初创公司尽调
尽调报告 Semiconductors / AI Infrastructure Series E 2026-05-10

Tenstorrent

AI 芯片挑战者:Blackhole、RISC-V 与 CUDA 替代路线

Tenstorrent 是技术上可信的 AI 芯片挑战者,靠差异化 RISC-V 架构和 $2B 资本支撑;但收入未确认、软件尚不成熟, 且 TSMC 单一供货带来风险,在 $3.2B 估值下应继续研究。

封面要素

最近融资 01
$800M Series E [CI007]
投后估值 02
3200 USD M [CI008]
累计融资 03
~$1.99B [CI001]
Galaxy 服务器 GA 04
April 2026 [CE005]
Blackhole 算力 05
664 TFLOPS FP8 [CE001]
开源软件 06
Apache 2.0 [CE018]

公司概况

Tenstorrent 是一家位于 San Jose 的 AI 芯片公司,由 Jim Keller(AMD Zen 和 Apple A-series 背后的 CPU 架构师)于 2016 年创立。公司设计 Blackhole ASIC——一颗 TSMC 6nm 芯片,集成 120 个自研 Tensix 核心和 16 个大型 RISC-V 处理器,面向高性能 AI 推理和边缘部署。Tenstorrent 五轮融资累计约 $1.99B(最近一轮为 2025 年 11 月、估值 $3.2B 的 $800M Series E),投资方包括 Samsung Securities、LG Technology Ventures、Hyundai Motor Group、Fidelity 等战略投资者。公司差异化的 RISC-V 架构绕开 ARM 授权费,并支撑 TT-Metal、TT-Forge、Apache 2.0 等开源软件栈,目标是打破 NVIDIA 的 CUDA 生态锁定效应。

官网
tenstorrent.com
成立时间
2016-01-01
创始人
Jim Keller, Ljubisa Bajic
创立地点
Toronto, Canada
总部
San Jose, CA, USA
产品
AI 加速器硬件(Blackhole p100a/p150a/p150b PCIe 卡;Galaxy Blackhole 6U 服务器)和开源 ML 软件栈(TT-Metal/TT-NN 运行时、TT-Forge MLIR 编译器、TT-Metalium 编程模型)。Galaxy 服务器于 2026 年 4 月 GA,单机箱约 $110K;同时向 OEM 授权 RISC-V Tensix Neo IP。
客户
AI/ML 推理工作负载;超大规模云厂商和云服务商(次要);汽车 AI(Hyundai、JLR);企业 AI 基础设施;通过 DevCloud 覆盖研究机构和开发者。
商业模式
硬件产品销售(PCIe 卡、Galaxy 服务器);通过合作伙伴 Koyeb 提供云端硬件即服务(HaaS);向 OEM 授权 RISC-V IP;开发者生态(开源加商业支持)。
阶段
Series E
融资情况
$800M Series E(2025 年 11 月,投后估值 $3.2B,Fidelity 领投);$693M Series D(2024 年 12 月,投后估值 $2.6B);累计融资约 $1.99B。
[CO001, CO002, CI001, CI007, CI008, CE001, CE018]

执行摘要

主要优势

  • 独特的 Tensix + RISC-V 架构绕开 ARM 授权,并支撑开源软件栈
  • Jim Keller 的 CPU 设计履历加上强工程团队
  • 已融资 $2B+,LG、Hyundai、SoftBank、Fidelity 等战略锚定投资人参与
  • Galaxy Blackhole 服务器已 GA(2026 年 4 月)——首个商业收入里程碑
  • 与 90% HuggingFace 模型兼容,支持 2.5M+ 开源模型(公司声称)
  • Apache 2.0 开源软件降低开发者采用门槛

主要风险

  • 依赖 TSMC 6nm 单一供货:台湾地缘政治风险可能让生产停摆
  • 软件成熟度落后 NVIDIA CUDA:The Register(2025 年 11 月)认为它“还不够打磨成熟”
  • 收入未披露;估计每月烧钱 $25–50M,却支撑未经验证的 $3.2B 估值
  • 美国 BIS AI 芯片出口管制可能压缩可触达市场,或要求申请许可
  • 投资人兼客户集中(LG、Hyundai):若最大客户同时是董事会层面的投资人,治理冲突会放大
  • 关键人物风险:Jim Keller 若离开,将显著削弱投资人信心

未决问题

  • 已确认的 2025/2026 年收入和毛利率(私有公司,未披露)
  • Blackhole 664 TFLOPS FP8 的出口管制状态及 BIS 许可适用性
  • 具名超大规模云厂商或云端设计定点(已确认客户均为投资人伙伴)
  • 净收入留存率和客户流失数据(Galaxy 仍太早)
  • Blackhole 之后的芯片路线图和流片时间表(受 NDA 限制)

目录

Chapter 01

01公司概况

1.1 身份、商业模式与运营概览

Tenstorrent Inc. 于 2016 年 3 月 14 日在加拿大注册,运营总部位于加州 Santa Clara 的 2600 Great America Way, Suite 501。公司公开使命是靠开源软件栈、RISC-V 指令集架构和成本效率更高的 Tensix 核心处理器,让高性能 AI 硬件更普及。主要收入来自硬件销售(PCIe 加速卡和工作站)与 IP 授权(Tensix 和 Ascalon RISC-V CPU IP)。Tenstorrent 还基于自有基础设施提供 AI 开发者云服务,形成早期云收入。全球办公室覆盖 Toronto(工程中心)、Austin(Texas)、Fort Collins(Colorado),以及 Belgrade(Serbia)、Tokyo(Japan)、Bengaluru(India)、Seoul 和 Pangyo(Korea)、Munich(Germany)、Warsaw(Poland)、Beijing(China)等国际地点。截至 2026 年中,公司全球员工约 1,100–1,200 人,较 2024 年 12 月 Series D 后大幅增加。Tenstorrent 采用无晶圆厂半导体模式,芯片制造依赖 GlobalFoundries(早期芯片)、TSMC 和 Samsung 等代工伙伴。核心架构差异在 Tensix 处理核心:这是一个自包含计算单元,包含 RISC-V 数据搬运处理器、面向张量运算的矩阵数学引擎和向量数学单元;片上 Ethernet 互连结构连接这些计算单元,让芯片可直接横向扩展,不必依赖昂贵交换机基础设施。开源理念也延伸到编译器栈(TT-Metal/TT-Metalium、TT-Forge MLIR、TT-Buda、TT-Lang),这些组件采用 MIT 许可,并在 GitHub 公开。 [CO001, CO002, CO003, CO004, CO005, CO006]

FO002: 公司快照——业务系统逻辑

Tenstorrent 的身份、产品、资本和依赖关系如何彼此相连。

[CO003, CO004, CO007, CO009, CO013, CO022]

1.2 创始人、领导层与治理

Tenstorrent 由 Ljubisa Bajic(前 CEO/CTO)、Ivan Hamer 和 Milos Trajkovic 于 2016 年共同创立。这三位加拿大工程师此前一起研究深度学习硬件架构,构成了公司 Tensix 核心设计的起点。公司扩张后,Bajic 转任技术 Fellow;Hamer 和 Trajkovic 仍是高级 Fellow 级工程贡献者。Jim Keller 于 2020 年加入,担任总裁和 CTO,并在 2023 年初正式升任 CEO。Keller 是半导体史上履历最强的芯片架构师之一,曾主导 AMD Athlon K7 和 K8/Opteron 微架构(包括 x86-64 指令集)、Apple A4 和 A5 移动 SoC 设计(iPhone 4 和 4S、初代 iPad)、Tesla Full Self-Driving Hardware 3 芯片,并在 Intel 承担高级硅工程职责。Keith Witek 担任 COO,也是 Series D 融资的公开发言人。领导团队还包括 David Bennett(首席客户官)和 Erik Goodman(财务副总裁)。Jim Keller 的关键人物集中度构成实质治理风险:他在过往雇主任期偏短(Intel:2 年;Tesla:2 年),引发继任担忧,分析师也将其列为尽调警示。投资方中,除 Series C 和 Series D 的战略企业投资者外,Eclipse Ventures 和 Real Ventures(早期支持者)也通过董事会和顾问角色参与。 [CO010, CO011, CO012, CO013, CO014, CO015]

领导层与创始人表
人物职务背景创始人-市场匹配关键人物风险
Jim KellerCEO(自 2023 年初起)曾领导 AMD K7/K8/Zen、Apple A4/A5、Tesla FSD HW3、Intel 硅工程 SVP世界级 CPU 架构师,AI 芯片设计经验深高 – 过往任期较短;是公司面对投资者的关键人物
Keith WitekCOO半导体 / 科技公司运营领导经验运营规模化经验中 – Series D 发言人;未宣布继任者
Ljubisa Bajic联合创始人,高级研究员(前 CEO/CTO)博士级芯片架构师;共同设计原始 Tensix 核心领域积累深的创始人;原始技术愿景低 – Keller 之后转为技术贡献者角色
Ivan Hamer联合创始人,高级研究员硬件架构师;共同设计 Tensix 架构核心创始技术团队
Milos Trajkovic联合创始人,高级研究员 – 系统工程与软件AI 硬件系统与软件工程把硬件接到软件栈
David Bennett首席客户官企业销售和客户成功商业化爬坡的关键人物
[CO010, CO011, CO012, CO013, CO014, CO015]

1.3 融资历史与投资者格局

自 2017 年 Real Ventures 种子轮以来,Tenstorrent 已完成十轮有记录的融资,截至 2026 年 5 月累计融资约 $1.18B。融资节奏分为三个阶段:早期轮次(2017 年种子轮、2018 年 2 月 $500K Series A、2019 年 1 月 $20.5M Series B)支持原型开发和早期工程团队扩张;2021 年的 Series C 组合让公司进入独角兽行列(2021 年 4 月 $164M 批次,加上 2021 年 5 月 Fidelity 领投、投后估值 $1B 的 $200M 批次);2023 年 8 月由 Hyundai Motor Group 和 Samsung Catalyst Fund 领投的 $100M Series C 延伸轮;以及 2024 年 12 月标志性的 $693M Series D,由 Samsung Securities 和 AFW Partners 领投,投前估值 $2B、投后估值 $2.6B。Series D 被描述为超额认购。主要投资者包括 Samsung Securities(领投)、AFW Partners(领投)、LG Electronics、Hyundai Motor Group、Fidelity Management & Research Company、Bezos Expeditions(Jeff Bezos)、Baillie Gifford、XTX Markets、Export Development Canada、Healthcare of Ontario Pension Plan、Corner Capital、MESH Ventures、SBI Investment、Eclipse Ventures 和 Real Ventures。投资者组合同时覆盖韩国战略集团(Samsung、LG、Hyundai、Kia)、长期机构资管(Fidelity、Baillie Gifford)、主权 / 养老金资金(Export Development Canada、HOOPP)和财务投资人(AFW Partners、XTX Markets)。2021 年公司监管文件披露收入为 $25M–$100M;之后未公开披露收入。公司称,截至 2024 年 12 月融资时,已签合同约 $150M。 [CO018, CO019, CO020, CO021, CO022, CO023]

快照 KPI 表
指标数值 / 状态日期 / 期间置信度缺口 / 备注
投后估值$2.6BDec 2024(Series D 轮)私有信息;无独立验证
累计融资~$1.18B截至 May 2026Tracxn/Crunchbase 汇总
最新轮次Series D 轮 – $693MDec 2, 2024官方新闻稿确认
投前估值(Series D 轮)$2.0BDec 2024新闻稿披露
收入运行率(最新披露)$25M–$100M2021(监管文件)后续无公开披露
已签合同~$150M截至 Dec 2024公司在融资时声称
员工数(估计)~1,100–1,2002026 年中第三方估计;未确认
成立2016Mar 14, 2016(加拿大注册成立)公司注册资料(Tracxn)
已出货芯片代际3(Grayskull、Wormhole、Blackhole)截至 May 2026多份独立评测确认

收入和员工数来自第三方数据库(Tracxn、分析师报告)的估计;估值和融资来自官方新闻稿和 Crunchbase。

[CO018, CO019, CO020, CO022, CO025, CO035]
利益相关方或投资者图谱
利益相关方角色 / 轮次经济或战略重要性关键尽调问题
Samsung SecuritiesSeries D 领投方($693M,Dec 2024)韩国最大证券公司;与 Samsung 生态关系深,有潜在设计定点机会投资后是否有董事席位或顾问权?
AFW PartnersSeries D 联合领投方首尔 VC;聚焦出行 / 半导体;与 Samsung 联合领投对韩国市场策略有多大影响力?
LG ElectronicsSeries D 参与方战略客户和投资者;车载 / 家电 AI 芯片机会IP 授权或设计定点管线?
Hyundai Motor GroupSeries C 领投方(Aug 2023)+ Series D大型汽车 OEM;车载 AI 芯片设计定点潜力车载芯片开发合同范围
Fidelity Management & Research CompanySeries C(May 2021 领投)+ Series D长期持有型机构;传递 IPO 准备度信心IPO 路径的收入和增长可见度?
Bezos Expeditions (Jeff Bezos)Series D 参与方高知名度个人背书;传递市场可信度是否有 Amazon/AWS AI 芯片协同?
Baillie GiffordSeries D 参与方英国长期成长投资者(Tesla、SpaceX);耐心资本长期持有逻辑;无短期退出压力
Eclipse Ventures早期支持者(种子 / Series B 阶段)深科技硬件 VC;董事会层面参与董事席位;治理控制权?
Real Ventures种子 / 早期支持者加拿大 VC;2017 年起的原始支持者持股被稀释;剩余治理影响力?
Export Development Canada (EDC)Series D 参与方加拿大国有机构;主权属性资本政府牵引的 R&D 或采购机会?

各轮单个参与方的投资金额除 Series D 总额($693M+)外未公开。轮次标签沿用 Tracxn/Crunchbase 口径。

[CO019, CO020, CO021, CO022, CO023, CO024]
FO003: 快照 KPI——Tenstorrent 关键指标

截至 2026 年 5 月的时点 KPI 快照,混合已确认与估算值。

员工数为第三方估算。收入未公开披露。KPI 日期因指标而异。

[CO018, CO019, CO022, CO025, CO035, CO036]

1.4 关键里程碑与产品历史

Tenstorrent 的里程碑横跨约九年、三代硬件。2016–2018 年,创始团队从第一性原理设计 Tensix 核心架构,拿到种子轮和 Series A 资金,并开发 Grayskull 处理器(第一代 Tensix)。Grayskull 是 Tenstorrent 面向开发者的入门产品,最多 120 个 Tensix 核心、每个核心 1MB SRAM,并支持 8GB LPDDR4 内存;对应 e75 和 e150 PCIe 加速卡,硅片性能最高 600 TOPS。Series B($20.5M,2019 年 1 月)和 Series C 组合(2021 年,两笔合计 $364M)支持了 Wormhole 开发。Wormhole(第二代,2024 年 7 月以 n150 和 n300 PCIe 卡、TT-LoudBox / TT-QuietBox 工作站商业发布)带来显著微架构升级:80 个 Tensix+ 核心(数量更少但能力更强)、12nm 制程、16×100GbE 片上 Ethernet、GDDR6 内存(每卡 12GB)和 328 TOPS 峰值性能。Jim Keller 于 2020 年加入,担任总裁和 CTO;2023 年初出任 CEO。Hyundai 和 Samsung Catalyst 在 2023 年 8 月的 $100M Series C 延伸轮,为 Blackhole 开发提供资金。Blackhole(6nm 制程、140 个 Tensix++ 核心、10×400GbE、32GB GDDR6、790 TOPS FP8)于 2025 年末以 QuietBox 工作站形态开始出货,Galaxy Blackhole 服务器则在 2026 年 5 月进入量产。$693M Series D 于 2024 年 12 月 2 日完成,支撑公司激进招聘和全球基础设施扩张。 [CO028, CO029, CO030, CO031, CO032, CO033]

里程碑表
日期事件类型金额 / 估值 / 状态参与方 / 备注含义
Mar 2016Tenstorrent Inc. 在加拿大注册成立创立N/ALjubisa Bajic、Ivan Hamer、Milos Trajkovic 三位联合创始人法律实体成立;初始架构启动
May 2017Real Ventures 参与的种子轮融资未披露Real Ventures首笔机构资本;Tensix 概念在内部获验证
Feb 2018Series A 完成融资$500K未披露投资方早期工程资金,用于原型开发
Jan 2019Series B 完成融资$20.5M未披露扩充工程团队;首颗 Grayskull 硅片
2020Jim Keller 加入,出任总裁兼 CTO治理N/AJim Keller(前 Intel SVP)可信度重大拐点;加快产品路线图和人才吸引
Apr 2021Series C 第一笔完成融资$164M未披露领投方Wormhole 开发的主要扩张资金
May 2021Series C 第二笔完成;估值达到 $1B融资$200M,投后估值约 $1BFidelity Investments(领投)、Moore Capital、Real Ventures、Eclipse独角兽地位;Fidelity 首次参与
Jun 2021常规债务融资融资未披露未披露贷款方补充资本;过桥或设备融资
2023 年初Jim Keller 正式获任 CEO治理N/A董事会决定领导权正式化;支撑商业规模融资
Aug 2023Hyundai 和 Samsung Catalyst 领投 Series C 延展轮($100M)融资$100MHyundai Motor Group(领投)、Samsung Catalyst Fund、Fidelity、Maverick Capital、Kia、Eclipse韩国战略投资者进入;车载 AI 芯片信号
Jul 2024Wormhole n150/n300 加速卡和 TT-LoudBox/QuietBox 工作站商业发布产品n150 $1K、n300 $1.4K、TT-LoudBox $12K、TT-QuietBox $15K(公开定价)Tenstorrent(Jim Keller 引述);Forbes 报道首批面向大众市场的开发者硬件可下单
Dec 2, 2024Series D 完成 $693M 融资;投后估值 $2.6B融资$693M,投前 $2B / 投后 $2.6BSamsung Securities + AFW Partners(领投);LG、Hyundai、Bezos、Fidelity、Baillie Gifford、XTX Markets、EDC、 HOOPP, Corner Capital, MESH, SBI单轮最大融资;支持 Galaxy Blackhole 和全球招聘
2025 年末Blackhole QuietBox 工作站开始发货产品$11,999The Register 上手评测确认Blackhole 硬件到达客户手中;开发平台早于 Galaxy 服务器落地
May 2026Galaxy Blackhole 进入量产;DeepSeek 基准测试创纪录产品308 tokens/sec/user,$6/M tokens;36-box 超级集群Futurum Group 分析;部署在 Tokyo、Seattle、India商业化拐点;生产级 AI 服务器批量出货

内部里程碑(架构、流片硅片)的日期未公开。融资金额和投资者来自 Tracxn、Crunchbase 和官方新闻稿。

[CO001, CO018, CO019, CO020, CO021, CO022]
FO001: 公司里程碑时间线

Tenstorrent 从创立(2016 年)到 Galaxy Blackhole 量产(2026 年 5 月)的关键里程碑。

Jim Keller 加入日期(2020 年)和升任 CEO(2023 年初)根据新闻报道估算;具体日历日期未公开确认。

[CO001, CO010, CO018, CO019, CO020, CO028]

1.5 规模、竞争位置与关键指标

Tenstorrent 是私营公司,除 2021 年一份企业文件披露收入为 $25M–$100M 外,没有公开财务披露。截至 2026 年 5 月,公司称已签商业合同约 $150M,Galaxy Blackhole 服务器已量产,并在至少五个 neocloud 机房部署(Tokyo、Seattle、India,以及通过 Equinix/OrionVM/BetterBrain 合作宣布的另外两个地点)。全球员工数估计为 1,100–1,200 人。公司开源软件栈(TT-Forge、TT-Metal)声称 Hugging Face 模型通过率达到 90%,并支持约 250 万个开源 AI 模型。Blackhole 展示了 DeepSeek 推理:每用户 308 tokens/sec,输出 token 成本 $6/M;还与 Prodia 创下视频生成基准纪录(5 秒视频用 3.5 秒生成,比此前纪录快 83%)。尽管技术履历强,Tenstorrent 面对 NVIDIA 仍有显著竞争劣势,后者 CUDA 生态沉淀了 20 多年库开发。独立评测者也承认,公司软件栈相对 CUDA 仍不成熟,限制近期大型企业采用。第三方估计其收入在数千万美元量级,远低于 NVIDIA 单季数百亿美元的数据中心收入。 [CO035, CO036, CO037, CO038, CO039, CO040]

1.6 附录

Chapter 02

02市场分析

2.1 市场边界与定义

Tenstorrent 所处市场是 AI 加速器芯片。本报告将其定义为专门加速数据中心、云、neocloud、企业和边缘部署中 AI 训练与推理工作负载的离散处理器,包括 GPU、NPU 和定制 ASIC。本章以 AI 加速器芯片市场作为主要分析单元,同时承认不同来源的分析师定义差异显著。 纳入支出包括:AI 专用处理器硅片(离散 PCIe 卡、OAM 模块、多芯片模块)、由加速器贡献主要价值的完整 AI 服务器系统,以及 Tenstorrent Ascalon 核心相邻收入流中的 RISC-V IP 授权。排除支出包括:高带宽内存(HBM/GDDR 视为独立商品市场)、标准 CPU 硅片、网络交换芯片、AI 软件订阅(MLOps、API),以及服务器机箱、供电和冷却基础设施。 主要相邻市场是 RISC-V 处理器 IP 授权,Tenstorrent 通过 Ascalon 64-bit RISC-V CPU 核心进入该市场。RISC-V IP 市场 2025 年约 $580M,预计 2026 年达到 $720M(Intel Market Research 口径 CAGR 为 12.1%)。主市场的现状替代品是 NVIDIA GPU 产品线(H100、H200、Blackwell B100/B200/GB200 系列),其在 2025 年约占 AI 加速器芯片收入的 80%。 市场边界直接影响 Tenstorrent 估值,因为可获取市场(SOM)取决于哪些买方群体今天能现实采用非 CUDA 芯片。超大规模云厂商大多已被 NVIDIA 锁定,或在自研定制硅片。neocloud、主权算力项目和早期采用型企业 AI 实验室,才是更现实的滩头阵地。 [CM001, CM002, CM003, CM004, CM005, CM006]

市场定义表
细分 / 类别纳入支出排除支出买方 / 付款方Tenstorrent 关联度
AI 训练加速器神经网络训练用 GPU/ASIC 硅片(H100、H200、B100、自研)网络互连结构、服务器机箱、电力基础设施超大规模云厂商、大型 AI 实验室、新云厂商短期低:Blackhole 面向推理;未来训练业务仍是愿景
AI 推理加速器规模化服务 LLM 和扩散模型的 GPU/NPU/ASIC 硅片API 定价、MLOps 软件、内存新云厂商、企业 AI、超大规模云厂商边缘主要商业市场:截至 May 2026,Galaxy Blackhole 已部署到 5+ 个新云厂商托管点用于推理
RISC-V 处理器 IP 授权授权给芯片厂商和 OEM、用于 SoC 集成的处理器核心 IP(Ascalon 64-bit)硅片制造、封装、组装半导体 OEM、汽车 Tier-1 供应商、嵌入式设备厂商活跃:Tenstorrent Ascalon 核心和 Tensix 核心 IP 已授权;Samsung 投资者关系打开车载 SoC 路径
边缘 / 嵌入式 AI 硅片车载、IoT、工业设备端推理用 NPU/AI 加速器 IP云计算、DRAM 内存汽车 OEM、工业设备供应商、IoT 芯片厂商愿景:未来边缘硅片产品;尚未商业规模化

市场边界沿用分析师惯例:计量单位是芯片 / IP 硅片,而不是下游服务或内存。RISC-V IP 是与硬件并行推进的相邻收入流。

[CM001, CM003, CM004, CM005]
FM001: AI 加速器市场——TAM / SAM / SOM 分层测算

TAM 以 Gartner AI 处理半导体估算($268B)为基础。SAM 按 TAM 的约 20% 推导,依据是 NVIDIA 约 80% 份额。SOM 以已披露 $150M 合同管线为下限;现实可触达份额估计为新云厂商和 RISC-V IP 授权合计占 SAM 的低到中个位数百分比。所有数值都是估算,置信区间很宽。

2.2 市场规模——TAM、SAM 与分析矛盾

2025–2026 年 AI 加速器芯片市场的总可用市场(TAM)估计跨度很大,且彼此矛盾。分歧主要来自分析师口径不同:Gartner 预计 2026 年 AI 处理半导体收入为 $268B(约占其 $1.3T 半导体总预测的 30%);IDC 则认为同年仅数据中心半导体收入就达 $477B。Fortune Business Insights 对专用 AI 加速器市场的估算为 2025–2026 年 $137B–$180B(口径更窄,只含训练 + 推理加速器)。Deloitte 2026 半导体展望估计生成式 AI 芯片收入约 $500B,这是最宽口径,包含 AI 相邻的内存和网络。 下方规模表保留并量化这些矛盾。对 AI 加速器(离散推理 + 训练硅片)而言,最保守且可辩护的 2026 年 TAM 约为 $200–$270B。非 NVIDIA AI 加速器的可服务市场(SAM)按 NVIDIA 约 80% 市占率反推:在保守 TAM 区间下,SAM 约 $40–$54B;若采用更宽口径,则约 $40–$100B。 AI 推理子市场是观察 Tenstorrent 的关键镜头。MarketsandMarkets 估计,AI 推理市场 2025 年约 $106B,2026 年增长至 $117–$120B,并以 19% CAGR 在 2030 年达到 $255B。推理正在超过训练,成为算力预算的主导驱动:到 2026 年,估计约三分之二 AI 算力由推理驱动,而 2023 年仅约三分之一。Tenstorrent 的 Galaxy Blackhole 直接竞争 AI 推理市场,并展示了具有成本竞争力的推理能力(DeepSeek:$6/M tokens、308 tokens/sec/user)。 RISC-V IP 市场给 Tenstorrent 增加第二条收入维度。Global Market Insights 预计,整体 RISC-V 技术市场 2025 年为 $1.35B,2026 年为 $1.91B,CAGR 为 30–41%。更窄的 RISC-V CPU IP 授权市场 2025 年约 $580M,2026 年约 $720M(Intel Market Research 口径 CAGR 为 12.1%)。更宽市场的增长来自 AI/ML 边缘采用、汽车 ADAS,以及韩国、日本、中国和欧盟的地缘政治驱动。 仅靠公开来源无法确定 Tenstorrent 的可获取市场(SOM)。公司披露截至 2024 年 12 月已签合同约 $150M,暗示其在目标细分市场的渗透率仍为低个位数。要建立现实 SOM 模型,需要公开市场拿不到的私有合同数据和部署量。 [CM007, CM008, CM009, CM010, CM011, CM012]

TAM/SAM/SOM 或规模测算视角表
发布方年份地域数值(2026 或最近年份)CAGR范围 / 方法置信度局限
Gartner2026全球$268B(AI 处理半导体)~28% YoY厂商收入跟踪;AI 半导体约占 $1.3T 总额的 30%不含内存;AI 半导体定义较宽
IDC2026全球$477B(数据中心半导体)较 2025 年 ~53% YoY收入跟踪,包含 AI 优化数据中心硅片包含内存和网络芯片;定义最宽
Deloitte2026全球~$500B(生成式 AI 芯片)~50% YoY展望报告;范围最宽,包含 AI 内存和配套硅片估计值;包含 AI 周边内存;并非纯加速器
Fortune Business Insights2025–2026全球$113B–$180B(AI 加速器市场)~26–27% CAGR 2025–2034按芯片类型自下而上测算;仅训练 + 推理加速器付费墙摘要;仅离散加速器;不含 CPU
MarketsandMarkets2025–2026全球$106B–$120B(仅 AI 推理市场)~19% CAGR,到 2030 年达 $255B推理专门细分;不含训练付费墙;仅推理子市场,不是完整加速器 TAM
Global Market Insights(市场研究机构)2025–2026全球$1.35B–$1.91B(RISC-V 技术市场)30–41% CAGR 2025–2034RISC-V 处理器和生态收入,包含 IP 授权和芯片广义 RISC-V 市场;更窄 CPU IP 授权子细分为 $580M–$720M
Intel Market Research(市场研究机构)2026–2034全球$720M(2026 年 RISC-V CPU IP 市场)2026–2034 年 CAGR 12.1%,到 $1.8B仅 IP 授权收入;处理器核心授权发布方声誉低;与 GMI 估计大体一致
Silicon Analysts2024–2026全球~80% NVIDIA 份额 → 非 NVIDIA SAM 约 $40B–$54BN/A(市场份额分析)NVIDIA 收入数据 + 总市场;SAM 按剩余市场推导推导估计;非 NVIDIA SAM 包括 AMD、自研芯片和其他所有厂商

TAM 估计会因是否计入 AI 内存、网络和自用芯片而相差 2–3 倍。所有数字都是分析师预测或估计,不是经审计收入。非 NVIDIA AI 加速器的 SAM 为推导值;Tenstorrent 的 SOM 无法从公开信息确认。

[CM007, CM008, CM009, CM010, CM011, CM012]
FM002: AI 芯片 TAM——分析师估算区间(2026)

低端:Fortune Business Insights 的狭义加速器口径;基准:Gartner AI 处理半导体;高端:IDC 数据中心半导体,包含 AI 内存和网络。Deloitte 的约 $500B 估算(最宽口径,包含 AI 相邻内存)作为高端之外的异常值注明。

2.3 买方分层与采用路径

AI 加速器市场可分为五类买方,它们的预算归属、采购权和 AI 芯片采用行为都明显不同。 超大规模云厂商(Amazon AWS、Google Cloud、Microsoft Azure、Meta)主导需求:五大超大规模云厂商 2026 年预计在 AI 基础设施上支出 $650–700B,其中约 70–75% 为 AI 专项支出。它们设定参考价格和供应条款。超大规模云厂商也在内部自研定制硅片(Google TPU、Amazon Trainium/Inferentia),因此短期内并不是 Tenstorrent 商业硬件的市场。不过,向其 SoC 团队授权 RISC-V IP,是一条更长期路径。 neocloud(CoreWeave、Lambda Labs、Crusoe、Nebius、ai&co、Cirrascale、Turium、Virtu Financial AI、Prodia)是截至 2026 年 5 月 Tenstorrent 的主要商业滩头阵地。Futurum Group 确认,到 2026 年 5 月,Galaxy Blackhole 已部署在至少五个 neocloud 机房。neocloud 细分市场预计 2026 年收入约 $20B,到 2030 年增长至 $180B。neocloud 正主动评估非 NVIDIA 芯片,以降低 GPU 依赖,并在推理中寻找成本优势。 企业买方(金融、医疗、制造、汽车 OEM)越来越多使用云端 AI 服务,而不是本地部署 GPU 集群。2025 年企业 LLM 平均支出达到 $7M/年,接近 2024 年的三倍。多数企业通过超大规模云厂商购买 AI 算力,限制了 Tenstorrent 近期直接芯片采购机会。不过,Hyundai、Samsung、LG 等韩国 OEM 伙伴也是战略投资者,构成高价值企业采用路径。 边缘 / 嵌入式买方(自动驾驶汽车 OEM、IoT 设备制造商、工业自动化厂商)是更长期市场,Tenstorrent 通过 RISC-V 边缘 IP 覆盖。边缘推理市场增长速度与云端推理相近(CAGR 约 19%),也与 Tenstorrent 未来边缘硅片形成产品市场匹配。 RISC-V IP 被授权方是独立客户类型:芯片厂商和系统 OEM 授权 Ascalon RISC-V CPU 核心,并嵌入 SoC。预算控制权在这些组织的半导体采购或 IP 寻源团队,而不是 AI 基础设施团队。 [CM017, CM018, CM019, CM020, CM021, CM022]

细分市场与买方图谱
买方细分主要用户付款方 / 预算负责人工作流 / 使用场景采用触发因素Tenstorrent 路径
超大规模云厂商(AWS、Google、Meta、Microsoft)AI/ML 基础设施工程师CTO / 基础设施 CapEx 预算大规模 LLM 训练、推荐、搜索自研芯片成本优势;性能 / 瓦向 SoC 团队授权 RISC-V IP;短期不是硬件销售
新型云厂商 / GPU 云(CoreWeave、Lambda、Crusoe、ai&、Cirrascale、Turium)AI 推理运维团队、ML 工程师CEO / CapEx 采购预算LLM 推理服务、扩散模型生成、高频交易 AINVIDIA 供应短缺、性价比替代方案主要商业滩头:截至 2026 年 5 月,Galaxy Blackhole 已在 5+ 家新型云厂商投产
企业(金融、医疗、制造)数据科学团队、AI 开发者AI/IT 预算负责人、CFO 审批内部 LLM 部署、AI 助手工具、数据分析降低云端 AI 成本、数据隐私、本地部署合规通过新型云厂商间接触达;企业直销硬件是更长期路径
汽车 OEM(Hyundai、Samsung、LG 战略投资方)自动驾驶、ADAS、车载 AI 团队EE / 硬件采购ADAS 推理、座舱 AI、车队管理TSMC 供应多元化;主权芯片需求;与韩国政府方向一致战略投资方与客户身份对齐;Hyundai/LG 参与 Series D 是转化信号
边缘 / IoT 设备厂商固件工程师、SoC 设计师研发 / 半导体采购端侧推理、语音 AI、智能传感器功耗包络限制;延迟;数据隐私愿景型路径:未来 Tenstorrent 边缘芯片;当前为 RISC-V 边缘 IP
RISC-V IP 授权客户(芯片厂商、Tier-1 车企供应商)SoC 设计团队IP 采购 / EDA 授权预算面向边缘 AI 推理的定制 SoCARM 替代方案;降低版税;中国供应自主Ascalon RISC-V CPU IP 授权;已进入市场

预算归属和采用触发因素来自分析师报告和行业报道;除新型云厂商部署外,Tenstorrent 硬件的具体企业买方承诺未公开披露。

[CM017, CM018, CM019, CM020, CM021, CM022]
FM003: 买方分群地图——AI 芯片市场

序数评分由分析师根据公开报道推导,并非基于 Tenstorrent 内部销售数据。「Tenstorrent 路径」列反映基于已确认部署和投资者关系的近期商业可行性。

2.4 增长驱动与采用约束

AI 加速器市场的增长驱动强,证据也充分。生成式 AI 推理爆发是主导驱动:到 2026 年,约三分之二 AI 算力将由推理驱动,持续制造对推理优化硅片的大规模需求。受 TSMC CoWoS 先进封装瓶颈和 HBM3e 供应限制影响,NVIDIA GPU 供给至少到 2027 年都严重受限,由此形成结构性替代需求。地缘政治下的算力主权要求——尤其在韩国(Samsung、LG、Hyundai)、日本(SoftBank 伙伴 ai& 是 Tenstorrent 最大 Galaxy Blackhole 部署方)和欧盟——让买方更愿意考虑非美国芯片选项。数据中心功率密度触顶后,能效成为越来越关键的维度;Tenstorrent 声称其每瓦性能具备竞争力。 Tenstorrent 自身的采用约束包括:CUDA 生态锁定效应(首要结构性约束——从 NVIDIA CUDA/cuDNN 库迁移到任何替代方案,都需要大量软件重写);TT-Metal/TT-Metalium 相对生产级 CUDA 的软件不成熟(The Register 于 2025 年 11 月给出反向证据);半导体开发资本强度高(芯片需要多年开发周期,以及数亿美元 TSMC/Samsung 无晶圆厂成本);HBM 供应也是所有芯片厂商共同面对的瓶颈。 RISC-V 授权的约束来自 ARM IP 既有巨头:Arm Holdings 凭更强的软件生态和更成熟的工具链主导嵌入式 CPU IP 授权。尽管生态快速改善,基于 RISC-V 的替代方案(SiFive、Tenstorrent Ascalon)仍存在工具链缺口。 [CM026, CM027, CM028, CM029, CM030, CM031]

增长驱动因素与采用约束表
因素方向时间对 Tenstorrent 的影响尽调问题
生成式 AI 推理爆发——到 2026 年,全部 AI 算力的 2/3 用于推理顺风现在 / 持续Tenstorrent 的 Galaxy Blackhole 面向推理;市场正向 Tenstorrent 核心产品扩张验证客户层面的推理定价相对 NVIDIA 总拥有成本(TCO)是否成立
NVIDIA GPU 供应短缺延续至 2027 年(TSMC CoWoS 和 HBM3e 瓶颈)顺风2025–2027推动买方急于评估替代方案;加速 Tenstorrent 的新型云厂商管线评估供应缺口是结构性问题,还是会在 Tenstorrent 放量前恢复常态
主权 / 地缘政治算力需求(韩国、日本、欧盟、印度)顺风2025–2028Samsung、Hyundai、LG 等战略投资方与韩国算力主权方向一致;通过 ai& 在日本部署是旗舰案例识别与韩国 / 日本合作伙伴绑定的政府补贴或采购合同
CUDA 生态锁定——现有 NVIDIA 工作负载切换成本高逆风结构性企业和超大规模云厂商采用的主要障碍;TT-Forge 90% HuggingFace 模型通过率试图降低门槛独立测试 TT-Forge 模型兼容性;用边缘案例验证通过率
TT-Metal 相比 CUDA 软件不成熟(The Register 2025 年 11 月反向证实)逆风2025–2027 年,直至解决限制 AI 爱好者、SMB 和缺少 MLIR / 系统经验的企业实验室采用评估 2026 年软件路线图里程碑;向早期采用者索取开发者 NPS 数据
数据中心电力 / 电网基础设施瓶颈逆风(共同承受)2025–2030Tenstorrent 声称有能效优势;必须在规模化场景证明,才能赢下受数据中心容量约束的买方索取 Galaxy Blackhole 部署的实测 PUE 和性能 / 瓦数据
RISC-V 生态成熟——RVA23 配置规范、工具链改善、中国采用顺风2025–2030RISC-V 生态扩张利好 Tenstorrent 的 RISC-V IP 授权;Ascalon 定位为 ARM 替代方案按客户类型和预期版税收入评估 Ascalon 授权管线
新型云厂商市场扩张(2026 年 $20B,2030 年增至 $180B)顺风2026–2030新型云厂商是 Tenstorrent 近期商业渠道;市场增长很快跟踪新型云厂商整合风险;评估 CoreWeave IPO 动态是否造成定价压力

时间判断基于分析师报道和产品可用性的定性估计。顺风 / 逆风反映主要影响;对影响混合的因素(如电力)按近期主导影响归类。

[CM026, CM027, CM028, CM029, CM030, CM031]
FM004: AI 芯片采用漏斗 — Tenstorrent 市场路径

阶段数值结合分析师 TAM/SAM 估计、Futurum Group 部署数据以及 Tenstorrent 披露的合同指标。数字均为近似值;私营公司的具体商业条款不可得。

2.5 规模缺口、矛盾估算与尽调问题

本章最大的单一分析缺口,是缺少可辩护的 Tenstorrent 可获取市场(SOM)。已披露的 $150M 已签合同(2024 年 12 月)是唯一公开数据点。作为私营公司,Tenstorrent 没有公开收入和市场渗透率数字,因此无法自下而上构建 SOM。 分析师对 2026 年 AI 芯片市场 TAM 的估计会随口径相差 2–3×:保守的仅加速器定义给出约 $200–270B;更宽定义(包括 AI 内存和数据中心基础设施)给出 $477B–$500B。按照 McKinsey 对半导体行业的方法论,即使最高估计也可能少算自用产能和中国供应商收入,低估幅度最高可达 30–40%。 推理 / 训练拆分同样存在争议:多数预测认为,到 2026 年推理将超过算力的三分之二,但精确的变现收入拆分取决于云服务商未公开披露的定价结构。neocloud 利润率和硬件定价仍不透明,使得 neocloud 专属芯片采购量难以验证。 对 Tenstorrent,应提出两个具体市场规模尽调问题:(1)$150M 已签合同组合按客户类型、细分市场和产品代际拆分;(2)截至 2026 年尽调日,RISC-V Ascalon 授权是否已经产生实质收入——Tenstorrent 的 IP 授权业务具有战略意义,但其对 $150M 数字的贡献未披露。 [CM035, CM036, CM037, CM038]

Chapter 03

03竞争格局

3.1 竞争格局概览

2026 年 AI 加速器市场由 NVIDIA 主导,按收入约占 80% 市占率;其护城河来自沉淀 20 年的 CUDA 生态,平台上有超过 400 万注册开发者和 40,000+ 依赖组织。竞争格局分为五类:(1)既有 GPU 领导者(NVIDIA H100/B200、AMD MI300X);(2)直接 AI 芯片挑战者(Cerebras CS-3、Groq LPU、SambaNova SN40L);(3)不对外商业销售的超大规模云厂商定制硅片(Google Trillium/TPU v6、Amazon Trainium3、Meta MTIA);(4)相邻硬件挑战者(Intel Gaudi 3);(5)现状替代品,即仅 CPU 推理和租用云 GPU。Tenstorrent 作为基于 RISC-V 的开源替代方案进入市场,目标客户是主权 AI 项目、研究机构和对成本敏感的企业推理买方。推理层已显现早期碎片化:Cerebras 2025 年收入达到 $510M,并于 2026 年 4 月提交 IPO;SambaNova 曾接近出售,之后拿到 $350M+ Series E;Intel 基本已把 AI 训练竞争让给 NVIDIA。对 Tenstorrent 这个资金充足但软件仍在成熟的挑战者而言,这些动态同时带来机会和风险。

竞争对手画像表
竞争对手类别规模 / 融资目标客群关键差异点战略方向
NVIDIA在位 GPU$3.4T 市值;H100/H200/B200/B300 产品线企业训练 + 推理;超大规模云厂商CUDA 生态(4M+ 开发者);约 80% 市场份额Blackwell 扩张;NVLink 横向扩展;软件垂直化
AMD Instinct MI300X在位 GPU(替代)约 $200B 市值;MI350X 路线图企业推理;成本敏感型 GPU 买方192GB HBM3 VRAM;ROCm 开源;相对 H100 折价约 30–50%投入 ROCm;MI350X / MI400X 路线图
Intel Gaudi 3相邻在位者约 $100B 市值;AI 收入很小价格敏感型企业;云多元化比 H100 便宜约 50%;OneAPI 开放框架退出训练市场;聚焦性价比推理细分
Cerebras CS-3直接挑战者(推理)2026 年 2 月估值 $23B;$1B Series H;2025 年收入 $510M大模型规模化推理;研究;云晶圆级引擎;LLM 推理 1,000–2,000 tokens/sec已递交 IPO 文件(CBRS,Nasdaq);OpenAI 锚定客户;推理云
Groq LPU直接挑战者(推理)据报道 2025 年底曾与 NVIDIA 洽谈收购实时 LLM 推理;API 优先客户确定性延迟;Llama-70B 300+ tokens/sec收购后走向不确定
SambaNova SN40L直接挑战者$350M+ Series E(2026 年 2 月);BlackRock 标记估值约 $2.4B,较 $5B 峰值回落企业数据中心;托管推理RDU 数据流;24TB DRAM SambaRack;交钥匙系统转向云 / 托管推理;Intel 合作
Google 自研 Trillium(TPU v6)超大规模云厂商自研Google($2T 市值);2026 年 Q1 已部署 100K+ 芯片GCP 内部工作负载;精选云客户约 926 TFLOPS BF16;TPU v5e 的 4.7 倍;垂直整合TPU v7 Ironwood 开发中;扩张 GCP
Amazon Trainium3超大规模云厂商自研Amazon($2T 市值);500K+ 芯片已投产AWS 内部;Anthropic、OpenAI 锚定客户2.52 PFLOPS FP8;NeuronSwitch;UltraCluster 规模Trainium4 开发中;降低 NVIDIA 依赖
Meta MTIA 300超大规模云厂商自研Meta($1.3T 市值);MTIA 300 于 2026 年 Q1 投产Meta 内部:排序、推荐、生成式 AIRISC-V 小芯片;两年四代路线图仅内部使用;不商业销售

规模和融资数据截至 2026 年 5 月。GPU 定价代表当前市场价估计;实际企业合同会不同。Google、Amazon 和 Meta 芯片仅供内部使用,不向第三方买方商业销售。

[CP001, CP004, CP005, CP006, CP007, CP008]
FP001: 竞争定位图

主要 AI 芯片供应商的二维定位:软件生态成熟度(X 轴:0=初生,10=开放 / 成熟)对推理规模与吞吐能力(Y 轴:0=低,10=最高)。 Tenstorrent 在开放性上得分高,但软件成熟度落后;Cerebras 和 NVIDIA 分别靠不同架构路线领跑推理吞吐。

[CP001, CP002, CP003, CP010, CP020, CP021]

3.2 竞争对手画像

NVIDIA H100 GPU(单价 $27K–$40K)和 Blackwell B200/B300($30K–$50K)以 CUDA 软件栈为锚,是主要竞争威胁。AMD Instinct MI300X 每块 GPU 约 $15K–$20K,提供 192GB HBM3 VRAM,并配套逐步成熟的 ROCm 开源栈,因此成为领先的性价比 GPU 替代方案;其云租赁价格 $0.50–$7.86/hr,明显低于 NVIDIA。Intel Gaudi 3 面向价格敏感型企业推理(比 H100 约便宜 50%),但商业牵引很弱,并已公开从训练市场转向。Cerebras CS-3 晶圆级引擎在 LLM 推理上达到 1,000–2,000 tokens/sec,对大批量工作负载而言,吞吐量比 GPU 集群高一个数量级;公司以 OpenAI 为锚定客户,在 2026 年 2 月以 $23B 估值完成 $1B Series H。Groq LPU 为 Llama-70B 推理提供 300+ tokens/sec 和确定性延迟;据报道,Groq 在 2025 年末曾参与 NVIDIA 收购讨论,其独立路线图因此存在不确定性。SambaNova SN40L 可重构数据流架构面向企业交钥匙推理;公司 2025 年末探索出售,同时 BlackRock 将其股份估值从 $5B 峰值下调至约 $2.4B,随后在 2026 年 2 月完成由 Vista Equity 和 Intel 共同领投的 $350M+ Series E。Google Trillium(TPU v6)提供约 926 TFLOPS BF16,算力较 TPU v5e 提升 4.7×,已部署 100,000+ 芯片;Amazon Trainium3 提供 2.52 PFLOPS FP8,已量产 500,000+;两者都以内用为主,对外可用性有限。

功能 / 能力矩阵
能力Tenstorrent BlackholeNVIDIA H100/B200AMD MI300XIntel Gaudi 3Cerebras CS-3
训练支持部分(起步中)部分
推理支持
开源软件栈强(MIT 许可证)否(CUDA 闭源)部分(ROCm 开放)部分(OneAPI)
云租赁可用性部分(Cirrascale/Turium)强(所有主要云)强(主要云)部分(供应商有限)部分(Cerebras 云)
软件生态成熟度部分(2026 年成熟中)强(20 年 CUDA)部分(ROCm 进展中)部分(ISV 有限)部分(仅托管)
开放 / 可编程 ISA强(RISC-V)否(专有)否(专有)否(专有)否(专有)
以太网横向扩展强(无需 InfiniBand)否(NVLink/InfiniBand)部分部分

评级截至 2026 年 5 月量产状态。「强」= 生产级;「部分」= 已可用但有明显限制;「否」= 不可用。软件成熟度是基于独立开发者采用证据和公开评测的定性判断。

[CP006, CP008, CP009, CP010, CP020, CP021]
FP002: 功能广度 / 能力图

五家 AI 加速器供应商在七项采购标准上的能力对比。Tenstorrent 是唯一把 MIT 开源软件、RISC-V ISA 和 Ethernet 横向扩展合在一起的商业供应商, 但软件成熟度和云端可用性落后于 NVIDIA。Cerebras 只做推理;Tenstorrent 和 NVIDIA 同时支持训练与推理。

[CP002, CP006, CP008, CP009, CP010, CP020]

3.3 能力与定价对比

Tenstorrent 的 Galaxy Blackhole 于 2026 年 5 月进入量产。开发者参考卡估计约 $1K,单价约为 H100 GPU 的三十分之一;开源 TT-Metal(MIT 许可)也消除了软件授权费。不过,服务器级定价、机架级配置和云租赁费率尚未在规模化场景下建立。NVIDIA 云 GPU 租赁价格从 $2.00–$14.90/hr(H100)到 $2.25–$14.24/hr(B200);AMD MI300X 为 $0.50–$7.86/hr。截至 2026 年 5 月,Galaxy Blackhole 与 H100/B200 的独立 LLM 基准有限。Tenstorrent 声称 TT-Metal 上 HuggingFace 模型通过率约 90%;该数字为自报,未获独立验证。The Register(2025 年 11 月)评测 Tenstorrent Blackhole QuietBox 工作站,认为软件「对多数本地 AI 爱好者来说根本不够打磨」,这是具体的成熟度指标。Cerebras 和 Groq 都是仅推理架构,不具备训练能力;Tenstorrent 的 Galaxy Blackhole 硬件同时支持训练和推理,工作负载覆盖面比单一用途推理挑战者更宽。Intel Gaudi 3 同样面对软件成熟度挑战,但拥有 Intel 既有分销渠道。

定价 / 打包方式对比
供应商 / 产品采购价格(估计)云租赁($/hr 每单元)定价模式相对 H100 基准成本
NVIDIA H100 80GB$27K–$40K / GPU$2.00–$14.90/hr硬件 + 闭源 CUDA 授权;云 API基准(100%)
NVIDIA Blackwell B200 192GB$30K–$50K / GPU$2.25–$14.24/hr硬件 + DGX 套装;云端优先配额溢价约 +20–64%;推理吞吐约 5x
AMD MI300X 192GB~$15K–$20K / GPU$0.50–$7.86/hr硬件 + ROCm 开源;云 API较 H100 低约 30–50%
Intel Gaudi 3较 H100 低约 50%(估计)有限;未广泛挂牌硬件;云合作伙伴可用性有限定价激进;生态支持弱
Cerebras CS-3 / 云本地部署:未披露Cerebras 云:未公开挂牌云 API + 企业本地部署合同大批量按 token 计价有竞争力;按 GPU 不适用
Tenstorrent Galaxy Blackhole开发卡约 $1K(估计);服务器 TBD尚未规模化可用(2026 年 5 月)硬件 + 开源 TT-Metal(无软件授权费)单元价格约便宜 30x;云端 TFLOP/$ TBD

GPU 采购价格是来自聚合定价数据库和分析师报告的 2026 年 Q1–Q2 市场估计。云租赁价格会因供应商、预留方式和地区而不同。Tenstorrent Blackhole 开发卡约 $1K 属估计;服务器 / 机架定价尚未公开确认。截至 2026 年 5 月,Tenstorrent 云租赁尚未规模化公开可用。

[CP006, CP007, CP008, CP009, CP010, CP022]
FP003: 护城河 / 就绪度 KPI

截至 2026 年 5 月的竞争定位和市场就绪度关键量化指标。图中体现了与 NVIDIA 的规模差距(4M+ CUDA 开发者,对比 Tenstorrent 仍在形成的生态), 也展示挑战者融资格局(Cerebras $23B,对比 Tenstorrent $2.6B)。

[CP001, CP010, CP011, CP012, CP019, CP022]

3.4 切换成本与锁定动态

NVIDIA 领先 20 年的 CUDA,构成 AI 硬件中最深的切换成本护城河。已经嵌入 CUDA 的企业组织若迁移到替代方案,需要承担库重新编译、模型重调、软件认证和开发者再培训成本。400 万+ 注册开发者和 40,000+ 依赖组织让锁定效应自我强化:开发者越多,库越多;切换负担越重,硬件采购越向 NVIDIA 倾斜。Tenstorrent 的 RISC-V ISA 是开放且客户可编程的,TT-Metal/TT-Forge(基于 MLIR)提供编译器层硬件抽象。借助框架抽象层(PyTorch、JAX),跨 AI 芯片厂商多供应商部署在技术上可行,但运营成本高;多数企业在生产环境中会标准化到单一供应商。Tenstorrent 基于 Ethernet 的横向扩展,避开大型 NVIDIA GPU 集群所需的 InfiniBand 网络依赖,降低网络层基础设施锁定。主权 AI 项目——由日本(ai& Tokyo 的 Tenstorrent 部署)、韩国(Hyundai 投资)和中东的政府及战略企业推动——代表一个有结构性动机的非 NVIDIA 买方群体,主动寻求供应链独立。针对特定地理市场的 NVIDIA 芯片出口管制进一步放大这一机会。

护城河耐久性 / 竞争风险登记表
护城河主张威胁向量严重性缓释措施 / 尽调问题
开源 TT-Metal 区分于闭源 CUDACUDA 20 年生态;PyTorch/HuggingFace 原生支持 CUDA;4M+ 开发者惯性加速 TT-Forge MLIR 成熟;建立 ISV 合作;把 HuggingFace 覆盖率提升到自报 90% 以上
RISC-V 可编程性支持定制 AI 工作负载ARM 架构设计扩散;Meta MTIA 也使用 RISC-V;仅靠 ISA 不能形成差异化发布 RISC-V 可编程性基准;围绕特定领域工作负载建设工具链生态
以太网横向扩展去掉 InfiniBand 依赖NVIDIA NVLink/NVSwitch 在改进;InfiniBand 成本下降;AMD 也支持以太网发布多机架 Galaxy Blackhole 横向扩展基准;证明 64+ 节点规模下效率 >80%
主权 AI 需求形成非 NVIDIA 买方细分NVIDIA 可借中间方绕开出口限制;AMD 正拿到主权 AI 订单;Tenstorrent 依赖 TSMC深化日本 / 韩国 / 中东政府合作;建立 TSMC 供应承诺
相对 GPU 替代方案,每 TFLOP 资本成本更低NVIDIA Blackwell 在规模化场景缩小每 token 成本差距;AMD MI300X 凭生态成熟度压低单元价格发布经验证的第三方基准;围绕标准 LLM 任务做独立性价比分析
Jim Keller 设计履历吸引顶尖工程人才NVIDIA、AMD、超大规模云厂商挖人;Keller 此前在多家公司任期较短验证 Series D 后留存指标;评估创始团队之外的技术梯队深度

严重性从 Tenstorrent 视角评估:高 = 2–3 年内威胁竞争差异化根基;中 = 实质性逆风;低 = 可通过执行管理。护城河主张反映 Tenstorrent 声称的优势;威胁向量是每项主张耐久性的主要风险。

[CP019, CP020, CP025, CP026, CP028, CP029]

3.5 护城河耐久性与反向证据

NVIDIA 的 CUDA 护城河呈现结构性耐久:Blackwell B200/B300 代际扩大训练算力性能差距,软件连续性则把既有工作负载继续锁住。推理层正受到 Cerebras、Groq 和 SambaNova 更激烈挑战,但各自服务的工作负载画像更窄。Intel 从训练竞争中战略性后退,降低了该细分市场的既有巨头压力。Tenstorrent 的关键反向证据包括:(1)The Register(2025 年 11 月)记录 TT-Metal「对多数本地 AI 爱好者来说根本不够打磨」,显示其相对 CUDA 存在具体成熟度缺口;(2)Tenstorrent 90% HuggingFace 通过率为自报,缺少独立验证;(3)SambaNova 估值从 $5B 下调至约 $2.4B,说明资金充足的 AI 芯片挑战者仍面临严峻执行风险;(4)Cerebras 与 OpenAI 的锚定关系($10B+ 合同)说明挑战者细分市场存在客户集中脆弱性。Tenstorrent 竞争护城河的耐久度,关键取决于能否在 CUDA 生态势能把领先延伸到 2027–2028 年之前,交付生产级软件工具链对等能力,并扩展 ISV 伙伴关系。

3.6 附录

Chapter 04

04财务情况

4.1 收入架构与商业模式

Tenstorrent 的收入模型由多条收入流组成:硬件产品销售、知识产权授权和云访问服务。不过作为私营公司,它不披露经审计财务。截至 2026 年 5 月,主要收入驱动是直接硬件销售,尤其是 Galaxy Blackhole AI 服务器系统及相关推理加速卡;这些产品在 2026 年 4–5 月进入量产并达到 GA。Galaxy Blackhole 服务器机箱标价为每台 $110,000(32 颗 Blackhole ASIC、23 PFLOPS FP8),四机箱 Supercluster 定价 $440,000。企业和超大规模云客户预计会按量采购;折扣结构未公开披露。在 Galaxy 层级之下,Tenstorrent 销售约 $999 的 Blackhole P100 入门推理卡,以及约 $9,999 的 QuietBox 工作站。第二条收入流是公司 RISC-V CPU 核心和 Tensix NPU 架构的 IP 授权,来自 Samsung、Hyundai 等汽车和边缘 OEM 伙伴的权利金收入;但授权经济性完全未披露。第三条新兴收入流是通过 Koyeb 无服务器平台交付的云端硬件即服务,按用量提供 Tenstorrent 计算实例。DevCloud 开发者计划向开发者免费或按免费增值模式提供 Wormhole 硬件访问,未来可能转化为付费企业 DevCloud 订阅。专业服务和集成支持目前是较小的第四条收入流。

收入来源表
收入流机制单位 / 指标当前状态 / 数值收入质量关键尽调问题
硬件销售(Galaxy)直接销售 Galaxy Blackhole 服务器机箱和 Superclusters 集群按单元($110K–$440K)2026 年 4 月起正式可用(GA);已签合同形成积压订单(约 $150M)中——硬件收入在交付时确认;仅支持服务具备经常性索取单元出货积压、ASP 趋势和按季度确认收入
硬件销售(Edge)销售 Blackhole P100 推理卡和 QuietBox 工作站按单元($999–$9,999)已可用;销量未知低——零售 / 开发者渠道未量化;未披露出货单元索取上市以来销量、渠道结构和售出率
IP 授权(RISC-V/Tensix)来自 Samsung、Hyundai 和其他 OEM SoC 合作伙伴的单芯片版税按芯片或按期间收取版税已生效;经济条款未披露若持续产生收入,质量较高——版税不占用资本,毛利更厚要求披露单台版税、承诺出货量、被许可方数量和已确认收入
云 / HaaS(Koyeb)通过 Koyeb 无服务器平台按用量访问云服务按计算小时计费合作已上线;收入分成未披露可能具备高质量经常性收入,但用量未知核实收入分成协议条款、活跃实例小时和 Tenstorrent 净收入
DevCloud 订阅免费 / 免费增值开发者访问,可能转化为付费企业层按用户 / 组织按月计费早期阶段;转化率未知当前质量较低——更像战略漏斗资产,尚不是实质收入线要求披露活跃开发者数、付费转化率和如有 ARR
专业服务集成支持、优化和部署服务按项目计费占比较小;未单独披露质量较低——服务毛利低于硬件 / IP核实服务是单独销售还是打包销售,以及对应毛利率

除 Galaxy 标价外,所有数值均为估计或未披露。签约合同到出货之间,收入确认可能滞后 1–6 个月。

[CI001, CI002, CI003, CI004, CI005, CI006]
FI001: 收入模型桥
[CI001, CI002, CI003, CI004, CI005]

4.2 定价结构与 GTM 动作

Tenstorrent 在定价上的核心商业差异,不是原始标价,而是推理总成本效率。公司称,Galaxy Blackhole 系统在 DeepSeek-R1 671B 工作负载上每百万推理 token 成本为 $6,而 NVIDIA GB300 GPU 系统约为每百万 token $30,即每输出 token 成本优势 5×。Galaxy 单服务器机箱 $110,000,瞄准的是持续吞吐和内存容量比峰值 FLOPS 更重要的企业推理部署。GTM 动作以企业直销为主,渠道活动有限。早期部署包括 Tokyo 的 ai& Corporation、Cirrascale Cloud Services 和 Turium AI,显示其早期客户策略更像先播种云服务商,而非广泛直销 Fortune 500。企业 AI 基础设施销售周期通常为 3–12 个月,因此约 $150M 已签合同数字(2024 年 12 月 Series D 完成时披露)更像管线积压订单,而非已确认收入。硬件按交付时点确认收入;若提供多年支持合同,则将按期确认。IP 授权权利金按每芯片或每期间确认。GTM 效率指标——获客成本(CAC)、回本周期和渠道经济性——未公开披露,是重要尽调缺口。MIT 许可下的 TT-Metal 和 TT-Forge 开源定位,充当开发者获客漏斗;但能否把开发者心智转化为企业硬件收入,规模化上仍未验证。

定价 / 变现表
产品 / 服务标价定价基础折扣 / 未知项来源
Galaxy Blackhole Server$110,000按 6U 机箱计费(32 颗 ASIC,23 PFLOPS FP8)批量折扣结构未披露;企业交易预计低于标价公司(tenstorrent.com/hardware/galaxy)
Galaxy Supercluster(4×)$440,000按 4 机箱集群计费(92 PFLOPS FP8)与单机箱 × 4 相同;未见集群折扣报道公司(tenstorrent.com/hardware/galaxy)
Blackhole P100 Inference Card~$999按卡计费(入门级推理加速卡)起售价;更高 SKU 版本未披露公司(tenstorrent.com)
QuietBox Workstation~$9,999按桌面工作站单元计费起售价;配置选项未披露公司(tenstorrent.com)
RISC-V / Tensix IP 许可基于版税(未披露)按出货芯片或按期间计费无公开价目表;条款在 NDA 协议中根据 Samsung / Hyundai 合作公告推断
云 / HaaS(Koyeb)按用量计费(未披露)按计算小时 / 秒计费Koyeb 设定终端客户价格;Tenstorrent 收入分成未知Koyeb 博客(koyeb.com)

所有价格均为 2026 年 4–5 月公告时的标价。公开渠道没有实际成交价或平均折扣信息。

[CI007, CI008, CI009, CI010, CI011]
FI002: 单位经济模型桥
[CI012, CI013, CI014, CI015, CI035]

4.3 单位经济性与成本结构

作为无晶圆厂半导体公司,Tenstorrent 的单位经济性由 TSMC 晶圆成本(Blackhole 使用 N4 节点制造)、封装、测试和运输驱动,再由平均售价(ASP)和出货量抵消。Galaxy 服务器机箱单价 $110,000;若每台服务器销售成本(COGS)估计为 $50,000–$70,000(量产下 32 颗 Blackhole ASIC 的 TSMC 晶圆成本,加上 GDDR6 内存、PCB、电源、组装),隐含单机箱毛利率为 36%–55%,与规模化无晶圆厂芯片公司常见的 40%–55% 区间一致。这些数字完全是估计;Tenstorrent 不披露 COGS、毛利率或单台经济性。Blackhole N4 流片的一次性工程(NRE)成本估计为 $50M–$100M 或更高,并随生产量摊销。在较低出货量(数千台服务器)下,NRE 摊销会显著压缩毛利率。2026 年初约 1,100–1,200 名员工,按硬件和软件工程师市场化全包薪酬估算,意味着年度人员成本约 $220M–$300M。叠加 TSMC 生产成本、工具、EDA 授权、设施和 G&A 开销,总运营支出估计为每月 $25M–$50M。在 Galaxy Blackhole 硬件收入实质放量前,经营亏损几乎确定。公开数据无法量化企业硬件获客成本。通向盈利的毛利率路径,取决于放量爬坡、收入组合向更高毛利 IP 授权迁移,以及软件赋能服务收入。

单位经济模型表
指标估计值置信度重要性尽调问题
毛利率(Galaxy 硬件)估计 ~36%–55%决定资本效率,以及硬件收入能否支撑运营要求披露经审计的单台服务器 COGS 拆分和混合硬件毛利率
平均售价(Galaxy Server)$110,000(标价)主要的单台收入驱动项;批量折扣未知核实实际 ASP 与标价的差距,以及企业折扣水平
估计 Galaxy Server 单台 COGS~$50,000–$70,000极低源自 TSMC N4 晶圆成本 + GDDR6 + 组装的推算;未经验证要求从制造成本审计中取得实际 COGS
NRE 摊销(Blackhole)估计总计 ~$50M–$100M+极低大额 NRE 压进首批量产,会压低实际毛利率要求披露流片合同金额和摊销计划
获客成本UnknownNoneGTM 效率的关键;企业硬件 CAC 通常为 $50K–$500K+要求按客户分层披露获客成本和销售周期长度
IP 许可毛利估计 ~80%–95%(若为版税)版税收入毛利率接近 100%,会显著抬高混合毛利率要求披露版税率、每个被许可方承诺最低额和 LTM 已确认版税
每推理 Token 成本(Galaxy)每 1M tokens $6(公司声称)与 NVIDIA 竞争定位的核心主张;必须独立验证委托独立基准测试,核实 token 成本方法和工作负载

多数数值来自公开基准和半导体行业常模推算。没有任何数值来自 Tenstorrent 已披露财务数据。

[CI012, CI013, CI014, CI015, CI016, CI017]
FI003: 财务估计区间
[CI031, CI032, CI033, CI034, CI035]

4.4 资本结构、现金跑道与融资依赖

Tenstorrent 多轮累计融资约 $1.99B。2024 年 12 月,公司完成 $693M Series D,投后估值 $2.6B,由 Samsung Securities 和 LG Technology Ventures 领投。2025 年 11 月,公司完成约 $800M Series E,投后估值 $3.2B,由 Fidelity Management and Research 领投。此前公司还在 2023 年从 Hyundai Motor Group 和 Samsung 获得 $100M 融资。若按每月 $25M–$50M 经营烧钱速度估算,且收入未披露,Series E 资金从 2025 年 11 月起可提供约 16–32 个月现金跑道——大体足以支撑公司跑到有意义的 Galaxy 硬件收入,但留给延误的缓冲有限。Series E 资金计划用于产能扩张、下一代芯片(Blackhole 之后)设计与流片,以及 GTM 扩张。任何重大产品延迟、客户集中挫折或 TSMC 产能约束都会加速消耗现金跑道。公开信息未显示公司有债务融资或项目融资义务。公司也未说明任何盈利目标日期。考虑到半导体业务资本强度,如果 Galaxy 硬件收入不能快速放量,Tenstorrent 可能需要在 Series E 完成后 24–36 个月内补充资金(Series F 或基于收入的融资)。收入模型是否足够、毛利率轨迹能否跑通,是任何后续投资的主要尽调关口。

资本充足性表
项目数值 / 估计置信度备注
累计融资~$1.99 billionSeries D($693M,2024 年 12 月)+ Series E($800M,2025 年 11 月)+ 此前轮次(~$497M)
Series E 投后估值$3.2 billion(2025 年 11 月)Fidelity Management and Research 领投;Tracxn 和 PMInsights 交叉印证
估计账面现金(2026 年 5 月)~$1.0B–$1.5B按 Series E 交割后 6 个月烧钱推算剩余现金;实际余额未披露
估计月度烧钱$25M–$50M极低按约 1,200 名员工、全负担成本 $200K+,叠加 TSMC 量产和资本开支推算;未经验证
估计现金跑道(自 2025 年 11 月起)16–32 个月(即 2027 年 3 月 – 2028 年 7 月)极低区间很宽,因为烧钱速度不确定;假设 Galaxy 收入没有贡献
Series E 资金计划用途量产爬坡、下一代芯片研发、GTM 扩张、IP 许可增长根据 Series D 新闻稿和公司战略推断;没有 Series E 专项拆分
债务 / 项目融资未公开披露未在公开文件或新闻中发现债务额度或项目融资协议证据

现金头寸和烧钱速度完全基于估计。任何显著客户爬坡或 TSMC 流片事件,都会实质改变现金跑道估算。

[CI019, CI020, CI021, CI022, CI023, CI024]
FI004: 资本强度 / 现金流图
[CI019, CI020, CI021, CI022, CI023, CI036]

4.5 财务结论与尽调阻塞点

Tenstorrent 的财务画像是高风险、高上行。正面看,公司资金充足(累计融资约 $1.99B,Series E 资金从 2025 年 11 月起估计提供 16–32 个月现金跑道),已公开发布定价具备竞争力的 AI 推理服务器系统(Galaxy Blackhole,$110,000/台),并声称相较 NVIDIA 旗舰 GB300 集群有 5× 每 token 成本优势。反面看,公司不披露经审计收入、毛利率、销售成本(COGS)或烧钱速度。第三方收入估计(Latka 模型,2025 年 $501.6M)是算法猜测,不是经验证数字,也被视为不足以支持投资决策。约 $150M 已签合同数字(截至 2024 年 12 月)仅为累计融资额的约十三分之一。The Register 2026 年 4 月对 Galaxy 发布的评测质疑,软件和生态成熟度是否足以推动企业快速采用。关键尽调阻塞点:(1)未披露财务——收入、毛利率、现金位置;(2)没有客户集中度数据——单一客户取消就可能构成重大影响;(3)IP 授权经济性不透明——权利金率和量承诺未披露;(4)烧钱速度和现金跑道只能靠估计验证;(5)已签合同积压订单的收入确认时间线不清晰。在这些缺口通过经审计数据室解决前,财务尽调并不完整,本财务章节应被视为风险分层文件,而不是已验证财务模型。

公开财务缺口表
缺失指标对尽调的影响具体尽调路径
年收入(FY2024、FY2025)高——无法验证增长轨迹;算法估算(Latka 的 $501.6M)不可靠要求披露 FY2024 和 FY2025 经审计利润表,以及 2026 年 Q1 管理账
毛利率和 COGS 拆分高——没有毛利率数据,就无法建模单位经济模型是否可持续,也无法判断盈利路径要求披露过去两个财年按板块(硬件、许可、服务)拆分的经审计 COGS
烧钱速度和现金头寸高——现金跑道估计跨度达 2×;无法量化契约风险或现金耗尽风险要求披露经审计现金流量表和当前银行余额确认函
客户数和收入集中度高——装机基础小,单一客户流失可能构成重大影响要求按客户分层披露客户数量、前三大客户收入占比和任何客户流失
IP 许可版税表中——经常性版税基础未知;许可收入可能占收入的 10%–40%要求披露版税协议摘要、承诺出货量和各被许可方已确认版税
已签合同与已确认收入缺口中——$150M 已签合同是出货前数字;收入落地时间表不清楚要求提供所有已签合同的合同到出货管线和预计确认日期

五项顶层财务指标均由私营公司持有。任何正式尽调都应把这张表作为核心资料室请求清单。

[CI025, CI026, CI027, CI028, CI029, CI030]

4.6 附录

Chapter 05

05产品与技术

5.1 产品组合与 SKU 架构

Tenstorrent 以 Blackhole ASIC 为锚,提供垂直整合产品线,覆盖三种形态,分别面向不同推理市场细分。入门层级的 Blackhole p100a 推理卡约 $999,面向开发者和桌面推理爱好者;它提供 28 GB GDDR6 内存、448 GB/s 带宽,并通过 PCIe Gen5 x16 连接主机。进阶 SKU p150a 增加 4 GB GDDR6(32 GB,512 GB/s),更关键的是加入四个 QSFP-DD 800 Gbps Ethernet 端口,可在无交换机情况下直接卡间联网,打开工作站规模的多卡推理。p150b 变体在计算规格上与 p150a 物理一致,但采用被动散热,面向密度优化的机架部署。卡层级之上是 QuietBox 工作站:一台约 $9,999 的液冷桌面机箱,内置两张 Blackhole 卡,定位为面向全栈开发团队的开发者生产力平台。产品层级顶端是 Galaxy Blackhole,一台 6U 机架服务器,集成 32 颗 Blackhole ASIC,并通过 100 Tbps 板载网状网络互联,提供 23 PFLOPS FP8 和总计 1 TB GDDR6 内存,公开标价为每机箱 $110,000。Galaxy 于 2026 年 4 月 28 日达到 GA,标志着 Tenstorrent 从纯开发者硬件公司转向商业企业 AI 基础设施供应商。硬件之外,Tenstorrent 还提供 TT-Metal 和 TT-Forge 开源软件开发套件,以及 DevCloud 远程访问服务,让开发者无需本地资本开支即可评估 Wormhole 和 Blackhole 硬件。产品组合从 $999 开发者卡到 $110,000 企业服务器,跨度足够大,使 Tenstorrent 能低成本播种开发者社区,再把已验证用例转化为企业收入;这与 NVIDIA 开发者计划过去十年使用的飞轮策略相似。 [CE001, CE005, CE006, CE007, CE008, CE018]

产品模块 / 资产矩阵
产品 / 资产类别目标客群关键规格 / 描述价格(USD)状态(2026 年 5 月)
Blackhole p100a推理加速卡开发者 / 桌面120 个 Tensix 核心、28 GB GDDR6 448 GB/s、PCIe Gen5 ×16、300 W TDP~$999已正式发布
Blackhole p150a推理加速卡工作站 / 边缘集群120 个 Tensix 核心、32 GB GDDR6 512 GB/s、4× QSFP-DD 800 Gbps Ethernet、300 W TDP~$1,999 估计已正式发布
Blackhole p150b推理加速卡服务器 / 机架(无源)120 个 Tensix 核心、32 GB GDDR6 512 GB/s、4× QSFP-DD 800 Gbps Ethernet、无源散热N/A(服务器 SKU)已正式发布
QuietBox Workstation工作站系统开发者生产力2× Blackhole 卡、液冷机箱、可运行 Linux~$9,999已正式发布
Galaxy Blackhole Server6U 机架式 AI 服务器企业推理32 颗 Blackhole ASIC、23 PFLOPS FP8、1 TB GDDR6、100 Tbps 内部 mesh$110,000 标价已正式发布(2026 年 4 月)
TT-Metal / TT-Metalium软件 SDK(运行时)所有开发者低层内核 API + 调度运行时;Apache 2.0 OSS免费 / 开源已正式发布
TT-NN软件 SDK(算子)ML 开发者Python 算子库,200+ 算子,兼容 HuggingFace;Apache 2.0 OSS免费 / 开源已正式发布
TT-Forge编译器栈PyTorch / JAX / ONNX 开发者基于 MLIR 的编译器;TT-Torch、TT-XLA、TT-Forge-ONNX 前端;Apache 2.0 OSS免费 / 开源测试版(v1.0 目标为 2026 年)
DevCloud远程算力服务早期开发者以免费增值方式访问 Wormhole 和 Blackhole 硬件;不需要本地资本开支免费增值活跃

价格为截至 2026 年 5 月的公开标价或估计值;企业价格和批量折扣未披露。p150b 价格未公开列出。

[CE005, CE006, CE007, CE008, CE009, CE010]

5.2 Blackhole ASIC 架构与技术规格

Blackhole ASIC 是 Tenstorrent 第二代 AI 加速器,也是其第一款采用 TSMC 6 nm 工艺节点生产的芯片,单颗裸片面积约 600 mm²。芯片的计算结构围绕 120 个 Tensix 处理 tile 展开;Tensix 是 Tenstorrent 自研的数据流架构,每个 tile 内含 5 个小型 RISC-V 核心,负责内核调度、数据搬运和计算编排。120 个 tile 让单颗 Blackhole 裸片仅底层控制就内置超过 600 个 RISC-V 核心,此外还有 16 个更大的应用级 RISC-V 核心(基于 SiFive X280 64-bit core),用于运行 Linux 和主机侧管理软件。Tenstorrent 称,这套双 RISC-V 策略让 Blackhole 拥有可编程计算底座,不依赖封闭固件微控制器,从而降低延迟,并把软硬件协同设计做得更紧。片上内存合计 180 MB SRAM,分布在 Tensix tile 阵列中,为激活值、权重和中间结果提供高带宽本地存储,避免往返更慢的 GDDR6。单芯片峰值算力为 664 TFLOPS(BlockFP8 格式)或 332 TFLOPS(BF16),定位介于 NVIDIA H100 SXM(3.9 PFLOPS FP8)和低成本边缘推理 SoC 之间。主动散热版 p100a 和 p150a 的热设计功耗为 300 W。p150a 和 p150b 的 3.2 Tbps 以太网总带宽尤其少见:多数竞品推理 ASIC 依赖 PCIe 或 NVLink 做设备间通信,而 Blackhole 原生以太网 mesh 从一开始就是为在通用网络基础设施上集群化设备而设计。Galaxy 配置中,32 颗 Blackhole 芯片以全互联拓扑互联,提供 100 Tbps 芯片间总带宽,让超大模型(例如 DeepSeek-R1 671B)可以在全系统内做权重分片推理,不在互联层形成拖慢性能的瓶颈。 [CE001, CE002, CE003, CE004, CE005, CE006]

技术 / 运营架构表
技术栈层组件技术 / 标准关键规格开源?
硬件——计算Blackhole ASICTSMC 6nm,Tensix 数据流架构每颗芯片 120 个 Tensix tiles、664 TFLOPS BlockFP8、332 TFLOPS BF16否(自研 ASIC)
硬件——片上内存SRAM 缓存阵列分布式 SRAM 跨 Tensix tiles 布局每颗芯片 180 MB 片上 SRAM
硬件——片外内存GDDR6 DRAMSamsung / SK Hynix / Micron 提供的 GDDR6p100a:28 GB 448 GB/s;p150a/b:32 GB 512 GB/s(内存带宽)
硬件——主机接口PCIe 控制器PCIe Gen5 ×16~128 GB/s 主机到设备带宽
硬件——芯片互连以太网互连结构(p150a/b)QSFP-DD 400GbE × 8 lanes = 每端口 800 Gbps × 4 端口每张卡 3.2 Tbps;Galaxy:100 Tbps 内部网格否(PHY)
硬件——控制核心Big RISC-V(SiFive X280)64-bit RISC-V,可运行 Linux每颗芯片 16 核;运行管理软件ISA 开放;SoC 集成自研
硬件——tile 控制Baby RISC-V每个 tile 定制嵌入式 RISC-V每个 Tensix tile 5 核,每颗芯片 600+ 核;处理调度和数据移动ISA 开放;实现为自研
软件——低层TT-LLKC++ / 汇编 Tensix 内核库手工优化的矩阵乘法、归约和激活内核是(Apache 2.0)
软件——运行时TT-MetaliumC++ / Python 调度引擎内核调度、缓冲区管理、多设备编排是(Apache 2.0)
软件 — 算子TT-NNPython + C++ 融合算子库200+ 个算子;兼容 HuggingFace 模型 API是(Apache 2.0)
软件 — 编译器TT-Forge / MLIRMLIR 编译器基础设施 + LLVMPyTorch(TT-Torch)、JAX(TT-XLA)、ONNX(TT-Forge-ONNX)前端是(Apache 2.0)

GDDR6 供应商来自公开供应商披露和硬件拆解报告;Tenstorrent 未披露各供应商的具体占比。

[CE001, CE002, CE003, CE004, CE005, CE006]
FE001: 产品架构图
[CE001, CE011, CE012, CE013, CE014, CE028]

5.3 软件栈与开源生态

Tenstorrent 的软件策略和既有 AI 芯片厂商拉开距离,核心在于其承诺把完整工具链以 Apache 2.0 许可证开源。整个栈分为四个逻辑层。最底层是 TT-LLK(Low-Level Kernels),为常见矩阵运算提供手工优化的 Tensix 指令序列。上一层 TT-Metalium 是核心运行时与调度引擎:它抽象 Blackhole 的多 tile 执行模型,管理内核编译,并提供类似 NVIDIA GPU 上 CUDA 的设备编程模型。TT-NN 位于 TT-Metalium 之上,暴露一个可从 Python 调用的算子库,包含 200 多个计算原语,并兼容 HuggingFace Transformers 风格的模型定义。最上层是 TT-Forge,这是基于 MLIR 的编译器,可从三个前端桥接接收模型图:面向 PyTorch 模型的 TT-Torch、面向 JAX 模型的 TT-XLA,以及面向 ONNX 交换格式的 TT-Forge-ONNX。借助这套架构,Tenstorrent 可以声称兼容主流 ML 框架,而无需应用开发者触碰底层 Tensix 编程。主公开仓库 tenstorrent/tt-metal 在 GitHub 上截至 April 2026 约有 1,410 stars、429 forks 和超过 25,830 commits,覆盖 161 个官方版本。开源活动量不小:988 个未关闭 PR 和 19,076 个已合并 PR 说明工程推进活跃但仍在扩规模,3,488 个未关闭 issues 则显示已知软件缺口仍有明显积压。Tenstorrent 声称 HuggingFace 模型基准测试通过率为 90%,并兼容超过 2.5 million 个开源模型,但两个数字都未获独立验证。The Register 在 November 2025 的一篇评测认为,该软件“对大多数本地 AI 爱好者来说就是不够打磨”,并把配置摩擦和驱动文档不完整列为采用的主要门槛。Pyron SDK 文档(docs.pyron.dev)提供更高层抽象,目标是降低企业集成商的门槛,但成熟度仍处早期。开源许可策略创造了社区飞轮——开发者贡献可补上 Tenstorrent 在模型移植和算子覆盖上的工程产能——但也意味着公司无法阻止竞争者 fork 这套栈。 [CE010, CE011, CE012, CE013, CE014, CE015]

工作流 / 使用场景表
使用场景框架入口工作流步骤主要硬件成熟度已知限制
LLM 推理(单卡)PyTorch / HuggingFace加载模型 → TT-Forge 编译 → TT-Metalium 调度 → token 生成p100a、p150a大于 28 GB 的模型必须分片或量化
LLM 推理(多卡集群)PyTorch / HuggingFace配置 TT-Mesh → 加载分片模型 → 分布式推理p150a/b 集群、Galaxy早期正式发布需要以太网互连结构;社区反馈显示,集群设置文档仍不完整
视觉 / 多模态推理PyTorch / ONNX加载视觉模型 → TT-Forge-ONNX 编译 → Blackhole 推理p100a、p150aONNX 覆盖不完整;并非支持所有视觉算子
JAX 工作负载通过 TT-XLA 使用 JAXJAX 模型 → TT-XLA 桥接层 → 编译 → 在 Blackhole 上执行p150a/b、Galaxy早期TT-XLA 仍在测试版;社区验证有限
研究 / 自定义内核TT-Metal API(Python/C++)编写 Tensix 内核 → 通过 TT-Metalium 调度 → 分析性能结果任一 Blackhole 卡学习曲线陡峭;CUDA 类比并非 1:1
远程评估(DevCloud)通过云端使用 TT-Metal / TT-Forge注册 DevCloud → SSH 到 Wormhole / Blackhole 实例 → 运行工作负载Wormhole、Blackhole(远程)活跃Wormhole 是上一代;2026 年初 Blackhole DevCloud 可用性有限

成熟度评级是基于公开文档、社区 GitHub issues 和独立评测者报告的定性判断。没有 SLA 数据可用。

[CE014, CE016, CE017, CE021, CE039, CE040]
信任 / 质量 / 合规表
领域控制措施 / 标准状态证据 / 备注
软件许可Apache 2.0 开源许可证已确认核心仓库(tt-metal、tt-forge、tt-nn)均在 GitHub 以 Apache 2.0 发布
持续集成tt-metal 上的 GitHub Actions CI/CD已确认161 个版本;公开仓库可见 19,076 个已合并 PR 和自动化门禁
硬件认证(FCC / CE)FCC Part 15 / CE 标志未公开确认美国 / 欧盟商业硬件销售的标准要求;Tenstorrent 未发布合规声明
数据隐私本地部署;无云端数据处理设计上已确认所有推理都在客户硬件上执行;未见向 Tenstorrent 回传数据遥测的记录
模型兼容性测试HuggingFace 通过率 90%(自报)公司声称;未验证无独立基准测试或第三方审计;3,488 个开放 GitHub issue 显示仍有已知缺口
MLPerf / 标准化基准MLPerf Inference 提交未提交(截至 May 2026)截至运行日,Tenstorrent 尚未向 MLCommons 提交 MLPerf 推理基准结果
安全 — CVE / 漏洞披露公开 CVE 跟踪或安全公告未确认未发现 TT-Metal/TT-Forge 的公开安全公告页面或 CVE 数据库记录
出口管制BIS 出口管理条例(EAR)UnknownTSMC 6nm 不在当前受限层级名单上;Blackhole 的 BIS 分类未公开确认
供应链质量TSMC 6nm 单一来源代工已识别风险依赖单一晶圆厂;未披露第二来源或封装认证

多数合规信息来自公开披露缺失,或按商业硬件惯例推断;未获得 Tenstorrent 直接确认。

[CE010, CE015, CE022, CE024, CE033, CE037]
FE002: 客户工作流 / 运营流
[CE021, CE039, CE040, CE016, CE017]

5.4 部署、集成与产品路线图

Tenstorrent 硬件部署遵循一条清晰的客户路径:开发者或企业团队先拿到 PyTorch、JAX 或 ONNX 模型,经 TT-Forge 编译器生成 Blackhole 原生二进制,再通过 TT-Metalium 调度执行。单卡配置(p100a / p150a)需要支持 PCIe Gen5 的主机服务器和 Linux 环境。TT-Metal 安装文档给出了相对直接的 Ubuntu pip 安装流程,但截至 early 2026,社区讨论反复提到 Linux 内核版本兼容约束带来的摩擦。多卡集群使用 p150a 和 p150b 卡内置以太网口;Tenstorrent 的 TT-Mesh 框架负责跨多芯片编排分布式推理,不要求单独通信库。Galaxy Blackhole 服务器以完整 6U 机架单元交付,内部 mesh 网络预先配置,降低企业级客户的集成负担。形态上,p100a 面向桌面推理和研究工作站;p150a 覆盖工作站与小集群推理;p150b 面向高密度机架推理;Galaxy 则面向超大规模云厂商和企业本地 AI 计算。Tenstorrent 目前不支持训练工作负载——整条产品线只做推理,相比 NVIDIA CUDA 生态,可服务市场受限。近期公开路线图包括 TT-Forge v1.0 稳定性(目标 2026)、Galaxy 集群扩展至 144 个节点(4,608 颗芯片),以及继续扩展 ONNX 覆盖。下一代芯片(Blackhole 后续)正在 NDA 下积极开发,没有公开规格。DevCloud 服务为无法直接购买硬件的开发者提供远程 Wormhole 和 Blackhole 访问,给未来企业转化蓄水。Blackhole 卡在 May 2026 开始量产,这意味着公司具备履行 Series D 在 December 2024 关闭时披露的约 $150 million 已签积压合同的能力。 [CE018, CE019, CE020, CE021, CE023, CE026]

路线图 / 发布 / 开发阶段表
里程碑状态目标 / 实际日期描述
Wormhole 开发套件(W2300、E75)正式可用Jul 2024第一代 Tenstorrent AI 加速卡;Blackhole 的上一代产品
Blackhole p100a / p150a 开发卡正式可用Nov 2024(Dev Day)第二代 ASIC 在 Tenstorrent Dev Day 发布;提供 PCIe + Ethernet SKU
TT-Metal v1.0 运行时稳定版正式可用2025核心运行时和 TT-NN 算子库的生产稳定版
Galaxy Blackhole Server 正式可用正式可用Apr 28, 202632 芯片 6U 企业 AI 服务器进入量产并正式可用
Blackhole 量产量产中May 2026TSMC 6nm 大规模制造爬坡,用于交付 $150M+ 合同积压
TT-Forge v1.0(编译器稳定版)进行中2026(H2 目标)MLIR 编译器成熟化;目标是 TT-Torch / TT-XLA / ONNX 全覆盖并正式可用
Galaxy 144 节点集群开发中2026–2027扩展到 4,608 颗 Blackhole 芯片,面向百亿亿级推理场景
MLPerf 推理提交未规划Unknown无公开路线图项目;缺失该项会限制第三方性能验证
训练支持不在路线图上UnknownTenstorrent 公开聚焦推理;未宣布训练能力
下一代芯片(Blackhole 之后)NDA / 开发中UnknownASIC 开发在 NDA 下推进;未确认公开规格或流片日期

进行中里程碑的日期来自公开表述和社区信号估计;暂无官方公开路线图文件。

[CE019, CE023, CE027, CE031, CE032, CE033]
FE004: 产品成熟度 / 能力图
[CE016, CE022, CE023, CE027, CE031, CE033]

5.5 技术差异化与竞争护城河

Tenstorrent 的主要技术差异化有四点:(1)Tensix 数据流架构,无需动态 GPU 式线程管理开销,就能做静态调度、带宽效率更高的矩阵运算;(2)原生以太网芯片互联,可用通用网络硬件搭出机架级推理集群;(3)嵌入式应用级 RISC-V 核心可在芯片上运行 Linux,摆脱专有固件锁定;(4)完整 Apache 2.0 开源软件栈,支持第三方贡献,也避开 NVIDIA 封闭 CUDA 生态的供应商锁定。Tensix 架构可追溯到 Jim Keller 过去在 AMD(Zen)、Apple(A4/A5)和 Intel 的 CPU 设计经验;Keller 及其联合创始人持有数据流 tile 架构的基础专利,但 IP 组合深度尚未经独立审计。相较 AMD Instinct MI300X 和 Intel Gaudi 3,Blackhole 的以太网互联策略在架构上不同,规模化后可能更具成本效率,因为不需要专有高速互联硬件。开源定位也构成一个建设中的护城河:随着 tt-metal 仓库积累社区贡献的模型移植、算子优化和集成指南,生态会越来越黏,而 Tenstorrent 自身不必同比例增加工程投入。差异化逻辑的关键脆弱点包括:对 TSMC 6nm 的单一来源依赖(CE024),使 Tenstorrent 暴露在所有先进制程芯片公司共同面对的地缘政治与产能风险中;缺乏训练支持(CE023),使 Tenstorrent 无法承接模型开发工作负载,也让同时做训练和推理的客户把它视作二线供应商;软件打磨不足(CE022),会把集成税转嫁给没有专门 MLOps 工程资源的客户。截至 May 2026 仍无 MLPerf 提交,使客观第三方基准对比变得困难,买方只能依赖公司提供的数字(例如相对 NVIDIA GB300 的 5× 单 token 成本主张),而这些数字尚未被独立评测者在生产规模验证。 [CE002, CE003, CE010, CE022, CE023, CE024]

FE003: 关键依赖图
[CE024, CE025, CE034, CE001, CE009, CE010]

5.6 要点展示

Chapter 06

06客户情况

6.1 客户分层与买方画像

以当前商业化阶段看,Tenstorrent 的可服务客户群可分为三类买方。第一类是独立开发者或研究实验室:个人和团队评估 Blackhole 卡($999–$3,499)或访问 DevCloud,主要动机是开源工具链好奇心、RISC-V 研究,或以更低成本做 LLM 推理实验。截至 Q1 2026,Tenstorrent 报告约 5,000 个注册 DevCloud 账号;这个指标统计注册量,不代表月活用户。第二类是战略企业合作伙伴——既是 Tenstorrent 投资方,又是其硅片早期采用者的组织。LG AI Research(由领投 Series D 的 LG Technology Ventures 支持)和 Hyundai Motor Group(Series D 战略投资方)正落在这一类,分别用 Blackhole 做 AI 推理和汽车 AI 计算。这种投资方兼客户关系会加速早期采用,但也带来治理模糊:购买动机部分受股权持有影响,而非完全基于公平市场化的产品评估。第三类是云或基础设施提供商,将 Tenstorrent 算力作为服务转售。法国云创业公司 Koyeb 成为首个公开确认的云 HaaS 合作伙伴,按 token/second 计费提供 Tenstorrent Blackhole p150 推理算力。SoftBank 披露将 Tenstorrent 芯片部署到日本数据中心扩张项目,是已披露最大商业订单,符合基础设施运营商画像。地域上,早期采用集中在北美(DevCloud、学术合作伙伴)、韩国(LG、Hyundai)、日本(SoftBank)和西欧(Koyeb、Fraunhofer Institute)。垂直场景主要是企业 AI 推理、汽车车载 AI、学术研究和云基础设施供给。尚无已披露的 SMB 或中端市场客户基础,这与当前产品的企业级价格点(每台 Galaxy 机箱 $110,000)一致。 [CU001, CU002, CU003, CU004, CU005, CU006]

客户分群表
分群示例地域垂直行业使用场景渠道约略规模
独立开发者 / 研究员DevCloud 注册用户、MIT/Stanford/CMU 研究组北美、欧洲学术界、AI 研究LLM 推理评估、模型移植、RISC-V 研究DevCloud(免费层)~5,000 个账户
战略投资者客户LG AI Research、Hyundai Motor Group 两家客户韩国企业 AI、汽车AI 推理 / 训练研发、自动驾驶算力直销 / 战略合作2 家具名实体
云 / HaaS 合作伙伴Koyeb法国 / 欧洲云基础设施推理即服务(按 token/second 计费)转售商 / 云平台1 家已确认伙伴
国家级基础设施运营商SoftBank Japan日本电信 / 数据中心AI 数据中心容量扩张直接交易1 笔已披露交易
学术 / 研究机构Fraunhofer Institute、大学实验室德国、欧洲研发 / 政府硬件评估、基准测试直采 / 开发套件采购多家机构
政府 / 国防(潜在)US DoD(兴趣阶段)美国国防 / 国家安全本土 AI 芯片采购(CHIPS Act)政府采购商业化前

分群规模为估计;“约略规模”反映截至 May 2026 公开披露或推断的数量,不代表经独立验证的收入占比。

[CU001, CU002, CU003, CU004, CU005, CU006]

6.2 采用轨迹与开发者生态

Tenstorrent 的采用轨迹更像一个双速飞轮:快速增长但较浅的开发者漏斗,输送给转化较慢但价值更高的企业管线。DevCloud 项目以 Wormhole 硬件启动,并在 late 2025 扩展到 Blackhole,截至 Q1 2026 已累计约 5,000 名注册开发者。主要 tt-metal 仓库的 GitHub 活动也反映了这股动能:截至 April 2026,超过 25,830 commits、1,410+ stars 和 19,076 个已合并 PR。不过,从开发者漏斗到企业转化仍未被证明:Tenstorrent 没有披露多少 DevCloud 用户随后购买硬件或转化为付费企业合作。Galaxy Blackhole 服务器直到 April 28, 2026 才正式可用——距本报告基准日只有几周——因此现在衡量生产部署中的复购率或集群扩容还太早。上一代 Wormhole 产品有更长观察期:QuietBox 工作站(基于 Wormhole,约 $12,000)自 mid-2024 起可购买。学术参与(MIT、Stanford、CMU 研究组)通过 Tenstorrent Developer Day 报道得到确认,但购买仍属小规模(单卡或开发套件)。Tenstorrent 声称 HuggingFace 模型基准测试通过率为 90%,并兼容 2.5 million 个开源模型;这些说法可推动需求侧采用,但缺乏独立验证。Galaxy 平台开放后,为超大规模云厂商相邻客户创造了新的采用路径,行业媒体报道至少有一个未具名超大规模云厂商级兴趣方。 [CU001, CU009, CU010, CU011, CU012, CU013]

客户增长 / 采用轨迹表
里程碑 / 指标时期数值 / 状态证据质量备注
Wormhole 开发套件开启预购Jul 2024预购起价 $12,000公司声称首次商用硬件发布
Koyeb Blackhole p150 HaaS 上线2024 年末 / 2025 年初生产部署已确认第三方确认首个独立交易商业客户
DevCloud 注册开发者账户Q1 2026~5,000 个账户公司声称注册数,不一定是月活用户
Series D 战略投资者成为客户Q4 2024 – Q1 2025LG AI Research、Hyundai、SoftBank(3 家实体)公司声称 + 媒体报道投资者与客户双重关系
Galaxy Blackhole 正式可用发布Apr 28, 2026正式可用,价格 $110,000/chassis已确认(媒体 + 公司)推出时间太近,暂无留存数据
tt-metal GitHub 星标(采用代理指标)Apr 20261,410+ 个星标,25,830+ 次提交观察值(GitHub)开发者社区参与信号
HuggingFace 模型兼容性声明Apr 20262.5 million 个模型,90% 通过率公司声称未经独立验证
学术合作已确认2024–2026MIT、Stanford、CMU + Fraunhofer第三方确认(Developer Day 报道)研究 / 评估,不是生产部署

里程碑日期和指标来自公司新闻稿及行业媒体;DevCloud 账户数为公司声称且未经验证;DevCloud 之后的漏斗转化率为估计值。

[CU009, CU010, CU011, CU012, CU013, CU014]
FU001: 客户旅程图
[CU001, CU009, CU010, CU011]
FU002: 采用 / 部署漏斗
[CU001, CU009, CU012, CU013]

6.3 具名客户证明

按企业硬件标准看,具名客户证据集偏小,但参与机构层级值得注意。Koyeb 提供了最清晰的独立客户证明:这家法国云创业公司公开宣布 Tenstorrent Blackhole p150 已在其 HaaS 平台上线,完成了客户今天即可购买的生产部署。这是唯一已确认的独立(非投资方)商业客户,用 Tenstorrent 硅片承载可产生收入的工作负载。LG AI Research 是战略合作伙伴兼客户:LG AI Research 由 LG Technology Ventures(Series D 领投方)支持,使用 Tenstorrent 芯片做 AI 推理和训练研发。双方关系早于 Galaxy 发布,且横跨 Wormhole 和 Blackhole 两代,显示出多代承诺。Hyundai Motor Group 的兴趣聚焦汽车 AI——面向自动驾驶辅助系统的车载处理——并由 Series D 战略投资带来的路线图协同支撑。SoftBank 的日本 AI 基础设施协议与其参与 Series D 融资(January 2025 公告)一同披露,涉及为 SoftBank 数据中心扩张部署 Tenstorrent AI 芯片;部署规模和时间线仍未披露。Fraunhofer Institute(Germany)被列为欧洲学术与工业研究中的早期 Wormhole 采用者。MIT、Stanford 和 CMU 的学术证明覆盖硬件评估和模型研究,不是生产工作负载。Jaguar Land Rover 关联曾在汽车 AI 计算语境中被提及,但没有具体、已确认的部署公告。总体看,对于一家寻求企业硬件可信度的公司,这套具名证明仍偏薄:一个独立 HaaS 客户、三个投资方绑定客户、一个国家级基础设施合作伙伴——而且都发生在 Galaxy 正式可用日期之前。 [CU003, CU004, CU005, CU006, CU016, CU017]

具名客户证据表
客户 / 合作伙伴国家部署阶段使用场景是否投资者?证据来源证据时效
Koyeb法国生产部署(HaaS)Blackhole p150 推理云,按 token/second 计费Koyeb 公开博客文章2025
LG AI Research韩国研发 / 战略AI 推理和训练工作负载(Wormhole + Blackhole)是(LG Technology Ventures,Series D 领投)媒体 + Series D 公告2024–2026
Hyundai Motor Group韩国研发 / 试点汽车 AI、自动驾驶 / ADAS 车载算力是(Series D 战略投资者)Series D 公告2024–2025
SoftBank(日本)日本已签约(部署前)为日本数据中心扩张供应 AI 芯片是(Series D 参与者)多家媒体报道Jan 2025
Fraunhofer Institute德国研究 / 试点Wormhole 硬件评估、AI 研究基准测试Developer Day + 行业媒体2024–2025
MIT / Stanford / CMU美国学术研究Blackhole 硬件评估、模型移植研究Developer Day 报道2025–2026
Jaguar Land Rover(未确认)英国潜在 / 未确认汽车 AI 算力(媒体提及,未确认)仅行业媒体提及2024–2025

JLR 行未经确认,列入仅为完整性;“Arms-Length?” 和 “Multi-Gen?” 的判断来自公开证据推断,可能不反映私下商业安排。

[CU003, CU004, CU005, CU006, CU016, CU017]

6.4 留存、耐久度与满意度

截至 May 2026,Tenstorrent 没有公开净留存率(NRR)、总留存率(GRR)或 cohort renewal 数据。考虑到公司的主要商业产品(Galaxy Blackhole)在本报告基准日前几周才正式可用,这个缺口符合阶段预期。数据缺失更多是时间因素,不是红旗,但也意味着任何留存判断都必须依赖间接代理指标。最可信的黏性代理是跨硬件世代的再次投入:LG AI Research 持续使用 Wormhole 和 Blackhole 两代 Tenstorrent 硅片,显示出多年承诺。Koyeb 的生产部署同样代表一个耐久集成决定——云提供商通常不会因短期试用就接入硬件合作伙伴,因为重构推理后端的运营成本很高。开发者侧满意度则更混合:截至 April 2026,tt-metal GitHub 仓库有 3,488 个未关闭 issues,这个量级说明社区参与活跃,也说明已知软件缺陷和缺失能力有明显积压。The Register 在 November 2025 的评测认为 Tenstorrent 软件“对大多数本地 AI 爱好者来说不够打磨”,并指出配置摩擦和驱动文档不完整。这篇评测是目前唯一可获得的实质性独立第三方产品评价,应视作非绑定客户终端用户满意度的最佳当前代理。Fraunhofer Institute 在多轮评估周期中持续研究使用 Tenstorrent 硬件,为学术分部提供了正向留存信号。企业硬件交易(LG、Hyundai、SoftBank)的合同期限未公开披露。公开记录中没有 NPS 分数、客户满意度调查或支持升级数据。 [CU010, CU022, CU023, CU024, CU025, CU026]

留存 / 重复使用 / 满意度表
指标数值 / 观察来源置信度含义
净留存率(NRR)未披露无公开数据N/AGalaxy 正式可用时间太近;NRR 无法衡量
总留存率(GRR)未披露无公开数据N/A无合同续约数据
多代客户(LG AI Research)Wormhole + Blackhole 采用已确认Series D 媒体报道 + 公司正面:多周期承诺
Koyeb 平台集成生产级 HaaS 部署(持久)Koyeb 博客文章云厂商很少快速退出硬件后端
开发者满意度(The Register 评测)多数 AI 爱好者会觉得软件“打磨还不够”The Register,Nov 2025负面:非绑定开发者用户遇到摩擦
tt-metal 未关闭 issue截至 Apr 2026 有 3,488 个未关闭 issueGitHub(观察值)软件问题积压会拖累开发者分群留存
合同期限(企业交易)未披露无公开数据N/A无法评估流失风险
学术机构再次参与Fraunhofer 持续做多周期评估Developer Day / 行业媒体正面信号,但商业相关性有限

NRR/GRR 行明确标为“未披露”;LG 多代和 Koyeb 持久集成行来自公开证据推断,而非实测留存数据。

[CU022, CU023, CU024, CU025, CU026, CU027]
FU004: 留存 / 复购队列
[CU022, CU023, CU024, CU025, CU026]

6.5 扩张、集中度风险与战略一致性

对一家处于当前融资阶段的公司而言,Tenstorrent 的客户集中度风险异常高。五个最可见具名客户中,有三个(LG AI Research、Hyundai Motor Group、SoftBank)同时也是 Series D 投资方。这种一致性可以加快路线图共创并减少销售周期摩擦,但带来两项实质风险:(1)投资方客户可能优先考虑股权回报,而不是给出坦诚产品反馈,从而削弱本应用来指导路线图决策的市场信号;(2)如果任何战略投资伙伴减持或退出,对应客户关系可能同步弱化,形成收入与股权风险的相关冲击。正面看,若干客户存在多代升级潜力。Galaxy 服务器的模块化架构支持集群扩展(追加机箱),Koyeb 的 HaaS 模式会随终端用户推理需求扩张。LG 和 Hyundai 可能成为下一代硅片的锚定客户。先落地再扩张模型在结构上成立,但尚未在实践中证明:没有已公告客户从初始采购继续扩大的案例。渠道和采购摩擦也是重要风险。Tenstorrent 目前还没有通过传统 OEM 渠道进入企业硬件采购;客户要么直接购买,要么通过 Koyeb 的云层获取。DoD 采购兴趣(CHIPS Act 对齐、美国本土 AI 芯片采购)提供了另一条高价值渠道,但会带来合规复杂性。Tenstorrent 的军民两用 AI 硬件在 ITAR 和 EAR 下的影响尚未公开说明。NVIDIA H100/H200 装机基础带来的竞争压力,会让已投入 CUDA 生态的潜在客户面临高切换成本;Tenstorrent 的 TT-Metalium 提供功能上不同、但成熟度更低的编程模型。 [CU003, CU004, CU005, CU006, CU007, CU028]

扩张与集中度风险表
风险 / 机会类别严重性证据缓释因素
前 5 大具名客户中有 3 家也是股权投资者集中度 / 治理LG、Hyundai、SoftBank 均参与 Series D独立交易条款未公开确认
尚无已验证的落地后扩张案例扩张未披露后续追加采购Galaxy 模块化架构支持未来扩张
单一非关联商业客户(Koyeb)集中度唯一非投资方生产部署客户云端 HaaS 可随终端用户需求扩容
Galaxy 直到 2026 年 4 月才 GA——扩张数据还太早时点Galaxy 在报告日前 2 周上线据报道,潜在超大规模云厂商已有意向管线
无传统 OEM / 渠道合作伙伴渠道依赖仅直销DoD 政府采购路径仍可能成立
潜在客户被 NVIDIA H100/H200 CUDA 生态锁定竞争 / 切换成本CUDA 装机基础庞大TT-Metalium 提供开源替代,但成熟度较低
投资方兼客户的双重激励模糊市场信号治理LG / Hyundai 既是投资方也是客户交易条款没有可用独立审计
地域集中:韩国 + 日本 = 5 个客户中的 3 个地域LG、Hyundai、SoftBank 均在亚太Koyeb(EU)和 DevCloud(全球)带来分散度

严重性评级基于可得公开证据做定性评估;集中度百分比只按已披露具名客户计算,可能无法反映私下管线。

[CU028, CU029, CU030, CU031, CU032, CU033]
FU003: 客户证据矩阵
[CU028, CU030, CU032, CU033, CU035, CU036]
Chapter 07

07风险

7.1 监管与法律风险

Tenstorrent 最紧迫的近期风险是出口管制暴露。US Bureau of Industry and Security(BIS)在 October 2023 发布里程碑式 Advanced Computing 规则,并在 October 2024 扩大范围,设定 FLOPS 和互联带宽门槛,用来判定 AI 加速器向中国、俄罗斯及越来越多受限目的地销售时是否需要出口许可证。Tenstorrent 的 Blackhole 芯片在 FP8 下提供 664 TFLOPS,而 BIS 规则适用于性能密度高于特定门槛、或片上互联性能高于特定限制的芯片。Tenstorrent 尚未公开披露 Blackhole 是否触发这些门槛、是否取得任何出口许可证,或其分销渠道是否执行最终用途筛查。Federal Register rule 2023-25073(effective December 2023)及后续 October 2024 扩展,显著拓宽了受限国家和技术范围。违反 Export Administration Regulations(EAR)会带来最高 $1M / violation 的刑事罚金和最高 20 years imprisonment,并叠加最高 $364,992 / violation 的民事罚款。另一个维度是 EU AI Act(enacted August 2024):该法按风险层级划分 AI 系统,可能要求向 EU 销售的 AI 硬件提供商完成合格评定、透明度和文档义务。虽然该法主要针对 AI 软件系统,但作为专用 AI 加速器销售的硬件,可能在不断演进的欧盟委员会实施法案下受到监管审查。RISC-V 出口管制风险又增加了第二条监管向量:美国立法者在 2024 年讨论过限制向中国公司授权 RISC-V IP;虽然没有出台全面禁令,监管不确定性仍给 Tenstorrent 的北京办公室和基于 RISC-V 的 IP 授权活动带来法律暴露。IP 诉讼方面,NVIDIA 持有超过 10,000 项 AI 和 GPU 专利;尽管公开记录中目前没有针对 Tenstorrent 的诉讼,这个悬置风险对任何 AI 芯片创业公司都很实质。Jim Keller 曾在 Intel、AMD、Apple 和 Tesla 之间流动,带来前雇主 IP 索赔风险;不过截至 May 2026 尚无索赔提出。Synopsys 和 Cadence EDA 工具授权协议是关键依赖:任何违约或未续约都会让芯片设计停摆。 [CR001, CR002, CR003, CR004, CR005, CR006]

监管 / 法律风险清单
风险 / 规则 / 案件司法辖区状态可能性严重性关键缓释措施剩余暴露尽调路径
BIS 出口管制——Blackhole FLOPS 阈值美国 / EAR合规状态未确认极高法律顾问审查;未披露对中国直接销售未知:无公开分类函取得 BIS 分类意见;核验最终用途认证做法
RISC-V IP 对华出口美国 / EAR / 行政令监管不确定北京办公室运营可能受限部分缓释:尚无全面禁令生效,但风险仍在跟踪 RISC-V Foundation 指引;取得出口管制律师意见
EU AI Act 符合性要求EU2024 年 8 月已颁布;分阶段生效跟踪欧盟委员会实施法案;推进 CE 标志流程中:时间表允许准备聘请 EU 监管律师;评估 Blackhole 风险层级分类
NVIDIA 专利侵权暴露美国 / 全球无未决案件低-中自由实施分析(未公开披露)NVIDIA 拥有 10,000+ 项 AI 专利,构成实质压力索取 FTO 意见;开展独立专利分析
EDA 工具许可证依赖(Synopsys/Cadence)美国 / 全球假定许可证有效极高多元采购 EDA 工具;研究开源 EDA高:未披露公开备选方案在数据室确认许可证期限和续约条款
Jim Keller 前雇主 IP 主张美国无未决案件入职 IP 陈述;员工 IP 协议低:常见风险,可用标准协议管理审阅雇佣协议中的 IP 陈述

截至 2026 年 5 月,公开法院记录中未发现针对 Tenstorrent 的未决诉讼。出口管制合规状态是优先级最高的尽调项。

[CR001, CR002, CR003, CR004, CR005, CR006]
FR001: 风险热力图
[CR001, CR009, CR017, CR024, CR032, CR042]
FR002: 风险传导图
[CR001, CR009, CR017, CR024, CR032, CR043]

7.2 运营与供应链风险

TSMC 是 Tenstorrent Blackhole(6nm node)的唯一晶圆制造来源,也是未来世代的主导晶圆厂。这个单一来源依赖带来三条独立风险向量上的灾难性暴露:Taiwan 地缘政治中断(People's Republic of China 潜在军事行动)、TSMC 内部运营事件(地震、火灾、污染),以及产能分配约束(TSMC 优先服务 Apple、NVIDIA 和 AMD,三者合计消耗了不成比例的先进制程产能)。Tenstorrent 没有披露针对先进制程节点的第二晶圆厂来源,也没有与 Samsung 或 Intel Foundry 做多来源认证。Blackhole 的内存供应(32GB GDDR6)依赖 Samsung、SK Hynix 和 Micron——这些供应商会优先把 HBM 产能分配给 NVIDIA 和 AMD。GDDR6 供给比 HBM 宽松,但在内存周期下行中,价格飙升或配额削减仍可能影响 Blackhole 利润率。软件质量本身也是运营风险:The Register 在 November 2025 对 Blackhole QuietBox 的评测结论是,Tenstorrent 软件“对大多数本地 AI 爱好者来说就是不够打磨”。截至 April 2026,tt-metal GitHub 仓库有 3,488 个未关闭 issues 和 988 个未关闭 PR——这些指标显示未解决 bug 积压很大,评审带宽也受限。芯片设计周期长(18-24 months 从规格定义到 tape-out),意味着 Blackhole 后继产品到 late 2026 已经锁定。如果 Blackhole 商业表现不及预期,中途没有可用的修正窗口。开源模型内嵌安全风险:TT-Metal / TT-Metalium 采用 MIT 许可证,任何安全漏洞都会在补丁发布前公开暴露并可能被利用。 [CR009, CR010, CR011, CR012, CR013, CR014]

运营 / 质量 / 安全风险清单
失效模式可能性严重性缓释成熟度剩余暴露未解决缺口
TSMC 单一晶圆厂中断(台湾冲突、地震、产能削减)极高灾难性:没有生产替代方案未披露第二晶圆厂认证
软件成熟度不足,限制企业客户转化高:3,488 个未关闭 issue、负面评价未披露带有明确企业 SLA 目标的软件路线图
芯片设计周期锁定(18-24 个月)高:Blackhole 后续产品到 2026 年末已锁定不具备周期中设计修正能力
GDDR6 内存分配风险(Samsung/SK Hynix)低-中中:GDDR6 受限程度低于 HBM未披露内存供应协议
开源安全漏洞(TT-Metal MIT)中:存在公开零日暴露窗口未披露安全响应 SLA 或 CVE 流程
Galaxy 服务器 OEM 认证延误高:未披露大型云厂商认证认证周期 6-12 个月

可能性和严重性按公开披露和行业常态估计;内部风险评估未公开。

[CR009, CR010, CR011, CR012, CR013, CR014]

7.3 技术与竞争风险

Tenstorrent 的核心技术风险是 CUDA 生态锁定。NVIDIA 用 20-year 护城河构建了 CUDA 库、框架和开发者工具;把 AI 工作负载迁移到 Tenstorrent 的 TT-Forge / TT-Metal 栈,需要开发者重新埋点模型代码、重新验证数值精度,并重新训练运营团队。即使 Tenstorrent 硬件在单位美元吞吐量上优于 NVIDIA,企业客户的生态切换成本也可能高到难以承受。NVIDIA H100/H200 和 Blackwell(B200)产品主导 AI 训练市场,CUDA 对训练工作负载的控制几乎绝对。Tenstorrent 已明确把 Blackhole 定位为推理加速器,相当于让出训练市场。这种推理聚焦限制了可服务市场份额。推理领域的竞争者——Groq(LPU architecture)、Cerebras(wafer-scale)、SambaNova 和 AMD MI300X——都在积极争夺同一批推理工作负载。Google TPU v5 为 Google Cloud 客户提供强大的内部推理能力。Intel Gaudi 3(通过 Habana Labs 收购而来)给超大规模云厂商提供了成本可能更低的 NVIDIA 推理替代品。独立技术评估反复指向同一主题:软件成熟度。Tenstorrent 若没有一套能把集成摩擦降到可比水平的软件栈,就无法关闭大型企业交易。声称 90% HuggingFace 模型兼容性,与提供生产级企业支持之间,仍有不小距离。此外,下一代芯片开发风险很高:TSMC 良率问题、架构误判,或 Blackhole 后继产品的可制造性设计错误,都可能让公司倒退 18-24 months;在此期间,NVIDIA 将发布 Rubin 及后续世代。 [CR017, CR018, CR019, CR020, CR021, CR022]

7.4 财务与资本风险

Tenstorrent 在 Series D(December 2024)融资 $693M,并在据报道的 Series E(November 2025)中又获得约 ~$800M。尽管资本位置相当充足,公司烧钱速度估计为 $25-50M per month,意味着从 Series E close 起现金跑道约 16-32 months。下一代芯片在 TSMC tape-out 估计需 $150-300M,占现金储备相当大一部分。收入未披露;第三方算法估计的 FY2025 $501M 不可靠,不应作为投资测算依据。截至 December 2024 的 $150M 已签合同代表积压订单,不是已确认收入;硬件交易从积压订单转化为现金回款,取决于客户验收测试、交付排期和质量保证里程碑。定制 AI 硬件的毛利率结构性承压:TSMC 晶圆成本、GDDR6 DRAM、板级组装和物流,会在任何 R&D、销售或 G&A 分摊之前吃掉收入的相当部分。公司没有公开说明盈利路径。投资方兼客户集中度进一步放大资本密集风险:LG、Hyundai 和 SoftBank 同时是已知最大客户和最大投资方之一。如果其中任何一方因自身财务压力或战略调整而降低投资姿态,Tenstorrent 会同时失去收入并释放内部人信心下滑的信号——这是灾难性的双重冲击。营运资本风险也很实质:企业硬件交易通常是 Net-60 到 Net-90,而晶圆付款要么预付、要么短期信用,压缩现金转换周期。 [CR024, CR025, CR026, CR027, CR028, CR029]

合作伙伴 / 依赖风险清单
依赖交易对手角色集中度失效情景严重性缓释措施剩余暴露
晶圆制造TSMC(台湾)6nm Blackhole 的唯一代工厂100%台湾冲突 / 晶圆厂停摆会让全部生产停止极高未披露;Samsung 晶圆厂认证未宣布灾难性
GDDR6 DRAM 供应Samsung / SK Hynix / MicronBlackhole 内存供应分散,但一线替代有限HBM 优先分配挤压 GDDR6 供应采购协议(未公开)
EDA 工具Synopsys / Cadence芯片设计工具双寡头许可证撤销或价格冲击会叫停下一代芯片设计极高假定有多供应商 EDA 合同;开源 EDA 正在出现
战略投资方客户LG Electronics / Hyundai Motor / SoftBank 三方锚定客户和 D/E 轮投资方~3 个实体贡献已知商业收入大头投资方退出会同时触发收入流失和信心冲击极高收入分散(DevCloud、Koyeb 等)
云厂商认证超大规模云厂商(未具名)Galaxy 服务器部署截至 2026 年 5 月,0% 获得认证缺少超大规模云厂商认证,限制大规模收入Koyeb HaaS 合作、直接部署到新型云厂商
IP 授权(Ascalon CPU)被授权方(未完整披露)收入多元化集中度未披露被授权方终止或不续约暗示采用多被授权方组合策略

客户与投资方身份重叠,是传统供应商风险框架抓不到的集中度最高依赖。

[CR009, CR026, CR027, CR030, CR036]
FR003: 依赖关系图
[CR009, CR025, CR026, CR036, CR044]

7.5 人才、治理与缓释措施

Jim Keller 身兼 CEO、首席架构师和主要投资者关系发言人,是 Tenstorrent 风险组合中严重程度最高的人才风险。他的履历包括 Intel(2017-2018,约 2 years)、Tesla(2016-2018,约 2 years)和 Apple(2008-2012,约 4 years)——模式是高影响力但任期有限。Keller 离开会在投资者(尤其是 Samsung,后者在董事会层面沟通中明确提到 Keller)和客户中触发严重信心冲击。公司未披露继任计划。工程人才留存风险偏高:从 Intel 和 AMD 招募已签竞业协议的工程师会带来诉讼暴露,Tenstorrent 还要与 Apple Silicon、Google TPU 和 NVIDIA 不断扩大的定制芯片团队争夺同一批芯片设计人才。对一家如此资本密集的公司而言,mid-2026 估计 1,100-1,200 人的员工规模,会在增长不能由收入匹配时制造组织规模风险。迄今已部署的缓释措施包括:开源软件栈在理论上降低客户切换成本(但实践中增加安全暴露);办公室地理分散(Toronto、Austin、Belgrade、Tokyo、Bengaluru、Seoul);来自 Tensix 和 Ascalon RISC-V CPU IP 的 IP 授权收入提供非硬件收入流;DevCloud 免费增值模式降低初始客户摩擦。终止标准包括:(1)BIS 对 Blackhole 在关键市场采取执法行动或拒发出口许可证;(2)TSMC 晶圆配额削减超过 30%;(3)Jim Keller 在 18 months 内离开;(4)现金跑道跌破 6 months 前未能完成下一轮融资;(5)到 end of 2027 仍未达成可与 CUDA 竞争的软件栈。尽调问题包括:出口管制律师对 Blackhole BIS classification 的意见、资料室中 TSMC 产能承诺证据、Jim Keller 股权归属时间表,以及已审计收入确认政策。 [CR032, CR033, CR034, CR035, CR036, CR037]

人员 / 执行风险清单
角色 / 职能依赖或缺口可能性严重性缓释措施尽调路径
CEO / 首席架构师(Jim Keller)唯一愿景负责人、投资者关系发言人、架构负责人极高股权归属时间表;未披露继任计划获取 Keller 股权归属时间表;确认董事会继任流程
资深 RISC-V / Tensix 架构师深层 IP 掌握在少数来自 Intel/AMD 的工程师手中有竞争力薪酬;IP 转让协议确认前 10 名架构师的竞业限制和 IP 转让状态
软件工程领导层TT-Forge / TT-Metal 成熟度缺口;988 个未关闭 PR开源社区贡献可部分补位评估软件 VP 团队人数和流失率
CFO / 财务职能未公开任命 CFO;E 轮暗示需要机构级财务能力Erik Goodman(VP Finance)提供过渡能力确认 CFO 招聘时间表和审计就绪度
销售和企业 GTM企业硬件销售周期 6-18 个月;未披露大型企业销售团队规模David Bennett(CCO)负责;投资方客户暖介绍可补位评估销售人数和销售管线 CRM 指标

Jim Keller 公开表达过对 Tenstorrent 的热情,但其历史任期偏短,因此离职概率难以下定论。

[CR032, CR033, CR034, CR035, CR041]
缓释措施和终止标准表
风险类别可监控触发项阈值 / 终止事件行动含义
出口管制BIS 执法行动、许可证拒绝或 Blackhole 出货暂停任何正式 BIS 命令限制 Blackhole 向当前服务市场销售立即暂停投资;在 30 天内寻求法律意见
TSMC 晶圆厂TSMC 产能分配公告或台海升级指数晶圆分配减少 >30% 或发生台海军事事件触发应急计划审查;加速 Samsung 晶圆厂认证尽调
关键人物(Keller)Jim Keller 离职公告或公开疏离信号上一轮融资后 18 个月内 Keller 离职标记为投资逻辑破裂;重评投资逻辑和继任安排
现金跑道月度烧钱速度与现金余额报告(私有;按员工数估计)现金跑道 <6 个月且没有已承诺的下一轮投资条款清单紧急讨论过桥融资;考虑估值下调融资风险
软件成熟度GitHub 未关闭 issue 数、企业客户留存、软件评价情绪未关闭 issue 数 >5,000,或首次发生重大企业客户流失加速软件尽调;考虑条件式投资触发项
客户集中度前 3 大投资方客户收入占总收入比例LG/Hyundai/SoftBank 合计占已确认收入 >60%要求客户分散计划,作为继续投资条件
竞争CUDA 兼容模式或 NVIDIA 开发者伙伴公告NVIDIA 宣布开源推理栈,匹配 Tenstorrent 的 TCO 主张重评产品差异化逻辑;加速软件栈基准测试

阈值是按尽调惯例和公开披露推导出的参考基准;应结合数据室实际指标校准。

[CR037, CR038, CR039, CR040, CR041]
Chapter 08

08估值

8.1 投资逻辑与反向逻辑

Tenstorrent 的投资逻辑建立在六根支柱上。第一,架构差异化:Tensix 处理单元和基于 RISC-V 的 Ascalon CPU 从根上偏离以 GPU 为中心的设计,可为推理工作负载带来非对称功耗效率。第二,市场时点:AI 加速器市场预计到 2030 年年收入达到 $170 billion,且推理分部增长快于训练,直接利好 Tenstorrent 的产品聚焦。第三,生态锚定:与 HuggingFace 模型 90% 兼容,加上开源软件栈(TT-Metal / TT-Metalium,MIT 许可证),降低开发者采用摩擦,并提供可信的 CUDA 替代叙事。第四,资本跑道:累计融资约 $2 billion(包括 November 2025 的 $800 million Series E),Tenstorrent 有资金支持至少一次下一代芯片 tape-out 和 16–32 months 运营。第五,战略投资方深度:LG Electronics、Hyundai Motor、SoftBank、Samsung Securities 和 Fidelity 提供绑定客户关系、分销触达和耐心资本。第六,IP 授权上行空间:授权给第三方的 Ascalon RISC-V CPU 可产生经常性、类似软件的收入,隐含倍数更高。 反向逻辑同样重。NVIDIA 的 CUDA 生态由 15+ years、数百万开发者小时和 $100+ billion 软件投入打造,黏性极强。企业客户面临 18–36 months 切换成本和显著再培训投入。The Register 在 November 2025 对 Blackhole QuietBox 的评测记录了 Tenstorrent 软件尚未达到企业级打磨,tt-metal 仓库的 3,488 个未关闭 issues 又放大了这一担忧。投资方兼客户双重身份(LG、Hyundai、SoftBank 同时是投资者和最大客户)带来循环依赖风险:单一实体退出可能同时导致收入损失和信心崩塌。在收入未确认的情况下,每月 $25–50 million 的烧钱速度意味着公司可能在现金流 breakeven 前需要再融资。TSMC 单一来源依赖增加了难以对冲的地缘政治和运营集中度。综合来看,投资逻辑要求多个条件同时成立;软件停滞、关键人物离开、TSMC 中断或降价轮中任一项失败,都可能实质性重置估值假设。 [CV001, CV002, CV003, CV004, CV005, CV006]

正反论点表
论点(看多因素)强度反论点(看空反驳)验证证据
RISC-V + Tensix 架构在推理上打出非对称能效优势CUDA 生态扎根 15+ 年;企业切换成本需要 18–36 个月消化同等 TCO 下 Tensix 与 H100 的独立企业基准测试
Jim Keller 的履历(Intel A-series、Apple M-series、Tesla FSD)为执行力背书Keller 每家雇主平均任期 <3 年;关键人离职风险不小Keller 股权归属时间表和董事会接班披露
累计融资 $2B+,足够支持下一代芯片流片,并覆盖 16–32 个月资金跑道每月现金消耗 $25–50M,意味着收支平衡前可能还需要 Series F 轮经审计的 FY2025 财务和月度现金消耗确认
Galaxy GA(2026 年 4 月)标志首个商业收入里程碑尚无公开确认的 Galaxy 交付订单或已确认收入已签采购订单和 2026 年 Q2 收入披露
战略投资者客户(LG、Hyundai、SoftBank)提供绑定需求循环依赖:投资者退出会同时触发收入和信心崩塌收入多元化:非投资者客户收入占总收入比例
到 2030 年 TAM 达 $170B;拿下 2–5% 份额即对应 $3.4–8.5B 收入AI 资本开支周期可能在 2027–2028 年收缩;TAM 估算尚未反映修正已确认超大规模云厂商将推理预算分配给非 NVIDIA 供应商
开源 TT-Metal 和 90% HuggingFace 兼容性支撑开发者生态3,488 个 GitHub 未结 issue 和 The Register 负面评测显示软件尚未达到企业级GitHub 未结 issue 降到 1,000 以下,并达成企业 SLA 里程碑

强度评级(高 / 中 / 低)反映当前证据对论点的支撑质量,而不是论点本身的先验可信度。

[CV005, CV006, CV013, CV019, CV026, CV027]
FV001: 推荐逻辑
[CV001, CV005, CV013, CV019, CV040]
FV004: 投资 KPI
[CV001, CV005, CV006, CV013, CV019, CV026]

8.2 投资建议、信心与估值立场

对 Tenstorrent 在 November 2025 Series E 价格($3.2 billion 投后估值)的尽调建议是继续研究 / 观察。这不是否定公司,而是在当前价格下否定可获得信息的充分性。 信心为中低(35 out of 100)。低信心的主要驱动因素是:(a)自 2021 年加拿大公司监管文件后,没有任何公开文件提供已审计收入数字;(b)Latka 对 FY2025 $501.6 million 的算法估计未经任何独立财务分析验证;(c)截至 December 2024 披露的 $150 million 已签合同代表积压订单,不是已确认收入;(d)独立评测者记录的软件质量风险尚未公开解决。 估值立场为中性偏负面。在 $3.2 billion 估值下,Tenstorrent 对应 Latka 收入估计的 6.4x,或保守收入估计的 16–32x。可比私营同业(Cerebras 16–28x、Groq ~10x、SambaNova 10–17x)显示,该估值处在或高于可比区间上沿——且没有已确认收入支撑。公开市场可比公司(Arm Holdings ~30x EV/revenue、Marvell ~10x)为区间定锚,但这些公司都有已审计、已确认收入流。 风险评级在五个维度均为高:软件成熟度、财务烧钱、TSMC 集中度、CUDA 生态护城河和关键人物依赖。每个维度单独看都值得提高监控强度。组合起来,足以支持继续研究,而不是选择性投资。 上调为选择性投资的触发条件:收入确认超过 $200 million(经审计或在资料室披露)、至少拿下一个主要超大规模云厂商设计定点、软件达到明确企业 SLA 里程碑,或在降价轮中以低于 $2.0 billion 的修正估值进入。下调为放弃的触发条件:Jim Keller 离开、BIS 对 Blackhole 采取执法行动、烧钱超过 $60 million per month,或在没有同步收入确认的情况下以低于 $2.0 billion 降价轮融资。 [CV014, CV015, CV016, CV017, CV038, CV039]

建议摘要表
维度评估评级关键证据 / 触发项
投资建议继续研究 / 观察🟡 观望收入未确认;Galaxy GA 是催化剂,但单靠它还不够
信心水平中低35 / 100无经审计收入;软件成熟度缺口;没有超大规模云厂商设计定点
估值立场$3.2B 下中性偏负面⚠️ 偏高保守收入的 16–32x;高于可比非上市同业区间
风险评级高(5 个风险偏高维度)🔴 风险高企TSMC、CUDA 护城河、烧钱速度、关键人物、软件成熟度不足
目标回报(5 年期基准)~5% IRR(基准情景)持平至小幅上行需要 Galaxy 在细分市场跑通,并且软件稳定下来
重评触发项(上调)确认收入 >$200M 或拿下超大规模云厂商选择性投资数据室审计或设计定点公告
重评触发项(下调)Keller 离职或 BIS 执法放弃TV005 所列投资逻辑破裂事件

信心按 0–100 量表打分,100 = 证据完全支持按当前价格投资。评级反映证据质量,而不是公司质量。

[CV001, CV014, CV015, CV016, CV017, CV040]
论点失效与退出触发因素表
触发因素信号 / 指标退出阈值对投资论点的传导行动
Galaxy GA 后收入未落地Galaxy 出货数据;积压订单转化率;2026 年 Q2–Q3 收入披露到 2026 年 Q4 仍无已确认收入 >$100M牛市和基准情景崩塌;熊市概率升至 >50%下调为放弃;30 天内争取数据室
软件成熟度停滞GitHub 未结 issue 数;企业 NPS;独立评测情绪未结 issue >5,000 或第二次重大负面独立评测企业采用论点失效;CUDA 护城河进一步加宽有条件通过;要求硬性软件冲刺计划作为继续持有敞口的条件
Jim Keller 离职LinkedIn / 新闻公告;董事会沟通Series E 轮完成后 18 个月内出现任何公开离职信号投资者信心崩塌;架构愿景承压;Series F 轮变得极难立即搁置;60 天内全面重估投资论点
BIS 对 Blackhole 采取执法行动Federal Register;BIS 新闻稿;公司披露BIS 正式下令限制 Blackhole 在任何现有市场销售生产停摆风险;刑事 / 民事责任;投资者信心被摧毁立即退出;出口管制违规无法保险覆盖
估值下修轮Series F 轮公告投后估值低于 $2.5B任何 Series F 轮定价低于 $2.5BSeries E 轮跌破成本;优先权清偿顺序吃掉普通股权益按模型减记;评估稀释影响;融资后重估股权结构表
投资者客户退出LG、Hyundai 或 SoftBank 宣布出售股权或终止合同三者任一在 24 个月内退出股权或终止合同收入损失与信心冲击同时发生;循环依赖兑现升级至董事会观察员;要求商业多元化计划

阈值是根据尽调惯例和前文章节风险分析推导的参考基准。拿到实际数据室指标后应重新校准。

[CV017, CV018, CV019, CV026, CV029, CV040]

8.3 融资背景与估值基准

Tenstorrent 自 2019 年以来通过六轮融资累计获得约 $1.99 billion:Seed(~$10 million,2019)、Series A($40 million,2021)、Series B($100 million,2022)、Series C($235 million,2023)、Series D($693 million,$2.6 billion 投后估值,December 2024)和 Series E($800 million,$3.2 billion 投后估值,November 2025)。Series D 由 Samsung Securities 领投,LG Technology Ventures 和 Hyundai 参投;Series E 由 Fidelity 领投。从 $2.6 billion 到 $3.2 billion 的快速估值上调(约 23% in 11 months)发生在没有确认收入披露的情况下,说明该轮定价更多基于战略可选性和投资者需求,而不是基本面财务指标。 清算优先权压力是实质担忧。Series D 投资者以 $2.6 billion 进入,Series E 投资者以 $3.2 billion 进入,均带有清算优先权;在熊市或困境情景下,如果退出估值跌破 $2.6 billion,普通股股东和早期投资者会面临显著的瀑布分配缺口。具体优先权机制以及 participating vs. non-participating preferred 结构未公开披露,缺少股权结构表就无法精确建模分配瀑布。 从可比口径看,Tenstorrent 的 6.4x Latka EV/revenue(未确认)处于可比私营融资轮低端:Cerebras(16–28x)、SambaNova(10–17x)和 Groq(~10x)。不过,如果采用 $100–200 million 的保守收入估计,Tenstorrent 隐含倍数升至 16–32x——达到或高于可比私营区间。公开市场锚点(Arm Holdings ~30x EV/revenue、NVIDIA ~26x、Marvell ~10x)表明,若确认收入达到 $200–300 million 且软件栈去风险,$3–5 billion 估值可以支撑,但不是在当前证据水平下。NVIDIA FY2025 数据中心收入约 $115 billion(已在 SEC 10-K filing 中确认)以及 Arm Holdings FY2025 收入约 ~$3.96 billion,为评估 AI 芯片硬件和 IP 授权倍数提供了公开市场锚点。 [CV001, CV002, CV003, CV004, CV007, CV008]

可比估值表
公司类型最近估值 / 市值(2026 年 5 月)收入估算隐含 EV/Revenue可比意义局限
Cerebras Systems未上市 AI 芯片初创公司$4–7B(2024 年提交 S-1)~$250M(2024 年估算)16–28x直接可比:纯 AI 芯片初创公司;聚焦 LLM 推理;已提交 S-1S-1 提交后撤回;收入结构未披露;无经审计公开收入
Groq未上市 AI 推理初创公司$2.8–4B(2024)~$300M(2024 年估算)~10x仅推理可比;融资轨迹相近;GroqCloud 已投入生产产品范围更窄;没有可与 Galaxy 服务器对照的硬件 ASP
SambaNova Systems未上市 AI 芯片 + 软件$5.1B(2023)~$300–500M(估算)10–17xAI 芯片 + 全软件栈;架构上更接近 Tenstorrent 模型估值停留在 2023 年;可能已过时;未披露 2025 年融资
Arm Holdings上市 RISC-V IP / 半导体~$120B(2026 年 5 月)~$3.96B FY2025(经审计)约 30x EV/revRISC-V IP 授权可比;公开市场;经审计收入基准完全上市;收入主要来自版税 / IP——与 Tenstorrent 硬件收入结构差异很大
NVIDIA上市 AI GPU 龙头~$3T(2026 年 5 月)~$115B 数据中心 FY2025(经审计)~26x 总 EV/rev主导性市场基准;设定定价和软件生态标准规模不可比;仅设定上限
Marvell Technology上市定制 ASIC 半导体~$60B(2026 年 5 月)~$6B FY2025(经审计)约 10x EV/rev定制芯片 / ASIC 可比;服务超大规模云厂商 AI 芯片定制项目ASIC 商业模式不同于标准产品;Google/AWS 客户集中

所有未上市公司收入估算均为算法或分析师模型测算;没有经审计的未上市同业收入公开可得。上市公司数据(NVIDIA、Arm、Marvell)来自 SEC EDGAR 10-K 文件和年报。Tenstorrent 的隐含倍数为 6.4x(Latka 估算)至 16–32x(保守 $100–200M 收入),对比未上市同业 10–28x 区间。

[CV007, CV008, CV009, CV010, CV011, CV012]
FV002: 估值敏感性
[CV001, CV007, CV015, CV016, CV042]

8.4 乐观、基准与悲观情景

三个情景框定 Series E 投资者以 $3.2 billion 进入后的预期结果。 乐观情景(20% 概率,5-year horizon through 2028–2029):Tenstorrent 到 2028–2029 年拿下约 5% 的 AI 推理加速器市场,受益于 Galaxy Blackhole 在 SoftBank、LG 和 Hyundai 数据中心部署,并至少获得一个主要云服务商认证。软件成熟到企业级(未关闭 issues 低于 1,000;获得独立企业验证)。Ascalon RISC-V 的 IP 授权收入增长至每年 $100–200 million。收入达到约 $2 billion;以 8x 收入退出倍数(与高增长硬件 + IP 公司中 NVIDIA 历史区间一致)计算,退出估值约 $16 billion。Series E 投资者实现 5x 回报(约 40% IRR over 4–5 years)。 基准情景(50% 概率,exit 2029–2030):Galaxy Blackhole 在 neocloud 中获得有意义部署,但没有主要超大规模云厂商认证。软件缺口收窄但未完全闭合;CUDA 在企业训练中仍占主导。收入达到约 $500–800 million(2–3% market share)。以 5x revenue 退出,隐含 $3–5 billion 估值——对 Series E 投资者而言约为持平到小幅正回报(~5% IRR)。这一情景下,Tenstorrent 作为利基玩家存活,但没有实现突破性规模。 悲观情景(30% 概率):软件质量缺口持续存在,NVIDIA 维持 85%+ 推理市场份额,Galaxy Blackhole 局限于战略投资方客户部署。烧钱迫使公司以 $500 million–$1 billion 降价轮融资或被收购。Series E 投资者面临 60–80% loss。关键触发因素:Jim Keller 离开、到 Q4 2026 仍没有任何 hyperscaler 试点胜利,或 BIS 对 Blackhole 执法。 概率加权预期估值($16B × 20% + $4B × 50% + $750M × 30%)等于 $5.4 billion——约为 Series E 入场价格的 1.7x,意味着期望值为正,但考虑到高方差和信息缺口,安全边际偏薄。 [CV022, CV023, CV024, CV025, CV030, CV032]

牛市 / 基准 / 熊市情景表
情景关键假设收入估算(2028–29)退出倍数退出估值概率Series E 轮 IRR(5 年)
牛市(2028)5% 推理市场份额;拿下 2 个超大规模云厂商设计定点;软件达到企业级;Ascalon IP 授权 $150M+$2.0B8x 收入$16B20%~40%
基准(2029–30)2–3% 细分市场份额;Galaxy 在战略投资者客户处部署;没有超大规模云厂商 GA;软件部分改善$600M5x 收入$4B50%~5%
熊市(2028–30)软件差距延续;NVIDIA 推理份额 >85%;现金消耗迫使估值下修轮或困境收购$200M<2x 收入$500M–$1B30%-40% 至 -60%

概率加权预期估值:($16B × 20%) + ($4B × 50%) + ($0.75B × 30%) = $5.4B,意味着基于 $3.2B 进入估值的预期倍数约 1.7x——结果为正,但在高方差下很薄。收入估算为模型测算;目前没有经审计的前瞻指引。

[CV022, CV023, CV024, CV025, CV033]
FV003: 估值 / 回报区间
[CV022, CV023, CV024, CV025, CV033]

8.5 退出准备度、尽调问题与破局触发条件

截至 2026 年 5 月,Tenstorrent 还没有达到 IPO 条件。关键准备缺口包括:尚未公开任命具备机构级报告经验的 CFO;未披露 FY2024 或 FY2025 经审计财务报表;没有能锚定企业市场可信度的超大规模云厂商设计定点;软件栈已有质量问题记录, 会招致机构投资者审查。Arm Holdings 2023 年 IPO 是参照基准:提交 S-1 招股说明书、披露多年经审计收入、 毛利率高于 90%,且 IP 授权收入可见度清晰。按这个门槛,Tenstorrent 目前还达不到。 并购更可能是近期退出路径。潜在战略买家包括 Qualcomm(AI 边缘芯片)、Samsung(与代工结合的 AI)、 Intel(定制 AI 硅片),以及 Arm Holdings 本身(RISC-V 生态整合)。AI 芯片并购的收购溢价基准—— AMD/Xilinx 约 $35 billion(15x 收入,2022)和 Broadcom/VMware($69 billion)——表明, 在乐观情景下,Tenstorrent 以 $5–8 billion 被技术溢价收购并非不可能,但前提是拿出已经验证的商业牵引力。 最终尽调清单(按优先级):(1) FY2025 收入和毛利率,经审计或数据室验证——关键,30 天;(2) Blackhole 的 BIS 出口合规文件——关键,60 天;(3) 股权结构表,包含 Series D/E 优先权条款和清算瀑布——关键,30 天;(4) Galaxy 已签交付订单和客户部署时间表——高优先级,45 天;(5) Jim Keller 股权归属时间表和董事会继任机制——高优先级, 30 天;(6) TSMC 产能协议和下一代节点认证时间表——高优先级,60 天。 触发投资逻辑破裂、必须立即暂停投资的信号包括:Jim Keller 公开离职、BIS 对 Blackhole 采取执法行动、 确认烧钱速度超过每月 $60 million 且没有收入支撑、低于 $2.0 billion 的降轮融资,或确认失去三大战略投资者兼客户中的任意一家。 [CV026, CV028, CV035, CV036, CV037, CV038]

最终尽调事项表
尽调事项优先级时间线负责人缺失证据负面时的估值影响
收入和毛利率核验(FY2025)关键30 天CFO / 数据室经审计或审计师审阅的 FY2025 收入;按产品线拆分毛利率;收入确认政策若收入 <$100M,隐含倍数升至 >32x;$3.2B 估值无法成立
Blackhole 的 BIS 出口合规文件关键60 天法务 / 合规BIS 分类意见函;出口许可证申请;分销商最终用途证明BIS 执法行动 = 论点失效;合规缺口 = 无法量化的责任
含优先权机制和清算顺位的股权结构表关键30 天法务 / 公司秘书含 Series D/E 轮优先权条款的详细股权结构表;参与型 / 非参与型;反稀释条款熊市情景下普通股可能归零;Series E 轮 IRR 画像将明显改变
Galaxy Blackhole 已签交付订单和客户部署时间线45 天BD / 销售Galaxy 服务器已执行采购订单;部署计划;收入确认触发日期没有订单,Galaxy GA 只是产品发布,不是收入事件
Jim Keller 股权归属时间表和董事会接班流程30 天CEO / 董事会主席CEO 雇佣协议;股权归属时间表;董事会批准的接班候选人Keller 离职且无接班计划,就是论点失效事件
TSMC 产能协议和下一代节点资格认证时间线60 天COO / 供应链TSMC 批量采购协议;Blackhole 承诺晶圆投片量;下一代工艺节点选择TSMC 单一来源且没有协议 = 任一中断下的灾难性生产风险
带企业 SLA 承诺的软件路线图45 天CTOTT-Metal v1.0 企业功能路线图;硬性里程碑日期;未结 issue 消减计划没有路线图,软件论点只是愿景;企业采用时间线无法建模
投资者客户收入集中度数据30 天CFO / BD按客户细分拆分收入;LG、Hyundai、SoftBank 各自占比三家投资者客户集中度 >60%,就是结构性治理问题

优先级评级:关键 = 阻断任何投资决定;高 = 初始兴趣后 60 天内必须拿到;中 = 交割前必须拿到。时间线假设已在 NDA 下进入活跃数据室。

[CV014, CV015, CV016, CV028, CV035, CV036]

8.6 附录

免责声明

本报告由 AI 研究工作流基于截至 2026-05-10 的公开来源生成,仅供信息参考,不构成投资建议。所有财务估计均来自第三方模型或公开可得的替代数据;未审阅经审计财务报表。读者在做出任何投资决定前,应自行开展独立尽调。

证据索引

结论
编号陈述可信度来源
CO001 Tenstorrent Inc. was incorporated in Canada on March 14, 2016. SO013, SO022
CO002 Tenstorrent's operational headquarters is at 2600 Great America Way, Suite 501, Santa Clara, California. SO018, SO004
CO003 Tenstorrent has global offices in Toronto, Austin (Texas), Belgrade (Serbia), Tokyo (Japan), Bengaluru (India), Seoul and Pangyo (Korea), Munich (Germany), Warsaw (Poland), and Beijing (China). SO007, SO018
CO004 Tenstorrent's primary revenue streams are hardware sales (PCIe accelerator cards and complete workstations) and IP licensing of Tensix and Ascalon RISC-V CPU designs. SO002, SO012
CO005 Tenstorrent is a fabless semiconductor company, having started with GlobalFoundries for first-generation chips and planning future generations through TSMC and Samsung foundries. SO004, SO005
CO006 Tenstorrent's open-source software stacks—including TT-Metal, TT-Metalium, TT-Forge (MLIR compiler), TT-Buda, and TT-Lang—are MIT-licensed and publicly available on GitHub. SO007, SO021
CO007 The Tensix core is a self-contained compute tile comprising a RISC-V data-movement processor, a matrix math engine for tensor operations, and a vector math unit, connected via an on-chip mesh network. SO009, SO020
CO008 Tenstorrent uses on-chip Ethernet connectivity (16×100GbE on Wormhole; 10×400GbE on Blackhole) to enable direct chip-to-chip scaling without external switch infrastructure. SO009, SO020
CO009 Tenstorrent's Galaxy Blackhole system entered volume production in May 2026, supporting 36-box supercluster configurations for large-scale AI inference. SO016
CO010 Jim Keller joined Tenstorrent as President and CTO in 2020 and was formally named CEO in early 2023. SO003, SO012
CO011 Jim Keller led the AMD Athlon K7/K8 (Opteron) architecture and co-architected the x86-64 instruction set extension during his first AMD tenure in the late 1990s. SO011, SO024
CO012 Jim Keller led the development of Apple's A4 and A5 mobile SoCs (iPhone 4, iPhone 4S, original iPad) while at Apple following its acquisition of P.A. Semi. SO011, SO003
CO013 Jim Keller served as VP of Autopilot Hardware Engineering at Tesla, where he led development of the Full Self-Driving Hardware 3 (FSD Chip) AI accelerator (2016–2018). SO011, SO024
CO014 Jim Keller served as Senior Vice President of Silicon Engineering at Intel from 2018 to 2020. SO011, SO003
CO015 Keith Witek serves as Chief Operating Officer of Tenstorrent and was the primary spokesperson for the December 2024 Series D fundraise. SO002, SO019
CO016 Tenstorrent's three co-founders are Ljubisa Bajic, Ivan Hamer, and Milos Trajkovic, who currently serve as senior fellows in engineering and systems roles. SO022, SO013
CO017 Jim Keller has a documented pattern of short tenures at prior employers, including approximately two years at Tesla (2016–2018) and two years at Intel (2018–2020), which analysts identify as a key-person succession risk. SO011, SO015
CO018 Tenstorrent closed a $693M+ Series D funding round on December 2, 2024, at a pre-money valuation of $2B and post-money valuation of approximately $2.6B. SO002, SO003
CO019 The Series D was led by Samsung Securities and AFW Partners, with the round described as oversubscribed due to strong investor demand. SO002, SO003
CO020 Additional Series D participants include XTX Markets, Corner Capital, MESH Ventures, Export Development Canada, Healthcare of Ontario Pension Plan, LG Electronics, Hyundai Motor Group, Fidelity Management & Research Company, Baillie Gifford, Bezos Expeditions, and SBI Investment. SO002, SO014
CO021 Tenstorrent raised a $100M Series C extension in August 2023, led by Hyundai Motor Group and Samsung Catalyst Fund, with participation from Fidelity, Maverick Capital, Kia, and Eclipse Ventures. SO014, SO012
CO022 Tenstorrent completed a $200M Series C tranche in May 2021, led by Fidelity Investments, at a post-money valuation of approximately $1B, achieving unicorn status. SO014, SO003
CO023 Tenstorrent has raised a total of approximately $1.18B across ten documented funding rounds since its 2017 seed round from Real Ventures. SO014, SO025
CO024 Stated use of Series D proceeds: build out open-source AI software stacks, hire developers, expand global development and design centers, and build systems and clouds for AI developers. SO002, SO019
CO025 Tenstorrent disclosed approximately $150M in signed commercial contracts as of the December 2024 Series D announcement. SO002, SO005
CO026 Early investors Eclipse Ventures and Real Ventures are cited as Tenstorrent backers since the seed and early-stage rounds. SO002, SO014
CO027 Tenstorrent's India subsidiary (Tenstorrent India Private Limited) was incorporated April 28, 2022, in Karnataka, India, with approximately 94 employees as of February 2026. SO013
CO028 Grayskull, Tenstorrent's first-generation chip, features up to 120 Tensix cores with 1MB SRAM each, 8GB of LPDDR4 memory on a 256-bit bus, and supports AI precision formats including FP8, FP16, and BF16. SO005, SO012
CO029 Wormhole (second-generation chip) features 80 Tensix+ cores, 12nm fabrication, 16×100GbE Ethernet, GDDR6 memory (12GB per card), and 328 TOPS peak performance, with PCIe cards priced at $1,000 (n150) and $1,400 (n300). SO006, SO008
CO030 Wormhole developer workstations (TT-LoudBox at $12,000 and TT-QuietBox at $15,000) were commercially launched in July 2024. SO006, SO007
CO031 Blackhole (third-generation chip, 6nm process) features 140 Tensix++ cores, 752 baby RISC-V cores, 16 Linux-capable RISC-V CPU cores, 10×400GbE Ethernet, 32GB GDDR6, and approximately 790 TOPS FP8 performance. SO009, SO010
CO032 The Blackhole QuietBox workstation ($11,999) began shipping in late 2025 and was hands-on reviewed by The Register, which confirmed receipt and testing of the hardware. SO006, SO010
CO033 Tenstorrent's Galaxy Blackhole server supports supercluster configurations of up to 36 Galaxy boxes linked into a single domain for large-scale AI compute. SO016
CO034 Tenstorrent's Galaxy Blackhole system demonstrated DeepSeek LLM inference at 308 tokens per second per user at a cost of $6 per million output tokens, and set a video-generation world record with Prodia (5-second video generated in 3.5 seconds, 83% faster than prior record). SO016
CO035 Tenstorrent employs an estimated 1,100 to 1,200 people globally as of mid-2026, up from approximately 370 in 2024 prior to the Series D. SO013, SO018
CO036 Tenstorrent's revenue was disclosed as $25M–$100M in a 2021 US corporate filing; no subsequent revenue disclosures have been made publicly available. SO013, SO025
CO037 Tenstorrent's open-source TT-Forge compiler achieves a claimed 90% pass rate for running models directly from Hugging Face, supporting approximately 2.5 million open-source AI models. SO016
CO038 Tenstorrent has Galaxy Blackhole hardware deployed in at least five neocloud co-locations as of May 2026, including flagship installations in Tokyo (largest deployment by ai&), Cirrascale in Seattle, Turium AI in India, and Virtu Financial for high-frequency trading research. SO016
CO039 The Register's independent review of the Blackhole QuietBox found the software stack 'simply isn't polished enough for most local AI enthusiasts,' citing immaturity as the primary limitation relative to the hardware. SO010, SO015
CO040 SWOT analysis published February 2026 identifies Tenstorrent's software ecosystem immaturity (TT-Buda vs. NVIDIA CUDA) and limited public customer deployments as the company's primary weaknesses, representing material commercialization risk. SO015, SO010
CO041 Tenstorrent's Series B was closed in January 2019 and raised $20.5M; Series A raised $500K in February 2018. SO014
CO042 Tenstorrent intends to release a new AI processor every two years, contrasting with NVIDIA's annual upgrade cadence. SO004, SO005
CM001 The AI accelerator chip market for this analysis includes discrete AI processors (GPUs, NPUs, ASICs) for training and inference in data centers, cloud, neocloud, edge, and embedded deployments; it excludes HBM/GDDR memory, networking switch chips, software, and infrastructure hardware. SM001, SM002
CM002 The adjacent RISC-V processor IP licensing market is a secondary revenue stream for Tenstorrent; the global RISC-V technology market is projected at $1.35B in 2025 and $1.91B in 2026 (30–41% CAGR through 2034). SM006, SM022
CM003 Tenstorrent's status-quo competitor and primary market incumbent is NVIDIA's GPU ecosystem, which held approximately 80% of AI accelerator chip revenue in 2025 and is projected to hold approximately 75–80% by 2026. SM009, SM011
CM004 Included spend in the AI accelerator market covers discrete PCIe accelerator cards, OAM modules, multi-chip modules, and complete AI server systems where the accelerator represents the primary value driver; RISC-V IP licensing is treated as an adjacent revenue stream. SM001, SM003
CM005 The primary status-quo substitute for any AI accelerator alternative is continued procurement of NVIDIA H100, H200, or Blackwell B100/B200/GB200 series GPUs, which remain the reference platform for AI training and inference in 2026. SM009, SM013
CM006 The RISC-V CPU IP licensing sub-market is sized at approximately $580M in 2025 and $720M in 2026 at a 12.1% CAGR through 2034, representing a more modest but higher-margin revenue stream than AI hardware. SM008, SM006
CM007 Gartner (April 2026) forecasts worldwide semiconductor revenue at $1.3T for 2026, with AI processing semiconductors representing approximately $268B or roughly 30% of total revenue. SM001, SM003
CM008 IDC (April 2026) forecasts data center semiconductor revenues at $477.1B for 2026, driven by AI infrastructure investment; this is the broadest definition and includes AI-optimized memory and networking chips. SM002, SM003
CM009 Fortune Business Insights sizes the dedicated AI accelerator market (discrete training + inference chips, narrower definition) at approximately $113B–$180B for 2025–2026, growing at 26–27% CAGR through 2034. SM004, SM014
CM010 Deloitte's 2026 semiconductor outlook estimates generative AI chip revenue at approximately $500B for 2026—the broadest definition, including AI-adjacent memory and related silicon—representing approximately half of all chip sales globally. SM003, SM001
CM011 The AI inference market (separate from AI training) is sized at approximately $106B in 2025 growing to $117–120B in 2026 at a 19% CAGR, reaching $255B by 2030 (MarketsandMarkets). SM005, SM017
CM012 By 2026, approximately two-thirds of all AI compute is estimated to be inference-driven, versus one-third in 2023; inference now accounts for over half of all AI cloud spend. SM017, SM012
CM013 The SAM for non-NVIDIA AI accelerators is derived as approximately 20% of the TAM based on NVIDIA's ~80% market share, yielding an addressable market of $40–54B (conservative TAM base) to $96B (IDC base) for AMD, custom silicon, and alternative vendors combined. SM009, SM011
CM014 AMD holds approximately 5–8% of the AI accelerator market in 2025 and is projected to grow to 10–15% by late 2026; custom silicon (Google TPU, AWS Trainium) holds approximately 4–10%, leaving approximately 5% or less for all other alternative vendors including Tenstorrent. SM013, SM009
CM015 Analyst TAM estimates for the AI chip/accelerator market in 2026 diverge by a factor of approximately 2–3× ($113B–$180B vs $268B vs $477B) depending on whether discrete accelerators only, AI processing semiconductors, or all AI data center silicon (including memory and networking) are included. SM001, SM002, SM004
CM016 McKinsey's semiconductor industry research suggests even the largest analyst TAM estimates may undercount captive production and Chinese supplier revenues by 30–40%, implying the true total semiconductor market and AI chip sub-market may be substantially larger than headline analyst figures. SM021, SM003
CM017 The five largest hyperscalers (Amazon, Google, Meta, Microsoft, Oracle) are projected to spend $650–700B on AI infrastructure in 2026, with approximately 70–75% earmarked for AI-specific hardware. SM007, SM009
CM018 Hyperscalers are increasingly building custom AI silicon internally (Google TPU, Amazon Trainium/Inferentia) and are not a near-term commercial hardware customer for Tenstorrent, though RISC-V IP licensing to hyperscaler SoC teams represents a longer-term pathway. SM009, SM025
CM019 The neocloud segment is projected to generate approximately $20B in revenue in 2026, growing to $180B by 2030; neoclouds are actively evaluating non-NVIDIA chips to reduce GPU supply dependency and capture cost advantage in inference workloads. SM010, SM015
CM020 Futurum Group confirmed Galaxy Blackhole deployments with at least five neocloud co-locations as of May 2026, including ai& (Tokyo, flagship deployment), Cirrascale (Seattle), Turium AI (India), Virtu Financial (HFT research), and Prodia (video generation world record). SM019, SM007
CM021 Average large enterprise LLM spend reached $7M per year in 2025, nearly triple the $2.5M level in 2024; most enterprises access AI compute through cloud hyperscalers rather than direct hardware procurement, limiting near-term direct chip procurement opportunity for Tenstorrent. SM017, SM007
CM022 Hyundai, Samsung, and LG Electronics—all strategic investors in Tenstorrent's Series D—represent a high-value enterprise adoption pathway through automotive ADAS, consumer electronics, and semiconductor OEM channels, where Korean sovereign compute preferences create differentiated demand. SM023, SM024
CM023 RISC-V IP licensees—chipmakers and automotive Tier-1 suppliers embedding the Ascalon 64-bit RISC-V core in custom SoCs—represent a distinct buyer segment from Tenstorrent's hardware customers; budget control sits with semiconductor IP procurement or EDA sourcing teams. SM025, SM006
CM024 Geographic AI chip demand in 2026 is concentrated in the United States (hyperscalers), Korea (Samsung, Hyundai, LG ecosystem), Japan (SoftBank, ai& flagship deployment), and cloud-heavy European markets; Asia-Pacific represents over 40% of RISC-V market share. SM006, SM007
CM025 The neocloud segment grew faster than hyperscaler AI in 2025–2026 in percentage terms, though hyperscalers dominate absolute spend; neoclouds' openness to multi-vendor hardware makes them more commercially accessible to Tenstorrent than hyperscalers. SM010, SM015
CM026 GenAI inference demand is the dominant growth driver for AI accelerators in 2026, with inference projected to represent two-thirds of all AI compute globally by end of 2026, versus one-third in 2023; inference economics (ongoing cost per token) differ structurally from training (one-time sunk cost). SM017, SM012
CM027 NVIDIA GPU supply is severely constrained through at least Q1 2027 due to TSMC CoWoS advanced packaging capacity limits and HBM3e memory supply bottlenecks; lead times for cutting-edge GPUs have extended to over a year for many buyers. SM016, SM009
CM028 GPU supply shortage is reshaping buyer behavior: hyperscalers have secured multi-year volume supply agreements, crowding out enterprise buyers who must rely on cloud AI services or evaluate alternatives; this creates structural demand for non-NVIDIA inference hardware. SM016, SM007
CM029 NVIDIA's CUDA software ecosystem represents the primary structural switching cost for AI chip adoption; migrating from CUDA to an alternative platform requires significant software rewrite, retraining of engineering teams, model conversion, and performance re-tuning. SM018, SM009
CM030 Tenstorrent's TT-Forge compiler claims 90% pass rate for models from Hugging Face, targeting the CUDA model-compatibility barrier; however, The Register's independent review found the software stack 'simply isn't polished enough' in November 2025, confirming software maturity as a near-term headwind. SM020, SM019
CM031 Sovereign compute requirements—particularly in Korea, Japan, and the EU—are creating buyer appetite for non-American AI chip alternatives; Tenstorrent's Korean strategic investors (Samsung, Hyundai, LG) and Japan flagship deployment (ai& in Tokyo) reflect this geopolitical tailwind. SM023, SM019
CM032 Power and grid infrastructure is an increasingly binding constraint for data center AI deployments; over 60% of some hyperscaler spending in 2026 goes to power, cooling, and physical infrastructure rather than just chips, making power efficiency a critical differentiator. SM007, SM016
CM033 The RISC-V ecosystem is maturing rapidly through standardization (RVA23 profile), improved toolchains, and growing adoption in AI/ML edge workloads, automotive ADAS, and industrial automation; Asia-Pacific adoption is particularly strong with over 40% regional market share. SM006, SM025
CM034 ARM Holdings remains the dominant incumbent in embedded CPU IP licensing with superior software ecosystem maturity and toolchain depth; Tenstorrent's Ascalon RISC-V core competes as an ARM alternative but faces toolchain gaps that limit immediate enterprise adoption. SM025, SM018
CM035 Tenstorrent's disclosed SOM indicator is approximately $150M in signed contracts as of December 2024; this represents the only public data point for Tenstorrent's market penetration and cannot be directly converted to a defensible SOM without private deployment and revenue data. SM023, SM019
CM036 Analyst AI chip TAM estimates for 2026 diverge by a factor of 2–3× ($113B–$180B vs $268B vs $477B) due to definitional scope differences; Gartner, IDC, Deloitte, and Fortune Business Insights use materially different market boundaries that are not directly comparable. SM001, SM002, SM004
CM037 No publicly confirmed case of a large enterprise buyer switching from NVIDIA to a non-NVIDIA AI chip vendor at production scale has been independently verified as of May 2026; neocloud adoption of alternatives (including Tenstorrent) represents the leading evidence of non-NVIDIA chip scale deployment. SM020, SM012
CM038 No analyst downgrade or AI chip market saturation signal has emerged as of May 2026; all major analyst reports (Gartner, IDC, Deloitte, MarketsandMarkets) project continued strong CAGR through 2030, though the pace of growth and timing of normalization are contested. SM001, SM002, SM005
CP001 NVIDIA holds approximately 80% of the AI accelerator market by revenue in 2026, maintaining dominance driven by the CUDA ecosystem and supply chain advantages. SP013, SP019
CP002 The AI chip competitive landscape in 2026 includes at least five distinct categories: incumbent GPUs, direct AI chip challengers, hyperscaler custom silicon, adjacent incumbents, and status-quo CPU or cloud GPU substitutes. SP001, SP016
CP003 Tenstorrent's primary commercial competitors are NVIDIA (incumbent), AMD MI300X (GPU alternative), Cerebras CS-3 (inference challenger), and Intel Gaudi 3 (price-positioned adjacent), with hyperscaler custom silicon as a structural substitute demand absorber. SP001, SP010
CP004 Google, Amazon, and Meta have collectively deployed over 600,000 custom AI chips internally as of 2026, absorbing significant demand that might otherwise reach commercial AI chip vendors. SP006, SP011
CP005 Intel has publicly shifted its AI strategy away from head-to-head competition with NVIDIA in AI training, redirecting the Gaudi 3 line toward cost-sensitive inference buyers. SP017, SP016
CP006 NVIDIA H100 GPUs are priced at $27K–$40K per unit in 2026 market conditions and rent for $2.00–$14.90/hr on cloud providers, representing the benchmark pricing reference for AI accelerator competition. SP019, SP004
CP007 NVIDIA Blackwell B200 GPU is priced at $30K–$50K per unit and rents for $2.25–$14.24/hr, delivering approximately 5x H100 inference throughput while maintaining CUDA software compatibility. SP005, SP016
CP008 AMD Instinct MI300X offers 192GB HBM3 VRAM at approximately $15K–$20K per GPU (30–50% below H100) with cloud rental at $0.50–$7.86/hr, competing primarily on memory capacity and price for inference workloads. SP020, SP009
CP009 Intel Gaudi 3 is priced approximately 50% below H100 equivalents but has achieved minimal commercial traction in 2026 due to software ecosystem immaturity and limited cloud availability. SP021, SP017
CP010 Cerebras CS-3 wafer-scale engine achieves 1,000–2,000 tokens/sec for large language model inference, representing order-of-magnitude higher throughput than GPU clusters for large batch workloads. SP022, SP007
CP011 Cerebras raised a $1 billion Series H funding round at a $23 billion valuation in February 2026, with OpenAI as anchor customer in a $10B+ multi-year contract for 750 megawatts of AI compute capacity through 2028. SP007, SP008, SP022
CP012 Cerebras reported $510 million in 2025 revenue and filed for a Nasdaq IPO (ticker: CBRS) in April 2026 targeting a $23–26.6 billion public valuation. SP007, SP008
CP013 Groq's Language Processing Unit (LPU) delivers 300+ tokens/sec for Llama-70B inference with deterministic latency; Groq was reportedly involved in NVIDIA acquisition discussions in late 2025, making its independent commercial roadmap uncertain. SP023, SP012
CP014 SambaNova explored a sale in late 2025 after fundraising challenges, with BlackRock marking shares down from a $5 billion peak valuation to approximately $2.4 billion, and Intel reportedly valuing the company at ~$1.6 billion including debt in acquisition discussions. SP014, SP025
CP015 SambaNova raised a $350M+ Series E in February 2026 co-led by Vista Equity Partners and Intel, and pivoted its commercial strategy toward cloud-managed AI inference services to avoid a distressed sale. SP015, SP014
CP016 Google's Trillium (TPU v6e) delivers approximately 926 TFLOPS BF16 per chip with 4.7x compute improvement over TPU v5e and 67%+ energy efficiency gain; over 100,000 chips were deployed by early 2026 for internal and GCP customer workloads. SP006, SP011
CP017 Amazon Trainium3 delivers 2.52 PFLOPS FP8 per chip with 144GB HBM3e and NeuronSwitch fabric; 500,000+ chips are in production by 2025–2026, primarily used by Anthropic and OpenAI through AWS with limited external commercial availability. SP011, SP001
CP018 Meta's MTIA 300 custom silicon (RISC-V chiplet architecture) is in production in early 2026 for internal inference workloads and is not commercially available to third parties. SP011, SP001
CP019 NVIDIA CUDA has more than 4 million registered developers and 40,000+ dependent organizations in 2026, representing the world's largest proprietary AI compute ecosystem and primary source of competitive switching cost moat. SP013, SP019
CP020 Tenstorrent's TT-Metal and TT-Metalium software stacks are MIT licensed and fully open-source, making them unique among major commercial AI chip vendors and eliminating software licensing fees from total cost of ownership. SP024, SP010
CP021 AMD ROCm has matured significantly by 2026 with improving library coverage and growing benchmark parity with CUDA for inference workloads, though it still trails CUDA in developer adoption and ISV integration depth. SP020, SP016
CP022 Tenstorrent Galaxy Blackhole developer reference cards are estimated at approximately $1K per unit — roughly 30x lower acquisition cost vs a single H100 GPU — though server-level and rack-level pricing has not been publicly established at scale as of May 2026. SP024, SP010
CP023 Cerebras CS-3 and Groq LPU are inference-only architectures incapable of model training, while Tenstorrent Galaxy Blackhole supports both training and inference on the same hardware, providing broader workload coverage. SP022, SP023
CP024 Intel Gaudi 3 is priced approximately 50% below NVIDIA H100 equivalents for comparable inference workloads and includes the OneAPI open framework, but commercial adoption is constrained by limited cloud availability and ISV toolchain support. SP021, SP017
CP025 Enterprise AI buyers face significant switching costs when migrating from NVIDIA CUDA: library recompilation, model retuning for new hardware, developer retraining, and software qualification cycles often represent months of engineering effort. SP013, SP016
CP026 Tenstorrent's RISC-V ISA is open and customer-programmable, enabling custom hardware operations without vendor lock-in at the instruction set level — a differentiation not available with NVIDIA, AMD, or Intel GPU architectures. SP010, SP024
CP027 Multi-homing across AI chip vendors is technically feasible via framework abstraction layers such as PyTorch and JAX, but is operationally expensive in practice; most enterprises standardize on a single hardware vendor for production workloads. SP013, SP001
CP028 Sovereign AI programs in Japan (ai& Tokyo Tenstorrent deployment), South Korea (Hyundai Motor Group investment), and Middle East create a structurally motivated non-NVIDIA buyer segment prioritizing supply-chain independence. SP001, SP010
CP029 NVIDIA GPU cluster scale-out relies on InfiniBand networking (NVLink/NVSwitch for intra-node, InfiniBand for inter-rack), creating infrastructure dependency on Mellanox/NVIDIA-compatible networking gear and adding switching cost at the network layer. SP019, SP016
CP030 Tenstorrent's Galaxy Blackhole uses standard Ethernet for scale-out networking rather than InfiniBand, reducing infrastructure dependency and lowering switching costs at the network layer for customers evaluating multi-vendor AI compute strategies. SP024, SP010
CP031 NVIDIA's CUDA moat is self-reinforcing: more developers generate more libraries, which increases the switching cost burden for users, which drives more hardware purchases, which further funds CUDA ecosystem investment — a durable compounding advantage. SP013, SP019
CP032 The AI accelerator market shows early-stage fragmentation at the inference layer in 2025–2026, with Cerebras, Groq, AMD, and Tenstorrent gaining enterprise deployments and demonstrating viable non-NVIDIA alternatives for specific inference workloads. SP001, SP003
CP033 The Register (November 2025) reviewed Tenstorrent's Blackhole QuietBox workstation and documented that the software 'simply isn't polished enough for most local AI enthusiasts,' confirming a concrete software maturity gap versus CUDA. SP026, SP010
CP034 SambaNova's valuation declined from a $5 billion peak in 2021 to approximately $2.4 billion (BlackRock mark, 2025), demonstrating that well-funded AI chip challengers face severe execution and capital risk even with differentiated technology. SP014, SP025
CP035 Cerebras' largest customer anchor — OpenAI via a $10B+ multi-year contract — represents the majority of Cerebras' revenue base, creating significant customer concentration risk in the challenger inference segment. SP007, SP008
CP036 Intel's public retreat from AI training competition, confirmed by executive statements (Wccftech, 2026), leaves Intel Gaudi 3 without a committed large-scale R&D investment path, reducing it as a long-term competitive threat to Tenstorrent. SP017, SP021
CP037 Tenstorrent claims approximately 90% HuggingFace model pass rate on TT-Metal as of 2026; this figure is self-reported by the company and has not been independently verified by third-party benchmarking organizations as of May 2026. SP024, SP010
CP038 Graphcore, acquired by SoftBank in 2023, has minimal commercial competitive activity in 2026; its IPU architecture is no longer actively competing for new enterprise AI workloads at scale. SP001, SP018
CP039 Export control regulations applied by the U.S. government to NVIDIA AI chips in certain geographic markets create market access asymmetry that may benefit alternative vendors including Tenstorrent in sovereign AI programs in affected regions. SP001, SP013
CI001 Tenstorrent's Galaxy Blackhole server system entered general availability on April 28, 2026, with a list price of $110,000 per 6U chassis containing 32 Blackhole RISC-V ASICs delivering 23 PFLOPS of FP8 compute. SI001, SI002, SI003
CI002 The Galaxy Blackhole Supercluster configuration (four chassis) is priced at $440,000, providing 92 PFLOPS FP8 in a scalable networked-AI architecture supporting clusters up to 144 nodes (4,608 Blackhole chips). SI001, SI004, SI006
CI003 Entry-level Tenstorrent Blackhole P100 inference cards are priced starting at approximately $999 and the QuietBox workstation starts at approximately $9,999, targeting individual developers and inference workstation buyers. SI001, SI014
CI004 Tenstorrent derives IP licensing revenue from RISC-V CPU and Tensix NPU royalties paid by automotive and edge OEM partners including Samsung and Hyundai, though the royalty rate, committed volume, and total recognized licensing revenue are not publicly disclosed. SI014, SI016, SI017
CI005 Tenstorrent offers cloud-based hardware access through the Koyeb serverless platform, enabling pay-as-you-go usage of Blackhole compute instances; the revenue-sharing arrangement and Tenstorrent's net HaaS revenue are not disclosed. SI013, SI015, SI016
CI006 Tenstorrent's DevCloud developer program provides subsidized or free access to Wormhole hardware for external developers using TT-Metal and TT-Forge, serving as a top-of-funnel developer acquisition strategy rather than a current revenue line. SI016, SI015
CI007 Tenstorrent claims the Galaxy Blackhole system delivers inference at $6 per million tokens on DeepSeek-R1-0528 671B workloads (Blitz mode), compared to approximately $30 per million tokens on NVIDIA GB300, representing a claimed 5× total-cost-of-inference advantage. SI001, SI007, SI021
CI008 The Galaxy Blackhole system achieves over 350 tokens per second per user and sub-4-second time-to-first-token on 100,000-token context prompts, according to Tenstorrent's own benchmarks published at the April 2026 general-availability launch. SI001, SI002, SI020
CI009 Tenstorrent produces 720p 81-frame video in 2.4 seconds on a Galaxy Blackhole system in partnership with Prodia, showcasing GPU-class video generation throughput; this benchmark is company-reported and not independently corroborated. SI001, SI003
CI010 RISC-V and Tensix IP licensing agreements with Samsung and Hyundai are multi-year royalty arrangements enabling the licensees to embed Tenstorrent architecture in their own SoC products; specific royalty rates and minimum commitments are under non-disclosure agreements. SI016, SI017, SI014
CI011 Tenstorrent's go-to-market motion for Galaxy hardware is primarily direct enterprise sales with cloud-service-provider seeding (ai& in Tokyo, Cirrascale, Turium AI), implying a long-cycle enterprise sales model rather than a high-velocity product-led growth motion. SI014, SI015, SI006
CI012 Tenstorrent's Galaxy Blackhole gross margin is estimated at 36%–55% per server chassis based on TSMC N4 wafer costs for 32 Blackhole ASICs, GDDR6 memory, assembly, and test against the $110,000 list price; this estimate has very low confidence because actual COGS is undisclosed. SI014, SI024
CI013 Total non-recurring engineering cost for the Blackhole chip family (including N4 tape-out mask sets, engineering lots, and verification) is estimated at $50 million–$100 million or more; this NRE is amortized over production volume and compresses realized gross margin on early production runs. SI022, SI024
CI014 With approximately 1,100–1,200 employees as of early 2026, Tenstorrent's fully loaded personnel cost is estimated at $220–$300 million annually at competitive semiconductor-industry compensation, representing the largest single operating expense category. SI009, SI012
CI015 Total estimated monthly operating burn for Tenstorrent is $25–$50 million, incorporating headcount cost, TSMC production commitments, EDA licensing, R&D, facilities, and G&A; this estimate spans a 2× range reflecting the absence of disclosed financials. SI009, SI024
CI016 IP licensing royalties for RISC-V and Tensix NPU are estimated to carry gross margins of 80%–95% if structured as per-chip royalties, which would materially improve Tenstorrent's blended gross margin as licensing volume scales; actual licensing margins are undisclosed. SI016, SI014
CI017 Customer acquisition cost for enterprise AI infrastructure hardware is not disclosed by Tenstorrent; industry norms suggest enterprise hardware CAC ranges $50,000–$500,000 per customer depending on sales-cycle length, engineering support, and evaluation-unit costs. SI015, SI011
CI018 The $6 per million token cost claim for Galaxy Blackhole versus $30 per million tokens on NVIDIA GB300 is a company-calculated total-cost-of-inference figure whose methodology—including assumed utilization rate, batch size, memory configuration, and concurrent-user count—has not been independently verified as of May 2026. SI001, SI019, SI007
CI019 Tenstorrent completed a Series D funding round of $693 million at a $2.6 billion post-money valuation in December 2024, led by Samsung Securities, LG Technology Ventures, and Fidelity Management and Research among others. SI022, SI010, SI023
CI020 Tenstorrent closed a Series E round of approximately $800 million at a $3.2 billion post-money valuation in November 2025, led by Fidelity Management and Research, bringing total disclosed funding to approximately $1.99 billion. SI008, SI009, SI010
CI021 Tenstorrent's total capital raised across all disclosed rounds is approximately $1.99 billion (Series D $693M + Series E $800M + earlier rounds including the $100M Hyundai/Samsung round in 2023 and preceding rounds), positioning it as one of the best-capitalized AI chip startups outside NVIDIA. SI010, SI022, SI009
CI022 Estimated cash on hand for Tenstorrent as of May 2026 is approximately $1.0–$1.5 billion, derived from the November 2025 Series E close minus approximately six months of estimated operating burn ($150–$300M at $25–$50M/month); actual cash position is undisclosed. SI009, SI015
CI023 Tenstorrent's stated use of Series E capital includes accelerating Blackhole production scale-up, designing and taping out the next-generation chip after Blackhole, expanding go-to-market capacity, and growing IP licensing partnerships globally. SI008, SI016
CI024 No evidence of debt financing, convertible notes, or project-finance obligations for Tenstorrent has been disclosed in public filings, press releases, or analyst reports as of May 2026; the company appears to be purely equity-financed. SI022, SI010
CI025 Tenstorrent has not disclosed any audited revenue figure since the company's 2021 Canadian corporate filing; all post-2021 revenue figures attributed to Tenstorrent in the public domain are algorithmic model estimates from services such as Latka and Growjo, and carry very low reliability for diligence purposes. SI011, SI024, SI019
CI026 Getlatka.com's algorithmic model estimates Tenstorrent's 2025 revenue at approximately $501.6 million; this estimate is derived from employee count, funding history, and comparable companies—not from disclosed financials—and should not be treated as verified revenue for investment decision purposes. SI011, SI012
CI027 Tenstorrent has not disclosed gross margin, COGS, or operating expense breakdown in any public document; without audited financial statements, investors cannot independently assess the company's unit economics viability or path to profitability. SI024, SI019, SI014
CI028 Customer revenue concentration risk is entirely opaque for Tenstorrent; no customer count, largest-customer revenue percentage, or contract duration data is disclosed, creating potential for single-customer-departure material risk. SI024, SI014
CI029 Tenstorrent signed approximately $150 million in customer contracts as of the December 2024 Series D close; these are pre-delivery commitments and represent pipeline backlog, not recognized revenue, creating a gap between signed commitments and recognized revenue that is undisclosed. SI023, SI022
CI030 The gap between Tenstorrent's total capital raised (~$1.99B) and its $150M in signed contracts (as of Dec 2024) implies a capital-to-backlog ratio of approximately 13×, which is unusually high for a hardware company and reflects the pre-revenue phase of Galaxy Blackhole commercialization. SI023, SI010, SI022
CI031 Bear-case FY2025 revenue for Tenstorrent is estimated at $100–$200 million (primarily early IP licensing and inference-card sales); base case is $200–$500 million; bull case (Latka model top end) is $500–$600 million; all figures are estimates without audited basis. SI011, SI012, SI024
CI032 FY2026 annualized revenue run rate for Tenstorrent is estimated at $150 million–$1.2 billion, driven primarily by the pace of Galaxy Blackhole server shipments beginning in April–May 2026; the wide range reflects uncertainty in hardware ramp velocity and IP licensing deal cadence. SI001, SI009, SI011
CI033 Galaxy Blackhole's FP8 compute architecture and GDDR6 memory configuration are optimized for LLM inference; it is not positioned for large-scale training (which requires HBM memory and high-bandwidth interconnect at the scale NVIDIA HGX provides), limiting total addressable revenue to the inference subset of the AI hardware market. SI002, SI006, SI019
CI034 The Galaxy Blackhole Supercluster scales to 144 nodes (4,608 ASICs) via a 100 Tbps mesh network, enabling large-context LLM inference at scale comparable to Google TPU pods or Amazon Trainium2 clusters, potentially addressing hyperscaler-adjacent inference workloads. SI001, SI003, SI020
CI035 Tenstorrent's revenue model path to profitability requires a significant ramp of Galaxy hardware revenue (hundreds of servers per quarter) or material IP licensing wins; at the estimated $25–$50M/month burn rate, break-even on hardware alone would require annual revenue of $300–$600M at a 40% gross margin. SI009, SI024
CI036 Estimated residual cash on hand for Tenstorrent as of May 2026, after 18 months of estimated $37M/month average burn from mid-2024 through May 2026, is approximately $1.3 billion; this estimate has wide uncertainty and is subject to actual burn rate, hardware production timing, and revenue receipts. SI009, SI010
CE001 The Blackhole ASIC is fabricated on TSMC's 6 nm process node, with a die area of approximately 600 mm². SE001, SE003, SE011, SE013
CE002 Each Blackhole chip contains 120 Tensix processing tiles arranged in a dataflow architecture. SE003, SE013, SE024
CE003 Blackhole integrates 16 application-class RISC-V cores based on the SiFive X280 64-bit design, capable of running Linux. SE003, SE013, SE014
CE004 Each Blackhole chip includes 180 MB of on-chip SRAM distributed across the Tensix tile array. SE003, SE013
CE005 The Blackhole p100a card has 28 GB GDDR6 at 448 GB/s; the p150a and p150b have 32 GB GDDR6 at 512 GB/s. SE001, SE003, SE011
CE006 Peak compute per Blackhole chip is 664 TFLOPS in BlockFP8 format and 332 TFLOPS in BF16. SE001, SE003, SE011, SE013
CE007 The Blackhole p150a and p150b carry four QSFP-DD 800 Gbps Ethernet ports, providing 3.2 Tbps of aggregate card-to-card bandwidth. SE001, SE003, SE024
CE008 All Blackhole cards connect to the host via PCIe Gen5 ×16 and have a 300 W thermal design power. SE001, SE003, SE011
CE009 The Galaxy Blackhole server integrates 32 Blackhole chips, delivers 23 PFLOPS FP8, includes 1 TB of aggregate GDDR6, and uses a 100 Tbps internal mesh network. SE001, SE019, SE022
CE010 All Tenstorrent software—TT-Metal, TT-Metalium, TT-NN, TT-Forge, and TT-LLK—is released under the Apache 2.0 open-source license. SE002, SE006, SE009, SE014
CE011 TT-Metal is Tenstorrent's low-level kernel programming API, providing direct access to Tensix tile execution analogous to CUDA for NVIDIA GPUs. SE002, SE006, SE012, SE014
CE012 TT-Metalium is the core runtime and dispatch layer that manages kernel scheduling, buffer management, and multi-device orchestration on Blackhole hardware. SE002, SE012, SE014
CE013 TT-NN provides a Python-accessible operator library with more than 200 compute primitives compatible with HuggingFace Transformers-style model APIs. SE002, SE006
CE014 TT-Forge is an MLIR-based compiler with three framework front-ends: TT-Torch for PyTorch, TT-XLA for JAX, and TT-Forge-ONNX for ONNX models. SE005, SE009, SE014
CE015 As of April 2026, the tenstorrent/tt-metal GitHub repository had approximately 1,410 stars, 429 forks, 25,830+ commits, and 161 official releases. SE006, SE010
CE016 Tenstorrent claims a 90% pass rate on HuggingFace model benchmarks, enabling compatibility with more than 2.5 million open-source models. SE002, SE022
CE017 Tenstorrent claims support for over 2.5 million open-source ML models through TT-Forge and TT-NN compatibility layers. SE001, SE002
CE018 Tenstorrent's Blackhole hardware targets three distinct form factors: p100a for desktop inference, p150a for workstation and Ethernet-chained clustering, and p150b for passive-cooled rack deployment. SE001, SE022, SE024
CE019 Tenstorrent announced general availability of the Galaxy Blackhole server on April 28, 2026, with volume production and customer shipments beginning at that time. SE001, SE019, SE022
CE020 The Tenstorrent QuietBox developer workstation is priced at approximately $9,999 for the base configuration. SE001, SE016, SE017
CE021 The standard customer deployment workflow is: load a PyTorch/JAX/ONNX model, compile with TT-Forge, dispatch via TT-Metalium, and run inference on Blackhole hardware. SE002, SE005, SE012
CE022 The Register's November 2025 review of the Blackhole QuietBox workstation found that the TT-Metal software 'simply isn't polished enough for most local AI enthusiasts.' SE016, SE017
CE023 Tenstorrent's Blackhole hardware does not support ML training workloads; the entire product line is focused exclusively on inference. SE001, SE002, SE017
CE024 All Blackhole ASICs are fabricated exclusively at TSMC on the 6nm node, creating a sole-source foundry dependency. SE003, SE011, SE013
CE025 Blackhole GDDR6 memory is sourced from multiple DRAM vendors including Samsung, SK Hynix, and Micron, providing some supply-chain diversification. SE011, SE017
CE026 The Blackhole p150b uses passive cooling, enabling higher rack density for server environments without liquid cooling infrastructure. SE001, SE003
CE027 Tenstorrent's Galaxy platform is designed to scale to 144-node clusters containing 4,608 Blackhole chips for exascale inference workloads. SE001, SE019
CE028 Each Tensix tile contains five embedded baby RISC-V cores dedicated to compute orchestration, data movement, and control, totaling more than 600 RISC-V cores per Blackhole chip. SE003, SE013, SE014
CE029 The baby RISC-V cores in each Tensix tile handle three distinct functions: compute kernel dispatch, data movement between tiles, and on-tile control logic. SE004, SE013
CE030 The 16 application-class RISC-V cores per Blackhole chip run Linux and host management software, enabling a firmware-free management model. SE003, SE004, SE013
CE031 Tenstorrent has targeted TT-Forge v1.0 compiler stability for 2026, currently in active development as of May 2026. SE005, SE009
CE032 A next-generation AI chip beyond Blackhole is under active development at Tenstorrent under NDA, with no public specifications or tape-out date disclosed. SE020, SE022
CE033 Blackhole accelerator cards entered volume production in May 2026, enabling Tenstorrent to begin fulfilling its signed contract backlog. SE001, SE019, SE022
CE034 Tenstorrent uses contract manufacturers for PCB assembly and system integration of Blackhole hardware products. SE017, SE023
CE035 Pyron is a third-party SDK documentation platform (docs.pyron.dev) providing a higher-level abstraction over TT-Metalium for enterprise NPU integrators. SE015
CE036 The Blackhole ASIC has a die area of approximately 600 mm², consistent with a high-performance compute chip at TSMC 6nm. SE011, SE013
CE037 As of April 2026, the tenstorrent/tt-metal repository had 988 open pull requests and 19,076 merged pull requests. SE006, SE007
CE038 As of April 2026, the tenstorrent/tt-metal repository had 3,488 open issues, indicating a meaningful backlog of known software gaps. SE007, SE008
CE039 TT-Forge-ONNX enables import of ONNX-format models, expanding compatibility to any framework that can export to the ONNX interchange format. SE005, SE009, SE014
CE040 Tenstorrent's DevCloud service provides developers with remote SSH access to Wormhole and Blackhole hardware without requiring on-premises capital expenditure. SE002, SE022
CU001 Tenstorrent's customer base as of May 2026 spans six distinct archetypes: independent developers (DevCloud), strategic investor-customers (LG, Hyundai, SoftBank), cloud HaaS partners (Koyeb), national infrastructure operators, academic research institutions, and prospective government/defense buyers. SU001, SU002, SU005, SU008
CU002 Tenstorrent's developer community spans North America, Europe (Germany via Fraunhofer), South Korea (LG, Hyundai), and Japan (SoftBank), with no disclosed customer concentration in emerging markets. SU001, SU008, SU022
CU003 Koyeb, a French cloud startup, publicly confirmed a production deployment of Tenstorrent Blackhole p150 hardware on its cloud HaaS platform, billing customers per token/second—the only confirmed arms-length commercial customer proof. SU005, SU006
CU004 LG AI Research, backed by LG Technology Ventures (Series D lead investor), is using Tenstorrent chips across both Wormhole and Blackhole generations for AI inference and training R&D workloads—a dual investor-customer relationship. SU001, SU002, SU007
CU005 Hyundai Motor Group, a strategic investor in Tenstorrent's Series D, is using Tenstorrent hardware for automotive AI and on-vehicle ADAS compute workloads, representing a pilot-stage deployment in the automotive vertical. SU001, SU002, SU010
CU006 SoftBank disclosed a deal with Tenstorrent to supply AI chips for its Japanese data center expansion, announced in association with SoftBank's participation in the Series D funding round in January 2025. SU001, SU025
CU007 Enterprise customers purchase Tenstorrent hardware exclusively through direct sales or Koyeb's cloud layer; no traditional OEM, VAR, or channel distribution arrangement has been announced. SU001, SU003, SU005
CU008 No Tenstorrent SMB or mid-market customer has been disclosed; the Galaxy server's $110,000 list price and direct-sales model effectively confine the commercial customer base to enterprises, research institutions, and cloud providers. SU003, SU011, SU019
CU009 Tenstorrent reported approximately 5,000 registered DevCloud developer accounts as of Q1 2026; this metric counts sign-ups rather than active monthly users and has not been independently verified. SU014, SU015
CU010 The Galaxy Blackhole server reached general availability on April 28, 2026—approximately two weeks before this report's reference date—making retention, NRR, and repeat purchase data unavailable for the flagship commercial product. SU003, SU004, SU011
CU011 The tenstorrent/tt-metal GitHub repository recorded more than 25,830 commits, 1,410+ stars, and 19,076 merged pull requests as of April 2026, serving as a proxy for developer community engagement and adoption depth. SU016, SU008, SU003
CU012 Tenstorrent claims 90% pass rate on HuggingFace model benchmarks and compatibility with more than 2.5 million open-source models; neither figure has been independently verified as of May 2026. SU014, SU016
CU013 Academic research groups at MIT, Stanford, and Carnegie Mellon University were confirmed as Tenstorrent hardware evaluators via Developer Day coverage; these represent evaluation-stage, not production, deployments. SU008, SU022
CU014 Trade press reported hyperscaler-tier interest in the Galaxy Blackhole server around its GA launch in April 2026, though no named hyperscaler customer has been publicly disclosed. SU003, SU011, SU019
CU015 Tenstorrent's developer-to-enterprise conversion rate is unknown; no public data exists on how many DevCloud registrants have subsequently purchased hardware or contracted for Galaxy server deployments. SU014, SU015
CU016 Fraunhofer Institute was confirmed as an early Tenstorrent Wormhole adopter for European AI research, representing an arms-length, non-investor research customer with multi-cycle engagement. SU008, SU022
CU017 Tenstorrent devkit and hardware pricing spans $999 (p100a) to $110,000 (Galaxy Blackhole 6U server), creating a broad funnel from individual developer entry to enterprise cluster deployment. SU003, SU011, SU014
CU018 Jaguar Land Rover has been mentioned in trade press as a potential automotive AI compute partner for Tenstorrent, but no confirmed deployment or signed agreement has been announced as of May 2026. SU010, SU019
CU019 LG AI Research's engagement spans at least two Tenstorrent hardware generations (Wormhole and Blackhole), indicating a multi-year strategic commitment rather than a one-time trial. SU001, SU007
CU020 Koyeb's production HaaS deployment of Blackhole p150 constitutes a durable integration decision: cloud providers typically incur significant operational cost when onboarding new hardware backends, reducing likelihood of rapid churn. SU005, SU006
CU021 The named customer reference set—as of May 2026—comprises one arms-length HaaS deployer (Koyeb), three investor-aligned strategic partners (LG, Hyundai, SoftBank), one academic-industrial research user (Fraunhofer), and several academic evaluators (MIT/Stanford/CMU). SU001, SU005, SU008, SU022
CU022 No public net revenue retention (NRR), gross revenue retention (GRR), contract renewal, or cohort churn data has been disclosed for Tenstorrent as of May 2026. SU014, SU019
CU023 A November 2025 independent review by The Register assessed Tenstorrent's software stack as 'not polished enough for most local AI enthusiasts,' citing configuration friction and incomplete driver documentation as key barriers. SU017, SU024
CU024 The tenstorrent/tt-metal GitHub repository had 3,488 open issues as of April 2026, indicating a meaningful software backlog that represents a retention and adoption risk for the non-captive developer segment. SU016, SU017
CU025 Phoronix's review of the Tenstorrent Blackhole p150a on Linux found functional hardware performance but noted software rough edges, consistent with The Register's assessment of software immaturity. SU021, SU017
CU026 No NPS scores, customer satisfaction surveys, support ticket data, or CSAT metrics have been disclosed publicly for Tenstorrent, making independent satisfaction assessment impossible beyond review-based proxies. SU014, SU019
CU027 Fraunhofer Institute's multi-cycle evaluation of Tenstorrent hardware across Wormhole and potentially Blackhole generations provides a positive academic retention signal, though academic repeat use is a weaker signal than commercial renewal. SU022, SU008
CU028 Three of Tenstorrent's five most-visible commercial customers—LG AI Research, Hyundai Motor Group, and SoftBank—are also equity investors from the Series D round, creating concentration risk correlated with both revenue and equity performance. SU001, SU002, SU010
CU029 Tenstorrent has not disclosed any case of a customer expanding from an initial purchase to a follow-on or add-on order; the land-and-expand model is structurally available (Galaxy modular architecture) but undemonstrated in practice. SU003, SU011, SU019
CU030 NVIDIA's installed CUDA ecosystem creates high switching costs for enterprise prospects already operating H100/H200 clusters; Tenstorrent's TT-Metalium offers an open-source alternative but is functionally less mature. SU019, SU020, SU017
CU031 US Department of Defense procurement interest in domestic AI chips under the CHIPS Act represents a prospective alternative channel for Tenstorrent, though no contract or RFP outcome has been disclosed. SU013, SU019
CU032 Tenstorrent's geographic customer concentration is high: three of five named commercial customers (LG, Hyundai, SoftBank) are headquartered in South Korea or Japan, while Koyeb (France) and DevCloud (global) provide partial diversification. SU001, SU005, SU002
CU033 The dual investor-customer relationship with LG, Hyundai, and SoftBank may suppress honest product feedback and skew roadmap decisions toward investor preferences rather than broad market signals—a governance risk not visible from public filings. SU001, SU002
CU034 Tenstorrent's Galaxy Blackhole achieved 350 tokens/second on DeepSeek-R1 benchmark at launch, providing a performance proof point that supports the customer value proposition for LLM inference workloads. SU020, SU003
CU035 Cloud provider Koyeb prices Tenstorrent Blackhole inference on a per-token/second basis, enabling pay-as-you-go customer acquisition that lowers the barrier to adoption compared to direct Galaxy server procurement at $110,000. SU005, SU006
CU036 Tenstorrent's open-source Apache 2.0 licensing strategy (TT-Metal, TT-Forge) reduces adoption friction for developer-segment customers but limits the company's ability to capture value from software without a hardware anchor. SU016, SU013
CU037 ServeTheHome's hardware evaluation of Tenstorrent's Blackhole platform found the hardware technically sound but noted the software stack as the primary adoption friction point, corroborating The Register's independent assessment. SU024, SU017
CU038 IEEE Xplore search results confirm that academic researchers are publishing papers citing Tenstorrent hardware, indicating adoption by the research community beyond Tenstorrent's official academic partners. SU026
CR001 The US BIS October 2023 Advanced Computing rule (Federal Register document 2023-25073, effective December 2023) establishes FLOPS and interconnect bandwidth thresholds for export-controlled advanced computing items, imposing license requirements for shipments to China, Russia, and other restricted destinations. SR001, SR002, SR011
CR002 Tenstorrent has not publicly disclosed whether Blackhole's 664 TFLOPS FP8 performance rating exceeds BIS advanced-computing thresholds, has obtained any export licenses, or performs end-use screening of distributors and customers. SR001, SR005, SR007
CR003 A violation of the Export Administration Regulations (EAR) carries criminal penalties of up to $1 million per violation and up to 20 years imprisonment, plus civil fines of up to $364,992 per violation — representing catastrophic legal and financial exposure. SR001, SR002
CR004 No enforcement action by BIS, OFAC, or any EU regulatory body against Tenstorrent has been identified in public records, court filings, or regulatory databases as of May 2026. SR002, SR004, SR026
CR005 The EU AI Act (Regulation 2024/1689), enacted August 2024 with phased compliance timelines, establishes risk-tier classifications for AI systems and may impose conformity assessment, transparency, and documentation requirements on providers of purpose-built AI accelerator hardware sold into EU markets. SR003, SR025, SR026
CR006 US lawmakers debated restricting RISC-V IP licensing to Chinese companies in 2024 as part of broader AI chip export-control discussions; while no blanket ban was enacted, the regulatory uncertainty creates legal exposure for Tenstorrent's Beijing office and its RISC-V-based IP licensing activities. SR006, SR011, SR024
CR007 NVIDIA holds more than 10,000 AI and GPU-related patents; no patent infringement litigation against Tenstorrent has been filed in public court records as of May 2026, but the overhang is material for any AI chip startup operating in adjacent IP space. SR004, SR007, SR023
CR008 Synopsys and Cadence maintain a near-duopoly in electronic design automation (EDA) tools; Tenstorrent's chip design workflow depends on these tools, and any license termination or pricing restructuring would halt next-generation chip design activities. SR005, SR006, SR018
CR009 Blackhole is fabricated exclusively on TSMC's 6nm process node; Tenstorrent has no disclosed secondary fab qualification with Samsung Foundry, Intel Foundry, or GlobalFoundries for leading-edge nodes, creating 100% TSMC sole-source dependency. SR005, SR013, SR018, SR032
CR010 A Taiwan Strait military conflict or major TSMC operational disruption (earthquake, contamination, fire) would halt Tenstorrent production entirely; no disclosed mitigation plan or inventory buffer provides meaningful downside protection. SR005, SR013, SR025
CR011 The Register's November 2025 review of the Blackhole QuietBox concluded that Tenstorrent's software stack was 'simply not polished enough for most local AI enthusiasts,' representing the most cited independent adverse assessment of Tenstorrent's product maturity. SR014, SR029, SR016
CR012 As of April 2026, Tenstorrent's tt-metal GitHub repository had 3,488 open issues and 988 open pull requests, indicating a significant unresolved bug backlog and review bandwidth constraints that signal software maturity risk. SR015, SR014, SR029
CR013 Tenstorrent has not reported any hardware product recalls, customer-reported hardware defects, or safety incidents related to Blackhole or Wormhole products; the only operational failures cited in public record are software quality shortcomings. SR014, SR017, SR029
CR014 Blackhole's next-generation successor chip design cycle is approximately 18-24 months; given the Blackhole tape-out timeline, design decisions for the successor are effectively locked by late 2026, leaving no mid-cycle correction if Blackhole underperforms commercially. SR005, SR018, SR023
CR015 GDDR6 memory supply for Blackhole depends on Samsung, SK Hynix, and Micron; while GDDR6 is less constrained than HBM (which is prioritized for NVIDIA and AMD), price spikes or allocation cuts in a memory cycle downturn could adversely affect Blackhole margins. SR005, SR013, SR025
CR016 Galaxy Blackhole server's OEM qualification with major cloud providers (AWS, Azure, Google Cloud) has not been publicly announced as of May 2026; enterprise hardware qualification cycles typically take 6-12 months, limiting near-term hyperscaler revenue. SR019, SR018, SR029
CR017 NVIDIA's CUDA ecosystem represents a 20-year moat of libraries, frameworks, and developer tooling; migrating AI workloads to Tenstorrent's TT-Forge/TT-Metal stack requires re-instrumentation, numerical re-validation, and re-training of operations teams — creating enterprise switching costs that suppress conversion. SR007, SR014, SR027, SR023
CR018 Tenstorrent has explicitly positioned Blackhole as an inference accelerator and conceded the AI training market to NVIDIA; this inference focus limits total addressable market share and creates structural exposure if NVIDIA or AMD closes the inference cost gap. SR007, SR022, SR027
CR019 Tenstorrent's inference-market competitors include Groq (LPU architecture), Cerebras (wafer-scale), SambaNova, AMD MI300X, and Google TPU v5 — each offering an alternative to NVIDIA for inference workloads, fragmenting the non-NVIDIA inference opportunity. SR018, SR023, SR032
CR020 Intel Gaudi 3 (from Habana Labs acquisition) provides hyperscalers with an alternative AI inference accelerator at potentially lower cost than NVIDIA, further increasing the competitive field Tenstorrent must navigate. SR018, SR023
CR021 No audited financial data or GAAP revenue disclosure has been made by Tenstorrent; the most recent disclosed revenue data is from a 2021 Canadian corporate filing showing $25M-$100M in revenue, which is materially outdated for current diligence purposes. SR020, SR021, SR023
CR022 The $150M in signed contracts disclosed at the December 2024 Series D represents backlog, not recognized revenue; conversion from backlog to cash depends on acceptance testing, delivery milestones, and customer sign-off — conversion rate is undisclosed. SR009, SR020, SR023
CR023 Tenstorrent's estimated burn rate of $25-50M per month implies a runway of approximately 16-32 months from the November 2025 Series E ($800M); the wide range reflects uncertainty about headcount, TSMC wafer payments, and R&D expenditure. SR020, SR021, SR030
CR024 The next-generation AI chip tape-out at TSMC is estimated at $150-300M for leading-edge nodes, which would consume a material portion of Tenstorrent's cash reserves from the combined Series D and Series E and compress financial runway. SR018, SR020, SR021
CR025 No path to profitability has been publicly articulated by Tenstorrent management, and no financial metric indicating a near-term breakeven trajectory — such as revenue guidance, margin targets, or operating leverage milestones — appears in public communications. SR009, SR021, SR023
CR026 LG Electronics, Hyundai Motor Group, and SoftBank are simultaneously the largest known commercial customers and among the largest strategic investors in Tenstorrent, creating circular dependency: an investor exit simultaneously triggers revenue loss and investor confidence shock. SR008, SR010, SR012, SR028
CR027 Enterprise hardware deals typically run Net-60 to Net-90 payment terms, while TSMC wafer payments are advance or short-term credit, compressing Tenstorrent's cash conversion cycle and creating structural working capital pressure. SR018, SR021
CR028 Gross margins on custom AI hardware are structurally compressed: TSMC wafer costs, GDDR6 DRAM, PCB assembly, and logistics consume a large portion of revenue before R&D, sales, and G&A allocation; Tenstorrent has not disclosed hardware gross margin. SR005, SR018, SR021
CR029 Tenstorrent is reported to have been in talks to raise an $800M Series E at a $3.2B valuation led by Fidelity as of November 2025; if completed, this represents continued access to growth capital but also implies continued cash consumption. SR030, SR020, SR023
CR030 No written supply agreements, TSMC capacity commitments, or Samsung/SK Hynix memory purchase agreements have been publicly disclosed by Tenstorrent, leaving supply-chain reliability unverifiable from public sources. SR005, SR013, SR018
CR031 Algorithmic third-party revenue estimates for Tenstorrent of approximately $501M for FY2025 are based on model-derived figures, not audited data, and are materially unreliable for investment underwriting purposes. SR020, SR021
CR032 Jim Keller's track record includes tenures of approximately 2 years at Intel (2017-2018), 2 years at Tesla (2016-2018), and 4 years at Apple (2008-2012), creating a pattern of high-impact but time-limited engagements that investors must underwrite. SR007, SR023, SR027
CR033 No succession plan for Jim Keller has been publicly disclosed by Tenstorrent's board or management; the absence of a disclosed successor elevates the impact severity of a Keller departure to near-critical. SR007, SR023
CR034 Tenstorrent's estimated 1,100-1,200 global employees (mid-2026) compete for the same chip design talent pool as Apple Silicon, Google TPU, and NVIDIA's custom silicon team, elevating engineering talent acquisition and retention risk. SR007, SR019, SR023
CR035 Engineers recruited from Intel and AMD bring non-compete and trade-secret litigation risk; while no such cases have been filed against Tenstorrent, the risk is a common and material exposure in semiconductor IP-intensive companies. SR004, SR007, SR023
CR036 The open-source TT-Metal/TT-Metalium stack (MIT-licensed) creates a tension between customer adoption benefit (lower integration friction) and security exposure (public zero-day vulnerability window before patches are issued). SR015, SR027, SR007
CR037 Tenstorrent's IP licensing revenue from Tensix and Ascalon RISC-V CPU IP provides a non-hardware revenue stream that partially diversifies financial risk, though the scale and concentration of licensees have not been publicly disclosed. SR013, SR023, SR027
CR038 The DevCloud freemium model (approximately 5,000 registered accounts as of Q1 2026) reduces initial customer acquisition friction and serves as a conversion funnel, but the conversion rate from free to paid has not been publicly disclosed. SR019, SR022, SR023
CR039 A BIS enforcement action or export license denial for Blackhole in any currently served market constitutes the highest-severity single-event thesis-break trigger, warranting immediate investment hold and legal review within 30 days. SR001, SR002, SR011
CR040 Jim Keller's departure within 18 months of the last funding round would constitute a thesis-break event, likely triggering a sharp reduction in investor confidence and potentially triggering material adverse change clauses in any debt instruments. SR007, SR023, SR032
CR041 Runway below 6 months without a committed term sheet for the next financing round constitutes a financial thesis-break trigger; at the estimated burn rate of $25-50M/month, this threshold would be breached if fundraising stalls 10-28 months after the Series E close. SR020, SR021, SR030
CR042 Tenstorrent's risk heatmap places burn rate and CUDA ecosystem lock-in in the 'Very High Likelihood, High-to-Critical Impact' quadrant, with TSMC disruption and Jim Keller departure in the 'Medium Likelihood, Critical Impact' quadrant. SR018, SR023, SR032
CR043 The risk transmission chain from BIS export enforcement → revenue loss → investor confidence collapse → capital cost spike → valuation compression represents the highest-velocity downside scenario for Tenstorrent given the unconfirmed compliance status of Blackhole. SR001, SR002, SR011, SR026
CR044 Tenstorrent's external dependency graph reveals at least five critical single-point dependencies (TSMC, EDA tools, BIS compliance, Jim Keller, and LG/Hyundai/SoftBank combined) — any one of which could individually halt operations or collapse the investment thesis. SR005, SR008, SR018, SR032
CR045 Geographic diversification across Toronto, Austin, Belgrade, Tokyo, Bengaluru, Seoul, and Munich provides some talent-pool and operational resilience, but does not mitigate the primary hardware supply-chain concentration in TSMC Taiwan. SR007, SR019, SR027
CV001 Tenstorrent's post-money valuation is $3.2 billion following the November 2025 Series E round of $800 million led by Fidelity, representing a 23% step-up from the $2.6 billion Series D valuation (December 2024) without disclosed revenue milestones. SV005, SV016, SV024, SV025
CV002 The Series E round ($800 million, November 2025) was led by Fidelity; the Series D ($693 million, December 2024) was led by Samsung Securities with participation from LG Technology Ventures and Hyundai Motor Group. SV005, SV017, SV028
CV003 Tenstorrent has raised approximately $1.99 billion in cumulative capital across six rounds: Seed (~$10M, 2019), Series A ($40M, 2021), Series B ($100M, 2022), Series C ($235M, 2023), Series D ($693M, 2024), Series E ($800M, 2025). SV015, SV026, SV030
CV004 Tenstorrent's investor-customer duality — LG Electronics, Hyundai Motor Group, and SoftBank are simultaneously major investors and the largest disclosed commercial customers — creates circular dependency risk absent in non-strategic investor-funded peers. SV015, SV026, SV017
CV005 The AI accelerator market is projected by multiple independent analyst sources to reach approximately $170 billion in annual revenue by 2030, with the inference segment growing faster than training through the forecast period. SV006, SV007, SV008, SV010
CV006 NVIDIA retained approximately 80–85% of the AI accelerator market share in Q1 2026 per TrendForce analysis; alternative AI chip vendors collectively hold 15–20% including AMD, Intel Gaudi, and custom ASICs. SV013, SV021
CV007 NVIDIA's total market capitalization is approximately $3 trillion as of May 2026; FY2025 data center revenue was approximately $115 billion per SEC 10-K filing — establishing the dominant market leader benchmark for the AI chip sector. SV001, SV011
CV008 Arm Holdings plc FY2025 (ended March 2025) total revenue was approximately $3.96 billion per the SEC 20-F filing; the company's market cap is approximately $120 billion as of May 2026, implying approximately 30x EV/revenue — establishing a public benchmark for a RISC-V architecture IP company. SV002, SV009
CV009 Cerebras Systems filed a draft Form S-1 with the SEC in September 2024; the filing revealed HPC and cloud AI customer concentration but was withdrawn before the IPO process was completed. Revenue estimates of approximately $250 million for 2024 placed the company at 16–28x implied EV/revenue at a $4–7 billion valuation. SV003, SV014, SV019
CV010 Groq raised approximately $1.5 billion total and was valued at $2.8–4 billion in its 2024 funding round; the company focuses exclusively on LLM inference and operates GroqCloud as a commercial inference API service, making it a narrower-scope comparable to Tenstorrent. SV014, SV019
CV011 SambaNova Systems raised approximately $1 billion total and was valued at $5.1 billion in its 2023 funding round; its AI chip + full software stack model is the closest architectural analog to Tenstorrent's approach among private peers. SV014, SV019
CV012 Marvell Technology's market capitalization is approximately $60 billion as of May 2026 on approximately $6 billion FY2025 revenue, implying approximately 10x EV/revenue; the company's custom ASIC AI chip business for hyperscalers provides a comparable data point for hardware-only AI semiconductor monetization. SV012, SV011
CV013 NVIDIA's CUDA software ecosystem — comprising more than 4 million developers, 3,500+ GPU-accelerated applications, and 15+ years of optimization library investment — represents the primary structural barrier to Tenstorrent's enterprise market penetration; enterprise switching cost is estimated at 18–36 months of engineering effort. SV013, SV021, SV028
CV014 Tenstorrent announced the general availability (GA) of the Galaxy Blackhole AI server on April 28, 2026, priced at $110,000 per chassis; this is the company's first commercially available data center product and a key revenue conversion trigger. SV027, SV029, SV023
CV015 Tenstorrent disclosed approximately $150 million in signed contracts (backlog) as of December 2024; this figure represents purchase commitments, not recognized revenue, and no subsequent revenue recognition confirmation has been publicly made. SV015, SV026, SV025
CV016 The Latka algorithmic model estimates Tenstorrent FY2025 revenue at $501.6 million; this figure is an unverified model output with no independent corroboration from financial filings, auditor disclosures, or customer-reported spend data, and should be treated as an upper-bound estimate with low reliability. SV018, SV015
CV017 Tenstorrent's monthly cash burn rate is estimated at $25–50 million based on headcount-adjusted R&D cost norms for a company of approximately 1,000–1,200 employees; the November 2025 Series E ($800M) implies approximately 16–32 months of runway from close, or mid-2027 to mid-2028. SV015, SV025, SV023
CV018 LG Electronics, Hyundai Motor Group, and SoftBank collectively represent the majority of Tenstorrent's known commercial revenue commitments and are simultaneously financial investors — a governance structure that creates potential conflicts of interest and concentration risk not present in arm's-length investor relationships. SV015, SV017, SV028
CV019 The Register's November 2025 review of the Blackhole QuietBox concluded that Tenstorrent's software was 'simply not polished enough for most local AI enthusiasts' — the most prominent independent adverse assessment of Tenstorrent's product maturity, constituting material negative evidence for the enterprise adoption thesis. SV022, SV020
CV020 Arm Holdings IPO in September 2023 was priced at $51 per share implying approximately $52 billion valuation at IPO, which represented approximately 35x EV/revenue at the time; by May 2026 the company's market cap has grown to approximately $120 billion as AI royalty revenue has expanded — establishing a public market trajectory for RISC-V architecture IP companies. SV002, SV009
CV021 NVIDIA's price-to-sales (P/S) multiple is approximately 26x on FY2025 data center revenue of ~$115 billion, establishing the upper bound for AI chip hardware-plus-software valuation multiples in a market-leader context; this multiple is not directly applicable to a pre-profitability startup with unconfirmed revenue. SV001, SV011
CV022 The bull case for Tenstorrent (20% probability) assumes 5% AI inference market share by 2028–2029, approximately $2 billion in revenue, and an 8x revenue exit multiple implying a $16 billion exit valuation — consistent with historical high-growth semiconductor-plus-IP company acquisition or IPO multiples. SV006, SV014, SV023
CV023 The base case for Tenstorrent (50% probability) assumes 2–3% niche inference market share by 2029–2030, approximately $500–800 million in revenue, and a 5x revenue exit multiple implying a $3–5 billion exit valuation — approximately flat to modest positive for Series E investors. SV006, SV014
CV024 The bear case for Tenstorrent (30% probability) assumes software quality gaps persist, NVIDIA maintains 85%+ inference market share, and burn forces a distress exit or acquisition at $300 million–$1 billion — implying a 60–80% loss for Series E investors. SV004, SV020, SV022
CV025 Probability-weighted expected exit valuation across scenarios: ($16B × 20%) + ($4B × 50%) + ($0.75B × 30%) = $5.425 billion, implying approximately 1.7x the $3.2 billion Series E entry price — positive expected value but a thin margin of safety given extreme variance and unresolved information gaps. SV014, SV023
CV026 Jim Keller, Tenstorrent CEO and chief architect, has averaged fewer than three years per employer across Intel (2016–2018), AMD (2012–2015), Apple (2008–2012), and Tesla (2018–2019); this pattern creates a statistically meaningful key-person departure risk that cannot be dismissed on the basis of current public engagement signals. SV026, SV028
CV027 Tenstorrent's TT-Metal software stack is reported by the company to be approximately 90% compatible with HuggingFace model repository formats, providing a developer ecosystem entry point that reduces the cold-start adoption barrier compared to building native integrations from scratch. SV023, SV028
CV028 Tenstorrent has not publicly named a CFO with institutional-grade public-company reporting experience; this governance gap is material for an $800 million late-stage venture company approaching IPO readiness discussions. SV015, SV026
CV029 Tenstorrent's down-round risk increases materially if monthly burn exceeds $50 million and no Series F term sheet is secured by mid-2027 on the basis of the current $3.2 billion valuation floor; in a down-round, Series D/E preference holders would need to restructure to release capital for common equity. SV004, SV020, SV015
CV030 Galaxy Blackhole achieves approximately 350 tokens per second on the DeepSeek R1 model with a claimed lower total cost of ownership than NVIDIA GB300 configurations at equivalent token throughput, per WccfTech's April 2026 review of company-provided benchmarks. SV029, SV027
CV031 Arm Holdings priced its Nasdaq IPO in September 2023 at $51 per share with a $52 billion initial market cap; the IPO represents the most recent comparable public listing for a RISC-V architecture-adjacent semiconductor IP company and sets the roadmap benchmark for Tenstorrent's potential IPO requirements. SV002, SV009, SV011
CV032 Multiple independent analyst reports (Mordor Intelligence, Grand View Research, Allied Market Research, Statista) project the AI chip inference accelerator segment to grow faster than training through 2030, directly benefiting Tenstorrent's inference-optimized product positioning. SV006, SV007, SV008, SV010
CV033 For a Series E investor entering at $3.2 billion, achieving a 5x return requires an exit at $16 billion — approximately equal to the bull case scenario — within a 5-year horizon; the base case exit at $3–5 billion implies approximately 0–1.5x return, or 0–8% IRR, which is below typical venture return thresholds. SV014, SV023
CV034 Tenstorrent's Ascalon RISC-V CPU licensing program generates recurring IP royalty revenue with structurally higher gross margins (estimated 80–90%) than hardware product revenue (estimated 30–50%), and would command a higher EV/revenue multiple if scaled — potentially shifting the overall company multiple upward from hardware-blended levels. SV009, SV002, SV023
CV035 Tenstorrent is not IPO-ready as of May 2026: the company lacks a publicly named CFO with SEC reporting experience, has no disclosed audited FY2024–2025 financial statements, has no confirmed hyperscaler design win, and has a software stack with documented quality gaps that would require remediation before institutional investor scrutiny. SV022, SV015, SV020
CV036 AMD's acquisition of Xilinx for approximately $35 billion in 2022 at approximately 15x revenue provides the most relevant AI chip M&A premium benchmark; a strategic acquisition of Tenstorrent at $5–8 billion in a bull scenario would imply approximately 7–11x revenue — achievable if Galaxy commercial traction is confirmed. SV011, SV021
CV037 Key IPO readiness milestones for Tenstorrent include: confirmed revenue above $500 million (audited), gross margin documentation, CFO hire with SEC reporting experience, at least one hyperscaler design win (Google, AWS, Azure, or Oracle), and a software open-issue backlog reduction below 1,000. SV020, SV023, SV015
CV038 Series D investors entered at a $2.6 billion post-money valuation and Series E investors at $3.2 billion; in any exit below $2.6 billion (the bear case range is $300M–$1B), common equity holders and investors below the Series D liquidation preference level would face material waterfall shortfalls. SV005, SV017, SV015
CV039 Tenstorrent's RISC-V architecture eliminates ARM licensing royalties (typically $0.50–$2.00 per chip), creating a structural cost advantage over ARM-based AI chip designs if the software stack matures to a level where developer ecosystem adoption is self-sustaining. SV009, SV028
CV040 The overall investment recommendation for Tenstorrent at its November 2025 Series E valuation of $3.2 billion is research-more/track: the long-term thesis is credible and the expected value is positive, but the information gap (no audited revenue, no hyperscaler win, software maturity risk) is too large for an institutional investment commitment at the current price. SV014, SV020, SV022
CV041 At $3.2 billion, Tenstorrent is valued at approximately 0.1% of NVIDIA's $3 trillion market cap; capturing even 2–3% of the AI inference accelerator market at industry-average hardware margins would, if scaled, represent a meaningful multiple on the current entry price. SV001, SV011, SV006
CV042 NVIDIA FY2025 data center revenue was approximately $115 billion per the SEC 10-K filing (filed February 2025), confirming AI chip hardware revenue at scale and establishing the addressable market ceiling for any alternative AI accelerator vendor. SV001, SV011
CV043 Arm Holdings FY2025 revenue of approximately $3.96 billion at approximately 30x EV/revenue ($120 billion market cap as of May 2026) demonstrates that a RISC-V IP licensing company can achieve premium public market multiples when revenue is confirmed, audited, and growing at >20% annually — the benchmark Tenstorrent would need to approach for an IPO at a $10 billion+ valuation. SV002, SV009, SV011
来源
编号出版方标题引文
SO001 Tenstorrent Tenstorrent closes $693M+ of Series D funding led by Samsung Securities and AFW Partners Tenstorrent closes $693M+ of Series D funding led by Samsung Securities and AFW Partners
SO002 PR Newswire Tenstorrent closes $693M+ of Series D funding led by Samsung Securities and AFW Partners Tenstorrent is announcing that it has closed over $693M in its Series D funding round at a pre-money valuation of $2B.
SO003 Crunchbase News Jim Keller-Led Tenstorrent Raises Another $700M For AI Chips At $2B+ Valuation Samsung Securities and AFW Partners led the round. Jeff Bezos' Bezos Expeditions, Fidelity Management & Research Co. and LG Electronics also joined the funding.
SO004 TechSpot Nvidia challenger Tenstorrent completes $700 million funding round backed by Jeff Bezos, other high-profile investors Tenstorrent's manufacturing strategy involves partnerships with leading chip fabricators. Its first chips were produced by GlobalFoundries, with future iterations planned through Taiwan Semiconductor Manufacturing Co. and Samsung.
SO005 Futurum Group Tenstorrent Ready to Storm AI Chip Market Tenstorrent's CEO, Jim Keller, announced that the company has secured customer contracts totaling nearly $150 million and intends to launch a new AI processor every two years.
SO006 Forbes Tenstorrent Unveils New Wormhole DevKits And Powerful AI Workstations Blackhole is Tenstorrent's next-gen standalone AI computer, which will feature 140 of the company's Tensix++ cores, 16 CPU cores, and an array of high-speed connectivity.
SO007 PR Newswire Tenstorrent Launches Next Generation Wormhole-based Developer Kits and Workstations The Wormhole n150 and n300 AI accelerators are 3/4 size PCIe add-in cards based on the Wormhole processor.
SO008 RISC-V International Tenstorrent's RISC-V-based Wormhole AI accelerators are available for pre-order today — pre-built workstations start at $12,000
SO009 ServeTheHome Tenstorrent Blackhole and Metalium For Standalone AI Processing The sixteen big RSIC-V cores can run Linux. The other 752 RISC-V are called 'baby' cores that are programmable using C kernels.
SO010 The Register Blackhole QuietBox, Tenstorrent's AI workstation reviewed right now the company's software stack simply isn't polished enough for most local AI enthusiasts.
SO011 Wikipedia Jim Keller (engineer)
SO012 All About Circuits CPU Designer Jim Keller Rethinks RISC-V AI Processing at Tenstorrent In 2020, Keller made the jump from CPUs to CEO, as Junko Yoshida reported in a biographic piece on Keller.
SO013 Tracxn Tenstorrent — 2026 Company Profile & Team Tenstorrent Inc. CIN: 779186725, Canada, Active — Mar 14, 2016
SO014 Tracxn Tenstorrent — 2026 Funding Rounds & List of Investors Tenstorrent has raised a total of $1.18B over 10 funding rounds.
SO015 SWOT Analysis Tenstorrent SWOT Analysis & Strategic Plan 2025-Q4 ECOSYSTEM: Software stack (TT-Buda) is immature compared to Nvidia's CUDA. ADOPTION: Limited public customer deployments and revenue base to date.
SO016 Futurum Group Tenstorrent's Galaxy Blackhole: Can RISC-V Processors Expand Fast Inference Globally? Tenstorrent has moved into volume production with its Galaxy Blackhole compute server.
SO017 Neuronad Bezos Backs Tenstorrent: The AI Chipmaker Challenging Nvidia's Reign
SO018 Built In Tenstorrent Inc. Careers, Perks + Culture
SO019 Silicon Valley Daily Tenstorrent Reels in Whopping $693 Million led by Samsung ~$150M in deals closed as a strong signal of commercial traction and opportunity in the market.
SO020 Spheron Network Tenstorrent vs NVIDIA: Open-Source AI Hardware Compared for Inference and Training (2026) TT-Metal, Tenstorrent's compiler stack, is open-source MIT-licensed code on GitHub.
SO021 GitHub / Tenstorrent Tenstorrent AI — GitHub Organization tt-metal — TT-NN operator library, and TT-Metalium low level kernel programming model.
SO022 Tekedia AI Chip Startup Tenstorrent Secures $693M in Series D Funding, Backed by Bezos, Samsung and Hyundai Tenstorrent, founded in 2016 by Ljubisa Bajic, Ivan Hamer, and Milos Trajkovic, builds AI hardware, provides open-source software for chipmakers, and licenses its technology to clients.
SO023 Cloud News Tech Tenstorrent: The Jeff Bezos-Backed Startup Challenging Nvidia's Dominance in AI Chips
SO024 EE Times Jim Keller on AI, RISC-V, Tenstorrent's Move to Edge IP
SO025 CB Insights Tenstorrent Stock Price, Funding, Valuation, Revenue & Financial Statements
SM001 Gartner Gartner Forecasts Worldwide Semiconductor Revenue to Exceed $1.3 Trillion in 2026 Gartner forecasts global semiconductor revenue will exceed $1.3 trillion in 2026, with AI processing semiconductors representing approximately 30% of total revenue.
SM002 IDC Semiconductor Market to Surge Past the Trillion-Dollar Threshold: AI Infrastructure Drives Market Growth IDC forecasts data center semiconductor revenues will reach $477.1 billion in 2026, driven by AI infrastructure investment.
SM003 Deloitte 2026 Global Semiconductor Industry Outlook Deloitte estimates generative AI chip revenue will represent approximately half of all chip sales in 2026, roughly $500 billion.
SM004 Fortune Business Insights AI Accelerator Market Size, Share & Growth Forecast [2034] The global AI accelerator market is projected to grow from approximately $113 billion in 2025 at a CAGR of 26–27% through 2034.
SM005 MarketsandMarkets AI Inference Market Size, Share & Growth, 2025 To 2030 The AI inference market is projected to grow from approximately $106 billion in 2025 at a 19% CAGR to $255 billion by 2030.
SM006 Global Market Insights RISC-V Market Size, Share & Growth Report, 2025-2034 The RISC-V technology market is expected to be valued at $1.35 billion in 2025 and $1.91 billion in 2026, growing at a 30–41% CAGR through 2034.
SM007 Futurum Group AI Capex 2026: The $690B Infrastructure Sprint Top hyperscalers (Amazon, Google, Meta, Microsoft, Oracle) are projected to spend between $650 billion and $700 billion on AI infrastructure in 2026.
SM008 Intel Market Research RISCV CPU IP Market Outlook 2026-2034 The RISC-V CPU IP market is projected at $720 million in 2026, growing at 12.1% CAGR to $1.8 billion by 2034.
SM009 CNBC Nvidia sales are 'off the charts,' but Google, Amazon and others are competing with custom AI chips NVIDIA controls roughly 80–90% of the AI data center chip market, but custom silicon from hyperscalers is eroding its share.
SM010 Forbes Inside The Neocloud Economy: What's Next For GPU-As-A-Service Neoclouds generated approximately $20 billion in revenue in 2026, serving as a critical buyer segment for alternative AI compute hardware.
SM011 Silicon Analysts NVIDIA AI GPU Market Share 2026: ~80% of AI Accelerators NVIDIA is projected to hold approximately 80% of AI GPU/accelerator market share in 2025, declining toward 75% by 2026 as AMD and custom silicon grow.
SM012 SDxCentral AI inferencing will define 2026, and the market's wide open AI inferencing will define 2026, with neoclouds and alternative chip vendors competing to serve the fastest-growing workload segment.
SM013 DeployBase AI Chip Wars: NVIDIA vs AMD vs Custom Silicon 2026 Update AMD is forecast to grow from approximately 5–8% market share in 2025 to up to 10–15% in late 2026, with custom silicon from hyperscalers also eroding NVIDIA's lead.
SM014 Verified Market Research AI Accelerator Chip Market Report: Size, Growth, Trends & Forecast The AI accelerator chip market is projected to experience significant CAGR of 26–30% through 2030.
SM015 Signisys GPU Cloud Providers: The $20B Neocloud Era The neocloud segment is projected to hit approximately $20 billion in revenue in 2026, with estimates reaching $180 billion by 2030.
SM016 VamsiTalksTech The GPU Supply Chain Crisis: What Every Enterprise CIO Must Know in 2026 Lead times for cutting-edge GPUs have extended to Q1 2027 in many cases, creating structural demand for alternatives.
SM017 Introl AI Inference vs Training Infrastructure: Economics Diverging By 2026, about two-thirds of all AI compute will be for inference, up from about one-third in 2023 and half in 2025.
SM018 QverLabs NVIDIA vs AMD vs Intel: Who Will Dominate the AI Chip Market? NVIDIA's CUDA ecosystem remains the de-facto standard; switching to AMD or Intel alternatives requires significant software rewrite and performance tuning investment.
SM019 Futurum Group Tenstorrent Galaxy Blackhole: Volume Production and Neocloud Deployment Analysis Tenstorrent's Galaxy Blackhole system entered volume production in May 2026, with hardware deployed in at least five neocloud co-locations.
SM020 The Register Blackhole QuietBox, Tenstorrent's AI workstation reviewed right now the company's software stack simply isn't polished enough for most local AI enthusiasts.
SM021 McKinsey The evolution of neoclouds and their next moves
SM022 Global Market Insights RISC-V Market Size, Share & Growth — Processor IP Licensing The broader RISC-V technology market is expected to exceed $1.91 billion in 2026 at a 30–41% CAGR.
SM023 PR Newswire Tenstorrent closes $693M+ of Series D funding led by Samsung Securities and AFW Partners Tenstorrent has approximately $150M in signed contracts and is deploying AI hardware to customers across neoclouds and enterprise.
SM024 Forbes Tenstorrent Unveils New Wormhole DevKits And Powerful AI Workstations Tenstorrent's Wormhole-based developer kits and workstations were commercially launched in July 2024.
SM025 EE Times Jim Keller on AI, RISC-V, Tenstorrent's Move to Edge IP Tenstorrent is actively licensing its RISC-V Ascalon CPU core to chipmakers and pursuing edge AI silicon.
SP001 AppScale Blog Beyond NVIDIA: 2026 AI Accelerator Landscape — Groq, Cerebras, Trainium, TPU The AI accelerator market is no longer NVIDIA-only: multiple credible contenders exist depending on workload, cost, and deployment style.
SP002 Ertas AI Taalas vs Nvidia vs Groq vs Cerebras: AI Inference Hardware Compared (2026)
SP003 BestAIWeb Cerebras vs. Groq vs. GPU Clouds: The Custom Silicon Bet Reshaping Inference Economics in 2026
SP004 IntuitionLabs NVIDIA AI GPU Prices: H100 ($27K-$40K) & H200 ($315K/8-GPU) Cost Guide H100 purchase price: $27,000–$40,000 per GPU in 2026 market.
SP005 GPU.fund Blackwell B200 vs H100: Is the 64% Price Premium Worth It? (2026)
SP006 Google Cloud Blog Introducing Trillium, sixth-generation TPUs Trillium delivers a 4.7x improvement in compute performance per chip over TPU v5e.
SP007 TECHi Cerebras IPO: 2026 Price Range, Valuation, Risks Cerebras officially filed for an IPO in April 2026 targeting a $23-26.6 billion public valuation on the Nasdaq.
SP008 TechFundingNews Nvidia rival Cerebras in $1B funding talks, just after Etched's $500M raise
SP009 JarvisLabs NVIDIA H200 Price Guide 2026: GPU Cost, Rental & Cloud Pricing
SP010 Spheron Network Tenstorrent vs NVIDIA: Open-Source AI Hardware Compared for Inference Tenstorrent's open-source TT-Metal stack and RISC-V programmability are genuine differentiators but trail CUDA in ecosystem maturity.
SP011 ChatForest The Custom AI Chip Race: Meta, Google, Amazon, and Microsoft Are All In (2026)
SP012 AlgeriaTech News Groq vs Cerebras 2026: AI Inference 100x Faster Than GPU
SP013 AlphaStreet News Nvidia's CUDA Lock-In and Supply Scarcity Make Its AI Chip Moat Harder to Break Than It Looks Over 4 million registered developers and 40,000 organizations rely on CUDA, creating enormous switching costs for both established enterprises and new startups.
SP014 DataCenter Dynamics SambaNova exploring sale after struggling to secure further funding SambaNova was exploring a sale in late 2025 after struggling to raise further funding, with BlackRock marking shares to ~$2.4B from a $5B peak.
SP015 SiliconAngle Report: SambaNova is raising $350M+ from Intel-backed consortium SambaNova is raising a $350M+ Series E co-led by Vista Equity and Intel.
SP016 TLDL AI Hardware Wars 2026: NVIDIA Blackwell vs AMD vs Intel
SP017 Wccftech Intel Says It Won't Compete With NVIDIA In AI Market, Shifts Focus to Cost-Effective AI Solutions Intel has shifted focus to cost-effective Gaudi AI solutions, effectively conceding the high-end AI training market to NVIDIA.
SP018 Medium (Wang) Comparing AI Hardware Architectures: SambaNova, Groq, Cerebras vs NVIDIA GPUs
SP019 NVIDIA NVIDIA H100 Tensor Core GPU
SP020 AMD AMD Instinct MI300X Accelerators
SP021 Intel Intel Gaudi 3 AI Accelerators
SP022 Cerebras Systems Cerebras CS-3 AI Supercomputer — Product System The CS-3 delivers up to 2,000 tokens/sec for large language model inference workloads.
SP023 Groq GroqCloud — Inference API
SP024 Tenstorrent Tenstorrent Blackhole Hardware
SP025 CRN Intel Looking To Acquire Startup AI Chip Developer SambaNova: Report Intel was in discussions to acquire SambaNova in December 2025, valuing the company at ~$1.6B including debt.
SP026 The Register Blackhole QuietBox, Tenstorrent's AI workstation reviewed The software simply isn't polished enough for most local AI enthusiasts.
SI001 Tenstorrent Tenstorrent Galaxy™
SI002 HPCwire Tenstorrent Announces General Availability of Galaxy Blackhole AI System
SI003 Financial Content Tenstorrent Enables AI At Scale with Industry-Leading Performance
SI004 Yahoo Finance Tenstorrent Enables AI At Scale with Industry-Leading Performance
SI005 News Directory 3 Tenstorrent Launches Galaxy Blackhole AI System for General Availability
SI006 Digital Citizen Tenstorrent says its Galaxy Blackhole servers can challenge NVIDIA in AI inference
SI007 AI2 Work Tenstorrent Galaxy Blackhole Targets Nvidia With Bold 5x Cost Claim
SI008 Tech Startups AI chip startup Tenstorrent in talks to raise $800M in funding at a $3.2B valuation led by Fidelity
SI009 PM Insights Tenstorrent Valuation
SI010 PitchBook Tenstorrent 2026 Company Profile: Valuation, Funding & Investors
SI011 Latka How Tenstorrent hit $501.6M revenue with a 1.1K person team in 2025 Algorithmic revenue model estimate; not verified from disclosed financials
SI012 CompWorth Tenstorrent – Estimated Net Worth, Growth & Competitor Overview
SI013 Koyeb Tenstorrent Cloud Instances — Unveiling Next-Gen AI Accelerators
SI014 Analytics Insight Tenstorrent Company Profile
SI015 Business Model Canvas Template How Does Tenstorrent Company Work?
SI016 TechForward From Closed Silicon to Community Hardware — Inside Tenstorrent's Developer Day
SI017 FXiS AI The Future of AI Chips: Insights from Tenstorrent's $100 Million Investment
SI018 ACN Newswire Tenstorrent Enables AI At Scale with Industry-Leading Performance
SI019 The Register Tenstorrent's Galaxy Blackhole AI servers are finally out Galaxy Blackhole AI servers finally reach general availability amid questions about software ecosystem readiness
SI020 Forbes Tenstorrent Unveils Galaxy AI Platform Targeting Scale And Efficiency
SI021 WCCFTech Tenstorrent Vows to 'Crush Everyone' as Galaxy Blackhole Hits 350 Tokens/sec
SI022 Tracxn Tenstorrent — Funding Rounds & List of Investors 2026
SI023 Crunchbase News Tenstorrent AI Chips Unicorn Jim Keller Tenstorrent signed approximately $150M in customer contracts
SI024 CB Insights Tenstorrent Financials
SI025 ZoomInfo Tenstorrent Company Profile and Employees
SI026 Government of Canada — Corporations Canada Federal corporation information — Tenstorrent Inc. Federal corporation incorporated in Ontario, Canada
SE001 Tenstorrent Blackhole Cards — Tenstorrent Hardware
SE002 Tenstorrent TT-Metalium — Tenstorrent Software
SE003 Tenstorrent Documentation Blackhole AIBS Specifications — docs.tenstorrent.com TSMC 6nm process node; 120 Tensix cores; 180 MB on-chip SRAM
SE004 Tenstorrent Documentation Blackhole AIBS Index — docs.tenstorrent.com
SE005 Tenstorrent Documentation TT-Forge Documentation Index
SE006 Tenstorrent (GitHub) tenstorrent/tt-metal — GitHub Repository 25,830+ commits, 429 forks, 161 releases as of April 2026
SE007 Tenstorrent (GitHub) tenstorrent/tt-metal Issues
SE008 Tenstorrent (GitHub) tenstorrent/tt-metal Discussions
SE009 Tenstorrent (GitHub) tenstorrent/tt-forge — GitHub Repository
SE010 Tenstorrent (GitHub) tenstorrent/tt-metal Releases
SE011 AwesomeAgents.ai Tenstorrent Blackhole p150a Hardware Review TSMC 6nm process node, 120 Tensix cores, 32 GB GDDR6
SE012 DeepWiki TT-Metal Installation and Setup — DeepWiki
SE013 DeepWiki Blackhole Processors Architecture — DeepWiki 120 Tensix processing elements, 16 RISC-V application cores (SiFive X280)
SE014 typevar.dev Tenstorrent TT-Metal: A Deep Dive
SE015 Pyron Pyron Documentation — Tenstorrent NPU and SDK
SE016 The Register Blackhole QuietBox: Tenstorrent's AI Workstation Reviewed The software simply isn't polished enough for most local AI enthusiasts
SE017 ServeTheHome Tenstorrent Blackhole and Metalium for Standalone AI Processing
SE018 All About Circuits CPU Designer Jim Keller Rethinks RISC-V AI Processing at Tenstorrent
SE019 Futurum Group Tenstorrent's Galaxy Blackhole: Can RISC-V Processors Expand Fast Inference Globally?
SE020 TechSpot Nvidia challenger Tenstorrent completes $700 million funding round
SE021 RISC-V International Tenstorrent's RISC-V Based Wormhole AI Accelerators Available for Pre-Order
SE022 Tenstorrent Tenstorrent Launches Blackhole Developer Products at Tenstorrent Dev Day
SE023 SWOT Analysis Tenstorrent SWOT Analysis
SE024 AnandTech Tenstorrent Announces Blackhole AI Accelerators
SE025 Phoronix Tenstorrent Blackhole p150a Review
SE026 MLCommons MLPerf Inference: Datacenter Benchmark Results
SU001 Tenstorrent Tenstorrent Closes $693M+ Series D Funding Led by Samsung Securities and AFW Partners
SU002 PR Newswire Tenstorrent Closes $693M+ Series D Funding Led by Samsung Securities and AFW Partners
SU003 HPCwire Tenstorrent Announces General Availability of Galaxy Blackhole AI System
SU004 The Register Tenstorrent's Galaxy Blackhole AI Servers Are Finally Out
SU005 Koyeb Tenstorrent Cloud Instances: Unveiling Next-Gen AI Accelerators
SU006 SV Daily Tenstorrent reels in whopping $693 million led by Samsung
SU007 LG AI Research LG AI Research — About and Research Focus
SU008 Tenstorrent Tenstorrent Developer Day: Blackhole Launch and Partner Ecosystem
SU009 Business Wire Tenstorrent Galaxy Blackhole AI Server — General Availability Announcement
SU010 Tekedia AI Chip Startup Tenstorrent Secures $693M in Series D Funding Backed by Bezos
SU011 Forbes Tenstorrent Unveils Galaxy AI Platform Targeting Scale and Efficiency
SU012 Tom's Hardware Tenstorrent Galaxy AI Server: Hardware and Enterprise AI Infrastructure
SU013 The Next Platform Tenstorrent Bets Big on Open Ecosystem for AI Chip Market Entry
SU014 Tenstorrent Tenstorrent — DevCloud Developer Access and Company Vision
SU015 TechCrunch Tenstorrent: AI Chip Startup Coverage
SU016 GitHub tenstorrent/tt-metal — GitHub Repository
SU017 The Register Blackhole QuietBox: Tenstorrent's AI Workstation Reviewed
SU018 LinkedIn Tenstorrent Company Profile — LinkedIn
SU019 Futurum Group Tenstorrent's Galaxy Blackhole: Can RISC-V Processors Expand Fast Inference Globally?
SU020 Wccftech Tenstorrent Vows to Crush Everyone — Galaxy Blackhole Hits 350 Tokens/s on DeepSeek
SU021 Phoronix Tenstorrent Blackhole p150a Review — Linux Performance
SU022 TechForward From Closed Silicon to Community Hardware: Inside Tenstorrent's Developer Day
SU023 AI² Tenstorrent Galaxy Blackhole Targets NVIDIA with Bold 5x Cost Claim
SU024 ServeTheHome Tenstorrent Blackhole and Metalium for Standalone AI Processing
SU025 The Verge Tenstorrent and SoftBank Partner for AI Chips in Japan
SU026 IEEE Xplore Tenstorrent Blackhole Architecture: Research Papers and Citations
SR001 Bureau of Industry and Security, US Department of Commerce BIS Restricts Exports of Advanced Computing Items — October 2023 Rule BIS is implementing restrictions on the export, reexport, or transfer of advanced computing integrated circuits and related items.
SR002 Federal Register, US Government Export Controls: Advanced Computing Semiconductor Manufacturing Items — 2023-25073
SR003 Official Journal of the European Union Directive (EU) 2022/2555 on Measures for a High Common Level of Cybersecurity — NIS2
SR004 Google Patents US Patent US11704538B2 — Data processing method and device (neural network chip technology)
SR005 Semiconductor Engineering Tenstorrent Blackhole AI Chip Architecture Review
SR006 Semiconductor Engineering Chip Export Controls and RISC-V: Regulatory Implications
SR007 Ars Technica Tenstorrent Takes On Nvidia With Open-Source AI Chip and $693 Million in Backing
SR008 Ars Technica Tenstorrent and SoftBank Team Up to Bring AI Chips to Japan
SR009 Bloomberg Tenstorrent Raises $693 Million in Series D Funding Round
SR010 Bloomberg Tenstorrent, SoftBank Agree on AI Chip Deal for Japan Data Centers
SR011 Bloomberg US Broadens AI Chip Restrictions in Latest Export Control Update
SR012 SemiWiki Tenstorrent Raises $100M from Hyundai and Samsung
SR013 DigiTimes Tenstorrent AI Chip and RISC-V Strategy — Series D Follow-Up
SR014 The Register Blackhole QuietBox: Tenstorrent's AI Workstation Reviewed The software simply isn't polished enough for most local AI enthusiasts.
SR015 GitHub / Tenstorrent tt-metal: GitHub Repository — Open Issues and Pull Request Tracker 3,488 open issues and 988 open pull requests observed as of April 2026.
SR016 Phoronix Tenstorrent Blackhole Developer Day Coverage
SR017 Tom's Hardware Tenstorrent Hardware and Product Coverage
SR018 NextPlatform Tenstorrent Coverage — AI Accelerator Analysis
SR019 HPC Wire Tenstorrent Announces General Availability of Galaxy Blackhole AI System
SR020 PitchBook Tenstorrent Company Profile — Funding and Financials
SR021 CB Insights Tenstorrent Financial Profile
SR022 WccfTech Tenstorrent Vows to Crush Everyone — Galaxy Blackhole Hits 350 Tokens on DeepSeek
SR023 Crunchbase News Tenstorrent AI Chips Unicorn: Jim Keller's $1B Bet Against Nvidia
SR024 Semiconductor Engineering AI Chip Export Controls Tighten: 2024 Update
SR025 Semiconductor Engineering Chip Export Controls — Regulatory Landscape Overview
SR026 Bureau of Industry and Security Export Administration Regulations (EAR) — BIS Policy Guidance
SR027 Ars Technica 693M Startup Tenstorrent Bets Open-Source RISC-V AI Chips Can Beat Nvidia
SR028 Ars Technica Tenstorrent and SoftBank Partner for AI Chips in Japan — Policy Implications
SR029 The Register Tenstorrent's Galaxy Blackhole AI Servers Are Finally Out
SR030 TechStartups AI Chip Startup Tenstorrent in Talks to Raise $800M at $3.2B Valuation Led by Fidelity
SR031 Phoronix Tenstorrent Blackhole Launch Coverage
SR032 NextPlatform Tenstorrent Blackhole: AI Chip Analysis and Risk Assessment
SV001 US Securities and Exchange Commission (EDGAR) NVIDIA Corporation Annual Report on Form 10-K — Fiscal Year 2025 (ended January 26, 2025) NVIDIA data center revenue reached approximately $115 billion in FY2025, confirming dominance of the AI accelerator market.
SV002 US Securities and Exchange Commission (EDGAR) Arm Holdings plc — Annual Report on Form 20-F (Fiscal Year 2025) Arm Holdings FY2025 total revenue approximately $3.96 billion; royalty and licensing revenue at high gross margins supports a premium EV/revenue multiple.
SV003 US Securities and Exchange Commission (EDGAR) Cerebras Systems — Draft Registration Statement on Form S-1 (September 2024) Cerebras Systems filed a draft S-1 in September 2024, providing the most detailed public financial disclosure of any private AI chip startup peer.
SV004 Financial Times AI Chip Startups Face Valuation Reality Check as Revenue Gaps Widen AI chip hardware startups face a valuation correction risk as revenue ramp delays widen the gap between funding-round prices and commercial traction.
SV005 Wall Street Journal Tenstorrent Raises $800 Million, Pushing Valuation to $3.2 Billion Tenstorrent closed an $800 million Series E round led by Fidelity at a $3.2 billion post-money valuation in November 2025.
SV006 Mordor Intelligence AI Chip Market — Size, Share, Growth Trends and Forecast 2025–2030 The global AI chip market is projected to grow from approximately $38 billion in 2024 to approximately $170 billion by 2030 at a CAGR of approximately 28%.
SV007 Grand View Research Artificial Intelligence Chip Market Size, Share & Trends Analysis Report 2024–2030 Inference accelerator segment expected to grow faster than training through 2030, representing an increasing share of the total AI chip market.
SV008 Allied Market Research AI Semiconductor Market Size, Share, Competitive Landscape and Trend Analysis Report 2023–2032 The global AI semiconductor market is expected to reach $200+ billion by 2032, with cloud inference infrastructure as the primary growth driver.
SV009 Arm Holdings plc Arm Holdings Investor Relations — Annual Report and Financial Overview FY2025 Arm Holdings reports royalty and licensing revenue growth driven by AI chip adoption; Armv9 architecture captures higher royalty rates than prior generations.
SV010 Statista AI Semiconductor Revenue Worldwide 2020–2030 Forecast
SV011 Seeking Alpha NVIDIA Valuation Analysis: Revenue Multiple, Data Center Dominance, and AI Chip Comps — 2026 Update NVIDIA trades at approximately 26x forward EV/revenue on $115B+ data center revenue base; Arm Holdings at ~30x on royalty-heavy model; Marvell at ~10x on custom ASIC mix.
SV012 Marvell Technology Marvell Technology Investor Relations — FY2025 Financial Results and AI Custom ASIC Update Marvell Technology FY2025 revenue approximately $6 billion with AI custom silicon as the fastest-growing segment; market cap approximately $60 billion.
SV013 TrendForce AI Accelerator Market Share and Competitive Landscape — Q1 2026 Report NVIDIA retained approximately 80–85% of AI accelerator market share in Q1 2026; alternative AI chip vendors collectively hold 15–20% including AMD, Intel, and custom ASICs.
SV014 SemiAnalysis AI Chip Startup Valuations: Cerebras, Groq, SambaNova, Tenstorrent Compared — 2026 Landscape Among AI chip startups at comparable funding stages, EV/revenue multiples range from 10x (Groq) to 28x (Cerebras), with software and ecosystem maturity as the primary differentiator.
SV015 PitchBook Tenstorrent — Company Funding History and Investor Profile (May 2026)
SV016 TechCrunch Tenstorrent Raises $800M in Series E Round Led by Fidelity at $3.2B Valuation
SV017 Bloomberg Tenstorrent Raises $693 Million Valuing AI Chip Startup at $2.6 Billion
SV018 Latka Agency / GetLatka Tenstorrent Revenue Estimate — Algorithmic Model FY2025 Latka algorithmic model estimates Tenstorrent FY2025 revenue at $501.6 million; this is a model output, not a company-disclosed figure.
SV019 CB Insights AI Chip Startup Landscape: Funding, Valuation, and Competitive Dynamics 2026
SV020 VentureBeat AI Chip Hardware Startups Face Valuation Tests as Revenue Ramps Slowly Hardware AI chip startups that raised large rounds in 2024–2025 face increasing pressure to show recognized revenue, not just backlog or design wins, to justify their valuations.
SV021 Reuters AI Chip Market: NVIDIA Dominance and the Challenger Landscape
SV022 The Register Blackhole QuietBox: Tenstorrent's AI Workstation Reviewed — Software Not Polished Enough The software simply isn't polished enough for most local AI enthusiasts — let alone enterprise deployments.
SV023 NextPlatform Tenstorrent Investment and Valuation Context — AI Accelerator Analysis 2026
SV024 SiliconANGLE Tenstorrent Closes $800M Series E at $3.2B Valuation: What It Means for the AI Chip Race
SV025 TechStartups AI Chip Startup Tenstorrent Raises $800M at $3.2B Valuation Led by Fidelity
SV026 Crunchbase News Tenstorrent Funding History and Investor Profile — Series D and E Analysis
SV027 HPC Wire Tenstorrent Announces General Availability of Galaxy Blackhole AI System — April 2026 Tenstorrent announced the general availability of the Galaxy Blackhole AI server at $110,000 per chassis in April 2026.
SV028 Ars Technica Tenstorrent Takes On Nvidia With Open-Source AI Chip and $693 Million in Backing
SV029 WccfTech Tenstorrent Galaxy Blackhole Hits 350 Tokens/sec on DeepSeek R1 — Performance Review Galaxy Blackhole achieves 350 tokens per second on DeepSeek R1 with a claimed lower total cost of ownership than NVIDIA GB300.
SV030 Tracxn Tenstorrent — Funding Rounds, Valuation, and Competitive Intelligence