初创公司尽调
尽调报告 AI infrastructure / GPU cloud Series B 2026-10-21

TensorWave

AMD 战略协同与可见基础设施规模突出,但运营经济性仍大多未披露

TensorWave 的基础设施可信,也得到 AMD 的战略支持,但在缺少经审计的经济数据、利用率、客户质量和融资条款时,$1.55 billion 的估值参考仍显牵强。

覆盖范围与披露说明

公开资料审阅截至 2026-10-21。融资和产品公告的记录好于收入、利用率、客户集中度、融资义务、治理和单位经济性;融资额与员工数估计的冲突仍未解决。

封面要素

投后估值 02
1550 USD millions [CO007, CV002]
基础设施模式 03
AMD-focused AI cloud [CO003, CV008]
建议 04
Research more [CV016]

公司概况

TensorWave 是一家由创始人主导、总部位于 Las Vegas、2023 年成立的 AI 基础设施公司。它运营覆盖训练、微调、推理、裸金属容量、托管推理、存储、监控和技术支持的 以 AMD 为中心的 云。2026 年 6 月 Series B 据报融资 $350 million,投后估值为 $1.55 billion;但当前收入、利用率、客户集中度、留存、现金消耗、债务、治理权利和经独立验证的单位经济性均未公开披露。

官网
tensorwave.com
创始人
Darrick Horton, Jeff Tatarchuk, Piotr Tomasik
创立地点
Las Vegas, Nevada
总部
Las Vegas, Nevada
产品
基于 AMD Instinct 的 AI 训练、微调、推理和 HPC 云基础设施,包含裸金属算力、托管推理、存储、监控和工程支持。
客户
AI 开发者、企业、初创公司、学术实验室和成本敏感的 PyTorch 团队。
商业模式
以销售沟通为入口的基础设施服务,结合专用或按量计费算力、托管端点、存储和支持。
阶段
Series B
融资情况
2026 年 6 月 10 日宣布完成 $350 million Series B,投后估值据报为 $1.55 billion。
[CO001, CO002, CO003, CO007, CO015, CM002, CI001, CI002]

执行摘要

主要优势

  • 聚焦 AMD 的差异化基础设施定位,并获得 AMD Ventures 的战略支持。
  • 2026 年 6 月完成大额融资,且加速器部署可见,说明公司具备执行扩张计划的能力。
  • 产品覆盖训练、推理、裸机、存储、监控和技术支持,范围较广。

主要风险

  • 对 AMD 和 ROCm 的集中依赖,把供应、路线图、软件兼容性和融资观感都绑在单一生态上。
  • 资本密集型扩张可能跑在签约利用率前面,而债务、租赁、契约和现金续航仍未披露。
  • 客户数量、集中度、留存率、合同期限,以及生产环境与试点客户的构成,尚未得到公开确认。
  • 基准测试、定价、收入、毛利率和工作负载经济性的说法,缺少经审计或可由独立方复现的支持。

未决问题

  • 经审计的 2025-2026 年收入、毛利率、经营亏损、现金余额、现金消耗和续航。
  • 按客户拆分的签约利用率、集中度、留存率、积压订单质量和生产使用构成。
  • 债务、设备融资、租赁、电力、契约、已承诺资本开支和担保受偿顺序条款。
  • 针对目标工作负载的独立基准测试,包括迁移投入、可靠性和总成本。
  • 当前股权结构表、董事会构成、投资人控制权和继任安排。

目录

Chapter 01

01公司概览

1.1 公司定位与运营模式

TensorWave 是一家由创始人主导、总部位于 Las Vegas、2023 年成立的私有 AI 与高性能计算云公司。它围绕 AMD Instinct 加速器搭建云基础设施,覆盖训练、微调、推理、裸金属算力、托管推理、存储、监控和技术支持。公开资料足以确认其业务类别和总部,但不足以核实客户数、收入、利用率或盈利能力。[CO001, CO003, CO014, CO016]

KPI 快照与披露状态
指标数值或状态期间置信度证据缺口
成立年份2023当前公开资料中准确的注册成立和开始运营日期需查阅设立文件
总部Las Vegas,Nevada当前公开资料中确认主要执行办公室和运营地点
阶段与估值Series B;投后估值 $1.55B2026-06-10高取得融资文件和完全摊薄后的资本结构
累计融资$493M 对比 $450M+截至 2026-10-21低核对融资所得、债务、SAFE 及轮次分类
员工数177 vs 1002026-09 vs 2026-08低核实薪资员工数、承包商和地域拆分
收入或 ARR未公开披露截至 2026-10-21低取得经审计收入、ARR 桥接表和收入确认政策
客户数未公开披露截至 2026-10-21低取得客户名单、集中度和生产使用结构
商业模式以 AMD 为中心的 AI/HPC 云当前公开资料中核实定价、利用率和合同期限

截至运行日的公开资料快照;融资额和员工数的冲突口径并列展示,未披露的运营指标不作估算。

[CO001, CO003, CO007, CO008, CO009, CO010, CO011]

1.2 创始人领导与治理透明度

运营团队由创始人主导。现有资料显示,Darrick Horton 任 CEO,Jeff Tatarchuk 任 首席增长官,Piotr Tomasik 任总裁或 COO。三人分别覆盖技术领导、商业拓展和运营,也让公司对外叙事高度集中在创始团队。现有证据未披露现任董事会、投资人同意权、投票控制、关联方政策或继任规划,因此无法可靠评估机构监督和关键人风险。该领导层表只是公开资料样本,不是完整的管理层或董事会名册;首要尽调应取得公司认证的组织架构和治理记录。[CO002, CO015]

领导层与创始人表
人员公开职务相关背景信号职能覆盖关键人依赖
Darrick Horton联合创始人兼 CEOGrowth Engineer 描述的过往技术与基础设施领导经历战略与技术方向高:公开战略与 CEO 高度绑定
Jeff Tatarchuk联合创始人兼 首席增长官Growth Engineer 描述的过往公司创建经历商业增长与合作伙伴中高:具名创始人负责增长职能
Piotr Tomasik联合创始人;总裁兼 COOGrowth Engineer 描述的软件与运营背景运营与执行中高:创始人主导运营覆盖

保留资料中公开出现的创始人和高管职务样本;不包括未披露高管、董事、继任安排以及资料未反映的职务变化。

[CO002, CO015]

1.3 融资、公开规模与里程碑冲突

TensorWave 的融资轮次不断扩大,最终在 2026 年 6 月 10 日宣布完成 $350 million Series B,投后估值据报为 $1.55 billion。更早的融资记录并不一致:一家数据库将 2024 年 10 月的 $43 million 融资称为 Series A,另一份资料称其为 SAFE;现有资料给出的累计融资额也有 $493 million 和 $450 million-plus 两种说法。员工数估计同样分歧:一份 2026 年 8 月资料显示 100 人,Tracxn 在 9 月 26 日显示 177 人。这些口径不能互换,应等股权结构表、薪资和融资记录出具后再定论。AMD 既是平台供应商又是投资人,可能加快扩容,但独立报道也提示循环融资和供应商集中风险。与 Modular 及 AMD Helios 的产品公告说明生态仍在推进,却不能证明已签约的利用率或持久收入。[CO004, CO005, CO006, CO007, CO008, CO009, CO010, CO011, CO012, CO013, CO014]

利益相关方或投资人图谱
利益相关方公开职务经济或控制相关性尽调要求
AMD 或 AMD Ventures战略投资人兼技术供应商平台集中度与融资关系拆分商业采购、投资所得、返利和供应承诺
Magnetar2025 年和 2026 年轮次的领投方连续领投可能带来重要治理权利取得董事会、观察员、同意权、清算权和保护性权利条款
Nexus Venture Partners早期及后续投资人长期融资关系确认持股、按比例认购权、董事会历史和信息权
Maverick SiliconSeries A 和 Series B 参与方后续机构资本确认证券类别、持股、附函和参与权
Western FrontierSeries B 参与方最新轮次经济敞口确认配额、证券条款和任何战略义务

基于公开融资资料和 2026 年 6 月轮次的部分图谱;这不是 股权结构表,也不能证明当前持股比例或控制权。

[CO004, CO005, CO006, CO007, CO008, CO009, CO014, CO015]
FO001: TensorWave 公开里程碑时间线

公开里程碑显示融资和产品合作推进迅速,但第三方资料中的融资分类和规模数据仍不一致。

仅标注月份的日期沿用来源精度;里程碑数值未作插值,彼此冲突的融资轮次分类和员工数估计均予以保留。

公司成立 [CO001]

据报道完成 $43M 融资 [CO004, CO005]

据报道完成 $100M Series A [CO006]

宣布与 Modular 合作 [CO013]

宣布完成 $350M Series B [CO007]

宣布支持 AMD Helios [CO012]

独立资料称有 100 名员工 [CO011]

Tracxn 报道有 177 名员工 [CO010, CO011]

[CO001, CO004, CO005, CO006, CO007, CO010, CO011, CO012, CO013]

1.4 展品

Chapter 02

02市场分析

2.1 市场边界与替代方案

应把 TensorWave 放在专业 AI 基础设施中评估,而不是拿它去对比全部云计算支出。其核心品类是面向训练、微调、推理、存储、编排和监控的托管 AMD Instinct 算力;相邻支出包括通用云 GPU 服务,自建加速器和以 NVIDIA 为中心的新型云则是同一工作负载的替代方案。经济问题在于,客户需要高性能加速器,又不愿接受单一供应商的软件和供应链。边界仍然很宽:Tracxn 将 TensorWave 放在数百家活跃竞争者之中,说明市场拥挤,不等于存在收入机会。更窄的投资问题是:有多少生产需求既能技术性迁移到 ROCm,又愿意承担迁移和供应商集中风险。[CM001, CM003, CM010, CM013, CM014]

市场定义与替代边界
分部或类别纳入支出或角色排除支出买方或付款方对 TensorWave 的相关性
AMD 专用 GPU 云面向 AI/HPC 的托管 AMD Instinct 算力、存储、编排、监控和工程支持无关的通用算力和非加速云服务AI 基础设施负责人;企业或 AI 公司算力预算核心市场
通用云 GPU 服务可迁移至专业基础设施的可比训练或推理工作负载不含加速器工作负载的通用云支出云平台负责人;企业云预算相邻市场与现状替代方案
以 NVIDIA 为中心的专业云等效 GPU 容量、托管集群和推理服务与 TensorWave 在技术上可承载的工作负载无关的供应商收入ML 平台或基础设施负责人;算力预算直接替代方案与定价基准
自建或托管于机房的加速器面向愿意自持硬件并自行运营的组织的内部构建替代方案非 AI 数据中心基础设施CIO、基础设施负责人或研究机构资本密集型替代方案

边界是分析口径,而非出版商定义的市场规模;现有来源没有提供各细分市场可比的收入 数据。

[CM001, CM003, CM010, CM013, CM014]

2.2 受证据约束的市场规模

现有证据衡量的是 TensorWave 的供给规模,而不是整个市场的收入。2026 年 6 月公司公告称线上有 8,192 个 MI325X GPU,并拥有超过 2 GW 的长期数据中心容量;TechTimes 报道约 10,000 个 AMD GPU、三个数据中心、14 MW 运营容量、500 MW 租赁容量和 2 GW 目标;ClusterBid 则估计 2026 年第二季度两个站点合计约 3,000 个 MI300X GPU。这些快照使用了不同 GPU 型号、日期和容量口径,不能相加,也不能换算成 TAM。它们说明扩张野心和潜在需求,但已安装 GPU、电力权利和市场支出是不同分母。现有资料没有单独给出可服务收入或 TensorWave 可获得的市场份额。[CM004, CM005, CM006, CM007, CM014, CM015]

用于市场规模测算的容量与品类代理指标
出版商与期间披露指标指标含义置信度局限
TensorWave 公告;2026-06-108,192 个 MI325X GPU 在线;已锁定超过 2 GW 的长期数据中心容量当前公司自称的机群规模与长期供给雄心中公司发布;在线 GPU 和已锁定电力都不等于收入 TAM
TechTimes;2026-06-11约 10,000 个 AMD GPU 分布于三个数据中心,总计 14 MW;租赁 500 MW,目标为 2 GW关于运营规模与扩张承诺的独立报道中引用了另一出版物;租赁或目标电力不等于已利用容量
ClusterBid;2026 年第二季度达拉斯与菲尼克斯约有 3,000 个 MI300X GPU;法兰克福在建按模型划分的机群估算与地理覆盖低与更新或口径不同的机群快照冲突;方法未披露
Tracxn;访问于 2026-10-21579 家活跃竞争者,其中 89 家获得融资广义品类拥挤度与替代方案可得性低该分类远宽于 AMD 专用 GPU 云,也不是支出估算

这些指标的单位、日期、模型口径和分母各不相同,因此作为独立的规模观察窗口保留 ,不做汇总。

[CM004, CM005, CM006, CM007, CM014, CM015]
FM001: 受证据约束的可触达市场收窄

可触达市场从广义 AI/HPC 基础设施需求,收窄至可迁移到 AMD、且在 TensorWave 上具备经济吸引力的生产工作负载。

各层宽度仅表示层级,不代表数值比例。

仅作定性收窄;层宽和标签不代表美元金额、转化率或市场份额。

[CM001, CM003, CM009, CM010, CM015]

2.3 买方、用户与采用路径

证据最充分的买方是 AI 原生模型开发商、企业 AI 项目、成本敏感的 PyTorch 团队,以及需要专门训练或推理容量的研究用户。基础设施负责人或技术创始人选择供应商,机器学习工程师使用服务,最终由组织从云或算力预算中付款。采用路径从工作负载兼容性和基准测试开始,随后经过迁移、可靠性验证、安全审查和生产扩容。Fireworks AI 与 Luma AI 提供了公司自述的生产案例,但现有资料没有客户分母、留存指标或生产与试点的拆分。TensorWave 官网声称具备企业认证和支持能力,但这些说法不能替代独立采购证据或客户级利用率数据。[CM002, CM009, CM012, CM016]

细分市场、买方、用户与采用路径
细分市场买方与预算负责人主要用户工作流采用触发因素与约束
AI 原生模型构建者CTO 或基础设施负责人;算力预算ML 基础设施与模型工程师大模型训练与生产推理容量与内存经济性;受可移植性和可靠性证明约束
企业 AI 项目CIO、AI 副总裁或云平台负责人;企业云预算ML 平台与应用团队从试验转向生产部署支持、安全性与可预测成本;受采购流程和 CUDA 依赖约束
成本敏感型 PyTorch 团队技术创始人或工程负责人;产品基础设施预算训练与推理工程师在内存密集型模型上进行 PyTorch 训练或批量推理报价中的 GPU 小时成本更低;受特定工作负载性能约束
研究与学术用户首席研究员或算力负责人;机构或科研经费预算研究人员与 HPC 从业者训练、微调、嵌入与检索工作负载获得专用算力;受迁移工作量和支持要求约束

买方角色根据现有来源描述的客户群体与采用机制推断;没有 采购记录或预算调查可供参考。

[CM002, CM008, CM009, CM010, CM012, CM016]

2.4 增长驱动、约束与估值相关性

需求受到加速器稀缺、GPU 经济性压力、高显存工作负载以及买方降低 NVIDIA 依赖的意愿支撑。TensorWave 只做 AMD,可能改善部分工作负载的性价比,但独立基准证据薄弱且高度依赖具体负载。ClusterBid 报告了更低的挂牌价格和大批量推理优势;AMD 案例则复述 TensorWave 与 Modular 的性能说法,并明确表示 AMD 未独立验证。ROCm 兼容性、CUDA 专用 kernel、迁移工作量、可靠性,以及客户是否愿意让核心生产负载运行在其上,仍是采用约束。AMD 的投资可能强化生态,同时也集中供应商敞口并带来循环融资疑虑。因此,估值应跟踪独立验证的利用率、已签约生产需求、可迁移成本和电力到收入的转化,而不能只看公告容量。[CM003, CM008, CM009, CM010, CM011, CM013, CM015]

2.5 展品

Chapter 03

03竞争者

3.1 竞争格局与相对规模

TensorWave 面对的是拥挤的 AI 基础设施市场,而不是与 Nvidia 的简单双强竞争。Tracxn 将 CoreWeave、Together AI、FluidStack、Rescale 和 Crusoe 列为相关的融资竞争者;买方也可以继续使用通用云,或自行建设和运营基础设施。所选竞争者的商业模式和融资年份差异很大,因此融资额只能作为规模信号,不能作为产品质量排名。TensorWave 据报累计融资 $493 million,对年轻供应商而言并不少,但低于若干列出的同业。与此同时,Data Center Dynamics 和 TechTimes 报告其运营着数千个 AMD GPU 和多个设施。这些报道支持其确有部署足迹,但公司口径的多 GW 容量目标不应解读为已安装或已利用容量。[CP006, CP007, CP008, CP009, CP010, CP011]

选定竞争者与替代方案画像
替代方案类别公开规模或融资信号目标任务差异化现有证据的局限
TensorWaveAMD 专用新云总融资 $493M;B 轮训练、微调与推理AMD 专属优化与工程支持已部署容量利用率与客户集中度未公开
CoreWeave规模化新云总融资 $2.38B;已上市AI 训练与推理云据报道更大的资本规模未保留可比的产品、价格与服务数据
Together AIAI 平台与云总融资 $1.33B;C 轮模型开发、微调与部署Tracxn 描述中更广的模型平台范围未保留可比的基础设施与合同条款
FluidStackAI/HPC 基础设施总融资 $3.28B;B 轮大规模 AI 算力据报道更大的资本规模未保留可比的加速器组合与服务水平
CrusoeAI 云与数据中心运营商总融资 $9.8B;F 轮AI 基础设施工作负载据报道远大的资本规模未保留可比的产品与定价数据
通用云或内部构建既有方案 / 现状替代方案不可比沿用现有采购,或掌控整套技术栈现有工作流或直接运营控制没有保留可匹配的成本、性能或切换数据集

选定的非穷尽式竞争格局。融资与轮次信号来自 Tracxn 当前页面,未 按日期、债务、估值或商业模式差异做标准化;通用云 / 内部构建一行是 分析层面的买方替代方案。

[CP001, CP006, CP010, CP011, CP015]
FP001: 部分竞争者资本规模与 Tracxn 得分对比

按已披露融资额,TensorWave 小于所选的较大型 AI 基础设施公司;其出版方定义的得分也更低。

条目Tracxn 得分 (0–100)总融资额 (0–10,000)说明
1. TensorWave 59 493 在所引页面的 579 家活跃竞争者中排名第 30。
2. CoreWeave 80 2,380 Tracxn 所选名单中的公开 AI 云服务商。
3. Together AI 73 1,330 Series C 轮模型平台和云服务商。
4. FluidStack 70 3,280 Series B 轮 AI/HPC 基础设施服务商。
5. Rescale 71 369 Series D 轮面向工程和科学的 HPC 平台。
6. Crusoe 66 9,800 Series F 轮 AI 云和基础设施运营商。

融资额按 Tracxn 显示值以百万美元计;得分是 Tracxn 专有的出版方定义指标,不是业绩基准。坐标轴无法证明产品质量、估值、利用率或功能可直接比较。

[CP010, CP011]

3.2 能力切口与买方选择

TensorWave 最清晰的切入口是专注,而不是广度。官网将 AMD Instinct 与 ROCm 平台定位为覆盖训练、微调和推理,并列出四代加速器。AMD 的评估页面还提出显存容量、分区、安全和工程支持等说法。这使其对重视 AMD 显存、供应商多元化或直接优化支持的工作负载成为可信选项,但不能证明在所有负载上更优。现有证据没有为大多数竞争者提供可比的安全、可靠性、网络或支持数据,因此相关单元格仍为未知。AMD 案例同样复述了大幅吞吐和节省成本的说法,并明确表示这些说法由 TensorWave 或 Modular 提供、未经 AMD 独立验证。[CP001, CP002, CP003, CP004, CP005]

保留不受支持单元格的能力矩阵
购买标准TensorWave选定新云同业通用云内部构建
加速器策略仅用 AMD Instinct;列出 MI300X 至 MI455X现有证据未知现有证据未知客户自选硬件;经济性未知
工作负载范围训练、微调、推理与 HPC选定同业披露的训练 / 推理或模型平台范围现有证据未知取决于客户工程范围
软件可移植性ROCm,加上公司 / 合作伙伴声称的 Modular 可移植性现有证据未知现有证据未知客户自行负责移植与兼容性工作
安全态势公司 / AMD 页面称环境符合 SOC 2 Type II 和 HIPAA现有证据未知现有证据未知取决于客户控制措施与审计范围
优化支持公司与 AMD 宣传提供 24/7 AMD 专项工程支持现有证据未知现有证据未知客户自行配置人员并运营优化职能

单元格仅概括分配到的预取证据。“未知”表示现有来源集不支持 可比断言;公司与 AMD 合作伙伴的说法不视为独立基准。

[CP001, CP002, CP003, CP004, CP005, CP011]

3.3 定价、打包与市场路径

公开定价证据是比较中最薄弱的一环。ClusterBid 发布了 TensorWave、CoreWeave 和 Lambda 的具体小时价格,但同一页面中的规模和融资说法与更有力的 2026 年 6 月报道冲突。因此,这些价格只能作为低置信度信号,不能作为投资判断事实。本文审阅的 TensorWave 和 AMD 官方页面强调评估、托管支持、每美元性能和节省比例,而没有公开价目表、合同期限、最低利用率、出口费用、服务抵扣或实际折扣。这种销售介入式模式可能支持针对具体负载的优化,却无法进行可靠的同口径 TCO 比较。在把 GPU 小时价格视为经济优势前,应就相同型号、批处理画像、可用性目标、网络、存储和支持范围取得匹配报价。[CP004, CP005, CP012, CP013]

定价与打包信号
供应商或证据载体公开信号单位 / 合同基础置信度投资判断含义
TensorWaveClusterBid 声称按需价格为 $1.99,预留价格为 $1.25GPU 小时;预留价格对应 12 个月预订低取得当前签署版报价,并进行匹配工作负载基准测试
CoreWeaveClusterBid 声称按需价格为 $3.00,预留价格为 $2.00GPU 小时;预订期限未完整记录低不要据此推断实际价格或可比服务范围
LambdaClusterBid 声称按需价格为 $2.80,预留价格为 $1.60GPU 小时;预订期限未完整记录低不要据此推断实际价格或可比服务范围
AMD 案例研究相较 NVIDIA B200 GPU 节省约 40%–60%取决于模型的第三方结果;无合同计价单位低至中AMD 表示该说法未经独立验证,且未必具有代表性

这些是方向性公开信号,不是标准化价目表。ClusterBid 当前对 TensorWave 规模与 融资的说法与更有力的 2026 年 6 月来源冲突;AMD 明确否认对案例节省额进行独立验证 。

[CP005, CP012, CP013]

3.4 切换摩擦、供应商权力与护城河持久性

TensorWave 的护城河取决于 AMD 和 ROCm 缩小生态差距的速度是否超过 Nvidia 扩大 CUDA 装机基础的速度。TBPN Digest 将 CUDA 周边库、工具链和机构熟悉度视为 AMD 面临的长期阻力;TechTimes 则指出,生产迁移仍取决于可靠性、性能、兼容性和工程成本。TensorWave 可以靠托管支持和 Modular 的可迁移性降低摩擦,但这些服务也会把持续优化的人力变成产品的一部分。AMD 同时是供应商、生态支持者和投资人;这种绑定改善了获取和融资,却集中暴露出议价能力和路线图风险。可持续的上行来自专注的优化层和替代供给渠道;悲观情景则是通用云增加类似 AMD 容量后把服务商品化,或 CUDA 兼容性继续对客户工程团队明显更友好而导致替代。[CP013, CP014, CP015, CP016]

3.5 展品

Chapter 04

04财务

4.1 收入模式与变现

TensorWave 提供的是围绕 AMD GPU 容量构建的基础设施服务,而不是传统软件订阅。公开产品包括裸金属算力、托管推理、高速存储、监控和工程支持,因此收入可能由预留容量、按量使用、托管端点、存储和支持组成。审阅的网站把买方引导至销售沟通,没有公开牌价、合同期限、折扣表或实际价格。AMD 宣称每百万 tokens 的推理成本最高可降低 70%,但这是供应商合作伙伴的营销说法,不能证明 TensorWave 的实际价格、收入或毛利率。没有合同或发票,客户活动到确认收入的桥接只能定性,无法量化。[CI001, CI002, CI003]

收入来源与质量
收入来源机制公开单位或数值收入质量尽调要求
裸金属 GPU 云专用 AMD 加速器容量无公开费率或合同期限可能按合同或用量计费;未经验证提供实际 GPU 小时定价、预订条款、利用率与续约数据
托管推理自动扩缩容模型端点与工程支持合作伙伴声称每百万 token 成本最高可降低 70%供应商 / 合作伙伴说法;不代表 TensorWave 的收入或利润率提供客户发票、token 量、服务成本与毛利率
存储与监控高速网络存储与工作负载监控无单独公开价格可能打包销售,也可能单独计量;未知提供附加率、计价单位与增量利润率
解决方案工程与支持专属上线辅导、优化与 24/7 技术支持无公开服务费可能支持客户留存,但也可能增加人工成本提供每客户支持人员配置、收费结构与贡献利润率

公开产品描述说明了可能的收入机制,但审阅来源未披露实际的 定价、收入组合、确认政策或合同经济性。

[CI001, CI002, CI003]
定价与变现证据
产品方案公开定价证据标价与实际价格含义
GPU 云访问以联系销售为主;审阅的主页没有价目表标价与实际价格均不可得无法基准比较每 GPU 小时价格或承诺用量折扣
托管推理AMD 宣传每百万 token 成本最高可降低 70%合作伙伴说法,不是 TensorWave 客户价格不要把该说法转化为收入或毛利率假设
大型部署AMD 接入表将预期生产需求按少于 100 至 1,000+ GPU 分段仅用于评估分段;无合同价格企业交易规模与销售周期仍属非公开

本表将供应商营销与线索资格字段,与 TensorWave 的标价、实际 价格及已确认收入分开;这些信息均未披露。

[CI002, CI003]
FI001: 存在公开数据断点的收入模型桥接

商业流程基本可见,但公开证据在实际定价、确认收入和毛利之前中断。

节点 / 连接节点 / 起点终点说明
节点 1 客户工作负载 [customer-workload]Training, fine-tuning, or inference 公开的产品范围;工作负载规模和时长仍因客户而异。
节点 2 方案设计 [solution-design]Sales and engineering scoping 以接洽为主的销售路径,但销售周期、CAC 和转化率均未披露。
节点 3 产能合同 [capacity-contract]Bare metal or managed endpoint 合同形式、期限、承诺量、折扣和最低消费均不可得。
节点 4 已部署产能 [deployed-capacity]AMD GPU infrastructure 公开的产能说法无法证明付费利用率或客户集中度。
节点 5 确认收入 [recognized-revenue]Unavailable 未找到经审计的收入、收入结构、确认政策或经常性收入披露。
节点 6 毛利 [gross-profit]Unavailable 需要用实际实现价格扣除硬件、电力、设施、网络、融资和支持成本。

这是定性流程桥接,不是累计瀑布图;不可得数值未作估算。

[CI001, CI002, CI004, CI014, CI015]

4.2 业务进展与销售效率

公开运营信号很强,但不能直接换算成经常性收入。TensorWave 2026 年 6 月公告称线上有 8,192 个 MI325X GPU,已获得超过 2 GW 的长期数据中心容量,Fireworks AI 和 Luma AI 正在使用该平台。独立报道则称其在三个数据中心运营约 10,000 个 GPU、14 MW 容量,租赁覆盖 500 MW,并计划达到 2 GW。这些数字说明部署和管线规模,不等于付费利用率、已签约积压、留存或销售效率。公开资料没有给出 CAC、回本期、销售周期、订单转化、logo 流失、NRR、渠道抽成或客户集中度。[CI004, CI005, CI006, CI009]

单位经济性与资本充足度
指标公开数值或状态置信度重要性尽调要求
年度收入跑率2024 年 $5M;某画像将其归因于 NextPlatform低;二手归因找到的唯一公开跑率数据点;并非 ARR与经审计收入及月度退出跑率核对
另一收入估算$7,015,510;某画像将其归因于 Prospeo,期间未说明低;期间不明的估算无法与 2024 年跑率数据清晰比较提供经审计的期间收入,并解释估算方法
在线加速器基数8,192 个 MI325X GPU 在线;公司于 2026 年 6 月声称中;公司发布容量只有签约并被利用时才能产生收入提供付费利用率、正常运行时间、折旧或租赁成本,以及每 GPU 收入
运营与租赁电力独立报道:运营 14 MW,租赁覆盖 500 MW中;引用另一出版物的二手报道显示运营容量与已承诺扩张容量之间存在巨大差距按站点提供已通电、已签约、已付费及已获期权的容量
最近一轮股权融资募集资金指定用于全球基础设施与 MI355X 集群;剩余现金未披露中;公司发布的资金用途声明资金支持扩张,但不揭示剩余现金或债务义务提供现金余额、月度烧钱、现金续航期、债务、契约与已承诺资本开支

数值保留来源中的指标名称与期间。募集资金不等于可用现金,容量不等于 利用率;两项收入数据也不视为经审计收入或 ARR。

[CI004, CI007, CI009, CI010, CI011, CI014, CI015]

4.3 资本密集度与融资依赖

TensorWave 表示最新股权融资所得将用于扩大全球基础设施并部署 MI355X 集群。这笔融资支持增长,却不能证明当前可用现金:公开资料没有披露已花现金、债务余额、设备融资义务、租赁承诺、契约、烧钱速度或 现金续航。累计融资本身也未定案,两份独立资料分别给出 $450 million-plus 和 $493 million。下行情景具有结构性:独立报道指出,专业 AI 云面临高资本需求、设备承诺、融资成本和利用率风险,而 TensorWave 的 AMD-only 模式又集中供应商敞口。因此,判断远期流动性必须取得私有证据。[CI007, CI008, CI012, CI013, CI015]

4.4 财务结论与尽调重点

仅凭公开记录,尚不能评估收入质量和利润率持久性。一份独立资料将 2024 年 $5 million 年化收入 run rate 归因于 NextPlatform,同时列出 Prospeo 的另一项 $7,015,510 收入估计但未说明期间;两者都不是经审计的公司披露,也不应当作 ARR。现有证据没有经审计的 2025-2026 收入、毛利率、经营亏损、烧钱、现金余额或 现金续航。核心尽调要验证:已签约且已付费的利用率,能否覆盖加速器折旧或租赁费用、电力、设施、网络、支持和融资成本,同时保持续约质量。在管理层提供对账后的财务数据和客户级 cohort 证据前,财务判断只能是高增长潜力伴随显著资本和利润率不确定性。[CI010, CI011, CI014, CI016]

4.5 展品

Chapter 05

05产品与技术

5.1 产品界面与客户工作流

TensorWave 销售的是进入以 AMD 为中心的 AI 云的能力,而不是独立软件工具。其公开产品覆盖训练、微调、生产推理、裸金属基础设施、托管端点、存储、监控和工程支持。买方流程从工作负载评估和迁移开始,经过集群 sizing 与部署,并依赖 TensorWave 工程师做优化和扩容。公开资料提到 MI300X、MI325X、MI355X 和 MI455X,但没有稳定的 SKU 目录、合同级服务定义或普遍可用的价格。Moreh、Fireworks AI 和 Luma AI 的客户案例说明确有工作负载接入,但现有资料没有一致的生产与试点边界或可比结果指标。[CE001, CE002, CE007, CE010, CE011, CE012]

产品模块与资产矩阵
模块或资产买方或用户公开成熟度信号差异化尽调缺口
分布式训练云模型训练与平台团队公开宣传采用 RoCEv2 网络与托管 SlurmAMD 专属集群调优与面向内存的加速器无独立测量的作业完成率、扩展效率或可用性
推理基础设施AI 产品与推理团队公司引用生产推理用户与托管端点AMD 内存容量与供应商专属优化无可比的延迟、吞吐量、错误率或单位成本数据集
裸金属 GPU 访问需要专用容量的基础设施团队由独立公司名录列出,并由 AMD 描述物理 GPU 分区与专用部署选项隔离设计、租户控制与合同条款未公开
Helios 机架级平台前沿模型构建者TensorWave 于 2026 年 7 月发布的公告称该平台已获采用整合 72 个 GPU、EPYC、Pensando、ROCm 与 Instinct 的 AMD 机架部署计划、客户可用性与实际性能未披露
工程支持与监控迁移或优化 AMD 工作负载的团队公司与 AMD 页面描述亲自参与式及 24/7 支持深度 AMD 技术栈专长无响应时间 SLA、支持人员配置数据或问题解决历史

截至 2026-10-21,状态与差异化来自发行方、供应商或名录描述;定量证明 为空的部分以尽调缺口呈现,不做估算。

[CE001, CE002, CE006, CE007, CE010, CE011, CE012]
工作流与用例评估
用户任务当前工作流TensorWave 方案公开宣称的收益局限
训练大模型配置多节点算力、调度作业并写入检查点采用 RoCEv2、Slurm 与面向检查点路径的 AMD Instinct 集群公司描述可扩展的分布式训练无独立的强扩展曲线或作业失败率
微调模型分配加速器并适配框架工作负载托管 AMD 环境与工程支持降低直接运营 AMD 技术栈的负担未公开按模型与框架版本划分的兼容性
提供生产推理部署端点、监控负载并扩展容量托管端点与内存优化型加速器发行方与供应商声称具备自动扩缩容、更高吞吐量与更低成本无标准化基准测试或合同约定的延迟证据
评估迁移至 AMD移植依赖 CUDA 的工作负载并验证输出借助 TensorWave 与 AMD 资源进行亲自参与式评估和优化单一供应商 Nvidia 技术栈的潜在替代方案迁移工程成本和不支持的算子仍取决于具体工作负载
运行敏感工作负载评估控制措施、隔离、审计范围和数据处理公司自称符合 SOC 2、ISO 27001 和 HIPAA 的环境企业控制体系仅有公开自述报告、范围、例外、隐私架构和数据驻留信息未公开

除非找到独立来源,否则这些优势均按公司或供应商自述处理;不推断任何 转化、性能或成本指标。

[CE001, CE003, CE004, CE005, CE007, CE008, CE012]

5.2 架构与运营模式

公开的技术栈足以辨认主要层次。AMD Instinct 加速器之上是包含 Helios、EPYC CPU、Pensando 网络和 ROCm 的机架与集群层;TensorWave 训练页面还提到 RoCEv2 spine-and-leaf 网络、面向 checkpoint 的存储路径、拓扑感知放置、gang scheduling 和 Slurm 分区。TensorWave 发布的公告描述了 72-GPU Helios 基础单元,AMD 合作伙伴页面则描述了物理 GPU 分区和托管端点。这些都是发行方和供应商的说法,不是独立基准。因此,该架构作为产品蓝图可信,但吞吐、隔离、自动扩缩和成本说法仍需按工作负载验证,才能用于投资判断。[CE003, CE006, CE012]

技术与运营架构
层级或组件角色已披露依赖主要风险
AMD Instinct GPU 集群训练与推理算力MI300X 至 MI455X 系列单一供应商硬件集中度与更新执行
Helios 机架级规模机架级一体化算力配备 EPYC CPU、Pensando 网络和 ROCm 的 72 块 GPU采用公告不能证明已部署可用性或经济性
RoCEv2 网络低延迟多节点同步Spine-leaf 以太网与拓扑感知放置没有独立的拥塞、尾延迟或故障域数据
检查点路径训练状态写入与恢复高吞吐存储路径没有实测恢复点或恢复时间目标
Slurm 控制平面Gang 调度、分区与集群扩缩容托管式调度器配置队列公平性和邻居噪声行为未披露
监控与控制基础设施健康度与敏感工作负载控制态势主动/被动检查与自称取得的认证没有公开 SLA、审计范围、事件历史或控制例外

架构仅根据留存的发行方、供应商和登记信息描述重建;性能和 控制有效性仍未验证。

[CE002, CE003, CE004, CE005, CE006, CE012]
FE001: TensorWave 产品架构图

基于来源边界,展示已披露的工作负载、服务、控制平面、集群和硬件层。

客户工作负载
训练、微调和推理工作负载通过辅助部署流程接入。
条目
  • 模型训练
  • 微调
  • 生产推理
服务层
裸金属访问、托管端点、监控和工程支持。
条目
  • 裸金属
  • 托管端点
  • 专家支持
控制平面
托管 Slurm 提供放置、协同调度、分区和调整规模能力。
条目
  • Slurm 调度
  • 感知拓扑的放置
  • 分区
集群网络与存储
RoCEv2 spine-leaf 网络及面向检查点的数据路径连接各节点。
条目
  • RoCEv2 网络
  • 检查点路径
  • 健康检查
机架平台
已宣布的 Helios 设计整合 Instinct GPU、EPYC CPU、Pensando 网络和 ROCm。
条目
  • Helios 机架
  • EPYC CPU
  • Pensando 网络
  • ROCm
加速器底座
公开产品组合覆盖 MI300X、MI325X、MI355X 和 MI455X。
条目
  • MI300X
  • MI325X
  • MI355X
  • MI455X

基于公开描述的概念层级图;这不是完整部署图,也不代表已测得的性能、可用性或安全效果。

[CE002, CE003, CE006, CE012]

5.3 成熟度、可靠性、支持与控制

TensorWave 发布的是运营实践,而不是经审计的服务表现。官网描述持续基础设施监控以及主动、被动健康检查;TensorWave 和 AMD 都称其具备 SOC 2 Type II 认证和 HIPAA 合规环境,TensorWave 另称已获 ISO 27001 认证。 AMD 还营销 24/7 高级 AI/ML 工程师支持,TensorWave 则把 onboarding 和扩容描述为深度介入式服务。这些信号符合面向企业的运营模式,但不能证明可用性、事故响应质量、恢复目标、租户隔离效果或认证范围。现有证据没有公开 SLA、状态历史数据、渗透测试摘要、认证报告或事故复盘。[CE004, CE005, CE007, CE010, CE011]

5.4 差异化与技术风险

TensorWave 的差异化在于专注:AMD-only 机群、ROCm 专业能力、面向显存的加速器选择,以及可加快获取新一代 Instinct 的供应商关系。这种专注能降低 AMD-ready 工作负载的迁移摩擦,并提供 NVIDIA 中心化基础设施之外的选择,但也把执行风险集中在一个软硬件生态。TechTimes 强调 CUDA 工具链的根深蒂固和迁移成本,并指出 AMD 同时是供应商和投资人;另一份可信度较低的行业评测则认为 ROCm 在框架和算子覆盖上仍落后 CUDA。这些批评方向上合理,但不能替代可复现基准。尽调应在约定服务水平下测试代表性模型、故障切换、框架兼容性、安全范围和支持响应。[CE008, CE009, CE013, CE014]

5.5 展品

Chapter 06

06客户

6.1 客户分群与购买路径

公开证据指向专门买方,而不是广泛的自助式云客户。TensorWave 面向训练和推理营销 AMD Instinct 与 ROCm 容量,第三方资料则将用户描述为 AI 开发者、企业、研究机构和成本敏感的 PyTorch 团队。可能的经济买方是基础设施、工程或采购负责人;实际使用者是 ML 或平台工程师;付款方是为专用算力提供预算的组织。这只是方向性分群,不是已披露的收入结构:现有资料没有按行业、地域、公司规模或收入档位拆分客户。产品材料强调方案工程,公司招聘信息同时描述多租户云和单租户裸金属部署,说明部署架构和支持强度属于商业资格审查的一部分,而非偶然的实施细节。[CU001, CU002, CU007, CU008, CU013, CU014]

客户分层与购买假设
细分市场买方/付款方主要用户使用场景匹配证据尽调缺口
前沿模型开发商AI 基础设施或工程负责人;公司预算ML 与平台工程师大模型训练与推理TensorWave 面向这类工作负载销售 AMD Instinct 和 ROCm 算力未披露分部收入、账户数量或地域构成
企业 AI 团队CIO、平台负责人或采购部门;企业预算内部 AI 平台与应用团队专用生产算力与托管部署公开材料描述了现场支持,以及多租户和单租户架构没有具名企业客户群、合同规模或采购周期数据
研究机构与学术实验室实验室或机构;科研经费或研究预算研究人员与 ML 工程师对成本敏感的训练与实验ClusterBid 将学术实验室列为目标细分市场低可信度二手证据;留存来源中没有具名部署
AMD 优先的软件与模型公司工程或产品负责人;运营预算推理与优化团队在 AMD GPU 上进行早期接入、验证与优化Moreh 被列为在生产云中使用 MI355X只有一个具名且细节充分的结果;没有客户群规模的证明

分层结论综合自产品定位、公司发布的职位转发和二手描述; 并非已披露的客户或收入拆分。

[CU001, CU007, CU013, CU014]

6.2 采用情况与客户案例质量

在现有资料中,Moreh 是最有力的具名终端客户案例。TensorWave 官网称 Moreh 在真实生产云中使用 MI355X 容量进行 frontier-LLM 推理,并报告模型权重下载时间约减少 75%。配套客户引语又称项目尚早,无法提供更广泛的客户指标,这显著限制了可推断内容:它证明一个生产工作负载和一项技术结果,却不能证明全机群利用率、经常性收入、续约或普遍成本节省。RFP.wiki 提到 Moreh 和 Higgsfield AI 的公开证言,但没有经核验的评论样本,也没有在抓取文本中提供足够详细的 Higgsfield 部署。AMD 案例验证了供应商关系和评估路径,但提及未具名大型 AI 客户,不能替代具名的 TensorWave 客户证据。[CU003, CU004, CU005, CU006, CU010]

公开采用与持续性信号
指标或信号数值期间证据状态含义缺失分母或局限
具名生产客户Moreh访问当前主页:2026-10-21公司托管的客户引述确认至少有一个生产云工作负载没有起始日期、支出、续约或算力分配信息
模型权重下载改善~75% 减少访问当前主页:2026-10-21公司自称的客户结果表明一次部署带来了可衡量的工作流收益没有基线、测试方法、样本量或独立验证
客户数量截至 2026-10-21 的留存证据未披露采用广度无法量化没有活跃账户定义或客户分母
留存与集中度截至 2026-10-21 的留存证据未披露持续性与依赖风险无法评分没有 NRR、GRR、流失、续约、合同条款或头部账户占比

null 表示该指标未在留存的指定证据池中披露;这不代表客户数为零、 也不代表不存在流失或集中度。

[CU003, CU016]
具名客户证明表
客户细分市场部署/使用场景生产环境 vs. 试点结果局限
MorehAI 软件/模型基础设施在 MI355X 上进行前沿 LLM 推理被描述为真实生产云,但优化仍处于早期TensorWave 声称模型权重下载时间减少约 75%公司托管证据;更广泛的客户指标明确尚未发布
Higgsfield AI生成式 AIRFP.wiki 提及的推理工作负载留存的抓取文本未予证实RFP.wiki 称存在客户证言,但留存文本未提供可审计的部署指标没有直接客户来源、期间、合同细节或独立结果验证
没有第三个经证实的名称未知未知未知AMD 仅指未具名的大型 AI 客户;Logo 或未具名引述不能证明已部署

指定证据池中找到的客户名称仅为部分列举;只有 Moreh 有详细部署和结果 证据,且不声称覆盖完整。

[CU004, CU010]

6.3 持久性、扩张与集中度

表面上的客户旅程是咨询式的:先筛选 AMD 和 ROCm 适配度,再申请评估,选择专用或托管架构,由方案工程师完成迁移和调优,进入生产,最后在性能和可靠性达标时扩容。这种路径可能形成切换成本和扩张空间,但公开记录没有证明它确实如此。TechTimes 将可靠性、兼容性、迁移成本和性能列为迁移生产负载的决策变量;RFP.wiki 转述了中断和编排不成熟的历史报道;ClusterBid 则把 CUDA 移植工作量列为约束。更重要的是,现有资料均未披露客户数、毛留存或净留存、logo 流失、续约率、合同期限、头部客户占比或渠道收入。这些缺口阻碍了对持久性和集中度的可靠判断,必须取得直接的 cohort、合同和账单证据。[CU009, CU011, CU012, CU015, CU016]

FU001: 基于证据的客户采用路径

一条由销售协助的路径:从 AMD 工作负载筛选,到生产运行,再到可能扩容;初始部署后,证据缺口逐步扩大。

这是基于证据的流程模型,不是实测漏斗;转化率、阶段时长和扩容率均未公开披露。

[CU002, CU006, CU009, CU011, CU015]

6.4 展品

Chapter 07

07风险

7.1 风险排序与投资含义

核心投资风险是集中度,而不是已有记录的法律或运营失败。TensorWave 明确围绕 AMD 硬件和 ROCm 构建,独立报道也称 AMD 同时是供应商和投资人。这种绑定可以提升工程聚焦,却把产品竞争力、硬件可得性、融资观感和客户采用都连到同一生态。二阶风险是资本密集度:$350 million Series B 用于扩张,但已募集资金不等于可用现金,也没有披露设备融资、租赁承诺、利用率或契约余量。因此,投资逻辑应以已验证的工作负载可迁移性、已签约需求和融资条款为条件,而不是只看公告容量。[CR001, CR005, CR006, CR007, CR008, CR009]

FR001: 剩余风险优先级矩阵

矩阵优先呈现已观察到的结构性依赖和有限披露缺口;评级是尽调判断,不是实测事件概率。

序数评级综合所引公开证据和明确缺口;它们不是数量化概率、损失估计,也不能证明不利事件已经发生。

[CR002, CR008, CR009, CR010, CR011, CR014, CR015]

7.2 监管、法律、隐私与安全敞口

现有页面称 TensorWave 已获 SOC 2 Type II 和 ISO 27001 认证并符合 HIPAA,但资料集中没有审计报告、范围声明、数据处理条款、正常运行时间 SLA、事故历史、诉讼记录或监管往来,无法核实覆盖范围。一份无日期的第三方页面与当前认证说法冲突,更适合视为时效性问题,而不是认证不存在的证据。没有资料证明发生过数据泄露、执法行动或诉讼;同样,在这个有边界的资料池中没有记录,也不能证明这些事件不存在。投资前应取得底层报告和陈述保证,而不能依赖网站徽章。[CR002, CR012, CR015, CR016]

监管/法律风险登记表
风险或规则面司法辖区/范围证据状态可能性严重性缓释措施成熟度剩余暴露尽调路径/止损标准
安全认证与审计范围企业云运营;所述控制措施面向美国当前官方和服务商页面称符合 SOC 2 Type II、ISO 27001 和 HIPAA;底层报告 及其范围未留存中等披露风险高公开资质已有陈述;私人审计证据未审阅中高获取最新报告、例外、范围和整改函;若重大例外被隐瞒 或未解决,则停止
隐私、DPA 与数据所在地义务客户合同与各运营地区留存证据中没有 DPA、分处理者清单、保留期限表或数据驻留条款中高未知在审阅文件前为高审阅 DPA、隐私附表、分处理者、删除控制和驻留承诺;若目标 工作负载无法满足所需控制,则停止
事件、宕机与 SLA 暴露所有生产服务留存证据描述了监控和健康检查,但其中没有完整的事件历史或 正常运行时间 SLA中高已描述监控;合同保护措施未验证中高获取 24 个月事件日志、SLO 达成情况、赔偿额度、RTO/RPO 和根因报告;若存在未披露的 重大事件或长期未达标,则停止
治理、诉讼与监管披露公司、融资、雇佣、IP 与贸易管制该证据池未证实存在不利程序;股东名册和法律记录不完整中等披露风险高未知在律师审阅前为高开展律师主导的诉讼、IP、制裁、出口管制和公司记录审阅;若存在未披露的 重大法律主张或限制,则停止

严重性和可能性是基于所引依赖与披露证据的分析评级,不是 来源原文引用的概率;缺失记录属于尽调缺口,不证明不合规或不存在 事件。

[CR002, CR012, CR015, CR016, CR001]

7.3 运营、技术与质量执行

TensorWave 把专注定位为可靠性优势,并描述 24/7 基础设施监控,但面向客户的控制措施不能替代实测可用性、恢复表现或完整事故日志。技术风险集中在 ROCm 兼容性和 AMD 路线图速度。一份负面第三方分析称 ROCm 在部分框架和 kernel 上仍落后 CUDA;其具体数据在此没有独立佐证,应当测试,而不是当作基准接受。AMD 自己的案例也说明,引用的性能和节省结果来自 TensorWave 或 Modular,未经 AMD 独立验证。否决标准是:目标工作负载出现可重复的性能落后,或迁移工作量抹去声称的成本优势。[CR003, CR010, CR011, CR012, CR015]

运营/质量/安全风险登记表
失效模式可能性严重性已观察到的缓释措施缓释措施成熟度剩余暴露监控指标/论点失效触发条件
AMD / ROCm 路线图或兼容性缺口相对客户需求扩大高结构性暴露高单一技术栈专业化、ROCm 工程能力与 Modular 可移植性定位发展中;公开基准测试主张未经独立验证高目标工作负载迁移工时、不支持的内核、单位美元性能;若切换成本 或性能不足抹去节省,则论点失效
生产宕机或恢复失败中高公开描述了 24/7 监控以及主动和被动健康检查控制措施已有陈述;结果历史未验证中高SLO 达成情况、事件频率、MTTR、赔偿额度、RTO/RPO;若有未披露的重大事件或 反复未达 SLA,则论点失效
算力建设超过合同利用率中高大额融资轮次与据报道的长期算力获取扩张已有资金;租约、电力和客户承诺未披露高已安装算力与可计费算力、合同利用率、电力押金、租约 照付不议;若 利用率无法覆盖固定承诺,则论点失效
性能或成本主张无法通过独立复现中中高合作伙伴案例研究与工作负载可移植工具仅有合作伙伴验证的营销证据中高对目标模型进行独立基准测试,计入完整迁移人力和失败率;若可复现的 经济性不具竞争力,则论点失效

风险评级综合当前公开控制、明确的来源限定和未解决的私人证据缺口;不 声称已经发生宕机或基准测试失败。

[CR001, CR002, CR003, CR010, CR011, CR013, CR014, CR015]

7.4 财务模型、合作伙伴与客户依赖

独立报道将 2026 年 6 月 Series B 置于 $350 million,投后估值为 $1.55 billion;两家数据库和新闻报道则把累计融资放在约 $493 million。这些数字说明融资渠道,不代表 现金续航、流动性、收入质量或承受低利用率的能力。同一组证据描述了大规模扩容,以及同时包含硬件供给、软件依赖和投资关系的 AMD 绑定。公开所有权数据不完整,现有资料也没有头部客户集中度、合同期限、终止权、积压质量、电力义务或债务契约。投资批准应要求合同级需求桥接、债务和租赁明细,以及把已承诺收入与管线分开的交易对手图谱。[CR004, CR005, CR006, CR007, CR008, CR009, CR013, CR014, CR016]

合作伙伴/依赖风险登记表
依赖项交易对手方角色与集中度失效情景严重性已观察到的缓释措施剩余暴露尽调路径
加速器供应、路线图与核心软件AMD独家硬件平台;ROCm 是核心软件生态;AMD Ventures 也是投资方路线图延期、供应受限、不利条款或生态缺口会削弱服务经济性和客户 需求高深度单平台优化与战略协同高审阅供应协议、分配优先级、定价、路线图承诺、终止权及 关联方治理
推理可移植层Modular用于降低跨硬件迁移摩擦的具名技术合作伙伴合作伙伴性能、支持或商业条款无法兑现承诺的可移植性中高AMD 案例研究中描述的即插即用容器和可移植性主张中高检查协议、续约与终止权、支持 SLA、基准测试来源和已测试的备用路径
数据中心、电力与融资交易对手方留存证据未披露据报道有超过 2 GW 的长期算力,但交易对手方和义务未披露部署或利用率落后时,租约、电力、债务或建设义务仍保持固定高大额股权融资与分阶段扩张高获取逐站点的租约、电力、项目融资、建设、担保和利用率计划
客户需求与收入集中度客户与合同未披露报告中存在具名用户,但头部账户占比、续约条款、积压订单质量和终止权 在此不可得大客户延迟部署、退出,或主要把 AMD 算力作为谈判筹码高没有建立公开的集中度缓释措施高审阅客户层面收入、利用率、积压订单、押金、终止权、NRR/GRR 和管道 转化

集中度、合同和交易对手方字段并不完整,因为私人协议不可得; 融资金额不视为可用现金,也不视为义务已有资金支持的证明。

[CR001, CR005, CR006, CR007, CR008, CR009, CR010, CR013, CR014, CR016]

7.5 展品

Chapter 08

08估值

8.1 投资观点与建议

当前建议是 research-more:对判断方向的信心为中等,但对价格的确定性很低。正面逻辑有具体支撑:TensorWave 以披露的 $1.55 billion 投后估值融资 $350 million,得到 AMD Ventures 和 Magnetar 的战略支持,并营销面向训练和推理的 AMD-exclusive 云。这一定位能为寻找 NVIDIA 中心化基础设施替代方案的客户提供采购选择。反方同样明确:独立报道强调 hyperscaler 垂直整合、CUDA 软件护城河、高基础设施资本需求以及对 AMD 路线图的依赖。 公开资料没有经审计收入、毛利率、利用率、客户集中度、留存、烧钱或有说服力的收入倍数。因此可以判断公司质量,但不能据 6 月估值做出买入决策。[CV001, CV002, CV003, CV008, CV012, CV013, CV016]

建议摘要
维度当前判断证据基础决策含义
建议继续研究轮次条款公开;运营和安全条款未公开暂不按表面估值做投资判断
信心中多个独立来源相互印证该融资;经济性数据仍属私人信息信心仅适用于谨慎判断,不适用于目标价格
风险评级高资本密集度、供应商集中度、CUDA 摩擦与融资依赖要求下行保护,并按里程碑入场
估值立场偏高/缺乏支撑$1.55B 投后估值参考,但未披露 EV/收入或经审计运营指标公开证据无法支持精确区间
价格驱动因素取决于证据经审计经济性、利用率、客户质量、债务与优先权瀑布仅在上述尽调材料包齐备后重新评估

判断截至 2026-10-21;建议条款是分析结论,不是引用的证券建议,且没有公开的现金流或收入倍数可用。

[CV001, CV002, CV010, CV013, CV014, CV016]
论点与反论点
立场论点当前证据改变判断所需证据
论点战略生态位AMD 共同领投 Series B,并将 TensorWave 推介为 AMD 独家云证明战略协同改善供应、成本或客户赢单的独立证据
论点差异化基础设施主张TensorWave 销售针对内存优化的 AMD 训练与推理基础设施工作负载级第三方基准测试和生产参考客户
反论点软件迁移摩擦独立分析认为 ROCm 在工具链和装机基础优势上仍落后于 CUDA可重复的迁移证据,以及经测量的总成本和可靠性
反论点资本与融资负担外部分析估计每 MW 成本为 $15M-$20M,并预计还需继续募资与部署时点匹配的合同利用率和融资,且不会造成惩罚性稀释
反论点供应商与投资者集中度AMD 既是核心供应商,也是领投方跨多代产品的供应保障,以及应对路线图延期的合同救济

证据综合公司与合作伙伴主张及独立分析;技术和成本主张仍 须待审阅基准测试方法和客户生产数据后才能验证。

[CV003, CV008, CV009, CV010, CV011, CV013]
FV001: 建议逻辑

证据从战略规模信号出发,经由证明缺口和风险,最终指向“继续研究”决策。

节点 / 连接节点 / 起点终点说明
节点 1 $350M Series B / $1.55B 投后估值 [scale]这是当前融资参考,不是经营价值的证明。
节点 2 AMD 独家基础设施 [differentiation]形成战略差异化,但供应商暴露集中。
节点 3 公开证明缺口 [proof]没有经审计的收入、利润率、利用率、留存率或单位经济性。
节点 4 资本与软件风险 [risk]融资强度和 CUDA 迁移摩擦仍然显著。
节点 5 价格与条款纪律 [discipline]需要资本结构表、债务、优先权和部署证据。
节点 6 继续研究 [decision]没有证据支撑的估值区间可供采用。
连接 1 $350M Series B / $1.55B 投后估值 [scale]AMD 独家基础设施 [differentiation]
连接 2 AMD 独家基础设施 [differentiation]公开证明缺口 [proof]
连接 3 公开证明缺口 [proof]资本与软件风险 [risk]
连接 4 资本与软件风险 [risk]价格与条款纪律 [discipline]
连接 5 价格与条款纪律 [discipline]继续研究 [decision]

该序列只表达决策逻辑;不代表概率、规模,也不代表阶段之间存在数量化转化。

[CV002, CV008, CV009, CV013, CV016]

8.2 融资背景与入场纪律

Series B 的 headline 意味着新资本约占投后价值的 22.6%,但这不是完全摊薄后的持股估计,也没有说明清算优先权、跟投、顺位、期权池扩大、债务、认股权证或投资人控制。公开融资总额同样冲突:一份独立资料称三轮合计 $450 million-plus,另一份分析称约 $493 million。两者都不等于今天可用现金。AMD 同时作为供应商和领投方,带来战略验证,也带来循环融资和集中度风险,因为部分投资资本可能通过购买加速器回流 AMD。因此,在把 $1.55 billion 参考值视为可投资估值前,应要求当前股权结构表、证券清偿 waterfall、债务明细、现金余额、已承诺资本支出和剩余部署义务。[CV001, CV002, CV004, CV005, CV006, CV007, CV010]

8.3 乐观、基准与悲观框架

公开证据足以搭建有条件的情景框架,但不足以给出数值估值区间。乐观情景要求生产利用率、持久客户需求以及独立证明的 AMD 性价比优势,才能把容量转化为有吸引力的利润率。基准情景假设 TensorWave 仍是有意义的 AMD 专注型细分供应商,但扩张需要反复融资,使回报对稀释和证券条款敏感。悲观情景叠加较低利用率、持续的 CUDA 迁移摩擦、超大规模云厂商替代和对 AMD 路线图的依赖;在这种组合下,固定基础设施承诺和倍数压缩可能损害股权价值。外部分析估计 GPU 集群成本为每兆瓦 $15 million 至 $20 million,并称完整部署 2 GW 需要数百亿美元,但这只是分析师估计,不是公司披露的已签约资本支出,只能作为方向性压力测试。[CV008, CV009, CV011, CV012, CV013, CV014, CV016]

乐观、基准与悲观情景
情景假设估值/回报逻辑概率信号下行触发条件
乐观生产利用率高;客户续约稳健;AMD 的性价比有独立验证;融资压力可控这些因素或可支撑 6 月估值或上行空间,但没有经审计的收入和利润率,就无法合理给出数值区间经审计的增长,加上稳定的毛利率和客户群留存基准测试无法转化为客户经济性
基准情景可行的 AMD 细分市场;迁移进度不均;持续依赖外部融资;利用率中等回报取决于入场折价、稀释和优先权条款,而不是企业表面增长势头在利用率已披露的情况下,产能可转化为已签约收入新融资会显著使普通股或新优先股处于次级地位
悲观情景利用率低;CUDA 摩擦持续;超大规模云厂商将需求内化;AMD 路线图延期估值倍数压缩叠加固定承诺可能损害股权价值;仅凭公开数据无法量化下行空间部署延期、续约疲弱或实际实现价格下降扣除电力和折旧后贡献利润率为负

情景具有条件性,不代表概率判断;由于缺少经审计的收入、利润率、利用率、债务、优先权和现金流数据,刻意不提供估值区间。

[CV010, CV011, CV012, CV013, CV014, CV015, CV016]

8.4 可比公司、退出准备度与逻辑破坏测试

现有可比公司集合只能用于搭建框架。独立资料将 CoreWeave 和数家基础设施相关上市公司列为参照,但也指出 TensorWave 各轮融资没有披露 EV-to-revenue。审阅过的 seed deck 分析同样找不到财务预测或单位经济性。在没有经审计收入、增长、毛利率、利用率、合同期限和客户集中度的情况下套用上市公司销售倍数,只会制造虚假精确。退出准备度同样未解,因为没有资料提供 IPO 时间表、经审计报表、董事会构成或治理权利。投资逻辑的破坏触发点是可量化的:利用率长期偏低、扣除电力和硬件折旧后无法产生正贡献利润、AMD 路线图明显延误、失去锚定客户,或融资轮给予惩罚性顺位。正向推进需要独立生产基准、经审计经济性、客户 cohort 证据,以及完整建模的资本结构和优先权 waterfall。[CV010, CV013, CV014, CV015, CV016]

8.5 展品

免责声明

本报告基于有边界的公开来源研究,不构成投资、法律、税务或会计建议。私营公司披露可能不完整、相互冲突或未经审计。

证据索引

结论
编号陈述可信度来源
CO001 TensorWave is a Las Vegas-based company founded in 2023 and currently described as a Series B business. 中 SO003, SO004, SO005
CO002 Public profiles identify Darrick Horton, Jeff Tatarchuk, and Piotr Tomasik as the co-founders leading TensorWave's executive functions. 中 SO003, SO004, SO005, SO008
CO003 TensorWave is positioned as an AMD-focused AI and HPC cloud offering infrastructure for training, fine-tuning, and inference. 中 SO003, SO004, SO005, SO008
CO004 Tracxn reports that TensorWave raised a $43 million Series A on October 8, 2024. 中 SO003
CO005 Growth Engineer instead labels TensorWave's October 2024 $43 million financing as a SAFE. 低 SO008
CO006 Tracxn reports that TensorWave raised a $100 million Series A on April 8, 2025. 中 SO003, SO004, SO005
CO007 TensorWave announced a $350 million Series B on June 10, 2026 at a $1.55 billion post-money valuation. 高 SO001, SO003, SO004, SO005, SO009
CO008 Tracxn reports $493 million of cumulative TensorWave funding across three rounds. 中 SO003
CO009 Silicon Valley Investment Club reports $450 million-plus of cumulative funding, conflicting with Tracxn's $493 million total. 低 SO004, SO005
CO010 Tracxn reports that TensorWave had 177 employees as of September 26. 中 SO003
CO011 Silicon Valley Investment Club reports 100 employees as of August 2026, conflicting with Tracxn's later 177-employee figure. 低 SO004, SO005
CO012 TensorWave announced support for AMD's Helios rack-scale solution on July 23, 2026. 中 SO006
CO013 TensorWave announced an AMD-based AI inference collaboration with Modular on April 6, 2026. 中 SO007
CO014 Independent reporting identifies circular financing and exclusive AMD supplier dependence as material concentration risks. 中 SO009
CO015 Current board composition, investor control rights, and succession arrangements are not disclosed in the retained evidence. 低
CO016 A low-reputation independent review argues that ROCm software ecosystem maturity is TensorWave's fundamental execution challenge. 低 SO010
CM001 TensorWave's core market is specialized cloud infrastructure for AI and HPC workloads on AMD Instinct GPUs, spanning training, fine-tuning, inference, storage, orchestration, and monitoring. 中 SM003, SM004, SM008
CM002 Retained sources identify AI developers, enterprises, startups, academic labs, and cost-sensitive PyTorch teams as plausible customer segments for TensorWave. 中 SM002, SM003, SM007
CM003 The principal demand drivers are access to scarce specialized compute, memory-intensive workload economics, price pressure, and reduced dependence on NVIDIA. 中 SM003, SM004, SM005
CM004 TensorWave's June 2026 release says 8,192 AMD Instinct MI325X GPUs were online. 中 SM001
CM005 TechTimes reported that TensorWave operated about 10,000 AMD GPUs across three data centers with 14 MW of capacity in June 2026. 中 SM005
CM006 TechTimes reported 500 MW of signed leases and a 2 GW capacity objective, while TensorWave separately claimed more than 2 GW of long-term data-center capacity secured. 中 SM001, SM005
CM007 ClusterBid's Q2 2026 estimate of approximately 3,000 MI300X GPUs across two operating data centers is not directly reconcilable with the larger June fleet snapshots without model, date, and site-scope detail. 低 SM002
CM008 ClusterBid reports TensorWave MI300X pricing of $1.99 per GPU-hour on demand and $1.25 on a 12-month reservation, below the NVIDIA-cloud examples in its comparison. 低 SM002
CM009 ClusterBid identifies large-batch inference, embedding and retrieval, and cost-constrained training as stronger AMD fits than single-stream inference or CUDA-kernel-dependent workloads. 低 SM002
CM010 ROCm compatibility, CUDA-specific tooling, workload migration, reliability proof, and engineering effort remain material barriers to production adoption. 中 SM002, SM005
CM011 AMD's case study presents up to 2x inference throughput and approximately 40–60% cost savings for TensorWave and Modular configurations but states that AMD did not independently verify those claims. 中 SM003
CM012 TensorWave's June 2026 release names Fireworks AI and Luma AI as users of its infrastructure for generative-AI and production-inference workloads. 中 SM001
CM013 AMD's investment can strengthen the AMD cloud ecosystem while increasing TensorWave's supplier concentration and creating a circular-financing concern. 中 SM005
CM014 Tracxn classifies TensorWave among 579 active competitors, including 89 funded companies, but that broad taxonomy is not a direct measure of AMD-specialized GPU-cloud demand. 低 SM008
CM015 Public evidence retained for this chapter does not establish a comparable dollar TAM, a serviceable AMD-cloud revenue pool, TensorWave utilization, or an obtainable market share. 中
CM016 TensorWave's homepage claims SOC II Type 2 and ISO 27001 certifications plus HIPAA compliance, but the retained pool does not independently validate procurement acceptance. 中 SM004
CP001 TensorWave markets an AMD Instinct and ROCm cloud for training, fine-tuning, inference, and HPC workloads. 中 SP005, SP006
CP002 TensorWave's current site lists MI455X, MI355X, MI325X, and MI300X accelerator options. 中 SP005
CP003 AMD's evaluation page states that TensorWave offers up to 288GB HBM3e memory per GPU and chiplet-level physical partitioning. 中 SP006
CP004 AMD's evaluation page represents TensorWave as SOC 2 Type II certified with HIPAA-compliant environments and 24/7 AMD-focused engineering support. 中 SP006
CP005 AMD's case study relays claims of up to 2x inference throughput and approximately 40-60% savings versus Nvidia B200 GPUs while stating that AMD did not independently verify the claims. 中 SP007
CP006 Multiple June 2026 sources report that TensorWave raised a $350 million Series B at a $1.55 billion valuation. 中 SP001, SP002, SP003, SP008
CP007 Data Center Dynamics reports that TensorWave launched an 8,192-GPU AMD Instinct MI325X training cluster. 中 SP002
CP008 Data Center Dynamics reports TensorWave's claim of more than 2GW of secured long-term capacity and a January 2026 additional 20MW agreement with TecFusions. 中 SP002
CP009 TechTimes, citing The Wall Street Journal, reports that TensorWave operated three data centers with about 10,000 AMD GPUs and 14MW of capacity. 中 SP003
CP010 Tracxn ranks TensorWave 30th among 579 active competitors and fifth by total funding in its comparison set. 中 SP008
CP011 Tracxn's ranked competitive set includes CoreWeave, Together AI, FluidStack, Rescale, and Crusoe as funded AI or HPC infrastructure alternatives. 中 SP008
CP012 ClusterBid claims TensorWave pricing of $1.99 per GPU-hour on demand and $1.25 on a 12-month reservation, but no retained official rate card corroborates those figures. 低 SP004
CP013 ClusterBid's Q2 2026 claims of $45 million raised and about 3,000 GPUs conflict with contemporaneous sources reporting a $350 million Series B and materially larger deployed clusters. 中 SP004, SP001, SP002, SP003
CP014 TBPN Digest identifies the CUDA ecosystem of libraries, tooling, and institutional familiarity as the persistent barrier to AMD workload migration. 中 SP001
CP015 TechTimes concludes that production migration depends on reliability, performance, software compatibility, and the engineering cost of moving away from CUDA. 中 SP003
CP016 TechTimes identifies circular-financing concerns and concentrated supplier risk because AMD both invests in TensorWave and supplies the platform on which it exclusively depends. 中 SP003
CI001 TensorWave publicly offers AMD GPU infrastructure for training and inference through bare-metal compute, managed inference, storage, monitoring, and technical support. 中 SI004, SI005, SI008
CI002 The reviewed TensorWave homepage uses a contact-led sales motion and discloses no list price, contract term, discount schedule, or realized customer price. 中 SI005
CI003 AMD claims TensorWave can reduce inference cost per million tokens by up to 70% versus traditional inference stacks, but does not disclose the comparison methodology or TensorWave's realized customer price. 低 SI007
CI004 TensorWave claimed that 8,192 AMD Instinct MI325X GPUs were online in June 2026. 中 SI001
CI005 TensorWave claimed it had secured more than 2 gigawatts of long-term data-center capacity by June 2026. 中 SI001
CI006 TensorWave's June 2026 release named Fireworks AI and Luma AI as users of its infrastructure without disclosing contract values, utilization, or retention. 中 SI001
CI007 TensorWave said its latest equity-financing proceeds would expand global infrastructure and deploy MI355X clusters, without disclosing remaining cash or associated obligations. 中 SI001, SI002, SI003, SI006, SI008
CI008 TensorWave said the Series B proceeds would support global infrastructure expansion and deployment of MI355X GPU clusters. 中 SI001
CI009 Tech Times reported three operating data centers with about 10,000 AMD GPUs and 14 megawatts of capacity, plus leases covering 500 megawatts and an aim to reach 2 gigawatts. 中 SI006
CI010 An August 2026 Silicon Valley Investment Club profile reports a $5 million annual revenue run rate for 2024 and attributes the figure to NextPlatform. 低 SI003, SI004
CI011 The same profile lists a $7,015,510 revenue estimate attributed to Prospeo without identifying the measurement period. 低 SI003, SI004
CI012 Tracxn reports $493 million of total funding across three rounds for TensorWave. 中 SI008
CI013 Silicon Valley Investment Club reports $450 million-plus of total funding, creating a $43 million discrepancy against Tracxn's total. 中 SI003, SI004
CI014 The reviewed public sources provide no audited 2025-2026 revenue, gross margin, operating loss, cash burn, cash balance, runway, debt balance, or covenant schedule. 中 SI003, SI004, SI008
CI015 The disclosed operating scale, leased expansion capacity, and infrastructure-focused use of funds imply substantial forward capital needs, but available evidence does not quantify committed capex or financing obligations. 中 SI001, SI006
CI016 TensorWave's revenue quality, unit economics, and liquidity remain ununderwriteable without private evidence on paid utilization, customer concentration, realized pricing, service costs, asset obligations, and cash. 中 SI003, SI004, SI006, SI008
CE001 TensorWave offers an AMD-exclusive GPU cloud for model training, fine-tuning, and inference workloads. 高 SE002, SE004, SE006
CE002 As of the access date, TensorWave's public product surface names MI300X, MI325X, MI355X, and MI455X accelerators. 中 SE002
CE003 TensorWave's training page describes RoCEv2 spine-and-leaf networking, checkpoint-oriented paths, and topology-aware Slurm scheduling. 中 SE005
CE004 TensorWave and AMD describe continuous infrastructure monitoring and hands-on optimization support for customer workloads. 高 SE002, SE006
CE005 Both TensorWave and AMD describe the platform as SOC 2 Type II certified and HIPAA compliant. 高 SE002, SE006
CE006 A TensorWave-issued paid release says the company adopted a 72-GPU AMD Helios rack-scale design integrating Instinct GPUs, EPYC CPUs, Pensando networking, and ROCm. 中 SE003
CE007 TensorWave and AMD present onboarding, workload optimization, and scaling as hands-on engineering engagements. 高 SE002, SE006
CE008 Tech Times reports that CUDA-specific tooling can make migration to AMD infrastructure expensive even when accelerator performance or pricing is competitive. 中 SE004
CE009 Tech Times identifies supplier concentration because AMD supplies TensorWave's defining hardware and software while also investing in the company. 中 SE001, SE004
CE010 TensorWave's homepage carries a Moreh testimonial describing early production-cloud access to MI355X, but it says customer metrics were not yet ready for publication. 中 SE002
CE011 Pulse 2.0 reports TensorWave's claim that Fireworks AI and Luma AI use its infrastructure for generative-AI and production-inference workloads. 中 SE001
CE012 An independent company registry describes TensorWave's public offering as bare-metal access, managed inference, high-speed network storage, and direct liquid cooling. 低 SE007
CE013 A low-reputation sector review argues that ROCm trails CUDA in inference-engine support, framework compatibility, and operator coverage, but its quantitative assertions are not independently corroborated in the assigned evidence. 低 SE008
CE014 Independent uptime, incident-history, SLA, audit-scope, and developer-adoption evidence for TensorWave remains unavailable in the assigned public source pool. 低
CU001 TensorWave markets AMD Instinct GPUs and ROCm capacity for AI training and inference workloads. 中 SU003
CU002 TensorWave says dedicated solution engineers work with customers from onboarding through scaling. 中 SU003
CU003 TensorWave claims Moreh runs frontier-LLM inference on MI355X and reduced model-weight download time by approximately 75%. 中 SU003
CU004 Moreh's company-hosted quotation describes access to MI355X in a real production cloud while saying broader customer metrics were still too early to publish. 中 SU003
CU005 AMD describes TensorWave as an all-AMD AI cloud using Instinct GPUs, Ryzen PRO processors, and the Helios rack. 中 SU001
CU006 AMD provides a request-evaluation route for testing AI models and workflows on TensorWave. 中 SU001
CU007 A TensorWave job repost describes customer-facing dashboards and APIs for both multi-tenant cloud and single-tenant bare-metal deployments. 中 SU004
CU008 Tech Times frames pricing, accelerator supply, and single-vendor dependence as reasons customers may seek AMD-based infrastructure alternatives. 中 SU005
CU009 Tech Times reports that production adoption depends on reliability, performance, software compatibility, and migration cost. 中 SU005
CU010 RFP.wiki reports no verified TensorWave scores on G2, Capterra, Trustpilot, or Gartner Peer Insights, leaving broad customer sentiment unquantified. 低 SU002
CU011 RFP.wiki describes TensorWave production deployments as dedicated bare-metal infrastructure with optional managed Kubernetes or Slurm. 低 SU002
CU012 RFP.wiki relays independent testing that reported multiple multi-hour interruptions and immature multi-tenant controls during 2025. 低 SU002
CU013 Tracxn describes TensorWave's service as bare-metal accelerator access for machine-learning training and inference with orchestration, storage, and monitoring. 中 SU007
CU014 ClusterBid identifies cost-sensitive PyTorch training and batched-inference teams, including startups and academic labs, as TensorWave's ideal customer profile. 低 SU008
CU015 ClusterBid reports that workloads dependent on unported custom CUDA kernels may require significant migration engineering. 低 SU008
CU016 The retained assigned-pool evidence does not disclose customer count, retention metrics, contract duration, renewal rates, or top-customer concentration. 中 SU002, SU003, SU006, SU007
CR001 TensorWave's platform is explicitly concentrated on AMD Instinct accelerators and the ROCm ecosystem. 高 SR003, SR001, SR005, SR006
CR002 Current official and independent provider pages state that TensorWave has SOC 2 Type II and ISO 27001 certifications and HIPAA-compliant infrastructure. 高 SR003, SR002
CR003 TensorWave's CEO says specialization in one hardware ecosystem is intended to support more reliable cluster operations and customer experience. 中 SR001, SR003
CR004 ComputePrices reports that TensorWave does not publish public hourly pricing. 中 SR002
CR005 Multiple independent sources report a $350 million Series B at a $1.55 billion post-money valuation in June 2026. 中 SR004, SR005, SR008
CR006 Tech Funding News and Tracxn place TensorWave's aggregate funding at approximately $493 million. 中 SR005, SR008
CR007 Independent reporting says the 2026 financing is intended to support additional compute capacity, product development, hiring, and infrastructure expansion. 中 SR004, SR005
CR008 TechTimes identifies high capital requirements, changing equipment demand, financing costs, and lower utilization as material risks for specialized AI clouds. 中 SR004
CR009 Independent reporting flags AMD's dual supplier-and-investor role as a circular-financing concern that also concentrates TensorWave's supplier risk. 中 SR004, SR005
CR010 AMD's case study states that its cited TensorWave and Modular performance and cost-savings claims were not independently verified by AMD. 中 SR006
CR011 An adverse third-party analysis describes ROCm compatibility, framework support, and operator coverage as continuing disadvantages relative to CUDA. 低 SR009
CR012 An undated ClusterBid page says TensorWave had no SOC 2 certification yet, conflicting with current official and provider pages that state SOC 2 Type II certification. 低 SR009, SR003, SR002
CR013 Tech Funding News reports that TensorWave secured more than two gigawatts of long-term data-center capacity. 中 SR005
CR014 The retained capacity and financing reports do not disclose site-level lease obligations, power commitments, debt terms, or contracted utilization. 低 SR004, SR005, SR007, SR008
CR015 The retained official and provider pages describe monitoring and certifications but do not provide a complete incident history or contractual uptime SLA. 低 SR003, SR002
CR016 The public shareholder register and complete ownership structure are not disclosed in the retained company-data source. 中 SR007
CV001 TensorWave announced a $350 million Series B financing on June 10, 2026. 高 SV001, SV002, SV003, SV005, SV007
CV002 The Series B established a reported $1.55 billion post-money valuation. 高 SV001, SV002, SV003, SV005, SV007
CV003 Magnetar and AMD Ventures co-led the Series B. 高 SV001, SV002, SV003, SV005, SV007
CV004 Dividing the $350 million financing by the $1.55 billion post-money valuation yields approximately 22.6%, but this is headline dilution arithmetic rather than a fully diluted ownership estimate. 中 SV001, SV007
CV005 An August 2026 independent profile reports total funding of more than $450 million across three rounds. 中 SV007
CV006 Angel Investors Network reports approximately $493 million of funding across three rounds. 中 SV002
CV007 Public cumulative-funding totals conflict between more than $450 million and approximately $493 million. 中 SV002, SV007
CV008 TensorWave markets an AMD-exclusive cloud for training and inference workloads. 高 SV004, SV008
CV009 AMD claims TensorWave's optimized stack can reduce inference cost per million tokens by up to 70% versus traditional inference stacks. 中 SV008
CV010 Independent reporting identifies circular-financing risk because AMD is both TensorWave's accelerator supplier and a lead investor. 中 SV005
CV011 Angel Investors Network estimates GPU-cluster build and operating cost at $15 million to $20 million per megawatt and describes full deployment of 2 gigawatts as a tens-of-billions capital requirement. 低 SV002
CV012 Independent analysis identifies hyperscaler custom-silicon programs as a substitution risk for independent GPU-cloud demand. 中 SV002
CV013 Independent analyses report that ROCm still faces migration and tooling disadvantages relative to CUDA. 中 SV002, SV003, SV005
CV014 The reviewed 2024 seed deck does not provide financial projections or unit economics. 中 SV006
CV015 An independent profile lists public-market yardsticks including CoreWeave, Astera Labs, Arm, and Klaviyo, while stating that EV-to-revenue is not disclosed for any TensorWave round. 中 SV007
CV016 Without disclosed revenue, margins, utilization, customer concentration, debt, and preference terms, public evidence does not support a precise revenue-multiple, cash-flow, or return-based valuation range. 中 SV006, SV007
来源
编号出版方标题引文
SO001 TensorWave TensorWave Raises $350 Million Series B at $1.55B Valuation to Expand Global AMD-Powered AI Infrastructure
SO002 TensorWave The Hidden Price of Free Cloud Credits
SO003 Tracxn TensorWave TensorWave is a series B company based in Las Vegas (United States), founded in 2023
SO004 Silicon Valley Investment Club TensorWave Company Profile TensorWave, Inc. is the company behind tensorwave.com, a Las Vegas-based AI cloud and GPU infrastructure provider centered on AMD Instinct accelerators.
SO005 Silicon Valley Investment Club TensorWave — $1.55B · CEO Darrick Horton TensorWave, Inc. is the company behind tensorwave.com, a Las Vegas-based AI cloud and GPU infrastructure provider centered on AMD Instinct accelerators.
SO006 TensorWave TensorWave Powers Frontier AI Growth for Cloud Customers with AMD Helios Rackscale Solution
SO007 TensorWave TensorWave and Modular: Breaking AI Inference Barriers Beyond NVIDIA
SO008 Growth Engineer TensorWave — Review, Pricing & Alternatives TensorWave is an AI and HPC cloud platform built on AMD Instinct GPUs that accelerates LLM training, inference, and research workloads without the supply constraints and premium pricing of NVIDIA.
SO009 Tech Times TensorWave Raises $350M: AMD Funds an AI Cloud Challenger Built Without Nvidia The relationship can increase market competition while creating a circular-financing concern. Some investment capital may eventually return to AMD through hardware purchases, and TensorWave's exclusive dependence on AMD concentrates supplier risk.
SO010 ClusterBid TensorWave GPU Cloud: AMD-Focused AI Infrastructure Strategy TensorWave's fundamental challenge is not the hardware but the software ecosystem.
SM001 TensorWave via Business Wire TensorWave Raises $350 Million Series B at $1.55B Valuation to Expand Global AMD-Powered AI Infrastructure TensorWave is already seeing adoption across next-generation AI companies, including Fireworks AI and Luma AI, which are leveraging the company’s AMD-based infrastructure to power large-scale generative AI workloads and production inference systems.
SM002 ClusterBid TensorWave GPU Cloud: AMD-Focused AI Infrastructure Strategy TensorWave occupies a niche but growing segment: AI developers who want NVIDIA-competitive pricing and are comfortable with the ROCm ecosystem. Their ideal customer is a cost-sensitive AI team running PyTorch-based training or batched inference workloads that benefit from MI300X's larger memory.
SM003 AMD TensorWave Provides Compelling Reliability, Resiliency, and Cost Benefits With AMD Instinct GPUs All performance and/or cost savings claims are provided by TensorWave and/or Modular and have not been independently verified by AMD. Performance and cost benefits are impacted by a variety of variables. Results herein are specific to such 3rd party organizations and may not be typical.
SM004 TensorWave TensorWave AI Cloud for AMD Instinct Accelerators Leverage AMD Instinct GPUs and ROCm for cutting-edge, open AI acceleration. TensorWave positions you for future-ready AI performance without vendor lock-in.
SM005 Tech Times TensorWave Raises $350M: AMD Funds an AI Cloud Challenger Built Without Nvidia The harder test is whether businesses will use TensorWave for production workloads rather than treating AMD capacity as a backup option or negotiating tool. That decision will depend on reliability, performance, software compatibility, and the engineering cost of moving away from CUDA.
SM006 TensorWave About TensorWave Building seamless, secure, reliable, and resilient AI infrastructure at scale, eliminating barriers and challenging the status quo to empower builders and support AI innovation.
SM007 Silicon Valley Investclub TensorWave — $1.55B · CEO Darrick Horton Public profiles place TensorWave in Las Vegas, Nevada, and describe it as a private AI/HPC infrastructure company serving AI developers and enterprises.
SM008 Tracxn TensorWave The company has 579 active competitors, including 89 funded and 33 that have exited
SP001 TBPN Digest TensorWave raises $350M Series B to build AMD-exclusive AI supercomputing infrastructure The persistent objection to AMD has been the CUDA ecosystem — not the chip itself, but the decade-plus of libraries, developer tooling, and institutional familiarity Nvidia has built around it.
SP002 Data Center Dynamics AMD-based AI cloud TensorWave secures $350m Series B funding Since then, it has launched an AI training cluster with 8,192 AMD Instinct MI325X GPUs online, and is preparing larger MI355X deployments across several new data center regions in North America.
SP003 TechTimes TensorWave Raises $350M: AMD Funds an AI Cloud Challenger Built Without Nvidia The relationship can increase market competition while creating a circular-financing concern. Some investment capital may eventually return to AMD through hardware purchases, and TensorWave's exclusive dependence on AMD concentrates supplier risk.
SP004 ClusterBid TensorWave GPU Cloud: AMD-Focused AI Infrastructure Strategy TensorWave's fundamental challenge is not the hardware but the software ecosystem.
SP005 TensorWave TensorWave: AI Cloud for AMD Instinct Accelerators Leverage AMD Instinct GPUs and ROCm for cutting-edge, open AI acceleration.
SP006 AMD Test AI models and workflows with AMD Instinct GPUs on TensorWave We build exclusively on AMD Instinct™ GPUs—optimizing every layer of our platform to deliver maximum performance, efficiency, and deep architectural expertise.
SP007 AMD TensorWave Provides Compelling Reliability, Resiliency, and Cost Benefits With AMD Instinct GPUs All performance and/or cost savings claims are provided by TensorWave and/or Modular and have not been independently verified by AMD. Performance and cost benefits are impacted by a variety of variables. Results herein are specific to such 3rd party organizations and may not be typical.
SP008 Tracxn TensorWave TensorWave ranks 30th amongst 579 ... of which 89 are funded.
SI001 TensorWave via Business Wire TensorWave Raises $350 Million Series B at $1.55B Valuation to Expand Global AMD-Powered AI Infrastructure TensorWave, the all-AMD AI cloud specializing in high-performance, memory-intensive workloads, today announced it has raised $350 million in Series B funding.
SI002 TAMradar TensorWave Raises $350M Series B for AMD AI Infrastructure TensorWave (tensorwave.com), a Las Vegas-based provider of AMD-exclusive AI and HPC cloud infrastructure, has raised $350 million in Series B funding at a $1.55 billion valuation.
SI003 Silicon Valley Investment Club TensorWave Company Profile Annual revenue run rate — $5 million in 2024 NextPlatform
SI004 Silicon Valley Investment Club TensorWave - Company Profile EV/Rev is not disclosed for any round.
SI005 TensorWave TensorWave: AI Cloud for AMD Instinct Accelerators Run your AI workloads reliably with AMD accelerators, all while achieving leading price to performance.
SI006 Tech Times TensorWave Raises $350M: AMD Funds an AI Cloud Challenger Built Without Nvidia The risk is that these valuations depend on continued demand for expensive accelerator capacity. A slowdown in AI spending could leave specialized providers managing equipment commitments, financing costs, and lower utilization.
SI007 AMD Test AI Models and Workflows with AMD Instinct GPUs on TensorWave Up to 70% reduction in cost per million tokens versus traditional inference stacks.
SI008 Tracxn TensorWave TensorWave has raised $493M in funding
SE001 Pulse 2.0 TensorWave Raises $350 Million Series B At $1.55 Billion Valuation To Expand AMD-Powered AI Infrastructure According to the company, customers including Fireworks AI and Luma AI are already using TensorWave’s infrastructure to support large-scale generative AI workloads and production inference systems.
SE002 TensorWave TensorWave: AI Cloud for AMD Instinct Accelerators TensorWave AI Cloud engineers carefully monitor every datacenter location to keep your critical training and inference workloads running.
SE003 TensorWave via Business Wire and Yahoo Finance TensorWave Powers Frontier AI Growth for Cloud Customers with AMD Helios Rackscale Solution The 72-GPU AMD Helios rackscale solution integrates AMD Instinct GPUs, 6th Gen AMD EPYC™ CPUs, AMD Pensando™ networking, and AMD ROCm™ software into a unified, open rack-scale platform built for high-volume inference, frontier-model training, and fine-tuning at scale.
SE004 Tech Times TensorWave Raises $350M: AMD Funds an AI Cloud Challenger Built Without Nvidia AI accelerators compete through software as much as silicon. Developers need compilers, optimized libraries, debugging tools, machine-learning framework support, documentation, and engineers who understand how to tune workloads.
SE005 TensorWave AI Training | GPU Cloud Infrastructure for Model Training Run distributed AI training on networking, storage, and scheduling built for scale.
SE006 AMD AMD Instinct GPUs on TensorWave Cloud Leverages AMD physical GPU partitioning to enable dedicated chiplet-level isolation—unlocking higher utilization and throughput.
SE007 Silicon Valley Investclub TensorWave — $1.55B · CEO Darrick Horton Its site presents the business as an AMD-exclusive AI cloud for training, fine-tuning, and deploying large models, with offerings that include bare metal access, managed inference, high-speed network storage, and direct liquid cooling.
SE008 ClusterBid TensorWave GPU Cloud: AMD-Focused AI Infrastructure Strategy TensorWave's fundamental challenge is not the hardware but the software ecosystem.
SU001 Advanced Micro Devices TensorWave Advances AI Compute Infrastructure with AMD TensorWave delivers an all AMD technology-based AI cloud supercharged with AMD Instinct™ GPUs, AMD Ryzen™ PRO processors, and the AMD Helios rack.
SU002 RFP.wiki TensorWave vs NVIDIA DGX Cloud (2026): Data-driven comparison Independent testing reported multiple multi-hour outages and immature Slurm/Kubernetes multi-tenant controls in 2025.
SU003 TensorWave TensorWave: AI Cloud for AMD Instinct Accelerators TensorWave powers Moreh's frontier LLM inference on MI355X & reduces model weight download time by ~75%
SU004 TensorWave via Simplify Software Engineer - Full Stack Develop robust APIs supporting multi-tenant cloud architectures and single-tenant bare-metal deployments.
SU005 Tech Times TensorWave Raises $350M: AMD Funds an AI Cloud Challenger Built Without Nvidia The harder test is whether businesses will use TensorWave for production workloads rather than treating AMD capacity as a backup option or negotiating tool.
SU006 Silicon Valley Investment Club TensorWave - $1.55B - Darrick Horton, Piotr Tomasik Public profiles place TensorWave in Las Vegas, Nevada, and describe it as a private AI/HPC infrastructure company serving AI developers and enterprises.
SU007 Tracxn TensorWave This service offers bare metal access to hardware accelerators designed for machine learning training and inference tasks.
SU008 ClusterBid TensorWave GPU Cloud: AMD-Focused AI Infrastructure Strategy Their ideal customer is a cost-sensitive AI team running PyTorch-based training or batched inference workloads that benefit from MI300X's larger memory.
SR001 SiliconANGLE TensorWave targets focused AI cloud strategy to deliver better customer experience TensorWave has remained committed to AMD because a singular focus allows the company to deliver a better customer experience.
SR002 ComputePrices TensorWave GPU Pricing & Specs TensorWave does not publish public hourly pricing.
SR003 TensorWave TensorWave: AI Cloud for AMD Instinct Accelerators 24/7 monitoring of critical infrastructure
SR004 TechTimes TensorWave Raises $350M: AMD Funds an AI Cloud Challenger Built Without Nvidia The relationship can increase market competition while creating a circular-financing concern.
SR005 Tech Funding News The Forbes 30 Under 30 CEO who left Lockheed Martin's Skunk Works raises $350M at $1.55B to challenge Nvidia's grip on AI infrastructure What differentiates TensorWave is its all-in position on AMD, a single-vendor strategy that provides deep product integration and favourable commercial terms but concentrates technical and supply chain risk on one chipmaker’s roadmap.
SR006 AMD TensorWave Provides Compelling Reliability, Resiliency, and Cost Benefits With AMD Instinct GPUs All performance and/or cost savings claims are provided by TensorWave and/or Modular and have not been independently verified by AMD.
SR007 Silicon Valley Investclub TensorWave — $1.55B · CEO Darrick Horton The full shareholder register is not public.
SR008 Tracxn TensorWave TensorWave has raised $493M in funding
SR009 ClusterBid TensorWave GPU Cloud: AMD-Focused AI Infrastructure Strategy TensorWave's fundamental challenge is not the hardware but the software ecosystem.
SV001 TensorWave (syndicated by Magnetar Capital) TensorWave Raises $350 Million Series B at $1.55B Valuation to Expand Global AMD-Powered AI Infrastructure TensorWave, the all-AMD AI cloud specializing in high-performance, memory-intensive workloads, today announced it has raised $350 million in Series B funding. The round was co-led by Magnetar and AMD Ventures, with continued participation from Maverick Silicon, Nexus Venture Partners, and Western Frontier.
SV002 Angel Investors Network TensorWave $350M Series B: AMD AI Infrastructure Analysis Capital intensity. Building and operating GPU clusters costs $15M to $20M per megawatt. At 2 gigawatts of secured capacity, TensorWave is looking at potential capital requirements in the tens of billions to fully deploy. $493M in total funding does not get you close to full buildout. Expect more fundraising. Expect dilution.
SV003 Zlvox TensorWave Raises $350M at $1.55B Valuation: What It Means for AMD and the AI Industry A $1.55 billion valuation comes with high expectations, and TensorWave faces real challenges on the path to justifying it.
SV004 TensorWave TensorWave: AI Cloud for AMD Instinct Accelerators Run your AI workloads reliably with AMD accelerators, all while achieving leading price to performance. TensorWave ensures efficient scaling without compromise.
SV005 Tech Times TensorWave Raises $350M: AMD Funds an AI Cloud Challenger Built Without Nvidia The relationship can increase market competition while creating a circular-financing concern. Some investment capital may eventually return to AMD through hardware purchases, and TensorWave's exclusive dependence on AMD concentrates supplier risk.
SV006 Startup Fundraising TensorWave Pitch Deck (2024): 10-Slide Seed Deck There are no financial projections or unit economics showing the margin profile of renting AMD chips versus the cost of power and hardware depreciation.
SV007 Silicon Valley Investclub TensorWave Company Profile EV/Rev is not disclosed for any round.
SV008 AMD AMD Instinct GPUs on TensorWave Cloud We build exclusively on AMD Instinct™ GPUs—optimizing every layer of our platform to deliver maximum performance, efficiency, and deep architectural expertise.