要标准化和聚合的供给路径 |
| 自有或融资建设的集群 | GPU、服务器、网络、电力和设施支出,用于内部控制容量 | 任何外包市场服务层 | 使用 capex 或承诺租赁的 CFO、财务和基础设施团队 | 重要邻近领域,但不在 Andromeda 直接服务的市场内 |
| 模型 API 或应用支出 | 代币化模型访问、应用层 AI 订阅,以及构建在算力之上的软件工具 | 专用 GPU 预留和裸金属集群控制 | 产品、应用或实验预算 | 可替代部分推理任务,但位于经纪式 GPU 市场边界之外 |
纳入的支出仅限外包 GPU 容量及其周边做市服务;剔除的支出包括自有基础设施、芯片资本开支,以及不应并入 Andromeda 直接市场的模型 API 消费。
[CM001, CM002, CM003, CM004, CM005, CM006]2.2 受证据约束的规模测算视角与估计区间
自上而下的品类视角很直接,即便公司特定切片并不直接。The Business Research Company 估计,全球 GPUaaS 收入将从 2025 年的 $5.8B 增至 2026 年的 $7.39B,并在 2030 年达到 $19.34B,CAGR 为 27.2%。S&P Global 的 451 Research 买方洞察页面进一步说明这已经是一个真实品类:其 GPUaaS 市场监测器追踪全球 22 家供应商的收入和增长预期。这是宽口径 TAM 背景,但不能直接称为 Andromeda 的可服务市场,因为其中混合了超大规模云厂商、专业云和托管服务层,经济性差异很大。 更有用的市场视角来自 2026 年观察到的价格卡。官方定价和市场汇总显示,H100 等价容量在很宽区间内出清:Presenc 引用 2026 年 Q2 spot 约 $1.20-$2.00 / 小时,按需约 $1.80-$3.50;RunPod H100 标价 $2.89-$3.29;Lambda H100 标价 $3.99;CoreWeave 公开 HGX H100 节点定价折算约 $6.16 / H100-hour 按需;Google Cloud 公开 A3 High 和 A3 Mega 价格卡折算折扣前约 $11.06-$11.68 / H100-hour。Cyfuture 的市场文章保留了另一端极值,称部分高端 超大规模云厂商层级仍达到约 $14.19 / 小时。这不是应被平滑掉的噪音;它证明市场正按工作负载关键性、支持水平、配额访问和合同结构分层,而不是收敛为一个商品价格。 最窄的公开 Andromeda 相关视角只是下限,不是真正的 SAM。DCD 报道称,Andromeda 集群可以在数小时内提供最多 2,000 块 H100,价格为每 GPU-hour $2.40-$3.00。若这条公开报价全年满载,代表约 $42.0M-$52.6M 的年化原始算力支出。这个数字远不能作为 Andromeda 的 TAM,但可作为可见公开市场流动性底线。公开证据仍无法切分更广泛 GPUaaS 市场中有多少既是多供应商,又对信任敏感到需要经纪商;也无法判断 Andromeda 能把其中多少支出捕获为收入,而非转手 GMV。正确结论因此受证据约束:品类很大且在增长,但精确 SAM 和 SOM 仍披露不足。[CM007, CM008, CM012, CM013, CM014, CM015]
| 发布方 / 视角 | 年份 | 地域 | 数值 | CAGR / 价格区间 | 方法 | 置信度 | 局限 |
|---|---|---|---|---|---|---|---|
| The Business Research Company 自上而下 TAM | 2026 | 全球 | $7.39B 收入 | 到 2030 年 CAGR 27.2%($19.34B) | 宽口径 GPUaaS 收入模型,覆盖解决方案、服务、部署模式、应用和终端行业 | 中 | 口径太宽,不能等同于 Andromeda 经纪业务专属 SAM;混入了超大规模云厂商、专业云和服务 |
| S&P Global / 451 Research 供应商集合视角 | 2025-2026 | 全球 | 跟踪 22 家供应商 | 未披露公开金额 | 买方洞察页面提到一套 GPUaaS 市场监测,基于 22 家供应商的财务和市场数据搭建 | 中 | 能验证品类成熟度,但未发布公开美元口径市场规模 |
| Andromeda 可见供给下限 | 2025-2026 | 未披露 | $42.0M-$52.6M 年化原始支出 | 每 GPU-hour $2.40-$3.00 | 按 DCD 引述的 2,000 块可用 H100 乘以每年 8,760 小时估算 | 中 | 仅为下限;假设满利用率,并把一条公开报价视为代表性样本 |
| 专业云 H100 价格区间 | 2026 | 以北美为主 / 线上全球访问 | 每 GPU-hour $2.89-$6.16 | RunPod $2.89-$3.29;Lambda $3.99;CoreWeave 节点等价约 $6.16 | 对比官方价格卡中的自助或公开费率 | 中 | 混合了单 GPU 与整节点报价,也混合了不同支持和网络套餐 |
| 超大规模云厂商 H100 价格区间 | 2026 | 公有云区域 | GCP A3 High / A3 Mega 每 GPU-hour 约 $11.06-$11.68 | Azure 和 AWS 公开文档披露规模与配置,而非可保留对比的标价等价口径 | 用 Google 发布的 8-GPU 节点价格除以八,得到可比单位经济性 | 中 | 企业折扣、配额和区域条款不公开,因此完全同口径比较仍不完美 |
| 通缩与碎片化视角 | 2024-2026 | 全球云市场 | H100 从稀缺期约 $8-$10,降到现货 / 按需约 $1.20-$3.50;高端层级仍可能触及约 $14.19 | 活跃价差很宽,而不是单一出清价 | 市场分析文章对比高峰短缺价格与当前现货、新云厂商和高端超大规模云厂商层级 | 中 | 供应商撰写和独立汇总能提供方向性参考,但不是标准化交易数据库 |
本表刻意采用多视角,而非单一 TAM 汇总;各行混合了宽口径品类收入、下限年化支出和每 GPU-hour 价格区间,因为公开证据不足以支撑干净的 Andromeda 专属 SAM 或 SOM。
[CM007, CM008, CM012, CM013, CM014, CM015]从广义 GPUaaS TAM 到 Andromeda 对外营销的 H100 供给可见年化流动性下限,搭出一组保守规模栈。
三层都是以十亿美元计的年化支出或收入机会,但代表不同视角:自上而下的品类 TAM、单个大型买方预算能力,以及可见开放市场供给。最后一层刻意保守,并不声称是 Andromeda 的完整 SAM。
[CM007, CM013, CM029, CM038]2026 年,H100 价格在 spot、专业云和超大规模云等价报价之间仍大幅分散;这是市场核心事实,不是可以抹平的异常值。
各行统一为 $/GPU-hour。CoreWeave 和 Google 数值来自公开 8-GPU 节点价格再除以 8,因此图中按相同单位比较的经济性并不完美,但透明。
[CM012, CM014, CM015, CM017, CM018, CM021]2.3 买方、用户、付款方与采用路径
公开证据指向几个不同需求分段,而不是一个泛泛的「AI 公司」买方。前沿实验室和训练密集型模型构建者是最清晰的第一类:Andromeda 自己的需求表要求潜在客户填写 GPU 类型、集群规模、开始日期和所需周数;SiliconANGLE 报道称,部分当前目标客户每年在基础设施上花费 $250M-$500M。这与开发者刷卡买一块 GPU 完全不是同一种采购动作。第二类是风投支持的 AI 初创公司和扩张期公司,需要快速获得 H100 或 H200 级容量,不想等待超大规模云厂商分配或签僵硬的多年合约。第三类是企业 AI 平台和受监管 HPC 团队;它们可以直接使用 AWS、Azure 或 Google,但当配额、交付期或成本形成约束时,仍会在意溢出容量、地域多样性或替代供应商。 在这些分段中,用户、买方和付款方并不是同一个人。可能的终端用户是研究员、训练工程师、推理工程师和平台团队。买方通常是基础设施或平台负责人,决定一个工作负载应跑在 超大规模云厂商、专业云、自有集群还是经纪式供应上。真正为有意义承诺付款的一端更集中:云财务负责人、采购,或 CFO 支持的基础设施预算会批准预留或长周期容量。Andromeda 表单间接体现了这种分工:它询问时间、周期、最低可预订容量和价格;S&P 对企业 AI 决策者和买方标准的表述也呼应这一点。更小或容错性更高的工作负载走另一条路径,通常通过 Vast 或较低层级 RunPod 库存等 spot 或可中断市场出清。 Andromeda 的供应侧也分层。公司称,容量存在于电信和加密挖矿数据中心、传统 MSP、主权云、初创云,甚至其他实验室的资产负债表上。其供应商流程显示了为何这重要:卖方必须披露网络、存储、地点、最低周期和价格,这意味着这个市场实质上是在技术匹配、合同形态和时间跨度之间做匹配。因此,采用路径更接近基础设施采购,而不是软件自助服务:定义工作负载,指定 GPU 和周期,对候选集群做基准测试,确认合同与运营要求,在合适底座上部署,然后才考虑续约或多供应商扩张。[CM010, CM011, CM024, CM025, CM029, CM030]
| 细分市场 | 用户 | 买方 | 付款方 | 工作流 | 预算负责人 | 采用触发因素 |
|---|---|---|---|---|---|---|
| 前沿实验室 / 大模型构建方 | 研究员、训练工程师和集群运维团队 | 算力负责人或平台负责人 | 中央基础设施或云预算 | 大型预留训练集群,对可用性和网络要求严格 | 基础设施负责人,以及采购 / 财务 | 关键上线训练需求,或拿不到足够超大规模云厂商配额 |
| VC 支持的 AI 初创公司与规模化公司 | 创始工程师、ML 工程师和平台团队 | CTO 或基础设施负责人 | 风险投资支持的算力预算 | 快速获得集群,用于训练、微调或早期生产推理 | CTO 与财务负责人 | 需要速度、灵活性,以及比超大规模云厂商偏好的更短承诺期 |
| 企业 AI 平台 / 受监管 HPC 团队 | 内部 ML 平台、分析、制药、气象或金融建模团队 | 企业云或基础设施经理 | IT、业务线或采购预算 | 在既有云之外采购溢出、多元化或专业集群 | 中央 IT / 采购 | 配额摩擦、地域、合规或性价比压力 |
| 批量推理 / 实验团队 | 推理工程师、数据科学家,或运行可检查点任务的开发者 | 团队级工程经理 | 职能云预算或银行卡支付支出 | 可容错的突发工作负载,可使用现货或可中断容量 | 应用或实验预算负责人 | 降低单位成本比最大化可用性更重要 |
| 供给侧容量持有者 | 数据中心运营和基础设施团队 | 容量销售或合作负责人 | 资产所有者或运营公司 | 上架容量、披露技术参数、设定最低条款,并寻找利用率 | 资产所有者、运营方或资金管理职能 | GPU 利用不足、容量闲置,或希望更快确认收入 |
预算负责人字段是从 Andromeda 入驻表单、S&P 买方框架和替代供应商合同结构推断的公开代理;留存来源没有给出 Andromeda 逐客户采购组织结构图。
[CM010, CM011, CM024, CM025, CM029, CM030]GPU 容量的使用者通常不是付款方;Andromeda 的市场处在技术工作负载所有者与集中化基础设施预算的交接处。
买方和付款方字段根据 Andromeda 准入表单中的采购信号和替代供应商合同结构推断;它们并非来自已发布的 Andromeda 客户参考名单。
[CM010, CM011, CM024, CM029, CM030, CM031]买方旅程在信任和采购闸门处收窄最明显,而不是在最初意识到 GPU 需求时。
数值是方向性的指数权重,不是观测到的转化率。漏斗顺序有证据支撑,但公开来源没有披露各阶段胜率。
[CM010, CM011, CM026, CM027, CM032, CM036]2.4 增长驱动、采用约束与估值相关性
主要增长驱动清晰可见,且相互强化。S&P 称,在过去两年大部分时间里,AI 技术和云基础设施一直位居企业技术支出意向最高的类别;其买方洞察笔记称,专业 GPUaaS 供应增长,是因为许多企业买不到足够芯片来运营自有系统。Andromeda 自己的市场文章补充了时间错配:AI 研究周期以周计,而数据中心建设需要 12 到 24 个月。RunPod、Lambda、CoreWeave、Vast 和超大规模云厂商 价格卡带来更多供应和更高价格透明度,也进一步扩大买方集合,使溢出、实验和中端市场推理在经济上变得可行;这在 2023-2024 年短缺高峰期并不成立。 制动力不只是「短缺」,而是结构性的。Andromeda 的核心论点是,原始供应存在,但不可互换。网络、存储、集群健康度、地理位置、可预订周期和支持质量,都会实质改变一块名义相同 H100 的价值。CoreWeave 的基准测试简报和 NVIDIA 的 H100 架构页面说明了原因:真实训练价值取决于 MFU、可靠性、NVLink、InfiniBand 和周边软件栈,而不只是芯片名称。SesameDisk 还补充了买方现实:当发布日期很硬时,配额、区域限制、排队时间和账号状态往往比标价更重要。这有助于解释,为什么 超大规模云厂商等价价格能长期高于 neo-cloud spot 底价,而市场不会立刻把价差套利掉。 供应缓解也仍是局部的。Presenc 认为,瓶颈已从 H100 绝对稀缺转向 HBM3e 内存、CoWoS 封装、机柜电力和数据中心部署。这些最后一公里摩擦对 Andromeda 很重要,因为它们提高了已认证第三方供应的价值;但它们也抬高了信任门槛:买方需要证明集群能按规格运行并保持在线。估值含义因此是双向的。正面看,碎片化、不可互换、受信任约束的市场,正是中立经纪商能创造价值的地方。负面看,单位价格下行和难以横向比较的价格卡,使外部很难仅凭公开数据推断持久抽佣率。关键尽调保留意见 不是市场是否存在,而是 Andromeda 能否在单位价格通缩压缩价差前,从协调、认证和运营中捕获足够多的经常性经济性。[CM009, CM016, CM021, CM022, CM023, CM026]
| 驱动 / 约束 | 方向 | 时点 | 影响 | 尽调问题 |
|---|---|---|---|---|
| 企业 AI 与云支出意向 | 驱动 | 当前 | 即便单位价格下跌,也能支撑外包 GPU 容量需求 | 验证 Andromeda 是否真正拿到了这波支出意向中的预算,还是只受益于市场整体热度 |
| 企业无法直接买到足够芯片或服务器 | 驱动 | 当前 | 把买方推向 GPUaaS 专业供应商和经纪商,而不是自建集群 | 测试 Andromeda 管线中有多少来自超大规模云厂商或 OEM 渠道配额失败 |
| 更多专业供给与价格透明度 | 驱动 | 当前至近期 | 扩大可触达工作负载,尤其是溢出训练和成本敏感型推理 | 确认价格透明度是提高 Andromeda 转化率,还是只是压窄抽成价差 |
| 可靠性与基准测试溢价 | 驱动 | 当前 | 客户愿意为通过认证、支持更重的供给付费,而不只追求最低标价 | 要求提供证据,证明客户选择 Andromeda 是因为可用性 / 性能,而不只是因为能拿到容量 |
| 集群不可互换与架构差异 | 约束 | 结构性 | 市场难以商品化,也拖慢全自动匹配 | 核查 Andromeda 如何对互连结构、存储和性能做基准测试,并与买方真正关心的指标对齐 |
| 信任、安全、SLA 与采购负担 | 约束 | 结构性 | 拉长采用周期;即使成本更高,一些买方仍留在超大规模云厂商或自有集群默认方案 | 要求提供企业安全审查、合同周期长度和续约驱动因素案例 |
| HBM、CoWoS、机柜供电与数据中心建设瓶颈 | 约束 | 当前至中期 | 供给缓解并不完整,且受地点限制,容量可得性和价格仍会波动 | 追问 Andromeda 供给中有多少已经签约、通电并完成基准测试,多少仍停留在理论层面 |
| 单位价格通缩与不可比价格卡 | 约束 | 当前 | 可能扩大需求,同时压缩经纪商经济性,并模糊毛利率基准 | 取得管理层层面的披露:抽成率、透传 GMV、服务组合,以及跨供应商如何标准化定价 |
本表把需求驱动与采用约束放在一起,因为两者共同决定估值相关性;最关键的开放问题是,原始算力价格下跌时,Andromeda 是否能从信任和协调中抓住可持续经济性。
[CM009, CM016, CM021, CM023, CM026, CM027]2.5 图表
03竞争格局
3.1 格局与替代方案类别
Andromeda 的买方并不是在一张整齐的 GPU 初创公司清单中选择。真实选择集合至少分成五类。第一类是 CoreWeave 和 Lambda 这类供应商自有 AI 云,销售自己的容量、合同和支持栈。Vast.ai 在预算重叠上接近这一类,尽管结构不同:它本身也是市场,但让价格形成和许多商业条款更贴近底层主机。第二类是存量 超大规模云厂商——AWS、Google Cloud 和 Azure。对许多企业而言,它们仍是默认现状,因为它们把加速器库存、经审计的合规计划和既有预算关系打包在一起。第三类是相邻的编排和弹性产品,如 NVIDIA Run:ai、RunPod 和 Modal;它们为部分团队去掉足够多运营摩擦,使其完全避开经纪式采购层。第四类是使用 NVIDIA DGX SuperPOD 等系统做内部建设或 on-prem 部署。第五类是与供应商直接双边签约,也就是 Andromeda 称其通过基准测试、标准条款和一张发票来简化的经纪商出现前现状。 这种框架重要,因为 Andromeda 最强的公开差异化不是绝对价格领先,也不是独家拥有供应。它的主页和供应商流程描述的是一个中立市场层,跨 100+ 供应商和 1,000 多笔已完成交易采购、基准测试、认证并标准化容量。当买方重视供应商发现、资质验证和商业简化,超过忠于任何单一云时,这一点最强。当买方已经信任某一家供应商、能在几分钟内自助拿到所需 GPU,或规模足以支撑直接合同或内部建设时,这一点最弱。因此,潜在进入者集合与具名对手同样重要:任何供应商只要获得更好的编排、更好的合规包装或更好的销售覆盖,都能向复制经纪层某一部分再迈近一步。[CP001, CP002, CP003, CP004, CP009, CP012]
| 竞争对手 / 类别 | 品类 | 规模 / 融资信号 | 目标客户 | 差异化 | 局限 |
|---|---|---|---|---|---|
| Andromeda | 中立经纪商 / 市场层 | 100+ 家供应商;1,000+ 笔交易;公开信任资料有限 | 需要跨多个供应商获取外部 GPU 容量的 AI 实验室、初创公司、企业和平台团队 | 寻找、基准测试、认证并标准化第三方供给,提供一张发票和一个支持渠道 | 未公开展示超大规模云厂商级合规深度或透明标价 |
| CoreWeave | 直接 AI 原生云 | Nasdaq 上市公司;收入积压订单近 $100B;2026 年 Q1 在用电力 >1 GW | 希望从单一供应商获得大型专用 AI 基础设施的前沿实验室、企业和客户 | 全栈 AI 云,提供存储、可观测性、安全、透明按需或现货定价,以及承诺容量选项 | 客户集中、杠杆和租约错配风险;不是中立多供应商经纪商 |
| Lambda Cloud | 直接 AI 云 | 100,000+ 云注册;5,000+ 硬件或私有云客户;2025 年 Series D 后累计股权融资 $863M | 寻求自助服务加大型专用集群的 AI 开发者、训练团队和企业 | 清晰公开定价、自助实例、1-Click 集群,以及 SOC 2 Type II 信任姿态 | 主要销售自有容量,而不是聚合碎片化外部供给 |
| Vast.ai | 市场型替代方案 | 供需交易所上有 20,000+ 块 GPU、40+ 个数据中心、68+ 种 GPU 类型 | 价格敏感的研究员、初创公司和溢出买方,愿意自行管理更高波动 | 实时市场定价、无长期合约、主机多样性广 | 托管采购、支持和企业信任深度低于最强的直接云或超大规模云厂商 |
| SambaNova | 邻近既有厂商 / 推理栈 | 2026 年 Series E 超 $350M;与 Intel 战略合作;主权和 SoftBank 相关部署 | 更看重推理经济性、而非中立供给聚合的企业、主权项目和服务商 | 垂直整合推理平台,讲 GPU 替代叙事,主权角度很强 | 不太像通用外包多供应商训练采购的直接经纪替代 |
| NVIDIA Run:ai 编排平台 | 编排邻近方案 | NVIDIA 旗下企业软件平台,支持云托管和混合部署路径 | 在公有云、私有云、混合或本地环境管理 GPU 的企业 | 动态 GPU 编排、策略控制和异构资源池化 | 不直接供应算力,归属 NVIDIA 后中立性下降 |
| RunPod / Modal | 无服务器与弹性云邻近方案 | RunPod 称有 1M+ 开发者;Modal 宣传可从 0 即时自动扩展到 1000+ 块 GPU | 看重速度和按秒计费经济性的开发者、推理团队和溢出买方 | 低摩擦部署、透明或用量定价,以及强突发弹性 | 不太适合经纪式采购、长期多供应商寻源,以及最重监管账户 |
| AWS / Azure / GCP | 既有默认方案 | 全球云规模;本组最深的公开合规目录 | 已通过既有云账户采购的企业、政府和受监管团队 | 既有采购路径、已审计合规广度,以及直接访问加速器实例 | 不中立,且可能价格高、容量受限,或设计上就是单供应商 |
| 内部建设 / DGX SuperPOD | 内部替代方案 | 可扩展到数万块 GPU 的交钥匙 AI 数据中心路径 | 资金充足、需求持续且有平台人才的实验室、主权项目和企业 | 部署后控制力最高,且无中介抽成 | 资本开支重、部署复杂,对小买方来说见效更慢 |
本表覆盖留存来源能证明的主要直接、邻近、既有和替代路径;它是 2026 年买方格局选样,不是所有 GPU 云的完整普查。
[CP001, CP003, CP004, CP006, CP009, CP011]按多供应商抽象深度,以及直接基础设施控制或信任强度,对主要替代方案作序数映射。
坐标轴是基于留存的产品、定价、信任和监管页面综合得出的序数判断,而不是来自第三方市场份额数据。
[CP003, CP004, CP009, CP012, CP014, CP016]3.2 直接、相邻与替代定位
当买方想要大型企业 AI 云,而不是中立经纪商时,CoreWeave 是最危险的规模化直接替代方案。其公开材料把 AI 原生 基础设施、Kubernetes 访问、存储、可观测性、专用推理,以及透明的按需、spot 和预留定价组合在一起。Lambda 的重叠方式不同:它提供更清晰的自助经济性和更快的首次触达,公开实例定价和 1-Click clusters 可从数十块扩展到 2,000+ 块 GPU,同时还宣传 SOC 2 Type II 和单租户选项。Vast.ai 是最清晰的价格底线竞争者,因为它宣传一个覆盖 20,000+ 块 GPU 和 40+ 个数据中心的实时供需市场,且没有长期合约。SambaNova 与其中立市场对手属性较弱,更像面向推理密集型企业、主权部署,以及愿意采用垂直整合平台加 GPU 替代叙事客户的存量式替代栈。 相邻集合通过只移除 Andromeda 解决痛点的一部分来竞争。NVIDIA Run:ai 争夺的是编排心智,而不是原始容量所有权:它把公共、私有、混合和 on-prem 环境中的资源池化,现在也被更紧地打进 NVIDIA 软件栈。RunPod 和 Modal 攻击市场中低摩擦的一端,采用按用量、开发者优先的流程。RunPod 提供 pods、serverless 和 clusters,并公开按小时和按秒经济性;Modal 则用 serverless 按秒计费和初创公司抵扣额度,在云和区域之间路由工作负载。这些替代方案并不能完美替代大型专用集群的经纪式采购,但对溢出、实验、推理,以及更看重即时弹性而非托管式多供应商采购流程的团队而言,它们是可信替代品。[CP004, CP005, CP006, CP009, CP011, CP012]
| 能力 | Andromeda | CoreWeave | Lambda | Vast.ai | SambaNova | Run:ai | RunPod / Modal | 超大规模云厂商 |
|---|---|---|---|---|---|---|---|---|
| 中立多供应商寻源 | 强 | 有限 | 有限 | 中等 | 有限 | 中等 | 有限 | 有限 |
| 跨第三方供给的标准化合同和一张发票 | 强 | 有限 | 有限 | 有限 | 有限 | 否 | 否 | 否 |
| 专用多千卡 GPU 集群 | 中等 | 强 | 强 | 中等 | 强 | 否 | 中等 | 强 |
| 公开标价 | 有限 | 强 | 强 | 强 | 有限 | 有限 | 强 | 有限 |
| 合规 / 信任文档深度 | 有限 | 中等-强 | 中等-强 | 有限 | 中等 | 中等 | 中等 | 强 |
| 混合 / 本地编排 | 有限 | 中等 | 有限 | 有限 | 有限 | 强 | 有限 | 中等 |
| 小型或突发工作负载的自助速度 | 有限 | 中等 | 强 | 强 | 有限 | 有限 | 强 | 中等 |
| 证据最充分的用例 | 多供应商采购与资质审核 | 大型专用 AI 云部署 | 快速直连访问加企业集群 | 最低成本灵活市场容量 | 生产推理与主权技术栈 | 全资源池 GPU 效率与策略控制 | 弹性推理与开发者溢出 | 可信既有采购路径 |
单元格只反映留存产品、文档、定价、信任或监管来源明确写出的能力;没有支撑的边界标为有限,而不推断为完整强项。
[CP001, CP004, CP005, CP009, CP012, CP014]能力图显示,市场被切分为中立经纪、自有容量云、编排层,以及弹性无服务器替代方案。
[CP001, CP005, CP009, CP012, CP014, CP016]3.3 定价、信任与切换动态
公开定价透明度,是这个市场可能比 Andromeda 估值叙事更快商品化的最清晰原因之一。CoreWeave 发布了大量按需和 spot 价格卡,并为承诺用量提供最高 60% 的预留折扣。Lambda 发布直接的每小时费率和自助实例访问。Vast.ai、RunPod 和 Modal 都披露透明或按用量计费的经济性,让买方能快速对标替代方案;Vast 还明确宣传没有长期合约。Andromeda 自己的公开材料更强调流程、认证和统一支持,而不是公布标价。这本身并不说明模式弱——经纪式基础设施本就应靠更好的总结果而非原始标价来销售——但确实意味着,买方可以很快锚定直接云或 serverless 替代方案,并要求 Andromeda 用速度、质量和合同简化来证明任何价差。 信任和切换成本则朝另一个方向作用。CoreWeave、Lambda、Modal、RunPod,尤其是 AWS、Google Cloud 和 Azure,都比 Andromeda 当前稀疏的信任页面公开了多得多的安全与合规控制细节。超大规模云厂商的信任姿态明显最强,因为它们发布广泛的合规目录、第三方证明和受监管行业覆盖。多宿主在技术和商业上也可行。Run:ai 为编排异构环境而生;Modal 跨云路由;Vast、RunPod 和 Lambda 依赖短周期或按用量结构;最大买方仍可选择内部建设。结果是,这个市场存在有意义但并非不可逾越的锁定。只有当客户把 Andromeda 深度嵌入供应商资质验证、SLA 标准化和持续多供应商运营时,其切换成本才可能真实存在;若客户主要只是从一个已知供应商租一个集群,切换成本就弱得多。[CP002, CP005, CP009, CP010, CP013, CP019]
| 供应商 / 类别 | 公开定价信号 | 合同模式 | 包含能力 | 未知项 / 折扣 | 含义 |
|---|---|---|---|---|---|
| Andromeda | 留存来源未见公开标价 | 经纪式项目或企业交易,采用标准化条款和统一账单 | 供应商发现、基准测试、认证、路由、合同、支持 | 实际抽成率、折扣和成本转嫁口径仍未公开 | 必须靠速度、质量和流程简化支撑利润,而不是靠标价透明度 |
| CoreWeave | 按需和 spot 价格披露细,预留折扣最高 60% | 按量使用、预留容量和专用推理产品 | 全栈 AI 云、存储、可观测性、安全和专家服务 | 企业议价条款仍然重要;部分专用方案仍靠销售推进 | 大型集群买家的强参照,也直接威胁经纪价差 |
| Lambda | 透明的按小时实例价格和公开的 1-Click 集群卡片 | 自助先到先得实例,加上专用集群和私有云 | 实例、集群、超级集群、信任控制和直接云访问 | 企业折扣和非公开私有云价格仍不透明 | 能快速标准化到单一供应商的买家,搜寻成本被压低 |
| Vast.ai | 实时供需定价,覆盖按需、可中断和预留选项 | 市场式按量使用,不要求长期合同 | 搜索、API 驱动供给,以及跨主机的广泛 GPU 库存 | 可靠性和服务质量更受主机基础影响,波动大于全托管云 | 为非监管或灵活工作负载压出透明的商品化底价 |
| RunPod | 公开按小时 GPU 卡片,加上按秒计费的 serverless | Pods、serverless workers 和预留集群式选项 | 开发、推理和集群打包在同一自助入口下 | 企业支持和最重的合规打包不如超大规模云厂商可见 | 对突发推理或快速实验是强替代 |
| Modal | 按秒 serverless 计费、免费额度,以及明确的创业公司或研究资助 | Serverless 执行,向企业计划上售 | 跨云路由、自动扩缩、训练、推理和批处理执行 | 专用长周期集群经济性不是公开叙事核心 | 适合弹性工作负载,对经纪式多供应商采购的替代性较弱 |
| AWS / Azure / GCP | 公开标价卡与协商后的承诺用量经济性并存 | 既有云合同、企业协议和预留消费计划 | 原生加速器实例、网络、存储和经审计的合规控制 | 实际成交价会随承诺量和既有账户议价能力大幅变化 | 采购挂在既有预算上;即便原始标价更高,替换成本也高 |
| 自建 / DGX SuperPOD | 没有简单的公开用量价格;经济性更像资本开支加运营,而不是云端运营开支 | 自有或托管 AI 数据中心部署 | 全栈计算、网络、存储和工作负载管理环境 | 真实总成本取决于利用率、部署速度、人员配置和融资结构 | 只有持续需求足够大、能摊薄复杂度时才可信 |
直接云和 serverless 平台的公开价格透明度远高于 Andromeda 的经纪模式;承销关键是 Andromeda 的工作流价值能否抵消这道可见性缺口。
[CP005, CP010, CP013, CP019, CP020, CP021]3.4 护城河耐久性与反向证据
Andromeda 故事中可持久的部分,不是它拥有最多 GPU 或最广合规目录,而是外包 GPU 市场足够碎片化,中立协调层有存在价值。如果买方确实需要供应商发现、性能验证、标准化合同、一张发票,以及跨许多独立供应商的统一运营界面,那么 Andromeda 占据了一个真实切入口;单一 neo-cloud 或单一超大规模云厂商都很难干净复制。这个切入口 应在溢出采购、多供应商风险管理,以及不愿过早押注单一直接供应商的买方中最强。 问题在于,公开证据也显示这个切入口可能很快变薄。CoreWeave 的规模和积压订单证明,直接供应商可以建立强大的企业分销和基础设施控制优势;但 Fitch 的分析也显示,这些对手自身承担客户集中度、杠杆和租约错配风险。NVIDIA 对 Run:ai 的所有权,以及 European Commission 认为 Run:ai 目前并非主导性编排层,提醒我们:编排很可能向上打包进更大的基础设施栈,而不是永久作为稀缺独立层存在。同时,Lambda、Vast、RunPod、Modal 和 CoreWeave 的公开定价,为市场大部分区域设置了可见商品参照点。因此,竞争结论是有条件的:Andromeda 在中立聚合和采购简化上强于直接云,但在披露的信任深度、上游供应控制和直接预算所有权上更弱。只有当认证、路由和商业标准化转化为比客户直接采购更好的经济性或更快采购速度时,它的护城河才耐久。[CP002, CP006, CP007, CP008, CP015, CP017]
| 护城河主张 | 竞争威胁 | 严重性 | 证据 | 缓释措施 / 尽调问题 |
|---|---|---|---|---|
| 中立聚合与供应商发现 | 直接云和市场把供应商搜索做得更容易,也能绕过经纪商直接销售 | 高 | Lambda、Vast、RunPod 和 CoreWeave 都提供直接购买路径或公开价格锚 | 索取赢单 / 输单数据,证明 Andromeda 在速度、质量或总成本上能击败已知直接云 |
| 认证与集群资格评估 | 上游供应商可以加入自己的基准测试、支持或专用推理打包 | 高 | CoreWeave 和 Lambda 都主打集成平台控制;超大规模云厂商和 NVIDIA 也在持续加软件层 | 要求证明 Andromeda 的认证能实质提升部署成功率或缩短采购时间 |
| 一张发票、一条支持路径 | 通过 Run:ai、Modal 或直接云做多宿主,切换成本仍然适中 | 中 | Run:ai 编排混合算力池,Modal 跨云路由;按量云避免长期承诺 | 衡量账户渗透率、续约率,以及仍在其他地方直接采购的客户数量 |
| 供应商广度作为护城河 | 供应商自有云和超大规模云厂商控制上游容量,并可向前整合商业化 | 高 | CoreWeave 的规模、超大规模云厂商的信任基础和直接自有容量,都会在部分交易中削弱中介必要性 | 梳理头部供应商集中度、排他性,以及供应商是否也在直接争夺同一需求 |
| 信任与监管就绪度 | 超大规模云厂商、Lambda、Modal、RunPod 和 CoreWeave 发布的信任或合规材料比 Andromeda 更详细 | 高 | AWS、Google Cloud 和 Azure 在合规广度上领先;几家较新的对手也拿出比 Andromeda 稀疏信任页更清晰的审计信任材料 | 索取客户安全问卷、受监管账户参考,并证明赢下合规导向交易靠的是合规而非价格 |
| 耐久的独立编排层 | 编排被打包进上游技术栈后,未必还能作为独立护城河品类存在 | 中 | NVIDIA 现在拥有 Run:ai;欧洲委员会则称 Run:ai 今天本身并不具备显著市场地位 | 询问 Andromeda 控制层功能相对供应商或 NVIDIA 软件可打包能力,是否有实质差异 |
严重性衡量的是 Andromeda 定价权和留存风险,而不是外包 GPU 需求本身;核心问题是,直接云成熟后,市场碎片化带来的价值还能不能留住。
[CP002, CP007, CP008, CP017, CP018, CP023]压缩记分卡,展示当前帮助或削弱 Andromeda 防御性的竞争特征。
数值是基于留存来源证据综合出的定性判断,不是披露的基准分数或管理层提供的胜率数据。
[CP002, CP017, CP023, CP029, CP032, CP033]3.5 图表
04财务
4.1 收入模式、变现信号与公开牵引
公开证据足以描述 Andromeda 的变现表层,但不足以把这个表层转成清晰的收入质量模型。官方流程文案持续显示一个双边 GPU 采购层:买方说明 GPU 类型、数量、时间线和周期;供应商披露容量、最低可预订规模、周期和条款;Andromeda 对供应做基准测试、标准化合同、路由需求,然后提供一张发票加一个支持渠道。这让公司的经济形态更接近高接触度基础设施经纪商或市场层,而不是拥有大部分所售硬件的传统云。同一份材料也说明,收入确认是核心尽调问题。坐在买方和供应商之间的公司,可以按总额或净额确认经济性,取决于合同控制权;保留的公开来源没有说明 Andromeda 采用哪种政策。 牵引比会计处理更清楚。官方页面称有 100+ 供应商、1,000+ 笔已完成交易、跨众多供应商的数十个集群,以及超过 10 亿 GPU-hour 活动量。独立报道给出了最强收入锚点:Upstarts 和 SiliconANGLE 都把 2025 年收入 run-rate 放在约 $100M,2024 年收入超过 $50M,并称自推出以来盈利。Data Center Dynamics 还保留了本轮检索中唯一的公开 Andromeda 价格线索:面向非 NFDG 买方,最多 2,000 块 H100,价格为每 GPU-hour $2.40-$3.00。这些事实支持真实商业规模,但核心承保问题仍未解决:Andromeda 报告的收入,主要反映的是净经纪经济性,还是大得多的总额 转手基数。[CI001, CI002, CI003, CI004, CI005, CI006]
| 收入流 / 变现触点 | 公开证据 | 可能单位 | 公开状态 | 收入质量备注 | 尽调问题 |
|---|---|---|---|---|---|
| 经纪式承诺集群采购 | 买方和供应商表单以及官方工作流都指向跨第三方供应商的预留或排期容量采购。 | 合约 GPU 小时、集群预留或最低消费承诺 | 有公开依据 | 如果按净额确认且按账户续约,可能是有吸引力的经常性收入。 | 提供承诺采购与其他全部活动之间的产品级收入拆分。 |
| 紧急 / 溢出容量采购 | DCD 保留了一段公开报价:最多 2,000 个 H100 可在数小时内交付,价格为每 GPU 小时 $2.40-$3.00。 | GPU 小时或短期预留 | 有公开依据但范围窄 | 能证明 Andromeda 可以清掉紧急需求,但单个报价看不出利润率或可重复性。 | 披露短期溢出需求相对较长期预留合同的收入占比。 |
| 多供应商计费与支持层 | 官网称买家获得一张发票、一个支持渠道和标准化合同。 | 附着在基础设施支出上的服务包 | 有公开依据 | 即便原始 GPU 价格商品化,也可能支撑溢价经济性。 | 拆出不属于纯转手计算的支持、解决方案或平台收入。 |
| 基准测试 / 价格情报 | 官方和媒体来源描述了基准测试、认证和价格指数,但没有披露单独变现。 | 潜在嵌入式价差或订阅 | 公开信息不清 | 可能被打包进计算经济性,而不是单独销售。 | 明确基准测试和价格情报是单独变现,还是嵌入抽成率。 |
| 供应商侧激活与资格评估 | 供应商工作流要求详细入驻和标准化条款,但没有公开供应商费率表。 | Unknown | 未披露 | 可能提升供应质量,但未必产生直接收入。 | 披露供应商是否支付入驻、认证或成功费。 |
| 活动的会计列报 | 没有留存来源说明 Andromeda 是按总额确认客户支出,还是只确认净经济收益。 | GMV 与已确认收入 | 未披露 | 这是本章最大的单一收入质量问题。 | 提供 ASC 606 principal-agent 政策,以及 GMV 到收入的桥接。 |
各行反映的是公开工作流和报道能支持的变现触点;收入结构、抽成率和会计列报仍属私有信息。
[CI001, CI002, CI003, CI012, CI013, CI015]| 触点 / 可比对象 | 公开价格或合同信号 | 标价与实际成交价 | 收入确认含义 | 来源 / 备注 |
|---|---|---|---|---|
| Andromeda 公开报价 | DCD 保留了每 H100 小时 $2.40-$3.00、最多 2,000 个 H100 可在数小时内交付的信息 | 单一窄口径公开报价,不是挂牌目录 | 可能代表总额计算转售经济性、经纪式预留费率,或薄利转手报价 | 本轮留存的唯一 Andromeda 公开价格;没有抽成率披露 |
| CoreWeave 直连云 | H100 按需 $49.24/小时;spot $19.71/小时;承诺用量折扣最高 60% | 挂牌标价 | 主云服务商通常直接确认基础设施收入 | 高端打包、网络和服务等级让直接比较并不完美 |
| Lambda 直连云 | H100 $3.99/GPU 小时;集群规模 16 到 2,000+ 个 GPU;预留容量通过销售协商 | 挂牌标价加协商型预留交易 | 直接供应商经济性,不是中立经纪 | 对更大的预留训练集群是有用基准 |
| RunPod 开发者与集群路径 | H100 PCIe $2.89/小时,H100 SXM $3.29/小时,覆盖 pods、serverless 和集群 | 挂牌标价 | 用量驱动的直接云和 serverless 经济性 | 最适合作为自助买家的低摩擦价格底 |
| Vast 市场 | 按需、可中断和预留条款;可中断便宜 50%+;1/3/6 个月预留;不要求长期合同 | 实时市场价格 | 市场式经济性,锁定有限,按秒计费 | 说明透明交易所式供给会如何挤压价差 |
本表有意把 Andromeda 唯一公开报价与官方竞品价格卡放在一起,用来框定变现语境,不主张它们的实际成交价可以苹果对苹果比较。
[CI013, CI017, CI018, CI019, CI020, CI021]展示买方需求、供应商供给、合同和支持如何转化为 Andromeda 收入,同时标出总额确认还是净额确认这个未解步骤。
该流程有证据支撑,但不是定量图:未解步骤在于 Andromeda 是按总额确认基础设施支出,还是只确认净经济收益。
[CI001, CI002, CI003, CI015, CI016, CI035]混合使用公开区间和固定点,锚定规模、估值、定价背景和收入密度,同时不把它伪装成经审计的管理层指引。
收入和估值项混合了 2024 年收入下限和 2025 年运行率;该图用于界定公开证据范围,而不是把不可同比的指标统一成一个预测。
[CI007, CI008, CI013, CI021, CI040, CI045]4.2 GTM 动作、价格框架与销售效率代理指标
GTM 动作看起来更像企业级顾问式销售,而不是自助服务。Andromeda 的客户表单要求买方填写 GPU 类型、数量、期望开始日期和所需周数;供应商表单要求卖方填写最低可预订 GPU 和最短周期。这是协商式基础设施采购的语言,不是低摩擦 SaaS 扩张。Upstarts 和 SiliconANGLE 描述的买方集合也支持这一判断:其中包括每年在基础设施上花费约 $250M-$500M 的 AI 公司。若这些买方具有代表性,销售周期更可能取决于集群验证、法律条款和时间线保障,而不是数字化需求生成。因此,公开证据不支持经典 CAC 或回本期计算,但支持一个收入密度代理指标:公开员工数信号约 17 到 20 人,对应报告约 $100M run-rate,意味着每名员工收入约 $5.0M-$5.9M。 定价参照集也说明了 Andromeda 的可能角色。CoreWeave、Lambda、RunPod 和 Vast 都发布明确价格或合同结构,但这些结构差异很大:CoreWeave 的 H100 按需价格远高于 RunPod 和 Lambda 标价,而 Vast 强调按秒、可中断和短期预留条款。这种离散度说明,该市场的经济价值不只是原始 GPU 访问;它还围绕合同周期、在线保障、网络、支持和验证打包。Andromeda 的公开材料符合这种解释。公司不发布价目表,但主打标准化合同、基准测试和单一运营界面。可能的结果是,当客户为采购确定性和供应商标准化付费时,Andromeda 获胜;当买方只想买最便宜的可见 H100 hour 时,它并不占优。[CI002, CI003, CI010, CI013, CI014, CI017]
| 指标 | 公开数值 / 状态 | 置信度 | 为什么重要 | 尽调问题 |
|---|---|---|---|---|
| 2025 年化收入规模 | ~$100M | 中 | 当前商业相关性的最佳公开规模锚点。 | 将披露的年化收入规模对齐到经审计确认收入和月末 ARR 等价口径。 |
| 2024 收入基数 | >$50M | 中 | 说明 2025 年规模并非从零开始;有助于框定增长速度。 | 提供 2024 全年收入和进入 2025 年的月度变化。 |
| 上线以来盈利 | 据称上线以来一直盈利 | 中 | 若属实,说明基础设施经纪商可能有强单位经济性,或 opex 异常精简。 | 按季度披露 EBITDA、营业利润和自由现金流。 |
| 买方支出代理指标 | 公开目标包括每年在计算上花费约 $250M-$500M 的买家 | 中 | 支撑大账户、企业式销售动作,也带来集中度风险。 | 提供客户支出分层和头部账户集中度。 |
| 人均收入代理指标 | ~$5.0M-$5.9M / 员工 | 中 | 显示收入密度高,但也可能反映总额列报或账户集中。 | 把年化收入规模、员工数和岗位结构对齐到真实人均净收入。 |
| 销售周期代理指标 | 起始日期和周数需预先提交;供应商披露最短期限和最低可预订 GPU 数 | 中 | 暗示这是顾问式基础设施采购,而非无摩擦自助转化。 | 分享从首次通话到签约和部署的中位周期。 |
| 毛利率 | 未公开披露 | 低 | 这是承销收入质量最核心的缺失输入。 | 按产品或合同类型提供毛利率 % 和毛利额。 |
| CAC / 回本期 | 未公开披露 | 低 | 需要用它区分高效企业销售与不稳定的创始人主导大单。 | 提供销售和营销支出、新 ARR 或 GMV cohort 数据,以及回本期分析。 |
| 抽成率 / 价差 | 未公开披露 | 低 | 没有它,年化收入规模无法转译为 GMV、贡献利润或竞争耐久度。 | 按工作负载类别披露净抽成率或经纪价差。 |
| 营运资本强度 | 未公开披露 | 低 | 回款和应付决定了一个看似轻资产的模式是否仍然耗现金。 | 提供 DSO、DPO、预付款条款,以及任何供应商担保或预留承诺。 |
这些接近空白的状态反映私营公司不透明;公开类比和估算都已清楚标注,没有被转成虚假的精确度。
[CI007, CI008, CI009, CI010, CI014, CI038]把公开可见的需求和定价信号,与真正决定利润率质量的隐含变量连接起来——抽成率、支持负担和营运资本。
该图刻意保持概念化,因为公开文件没有披露 Andromeda 的抽成率、毛利率或回款条款。
[CI010, CI013, CI021, CI038, CI039, CI046]4.3 成本结构、利润率类比与资本充足性
公开材料表面上指向轻资产模型,但不能最终证明它拥有轻资产损益表或资产负债表。由于 Andromeda 似乎采购第三方容量,而不是运营自己的大型机队,最明显的直接云成本——自有 GPU 折旧、数据中心建设和超大规模云厂商级电力采购——应在结构上低于 CoreWeave 或 Amazon 等供应商。即便如此,公司似乎仍承担真实服务交付成本:工作负载验证、基准测试、合同标准化、支持、可观测性,以及潜在的预先安排或预留容量。这点重要,因为若经纪商预付供应商、担保承诺,或在供应商义务已固定后才向客户收款,它仍可能变成重营运资本模式。保留的公开来源没有披露 Andromeda 的这些机制。 上市公司类比让风险更尖锐。DigitalOcean 文件显示,其云模型远轻于 CoreWeave,但即便 DigitalOcean 也在 2026 年 Q1 因数据中心扩张成本先于收入到来,毛利率从 61% 压缩到 56%,并且仍承担有意义的资本开支 和长期机房租约。CoreWeave 是另一端极值:文件和 Fitch 显示 take-or-pay 收入可见性,但也显示严重集中度、庞大债务结构,以及专门为客户合同 资本开支融资而设置的资产级工具。Amazon 文件则显示,直接云所有权可以把资本需求、预付款和履约义务放大到多大。在这个背景下,Andromeda 2026 年 3 月融资提高了外界对其有资本增长的信心,但公开证据仍未披露手头现金、现金消耗、续航期、债务或任何供应商侧担保。因此,无法仅凭公开证据承保前瞻资本充足性。[CI006, CI014, CI022, CI023, CI024, CI025]
| 项目 | 公开数值 / 状态 | 为什么重要 | 可比对象 / 背景 | 尽调问题 |
|---|---|---|---|---|
| 账面现金 | Andromeda 未披露 | 缺少该数据,无法判断 2026 年 3 月融资后的资产负债表韧性。 | CoreWeave 和 DigitalOcean 在文件中明确披露流动性。 | 提供季末现金、受限现金和可用循环信贷额度。 |
| 月度烧钱 | Andromeda 未披露 | 仅凭估值或融资新闻推不出资金续航。 | DigitalOcean 和 Amazon 披露现金流和资本开支细节;Andromeda 没有。 | 提供月度运营开支消耗、任何资本开支,以及营运资本波动。 |
| 资金续航月数 | Andromeda 未披露 | 需要它判断下一轮融资是可选项还是被迫事项。 | DigitalOcean 称现有现金和信贷至少覆盖 12 个月;Andromeda 没有类似表述。 | 按董事会批准模型提供基准、计划和下行三种资金续航月数。 |
| 最近融资用途 | 公开表述是扩大客户基础和团队建设;具体分配未知 | 决定资金流向增长、预购、债务服务,还是防御性流动性。 | 招聘数据体现财务、采购、合作伙伴和计算市场岗位需求。 | 提供 2026 和 2027 年 sources-and-uses 预算。 |
| 下一轮触发条件 | 未披露 | 必须知道未来融资取决于增长、利润率还是隐藏承诺。 | CoreWeave 使用与合同支持资本开支绑定的资产级融资;如果 Andromeda 真正轻资产,可能不需要。 | 说明触发下一次融资或信贷工具的里程碑。 |
| 债务 / 项目融资义务 | 未发现 Andromeda 公开披露 | 即便是经纪商,如果担保供给或为预留融资,也会加杠杆。 | CoreWeave 的 DDTL 5.0 以 SOFR + 4.50% 为合同资本开支融资,并带 DSCR 契约。 | 提供债务时间表、担保、信用证和预留承诺。 |
| 营运资本与租赁敞口 | 未公开披露 | 如果回款滞后于供应商付款,可能成为隐藏资本强度驱动因素。 | DigitalOcean 披露长期 co-lo 租赁义务;Amazon 披露 AWS 预付款和履约义务。 | 提供 DSO、DPO、未实现收入、预付款和任何租赁承诺。 |
| 相对上市可比公司的披露深度 | 低 | 不透明本身就是融资风险,因为投资人只能盲承销。 | CoreWeave、DigitalOcean 和 Amazon 都披露了 Andromeda 缺失的合同、流动性或义务细节。 | 开放包含经审计财务和合同摘要的数据室。 |
本表关注前瞻资本充足性,而不是重复融资轮历史;反复出现的主题是,上市可比公司披露现金、债务和义务,因此能被承销,Andromeda 则不能。
[CI024, CI025, CI026, CI027, CI028, CI032]定性矩阵显示,在公开文件中,Andromeda 介于轻资产云软件姿态与供应商自有 AI 云供给的资本负担之间。
分类是定性的,并有证据支撑;Andromeda 项下的「未知」反映公开披露缺失,而不是证明风险不存在。
[CI024, CI027, CI028, CI031, CI032, CI033]4.4 财务结论、收入质量观点与承保阻碍
Andromeda 的公开财务案例有吸引力,但单独看还不到投资级。正面看,公司似乎找到了真实经济切入口:大型 GPU 买方想要更快访问、更多供应商选择,并减少直接云或超大规模云厂商未必总能解决的双边签约摩擦。如果 Andromeda 在已承诺工作负载上收取净经纪价差,同时避免大规模自有基础设施资本开支,那么报告所称盈利、约 $100M run-rate 规模和小团队结合起来,可能意味着非常有吸引力的收入质量和资本效率。公开 $1.5B 估值锚点,对应粗略约 15x run-rate 收入倍数,或约 30x 2024 年收入;若利润率确实很高且现金转换强,这对一家私营 AI 基础设施中介而言激进但并不荒唐。 问题在于,缺失项恰恰是判断这种乐观解释是否成立所需的输入。没有公开抽佣率披露,没有公开总额 / 净额会计政策,没有经审计毛利率,没有现金或现金消耗 披露,没有债务或担保明细,也没有客户集中度数据。这意味着投资人无法判断收入基础是否耐久、服务交付是增厚利润率还是依赖大量人力,或隐藏营运资本义务是否会迫使进一步融资。最干净的财务结论因此偏谨慎:相对于直接云,Andromeda 的模式可能更轻资本;但若没有管理层会议级数据集,公开证据仍不足以承保收入质量、利润率路径或资本充足性。[CI015, CI016, CI040, CI041, CI042, CI043]
| 缺失指标 | 对承销的影响 | 精确尽调路径 |
|---|---|---|
| 总额 / 净额收入确认政策 | 没有它,估值倍数和收入质量都不可信。 | 索取 ASC 606 备忘录,并按产品线调节 GMV 到已确认收入。 |
| 抽成率 / 经纪价差 | 无法把客户支出或交易数量转成贡献经济性。 | 索取按合同类型和主要账户 cohort 划分的抽成率历史。 |
| 账面现金、烧钱和资金续航 | 仅凭公开材料无法评估资本充足性。 | 索取月度管理账、现金桥和 13 周现金预测。 |
| 毛利率与服务交付成本 | 没有收入成本细节,利润率路径仍是猜测。 | 索取按工作负载类别划分的毛利率和支持成本分摊。 |
| 客户集中度与合同期限 | 如果一两个实验室贡献多数收入,小团队模式风险很高。 | 索取前 10 大客户明细、期限摘要和按 cohort 续约情况。 |
| 营运资本条款和担保 | 所谓轻资产经纪商仍可能因预购或担保消耗现金。 | 索取 DSO、DPO、预付款政策、供应商条款和任何信用证。 |
| 债务、信贷或项目融资义务 | 隐藏杠杆会实质改变承销和下行风险。 | 索取债务时间表、契约包和任何表外承诺。 |
| 经审计的 2024-2025 财务报表 | 媒体报道不足以确认盈利质量或资产负债表强度。 | 开放尽调室,提供经审计财务、董事会预算和与 cap table 联动的资金管理摘要。 |
这些缺口是具体阻碍,导致公开材料今天还无法支撑完整投资承销决策。
[CI015, CI016, CI041, CI043, CI044, CI046]4.5 图表
05产品与技术
5.1 产品定义与模块地图
从客户流程看,Andromeda 提供的是一条托管路径,把碎片化 GPU 供应变成可部署的 AI 基础设施。买方定义 GPU 类型、数量、区域和时间线等工作负载需求;Andromeda 随后在大型供应商网络中采购、基准测试并定价容量,标准化商业包,并给出可部署配置,配一张发票、一个支持渠道和一层统一可观测性。供应商从市场另一侧进入,提交集群细节,包括 GPU 类型、数量、价格、网络、存储、位置、接口类型和预订窗口。官方文案称,这些集群随后会被基准测试、规范化、认证,并按标准化条款匹配给合格需求。 因此,这个产品比泛泛的「GPU marketplace」标签更具体。可见模块包括买方需求录入、供应商接入、认证与基准测试、价格和合同标准化、部署路由,以及持续运营支持。公司自己的长篇洞察页尤其重要,因为它把 GPU 市场缺失的部分框定为交易基础设施,而不是原始硬件所有权:基准测试、标准合同、匹配、运营和信任。SiliconANGLE 与 Upstarts 也强化了这一解读:Andromeda 是一个验证基础设施、管理采购复杂性,并在众多供应商之间充当中立撮合者的层,而不是被某个单一云锁定的运营商。[CE001, CE002, CE003, CE004, CE005, CE006]
| 模块 / 资产 | 主要用户 | 功能 | 差异化 | 成熟度 / 状态 | 尽调缺口 |
|---|---|---|---|---|---|
| 买方需求录入与寻源工作流 | AI 实验室、创业公司或企业基础设施买家 | 收集 GPU 类型、数量、地区和时间线等工作负载需求,然后启动寻源和定价 | 把碎片化容量发现变成一个结构化录入动作 | 已上线且公开可见 | 没有公开证据显示其已有自助 API、调度器插件或客户可配置的策略界面 |
| 提供方接入和认证流程 | 数据中心、云厂商、电信运营商或其他算力提供方 | 采集 GPU、网络、存储、位置、接口和可用性数据,再把集群送入资质认证和需求匹配 | 在供给触达买方之前,先把异质供给标准化 | 已上线,且公开可见 | 基准测试套件、通过阈值和补救规则尚未公开 |
| 基准测试、定价和合同标准化 | 买方和提供方的商业运营人员 | 对供给跑基准测试,在网络内定价,标准化 SLA 和合同,并暴露价格指数信号 | 把市场情报和交易执行结合起来,而不是只做被动挂牌 | 公开说法中已上线,方法不透明 | 价格指数、基准测试权重或合同模板的方法尚未公开 |
| 多提供方编排和控制平面 | 内部平台团队和客户技术团队 | 在 VM、Kubernetes、裸金属和调度器上抽象一层,用于训练或推理任务的开通、生命周期管理和路由 | 基于异质算力底座建立一层统一控制层 | 正在积极建设,招聘信号强 | 没有公开架构文档、API 参考或发布说明证明功能完整度 |
| 运维、可观测性和事故支持层 | 客户平台团队和 Andromeda 运维人员 | 提供可观测性、故障排查、可靠性监控、事故响应和部署后优化 | 承诺提供类似超大云厂商的支持,但不要求客户直接管理每个提供方 | 真实存在,且在扩张 | 公开 MTTR、正常运行时间和事故历史指标缺失 |
| 解决方案工程和 POC 流程 | 技术买方、安全负责人和平台团队 | 负责演示、参考架构、评估计划、成本建模,以及从 POC 到生产的交接 | 让平台比单纯经纪式原始容量更容易购买和部署 | 活跃且团队配置清晰 | 关于完整部署流程的公开客户案例仍然稀少 |
各行反映公开工作流页面、招聘材料和独立报道中明确可见的产品界面,而不是所有内部子系统。
[CE001, CE002, CE004, CE005, CE006, CE007]| 用户任务 | 当前工作流痛点 | Andromeda 方案 | 公开收益信号 | 限制 |
|---|---|---|---|---|
| 快速拿到大规模训练集群 | 供给分散,多年期合同常常太僵,难以匹配突发需求 | 通过一个界面完成聚合采购、基准测试、定价、标准化合同和部署 | 更快接入合格容量,同时只有一张发票、一个支持关系 | 公开来源没有量化中位部署时间,也没有按工作负载类型披露成功率 |
| 在不自建机群的情况下运行推理或短周期工作负载 | 团队可能需要尖峰容量,而不是永久承诺 | 在全球供给之间路由训练和推理任务,并从多个提供方租用资源 | 客户先不用建自有集群,也能轻资产获取算力 | 公开资料没有拆分推理使用、训练使用和类似现货的工作流 |
| 把第三方容量推向市场 | 提供方硬件各不相同,也没有标准路径触达合格企业需求 | 接收集群规格,对环境跑基准测试并认证,再标准化交易结构、路由合格需求 | 让闲置或碎片化供给更快变成收入 | 公开资质标准和失败处理规则未披露 |
| 执行技术评估或 POC | 复杂买方在投入生产前,需要证明吞吐、成本、可靠性和安全性 | 解决方案工程师负责发现需求、演示、POC、参考架构和成功标准映射 | 技术信任更高,向生产交接更干净 | 公开证据没有显示转化率或平均 POC 周期 |
| 运维和调试生产 GPU 集群 | 故障可能来自硬件、存储、网络 Fabric、驱动、编排或 ML 框架 | SRE 和平台团队提供可观测性、GPU 健康检查、事故响应和复盘驱动的修复 | 买方把困难的集群调试工作外包给专业团队 | 自动化与专家人工介入各占多大比例,公开信息仍不清楚 |
这些用例优先覆盖官方工作流界面和当前技术招聘材料中描述的客户与提供方任务。
[CE001, CE002, CE007, CE013, CE015, CE018]Andromeda 的公开架构是一套六层栈,从买方和供应商接入,到认证、控制平面路由、异构执行底座和运营支持。
[CE001, CE002, CE006, CE013, CE015, CE018]面向客户的动作从工作负载定义和寻源开始,进入评估、标准化合同、部署,再到持续运营支持。
[CE001, CE005, CE007, CE018, CE024, CE029]5.2 架构、部署与运营模式
公开运营模式是一套分层控制与支持栈,位于客户、供应商和异构执行环境之间。官方流程页面显示市场两侧都会收集结构化元数据,当前工程岗位则补上缺失的技术细节:Andromeda 正在建设编排、资源配置、生命周期管理、API、服务和控制平面,用来抽象 VM、Kubernetes、bare metal 和基于调度器的环境。供应商表单明确接受 Kubernetes、Slurm 和 VM 接口;软件工程岗位又把范围扩展到「VMs, Kubernetes, bare metal, schedulers」。这有力证明,公司在规范化多个基础设施底座,而不是强行要求一个栈。 部署似乎也有意采用服务主导模式。解决方案工程岗位描述了技术发现、演示、POC、参考架构、基准测试和成本建模,然后才平滑交接到生产。站点可靠性岗位随后说明生产支持的形态:跨多个供应商配置 Kubernetes 集群,使用 Terraform、Helm 和自动化工具,运营可观测性栈,并主导事故和事后复盘。最详细的公开技术信号来自高级 SRE 岗位,点名 topology-aware scheduling、InfiniBand、RoCE、NVLink、NCCL、CUDA、distributed PyTorch、GPU telemetry 和自愈自动化是核心关注点。一位现任 Andromeda 基础设施工程师自发布的 GitHub CV 也方向性支持同一图景,提到 GitOps、KubeRay、Slurm、Weka、VAST、Prometheus、Grafana、Loki 和内部编排工具;不过,该来源应被视为从业者信号,而不是经审计的公司披露。[CE013, CE014, CE015, CE016, CE017, CE018]
| 层 / 组件 | 角色 | 公开证据 | 关键依赖 | 观察到的风险 |
|---|---|---|---|---|
| 接入和元数据层 | 以结构化形式采集买方需求属性和提供方基础设施属性 | 官方客户和提供方表单直接公开所需字段 | 市场两端都提供真实、完整的元数据 | 坏元数据可能误导供给路由,或让工作负载规格不足 |
| 资质审核、认证和市场标准化层 | 对供给跑基准测试、验证性能、规范化报价并标准化条款 | 官网、洞察页和独立报道都描述了基准测试和认证功能 | 可重复执行的基准测试,以及合同纪律 | 公开基准测试方法和阈值仍不透明 |
| 控制平面和路由层 | 把异质基础设施转成可开通容量,并将训练或推理工作负载调度到全球供给上 | 软件工程和 SRE 岗位描述了编排、开通、生命周期管理和多提供方集群演进 | 内部 API、服务、自动化和调度器抽象 | 路由或生命周期 bug 可能伤害成本、性能或可用性 |
| 执行底座 | 在 VM、Kubernetes、裸金属、Slurm 及相关调度环境上运行工作负载 | 提供方工作流和工程岗位都明确点名这些环境 | 提供方就绪度、驱动栈、调度器正确性和硬件健康 | 异质环境会抬高集成和支持复杂度 |
| 可靠性和可观测性层 | 监控 GPU 健康、Fabric 吞吐、系统指标和事故状态,并支持响应和补救 | 高级 SRE 和站点可靠性岗位列出 SLO、错误预算、遥测、复盘和深度可观测性 | GPU 遥测、Fabric 健康数据和有能力的值班人员 | 缺少公开状态或事故指标,外部验证很难 |
| 客户技术合作层 | 连接售前、评估、架构和生产交接 | 解决方案工程岗位要求演示、POC、参考架构,以及在工程和安全团队之间拉齐干系人 | 高质量客户互动和可重复使用的赋能资产 | 如果自动化落后于需求,服务密集型交付可能很难扩展 |
架构表强调公开证据中可见的层和依赖,不声称公司尚未披露的内部实现细节。
[CE004, CE006, CE013, CE014, CE015, CE016]产品依赖外部供应商供给、认证逻辑、编排正确性、异构执行环境和人工支持深度。
[CE006, CE011, CE015, CE021, CE031, CE038]5.3 成熟度、差异化与支持
最强的成熟度信号,不是已经发布的 API 手册,也不是一组认证徽章,而是商业和运营证据在真实环境里反复一致。官方页面已经展示出双边工作流、100 多家供应商,以及标准化的支持话术。独立报道进一步说明,Andromeda 会处理采购细节、监控可靠性问题,并且已经大规模处理真实交易。招聘信息也强化了这一点:这是一门正在运转的生意,不是一份宣传册。软件工程师在搭产品化的编排和控制平面,解决方案工程师在做结构化评估和参考架构,可靠性团队则要负责事件响应、GPU 专属可观测性,以及多供应商集群性能。 因此,Andromeda 的差异化更像是运营能力,而不是单纯靠硬件。公司声称把供应商资质审核、基准驱动的归一化、价格发现、合同标准化、工作负载路由和客户支持合到一个中立层里,在不拥有底层机群的情况下,提供接近超大规模云厂商的一致性。这个模式理论上比直接拥有每个集群更省资本,但也意味着产品依赖供应商质量、内部支持深度和编排正确性。公开成熟度因此并不均衡。买方和供应商工作流看起来已经上线,路由和支持栈也像是真实存在,长期的「流动性层」论点也成立。但公开材料仍缺正式路线图、公开 API 或 SDK 界面,也缺详细技术文档来解释:工作流里有多少已经自动化,又有多少仍靠专家人工介入。[CE009, CE010, CE018, CE019, CE020, CE021]
| 日期 / 阶段 | 功能 / 里程碑 | 状态 | 含义 | 来源 |
|---|---|---|---|---|
| 2025-06 拆分前商业化信号 | Andromeda Cluster 容量已经面向非 NFDG 公司开放,并公布 H100 可用量和价格区间 | 已上线的历史运营证明 | 说明这门业务在 2026 年 3 月公开发布前,已经从真实集群运营中长出来 | DCD |
| 2026-03 公开发布 | 公开官网、X 发布串和融资周报道都把 Andromeda 定位为中立算力市场,拥有 100 多家提供方和 1,000 多笔交易 | 已上线,外部可见 | 确立当前商业产品身份,而不再只是私有的组合公司专用集群 | 官网、X、SiliconANGLE |
| 2026 当前工作流界面 | 买方和提供方工作流页面、合同标准化、认证和支持表述都能在公开网站上看到 | 已上线 | 核心市场工作流看起来已成熟到足以承接外部需求 | 官网、客户页、提供方页、洞察 |
| 2026 当前控制平面建设 | 软件工程岗位聚焦编排、开通、生命周期管理、API、服务,以及跨异质底座的控制平面 | 从招聘推断,正在积极建设 | 暗示产品化推进已超出定制化运维工作 | Built In 软件工程岗位和 Ashby 软件工程岗位 |
| 2026 当前可靠性和机群运维扩张 | 站点可靠性岗位强调多提供方 Kubernetes 运维、GPU 遥测、高速 Fabric 健康、SLO 和复盘 | 从招聘推断,正在积极建设 | 可靠性是核心产品层,但也是高度依赖人员配置的环节 | Built In SRE 岗位、RemoteOK 和 Ashby SRE 岗位 |
| 2026 当前解决方案工程流程 | 解决方案工程岗位把演示、POC、评估计划、参考架构和生产交接正式化 | 从招聘推断,正在积极建设 | 说明 Andromeda 正在扩展可重复的企业部署流程,而不只是聚合供给 | BuiltInSF 解决方案工程岗位和 Built In 职位页 |
| 2026 当前信任和披露缺口 | 公开信任中心仍然稀疏;本次未发现第一方状态页或认证库 | 缺口 | 信任披露落后于产品和收入野心,会给企业买方制造尽调工作 | 信任中心和隐私政策 |
由于公司不发布正式路线图或发布日志,本表区分了有日期的公开里程碑和从当前招聘推断出的建设信号。
[CE015, CE018, CE019, CE022, CE031, CE035]从能力成熟度看,工作流商业化最强,公开基准和信任披露最弱。
[CE007, CE015, CE018, CE027, CE031, CE035]5.4 信任、质量和技术风险
信任与控制披露确实存在,但相对产品承诺仍偏薄。隐私政策是目前最实质的公开控制文件。文件称,Andromeda 运营的是纯商业平台,会收集详细的客户和供应商基础设施数据,用 Cookie 支持认证和核心功能,可按需提供 DPA,并维护访问控制、传输和静态加密,以及持续监控。官网和第三方报道也暗示,基础设施上架前会经过供应商资质审核和安全筛查。这些都是正面信号,但大多仍是公司自己的说法。公开信任中心除了标题页几乎没有更多内容,本轮也没有发现第一方公开状态页、具名外部认证、已发布 SLA 文件或事件复盘。 这个缺口很关键,因为 Andromeda 明确把产品卖点放在跨第三方基础设施的可靠性、标准化和信任上。如果基准方法、认证门槛、供应商整改规则和事件历史继续不透明,买方就只能很大程度上凭信任接受公司的质量主张。运营依赖图也不简单:供应商供给可能收紧,网络或存储缺陷可能打断训练任务,编排错误可能把作业路由错位,服务很重的支持模式在需求超过专业人员供给时也会变成瓶颈。正确结论不是产品不成熟,而是商业工作流走在公开信任界面前面。因此,投资人或企业买方应把基准方法、供应商质量控制、状态报告和自动化深度视为剩余技术尽调里最重要的议题。[CE011, CE021, CE025, CE026, CE027, CE028]
| 控制领域 | 公开状态 | 范围 | 缺口 | 来源信号 |
|---|---|---|---|---|
| 隐私和控制者 / 处理者模型 | 已公开声明 | 商用平台、控制者 / 处理者区分;在 GDPR 或 UK GDPR 场景下可按要求提供 DPA | 未发现公开模板 DPA 或已谈妥的企业隐私附件 | 隐私政策 |
| 核心安全控制 | 已公开声明 | 访问控制、传输和静态加密,以及持续监控 | 未发现公开 SOC 2、ISO 27001、渗透测试摘要或等效审计材料 | 隐私政策和信任中心框架页 |
| 提供方质量筛查 | 公开声称 | 上线前需满足性能和安全要求,并通过企业级基准验证 | 公开基准测试套件、阈值和重新认证节奏未披露 | 官网、提供方工作流和 SiliconANGLE |
| 可靠性工程纪律 | 招聘信号强 | SLO、错误预算、复盘、可观测性和面向 GPU 基础设施的事故响应 | 没有公开正常运行时间指标、状态页或面向客户的事故档案 | Built In 和 RemoteOK 职位描述 |
| 公开信任披露界面 | 稀疏 | 存在信任中心,隐私政策也较详细 | 相比公司对可靠性和标准化的主张,披露界面仍偏薄 | 信任中心和已审阅的官方页面 |
公司披露隐私的内容多于可由外部验证的安全或可靠性表现。
[CE006, CE011, CE025, CE026, CE027, CE028]5.5 证据展项
06客户
6.1 客户分层和购买中心形态
Andromeda 的公开客户画像,在细分层面最清楚,在具名账户层面并不清楚。官方文案、隐私政策和客户需求表都一致显示,买方动作瞄准的是 AI 团队:它们需要突发或大规模训练与推理容量,又不想签多年期超大规模云承诺。直接买方通常是能指定 GPU 类型、数量、区域、时间和时长的基础设施、研究或平台负责人。实际使用者看起来是 ML 研究员、平台工程师和 SRE 式基础设施团队;账户一旦足够大,经济买方又可能上移到财务、采购、安全或法务。这个拆分很重要,因为 Andromeda 卖的不是自助式商品 API,而是围绕集群的一层重服务采购和运营能力。 公开招聘补上了缺失的分层细节。Solutions Engineer 岗位明确面向前沿实验室、AI 原生创业公司和企业,并把工程领导、平台团队,以及安全或基础设施负责人列为核心利益相关方。隐私政策也单独说明,平台服务 AI 团队和企业客户,同时收集财务和账单数据,这支持了付款方与使用者分离的判断。在供给侧,Andromeda 也实际把供应商当作渠道客户或交易对手来服务:数据中心、云供应商、电信运营商和其他运营方通过同一平台把集群推向市场。因此,本章更合适的框架是一个双边市场,需求侧有三类尤其重要:NFDG 或 AI Grant 创业公司、保密的大型 AI 实验室,以及做结构化评估的企业技术买方。 地理上,公司以美国为中心,但不只服务美国。公开职位集中在 San Francisco,以及北美或全球远程覆盖;官方页面则承诺可提供全球算力,并收集供应商位置数据。由此形成的细分画像可信、商业逻辑也顺,但仍缺最影响承保判断的拆分:活跃客户数、按细分市场划分的收入、地域组合,以及需求里还有多少来自 NFDG 生态。[CU001, CU002, CU003, CU004, CU005, CU006]
| 细分 | 买方 / 用户 / 付款方 | 主要用例 | 公开证明 | 战略价值 | 缺口 |
|---|---|---|---|---|---|
| 前沿 AI 实验室 / 模型构建者 | 买方:基础设施 / 研究负责人;用户:训练工程师;付款方:财务 / 采购 | 大规模训练和高强度推理集群 | 官方页面和职位信息反复提到领先 AI 实验室和头部客户训练任务 | 支出大、紧迫性高、重复使用潜力强 | 未披露具名实验室客户或收入拆分 |
| AI 原生初创公司 / AI Grant 公司 | 买方:创始人 / 平台负责人;用户:研究和 ML 基础设施团队;付款方:初创公司财务 | 突发训练任务、早期推理、灵活集群访问 | AI Grant 提供 Andromeda Cluster 使用权;Upstarts 点名 ElevenLabs 和 Pika 是早期用户 | 早期用量渠道和创始人网络飞轮 | 当前活跃客户数和按批次留存不公开 |
| 企业 AI / 平台团队 | 买方:高管支持者,加上平台、安全和采购干系人;用户:内部 AI 平台团队 | 结构化评估、POC、生产交接 | 解决方案工程师岗位明确瞄准企业和战略账户 | 更高 ACV 和先落地再扩张潜力 | 未发现具名企业部署参考 |
| 推理密集型 AI 运营方 | 买方:平台或服务负责人;用户:推理 / SRE 团队;付款方:财务 | 短周期或可变需求推理工作负载 | 官方洞察和职位信息提到跨提供方路由训练和推理 | 匹配向灵活、非多年期需求迁移的趋势 | 公开尚未确认具名推理客户 |
| 算力提供方 / 供给侧交易对手 | 买方:提供方 BD 或基础设施负责人;用户:集群运维人员;付款方:提供方组织 | 把闲置或碎片化 GPU 容量推向市场 | 提供方表单、合作伙伴岗位和采购岗位都显示结构化提供方接入 | 供给广度是客户价值和利用率的核心 | 供应商集中度和按供应商划分的收入结构未披露 |
| VC / 加速器渠道 | 买方:基金或加速器合伙人;用户:被投公司基础设施团队;付款方:因交易结构而异 | 为被投公司的算力需求提供获取渠道 | 合作伙伴负责人岗位显示,VC 基金和加速器是有意布局的分销伙伴 | 高效获客和生态锁定 | 如果需求仍过度偏向 NFDG,会形成集中风险 |
各行拆分的是可见需求和渠道界面;Andromeda 没有公开完整客户数或收入占比,因此本表不暗示这些数据。
[CU001, CU002, CU005, CU006, CU007, CU008]展示 Andromeda 的创业公司、实验室和企业客户从线索来源到生产支持的高接触路径。
阶段综合自官方工作流页面、隐私政策和面向客户的招聘材料,而不是来自一张已发布的第一方流程图。
[CU003, CU004, CU009, CU011, CU021, CU036]6.2 采用轨迹和具名客户证据
现有证据足以说明 Andromeda 已有真实采用,但还不足以说明公开客户证据很强。官方页面撑住了漏斗顶端和运营规模叙事:100 多家供应商、1,000 多笔交易、超过 10 亿 GPU 小时活动,以及横跨多家供应商的数十个集群。独立报道随后补上重要商业化标记。DCD 报道称,到 2025 年 6 月,该集群已经向非 NFDG 客户开放,公布了报价、快速 onboarding,并可提供最高 2,000 张 H100。Upstarts 和 SiliconANGLE 又补充说,公司 2025 年运行收入约 $100M,同时服务 Moushey 大多不愿具名的知名创业公司和实验室。合在一起,这些足以证明真实生产需求已经存在,尽管证据仍以汇总指标为主,而不是具体账户。 具名证据薄得多。本轮最清楚的显式名称是 ElevenLabs 和 Pika:Upstarts 称两家公司都使用了早期投资组合集群,SaaS Sentinel 也重复了这一点。这些引用重要,因为它们把 Andromeda 连接到爆发式 AI 工作负载,而不是泛泛的创业公司 logo;但它们仍够不上最好的客户证据形式。Andromeda 没有第一方案例研究,没有 ElevenLabs 或 Pika 的客户引语把生产使用归因于 Andromeda,也没有部署时长、训练吞吐、节省金额、正常运行时间或合同扩张等公开部署结果。Cursor、Perplexity、Browserbase 和其他 AI Grant 公司对渠道分析有意义,因为 AI Grant 称更大额投资会附带 Andromeda Cluster 访问权;但这些名称的直接部署仍只与生态相关,并未被公开确认。 因此,正确结论应克制,而不是宣传式。Andromeda 看起来已经从只服务投资组合公司的实验,走向了一个真实的保密算力市场,既有创业公司需求,也有更大实验室需求;但对一家声称已有显著规模的公司来说,公开引用集仍异常稀疏。投资人应把汇总采用指标视为支撑证据,把具名客户证据视为真实但不完整。[CU012, CU013, CU014, CU015, CU016, CU017]
| 指标 / 里程碑 | 数值 / 状态 | 日期 | 来源 / 置信度 | 含义 | 缺失分母 |
|---|---|---|---|---|---|
| 初始托管集群需求 | 几乎瞬间被填满 | 2023-2024 | Upstarts / 中 | 早期需求超过初始供给 | 没有精确客户数或等待名单规模 |
| 早期内部容量 | 2024 年初已有 4,000+ 块 GPU | 2024-early | Upstarts + SaaS Sentinel / 中 | 规模足以在外部发布前支持多个组合公司用户 | 按客户分配了多少仍未知 |
| 更广泛商业化前的集群足迹 | 3,200 块 H100 + 432 块 H100 + 768 块 A100 由 DCD 引述 | 2025-06-21 | DCD / 中 | 显示生产级库存和异质工作负载支持 | 没有利用率或唯一账户数 |
| 外部市场开放 | 非 NFDG 公司可购买访问权;数小时内最多有 2,000 块 H100 可用,价格为 $2.40-$3.00/GPU-hour | 2025-06-21 | DCD / 中 | 相比封闭组合公司需求,边界明显拓宽 | 未披露转化的非 NFDG 买方数量 |
| 提供方广度 | 100+ 家提供方 | 2026 当前 | 官网 + SiliconANGLE / 高 | 支撑供给多样性和路由价值主张 | 没有提供方集中度或活跃提供方占比 |
| 交易数量 | 已完成 1,000+ 笔交易 | 2026 当前 | 官网 + SiliconANGLE / 高 | 最强公开重复使用替代指标 | 交易不等于客户或续约 |
| 运营吞吐量 | 十亿级以上 GPU-hours;跨许多提供方的数十个集群 | 2026 当前 | 官网洞察 + Upstarts 背景 / 高 | 显示真实部署量,而不只是堆客户标识 | 没有按训练与推理或客户细分拆分 |
| 商业规模 | $100M 2025 年化规模;知名初创公司和实验室大多未具名 | 2026-03 | Upstarts + SiliconANGLE / 中 | 即使公开参考很薄,采用也已有有意义的商业规模 | 没有客户数、ACV 或头部客户占比披露 |
本表混合采用、供给和商业化代理指标,因为 Andromeda 披露平台活动指标,却不披露干净的客户账户数。
[CU012, CU013, CU014, CU015, CU016, CU017]| 客户 | 细分 | 部署 / 用例证据 | 生产 / 试点 | 结果 / 参考质量 | 限制 |
|---|---|---|---|---|---|
| ElevenLabs | AI 原生初创公司 / 语音 AI | Upstarts 称 ElevenLabs 使用过 Andromeda 早期集群;SaaS Sentinel 也重复了这一点 | 有历史使用迹象;当前生产状态未披露 | 两篇独立的 2026 年 3 月报道都点名;公司相关性强,但仍不是第一方证明 | 没有客户引用、案例研究、支出、工作负载规模或 2026 年连续性确认 |
| Pika | AI 原生初创公司 / 生成式视频 | Upstarts 称 Pika 使用过早期 NFDG 支持的集群;SaaS Sentinel 独立呼应了这一引用 | 有历史使用迹象;当前生产状态未披露 | 两篇独立报道点名;GPU 密集型用例可信 | 没有第一方部署细节、结果或续约可见性 |
| Cursor | AI Grant 第 1 批 / AI 开发者工具 | AI Grant 确认 Cursor 入选该批次,并且较大投资可获得集群访问;BuildMVPFast 后来把 Cursor 归入 Andromeda 的组合用户圈 | 生态访问已确认;直接部署未确认 | 可作为渠道证据,但作为部署证明较弱 | 只有间接关联;不足以视为当前在线的具名生产账户 |
| Perplexity | AI Grant 第 1 批 / AI 搜索 | AI Grant 确认其组合公司身份和集群访问渠道;BuildMVPFast 后来把 Perplexity 列入 Andromeda 的算力圈 | 生态访问已确认;直接部署未确认 | 显示渠道可能触达知名 AI 公司 | 没有公开直接声明 Perplexity 已在 Andromeda 部署生产工作负载 |
本表刻意区分明确点名使用与较弱的生态关联证据;仅有标识或组合公司身份,不视为生产使用证明。
[CU027, CU029, CU030, CU031, CU047, CU038]展示从渠道驱动的线索生成,到已评估的生产部署和重复使用的推进过程。
这是数值漏斗的流程替代图,因为 Andromeda 没有发布逐阶段转化数量。
[CU015, CU019, CU020, CU021, CU023, CU034]对少数已审阅公开参考,比较具名客户证据的质量、新鲜度和生产可见度。
证据质量标签是基于来源具体程度和独立性的分析判断,并非公司发布的评分体系。
[CU025, CU026, CU029, CU030, CU031, CU047]6.3 持久性、扩张路径和集中度风险
持久性是公开材料最薄弱的地方。已审阅来源没有披露 NRR、GRR、流失率、续约率、合同期限、队列曲线,甚至没有披露当前客户数。这不代表业务没有留存,而是外部投资人无法从公开证据核验。可用的最好代理指标只有方向性。官方和第三方来源显示,公司有可重复的交易流,部署后仍持续介入的服务层,以及一位匿名创始人告诉 Upstarts:他们的公司继续与 Andromeda 合作,是因为它像外包工程团队,比超大规模云厂商更灵活。与此同时,同一位创始人也主动给出核心持久性警示:如果创业公司规模足够大,最终可能把这些能力收回内部。 扩张路径看得见,但转化率看不见。Solutions Engineer 岗位描述了一套 land-and-expand 动作:靠演示、POC、客户 ROI 和战略账户里的成功生产交接来扩张。合作岗位更进一步,明确点名 GPU 供应商、AI 实验室、OEM、VC 基金和加速器都是渠道放大器,并称合作 KPI 与收入贡献、利用率和增长挂钩。这说明 Andromeda 试图同时扩大直接企业式销售,以及把需求导向平台的投资组合或生态渠道。采购、算力交易和业务人员岗位也暗示,公司正在实时平衡客户需求、供应商经济性和合同结构。 同样的特征也制造集中度风险。业务起源于 NFDG 生态,Upstarts 称许多投资组合公司仍是客户,即使它们已不再享受优先待遇。客户不透明也让头部账户风险无法量化:最大的实验室很可能最保密,也最难作为引用。采购摩擦也不轻。法务、安全、出口、财务和定制供给岗位都说明,大账户的签约和留存是高接触流程,而不是低摩擦自助动作。并且,由于 Andromeda 往往凭借分数化采购和 SRE 层取胜,一些成熟客户最终可能把这项能力内化。因此,商业故事具备扩张能力,但公开层面还没有足够持久性,无法在没有管理层室数据的情况下承保。[CU032, CU033, CU034, CU035, CU036, CU037]
| 指标 | 数值 / 状态 | 分部 / 代理指标 | 置信度 | 尽调要求 |
|---|---|---|---|---|
| 净收入留存(NRR) | 全部分部 | 低 | 要求按初创公司、实验室和企业分组提供 NRR | |
| 毛收入留存(GRR) | 全部分部 | 低 | 要求提供 GRR、客户数留存和流失定义 | |
| 客户数 / 活跃账户 | 全部分部 | 低 | 要求提供当前活跃客户数和过去 12 个月活跃客户数 | |
| 重复使用代理指标 | 1,000+ 笔交易,加上持续的需求路由角色 | 混合分部 | 中 | 要求拆分重复账户与净新增账户贡献的交易占比 |
| POC 到生产转化 | 内部跟踪,但未公开披露 | 战略客户 / 企业 | 中 | 要求提供评估到生产转化率和部署中位时长 |
| 合约期限 / 使用节奏 | 公开证据显示,使用形态混合了突发使用、按小时计价和更大的战略交易 | 初创公司、实验室、企业 | 中 | 要求提供按需、预留和较长期承诺的组合 |
| 客户满意度 / 支持质量 | 匿名创始人称赞灵活性和工程支持;未找到公开评价语料 | 初创公司 / 实验室代理指标 | 低 | 要求安排客户访谈,并提供支持 KPI 历史(SLA、MTTR、续约驱动因素) |
持久性证据几乎全靠代理指标;null 表示该指标有商业重要性,但本轮审阅的公开材料不足以支撑。
[CU030, CU032, CU033, CU034, CU035, CU036]| 扩张驱动因素 / 集中风险 | 类型 | 影响 | 尽调路径 |
|---|---|---|---|
| VC 和加速器合作把组合公司导向 Andromeda | 扩张驱动因素 | 高效获取早期客户,并形成创始人信任飞轮 | 量化销售管线和收入仍有多少来自 NFDG 或 AI Grant,而非直销 |
| 向非 NFDG 买家开放后,可触达需求不再局限于自有组合公司 | 扩张驱动因素 | 降低纯生态依赖,支撑外部商业化 | 要求提供 2025 年中以来非 NFDG 账户分组及其收入贡献 |
| 由服务牵引的 POC 和战略客户打法,可能扩展为更大的生产合作 | 扩张驱动因素 | 如果评估能转化,支撑先落地后扩张 | 要求提供胜率、扩张 ARR,以及每名解决方案工程师撬动的人效 |
| 保密的大型 AI 实验室可能是体量超大的客户,但无法点名 | 集中风险 | ACV 可能很高,但无法用公开证据估算规模 | 要求提供前 5 大客户占比、最大客户占比和续约可见度 |
| 来自组合公司的客户基础可能仍过度绑定 NFDG 生态 | 集中风险 | 渠道集中会削弱议价能力和外部验证 | 按渠道映射客户来源,并验证非 NFDG 客户是否已主导增长 |
| 部分成熟客户可能逐步把采购和 SRE 能力内化 | 集中风险 | 可能限制最成熟客户的长期留存 | 询问流失原因,以及客户从平台「毕业」的案例 |
| 大额交易的采购、法务、安全和出口审查看起来都要大量人工介入 | 集中风险 / 摩擦 | 周期更长、销售支持负担更重,可能拖慢扩张 | 要求提供平均销售周期、安全审查通过率和合同签署瓶颈 |
扩张逻辑可信,但公开记录仍未披露头部客户集中度、渠道组合,或大型保密实验室账户粘性究竟如何。
[CU015, CU019, CU020, CU034, CU037, CU038]| 摩擦点 | 公开信号 | 对客户的影响 | 商业含义 | 尽调要求 |
|---|---|---|---|---|
| 安全 / 数据处理审查 | 隐私政策说明可按请求提供 DPA,并描述仅限业务用途的数据处理 | 企业买家部署前可能需要正式的法务和安全审查 | 可能拉长企业导入;但一旦成交,账户质量也更高 | 要求提供标准安全材料包、DPA 模板和合规路线图 |
| 多方参与的技术评估 | Solutions Engineer 岗位说明购买流程涉及高管、平台团队和安全负责人 | 交易需要协调技术验证,不是简单的自助下单 | 抬高售前成本,也更依赖稀缺的高级人才 | 要求提供每笔交易的利益相关方中位数,以及按分部的转化率 |
| 关键收入合同 | Built In 招聘摘要强调关键收入协议、隐私 / 出口和金融合作 | 大额交易可能需要定制合同工作和合规审查 | 可能拖慢成交,或缩小买家池 | 要求提供平均红线修订周期和标准条款例外 |
| 实时供需匹配 | Compute Trader 和业务岗位围绕把线索匹配到算力、最大化利用率 | 客户在生产启动前可能需要定制寻源 | 供给侧复杂度可能改善利润率,但也制造执行风险 | 要求提供即时满足与逐步撮合的需求占比 |
| 客户背书不透明 | 审阅材料未出现官方案例库或具名客户引语集 | 潜在客户缺少公开证据来降低采购决策风险 | 可能抬高 CAC,让企业销售更难 | 要求提供客户背书计划细节,以及 NDA 下可背书客户名单 |
| 公开信任中心内容稀疏 | 相比企业级主张,信任中心公开的实质内容很少 | 重视安全的客户可能需要更深入的线下尽调 | 削弱大客户公开转化支撑 | 询问信任中心内容、SLA 披露和事故历史透明度路线图 |
这张补充表把采购摩擦拆成具体尽调议程;由于未找到按时间分桶的留存数据,它替代留存队列图。
[CU021, CU022, CU023, CU035, CU036, CU041]6.4 证据展项
07风险
7.1 按严重度排序的风险栈和投资含义
Andromeda 的公开风险栈并不均匀。最高的剩余风险,是 Nat Friedman 和 Daniel Gross 退居幕后之后,治理和连续性仍不透明,公开控制权披露也很薄。这很重要,因为两位创始人曾是早期供给、赞助方和客户叙事的核心,而 Wil Moushey 现在看起来是运营重心。第二个风险是运营和供应商依赖:Andromeda 的价值主张建立在采购、基准测试、签约和支持第三方供给上,而不是自有标准化超大规模云栈。第三个风险是客户集中和多云并用,因为业务公开服务保密 AI 实验室,仍未披露客户数和留存数据,且至少一位早期具名用户现在公开在 Google Cloud 上扩张。第四个风险是资本容量和利润率复杂度。Strategic Compute Finance 岗位和 Moushey 自己的评论都说明,贷款方关系、承保和营运资本架构是正在发生的问题,不是后台细节。第五个风险是出口管制、制裁、隐私和信任上的合规执行。招聘和供应商广度显示了公开缓释动作,但成熟控制的公开证据仍有限。投资含义很直接:Andromeda 仍可能值得投,但前提是管理层能在资本投入前,用具体内部证据补上这些披露和控制缺口。[CR001, CR002, CR017, CR018, CR019, CR020]
这张定性热力图按发生可能性和影响程度梳理 Andromeda 的主要风险;治理不透明、供应商依赖和资本—产能复杂性占据右上角。
可能性和影响程度是基于公开来源作出的定性判断,并非量化概率模型。
[CR042, CR045, CR046, CR047, CR048, CR049]这张有向图展示上游治理、合规、供应商、客户和资本风险如何传导为收入、利润率、融资和估值压力。
边表示分析上的传导路径,而非实测因果权重。
[CR020, CR024, CR032, CR033, CR041, CR044]7.2 监管、法律、隐私和信任风险
即便没有确认的公开诉讼或执法行动直接指向 Andromeda,监管和法律风险仍然重要。核心原因在结构上:Andromeda 公开把自己描述为连接客户和基础设施供应商的全球 GPU 市场,而隐私政策确认公司会处理联系人、账户、账单和服务使用信息。这意味着公司不只是营销 GPU;它还集中管理敏感交易对手、合同和数据流。BIS 和 Federal Register 材料显示,先进计算控制现在已经直接覆盖规避、目的地风险和受控芯片扩散。OFAC 框架让制裁合规不仅适用于美国实体,也适用于使用美国原产货物或服务的外国实体。与此同时,FTC 指引让隐私承诺和事件响应都具备运营后果。Andromeda 的缓释姿态可见但不完整:公司正在招聘承担隐私和出口职责的商业律师,但本轮审阅的信任中心仍很稀疏,没有提供谨慎企业买方或投资人预期看到的认证、状态或事件历史细节。因此,正确的承保姿态不是假设存在隐藏问题,而是把出口筛查、制裁筛查、事件响应、DPA 流程和信任证据列为放行前尽调要求。[CR003, CR004, CR005, CR006, CR007, CR008]
| 规则 / 许可 / 案件 | 辖区 | 状态 | 发生概率 | 严重性 | 缓释措施 | 剩余敞口 | 尽调路径 |
|---|---|---|---|---|---|---|---|
| 先进计算出口管制 / AI 扩散规则 | 美国 BIS / Federal Register | 已生效且仍在演进;反规避和许可规则适用于敏感芯片流向及交易对手 | 中高 | 极高 | 专门法务招聘,加上可能的内部交易对手审查 | 高 — 未发现公开筛查流程、国家矩阵或责任归属 / 升级机制证据 | 要求提供出口管制政策、被拒方筛查步骤、受限辖区规则和升级责任人 |
| OFAC 制裁筛查和交易对手 | 美国财政部 / OFAC | 现行制裁项目和合规框架适用于与美国有关、以及使用美国原产商品或服务的外国用户 | 中 | 极高 | 市场平台集中化原则上让合规项目设计可行 | 高 — 项目证据未公开,全球供应商 / 客户路由抬高筛查复杂度 | 要求提供制裁手册、受益所有权核查、筛查工具输出和例外处理 |
| 隐私政策、DPA 和数据泄露响应义务 | FTC + 公司法律承诺 | 隐私政策已公开,FTC 商业指引仍有效 | 中 | 高 | 业务用途定位、单一支持层和可能的合同流程,能把合规控制集中起来 | 中高 — 信任披露细节稀疏,且没有公开事故历史材料包,增加尽调负担 | 获取 DPA 模板、子处理方、留存计划、安全材料包和事故响应手册 |
| 商业合同、IP、雇佣和治理法律负荷 | 公司法务体系 / 合同法 | 公开招聘显示,公司仍在通过专职法律顾问岗位补齐这些职责 | 中 | 高 | 商业法律顾问招聘和标准化市场平台流程 | 中高 — 法务职能成熟度尚无法从外部证明 | 要求提供法务 / 合规组织架构图、合同标准、IP 所有权审查和董事会治理日历 |
| 关键交易对手的设施、安全和环境准备度 | 供应商 / 设施特定制度 | 审阅来源未出现公开的设施级许可或安全材料包 | 中低 | 中高 | 轻资产模式可能把部分义务推给交易对手 | 中 — 若关键设施或供应商未通过安全或出口准备,缺口仍在 | 要求提供关键设施清单、许可、安全证明,以及任何环境或劳工事项升级记录 |
严重性和发生概率为定性判断,依据是一手监管 / 法律来源,以及本轮审阅的公司自有信任与隐私页面。
[CR003, CR004, CR005, CR006, CR007, CR008]7.3 运营、合作方和客户依赖风险
从运营上看,Andromeda 更像市场运营方和托管采购层,而不是自给自足的云。Upstarts 称,公司从 CoreWeave 等供应商和其他运营方租赁供给,而不是承担硬件所有权风险;采购和业务人员岗位则显示,供给获取、供应商经济性,以及从线索到容量的匹配仍是核心且活跃的职能。SemiAnalysis 提供了有用的反向视角,把 Andromeda 描述成架在多家 neocloud 之上的分数化 SRE 和采购层。这个模式有吸引力,因为它避开了部分固定资产敞口,但也把风险推向合作方可靠性、供应商集中度、替代速度和质量控制。客户侧风险遵循同样模式。DCD 显示 NFDG 之外已有真实商业化,但公开证据在市场活动上远强于在多元收入集中度上。Upstarts 明确表示大型实验室业务保密,同一篇文章也给出了公开材料里最强的反向留存信号:一位匿名创始人说,公司最终可能把这项能力收回内部。ElevenLabs 在 2026 年扩大使用 Google Cloud,也说明早期与 Andromeda 相邻的客户可以在别处多云并用。结果是,平台确实有需求和供给证据,但仍明显依赖外部供应商、不透明大客户和高接触人工执行。[CR017, CR018, CR019, CR020, CR021, CR022]
| 失效模式 | 发生概率 | 严重性 | 缓释成熟度 | 剩余敞口 | 未解决缺口 |
|---|---|---|---|---|---|
| 供应商质量或可用性故障扰乱客户工作负载 | 高 | 高 | 中 | 高 | 未找到公开的 SLA 材料包、替换矩阵或供应商集中度披露 |
| 供需匹配跟不上增长,利用率受损 | 中高 | 高 | 中 | 高 | 招聘显示公司在主动手工匹配并管理利用率,但履约率历史未公开 |
| 安全或事故响应弱于客户保密要求 | 中 | 高 | 中低 | 高 | 信任中心细节、认证证据和事故日志未公开可见 |
| 高度人工介入的 POC 和战略客户导入拖慢扩张,或压垮稀缺人才 | 中 | 中高 | 中 | 中高 | 未发现公开的转化、部署时间或支持容量指标 |
| 跨供应商基准测试和编排未能标准化异质供给 | 中 | 高 | 中 | 高 | 外部投资人无法查看供应商评分卡、基准测试方法或升级数据 |
本登记表聚焦于从市场平台流程、人员配置模式和稀疏的公开信任披露中可见的运营失效模式。
[CR007, CR014, CR015, CR017, CR018, CR019]| 依赖项 | 交易对手 | 角色 | 集中度 | 失败情景 | 严重性 | 缓释措施 | 剩余敞口 |
|---|---|---|---|---|---|---|---|
| GPU 供给和集群容量 | 第三方供应商 / 云厂商 / 新云厂商 | 为客户工作负载寻源容量,并保障正常运行时间 | 公开未知;公司称有 100+ 家供应商,但未披露头部供应商占比 | 少数关键供应商失效、重新定价上涨,或替换速度慢于客户时间表 | 极高 | 供应商广度、采购领导力和挂牌流动性 | 高 |
| 创始赞助方和渠道网络 | NFDG / AI Grant / 创始人关系 | 早期客户和可信度渠道 | 历史上重要;当前占比未披露 | 创始人退居幕后削弱转介绍流、人才吸引力或专有供给获取 | 高 | 非 NFDG 商业化和专职合作伙伴招聘 | 中高 |
| 具名和保密 AI 实验室客户 | 大型实验室和初创公司买家 | 收入和使用需求 | 公开未知;客户数和头部账户占比未披露 | 一两个大客户主导支出,或把工作负载迁回超大规模云厂商或自有集群 | 高 | 重服务和灵活采购模式 | 高 |
| 资本伙伴和贷款方 | Paradigm 加上未来债务或结构化融资提供方 | 扩大资本容量并为供给融资 | 公开未知;贷款方名单未披露 | 贷款方或资本伙伴无法随需求扩张,或要求不具吸引力的经济条款 | 高 | 专职算力金融招聘和据报已盈利的基础业务 | 高 |
| 硬件和设施生态 | OEM、ODM、集成商、零部件供应商、数据中心 | 支撑集群建设和供给经济性 | 分散但不透明 | 上游生态出现延误、短缺或经济性恶化 | 中高 | 专职采购负责人和全球寻源努力 | 中高 |
最大的盲点是集中度和替换速度,因为公开档案没有提供供应商、客户或贷款方占比数据。
[CR017, CR018, CR019, CR021, CR022, CR023]这张图梳理维持市场运转所需的关键外部依赖:供应商、渠道、监管者和资本。
这张图突出依赖方向,而不是依赖程度或排他性;集中度仍是需要私有数据核查的尽调事项。
[CR018, CR019, CR021, CR026, CR032, CR033]7.4 财务模型、团队和否决标准
财务和执行风险不是典型困境信号,而是规模与控制信号。Upstarts 报道公司已盈利,2025 年收入运行率为 $100M,但这些数字没有回答更重要的尽调问题:营运资本、交易对手融资、对冲、信用敞口或贷款方支持。事实上,公开招聘模式指向相反方向:Strategic Compute Finance 岗位明确提到 capex 规划、情景建模、交易结构设计和贷款方关系,这意味着平台扩张需要围绕容量做真正的金融工程,而不只是写更多软件。组织图在公开层面同样未完成。开放岗位横跨法务、财务、采购、合作、解决方案、软件和 SRE,而官方页面仍没有给出清晰董事会或控制权地图。因此,否决标准应具体、可监控:如果 Moushey 离开且没有继任图谱,就是治理失败;如果无法证明筛查、DPA 或事件控制,就是合规失败;如果少数客户、渠道或供应商主导交易量,就是集中度失败;如果管理层无法证明贷款方深度、流动性韧性和利润率承受力,就是资本失败。这些不是理论问题;它们正是一个高速增长算力市场停止复合增长、变得不可融资的条件。[CR001, CR002, CR008, CR021, CR031, CR032]
| 角色 / 职能 | 依赖或缺口 | 发生概率 | 严重性 | 缓释措施 | 尽调路径 |
|---|---|---|---|---|---|
| CEO / 运营重心 | 在创始人连续性下降时,Wil Moushey 似乎承担了公开运营叙事 | 中 | 极高 | 现有执行动能和跨职能活跃招聘 | 要求提供继任计划、关键人留任方案和授权运营图谱 |
| 董事会 / 治理层 | 审阅的官方页面未出现公开董事会名单或控制权图谱 | 高 | 高 | 如果治理强于公开披露所示,尽调中或可解决 | 要求提供董事名单、观察员权利、委员会结构和所有权 / 控制权摘要 |
| 商业法律顾问和合规责任 | 法务、出口、隐私和合同职责仍显示为开放岗位或新建角色 | 中 | 高 | 专职招聘加标准化市场平台流程 | 要求提供法务组织架构图、外部律师组合和合规 RACI |
| 算力金融负责人 | 公司明确在招聘或扩充资本容量和贷款方管理专长 | 中 | 高 | 角色定义异常具体、指向明确 | 要求提供当前财务负责人、贷款方销售管线和利润率敏感性模型 |
| SRE / 解决方案 / 供给执行梯队 | 专业岗位意味着执行依赖稀缺的技术和运营人才 | 中 | 中高 | SRE、解决方案、采购和软件方向有多个公开岗位 | 要求提供已填补与开放岗位对比、流失率、招聘速度和单点依赖职能 |
人员风险偏高:公开招聘显示公司在有计划地搭建团队,但公开治理披露落后于运营复杂度。
[CR001, CR002, CR008, CR021, CR032, CR033]| 风险 | 可监控触发信号 | 阈值 / 事件 | 行动含义 |
|---|---|---|---|
| 治理 / 连续性不透明 | 领导层和治理披露 | Moushey 离任、创始人控制权模糊未解决,或管理层在交割前无法提供董事会和继任图谱 | 在治理和控制权有书面文件前,视为投资逻辑破裂 |
| 合规执行 | 出口、制裁、隐私和信任证据 | 未见成文筛查流程、DPA / 安全材料包,或任何重大数据事件 / 执法事件 | 暂停或放弃,直到控制措施有证据支撑,且事件范围已厘清 |
| 供应商集中度与可靠性 | 容量替代能力与供应商份额 | 前三大供应商占据多数容量,且没有快速替代计划或 SLA 证据 | 下调确信度、放宽下行情景,或要求更强合同保护 |
| 客户集中度与多云使用 | 头部客户 / 渠道占比与访谈质量 | 少数保密实验室或 AI Grant 相关渠道主导支出,或已具名参考客户把战略工作负载迁往别处 | 如果多元化很弱,重新核算留存假设,并考虑不投 |
| 资本能力与利润率韧性 | 流动性与贷款方支持 | 管理层无法说明贷款方深度、营运资本政策,或采购冲击下的利润率承受力 | 视为估值和融资层面的否决触发;要求重新定价,否则退出 |
| 执行能力余量 | 关键岗位补齐进度与运营指标 | 关键空缺岗位长期未补,或增长放大时,部署 / POC / 利用率指标恶化 | 下调确信度,并在承销激进上行前要求运营证据 |
这些触发条件被设为可观察的尽调条件,而不是预测,因为公开披露仍基本没有量化集中度和流动性。
[CR042, CR045, CR046, CR047, CR048, CR049]08估值
8.1 建议:跟踪——公司重要,但当前价格仍需要分母和治理证据
基于公开证据,对当前 $1.5B 锚点下的 Andromeda,结论是跟踪,不是买入,也不是否定。正面案例真实存在。前文已经证明,市场确实需要更窄的 GPU 经纪和采购层;公开材料也仍支持有意义的牵引:100 多家供应商、1,000 多笔交易、超过 10 亿 GPU 小时活动、据报道约 $100M 的 2025 年运行率,以及公司为 AI 买方解决真实供给碎片化问题的迹象。在 2026 年,公开 AI 基础设施公司仍以异常宽、且有时偏高的倍数交易,15x 收入的标题数字并不会自动出局。 负面案例同样真实,而且对决策更关键。Andromeda 的估值仍压在一个没有公开审计、甚至没有清楚分类的分母上:它到底是净经纪经济性,还是总额转手收入?同一份材料仍缺公开 NRR、GRR、客户数、供应商集中度和股权结构细节;Nat Friedman 和 Daniel Gross 退居幕后之后,创始人连续性也明显弱了。这意味着,只有在隐藏变量都朝有利于投资人的方向发展时,当前价格才可能公平。没有这些证据,正确姿态是保持接触,坚持价格和条款纪律,不要把一门真实生意误当作已经完全承保的生意。[CV004, CV007, CV009, CV032, CV035, CV036]
| 决策字段 | 当前观点 | 依据 |
|---|---|---|
| 建议 | 跟踪 | 保持跟进,但不能仅凭公开证据就把 2026 年 3 月价格视为明显有吸引力。 |
| 确信度 | 中 | 估值锚是真实的,可比公司样本也能用,但关键分母和股权结构输入仍未披露。 |
| 风险评级 | 高 | 收入质量、集中度、治理和融资条款风险都可能很快伤到投资论点。 |
| 估值立场 | 偏高 | 按披露的 2025 年 run-rate 约 15x,在 2026 年 AI 基础设施里不算离谱,但仍假设隐含经济性有利。 |
| 进入纪律 | 价格和条款敏感 | 只有在尽调验证净经济性,且文件没有显示惩罚性优先条款或债务包袱时才投资。 |
| 持有 / 退出姿态 | 靠增长消化倍数 | 可支撑的情形是中期持有,而不是近期加价或轻松战略退出。 |
这份摘要刻意对价格敏感,而不只看公司质量;最大未解变量是已披露收入锚之下的经济质量。
[CV007, CV032, CV036, CV042, CV043, CV044]| 论点 | 当前证据 | 哪些证据会改变观点 |
|---|---|---|
| 中立经纪切入点 | 官方流程和独立报道支持一个真实需求:多供应商寻源、资质审核和统一签约。 | 证明客户续约和扩张,是因为 Andromeda 在速度、结果或经济性上明显胜过直接采购。 |
| 真实商业规模 | 公开报道支持 ~$100M 2025 年 run-rate,加上 100+ 家供应商、1,000+ 笔交易和超过 10 亿 GPU-hour 活动量。 | 经审计的已确认收入桥接和客户数量披露,会把规模从「看起来可信」变成「可以承销」。 |
| 轻资产上行 | 如果 Andromeda 不拥有大部分供给也能赚净价差,这个模式相对资本更重的云厂商可能值得溢价。 | 拿出 principal-agent 会计处理、毛利率和营运资本机制,证明收入在经济上干净。 |
| 反论点:分母质量仍未知 | 公开来源仍未显示,披露收入基数到底是总额转手收入,还是净经纪经济性。 | 管理层认证的收入确认备忘录和 cohort 经济性,可能显著提高确信度。 |
| 反论点:耐久性和多云风险仍未解决 | 具名客户证据很薄,没有公开 NRR 或 GRR,且至少一个早期用户正明显转向 Google Cloud 扩容。 | 留存、集中度和钱包份额数据能说明 Andromeda 是黏性平台,还是好用的过渡层。 |
| 反论点:治理和融资不透明度仍高 | 创始人似乎已不再活跃,控制权地图很薄,公开融资报道也缺少证券条款和优先权细节。 | 董事会、继任和融资文件会显示,名义估值能否干净映射到可投资经济性。 |
投资论点最强的地方是真实市场需求和表观规模;反论点最强的地方是隐藏经济性、隐藏集中度和隐藏条款。
[CV001, CV003, CV004, CV008, CV009, CV012]判断仍偏谨慎:真实牵引力和 2026 年可比公司行情虽有支撑,但经济性、集中度和条款仍是隐藏变量。
[CV001, CV004, CV007, CV035, CV037, CV038]Andromeda 在市场真实度和潜在资本效率上得分较好,但公开证据充分性和下行保护明显较弱。
得分是以引用的公开证据为锚的 0-10 序数判断,不是管理层披露的 KPI。
[CV004, CV009, CV013, CV016, CV032, CV042]8.2 融资背景和可比公司纪律
本章最难的地方在于,当前价格必须同时放进两套很不同的参照系里判断。一方面,按正常承保标准,Andromeda 仍然披露不足。公开来源集中指向 2026 年 3 月 Paradigm 支持的一轮融资,估值 $1.5B,Paradigm 总投资约 $60M;但它们没有披露该轮是否带有优先权、可转债、老股转让,或任何会实质改变标题数字之下经济性的投资人保护条款。SEC 的公开 Form D 数据基础设施存在,但本轮保留的公开证据仍没有为 Andromeda 本身找到清晰的证券类型或优先清偿栈桥接。 另一方面,2026 年公开市场里的 AI 和云参照,并没有给投资人一个简单的「私有市场显然太贵」答案。CoreWeave 在 StockAnalysis 上约为 8.7x P/S,但拥有更重的自有基础设施模式;DigitalOcean 到 2026 年 7 月约为 17.4x TTM 收入,尽管它是更宽的云业务;Nebius 则以极端的三位数收入倍数交易,反映的更多是 AI 基础设施期权价值,而不是当下现金收益纪律。这个宽区间不能证明 Andromeda 便宜;它证明,当市场相信增长和战略位置真实存在时,仍愿意为 AI 基础设施稀缺性付钱。因此,估值问题更窄也更尖锐:Andromeda 据报收入的经济性,是否比公开材料目前允许投资人核验的程度更干净、更持久、更不易受集中度拖累?[CV006, CV007, CV008, CV018, CV019, CV024]
| 可比对象 | 指标 | 倍数 / 估值 / 状态 | 相关性 | 局限 |
|---|---|---|---|---|
| Andromeda(标的) | 2026 年 3 月私募标记估值,对比披露的 2025 年 run-rate / 2024 年收入下限 | $1.5B;约为披露 2025 年 run-rate 的 ~15x,约为 2024 年 >$50M 下限的 ~30x | 正是本次承销要回答的问题。 | 收入口径、总额 vs 净额处理和融资条款仍未披露。 |
| CoreWeave | 2026 年 6 月 30 日 P/S 与 TTM 收入 | ~8.7x P/S,基于 ~$6.23B TTM 收入;市值 ~$54.3B | 最好的公开纯 AI 云可比,反映当前投资者胃口。 | 自有基础设施、take-or-pay 合同和极高集中度,让它成为非常不同的模式。 |
| DigitalOcean | 2026 年 7 月市值 / TTM 收入 | ~17.4x 收入;基于 ~$0.94B TTM 收入,市值 ~$16.38B | 有用的轻资产公开云参照,且有备案支持的多元化和合同结构。 | 更广泛的 SMB 云平台,披露更强,市场平台不透明度低得多。 |
| Nebius Group | 2026 年 7 月市值 / 当前收入 | ~135x 收入;基于 ~$0.52B 收入,市值 ~$70.11B | 显示 2026 年市场愿意为 AI 基础设施期权价值拉伸到多远。 | 投机性太强,垂直差异也太大,不能当作干净可比。 |
| Amazon / AWS 语境 | 2026 年 7 月全公司市值 | ~$2.563T 市值 | 只适合作为现有巨头的信任、采购和捆绑上限,说明它们如何压制经纪商价值。 | 全公司语境,不是中立 GPU 中介的直接估值倍数。 |
这是本章明确选出的完整可比篮子:标的标记估值,加上三个公开 AI / 云参照和一个现有巨头语境行。
[CV018, CV019, CV024, CV025, CV026, CV029]8.3 情景区间、下行触发因素和价格敏感性
牛市、基准和熊市情景,更多由营收线之下的东西驱动,而不是由顶线故事驱动。牛市情景不是 Andromeda 变成超大规模云厂商;而是当前约 $100M 规模大多反映的是净额、高价值的经纪经济性,建立在有黏性的采购和运营关系上,供应商和客户集中度可控,公司还能继续把供给碎片化转化成持久战略位置。在这种情况下,高端 AI 基础设施倍数仍能站住,当前进入价格也能成立。 基准情景更克制,也更符合公开证据。Andromeda 仍是一家真实且有用的公司,但市场发现,总额转手收入、客户多云并用、高接触服务负担和融资复杂度中的某种组合,让业务无法享受高端私有市场倍数偏爱的那种干净软件式持久性。在这个世界里,当前估值只算勉强站得住,主要还要靠公司长进估值。熊市情景很直接:如果据报收入基础的经济性弱于表面,如果直连云厂商补上工作流缺口,如果客户或供应商集中度很高,或者治理和融资不透明迫使公司接受不利于投资人的条款,下行空间会迅速打开,因为当前价格给分母失望留下的余地比叙事暗示的更小。[CV012, CV013, CV015, CV017, CV032, CV034]
| 情景 | 概率信号 | 收入假设 | 估值逻辑 | 隐含股权价值 |
|---|---|---|---|---|
| 乐观 | 管理层证明当前规模大多来自净额、高利润率的经纪经济性,同时集中度可控,贷款方 / 供应商杠杆增强。 | ~$120M-$150M 近期确认收入或等同 run-rate,且续约表现强 | 稀缺 AI 基础设施经纪商可拿 ~18x-24x 收入倍数,前提是信任和耐久性改善 | $2.2B-$3.6B |
| 基准 | 业务真实且仍在增长,但留存、集中度和融资条款上的公开缺口持续太久,压住热情。 | ~$100M-$120M,经济质量混合但可接受 | 仍享有溢价、但因证据不足打折的 ~10x-15x 收入倍数 | $1.0B-$1.8B |
| 悲观 | 收入被证明是总额口径或薄利率,直接云厂商补上流程缺口,或集中度 / 治理问题削掉溢价。 | ~$80M-$100M,隐含净经济性更低、耐久性更弱 | ~5x-8x 收入,更接近更严苛的基础设施或经纪商纪律区间 | $0.4B-$0.8B |
| 当前标记估值 | 当前公开锚已经假设公司至少落在基准情景上半区。 | ~$100M 披露的 2025 年 run-rate | $1.5B 名义估值 | $1.5B |
这些情景是公开证据下的纪律工具,不是管理层指引;由于收入质量、集中度和融资条款仍不透明,区间被拉宽。
[CV004, CV005, CV032, CV036, CV037, CV038]| 触发条件 | 阈值 | 对投资论点的传导 | 行动含义 |
|---|---|---|---|
| 领导连续性失效 | Wil Moushey 离任,或在没有可信继任与董事会地图的情况下失去运营控制 | 公开的创始人 / 运营者叙事失去剩下的连续性锚。 | 暂停,或从第一性原理重新核算。 |
| 收入质量低于预期 | 尽调显示披露收入大多是总额转手收入,或利润率显著低于投资者假设 | 即便名义收入真实,有效估值倍数也会急剧上升。 | 按更低净收入或更低倍数假设重新给公司估值。 |
| 集中度偏高 | 少数客户或供应商主导 GMV、确认收入或容量,替代路径薄弱 | 中立性和多元化不再是护城河,而变成口号。 | 下调确信度、压缩倍数,并重估下行。 |
| 直接云替代上升 | 具名客户越来越多地把耐久工作负载迁往直接供应商或内部采购 | 经纪层变成过渡便利,而不是耐久基础设施。 | 削减乐观情景概率,并把基准情景视为上限。 |
| 融资条款对投资者不友好 | 优先权、可转债或合同绑定债务,实质性压低新股权或普通股等价上行 | 名义估值不再反映可投资经济性。 | 拒绝当前条款,或要求重置。 |
这些触发条件旨在直接转化为估值受损,而不是泛泛的运营担忧。
[CV009, CV037, CV041, CV045, CV046, CV047]只有收入质量强到让投资者愿意基于接近报道称约 ~$100M 的收入基数给出溢价倍数,当前估值才更容易站住。
敏感性柱状图只用公开快照做简单估值算术,不调整稀释、总额 / 净额会计处理或优先权悬置。
[CV004, CV018, CV019, CV024, CV025, CV032]仅看公开证据,只有 Andromeda 的隐藏经济性落在投资者有利一侧,区间中枢才会围绕当前估值。
区间来自公开锚点和简单倍数纪律构建的情景输出,不追求 DCF 精度。
[CV032, CV036, CV037, CV039, CV040, CV041]8.4 退出准备度和真正影响决策的尽调问题
公开材料足以说明 Andromeda 为什么值得关注,却不足以说明新投资人为什么应无摩擦接受当前价格。缺失材料不是装饰性的。投资人仍需要一座经审计或由管理层认证的收入桥,区分总市场交易量、确认收入和净价差;还需要清晰了解客户和供应商集中度、真实续约和留存数据,以及定义 2026 年 3 月标题估值在普通股等价经济性上到底意味着什么的融资文件。没有这些材料,当前估值更应被描述为可能合理但证据不足。 这也限制了退出承保信心。能够支撑的持有逻辑,是中期长进倍数的案例,而不是近期上调估值或干净战略退出的案例。超大规模云厂商和直接 AI 云仍主导信任、采购关系和基础设施深度,而 Andromeda 周围公开可见的创始人板凳厚度已比最初叙事更薄。如果尽调能证明 Andromeda 拥有一个有黏性、多元化、具备净经济性且融资有纪律的经纪层,建议可以上调。否则,正确反应是耐心:跟踪公司、保持访问通道,并要求业务用比公开记录更硬的证据赚到高端估值。[CV008, CV016, CV017, CV039, CV042, CV043]
| 主题 | 缺失证据 | 为何重要 | 负责人或尽调路径 |
|---|---|---|---|
| 经审计收入桥接与会计政策 | 确认收入、GMV、take rate、principal-agent 政策,以及 2024 年至当前 2026 年退出 run-rate 的 cohort 桥接 | 如果投资者不知道收入实际代表什么,当前倍数就没有意义。 | 财务尽调、审计师材料包和财务总监备忘录。 |
| 客户留存与集中度 | NRR、GRR、流失、客户数量、头部账户占比,以及分细分市场支出扩张 | 分母质量之外,耐久性是最大未解变量。 | RevOps 导出、董事会 KPI 材料和客户访谈。 |
| 供应商集中度与 SLA 韧性 | 头部供应商占比、替代矩阵、利用率历史、故障历史,以及 SLA 或赔付信用框架 | 中立聚合只有能顶住供应商故障或重新定价,才有价值。 | 运营尽调和供应商侧合同审查。 |
| 股权结构与融资条款 | 证券类型、清算优先权、参与权、反稀释条款、老股出售、补充协议,以及任何与客户合同绑定的债务 | 名义估值可能显著高估真实股权价值。 | 律师审查融资文件和股权结构。 |
| 治理与继任 | 当前董事会、投票控制、创始人权利、运营者继任和升级责任归属 | 治理不透明是这个投资论点最快破裂的路径之一。 | 董事会材料包、章程文件和管理层访谈。 |
| 信任、合规与企业就绪材料包 | 出口管制、制裁、隐私、安全、事件和 DPA 材料,以及真实认证状态 | 退出就绪和溢价倍数延续都需要不止一个稀疏的信任页面。 | 法务、安全和企业客户尽调。 |
如果这些问题无法清楚回答,正确反应是继续跟踪公司,而不是追当前价格。
[CV008, CV013, CV016, CV017, CV042, CV043]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,都应直接向管理层并通过一手文件核实。
证据索引
| 编号 | 陈述 | 可信度 | 来源 |
|---|---|---|---|
| CO001 | Official home and social copy present Andromeda as a compute-market company that helps AI teams access high-performance compute quickly and at scale. | 高 | SO001, SO008 |
| CO002 | Official insights and privacy materials describe Andromeda as marketplace or market infrastructure connecting compute buyers with infrastructure providers. | 高 | SO002, SO006 |
| CO003 | Andromeda's public buyer and provider flows show a two-sided product in which customers specify workload needs and providers submit GPU, networking, location, and availability details. | 高 | SO003, SO004 |
| CO004 | Official copy says Andromeda benchmarks supply, standardizes contracts, consolidates billing, and provides a support and observability layer around third-party compute. | 高 | SO001, SO002, SO022 |
| CO005 | Official footers and the privacy policy identify the operating entity as Andromeda Cluster, Inc. | 高 | SO001, SO005, SO006 |
| CO006 | Built In, the company's X profile, and Gaebler all point to San Francisco as Andromeda's current operating base, with Gaebler listing 228 Grant Ave as a mailing address. | 高 | SO008, SO009, SO024 |
| CO007 | Andromeda was already operating as the Andromeda Cluster by late November 2023, and its March 2026 launch post said the team had spent three years building the market infrastructure for compute. | 高 | SO002, SO008, SO021 |
| CO008 | Reviewed public sources do not disclose a precise legal incorporation date, so the best-supported founding framing is late-2023 project origin rather than an exact day. | 中 | SO001, SO002, SO021 |
| CO009 | Built In and Gaebler both describe Andromeda Cluster as founded by Nat Friedman and Daniel Gross. | 中 | SO009, SO024 |
| CO010 | Official insights and Upstarts both indicate the business began as a cluster serving Nat Friedman and Daniel Gross's venture portfolio before broadening into a wider compute market network. | 高 | SO002, SO021 |
| CO011 | Upstarts reports that Daniel Gross recruited Wil Moushey in late November 2023 to step in and run the Andromeda project. | 中 | SO021 |
| CO012 | Raising.fi and Gaebler identify Wil Moushey as Andromeda's CEO in March 2026 coverage of the financing. | 中 | SO023, SO024 |
| CO013 | DCD and CB Insights describe Nat Friedman as a former GitHub CEO and NFDG cofounder. | 中 | SO025, SO027 |
| CO014 | Daniel Gross's personal site and DCD identify Gross as a veteran AI or search operator and NFDG cofounder, and his site now says he runs compute for Meta. | 高 | SO025, SO026 |
| CO015 | Reviewed official pages do not publish a board roster or a formal executive-team page beyond public job postings and founder or CEO mentions. | 中 | SO001, SO005, SO006, SO011 |
| CO016 | Built In and Ashby listings show active hiring across legal, finance, partnerships, procurement, compute markets, solutions engineering, software engineering, and site reliability. | 高 | SO010, SO011, SO012, SO013, SO014, SO015, SO016, SO017, SO018, SO019, SO020 |
| CO017 | The Commercial Counsel role shows the company is building explicit capability around corporate governance, privacy and export compliance, vendor matters, fundraising, and M&A. | 高 | SO010, SO012 |
| CO018 | The Strategic Compute Finance Lead role mentions lender relationships and capital allocation, implying financing complexity without proving a public debt facility. | 中 | SO010, SO013 |
| CO019 | Built In lists 17 employees while Upstarts describes a team of about 20, so the strongest public workforce signal is a high-teens team rather than one exact number. | 中 | SO009, SO021 |
| CO020 | Key-person dependence is high because the public record centers current operating execution on Moushey while the founder network still dominates the company's origin and sponsor narrative. | 中 | SO009, SO021, SO025 |
| CO021 | Upstarts says Nat Friedman and Daniel Gross no longer have an active relationship with Andromeda after moving to Meta, making founder withdrawal the clearest public leadership change. | 中 | SO021, SO026 |
| CO022 | Upstarts says Paradigm's March 2026 financing valued Andromeda at $1.5 billion and was the equivalent of a Series A, while Gaebler records the same event as undisclosed venture equity. | 中 | SO021, SO024 |
| CO023 | Upstarts, Raising.fi, and Gaebler each indicate Paradigm's total investment in Andromeda is $60 million. | 中 | SO021, SO023, SO024 |
| CO024 | Gaebler records the March 18, 2026 event as a $60 million venture-equity round with Paradigm as investor. | 中 | SO023, SO024 |
| CO025 | SiliconANGLE says the exact size of the newly closed March 2026 tranche was unclear even as Paradigm's total investment to date reached $60 million. | 中 | SO021, SO022 |
| CO026 | Upstarts and SiliconANGLE both report that Andromeda's annualized revenue run rate reached about $100 million in 2025 after exceeding $50 million in 2024. | 高 | SO021, SO022 |
| CO027 | Upstarts and SiliconANGLE both say Andromeda has operated profitably since launch. | 高 | SO021, SO022 |
| CO028 | Reviewed public sources do not cleanly disclose lifetime capital raised because NFDG's more than $100 million precursor cluster build appears separate from Paradigm's disclosed company investment. | 中 | SO021, SO025 |
| CO029 | Reviewed public sources do not disclose whether the March 2026 financing included secondaries or special governance terms. | 中 | SO021, SO022, SO024 |
| CO030 | No public credit facility is disclosed in reviewed sources even though the finance role mentions lender relationships and Moushey describes credit as a bottleneck in the market. | 中 | SO013, SO021 |
| CO031 | Official counters and SiliconANGLE say Andromeda works with more than 100 compute providers and has processed more than 1,000 transactions. | 高 | SO001, SO022 |
| CO032 | Official home and insights pages say Andromeda has billions of GPU-hour liquidity or a billion-plus GPU-hours of experience and now operates dozens of clusters across many providers. | 高 | SO001, SO002 |
| CO033 | Upstarts says Andromeda manages tens of thousands of GPUs at any given time and hopes to reach as many as 100,000 GPUs under management in 2026. | 中 | SO021 |
| CO034 | Upstarts says many of Andromeda's initial users came from AI Grant or NFDG portfolio companies such as ElevenLabs and Pika, while official and Built In copy now frames the buyer base more broadly as leading AI labs and startups. | 中 | SO009, SO021 |
| CO035 | Public sources identify customer type but not a numeric customer count: they describe leading AI labs, notable startups, and labs rather than disclosing a total number of accounts. | 高 | SO001, SO009, SO021, SO022 |
| CO036 | Upstarts says Andromeda recently opened a San Francisco office, while official career and jobs pages show US, North America, and global remote hiring around that base. | 高 | SO005, SO011, SO021 |
| CO037 | Official buyer or provider workflows and jobs copy show the product is operationally complex, covering bare-metal, Kubernetes or Slurm clusters, benchmarking, pricing, and deal structuring rather than only fixed cloud resale. | 高 | SO002, SO003, SO004, SO010 |
| CO038 | DCD reported in June 2025 that the Andromeda Cluster was already available to non-NFDG companies at $2.40-$3.00 per GPU-hour, showing commercialization before the March 2026 public brand launch. | 中 | SO008, SO025 |
| CO039 | DCD said Meta's talks around partially acquiring NFDG left what it would mean for Andromeda unclear, creating a public continuity risk tied to founder-side changes. | 中 | SO025 |
| CO040 | The March 18, 2026 X post said “Today we're announcing Andromeda” after three years building the market infrastructure for compute, anchoring the external launch date. | 高 | SO001, SO008 |
| CO041 | The privacy policy's March 16, 2026 revision date is the clearest public legal or compliance milestone in the reviewed official set. | 中 | SO006 |
| CO042 | Company materials and job postings consistently frame providers, customers, and standardized contracts as the core business, implying Andromeda is building a neutral market layer rather than a captive single-provider cloud. | 高 | SO001, SO002, SO004, SO010 |
| CO043 | Andromeda exposes a public trust center, but the public trust surface is sparse compared with the company's revenue and funding narrative. | 中 | SO006, SO007 |
| CM001 | Andromeda publicly positions itself as market infrastructure that connects AI builders and infrastructure providers through benchmarking, standardized contracts, routing, and operations rather than as a simple first-party GPU lessor. | 高 | SM001, SM003, SM004 |
| CM002 | The retained top-down GPUaaS definition is on-demand virtualized GPU access for machine learning, high-performance computing, and advanced visualization without the customer owning the hardware. | 中 | SM007 |
| CM003 | The retained GPUaaS market definition includes managed GPU hosting, workload orchestration, GPU resource provisioning, and training and optimization services. | 中 | SM007 |
| CM004 | Andromeda says the supply it aggregates can come from telco and crypto data centers, legacy MSPs, sovereign clouds, startup clouds, and even other labs' balance-sheet capacity. | 高 | SM001, SM006 |
| CM005 | Owned-cluster capex, chip purchases, data-center construction, and tokenized model-API consumption should be treated as excluded or adjacent spend rather than Andromeda's direct brokered-compute market. | 中 | SM001, SM007, SM018 |
| CM006 | Status-quo substitutes for Andromeda include hyperscaler accelerator instances, specialist GPU clouds, direct bilateral provider contracts, owned clusters, and spot or interruptible marketplaces for tolerant workloads. | 高 | SM013, SM014, SM017, SM018, SM019, SM020, SM021 |
| CM007 | The Business Research Company estimates the global GPUaaS market at $7.39 billion in 2026, up from $5.8 billion in 2025, reaching $19.34 billion in 2030. | 中 | SM007 |
| CM008 | S&P Global discloses that its 451 Research GPUaaS market monitor examines revenue and growth expectations from 22 providers worldwide. | 中 | SM008 |
| CM009 | S&P Global says strong GPU demand and difficulty sourcing chips pushed buyers toward GPUaaS specialists and alternative providers when hyperscalers responded relatively slowly. | 中 | SM008 |
| CM010 | Andromeda's customer intake form asks for GPU type, volume, start date, and number of weeks, signaling that the core demand unit is planned cluster capacity rather than casual ad hoc experimentation. | 中 | SM002 |
| CM011 | Andromeda's provider intake form asks sellers to disclose GPU-hour pricing, network fabric, node resources, location, and minimum bookable terms, showing that market fit depends on infrastructure quality and contract shape as well as chip type. | 中 | SM003 |
| CM012 | Data Center Dynamics reported that up to 2,000 H100s were available through Andromeda within hours at $2.40-$3.00 per GPU-hour. | 中 | SM005 |
| CM013 | The public Andromeda inventory quote implies a lower-bound annualized raw spend pool of about $42.0 million to $52.6 million if 2,000 H100s were fully utilized for a year at the quoted hourly rate. | 中 | SM005 |
| CM014 | RunPod's official pricing lists H100 PCIe at $2.89 per hour and H100 SXM at $3.29 per hour, with A100 capacity below $1.50 per hour. | 中 | SM013 |
| CM015 | Lambda's official pricing lists H100 instances at $3.99 per GPU-hour and offers clusters from 16 to 2,000+ H100 or B200 GPUs. | 中 | SM014 |
| CM016 | CoreWeave markets H100 and H200 clusters as purpose-built AI infrastructure with bare-metal Kubernetes, high-performance networking, expert support, and pricing starting at $2.23 per hour on its H100/H200 product page. | 中 | SM015, SM024 |
| CM017 | CoreWeave's public HGX H100 pricing is $49.24 per hour on-demand and $19.71 per hour spot for an 8-GPU node, implying about $6.16 per H100-hour before discounts. | 中 | SM016 |
| CM018 | Google Cloud's public A3 High and A3 Mega cards imply roughly $11.06 to $11.68 per H100-hour before discounts when their 8-GPU node prices are divided by eight. | 中 | SM019, SM020 |
| CM019 | Azure's ND H100 v5 starts with a single VM containing eight H100 GPUs, 96 vCPUs, 1.9 TB RAM, and 3.2 Tbps of interconnect bandwidth, scaling to thousands of GPUs. | 中 | SM021 |
| CM020 | AWS says P5, P5e, and P5en instances can scale in UltraClusters up to 20,000 H100 or H200 GPUs for training and inference workloads. | 中 | SM018 |
| CM021 | Presenc says Q2 2026 H100 rental pricing is roughly $1.80-$3.50 per hour on-demand with spot as low as $1.20 and direct-purchase lead times down to 6-12 weeks. | 中 | SM009 |
| CM022 | SesameDisk argues that 2026 buyer usefulness depends on accelerator type, region, quota, and queue time rather than abstract GPU availability. | 中 | SM010 |
| CM023 | Cyfuture says H100 pricing moved from about $8-$10 per hour in the shortage period to a 2026 range spanning roughly $1.38 per hour at neo-cloud floors and about $14.19 per hour in premium hyperscaler tiers. | 中 | SM011 |
| CM024 | GPUHosted characterizes RunPod and Modal as serverless leaders, Vast as a low-cost marketplace, and Lambda and CoreWeave as specialized AI-training providers, reinforcing segmentation by workload and service model. | 中 | SM012 |
| CM025 | Vast's official pricing emphasizes on-demand, interruptible, and reserved tiers, with interruptible capacity 50%+ cheaper and reserved terms of one, three, or six months. | 中 | SM017 |
| CM026 | Andromeda explicitly argues that compute is not fungible and that buyers need benchmarking, standardized contracts, observability, SLAs, and support before third-party capacity is trustworthy for frontier workloads. | 高 | SM001, SM003 |
| CM027 | CoreWeave's benchmark brief claims more than 50% MFU, 20% more useful compute per dollar, and 10x greater reliability on 1,024 H100 GPUs, illustrating why reliability and orchestration can justify a premium over raw spot access. | 高 | SM015, SM025 |
| CM028 | NVIDIA's H100 materials emphasize NVLink, InfiniBand, and large-scale cluster communication, supporting the view that architecture details materially change workload value even when the GPU label is the same. | 高 | SM018, SM023 |
| CM029 | Frontier labs and large model builders are the clearest first-buy segment because Andromeda says it serves leading AI companies and SiliconANGLE reports that notable customers can spend $250 million to $500 million per year on infrastructure. | 中 | SM001, SM006 |
| CM030 | Enterprise AI platform teams and regulated HPC users are an adjacent segment because AWS, Azure, and Google all market their highest-end GPU fleets around deep learning, analytics, pharma, weather, and financial modeling use cases. | 中 | SM018, SM019, SM021, SM022 |
| CM031 | In this market, the end user is typically a researcher, training engineer, inference engineer, or platform team, while the buyer is an infrastructure lead and the payer is a centralized cloud, procurement, or finance budget owner. | 中 | SM002, SM003, SM008 |
| CM032 | The adoption path for Andromeda-style brokerage runs from workload specification to supply benchmarking to contract standardization to deployment-substrate choice and then ongoing observability-backed operations. | 高 | SM001, SM002, SM003 |
| CM033 | S&P says AI technology and cloud infrastructure have ranked among the highest enterprise technology spending-intent categories for most of the last two years. | 中 | SM008 |
| CM034 | Andromeda says AI research cycles move in weeks while data centers take twelve to twenty-four months to build, making time-to-capacity a core market driver. | 中 | SM001 |
| CM035 | Presenc says 2026 supply bottlenecks have shifted across HBM3e memory, CoWoS packaging, power delivery, and rack deployment rather than disappearing entirely. | 中 | SM009 |
| CM036 | SesameDisk says quotas, regional limits, queue times, and enterprise relationship status often matter more than headline list pricing when buyers actually need capacity on schedule. | 中 | SM010 |
| CM037 | Falling unit prices do not eliminate the broker opportunity, but they do shift value from scarce hardware access alone toward routing, quality certification, reliability, and contracting. | 中 | SM001, SM011, SM025 |
| CM038 | No retained public source isolates an Andromeda-specific SAM, SOM, or take rate, so any precise company-level market model would require management disclosure rather than extrapolation from broad TAM figures. | 中 | SM005, SM007, SM008 |
| CM039 | Broad GPUaaS TAM estimates and current H100 price bands are different market lenses: the first measures category revenue while the second shows unit-price deflation and service-level fragmentation. | 中 | SM007, SM009, SM011 |
| CM040 | Public price cards do not support a clean apples-to-apples take-rate benchmark because they mix spot listings, self-serve on-demand nodes, reserved capacity, and enterprise bundles with different networking and support assumptions. | 中 | SM009, SM011, SM013, SM016, SM019 |
| CM041 | Spot and interruptible inventory are real substitutes for fault-tolerant batch or experimentation workloads, but they are not equivalent to stable reserved clusters for launch-critical training or customer-facing inference. | 中 | SM001, SM009, SM017 |
| CM042 | The most valuation-relevant market conclusion is that Andromeda is operating in a real and growing outsourced-GPU category, but its upside depends on monetizing trust, coordination, and operations in a fragmented market rather than simply renting scarce chips. | 中 | SM001, SM007, SM008, SM011 |
| CP001 | Andromeda publicly positions itself as a neutral market layer that sources, benchmarks, certifies, and standardizes third-party compute across 100+ providers with one invoice and one support channel. | 中 | SP001, SP002 |
| CP002 | Andromeda's retained public trust surface is materially thinner than the compliance-heavy materials published by hyperscalers and several direct or adjacent rivals. | 中 | SP003, SP011, SP026, SP030, SP032, SP033, SP034 |
| CP003 | The 2026 buyer landscape around Andromeda spans direct provider-owned clouds, incumbent hyperscalers, orchestration layers, serverless or elastic adjacencies, and internal-build substitutes. | 中 | SP004, SP009, SP014, SP020, SP024, SP028, SP032, SP035 |
| CP004 | CoreWeave overlaps directly with Andromeda's target buyer by combining AI-native infrastructure, Kubernetes access, storage, security, observability, and expert services in one provider-owned cloud. | 中 | SP004, SP005 |
| CP005 | CoreWeave publishes granular on-demand and spot pricing and says committed usage can reduce on-demand prices by up to 60%, making it a visible enterprise benchmark for direct capacity buying. | 高 | SP005, SP006 |
| CP006 | CoreWeave's Q1 2026 results show unusual scale for a direct rival, with nearly $100 billion of revenue backlog and more than 1 GW of active power. | 中 | SP007 |
| CP007 | Fitch says Microsoft represented 62% of CoreWeave's 2024 revenue and the top two customers represented 77%, showing high concentration beneath the growth story. | 中 | SP008 |
| CP008 | Fitch also highlights leverage and lease-term mismatch risk at CoreWeave, with leases typically running 3 to 15 years versus customer contracts of 3 to 5 years. | 中 | SP008 |
| CP009 | Lambda competes as a direct AI cloud with self-serve instances, 1-Click clusters, superclusters, and a public trust posture built around single-tenant architecture and SOC 2 Type II. | 高 | SP009, SP010, SP011 |
| CP010 | Lambda's published pricing makes it a low-friction direct alternative, including self-serve first-come access and public H100-class hourly rates. | 中 | SP010 |
| CP011 | Public scale and funding signals show Lambda is not a niche point player: BusinessWire said it had 100,000+ cloud sign-ups and 5,000+ hardware or private-cloud customers, while Verdict said the 2025 Series D brought total equity raised to $863 million. | 中 | SP012, SP013 |
| CP012 | Vast.ai is a direct substitute for many price-sensitive buyers because it presents itself as a real-time GPU marketplace across 20,000+ GPUs and 40+ data centers rather than as a premium managed cloud. | 高 | SP014, SP015, SP016 |
| CP013 | Vast.ai's marketplace mechanics lower switching costs because providers retain pricing and contract control and the platform explicitly markets no long-term contracts. | 高 | SP015, SP016 |
| CP014 | SambaNova is better understood as a vertically integrated inference and sovereign-AI alternative than as a neutral multi-provider broker, emphasizing fast inference, agentic AI, and its own systems. | 高 | SP017, SP019 |
| CP015 | SambaNova's 2026 strategic direction is to use a $350 million Series E round and an Intel collaboration to expand SN50 production, cloud capacity, and go-to-market reach. | 中 | SP019 |
| CP016 | NVIDIA Run:ai competes on orchestration rather than supply ownership, promising GPU pooling and policy-driven resource management across public cloud, private cloud, hybrid, and on-prem environments. | 高 | SP020, SP021 |
| CP017 | NVIDIA ownership broadens Run:ai's distribution but weakens its neutrality as an independent broker-adjacent layer. | 中 | SP020, SP022 |
| CP018 | The European Commission said Run:ai does not hold a significant position in GPU orchestration today and noted customers can use alternative software or build orchestration features themselves. | 中 | SP023 |
| CP019 | RunPod is a credible low-friction adjacent alternative because it combines pods, serverless, and clusters, claims 1M+ developers, and markets audited or partner-backed compliance controls. | 高 | SP024, SP025, SP026 |
| CP020 | RunPod's public hourly and per-second packaging makes it especially attractive for bursty inference or overflow workloads rather than for brokered large-cluster procurement. | 高 | SP025, SP027 |
| CP021 | Modal competes for elastic AI workloads through cross-cloud routing, autoscaling from 0 to 1000+ GPUs, serverless per-second pricing, and startup credits. | 高 | SP028, SP029, SP031 |
| CP022 | Modal's security materials indicate a stronger documented trust posture than many neo-clouds, including SOC 2 Type II and HIPAA-capable enterprise controls. | 高 | SP029, SP030 |
| CP023 | AWS, Google Cloud, and Azure retain the clearest public trust or regulatory lead because each publishes broad audited compliance catalogs, third-party attestations, and regulated-industry coverage. | 高 | SP032, SP033, SP034 |
| CP024 | Internal build remains a credible substitute for the largest buyers because NVIDIA markets DGX SuperPOD as a turnkey AI data center platform scaling to tens of thousands of GPUs. | 中 | SP035 |
| CP025 | Internal build lowers dependence on brokers or clouds but shifts the burden to capital, deployment, and operations complexity. | 中 | SP035 |
| CP026 | Most direct alternatives own or directly control supply, while Andromeda's defining difference is aggregating fragmented third-party capacity and standardizing the commercial layer around it. | 中 | SP001, SP004, SP009, SP014 |
| CP027 | Andromeda is structurally strongest when the buyer values supplier discovery, certification, SLA normalization, and unified billing more than a deep relationship with one direct cloud. | 中 | SP001, SP002, SP003 |
| CP028 | Andromeda is structurally weaker where the buyer prioritizes published compliance evidence, existing enterprise budget rails, or immediate self-serve instances from a known provider. | 中 | SP003, SP009, SP011, SP023, SP030, SP032, SP033, SP034 |
| CP029 | Switching costs look moderate rather than hard because buyers can combine direct clouds, serverless providers, or orchestration layers without abandoning standard ML tooling. | 中 | SP016, SP020, SP024, SP028, SP031 |
| CP030 | Multi-homing is commercially plausible because Run:ai is built to orchestrate heterogeneous environments while Vast, RunPod, and Modal rely on usage-based or low-commitment models. | 中 | SP015, SP016, SP020, SP021, SP025, SP028, SP031 |
| CP031 | Distribution power is strongest with CoreWeave and the hyperscalers because they pair direct infrastructure control with either large backlog and power scale or deeply embedded enterprise compliance and procurement channels. | 中 | SP007, SP023, SP032, SP033, SP034 |
| CP032 | Supply access is a double-edged sword for Andromeda: fragmentation creates demand for a broker, but provider-owned clouds and hyperscalers can bypass the broker and compress its take rate. | 中 | SP001, SP004, SP009, SP014, SP032 |
| CP033 | Price commoditization risk is real because CoreWeave, Lambda, Vast.ai, RunPod, and Modal all expose public price anchors or usage-based constructs that let buyers benchmark alternatives quickly. | 高 | SP006, SP010, SP015, SP025, SP027, SP029, SP031 |
| CP034 | CoreWeave's leverage and customer concentration are adverse evidence that even scaled direct rivals can have fragile economics beneath demand growth. | 中 | SP007, SP008 |
| CP035 | Run:ai's small current market position and bundling into NVIDIA are adverse evidence against assuming orchestration alone is a durable independent moat category. | 中 | SP022, SP023 |
| CP036 | Likely entrants into Andromeda-adjacent territory include direct clouds adding more orchestration, NVIDIA-linked software stacks, and providers that package their own compliance or qualification layers. | 中 | SP005, SP015, SP019, SP020, SP026 |
| CP037 | Andromeda's moat is durable only if neutral qualification, routing, and commercial normalization produce better economics or faster procurement than customers can get by going direct or building around standard orchestration tools. | 中 | SP001, SP002, SP020, SP033, SP035 |
| CI001 | Official Andromeda workflow describes a multi-provider procurement layer that benchmarks supply, standardizes contracts, and delivers one invoice plus one support channel to buyers. | 高 | SI001, SI003 |
| CI002 | Andromeda’s customer intake flow asks for GPU type, quantity, start date, and weeks of usage, which supports a reserved-workload procurement motion rather than a purely self-serve software checkout. | 高 | SI001, SI002 |
| CI003 | Andromeda’s provider intake flow asks suppliers for minimum bookable GPUs, minimum duration in weeks, and cluster details, implying that supply activation is contract-shaped and operationally specific. | 高 | SI001, SI003 |
| CI004 | Andromeda’s official site claims 100+ compute providers and 1,000+ completed transactions. | 高 | SI001, SI007 |
| CI005 | Andromeda’s insights page says the company has helped operate a constellation of dozens of clusters across many providers and more than a billion GPU-hours. | 高 | SI001, SI004 |
| CI006 | Andromeda’s insights essay says AI research moves in weeks while data centers take twelve to twenty-four months to build, framing speed and coordination as the company’s economic wedge. | 中 | SI004 |
| CI007 | Upstarts reported that Andromeda passed a 2025 revenue run rate of about $100 million. | 中 | SI006 |
| CI008 | Upstarts reported that Andromeda generated more than $50 million of revenue in 2024. | 中 | SI006 |
| CI009 | Upstarts reported that Andromeda has operated profitably since launch. | 中 | SI006 |
| CI010 | Upstarts described Andromeda’s customer sweet spot as companies with roughly $250 million to $500 million of annual compute spend. | 中 | SI006 |
| CI011 | SiliconANGLE independently reported that Andromeda’s annualized revenue run rate grew from about $50 million in 2024 to about $100 million in 2025 and that the company has been profitable since launch. | 中 | SI007 |
| CI012 | SiliconANGLE reported that the company intends to use its new funding to grow its customer base and that it claims more than 100 providers and more than 1,000 GPU transactions. | 中 | SI007 |
| CI013 | Data Center Dynamics reported that Andromeda’s earlier public cluster quote offered non-NFDG buyers up to 2,000 H100s at $2.40 to $3.00 per GPU-hour with access available within hours. | 中 | SI008 |
| CI014 | Andromeda’s careers page publicly shows open roles spanning partnerships, business staff, commercial counsel, procurement or supply chain, strategic compute finance, compute trading, solutions, and SRE. | 中 | SI005 |
| CI015 | No retained public source discloses an Andromeda list-price catalog, take rate, or formal discount schedule. | 中 | SI001, SI002, SI003, SI006, SI007, SI008 |
| CI016 | No retained public source discloses whether Andromeda recognizes revenue gross as principal or net as commission or spread, making revenue-quality interpretation incomplete. | 中 | SI001, SI002, SI003, SI024 |
| CI017 | CoreWeave’s public pricing page lists NVIDIA HGX H100 at $49.24 per hour on-demand, $19.71 per hour spot, and says committed usage can receive up to 60% discounts versus on-demand rates. | 中 | SI011 |
| CI018 | Lambda publicly lists H100 instances at $3.99 per GPU-hour and advertises 1-Click clusters that scale from 16 to 2,000+ GPUs plus reserved capacity through sales. | 中 | SI018 |
| CI019 | RunPod publicly lists H100 PCIe at $2.89 per hour and H100 SXM at $3.29 per hour across pods, serverless, and cluster products. | 中 | SI019 |
| CI020 | Vast.ai publicly offers on-demand, interruptible, and reserved pricing with per-second billing, 1-, 3-, or 6-month reserved terms, and no long-term contracts required. | 中 | SI020 |
| CI021 | Public H100 and adjacent price cards span from roughly $2.89 per GPU-hour on RunPod to $49.24 per GPU-hour on CoreWeave on-demand, implying that contract shape and managed-service content matter as much as nominal GPU type. | 高 | SI011, SI018, SI019, SI020 |
| CI022 | CoreWeave’s March 2026 10-Q says the company sells both committed contracts and on-demand access, and that committed contracts represented 98% of revenue in Q1 2026. | 中 | SI014 |
| CI023 | CoreWeave’s 2025 10-K says customers generally buy multi-year committed, take-or-pay contracts and that remaining performance obligations reached $60.7 billion with about a five-year weighted average contract duration at year-end 2025. | 中 | SI013 |
| CI024 | CoreWeave’s 2025 10-K says Microsoft was 67% of 2025 revenue, and that year-end liquidity included $3.127 billion of cash and cash equivalents plus $6.862 billion of total liquidity. | 中 | SI013 |
| CI025 | CoreWeave’s May 2026 results release said revenue backlog reached $99.4 billion and active power surpassed 1 gigawatt by March 31, 2026. | 中 | SI012 |
| CI026 | CoreWeave’s May 2026 results release said the company secured an $8.5 billion non-recourse DDTL 4.0 facility and closed a $2 billion NVIDIA equity investment. | 中 | SI012 |
| CI027 | CoreWeave’s May 15, 2026 8-K disclosed a $3.1 billion delayed draw term loan to finance customer-contract capex, priced at SOFR plus 4.50%, maturing in 2031, and tested to a 1.35x debt service coverage ratio. | 中 | SI016 |
| CI028 | Fitch’s July 2025 CoreWeave note said Microsoft represented 62% of 2024 revenue, the top two customers represented 77%, capital intensity was expected to peak in 2025, and lease terms of 3 to 15 years could outlast customer contracts of 3 to 5 years. | 中 | SI017 |
| CI029 | DigitalOcean’s 2025 10-K says pricing is primarily consumption-based, most customers are month-to-month, and some larger customers now sign committed minimum-spend contracts. | 高 | SI021, SI022 |
| CI030 | DigitalOcean’s 2025 10-K says it had about 21,000 digital native enterprise customers in 2025, ARR of $970 million, and no material customer concentration because its top 25 customers were only 7% of revenue. | 中 | SI021 |
| CI031 | DigitalOcean’s Q1 2026 10-Q says gross profit fell to 56% from 61% because data center expansion costs were incurred ahead of revenue ramp. | 中 | SI022 |
| CI032 | DigitalOcean’s Q1 2026 10-Q says property-and-equipment capex was $61.963 million in the quarter and that uncommenced data-center lease obligations totaled about $668.444 million with a 9.8-year weighted average term. | 中 | SI022 |
| CI033 | DigitalOcean’s 2025 10-K says remaining performance obligations were $134.085 million, with $72.799 million expected within 12 months, and management said cash and available credit should cover working capital, capex, and debt needs for at least the next 12 months. | 中 | SI021 |
| CI034 | DigitalOcean’s May 2026 8-K said its amended credit agreement added $112.5 million of revolving capacity and a $50 million larger letter-of-credit sublimit for working capital, capital expenditures, acquisitions, and refinancing. | 中 | SI023 |
| CI035 | Amazon’s 2025 10-K says the company recognizes gross revenue on inventory sales but only its net share of revenue on third-party seller service sales. | 中 | SI024 |
| CI036 | Amazon’s 2025 10-K says unearned revenue is driven largely by AWS prepayments and that long-term AWS-related performance obligations were about $244 billion with a 4.1-year weighted average life at December 31, 2025. | 中 | SI024 |
| CI037 | Amazon’s Q1 2026 10-Q says cash capital expenditures were $43.2 billion in the quarter and AWS sales were $37.587 billion, showing how capex-heavy direct-cloud ownership can be at hyperscale. | 中 | SI025 |
| CI038 | The retained official and press sources support a consultative, enterprise GTM motion centered on reserved workloads and provider matching rather than a low-touch, swipe-card software CAC loop. | 中 | SI001, SI002, SI003, SI006, SI008 |
| CI039 | Public evidence makes Andromeda look more asset-light than provider-owned GPU clouds, but only if the company is not pre-buying supply, extending guarantees, or warehousing significant working-capital risk off the public record. | 低 | SI001, SI003, SI017, SI022, SI025 |
| CI040 | Using a public headcount range of about 17 to 20 employees against a reported 2025 revenue run rate of about $100 million implies a rough revenue-per-employee range of about $5.0 million to $5.9 million. | 中 | SI006, SI009 |
| CI041 | Andromeda’s public traction file is strongest on revenue run-rate, provider breadth, transactions, and GPU-hour activity, but it does not disclose customer count, ARR, NRR, utilization, or concentration. | 中 | SI001, SI004, SI006, SI007 |
| CI042 | The most explicit public use-of-funds statement is that new capital will be used to grow Andromeda’s customer base, while careers data suggest some proceeds are also supporting finance, procurement, partnership, and operations buildout. | 中 | SI005, SI007 |
| CI043 | No retained public source discloses Andromeda cash on hand, monthly burn, runway, or audited gross margin as of the July 2026 run date. | 中 | SI006, SI007, SI008 |
| CI044 | No retained public source discloses Andromeda debt facilities, provider guarantees, prepayment obligations, DSO or DPO, or customer concentration. | 中 | SI001, SI005, SI006, SI007, SI008 |
| CI045 | Using the public $1.5 billion valuation anchor against about $100 million of 2025 run-rate revenue or more than $50 million of 2024 revenue implies a coarse public valuation-to-revenue range of roughly 15x to 30x. | 中 | SI006, SI007 |
| CI046 | Revenue quality may be attractive if Andromeda is earning net broker economics on recurring committed workloads, but public evidence is still insufficient to underwrite margin durability or cash efficiency because recognition policy, take rate, and service-cost absorption are undisclosed. | 中 | SI001, SI006, SI007, SI021, SI022, SI024 |
| CI047 | Compared with public comparables such as CoreWeave, DigitalOcean, and Amazon, Andromeda offers far less public disclosure on contract mix, liquidity, and financing obligations, which is itself a material underwriting handicap. | 高 | SI013, SI021, SI024 |
| CE001 | Official workflow copy shows the buyer motion begins with defining workload requirements such as GPU type, quantity, region, and timeline before Andromeda sources supply. | 高 | SE001, SE003 |
| CE002 | Official workflow copy shows the provider motion begins with onboarding infrastructure details and then moves into certification, qualified demand routing, and standardized deal execution. | 高 | SE001, SE004 |
| CE003 | The insights page and privacy policy both describe Andromeda as a marketplace or market-infrastructure layer connecting compute customers with providers. | 高 | SE002, SE005 |
| CE004 | Public customer and provider forms show that Andromeda collects detailed workload and infrastructure metadata including GPU type, quantity, timing, network, storage, location, provider affiliation, interface type, and booking constraints. | 中 | SE003, SE004 |
| CE005 | Official home, customer, and insights copy say Andromeda benchmarks and prices compute across more than 100 providers in real time. | 高 | SE001, SE002, SE003 |
| CE006 | Official home, provider, and insights copy say Andromeda certifies or validates provider environments against enterprise-grade benchmarks and standardizes terms before deployment. | 高 | SE001, SE002, SE004 |
| CE007 | Official home and insights copy say the buyer-facing experience includes standardized SLAs or contracts, one invoice, one support channel, and full observability. | 高 | SE001, SE002 |
| CE008 | Andromeda's insights essay frames the product as market infrastructure that combines benchmarks, standards, matching, operations, and trust rather than as a simple listing site. | 高 | SE002, SE008 |
| CE009 | Upstarts describes Andromeda as a Switzerland-like neutral matchmaker that earns a spread by connecting supply and demand while handling complex market operations. | 中 | SE022 |
| CE010 | Upstarts and SiliconANGLE both describe Andromeda as the layer that vets, verifies, and manages procurement details rather than merely exposing a list of unused GPUs. | 中 | SE022, SE023 |
| CE011 | SiliconANGLE says Andromeda checks whether provider hardware and supporting systems meet performance and security requirements before listing that infrastructure on the platform. | 中 | SE023 |
| CE012 | SiliconANGLE reports that Andromeda offers an Andromeda Pricing Index with region-specific GPU price data. | 中 | SE023 |
| CE013 | Public software-engineering hiring shows Andromeda is building orchestration, provisioning, lifecycle-management, APIs, services, and control planes that abstract over VMs, Kubernetes, bare metal, and schedulers. | 高 | SE010, SE016 |
| CE014 | Public site-reliability hiring shows Andromeda provisions and operates Kubernetes-based clusters for customers across multiple providers and expects Infrastructure-as-Code proficiency in tools such as Terraform, Helm, and Ansible. | 高 | SE011, SE019 |
| CE015 | Senior SRE hiring shows Andromeda designs and evolves multi-provider, multi-region GPU compute clusters optimized for large-scale training. | 高 | SE012, SE014, SE018 |
| CE016 | Senior SRE public role text shows topology-aware scheduling, networking, storage decisions, GPU health checks, self-healing, and firmware or driver lifecycle management are explicit operating concerns inside the product stack. | 中 | SE012, SE014 |
| CE017 | Senior SRE public role text explicitly names InfiniBand, RoCE, NVLink, NCCL, CUDA, distributed PyTorch, DeepSpeed, Megatron, FSDP, Slurm, and Kubernetes GPU orchestration as relevant technologies. | 中 | SE012, SE014 |
| CE018 | Public solutions-engineering hiring shows Andromeda runs demos, POCs, reference architectures, cost modeling, and performance-oriented technical validation before production handoff. | 高 | SE013, SE017 |
| CE019 | Official careers, Built In, and Ashby surfaces show active hiring across software engineering, solutions engineering, SRE, procurement, compute trading, and partnerships around the product. | 高 | SE007, SE009, SE015 |
| CE020 | Built In's jobs page says compute-trader and procurement roles are responsible for supplier negotiations, compute resource utilization, and GPU-cluster supply strategy. | 中 | SE009 |
| CE021 | Upstarts reports that Andromeda leases access from CoreWeave and many other providers rather than owning all of the infrastructure itself. | 中 | SE022 |
| CE022 | DCD's June 2025 reporting shows the pre-spinout Andromeda Cluster had concrete large-scale assets and pricing, including published H100 availability and per-GPU hourly ranges. | 中 | SE024 |
| CE023 | The insights page plus current software-engineering and senior-SRE hiring all indicate that Andromeda routes both training and inference workloads across global supply rather than only brokering one-off reservations. | 中 | SE002, SE010, SE012 |
| CE024 | The public deployment motion appears services-led because the homepage and solutions-engineering role both emphasize guided evaluation, architecture work, and production handoff rather than an entirely self-serve motion. | 高 | SE001, SE013 |
| CE025 | The privacy policy says the platform is intended for business use and collects detailed customer compute requirements, provider infrastructure information, and technical session or analytics data. | 中 | SE005 |
| CE026 | The privacy policy says Andromeda maintains access controls, encryption in transit and at rest, ongoing monitoring, and a DPA process for GDPR or UK GDPR contexts. | 中 | SE005 |
| CE027 | The public trust surface is thin because the trust center exposes only a shell page while the privacy policy remains the main concrete public control document. | 高 | SE005, SE006 |
| CE028 | Reviewed official pages in this run did not surface a first-party public status page, incident archive, or named external security certifications. | 中 | SE001, SE005, SE006 |
| CE029 | Upstarts quotes a customer saying Andromeda remains valuable when it removes the headache of debugging GPU clusters and provides better customer service than a hyperscaler. | 中 | SE022 |
| CE030 | SiliconANGLE reports that Andromeda engineers monitor platform infrastructure for reliability issues and make adjustments when necessary. | 中 | SE023 |
| CE031 | Public SRE role descriptions make clear that SLOs, error budgets, observability, incident response, and postmortems are expected parts of Andromeda's product and operating model. | 中 | SE011, SE012, SE014 |
| CE032 | A public GitHub CV from a current Andromeda AI infrastructure engineer describes platform engineering across AI and GPU bare-metal infrastructure, GitOps automation, observability, and one-click burn-in, connectivity, and security tooling. | 低 | SE025 |
| CE033 | The same self-published GitHub CV references tools and substrates such as ArgoCD, Kubernetes, RKE2, Helm, Go, Python, Terraform, Slurm, KubeRay, Weka, VAST, RoCE, InfiniBand, Prometheus, Grafana, and Loki. | 低 | SE025 |
| CE034 | Andromeda's clearest differentiation is the combination of provider qualification, benchmark-driven normalization, price discovery, contract standardization, workload routing, and ongoing operations support in one neutral layer. | 高 | SE001, SE002, SE022, SE023 |
| CE035 | Product maturity is uneven because buyer and provider workflows look live and scaled, while benchmark methodology, certification criteria, and public trust disclosures remain sparse. | 高 | SE001, SE002, SE005, SE006, SE022, SE023 |
| CE036 | This run did not surface a public API reference, SDK, or formal docs site, so deployment appears more operator-assisted than fully self-serve. | 中 | SE001, SE002, SE007, SE013 |
| CE037 | The provider workflow explicitly accepts Kubernetes, Slurm, and VM interfaces while the software-engineering role also mentions bare metal, indicating heterogeneous execution substrates. | 中 | SE004, SE010 |
| CE038 | The key technical dependencies include external provider supply, infrastructure health across network and storage layers, orchestrator correctness, and Andromeda's own support or control-plane execution. | 中 | SE012, SE022, SE023, SE025 |
| CE039 | Trust risk remains because public sources describe standardization and certification but do not publish the benchmark suite, pass thresholds, provider-remediation process, or independent audit artifacts. | 中 | SE001, SE004, SE006, SE023 |
| CE040 | The reviewed source set supports treating Andromeda as an operational intermediation layer across procurement, deployment, observability, and troubleshooting rather than as a passive marketplace. | 高 | SE001, SE002, SE022, SE023 |
| CE041 | Current hiring implies meaningful active build-out across control planes, AI-infrastructure operations, and repeatable enterprise deployment motion rather than a fully finished product surface. | 高 | SE007, SE015, SE021 |
| CE042 | March 2026 launch-week materials made Andromeda publicly legible as a neutral compute market with provider breadth, transaction history, and a formal workflow for both sides of the market. | 高 | SE001, SE008, SE023 |
| CU001 | Official home copy positions Andromeda as a platform connecting AI teams with high-performance compute at scale rather than as a single-provider cloud. | 高 | SU001, SU002 |
| CU002 | The privacy policy says the platform connects AI teams and enterprises as customers with infrastructure providers, confirming a two-sided marketplace structure. | 高 | SU001, SU006 |
| CU003 | The customer intake form captures GPU type, requested quantity, start date, and duration, showing that demand is scoped as workload-specific compute procurement rather than generic software seats. | 高 | SU003, SU006 |
| CU004 | Official workflow copy says buyers define workload parameters and then receive sourced, benchmarked, priced, and deployable compute through one relationship. | 高 | SU001, SU002 |
| CU005 | The provider intake flow collects price, networking, storage, geographic location, cloud affiliation, interface type, and booking constraints, which means provider attributes are part of customer matching and segmentation. | 高 | SU004, SU006 |
| CU006 | Official and recruiting materials explicitly target frontier AI labs, AI-native startups, and enterprises as demand-side customer segments. | 中 | SU001, SU018 |
| CU007 | Official surfaces and recruiting language show that Andromeda also serves compute providers, data centers, and cloud operators as structured counterparties in the same marketplace. | 中 | SU002, SU004 |
| CU008 | The partnerships role makes VC funds and accelerators an explicit customer-acquisition channel by routing portfolio-company compute demand into Andromeda. | 中 | SU021, SU009 |
| CU009 | Because Andromeda promises one invoice, one support channel, and centralized data handling, the operational user and economic payer can differ inside the same account. | 中 | SU001, SU006 |
| CU010 | Public location signals imply a San Francisco-centered company serving globally distributed demand and supply rather than a purely local customer base. | 中 | SU001, SU016, SU017 |
| CU011 | The Solutions Engineer role names engineering leaders, platform teams, and security or infrastructure owners as stakeholders, which implies a multi-threaded enterprise buying center. | 中 | SU018, SU020 |
| CU012 | Official home copy says Andromeda works across more than 100 compute providers. | 高 | SU001, SU011 |
| CU013 | Official home copy says Andromeda has completed more than 1,000 transactions, the strongest public repeat-usage proxy in the current file. | 高 | SU001, SU011 |
| CU014 | The insights page says the network has processed over a billion GPU-hours across dozens of clusters and many providers, which is evidence of operational throughput rather than a static listing catalog. | 高 | SU002, SU010 |
| CU015 | DCD reported that by June 2025 Andromeda was available to non-NFDG companies, with up to 2,000 H100s accessible within hours at $2.40-$3.00 per GPU-hour. | 中 | SU012 |
| CU016 | Upstarts says the original managed cluster filled almost instantly, which supports a real supply-constrained demand story before Andromeda broadened its market. | 中 | SU010 |
| CU017 | Upstarts says Andromeda reached roughly a $100 million 2025 revenue run-rate while serving notable AI startups and labs that Moushey would not publicly name. | 中 | SU010, SU011 |
| CU018 | SiliconANGLE says Andromeda's customer base includes notable AI startups and labs that often spend $250 million to $500 million per year on infrastructure. | 中 | SU011 |
| CU019 | AI Grant says larger investments come with access to the Andromeda Cluster, making the accelerator a formal customer-acquisition and adoption channel. | 中 | SU009 |
| CU020 | The Head of Partnerships role says Andromeda wants both VC or accelerator channels and provider distribution partnerships to expand what the company can offer and to whom. | 中 | SU021 |
| CU021 | Customer-facing recruiting shows that strategic accounts are expected to move through technical discovery, demos, POCs, and clean production handoffs rather than simple self-serve activation. | 中 | SU018, SU020 |
| CU022 | The Staff SRE role says top-customer incidents and direct work with sophisticated AI infrastructure customers are part of production operations, implying live high-stakes deployments. | 中 | SU025 |
| CU023 | Compute Trader and business-staff roles focus on matching incoming sales leads to internal and external capacity and maximizing utilization, which implies ongoing live demand-routing rather than one-off brokered deals. | 中 | SU023, SU024 |
| CU024 | The procurement role says Andromeda has already established a fundamental provider-capacity layer and is now widening network liquidity and services, a sign of expansion from an existing installed base. | 中 | SU022 |
| CU025 | Upstarts explicitly says portfolio companies ElevenLabs and Pika took advantage of Andromeda's early 4,000-plus GPU stockpile. | 中 | SU010 |
| CU026 | SaaS Sentinel independently repeated that ElevenLabs and Pika used the infrastructure, corroborating the existence of named early users even though it adds no first-party customer quote. | 中 | SU013, SU010 |
| CU027 | Upstarts says many initial users were AI Grant-backed startups and separately names Browserbase, Cursor, Granola, and Perplexity as relevant batchmates, but it does not directly confirm each as a live Andromeda deployment. | 中 | SU010, SU009 |
| CU028 | BuildMVPFast later summarized Andromeda's portfolio-user orbit as ElevenLabs, Pika, Cursor, Perplexity, and Browserbase, but that source is derivative and low-reputation rather than direct customer evidence. | 低 | SU015 |
| CU029 | No reviewed official Andromeda surface in this run named ElevenLabs, Pika, Cursor, or Perplexity in a case study, logo wall, or customer quote. | 中 | SU001, SU002, SU003, SU005 |
| CU030 | Upstarts quoted an anonymous founder saying their company keeps working with Andromeda because it serves as an excellent engineering team and is more flexible than a hyperscaler. | 中 | SU010 |
| CU031 | ClusterMAX rated Andromeda as Unavailable in November 2025 and said the service could not yet be verified, which is an adverse signal that public customer validation lagged the market narrative. | 中 | SU014 |
| CU032 | No reviewed source disclosed NRR, GRR, churn, logo retention, renewal rate, contract length, or a cohort curve for Andromeda customers. | 高 | SU001, SU002, SU005, SU010 |
| CU033 | The best public retention proxy is behavioral rather than contractual: 1,000-plus transactions plus live demand-routing roles suggest repeated use, but not how many accounts repeat or renew. | 中 | SU001, SU023, SU024 |
| CU034 | The Solutions Engineer role ties technical evaluations to successful expansions across strategic accounts, implying an intended land-and-expand motion even though conversion data is absent. | 中 | SU018, SU020 |
| CU035 | Built In's job summaries say POCs should convert reliably and be tied to customer ROI, another sign that Andromeda internally measures conversion quality without disclosing public benchmarks. | 中 | SU017, SU018 |
| CU036 | The privacy policy offers DPAs on request and frames the platform as business-use software handling billing and contractual data, which suggests enterprise procurement and legal review are part of onboarding. | 高 | SU006, SU017 |
| CU037 | Upstarts says NFDG portfolio companies do not get preferential treatment but that many remain customers, implying the installed base likely still carries meaningful ecosystem concentration. | 中 | SU010 |
| CU038 | DCD and Upstarts together show Andromeda started as a portfolio-only cluster and only later opened externally, so customer diversification is real but early-channel concentration remains unresolved. | 中 | SU012, SU010 |
| CU039 | The partnerships role says active partnership KPIs are tied to revenue contribution, utilization, and growth, showing that channel expansion is judged on measurable commercial output. | 中 | SU021 |
| CU040 | A competitor comparison from CompuX characterizes Andromeda as a procurement layer without financing, which suggests some compute buyers may still need separate capital solutions to scale usage. | 低 | SU026 |
| CU041 | Built In job material around commercial counsel, procurement, and solutions engineering shows that privacy/export compliance, revenue-critical agreements, and custom supply matching are material parts of closing larger accounts. | 中 | SU017, SU018, SU022 |
| CU042 | The same anonymous founder who praised Andromeda also said a large enough startup might eventually bring those capabilities in-house, creating a plausible durability ceiling for the most sophisticated accounts. | 中 | SU010 |
| CU043 | ClusterMAX describes Andromeda as a fractional SRE or fractional procurement team for trusted NFDG companies, which supports the services-led value proposition but also reinforces concentration and opacity concerns. | 中 | SU014 |
| CU044 | gpulist.ai, which identifies itself as being from andromeda.ai and says it is serving $1.84 billion of listings, suggests Andromeda is also cultivating a supply-liquidity and market-data surface around the marketplace. | 中 | SU008 |
| CU045 | The public trust center is extremely sparse relative to the company's enterprise-grade claims, which weakens external reference quality for security-conscious buyers. | 中 | SU007, SU006 |
| CU046 | ElevenLabs announced a multi-year February 2026 Google Cloud and NVIDIA Blackwell agreement, which shows the named early-user proof does not imply current exclusive reliance on Andromeda and highlights multihoming risk for breakout customers. | 中 | SU027 |
| CU047 | A September 2024 Lightspeed event featured Pika and ElevenLabs as breakout generative-AI companies, reinforcing that the two clearest named Andromeda references are strategically meaningful logos even if their Andromeda deployment details remain thin. | 中 | SU028 |
| CR001 | Upstarts reports that Nat Friedman and Daniel Gross now have no active relationship with Andromeda, making founder withdrawal the clearest public continuity change. | 中 | SR016 |
| CR002 | Data Center Dynamics reported that the effect of Meta's buyout discussions around NFDG on Andromeda was unclear, reinforcing public uncertainty around sponsor continuity. | 中 | SR017 |
| CR003 | Official home and privacy-policy pages describe Andromeda as a global GPU marketplace connecting buyers and providers rather than as a single-owner cloud operator. | 高 | SR001, SR003 |
| CR004 | The privacy policy states that Andromeda's services are intended for business use and that users consent to its data-processing terms by using the services. | 中 | SR003 |
| CR005 | The privacy policy says Andromeda collects identity, contact, account, billing, and service-interaction information from people using the platform or interacting with the company. | 中 | SR003 |
| CR006 | The customer and provider intake flows show that Andromeda brokers both workload-side and supplier-side data, including GPU type, quantity, timing, geography, cloud affiliation, and price. | 高 | SR005, SR006 |
| CR007 | The reviewed trust-center surface is sparse and does not itself present substantive public certification, incident-history, or control-detail content. | 中 | SR004 |
| CR008 | Built In's commercial-legal role description says Andromeda still needs dedicated support for revenue-critical contracts, governance, privacy and export compliance, IP, employment, vendor matters, and fundraising or M&A. | 中 | SR013 |
| CR009 | BIS's updated public-information page says advanced-computing controls retain licensing requirements for China, Hong Kong, and Macau and add anti-circumvention measures plus new due-diligence expectations. | 高 | SR025, SR026 |
| CR010 | The Federal Register's AI Diffusion rule says BIS revised controls on advanced-computing chips and added a new control on certain advanced closed-weight AI model weights. | 中 | SR026 |
| CR011 | OFAC's compliance framework explicitly applies to foreign entities that conduct business in or with the United States, use U.S. persons, or use U.S.-origin goods or services. | 中 | SR028 |
| CR012 | OFAC's sanctions-program page shows active China-related, cyber-related, non-proliferation, North Korea, and Russia-related programs that can intersect with cross-border compute commerce. | 中 | SR027 |
| CR013 | FTC guidance says businesses must honor their privacy-policy promises and maintain security appropriate to the nature of the data they hold. | 高 | SR030, SR032 |
| CR014 | FTC breach-response guidance says companies should secure systems quickly and remediate vulnerabilities after any incident, which raises the stakes of sparse public incident preparedness signals. | 中 | SR029 |
| CR015 | FTC's Start with Security guidance says service providers should implement reasonable security measures, making vendor-risk control relevant to Andromeda's cross-provider marketplace model. | 中 | SR031 |
| CR016 | FTC's personal-information guide says sensitive-data loss can lead to fraud, identity-theft harms, loss of customer trust, and litigation costs. | 中 | SR032 |
| CR017 | Upstarts reports that Andromeda leases access from CoreWeave and other providers, so it avoids direct hardware ownership and depreciation but remains exposed to third-party supply performance. | 中 | SR016 |
| CR018 | The procurement role says Andromeda acquires and facilitates compute resources across the company while working with providers, sales, and technical teams to match supply with demand. | 中 | SR009 |
| CR019 | The same procurement role says scaling requires widening provider network and liquidity while interacting with OEMs, ODMs, integrators, and parts suppliers, which confirms multi-layer supply-chain dependence. | 中 | SR009 |
| CR020 | The business-staff role is explicitly tasked with accelerating supply-demand matching on the platform, implying that fill-rate and utilization are still execution-sensitive processes. | 中 | SR011 |
| CR021 | The solutions-engineer role says Andromeda wins strategic accounts through technical discovery, demos, POCs, evaluations, and expansions across hardware plus software workflows. | 中 | SR012, SR015 |
| CR022 | Upstarts says AI-lab business is high-dollar and secretive, limiting public visibility into exact customer names, contract terms, and concentration. | 中 | SR016 |
| CR023 | Upstarts quotes an anonymous founder saying Andromeda serves as an outsourced engineering team and is more flexible than a hyperscaler, which supports real customer value but also a services-heavy model. | 中 | SR016 |
| CR024 | The same founder told Upstarts that if the startup grows big enough it may make sense to bring those capabilities in-house, showing a clear customer-repatriation risk. | 中 | SR016 |
| CR025 | A February 2026 PR Newswire release says ElevenLabs expanded its use of Google Cloud and Blackwell GPUs, demonstrating that at least one named early Andromeda user can multihome with major cloud providers. | 中 | SR022 |
| CR026 | Data Center Dynamics reported that by June 2025 non-NFDG companies could access Andromeda capacity with up to 2,000 H100s within hours at published pricing, confirming commercialization beyond the captive portfolio. | 中 | SR017 |
| CR027 | Official home copy and SiliconANGLE support that Andromeda works with 100-plus providers and has completed more than 1,000 transactions. | 高 | SR001, SR018 |
| CR028 | gpulist.ai says it is serving $1.84 billion of listings, which signals active marketplace-side liquidity even though listing volume is not the same as contracted revenue. | 中 | SR023 |
| CR029 | SemiAnalysis describes Andromeda as a fractional SRE and procurement layer sourcing from multiple neoclouds, which reinforces operational dependence on heterogeneous external supply. | 中 | SR019 |
| CR030 | SemiAnalysis also says the future of Andromeda is uncertain after Nat Friedman and Daniel Gross joined Meta, adding an independent adverse view on continuity. | 中 | SR019 |
| CR031 | CompuX frames Andromeda as lacking a financing or credit mechanism, highlighting the risk that customer or supplier economics may demand more balance-sheet support than the platform currently discloses. | 低 | SR020 |
| CR032 | The Strategic Compute Finance Lead role says Andromeda is building financial architecture around capex planning, scenario modeling, deal execution, and lender relationships. | 中 | SR010 |
| CR033 | The same finance role says the company is focused on expanding capital capacity and deepening lender relationships, implying capital access is an active operating dependency rather than a solved problem. | 中 | SR010 |
| CR034 | Upstarts reports that Andromeda was profitable since start, reached roughly a $100 million 2025 revenue run rate, and hoped to scale compute under management toward 100,000 GPUs. | 中 | SR016 |
| CR035 | Upstarts quotes Moushey saying only a handful of balance sheets can handle these investments and that the market still has to figure out how to underwrite them, underscoring structural capital intensity. | 中 | SR016 |
| CR036 | AI Grant offers seed-stage AI startups capital and large cloud-credit packages, supporting a founder-network channel that can help demand origination but also concentrate early customer acquisition. | 中 | SR021, SR016 |
| CR037 | Official and Ashby jobs pages list open legal, finance, procurement, partnerships, software, solutions, and SRE roles, showing that Andromeda is still building key control and execution functions in public view. | 中 | SR007, SR008 |
| CR038 | Reviewed official pages do not expose a board roster, formal executive map, or public control summary, leaving governance and succession externally under-disclosed. | 中 | SR001, SR007 |
| CR039 | Reviewed official home and trust surfaces do not expose a public incident log, detailed status history, or named certification set, which leaves trust disclosure notably thinner than the revenue narrative. | 中 | SR001, SR004 |
| CR040 | The privacy policy and both intake flows indicate that Andromeda centralizes contact, account, billing, workload, and supplier data across both sides of the marketplace. | 高 | SR003, SR005, SR006 |
| CR041 | Because Andromeda acts as one contract, billing, and support layer across counterparties, contract disputes, fraud screening, billing errors, and working-capital mismatches can accumulate at the broker rather than stay with a single provider. | 中 | SR003, SR006, SR013 |
| CR042 | Public hiring and provider-breadth claims show real mitigation effort on supply and commercialization, but the reviewed control stack is still process-building rather than publicly attested and mature. | 中 | SR004, SR007, SR009, SR010, SR012 |
| CR043 | Public sources confirm non-NFDG commercialization but still do not disclose customer count, retention metrics, or top-customer share, so diversification cannot be underwritten from public evidence alone. | 中 | SR016, SR017 |
| CR044 | Export-control, sanctions, and privacy obligations are material to Andromeda because the marketplace routes global supply and demand through U.S.-linked advanced-computing infrastructure while holding identifiable business data. | 高 | SR003, SR025, SR028, SR030 |
| CR045 | The highest-ranked residual risk is governance and continuity opacity after the founders stepped back because public control disclosure is missing while sponsor relationships historically mattered for supply, capital, and customer access. | 中 | SR007, SR016, SR017 |
| CR046 | The second-ranked residual risk is provider and operational dependence because delivery still requires active procurement, capacity matching, and quality management across third-party supply. | 中 | SR009, SR011, SR016, SR019 |
| CR047 | The third-ranked residual risk is customer concentration and multihoming because named references remain thin, confidential labs likely dominate spend, and at least one early user publicly scaled elsewhere. | 中 | SR016, SR017, SR022 |
| CR048 | The fourth-ranked residual risk is capital-capacity and margin complexity because public evidence points to lender relationships, deal structuring, and underwriting demands without disclosure of working-capital protections. | 中 | SR010, SR016, SR020 |
| CR049 | The fifth-ranked residual risk is regulatory and compliance execution because no public export-screening, sanctions-screening, incident-history, or certification package was surfaced despite a material ruleset. | 中 | SR004, SR025, SR028, SR030 |
| CR050 | A real mitigation is that Andromeda already claims foundational provider capacity and is hiring specialized leaders across procurement, finance, legal, and customer execution. | 中 | SR007, SR009, SR010 |
| CR051 | A leadership thesis-break trigger would be Wil Moushey departing or management failing to provide a credible governance, ownership, and succession map before investment close. | 中 | SR007, SR016, SR017 |
| CR052 | A regulatory thesis-break trigger would be Andromeda being unable to document export or sanctions screening, DPA or security posture, or a material data-security incident requiring external response. | 中 | SR003, SR013, SR029, SR031 |
| CR053 | A partner and customer thesis-break trigger would be evidence that a small set of providers, AI Grant-origin channels, or customers dominates volume without durable substitution or contract protection. | 中 | SR016, SR017, SR021 |
| CR054 | A capital-capacity thesis-break trigger would be management failing to evidence lender support, working-capital resilience, or margin tolerance as compute under management scales. | 中 | SR010, SR016, SR020 |
| CV001 | Andromeda's official site and independent March 2026 coverage both support a network of more than 100 providers and more than 1,000 completed GPU transactions. | 高 | SV001, SV006 |
| CV002 | Andromeda's insights page and Upstarts both support a record of operating large-scale training infrastructure across dozens of clusters and more than a billion GPU-hours. | 中 | SV002, SV005 |
| CV003 | Andromeda's official workflow and privacy surfaces show a high-touch broker model that handles provider, buyer, billing, and deployment data rather than a simple self-serve API checkout. | 高 | SV001, SV003 |
| CV004 | Upstarts and SiliconANGLE both place Andromeda around a roughly $100 million 2025 revenue run-rate and say the company has operated profitably since launch. | 中 | SV005, SV006 |
| CV005 | Upstarts reported that Andromeda generated more than $50 million of revenue in 2024. | 中 | SV005 |
| CV006 | Upstarts, Gaebler, and Raising.fi cluster around Paradigm as the investor and about $60 million of total Paradigm capital committed to Andromeda by March 2026. | 中 | SV005, SV007, SV008 |
| CV007 | Upstarts and SiliconANGLE both anchor Andromeda's March 2026 valuation at roughly $1.5 billion. | 中 | SV005, SV006 |
| CV008 | Public round coverage and the general availability of SEC Form D data still do not surface Andromeda's own security type, liquidation preferences, ratchets, or any secondary component. | 中 | SV005, SV006, SV007, SV008, SV019 |
| CV009 | Upstarts says Nat Friedman and Daniel Gross no longer have an active relationship with Andromeda, while Data Center Dynamics says the implications for the cluster are unclear. | 中 | SV005, SV009 |
| CV010 | SemiAnalysis describes Andromeda as procuring capacity from a range of neoclouds on behalf of startups, reinforcing that the company depends on external providers rather than owning a unified hyperscale stack. | 中 | SV010 |
| CV011 | AI Grant's portfolio-access model shows that Andromeda still benefits from founder-network and accelerator-style channel distribution rather than purely stand-alone demand. | 中 | SV011 |
| CV012 | ElevenLabs' 2026 Google Cloud partnership shows that at least one named early Andromeda-linked user can multihome significant AI infrastructure onto a large incumbent platform. | 中 | SV012 |
| CV013 | The Strategic Compute Finance role says Andromeda is focused on expanding capital capacity, deepening lender relationships, and building financial infrastructure for its next phase of growth. | 中 | SV014 |
| CV014 | The Business Staff role shows Andromeda is still actively hiring to accelerate supply and demand matching on the platform. | 中 | SV015 |
| CV015 | The Solutions Engineer role shows a land-and-expand motion built around demos, POCs, and expansions across strategic accounts rather than low-friction self-serve expansion. | 中 | SV016 |
| CV016 | No reviewed public source disclosed Andromeda's NRR, GRR, churn, active customer count, or revenue concentration. | 中 | SV001, SV005, SV006 |
| CV017 | No reviewed public source disclosed provider concentration, substitution coverage, or SLA depth across Andromeda's stated provider network. | 中 | SV001, SV004, SV013 |
| CV018 | CompaniesMarketCap lists CoreWeave at about $54.30 billion of market capitalization in July 2026. | 中 | SV017 |
| CV019 | StockAnalysis says CoreWeave had $6.23 billion of trailing revenue and an 8.72x P/S ratio as of June 30, 2026. | 中 | SV018 |
| CV020 | CoreWeave's Q1 2026 results release reported $2.078 billion of quarterly revenue, $99.4 billion of revenue backlog, and more than 1 gigawatt of active power. | 中 | SV020 |
| CV021 | CoreWeave's 2025 10-K says customers buy multi-year take-or-pay contracts and remaining performance obligations reached $60.7 billion at year-end 2025. | 中 | SV022 |
| CV022 | CoreWeave's 10-K and Fitch both show very high customer concentration, with the 10-K saying Microsoft was 67% of 2025 revenue and Fitch saying Microsoft was 62% of 2024 revenue while the top two customers were 77%. | 中 | SV022, SV023 |
| CV023 | CoreWeave's 10-Q says March 2026 cash, cash equivalents, and marketable securities were $2.2 billion and that future investments may require significant debt and or equity financing. | 中 | SV021 |
| CV024 | CompaniesMarketCap lists DigitalOcean at about $16.38 billion of market capitalization in July 2026. | 中 | SV024 |
| CV025 | CompaniesMarketCap lists DigitalOcean at about $0.94 billion of trailing-twelve-month revenue in 2026. | 中 | SV025 |
| CV026 | DigitalOcean's current public multiple is roughly 17.4x revenue using the July 2026 market-cap and revenue snapshots. | 中 | SV024, SV025 |
| CV027 | DigitalOcean's 2025 10-K says pricing is primarily consumption-based, larger workloads increasingly use committed contracts, 2025 ARR was $970 million, and the top 25 customers were only 7% of revenue. | 中 | SV026 |
| CV028 | DigitalOcean's 2026 10-Q says gross profit fell to 56% from 61% because expansion costs arrived ahead of the revenue ramp. | 中 | SV027 |
| CV029 | CompaniesMarketCap lists Nebius Group at about $70.11 billion of market capitalization in July 2026. | 中 | SV028 |
| CV030 | CompaniesMarketCap lists Nebius Group at about $0.52 billion of revenue on its current snapshot. | 中 | SV029 |
| CV031 | Nebius therefore trades at roughly 135x revenue on those July 2026 snapshots, making it an upside appetite reference rather than a sober operating comp. | 中 | SV028, SV029 |
| CV032 | CompaniesMarketCap lists Amazon at about $2.563 trillion of market capitalization in July 2026, underscoring how much trust and procurement scale still sit with hyperscaler incumbents. | 中 | SV030 |
| CV033 | Amazon's 10-K and 10-Q show gross-versus-net revenue mechanics, AWS prepayments, and $43.2 billion of Q1 2026 capex, illustrating the capital burden Andromeda can avoid only if its broker model is truly asset-light. | 中 | SV031, SV032 |
| CV034 | Andromeda's current private valuation equals roughly 15x the reported ~$100 million 2025 run-rate and roughly 30x the reported >$50 million 2024 revenue floor. | 中 | SV005, SV006 |
| CV035 | Current 2026 public AI and cloud infrastructure multiples are wide, spanning roughly 8.7x for CoreWeave, about 17.4x for DigitalOcean, and more than 100x for Nebius. | 中 | SV018, SV024, SV025, SV028, SV029 |
| CV036 | That dispersion means Andromeda's ~15x run-rate multiple is not obviously absurd in 2026 public-market context, but it is not clearly cheap either. | 中 | SV005, SV006, SV018, SV024, SV025, SV028, SV029 |
| CV037 | The decisive underwriting issue is denominator quality because if Andromeda's reported revenue is gross pass-through rather than net spread, the effective economic multiple could be materially higher than 15x. | 中 | SV003, SV005, SV006, SV031 |
| CV038 | The other decisive issue is durability because without disclosed retention, concentration, or provider substitution data investors cannot tell whether Andromeda is a sticky broker platform or a high-touch stopgap. | 中 | SV005, SV006, SV010, SV016 |
| CV039 | The bull case requires that Andromeda's current scale mostly reflects high-value net economics, that multihoming does not erode wallet share, and that capital-capacity buildout becomes a moat rather than a drag. | 中 | SV005, SV006, SV012, SV014, SV015 |
| CV040 | The base case is that Andromeda remains strategically relevant and keeps growing, but only grows into the current mark because proof on renewal quality, governance, and financing terms remains incomplete. | 中 | SV005, SV006, SV013, SV014, SV016, SV026, SV027 |
| CV041 | The bear case is that customers or providers concentrate, direct clouds close the workflow gap, or governance and financing opacity prevent premium valuation persistence. | 中 | SV009, SV010, SV012, SV013, SV022, SV023 |
| CV042 | Public evidence does not support a clean buy call at the current price because audited recognized revenue, gross-versus-net accounting, customer retention, provider concentration, and financing terms remain undisclosed. | 中 | SV005, SV006, SV007, SV008, SV026, SV027 |
| CV043 | The most decision-useful diligence asks are an audited revenue bridge, a cap-table and side-letter package, customer and provider concentration by GMV and revenue, and a fuller trust and compliance packet. | 中 | SV003, SV004, SV013, SV014, SV015, SV016, SV019 |
| CV044 | The only supportable hold logic from public evidence is a medium-duration grow-into-the-multiple case, not a near-term flip or clean exit-underwriting case. | 中 | SV005, SV006, SV024, SV025, SV030 |
| CV045 | If Wil Moushey left without a clear succession plan, the thesis would weaken immediately because the public founder bench is already thinner than the original narrative suggests. | 中 | SV005, SV009 |
| CV046 | If diligence shows top customers or providers are concentrated and hard to replace, the thesis breaks because neutrality and diversification are central to Andromeda's value proposition. | 中 | SV001, SV010, SV022, SV023 |
| CV047 | If diligence shows revenue is primarily gross pass-through or margins are compressed by service delivery or financing support, the valuation should compress materially. | 中 | SV003, SV005, SV026, SV027, SV032 |
| CV048 | If future financings add punitive preferences, convertibles, or contract-tied debt, headline valuation may materially overstate real equity upside. | 中 | SV014, SV019, SV021, SV023 |
| 编号 | 出版方 | 标题 | 引文 |
|---|---|---|---|
| SO001 | Andromeda | Andromeda · Access the world's compute | Andromeda connects AI teams with high-performance compute fast, at scale, and on terms that work |
| SO002 | Andromeda | Andromeda · Access the world's compute | For the last three years, we have helped leading AI companies source, onboard, and operate large-scale training infrastructure. |
| SO003 | Andromeda | Andromeda · Access the world's compute | The platform sources, benchmarks, and prices compute across 100+ providers in real time. |
| SO004 | Andromeda | Andromeda · Access the world's compute | Andromeda then structures and standardizes contracts to get capacity deployed. |
| SO005 | Andromeda | Andromeda · Access the world's compute | Andromeda exists to make high-performance compute available to every team building at the frontier. |
| SO006 | Andromeda Cluster, Inc. | Privacy Policy | Andromeda Cluster, Inc. and its affiliates operates a global GPU compute marketplace platform connecting Customers with Providers. |
| SO007 | Andromeda | Andromeda.ai Trust Center | Andromeda.ai Trust Center |
| SO008 | X | Andromeda (@andromeda_ai) / X | Today we're announcing Andromeda. For three years, we've built the market infrastructure for compute. |
| SO009 | Built In | Andromeda (andromeda.ai) Careers, Perks + Culture | Built In | Andromeda Cluster was founded by Nat Friedman and Daniel Gross to give early-stage startups access to the kind of scaled AI infrastructure once reserved only for hyperscalers. |
| SO010 | Built In | Andromeda (andromeda.ai) Jobs + Careers | Built In | Provide end-to-end commercial legal support: draft and negotiate revenue-critical agreements, support corporate governance, compliance (privacy, export), and assist with fundraising and M&A as needed. |
| SO011 | Andromeda Cluster | Andromeda Cluster Jobs | Andromeda Cluster Jobs |
| SO012 | Andromeda Cluster | Commercial Counsel @ Andromeda Cluster | Commercial Counsel @ Andromeda Cluster |
| SO013 | Andromeda Cluster | Strategic Compute Finance Lead @ Andromeda Cluster | Strategic Compute Finance Lead @ Andromeda Cluster |
| SO014 | Andromeda Cluster | Head of Partnerships @ Andromeda Cluster | Head of Partnerships @ Andromeda Cluster |
| SO015 | Andromeda Cluster | Head of Procurement/Supply Chain @ Andromeda Cluster | Head of Procurement/Supply Chain @ Andromeda Cluster |
| SO016 | Andromeda Cluster | Compute Trader @ Andromeda Cluster | Compute Trader @ Andromeda Cluster |
| SO017 | Andromeda Cluster | Member of the Business Staff - Compute Markets @ Andromeda Cluster | Member of the Business Staff - Compute Markets @ Andromeda Cluster |
| SO018 | Andromeda Cluster | Solutions Engineer @ Andromeda Cluster | Solutions Engineer @ Andromeda Cluster |
| SO019 | Andromeda Cluster | Software Engineer - AI Infrastructure @ Andromeda Cluster | Software Engineer - AI Infrastructure @ Andromeda Cluster |
| SO020 | Andromeda Cluster | Senior Site Reliability Engineer - AI Infrastructure @ Andromeda Cluster | Senior Site Reliability Engineer - AI Infrastructure @ Andromeda Cluster |
| SO021 | Upstarts | The $1.5B Compute Startup Helping AI's Hottest Companies Find GPUs | Spun out as its own startup with backing from NFDG and Paradigm, Andromeda quietly passed a revenue run rate of $100 million in 2025. |
| SO022 | SiliconANGLE | On-demand GPU startup Andromeda raises funding at $1.5B valuation | The company offers access to infrastructure from more than 100 providers. It claims to have processed more than 1,000 GPU transactions since launching about two years ago. |
| SO023 | Raising.fi | Andromeda AI Inc. Secures $60 Million in Latest Funding Round Led by Paradigm | Wil Moushey, CEO of Andromeda AI Inc., spearheads the company's mission to ease the complexities involved in AI infrastructure procurement. |
| SO024 | Gaebler / VentureDeal | Andromeda 3/18/2026 Capital Raise - Gaebler.com Venture Capital Database | Andromeda closed a $60 million funding round on 3/18/2026. Investors included Paradigm. |
| SO025 | Data Center Dynamics | Meta in talks to partially acquire VC fund NFDG, hire Nat Friedman and Daniel Gross for AI shakeup | What it would mean for the Andromeda Cluster is unclear. |
| SO026 | Daniel Gross | Daniel Gross | I run compute for Meta. |
| SO027 | CB Insights | Nat Friedman and Daniel Gross | Nat Friedman and Daniel Gross is an investor group run by angel investor Nat Friedman and entrepreneur Daniel Gross. |
| SM001 | Andromeda | A view from billions of GPU-Hours | What this market needs is the full infrastructure of trade: sourcing, quality certification, standardized contracts, structuring, matching, operations, and the trust to make it work at scale. |
| SM002 | Andromeda | Andromeda customer intake form | GPU Type; How many GPUs do you need?; When would you like to start?; For how many weeks do you need the GPUs? |
| SM003 | Andromeda | Andromeda provider intake form | Price GPU/Hr ($); Interconnect Network; Minimum Bookable GPUs; Min Bookable Duration (weeks). |
| SM004 | Andromeda | Andromeda · Access the world's compute | Andromeda facilitates commerce between AI builders and infrastructure providers. |
| SM005 | Data Center Dynamics | Meta in talks to partially acquire VC fund NFDG, hire Nat Friedman and Daniel Gross for AI shakeup | We currently have up to 2,000 H100s available and can give you access to GPUs within a few hours. |
| SM006 | SiliconANGLE | On-demand GPU startup Andromeda raises funding at $1.5B valuation | The company will use its new funding to grow its customer base, which reportedly includes multiple notable AI startups and labs that typically spend $250 million to $500 million per year on infrastructure. |
| SM007 | The Business Research Company | Global Graphics Processing Unit (GPU) As A Service Market Report 2026 | The graphics processing unit (gpu) as a service market size has grown exponentially in recent years. It will grow from $5.8 billion in 2025 to $7.39 billion in 2026. |
| SM008 | S&P Global Market Intelligence | Buyer insight: Product and company considerations for GPUaaS infrastructure | AI technology and cloud infrastructure have consistently placed among the three highest-scoring areas of enterprise technology spending intent... and demand ... has led to a boom in GPU-as-a-service cloud infrastructure offerings. |
| SM009 | Presenc | The State of AI GPU Supply in 2026 | NVIDIA H100 cloud rental rates fell from approximately $8/hr in early 2023 to $1.80-3.50/hr in Q2 2026, with spot pricing as low as $1.20/hr. |
| SM010 | SesameDisk | GPU Spot Price and Capacity Outlook for AI Workloads in 2026 | Quotas, regional limits, queue times, and enterprise relationship status often matter more than list pricing. |
| SM011 | Cyfuture AI | The Market Shock: GPU Pricing Undergoes Its Fastest Correction in Infrastructure History | By Q1 2026, that same H100 access is available from neo-cloud providers for $1.38–$2.63/hr. |
| SM012 | GPUHosted | GPU Cloud 2026 Guide | The GPU cloud market in 2026 features a diverse array of providers, each with unique strengths and pricing models. |
| SM013 | RunPod | How Runpod GPU pricing works | Runpod pricing is based on the type of GPU workload you run. Pods are dedicated GPU instances for development and long-running jobs, Serverless bills inference workers based on usage, and Clusters support multi-node workloads and reserved capacity. |
| SM014 | Lambda | AI Cloud Pricing | GPU Compute & AI Infrastructure | Production-ready clusters from 16 to 2,000+ NVIDIA B200 or H100 GPUs. |
| SM015 | CoreWeave | NVIDIA HGX H100/H200 | Products | CoreWeave | CoreWeave’s purpose-built AI cloud delivers up to 20% higher Model FLOPS Utilization (MFU) and 10x greater reliability on thousand-GPU clusters. |
| SM016 | CoreWeave | CoreWeave Cloud Pricing | NVIDIA HGX H100 ... On-Demand Price: $49.24 / Hour ... Spot Price: $19.71 / Hour. |
| SM017 | Vast.ai | Live GPU Prices | Interruptible — 50%+ cheaper. Best for batch training. Reserved — Up to 50% Off. |
| SM018 | Amazon Web Services | Amazon EC2 P5 instances | These instances are deployed in Amazon EC2 UltraClusters that enable scaling up to 20,000 H100 or H200 GPUs interconnected with a petabit-scale nonblocking network. |
| SM019 | Google Cloud | Accelerator-optimized VM Pricing | A3 High ... $88.490000119 / 1 hour. A3 Mega ... $93.400712807 / 1 hour. |
| SM020 | Google Cloud | GPU pricing - Google Cloud | Spot prices are variable and can change up to once every day, but provide discounts of up to 91% off of the corresponding default price for many machine types, GPUs, TPUs, and Local SSDs. |
| SM021 | Microsoft Learn | ND H100 v5 size series - Azure Virtual Machines | The ND H100 v5 series starts with a single VM and eight NVIDIA H100 Tensor Core GPUs ... and can scale up to thousands of GPUs with 3.2 Tbps of interconnect bandwidth per VM. |
| SM022 | Microsoft Learn | ND H200 v5 size series - Azure Virtual Machines | The ND H200 v5 series starts with a single VM and eight NVIDIA H200 Tensor Core GPUs ... and can scale up to thousands of GPUs with 3.2Tb/s of interconnect bandwidth per VM. |
| SM023 | NVIDIA | NVIDIA H100 Tensor Core GPU | The combination of fourth-generation NVLink ... and NDR Quantum-2 InfiniBand networking ... delivers efficient scalability from small enterprise systems to massive, unified GPU clusters. |
| SM024 | CoreWeave | The world's #1 AI cloud platform, purpose-built for what's next | CoreWeave Cloud is an AI-native platform purpose-built for AI. It combines next-generation infrastructure, intelligent tools, and expert support. |
| SM025 | CoreWeave | Performance Benchmarks Report | Explore how our large-scale training benchmarks delivered 20% greater MFU and 10x uptime on 1,024 NVIDIA H100 GPUs. |
| SP001 | Andromeda | Andromeda | The platform sources, benchmarks, and prices compute across 100+ providers in real time. |
| SP002 | Andromeda | Providers - Andromeda | Qualified demand routes to your capacity. Deals execute on standardized terms. |
| SP003 | Andromeda | Andromeda.ai Trust Center | |
| SP004 | CoreWeave | The Essential Cloud for AI | CoreWeave | CoreWeave Cloud is an AI-native platform purpose-built for AI. |
| SP005 | CoreWeave | CoreWeave Cloud Platform | CoreWeave offers flexible pricing models for reserved and on-demand GPU capacity, with transparent usage-based billing and no hidden or egress fees. |
| SP006 | CoreWeave | CoreWeave Cloud Pricing | CoreWeave offers up to 60% discounts over our On-Demand prices for committed usage. |
| SP007 | CoreWeave Investor Relations | Record First Quarter Revenue and Revenue Backlog Highlight Unprecedented Demand for CoreWeave Cloud | We surpassed 1 GW of active power and believe we are well on our way to more than 8 GW by 2030. |
| SP008 | Fitch Ratings | Fitch Rates CoreWeave's Proposed New Notes 'BB-'/'RR4' | In 2024, Microsoft represented 62% of CoreWeave's revenue, with the top two customers combined accounting for 77%. |
| SP009 | Lambda | The Superintelligence Cloud | Lambda | Protect sensitive data with a single-tenant, shared-nothing architecture. Achieve production-grade compliance with SOC 2 Type II certification. |
| SP010 | Lambda | AI Cloud Pricing | Lambda | Deploy NVIDIA B200, H100, A100, or GH200 instances in minutes with self-serve, first-come access. |
| SP011 | Lambda | Trust | Lambda | SOC 2 Type II attestation. |
| SP012 | BusinessWire / Lambda | Lambda Raises $320M to Build a GPU Cloud for AI | Lambda has amassed over 100,000 customer sign-ups on Lambda Cloud. |
| SP013 | Verdict | AI infrastructure company Lambda secures $480m in Series D | Lambda has secured $480m in Series D funding round, bringing its total equity raised to $863m. |
| SP014 | Vast.ai | Vast.ai | Prices set by supply and demand across 20,000+ GPUs. |
| SP015 | Vast.ai | GPU Pricing — Live Platform Rates - Vast AI | No long-term contracts required. Scale up, scale down, or switch GPU types anytime without penalties. |
| SP016 | Vast.ai | Documentation | Vast.ai | Providers retain full control over pricing and contracts. |
| SP017 | SambaNova | SambaNova | The Fastest AI Inference Platform | SambaCloud was the first platform to support all three variants of Llama 3.1 with fast inference. |
| SP018 | SambaNova | SambaNova Systems Trust Center | |
| SP019 | BusinessWire / SambaNova | SambaNova Unveils Fastest Chip for Agentic AI, Collaborates with Intel, and Raises $350M | SambaNova has obtained $350 million in strategic Series E financing to expand manufacturing and cloud capacity. |
| SP020 | NVIDIA | Accelerate AI and Machine Learning Workflows | NVIDIA Run:ai | With support for public clouds, private clouds, hybrid environments, or on-premises data centers, NVIDIA Run:ai provides unparalleled flexibility and adaptability. |
| SP021 | NVIDIA Run:ai | NVIDIA Run:ai Documentation | NVIDIA Run:ai accelerates AI operations with dynamic orchestration across the AI life cycle. |
| SP022 | TechCrunch | Nvidia completes acquisition of AI infrastructure startup Run:ai | Run:ai said its software, which currently only works with Nvidia products, will be open sourced. |
| SP023 | European Commission | Commission approves acquisition of Run:ai by NVIDIA | Run:ai does not have a significant position on the market for GPU orchestration software today. |
| SP024 | Runpod | The AI Developer Cloud | Runpod | Trusted by over one million developers at the world's leading AI companies. |
| SP025 | Runpod | GPU Cloud Pricing | Runpod | Pods are dedicated GPU instances for development and long-running jobs, Serverless bills inference workers based on usage, and Clusters support multi-node workloads and reserved capacity. |
| SP026 | Runpod | AI Infrastructure Security and Compliance | Runpod | Your endpoints can be isolated to run exclusively on data centers that tick every box on your compliance checklist. |
| SP027 | Runpod Docs | Serverless Pricing - Runpod Documentation | Serverless offers pay-per-second pricing with no upfront costs. |
| SP028 | Modal | Modal: High-performance AI infrastructure | Modal routes workloads across clouds and regions in real time. Get the GPUs you need in seconds, with no commitments or capacity planning. |
| SP029 | Modal | Plan Pricing | Modal | Early-stage startups can get free compute credits on Modal. |
| SP030 | Modal Docs | Security and privacy at Modal | Modal Docs | We have successfully completed a System and Organization Controls (SOC) 2 Type 2 audit. |
| SP031 | Modal Docs | Introduction | Modal Docs | You get full serverless execution and pricing because we host everything and charge per second of usage. |
| SP032 | AWS | AWS Compliance Programs | Compliance certifications and attestations are assessed by a third-party, independent auditor. |
| SP033 | Google Cloud | Compliance resource center | Google Cloud's industry-leading certifications, documentation, and third-party audits help support your compliance. |
| SP034 | Microsoft Azure | Azure compliance offerings | Microsoft Azure leads the industry with more than 100 compliance offerings. |
| SP035 | NVIDIA | NVIDIA DGX SuperPOD | Scaling to tens of thousands of NVIDIA GPUs, NVIDIA DGX SuperPOD tackles training and inference for state-of-the-art generative AI models. |
| SI001 | Andromeda | Andromeda | The platform sources, benchmarks, and prices compute across 100+ providers in real time. |
| SI002 | Andromeda | Andromeda customer intake form | GPU Type; How many GPUs do you need?; When would you like to start?; For how many weeks do you need the GPUs? |
| SI003 | Andromeda | Providers - Andromeda | Qualified demand routes to your capacity. Deals execute on standardized terms. |
| SI004 | Andromeda | A view from billions of GPU-Hours | What this market needs is the full infrastructure of trade: sourcing, quality certification, standardized contracts, structuring, matching, operations, and the trust to make it work at scale. |
| SI005 | Andromeda | Andromeda · Access the world's compute | Andromeda exists to make high-performance compute available to every team building at the frontier. |
| SI006 | Upstarts | The $1.5B Compute Startup Helping AI's Hottest Companies Find GPUs | Andromeda quietly passed a revenue run rate of $100 million in 2025, up from $50 million-plus the year before; operating profitably since its start. |
| SI007 | SiliconANGLE | On-demand GPU startup Andromeda raises funding at $1.5B valuation | Andromeda’s annualized revenue run rate reportedly grew from $50 million in 2024 to $100 million last year. It has operated profitably since launch. |
| SI008 | Data Center Dynamics | Meta in talks to partially acquire VC fund NFDG, hire Nat Friedman and Daniel Gross for AI shakeup | Andromeda is now available for non-NFDG companies, at $2.40-3.00/GPU/hour. We currently have up to 2,000 H100s available and can give you access to GPUs within a few hours. |
| SI009 | Built In | Andromeda (andromeda.ai) Careers, Perks + Culture | Built In | HQ San Francisco 17 Total Employees. |
| SI010 | CoreWeave | The Essential Cloud for AI | CoreWeave | CoreWeave Cloud is an AI-native platform purpose-built for AI. |
| SI011 | CoreWeave | CoreWeave Cloud Pricing | CoreWeave offers up to 60% discounts over our On-Demand prices for committed usage. |
| SI012 | CoreWeave Investor Relations | Record First Quarter Revenue and Revenue Backlog Highlight Unprecedented Demand for CoreWeave Cloud | Revenue backlog was $99.4 billion as of March 31, 2026. Secured first-of-its-kind DDTL 4.0 Facility, an $8.5 billion non-recourse delayed draw term loan facility. |
| SI013 | Securities and Exchange Commission | CoreWeave, Inc. Annual Report (Form 10-K) | Customers generally access our platform through multi-year committed contracts, under which they purchase a specified amount of capacity on a take-or-pay basis over the contract term. As of December 31, 2025, we had $60.7 billion of remaining performance obligations. |
| SI014 | Securities and Exchange Commission | CoreWeave, Inc. Quarterly Report (Form 10-Q) | We currently sell access to our platform either through committed contracts, which are take-or-pay, or on-demand, which are pay-as-you-go. For each of the three months ended March 31, 2026 and 2025, committed contracts accounted for 98% of our revenue. |
| SI015 | Securities and Exchange Commission | CoreWeave, Inc. Current Report (Form 8-K) — May 7, 2026 | On May 7, 2026, CoreWeave, Inc. issued a press release announcing its financial results for the fiscal quarter ended March 31, 2026. |
| SI016 | Securities and Exchange Commission | CoreWeave, Inc. Current Report (Form 8-K) — May 15, 2026 | The DDTL 5.0 Facility was entered into primarily to finance capital expenditures required to perform certain customer contracts, including the acquisition of GPU servers and related infrastructure. |
| SI017 | Fitch Ratings | Fitch Rates CoreWeave's Proposed New Notes 'BB-'/'RR4' | In 2024, Microsoft represented 62% of CoreWeave's revenue, with the top two customers combined accounting for 77%. |
| SI018 | Lambda | AI Cloud Pricing | Lambda | Deploy NVIDIA B200, H100, A100, or GH200 instances in minutes with self-serve, first-come access. |
| SI019 | Runpod | GPU Cloud Pricing | Runpod | Pods are dedicated GPU instances for development and long-running jobs, Serverless bills inference workers based on usage, and Clusters support multi-node workloads and reserved capacity. |
| SI020 | Vast.ai | GPU Pricing — Live Platform Rates - Vast AI | No long-term contracts required. Scale up, scale down, or switch GPU types anytime without penalties. |
| SI021 | Securities and Exchange Commission | DigitalOcean Holdings, Inc. Annual Report (Form 10-K) | While our pricing is primarily consumption-based and the majority of our customers use our platform on a month-to-month basis, a growing number of customers are using our platform for larger workloads and some of these customers are opting to enter into committed contracts. |
| SI022 | Securities and Exchange Commission | DigitalOcean Holdings, Inc. Quarterly Report (Form 10-Q) | Gross profit decreased to 56% for the three months ended March 31, 2026 from 61% for the three months ended March 31, 2025. The decline in gross margin resulted from incurrence of costs for data center expansions in advance of the ramp in revenue from new data centers. |
| SI023 | Securities and Exchange Commission | DigitalOcean Holdings, Inc. Current Report (Form 8-K) — May 5, 2026 | The proceeds of the revolving credit facility may be used for working capital, capital expenditures, permitted acquisitions, refinancing of indebtedness and other general corporate purposes. |
| SI024 | Securities and Exchange Commission | Amazon.com, Inc. Annual Report (Form 10-K) | Generally, we recognize gross revenue from items we sell from our inventory as product sales and recognize our net share of revenue of items sold by third-party sellers as service sales. |
| SI025 | Securities and Exchange Commission | Amazon.com, Inc. Quarterly Report (Form 10-Q) | Cash capital expenditures were $43.2 billion during Q1 2026, which primarily reflect investments in technology infrastructure (the majority of which is to support AWS business growth). |
| SE001 | Andromeda | Andromeda · Access the world's compute | Define your workload. GPU type, quantity, region, timeline. We source and benchmark. The platform sources, benchmarks, and prices compute across 100+ providers in real time. |
| SE002 | Andromeda | Andromeda · Access the world's compute | What this market needs is the full infrastructure of trade: sourcing, quality certification, standardized contracts, structuring, matching, operations, and the trust to make it work at scale. |
| SE003 | Andromeda | Andromeda · Access the world's compute | GPU Type. How many GPUs do you need? When would you like to start? For how many weeks do you need the GPUs? |
| SE004 | Andromeda | Andromeda · Access the world's compute | GPU Type, Number of GPUs, Price GPU/Hr, Interconnect Network, Node RAM, Geographic Location, Cloud Service Provider, Cluster Interface, Minimum Bookable GPUs, Min Bookable Duration. |
| SE005 | Andromeda Cluster, Inc. | Privacy Policy | Andromeda Cluster, Inc. and its affiliates operates a global GPU compute marketplace platform connecting Customers with Providers. |
| SE006 | Andromeda | Andromeda.ai Trust Center | Andromeda.ai Trust Center |
| SE007 | Andromeda | Careers at Andromeda | Andromeda exists to make high-performance compute available to every team building at the frontier. |
| SE008 | X | Andromeda (@andromeda_ai) / X | Today we're announcing Andromeda. For three years, we've built the market infrastructure for compute. |
| SE009 | Built In | Andromeda (andromeda.ai) Jobs + Careers | Built In | The Site Reliability Engineer will provision and manage Kubernetes clusters, build automation tools, debug customer issues, and improve infrastructure reliability. |
| SE010 | Built In | Software Engineer - AI Infrastructure - Andromeda (andromeda.ai) | Build robust APIs, services, and control planes that abstract over diverse infrastructure types (VMs, Kubernetes, bare metal, schedulers). |
| SE011 | Built In | Site Reliability Engineer - AI Infrastructure - Andromeda (andromeda.ai) | Provision, configure, and operate Kubernetes-based clusters for customers across multiple providers. |
| SE012 | Built In | Senior Site Reliability Engineer - AI Infrastructure - Andromeda (andromeda.ai) | Design and evolve multi-provider, multi-region GPU compute clusters optimized for large-scale training. |
| SE013 | Built In San Francisco | Solutions Engineer - Andromeda (andromeda.ai) | Own the technical POC process: define scope, success metrics, architecture, timeline, and stakeholder alignment; ensure clean handoffs into production. |
| SE014 | RemoteOK | Remote Senior Site Reliability Engineer AI Infrastructure at Andromeda Cluster | Reliability & Performance Engineering: Define SLOs and error budgets that account for the unique failure modes of GPU infrastructure. |
| SE015 | Andromeda Cluster | Andromeda Cluster Jobs | Andromeda Cluster Jobs |
| SE016 | Andromeda Cluster | Software Engineer - AI Infrastructure @ Andromeda Cluster | Software Engineer - AI Infrastructure @ Andromeda Cluster |
| SE017 | Andromeda Cluster | Solutions Engineer @ Andromeda Cluster | Solutions Engineer @ Andromeda Cluster |
| SE018 | Andromeda Cluster | Senior Site Reliability Engineer - AI Infrastructure @ Andromeda Cluster | Senior Site Reliability Engineer - AI Infrastructure @ Andromeda Cluster |
| SE019 | Andromeda Cluster | Site Reliability Engineer - AI Infrastructure @ Andromeda Cluster | Site Reliability Engineer - AI Infrastructure @ Andromeda Cluster |
| SE020 | Andromeda Cluster | Staff SRE, AI Infrastructure @ Andromeda Cluster | Staff SRE, AI Infrastructure @ Andromeda Cluster |
| SE021 | Andromeda Cluster | General Interest - Experience w/ AI Infrastructure @ Andromeda Cluster | General Interest - Experience w/ AI Infrastructure @ Andromeda Cluster |
| SE022 | Upstarts | The $1.5B Compute Startup Helping AI's Hottest Companies Find GPUs | It's a technical challenge to vet and verify access from a wide range of sources, then quickly handle the deal structuring and contractual agreements. |
| SE023 | SiliconANGLE | On-demand GPU startup Andromeda raises funding at $1.5B valuation | Before the company makes a GPU provider’s infrastructure available through its platform, it checks that the hardware meets performance and security requirements. |
| SE024 | Data Center Dynamics | Meta in talks to partially acquire VC fund NFDG, hire Nat Friedman and Daniel Gross for AI shakeup | The Andromeda Cluster launched with 2,512 H100s GPUs, and has since grown to 3,200 H100s on 400 nodes interlinked with 3.2Tbps InfiniBand. |
| SE025 | GitHub | site/files/Dan_Yasny_CV.md at 203e41f7e26f6a6ab9e4ac59f19622aa868780bc · dyasny/site | Platform Engineering around AI and GPU baremetal on-premise infrastructure across multiple top tier GPU providers and major cloud providers. |
| SU001 | Andromeda | Andromeda · Access the world's compute | Andromeda connects AI teams with high-performance compute fast, at scale, and on terms that work. |
| SU002 | Andromeda | Insights | For the last three years, we have helped leading AI companies source, onboard, and operate large-scale training infrastructure. |
| SU003 | Andromeda | Customer intake form | How many GPUs do you need? When would you like to start? For how many weeks do you need the GPUs? |
| SU004 | Andromeda | Provider onboarding form | GPU Type... Number of GPUs... Price GPU/Hr... Geographic Location... Cloud Service Provider... Cluster Interface... Minimum Bookable GPUs. |
| SU005 | Andromeda | Careers at Andromeda | Careers at Andromeda |
| SU006 | Andromeda | Privacy Policy | Andromeda... operates a global GPU compute marketplace platform... connecting AI teams and enterprises... with infrastructure providers. |
| SU007 | Andromeda | Andromeda.ai Trust Center | Andromeda.ai Trust Center |
| SU008 | gpulist | gpulist | Currently serving $1.84B of listings. |
| SU009 | AI Grant | AI Grant | AI Grant companies — batch 2 ... Pika ... batch 1 ... Perplexity ... Cursor. |
| SU010 | Upstarts Media | AI Compute Startup Andromeda Raises $60M At $1.5B Valuation | Fast-growing companies in the portfolio like voice unicorn ElevenLabs and video generation startup Pika took advantage. |
| SU011 | SiliconANGLE | On-demand GPU startup Andromeda raises funding at $1.5B valuation | The company will use its new funding to grow its customer base, which reportedly includes multiple notable AI startups and labs... |
| SU012 | Data Center Dynamics | Meta in talks to partially acquire VC fund NFDG, hire Nat Friedman and Daniel Gross for AI shakeup | With compute less constrained, Andromeda is now available for non-NFDG companies, at $2.40-3.00/GPU/hour. |
| SU013 | The SaaS Sentinel | GPU Startup Andromeda Reaches $1.5B Valuation After Paradigm Investment | Portfolio companies like voice startup ElevenLabs and video generation company Pika used the infrastructure. |
| SU014 | SemiAnalysis / ClusterMAX | Andromeda Review 2026: Unavailable Tier GPU Cloud | Interesting service we cannot verify yet (not launched, sold out, government-only, etc.). |
| SU015 | Build MVP Fast | Compute-for-Equity: a16z Oxygen GPU Seed Funding | Their portfolio reads like a who's who of AI breakout companies: ElevenLabs, Pika, Cursor, Perplexity, Browserbase. |
| SU016 | Built In | Andromeda (andromeda.ai) Careers, Perks + Culture | 17 Total Employees. |
| SU017 | Built In | Andromeda (andromeda.ai) Jobs + Careers | Provide end-to-end commercial legal support: draft and negotiate revenue-critical agreements... support corporate governance, compliance (privacy, export)... |
| SU018 | Built In San Francisco | Solutions Engineer - Andromeda (andromeda.ai) | We're a seed-stage AI infrastructure startup powering large-scale training and inference for frontier labs, AI-native startups, and enterprises. |
| SU019 | Ashby | Andromeda Cluster Jobs | Head of Partnerships... Head of Procurement/Supply Chain... Compute Trader... Solutions Engineer... Staff SRE. |
| SU020 | Ashby | Solutions Engineer @ Andromeda Cluster | Own the technical POC process... ensure clean handoffs into production. |
| SU021 | Ashby | Head of Partnerships @ Andromeda Cluster | Develop partnerships with VC funds and accelerators to channel their portfolio companies' compute needs to Andromeda. |
| SU022 | Ashby | Head of Procurement/Supply Chain @ Andromeda Cluster | Today we have already established the fundamental layer of capacity with providers. |
| SU023 | Ashby | Compute Trader @ Andromeda Cluster | Match incoming leads from our sales team with internal capacity and external capacity in the market. |
| SU024 | Ashby | Member of the Business Staff - Compute Markets @ Andromeda Cluster | Maximize utilization of our compute resources. |
| SU025 | Ashby | Staff SRE, AI Infrastructure @ Andromeda Cluster | When a top-customer training run degrades... you're the engineer who walks the stack... until the answer is found. |
| SU026 | CompuX | CompuX vs Andromeda AI: Credit Marketplace vs Neutral GPU Broker | Andromeda focuses on procurement efficiency at scale... but without any financing or credit mechanism. |
| SU027 | Google Cloud / PR Newswire | ElevenLabs Partners with Google Cloud for Cloud Services and the Latest NVIDIA Blackwell GPUs | ElevenLabs will utilize Google Cloud's G4 virtual machines (VMs), powered by NVIDIA RTX PRO 6000 Blackwell GPUs, to train and serve its voice models. |
| SU028 | Lightspeed Venture Partners | Generative NYC: Pika and ElevenLabs Share the Secrets of Their Success | Lightspeed chats with the founders of Pika and ElevenLabs about the future of AI audio and video. |
| SR001 | Andromeda | Andromeda · Access the world's compute | Andromeda connects AI teams with high-performance compute fast, at scale, and on terms that work. |
| SR002 | Andromeda | Andromeda · Insights | For the last three years, we have helped leading AI companies source, onboard, and operate large-scale training infrastructure. |
| SR003 | Andromeda | Privacy Policy | Andromeda Cluster, Inc. and its affiliates operates a global GPU compute marketplace platform connecting AI teams and enterprises with infrastructure providers. |
| SR004 | Andromeda | Andromeda.ai Trust Center | Andromeda.ai Trust Center. |
| SR005 | Andromeda | Customer intake form | How many GPUs do you need? When would you like to start? For how many weeks do you need the GPUs? |
| SR006 | Andromeda | Provider onboarding form | GPU Type, Number of GPUs, Price GPU/Hr, Geographic Location, Cloud Service Provider, Cluster Interface. |
| SR007 | Andromeda | Careers at Andromeda | Open positions include Commercial Counsel, Strategic Compute Finance Lead, Head of Procurement/Supply Chain, Solutions Engineer, and Staff SRE, AI Infrastructure. |
| SR008 | Ashby | Andromeda Cluster Jobs | The jobs board lists Commercial Counsel, Strategic Compute Finance Lead, Compute Trader, Head of Partnerships, Head of Procurement/Supply Chain, and multiple SRE/software roles. |
| SR009 | Ashby | Head of Procurement/Supply Chain | We are responsible for acquiring and facilitating compute resources across the company, working closely with compute providers, sales, and technical teams to match compute supply with demand. |
| SR010 | Ashby | Strategic Compute Finance Lead | We’ve established the foundational layer of our compute portfolio and are now focused on expanding capital capacity, deepening lender relationships, and building the financial infrastructure to support our next phase of growth. |
| SR011 | Ashby | Member of the Business Staff - Compute Markets | We’re hiring a Member of the Business Staff to accelerate supply and demand matching on our platform. |
| SR012 | Ashby | Solutions Engineer | We’re hiring a Sales Engineer to lead technical discovery, run world-class demos and POCs, and partner with Sales to drive successful evaluations and expansions across strategic accounts. |
| SR013 | Built In | Andromeda Jobs + Careers | Provide end-to-end commercial legal support: draft and negotiate revenue-critical agreements, support corporate governance, compliance (privacy, export), IP, employment, vendor matters, and assist with fundraising and M&A. |
| SR014 | Built In | Andromeda Careers, Perks + Culture | We began with a single managed cluster — but it filled almost instantly. |
| SR015 | Built In San Francisco | Solutions Engineer - Andromeda | Andromeda is powering large-scale training and inference for frontier labs, AI-native startups, and enterprises. |
| SR016 | Upstarts Media | The $1.5B Compute Startup Helping AI's Hottest Companies Find GPUs | Gross and Friedman, now at Meta, have no active relationship with Andromeda anymore. |
| SR017 | Data Center Dynamics | Meta in talks to partially acquire VC fund NFDG, hire Nat Friedman and Daniel Gross for AI shakeup | What it would mean for the Andromeda Cluster is unclear. |
| SR018 | SiliconANGLE | On-demand GPU startup Andromeda raises funding at $1.5B valuation | The company will use its new funding to grow its customer base, which reportedly includes multiple notable AI startups and labs. |
| SR019 | SemiAnalysis / ClusterMAX | Andromeda Review (Unavailable) | Their model now involves procuring capacity from a range of neoclouds on our list, on behalf of the startups. |
| SR020 | CompuX | CompuX vs Andromeda AI: Credit Marketplace vs Neutral GPU Broker | Andromeda focuses on procurement efficiency at scale, but without any financing or credit mechanism. |
| SR021 | AI Grant | AI Grant | AI Grant gives seed-stage AI startups capital and cloud credits, and its batches include Pika, Cursor, and Perplexity. |
| SR022 | Google Cloud / PR Newswire | ElevenLabs Partners with Google Cloud for Cloud Services and the Latest NVIDIA Blackwell GPUs | ElevenLabs expands use of Google Cloud's AI stack, including Gemini and Veo models, and NVIDIA Blackwell GPUs. |
| SR023 | gpulist | gpulist | Currently serving $1.84B of listings. |
| SR024 | Bureau of Industry and Security | Guidance / Frequently Asked Questions | An official website of the United States government. |
| SR025 | Bureau of Industry and Security | Updated Public Information Page: Export Controls Imposed on Advanced Computing and Semiconductor Manufacturing Items | The AC/S IFR retains the licensing requirements for the PRC (including Hong Kong and Macau) and establishes a worldwide licensing requirement for companies headquartered in countries of concern or parented there. |
| SR026 | Federal Register | Framework for Artificial Intelligence Diffusion | BIS revises the Export Administration Regulations controls on advanced computing integrated circuits and adds a new control on AI model weights. |
| SR027 | OFAC / U.S. Treasury | Sanctions Programs and Country Information | Active sanctions programs include Chinese Military Companies, Cyber-Related Sanctions, Non-Proliferation Sanctions, North Korea Sanctions, and Ukraine-/Russia-related Sanctions. |
| SR028 | OFAC / U.S. Treasury | Publication of 'A Framework for OFAC Compliance Commitments' | OFAC is publishing a framework on the essential components of a sanctions compliance program for organizations subject to U.S. jurisdiction and foreign entities using U.S.-origin goods or services. |
| SR029 | Federal Trade Commission | Data Breach Response: A Guide for Business | Move quickly to secure your systems and fix vulnerabilities that may have caused the breach. |
| SR030 | Federal Trade Commission | Privacy and Security | Even if you do not make specific claims, you still have an obligation to maintain security that is appropriate in light of the nature of the data you possess. |
| SR031 | Federal Trade Commission | Start with Security: A Guide for Business | Make sure your service providers implement reasonable security measures. |
| SR032 | Federal Trade Commission | Protecting Personal Information: A Guide for Business | If sensitive data falls into the wrong hands, it can lead to fraud, identity theft, or similar harms, and perhaps even defending yourself against a lawsuit. |
| SV001 | Andromeda | Andromeda · Access the world's compute | Andromeda connects AI teams with high-performance compute fast, at scale, and on terms that work. |
| SV002 | Andromeda | Andromeda · Insights | For the last three years, we have helped leading AI companies source, onboard, and operate large-scale training infrastructure. |
| SV003 | Andromeda | Privacy Policy | Andromeda Cluster, Inc. and its affiliates operates a global GPU compute marketplace platform connecting AI teams and enterprises with infrastructure providers. |
| SV004 | Andromeda | Andromeda.ai Trust Center | Andromeda.ai Trust Center. |
| SV005 | Upstarts Media | The $1.5B Compute Startup Helping AI's Hottest Companies Find GPUs | Spun out as its own startup with backing from NFDG and Paradigm, Andromeda quietly passed a revenue run rate of $100 million in 2025. |
| SV006 | SiliconANGLE | On-demand GPU startup Andromeda raises funding at $1.5B valuation | The company offers access to infrastructure from more than 100 providers. It claims to have processed more than 1,000 GPU transactions since launching about two years ago. |
| SV007 | Gaebler / VentureDeal | Andromeda 3/18/2026 Capital Raise - Gaebler.com Venture Capital Database | Andromeda closed a $60 million funding round on 3/18/2026. Investors included Paradigm. |
| SV008 | Raising.fi | Andromeda AI Inc. Secures $60 Million in Latest Funding Round Led by Paradigm | Andromeda AI Inc. secures $60 million in latest funding round led by Paradigm. |
| SV009 | Data Center Dynamics | Meta in talks to partially acquire VC fund NFDG, hire Nat Friedman and Daniel Gross for AI shakeup | What it would mean for the Andromeda Cluster is unclear. |
| SV010 | SemiAnalysis / ClusterMAX | Andromeda Review (Unavailable) | Their model now involves procuring capacity from a range of neoclouds on our list, on behalf of the startups. |
| SV011 | AI Grant | AI Grant | Larger investments come with access to the Andromeda Cluster to supercharge infrastructure growth. |
| SV012 | Google Cloud / PR Newswire | ElevenLabs Partners with Google Cloud for Cloud Services and the Latest NVIDIA Blackwell GPUs | ElevenLabs expands use of Google Cloud's AI stack, including Gemini and Veo models, and NVIDIA Blackwell GPUs. |
| SV013 | Ashby | Andromeda Cluster Jobs | The jobs board lists Commercial Counsel, Strategic Compute Finance Lead, Compute Trader, Head of Partnerships, Head of Procurement/Supply Chain, and multiple SRE/software roles. |
| SV014 | Ashby | Strategic Compute Finance Lead | We’ve established the foundational layer of our compute portfolio and are now focused on expanding capital capacity, deepening lender relationships, and building the financial infrastructure to support our next phase of growth. |
| SV015 | Ashby | Member of the Business Staff - Compute Markets | We’re hiring a Member of the Business Staff to accelerate supply and demand matching on our platform. |
| SV016 | Ashby | Solutions Engineer | We’re hiring a Sales Engineer to lead technical discovery, run world-class demos and POCs, and partner with Sales to drive successful evaluations and expansions across strategic accounts. |
| SV017 | CompaniesMarketCap | CoreWeave (CRWV) - Market capitalization | As of July 2026 CoreWeave has a market cap of $54.30 Billion USD. |
| SV018 | StockAnalysis | CoreWeave Revenue | CoreWeave had revenue of $2.08B in the quarter ending March 31, 2026, and revenue in the last twelve months of $6.23B with a P/S ratio of 8.72. |
| SV019 | SEC | Form D Data Sets | The Form D Data Sets below provide the structured data from Notices of Exempt Offerings of Securities filed with the Commission. |
| SV020 | CoreWeave Investor Relations | Record First Quarter Revenue and Revenue Backlog Highlight Unprecedented Demand for CoreWeave Cloud | Revenue backlog was $99.4 billion as of March 31, 2026. |
| SV021 | Securities and Exchange Commission | CoreWeave, Inc. Quarterly Report (Form 10-Q) | As of March 31, 2026, we had cash, cash equivalents, and marketable securities of $2.2 billion. |
| SV022 | Securities and Exchange Commission | CoreWeave, Inc. Annual Report (Form 10-K) | Customers generally access our platform through multi-year committed contracts, under which they purchase a specified amount of capacity on a take-or-pay basis over the contract term. |
| SV023 | Fitch Ratings | Fitch Rates CoreWeave's Proposed New Notes 'BB-'/'RR4' | In 2024, Microsoft represented 62% of CoreWeave's revenue, with the top two customers combined accounting for 77%. |
| SV024 | CompaniesMarketCap | DigitalOcean (DOCN) - Market capitalization | As of July 2026 DigitalOcean has a market cap of $16.38 Billion USD. |
| SV025 | CompaniesMarketCap | DigitalOcean (DOCN) - Revenue | According to DigitalOcean's latest financial reports the company's current revenue (TTM) is $0.94 Billion USD. |
| SV026 | Securities and Exchange Commission | DigitalOcean Holdings, Inc. Annual Report (Form 10-K) | While our pricing is primarily consumption-based and the majority of our customers use our platform on a month-to-month basis, a growing number of customers are using our platform for larger workloads and some of these customers are opting to enter into committed contracts. |
| SV027 | Securities and Exchange Commission | DigitalOcean Holdings, Inc. Quarterly Report (Form 10-Q) | Gross profit decreased to 56% for the three months ended March 31, 2026 from 61% for the three months ended March 31, 2025. |
| SV028 | CompaniesMarketCap | Nebius Group (NBIS) - Market capitalization | As of July 2026 Nebius Group has a market cap of $70.11 Billion USD. |
| SV029 | CompaniesMarketCap | Nebius Group (NBIS) - Revenue | According to Nebius Group's latest financial reports the company's current revenue is $0.52 Billion USD. |
| SV030 | CompaniesMarketCap | Amazon (AMZN) - Market capitalization | Market cap: $2.563 Trillion USD. |
| SV031 | Securities and Exchange Commission | Amazon.com, Inc. Annual Report (Form 10-K) | Generally, we recognize gross revenue from items we sell from our inventory as product sales and recognize our net share of revenue of items sold by third-party sellers as service sales. |
| SV032 | Securities and Exchange Commission | Amazon.com, Inc. Quarterly Report (Form 10-Q) | Cash capital expenditures were $43.2 billion during Q1 2026, which primarily reflect investments in technology infrastructure. |