d-Matrix
全栈推理硬件公司,产品进展可信,但 $2B 私募估值仍缺完整证明
d-Matrix 拥有差异化推理架构和可信商业化势头,但 $2B Series C 轮仍领先于公开证据:收入、客户深度和可持续部署经济性都还没证明。
封面要素
公司概况
d-Matrix 是一家位于 Santa Clara 的 AI 推理基础设施公司,由 Sid Sheth 和 Sudeep Bhoja 于 2019 年创立。公司现在围绕 Corsair 加速卡、JetStream 网络、Aviator 软件和开放标准机柜部署,给出了一套具体的全栈产品叙事,目标客户包括超大规模、企业和主权数据中心。公开融资历史有较强证据支撑:公司在 $2B 估值下完成 $275M Series C,累计披露融资 $450M;到 2025 年底,员工超过 250 人,并拥有多办公室布局。仍缺的是经营分母:公开资料尚未披露收入、ARR、广泛客户数或控制经济性,细节不足以像成熟后期基础设施公司那样承销这门生意。
- 成立时间
- 2019-01-01
- 创始人
- Sid Sheth, Sudeep Bhoja
- 总部
- Santa Clara, California
- 产品
- Corsair 加速卡、JetStream 透明网络卡、Aviator 软件和 SquadRack 参考架构,共同拼出一套低延迟推理栈,面向标准 PCIe 服务器、Ethernet 网络和存量数据中心部署。
- 客户
- 运行低延迟敏感推理负载的超大规模、企业、主权云和专业 AI 云运营商。
- 商业模式
- 经由 OEM、集成商和云伙伴销售硬件加软件平台,经济性大概率绑定在多组件部署上,而不是独立的自助式软件。
- 阶段
- Series C private company
- 融资情况
- 2023 年完成 $110M Series B,随后在 2025 年 11 月以 $2B 估值完成 $275M Series C,累计披露融资达到 $450M。
执行摘要
主要优势
- Corsair、JetStream、Aviator 和 SquadRack 已拼成完整产品栈,d-Matrix 不只是架构概念,也给买方一个可测试的部署面。
- 延迟逻辑有差异化:以内存为中心的 DIMC 设计和开放标准部署模型,瞄准 brownfield 数据中心的真实推理瓶颈,而不是要求整机柜重建。
- 资本和生态支持强:披露融资 $450M,Temasek、Bullhound、Triatomic、QIA、EDBI 和 M12 等可见股东,降低了短期生存融资风险。
- 早期商业化证据足够重要:Supermicro 渠道供货、GigaIO scale-up 集成、Gimlet 异构云工作,显示从送样到生产的可信路径。
主要风险
- 没有公开收入、ARR、毛利率、定价或客户数能支撑当前 $2B 估值,估值承销看不见分母。
- 执行依赖 Corsair、JetStream、Aviator、合作伙伴认证系统,以及商业化前 3DIMC/Raptor 路线图同步成熟。
- 先进封装、晶圆代工伙伴和 OEM / 集成伙伴带来的供应链与生态依赖,制造了真实认证、成本和时间风险。
- 客户证据仍由伙伴牵引、偏试点:公开材料还缺具名生产客户名单、留存指标和集中度数据。
- 公开安全、合规和出口管制姿态,比许多企业或主权买方要求的水平更薄。
未决问题
- 经审计收入或 ARR、毛利率、定价、现金跑道和已实现单元经济仍未披露。
- 具名生产部署、活跃客户数、重复下单节奏、留存和集中度指标仍未出现在公开资料中。
- 股权结构条款、清算优先权、老股交易或 409A 标记,以及董事会控制经济性均未公开。
- JetStream 爬坡和 3DIMC 到 Raptor 路线图的生产良率、成本曲线和交付时间,证据仍不足。
- SOC 2、ISO 27001、trust center 文档,以及具体出口管制敞口分析等公开安全 / 合规材料仍不完整。
目录
01公司概况
1.1 身份、版图与商业模式
应把 d-Matrix 看作一家专门围绕推理而非训练打造的 Series C AI 基础设施公司。官方网站首页、关于页面和 2025 年 Series C 新闻稿都把公司锚定在 Santa Clara,并将业务表述为向超大规模、企业和主权数据中心工作负载销售加速计算。公开材料称公司成立于 2019 年,当前产品栈结合了 Corsair 推理加速器、JetStream 网络和 Aviator 软件。首页还称,3DIMC 架构采用基于 chiplet、适配 PCIe 的设计,可扩展到超过 1000 亿参数的模型。后续章节可复用的底层事实因此很清楚:d-Matrix 正在把低延迟推理所需的硬件加软件基础设施商业化。公开记录仍未给出公司层面的客户数或收入数据,无法把当前商业规模完全量化。[CO001, CO002, CO003, CO008, CO009, CO010]
| 指标 | 数值 / 状态 | 日期 | 置信度 | 缺口 / 备注 |
|---|---|---|---|---|
| 成立时间 | 2019 | 2019 | 高 | 官方和投资人材料均指向 2019 年为成立年份。 |
| 总部 | Santa Clara, California | 2025-11-12 | 高 | Series C 和 Corsair 材料都把 Santa Clara 作为标准总部。 |
| 公开办公地点 | Toronto;Sydney;Bangalore;Belgrade | 2025-11-12 | 高 | 这是抓取材料中最新具名的非总部办公室。 |
| 阶段 | 未上市 Series C 公司 | 2025-11-12 | 高 | 阶段判断由 2025 年 11 月融资支撑,且没有公开上市证据。 |
| 核心产品 | Corsair 加速器;JetStream 网络;Aviator 软件 | 2025-11-18 | 高 | 到 2025 年底,公司公开描述的是三部分硬件和软件栈。 |
| 已披露累计融资 | $450M | 2025-11-12 | 高 | 官方、投资人和独立报道都相互印证了累计融资额。 |
| 最新披露估值 | $2B | 2025-11-12 | 高 | 估值专门绑定 Series C 交割。 |
| 最新披露员工数 | 全球 250+ 人 | 2025-11-12 | 中 | Series C 关键事实公开披露了该数据,但 2026 年没有独立刷新。 |
| 公开商业化证明 | GigaIO 合作加 Gimlet Labs 基准 | 2026-03 | 中 | 具名证明来自合作伙伴和工作负载公告,而不是广泛客户名单。 |
| 当前客户数 | 未公开披露 | 2026-05-26 | 低 | 应直接索取已签约客户数、部署数和复购数,而不是从合作伙伴叙事推断。 |
| 当前收入或 ARR | 未公开披露 | 2026-05-26 | 低 | 应直接从财务材料索取月度经常性收入或收入运行率。 |
本表把硬性的公开身份和融资事实,与抓取材料中仍缺失的重要指标拆开,尤其是客户数、收入或 ARR。
[CO001, CO002, CO003, CO009, CO010, CO015]创始人和资本支撑全栈推理平台,但商业化证据仍落后于融资叙事。
[CO004, CO008, CO010, CO015, CO025, CO026]1.2 创始人、领导层与治理
创始人、高管和董事姓名层面的公开图景异常清楚,但控制经济性仍不完整。关于页面将 Sid Sheth 标为创始人兼 CEO,将 Sudeep Bhoja 标为创始人兼 CTO;公开的投资人和产品新闻稿到 2025 年仍持续引用这两人,显示技术和商业领导层稳定。公司还展示了覆盖软件、产品、财务、法务和制造的运营班底。治理可见度则更混合。d-Matrix 公开列出一批董事,背后对应 Bullhound Capital、Playground Global、Triatomic Capital、M12、Nautilus Venture Partners 和 Temasek,与融资历史吻合,也为后续章节提供了一张可用地图,说明谁可能影响战略。但已抓取材料没有披露董事会委员会、投票门槛、保护性条款或当前持股比例,因此治理分析应把董事名单视为真实但不完整,而不是假定标准风投权利已经存在。[CO004, CO005, CO006, CO007, CO042, CO046]
| 人物 | 公开角色 | 背景或职能范围 | 重要性 | 关键人物或治理备注 |
|---|---|---|---|---|
| Sid Sheth | 创始人兼首席执行官 | Series B、Corsair 和 Series C 发布中的公开商业与战略声音 | 融资故事、上市路径和推理优先投资逻辑的核心负责人 | 关键人物依赖高,因为他同时锚定投资人沟通和外部战略 |
| Sudeep Bhoja | 创始人兼首席技术官 | 与架构和计算内存路线图绑定的联合创始人 | 平台以及 Corsair 之后后续产品的核心技术守门人 | 技术连续性强,但路线图执行仍然集中 |
| Peter Buckingham | 软件高级副总裁 | 关于页面上可见的软件高管 | 说明公司不只做芯片,也在建设软件能力 | 公开来源没有显示更深的软件组织厚度 |
| Sree Ganesan | 产品副总裁 | 关于页面上的公开产品负责人 | 支撑产品化和路线图包装的职能覆盖 | 未披露公开产品 P&L 或商业所有权细节 |
| PJ Jamkhandi | 财务与会计副总裁 | 关于页面具名的财务高管 | 公开融资节奏快于财务披露,因此对尽调重要 | 抓取材料中没有出现 CFO 级 KPI 包 |
| Richard Ogawa | 总法律顾问 | 关于页面具名的法务高管 | 融资和 IP 密集型合作依赖法务执行,因此相关 | 董事会委员会和法律风险监督结构未公开披露 |
| Jerry Qubain | 制造与运营副总裁 | 关于页面上可见的运营负责人 | 半导体业务必须放大制造和交付,因此重要 | 制造足迹和供应商依赖大多未披露 |
| Per Roman | Bullhound Capital 关联董事 | 关于页面披露、Series C 投资人材料呼应的投资人关联董事 | 说明新资本提供方在治理上有可见影响力 | 公开材料未披露委员会任命或投票门槛 |
| Michael Stewart | M12 关联董事 | 投资人关联董事,也是公司发布中反复出现的公开支持者 | 将长期 Microsoft 邻近关系与董事会可见度连在一起 | 席位背后的投资人权利和持股经济性未公开 |
| Russell Tham | Temasek 关联董事 | 公共董事名单上与 Series B 领投方绑定的董事 | 显示 Series B 资本延续进当前治理 | 抓取来源未披露 Temasek 是否持有保护性条款 |
覆盖范围有意保持部分化,因为公开记录对具名创始人、高管和董事的揭示,明显强于对委员会结构、持股比例或内部汇报线的揭示。
[CO004, CO005, CO006, CO007, CO042, CO046]1.3 融资形成、阶段与利益相关方
融资形成是公司概况中外部佐证最强的一部分。VentureBeat 报道了 2022 年围绕 Microsoft Project Bonsai 合作关系完成的 $44M Series A。官方 2023 年新闻稿随后披露 Temasek 领投的 $110M Series B;官方 2025 年 Series C 新闻稿、Bullhound 声明、QIA 声明和独立媒体又一致确认,d-Matrix 以 $2B 估值融资 $275M,累计披露融资达到 $450M。同一组 Series C 材料还给出了已抓取资料中最清楚的当前规模标记:全球 250+ 名员工、Santa Clara 总部,以及 Toronto、Sydney、Bangalore 和 Belgrade 办公室。上述组合支持一个务实阶段判断:这是一家后期私有基础设施创业公司,而不是早期研究项目。融资可见度仍高于经营披露。公司称新资金将支持全球扩张和大规模部署,但已抓取来源没有披露当前收入、ARR 或广泛具名客户名单,这些指标应继续列为明确尽调缺口。[CO011, CO012, CO013, CO014, CO015, CO016]
| 利益相关方 | 角色 | 公开信号 | 控制权或经济重要性 | 尽调要求 |
|---|---|---|---|---|
| Bullhound Capital | Series C 联合领投方和董事会关联投资人 | 领投 2025 年融资,并由 Per Roman 在公开董事会中代表 | 最新一轮中的新资金支持方,且具备可见治理存在感 | 核实董事席位条款、经济权益和清算优先权 |
| Triatomic Capital | Series C 联合领投方 | 被列为 2025 年联合领投方,并由 Jeff Huber 在公开董事会中代表 | 当前资本结构中的重要信号型投资人 | 确认出资金额,以及任何信息权或同意权 |
| Temasek | Series B 领投方和 Series C 联合领投方 | 支持了 2023 年 Series B,又在 2025 年 Series C 中出现,并由 Russell Tham 代表进入董事会名单 | 融资弧线中最可见的重复支持方 | 索取当前持股比例和任何特殊治理权 |
| Qatar Investment Authority | 新 Series C 投资人 | QIA 公开宣布参与 2025 年融资 | 增加主权资本信誉,并可能带来长久期支持 | 澄清该持仓是纯财务投资还是战略性投资 |
| M12 | 重复投资人和董事会关联 Microsoft 基金 | 在 Series B 和 Series C 材料中可见,并由 Michael Stewart 代表进入董事会名单 | 以董事会层面的可见度维持长期 Microsoft 邻近关系 | 澄清除风投持股外,是否存在商业或云渠道预期 |
| Playground Global | 拥有公开董事会可见度的早期投资人 | 在 Series B 材料中被引用,并由 Sasha Ostojic 代表进入董事会 | 可能对早期技术和战略成形故事重要 | 确认后续融资后的当前持股 |
| Nautilus Venture Partners | 董事会关联的重复投资人 | 出现在董事会名单和后续投资人名单中 | 在早期轮次和当前治理之间提供连续性 | 澄清经济权益、董事会投票权和跟投情况 |
| GigaIO | 部署与基础设施合作伙伴 | 2025 年 5 月合作将 Corsair 与 SuperNODE 集成,用于企业级推理 | 在 2026 年收购标题出现在新闻索引前,这是最可见的系统级商业化证明 | 索取已签约部署、管线价值和任何排他性条款 |
| Alchip | 路线图与封装合作伙伴 | 2025 年 11 月合作瞄准 3D DRAM 和 Raptor 后继平台 | 对下一代产品至关重要,而不是当前收入可见度 | 澄清商业化时间表、制造依赖和共同开发义务 |
| Gimlet Labs | 公开工作负载验证合作伙伴 | 2026 年 3 月基准在异构部署模型中使用 Corsair | 提供具体外部工作负载证据,即便它还不是广泛客户名单 | 索取该关系是否产生经常性生产支出,还是只处于评估状态 |
这不是完整股权结构表。它把投资人和战略商业化交易对手放在一起,因为两类主体都会影响公司的控制权故事和近期执行风险。
[CO015, CO016, CO020, CO026, CO027, CO028]公开可支撑的关键指标(KPI)集中在融资、覆盖范围和产品能力,而不是收入或客户数。
现有客户数、收入或年经常性收入(ARR)有意排除,因为抓取材料没有披露。
[CO003, CO009, CO015, CO018, CO022, CO024]1.4 里程碑、商业证明与压力点
里程碑记录显示,公司正从架构命题推进到产品化、伙伴集成和早期公开工作负载证明。时间线从 2019 年创立开始,随后是 2022 年 Series A 及 Microsoft 工具链关系,再到 2023 年为商业化提供资金的 Series B。Corsair 于 2024 年 11 月发布,面向早期访问客户采样,并设定 2025 年 Q2 广泛可用目标。d-Matrix 随后扩展产品栈:2025 年 5 月与 GigaIO 合作,2025 年 9 月发布 JetStream,2025 年 10 月推出 SquadRack 参考架构,2025 年 11 月宣布 Alchip 路线图,2026 年 3 月 Gimlet Labs 基准声称在异构设置中带来 2–10x 延迟收益。同一记录也暴露压力点。公开材料仍高度依赖供应商声明和伙伴主导的商业化证明;独立报道则强调,d-Matrix 这类创业公司必须突破 Nvidia 已安装生态、扩大制造规模,并证明客户采用具备韧性。因此,本章最大的剩余缺口是商业披露和控制可见度,而不是身份或资本化。[CO021, CO022, CO023, CO024, CO025, CO026]
| 日期 | 事件 | 类型 | 金额或状态 | 参与方 | 含义 |
|---|---|---|---|---|---|
| 2019 | 公司成立 | 创立 | 成立 | Sid Sheth;Sudeep Bhoja | 确定起始日期,也解释了为什么后续材料称公司到 2025 年已有六年历史 |
| 2022-04 | Series A 和 Microsoft Bonsai 关系公开 | 融资 | $44M Series A;Microsoft 工具支持 | 参与方 / 来源:d-Matrix;VentureBeat;Microsoft Project Bonsai | 显示 Corsair 公开前,公司已获得早期生态入口 |
| 2023-09-06 | Series B 完成 | 融资 | Temasek 领投 $110M | 参与方 / 投资方:d-Matrix;Temasek;Playground;M12;SK Hynix;Marvell;Entrada | 为商业化提供资金,并扩大了战略资本基础 |
| 2024-11-19 | Corsair 发布 | 产品 | 首个公开推理平台发布 | d-Matrix;M12;OEM 和集成商合作者 | 推动公司从架构论点走向出货产品故事 |
| 2025-05-01 | GigaIO 合作宣布 | 合作 | 面向机架级推理的 SuperNODE 集成 | d-Matrix;GigaIO | 提供企业部署证明和系统级分发支持 |
| 2025-09-08 | JetStream 发布 | 产品 | 400Gbps 透明 NIC;预计年底投产 | d-Matrix | 将 d-Matrix 拓展到网络层,并让平台更完整 |
| 2025-10-14 | SquadRack 宣布 | 扩张 | 基于开放标准的机架级参考架构 | 参与方 / 生态伙伴:d-Matrix;Arista;Broadcom;Supermicro | 显示生态牵引和基于标准的部署策略 |
| 2025-11-12 | Series C 完成 | 融资 | 融资 $275M、估值 $2B;累计融资 $450M | 参与方 / 投资方:d-Matrix;Bullhound Capital;Triatomic;Temasek;QIA;EDBI;M12 | 确认后期私营公司状态,并为全球扩张提供资金 |
| 2025-11-18 | Alchip 合作宣布 | 合作 | 3DIMC 和 Raptor 后继路线图 | d-Matrix;Alchip | 指向 Corsair 之后的产品路线图 |
| 2026-03 | Gimlet Labs 基准发布 | 合作 | 异构设置中 2–10x 延迟改善主张 | Gimlet Labs;d-Matrix | 在外部环境中用 Corsair 提供公开工作负载证明 |
| 2026-04 | GigaIO 数据中心业务收购出现在新闻索引中 | 扩张 | Series C 后收购里程碑 | d-Matrix;GigaIO | 暗示 2025 年融资后仍在推进系统级扩张 |
日期按抓取记录中的可见形式保留。只有月份的里程碑仍保持月份粒度,不从记忆或模型假设中回填过度精确的日期。
[CO001, CO011, CO012, CO021, CO023, CO024]2019 到 2026 年间,公司从融资和架构定位走向产品发布、伙伴集成和公开负载验证。
只有月份的里程碑使用该月第一天,方便渲染器放到时间线上;这不表示抓取来源支持更精确日期。
[CO001, CO008, CO011, CO012, CO013, CO021]1.5 图表
02市场分析
2.1 市场边界,以及从训练转向推理
d-Matrix 的市场边界不是抽象的「AI」,甚至也不是所有加速器支出。公司公开材料、技术文档和发布信息都持续把公司放在数据中心推理场景:以低延迟、高内存带宽和更好的成本及能效,服务已经训练好的生成式模型。因此,纳入的支出包括加速卡、推理服务器和机柜、扩展服务所需的互连与 NIC 层,以及让这些系统可在企业或 neocloud 环境中部署的软件栈。排除的支出包括模型训练集群、永远不会变成商用硬件的超大规模云厂商自研芯片、纯 API 支出和边缘设备推理。这个区分重要,因为训练完成后技术瓶颈会变化。d-Matrix 和第三方开发者证据都强调,prefill 受计算约束,而 decode 与智能体服务更受内存和延迟约束;这正是通用 GPU 经济性最薄弱、专业加速器有机会切入的栈位置。[CM001, CM002, CM003, CM004, CM005, CM006]
| 细分 / 类别 | 包含支出 | 排除支出 | 买方 / 付款方 | 与 d-Matrix 的相关性 |
|---|---|---|---|---|
| 广义生成式 AI 支出 | 模型软件、服务、云消耗、基础设施和邻近工具 | 无法清晰拆分训练、推理、软件和服务 | CIO / CTO / 业务线预算 | 用于承保 d-Matrix 过宽 |
| AI 加速器市场 | 面向训练和推理 AI 工作负载的商用加速器硬件 | 通用 CPU 支出和大多数软件 / 服务 | 云厂商、企业、模型实验室、OEM | 可作为硬件外边界,但仍比 d-Matrix 更宽 |
| 数据中心生成式推理基础设施 | 面向已部署 LLM 的推理卡、服务器、机架、内存、互连和服务软件 | 模型训练集群、边缘 AI 设备、无关 HPC | 基础设施团队、平台团队、AI 产品运营方 | 最接近的直接类别 |
| 开放标准低延迟推理系统 | 基于 PCIe 和 Ethernet 的推理部署,使用商用加速器和合作伙伴服务器 | 超大规模云厂商自用定制芯片和专用内部网络架构 | 企业、neocloud、OEM / 集成商渠道 | d-Matrix 最相关切入点 |
| 现状替代栈 | Nvidia / AMD GPU 服务器、Groq 式推理云和外部 API 推理 | 尚未被采用的专门化商用加速器 | 既有云或平台所有者 | 主要替代对象 |
边界逻辑从极宽的 AI 支出,收窄到 d-Matrix 实际切入的更小商用数据中心推理层。第四行是可操作的承保边界;其上各行是背景,不是 SAM。
[CM001, CM002, CM003, CM004, CM005, CM007]2.2 需求驱动与结构性瓶颈
公开证据强烈支持一个品类判断:推理需求正在跑赢最初为训练而建的基础设施。CNN、VentureBeat 和 TechTarget 都描述了 GPU 稀缺、数据中心扩张压力,以及为追上 AI 需求而增加能源和网络容量的必要性。d-Matrix 自己的架构叙事与这些独立证据方向一致:内存搬运消耗电力,主机中转网络增加延迟,batch size 直接拉扯用户体验与硬件 ROI。推理和智能体工作流放大了这些压力,因为它们增加了推理时计算量,又不允许运营商用简单吞吐指标掩盖延迟。市场因此从训练优先的讨论转向运营和服务讨论。约束栈不只是原始 FLOPS,而是电力、内存带宽、集群通信开销,以及在规模化交付快速答案时不让太多昂贵硬件闲置的经济性。[CM008, CM009, CM010, CM011, CM012, CM013]
| 因素 | 方向 | 时点 | 对 d-Matrix 的影响 | 尽调问题 |
|---|---|---|---|---|
| 推理型与智能体工作负载拉高推理时计算量 | 驱动因素 | 当前且在加速 | 更多服务任务开始受延迟和内存约束,因而支撑品类投资逻辑 | 量化管线需求中,推理型 / 智能体相对传统聊天占多少 |
| GPU 短缺与数据中心建设压力 | 驱动因素 | 当前 | 让买方更愿意评估替代方案和异构服务架构 | 要求提供因 GPU 供应不足而让 d-Matrix 进入评估的真实买方案例 |
| 内存搬运功耗负担 | 驱动因素 | 结构性 | 相比计算与 HBM 分离的堆栈,强化 d-Matrix 的 DIMC / SRAM 叙事 | 在代表性内存受限工作负载上测试每次请求功耗 |
| 主机中转的多节点网络开销 | 驱动因素 | 规模化后具结构性 | 支撑 JetStream 和透明 NIC 在机架级推理中的定位 | 验证真实多节点流量下的生产延迟和抖动 |
| 交互式应用的批大小经济性 | 驱动因素 | 当前 | 有利于在小批量下保持高利用率的硬件,而不是只优化峰值吞吐 | 获取客户证据,确认可接受延迟阈值和收入敏感性 |
| Nvidia 软件护城河和基准领先地位 | 约束 | 当前 | 限制 d-Matrix 在极特定工作负载楔子之外的采用 | 对比 GPU 堆栈,测试软件成熟度、工具摩擦和部署耗时 |
| 超大规模云厂商自用定制芯片增长 | 约束 | 当前且在上升 | 压缩初创公司可争取的广义推理市场商用份额 | 估算目标需求中已有多少锁定内部芯片或预留 GPU 供应 |
| 需要生产环境 TCO 证明,而不只是供应商说法 | 约束 | 眼下 | 独立的工作负载级证据仍弱于市场叙事 | 要求提供已签约生产部署、参考架构和客户 ROI 证据 |
这张表把结构性驱动因素和结构性约束放在一起,因为扩大推理需求的同一批力量, 也会抬高初创供应商的举证门槛。每一行都应当被当作投资判断问题,而不是认定需求能顺畅转化为收入。
[CM008, CM009, CM010, CM011, CM012, CM013]2.3 买方细分、预算所有者与采用路径
买方地图比泛泛的「企业 AI」表述更窄。d-Matrix 最现实的早期细分,是拥有低延迟敏感推理负载的企业或私有云运营商、neocloud 和托管推理供应商、试图控制服务经济性的 AI 产品公司,以及把卡封装进系统的 OEM 或集成商伙伴。超大规模云厂商仍具战略意义,但结构上更难拿下,因为它们已经控制大型 GPU 资产,并越来越多地建设自有芯片。d-Matrix 的开放 PCIe 和 Ethernet 定位,再加上风冷机柜叙事,更自然地匹配那些希望把推理塞进既有数据中心版图、而不愿围绕液冷巨型集群重建设施的客户。采用路径也相应偏重基础设施:先识别受内存约束或对延迟敏感的工作负载,在试点或单节点部署中验证匹配度,然后在网络、模型切分和运营成本跑通后,扩展到服务器或机柜级生产。集成商和 OEM 很重要,因为买方买的是可部署系统,而不只是一颗芯片。[CM016, CM017, CM018, CM019, CM020, CM021]
| 细分市场 | 买方 | 用户 | 付款方 | 工作流 | 预算负责人 | 采用触发因素 |
|---|---|---|---|---|---|---|
| 企业私有云 / 受监管运营方 | 基础设施或平台工程团队 | ML 平台、应用和运维团队 | 企业 IT / 基础设施预算 | 内部 Copilot、支持型 Agent、私有检索和推理 | CTO / 基础设施副总裁 / AI 平台负责人 | 低延迟、数据控制和可预测服务成本的需求 |
| 新云厂商或托管推理提供商 | 云产品与容量规划团队 | 外部开发者客户 | 云资本开支和服务利润率 | 托管模型服务和推理 API 产品 | 云业务总经理 / 产品副总裁 / 财务 | GPU 推理服务的利润率压力,以及差异化延迟需求 |
| AI 产品公司 / 模型 API 运营商 | 应用 AI 或平台团队 | 编程、客服或 Agent 产品的终端用户 | 产品 P&L | 交互式、内存受限的推理工作负载 | CPO / CTO / 基础设施负责人 | 用量快速冲高,推理经济性亏损 |
| OEM / 集成商渠道 | 服务器 OEM 或系统集成商 | 解决方案架构师和部署工程师 | 终端客户项目预算 | 将加速卡、服务器、网络和支持打包成可部署系统 | 业务单元负责人 / 解决方案业务负责人 | 客户对交钥匙部署的需求 |
| 超大规模云厂商 / 自用芯片买方 | 云基础设施组织 | 内部模型和平台团队 | 巨额数据中心资本开支 | 超大规模基础模型和云推理 | 基础设施高级副总裁 / 自研芯片组织 | 通常更难让 d-Matrix 赢单,因为既有 GPU 和内部芯片方案都很强 |
前四行是现实的商用市场目标。超大规模云厂商这一行也列入,因为它有战略意义; 但公开证据显示,这个细分市场最难被一家商用加速器初创公司直接触达。
[CM016, CM017, CM018, CM019, CM020, CM021]现实买方分段按用户画像、付款模式、预算负责人,以及最可能推动 d-Matrix 式部署进入生产的触发因素来映射。
[CM018, CM019, CM020, CM021, CM041, CM042]商用推理采用从负载痛点收窄到试点证明,再到生产推广;现有软件和容量替代方案始终构成摩擦节点。
该流程是定性图。它反映来源包暗示的基础设施采购路径,而非已发布转化漏斗;风险节点显示现有供应商和 API 抽象在哪些位置会阻止采用走到全面铺开之前。
[CM018, CM020, CM021, CM037, CM043, CM044]2.4 规模测算视角,以及真正可触达的市场
公开市场数字很大,但不能直接相加,也不都描述同一个商业表面。Gartner 的生成式 AI 支出预测是最宽的相邻口径,包含软件、服务和基础设施。MarketsandMarkets 的 AI 推理估计更窄,但仍比商用芯片本身更宽。Silicon Analysts 的 AI 加速器市场在硬件轴上再窄一些,但仍包含训练和多种现有或自研架构。来源包中最有决策价值的下限视角甚至不是 TAM 报告:CNBC 估计,仅 Bing AI 就至少需要 $4B 基础设施。这个部署成本视角抓住了该品类存在的原因。d-Matrix 可触达的 SAM 小于上述所有数字,因为它排除了训练、排除了大量超大规模云厂商自研芯片,也排除了满足于通过外部 API 消费推理的买方。公开数据足以证明品类很大,但不足以在没有管理层披露的情况下发布干净的 d-Matrix SOM。[CM022, CM023, CM024, CM025, CM026, CM027]
| 发布方 / 口径 | 年份 | 地理范围 | 数值 / 指标 | 增长 | 方法论视角 | 置信度 | 局限性 |
|---|---|---|---|---|---|---|---|
| Gartner 经 Forbes 转引 | 2025 | 全球 | $644B 生成式 AI 支出 | 同比 76.4% | 覆盖硬件、软件和服务的极宽口径支出预测 | 低 | 口径远宽于推理硬件或商用加速器 |
| MarketsandMarkets 经 Forbes 转引 | 2025-2030 | 全球 | $106.15B 至 $254.98B AI 推理市场 | 隐含多年强劲增长 | 2025 年市场评论引用的推理市场预测 | 低 | 范围不止商用数据中心硬件 |
| Silicon Analysts | 2025E-2026E | 全球 | $160B 至 $200B+ AI 加速器总市场 | 快速增长,份额也向更多玩家扩散 | 加速器硬件市场的收入视角 | 低 | 混合训练与推理,也包含既有厂商和自用芯片敞口 |
| CNBC 基础设施视角 | 2023 | 全球 / Bing 参考案例 | 服务 Bing AI 用户至少需要 $4B 基础设施 | N/A | 部署成本视角,而非 TAM | 中 | 单一服务案例,不是市场规模 |
| d-Matrix 产品架构 | 2025-2026 | 机架 / 数据中心 | 性能模式最高支持 100B 参数模型;容量模式支持 1T+ | 随加速卡、服务器和机架扩展 | 商用推理平台的技术部署边界 | 中 | 能力表述,不是收入 TAM |
| 分析师综合判断 | 2026 | 全球商用切片 | 公开资料中没有清晰的独立 SAM 或 SOM | N/A | 基于资料包证据约束得出的结论 | 中 | 需要管理层披露部署量、每个集群的卡数和买方结构 |
这些视角有意不做加总。前三行的市场定义分母各不相同;第四、第五行看的是部署经济性 / 能力边界; 最后一行记录公开资料里仍缺失的 SAM/SOM,而不是凭空造一个数。
[CM022, CM023, CM024, CM025, CM026, CM027]分层视角从广义生成式 AI 支出收窄到更小的部署成本和外售硬件切口;后者才真正关联 d-Matrix。
这些层不是嵌套数据集。它们都用 $B 单位展示逐步收窄的商业视角:广义支出、加速器硬件、推理市场, 以及一个具体大型部署成本案例。目的在于强制边界纪律,而不是暗示数值会相加或机械递进。
[CM022, CM023, CM024, CM025, CM026, CM040]公开 $B 视角差异很大,取决于分母是广义生成式 AI 支出、推理市场收入、加速器硬件, 还是单一大型部署成本案例。
所有行都使用同一个 $B 单位,但它们描述的市场面不同。AI 推理行的中点是算术中点,并非发布方提供; 这里只用于显示 2025 与 2030 端点之间的离散度。
[CM022, CM023, CM024, CM025, CM026, CM045]2.5 相互矛盾的基准与市场份额叙事
本章的核心矛盾是:推理机会显然在扩张,但公开的份额获取证据仍压倒性地偏向现有厂商。d-Matrix 自有材料和 Gimlet 伙伴支持的工作负载显示,在受内存约束或异构推理阶段,特别是小 batch 延迟重要的地方,产品可能具备真实性能优势。但标准化基准覆盖、主流 GPU 产品页和市场份额估计仍显示,Nvidia 在多数公开比较中设定可见前沿。即便市场份额数字彼此也不一致,因为分母会变化:一个来源谈 AI 芯片,另一个谈独立 GPU,第三个谈包含自研芯片在内的全部加速器。这并不意味着创业公司命题错误,而是意味着命题必须更具体。投资人应把 d-Matrix 承销为切入商用、低延迟敏感、开放标准推理部署的楔子,目标买方重视可预测 TCO 和部署匹配;而不是承销成一个泛泛声称将取代所有推理负载 GPU 栈的公司。[CM028, CM029, CM030, CM031, CM032, CM033]
2.6 图表
03竞争格局
3.1 竞争格局概览
d-Matrix 身处一个拥挤但仍在流动的 AI 推理市场。最重要的直接替代品不是一组同质化同业:NVIDIA 和 AMD 销售可同时跑训练与推理的通用加速器栈;Groq、Cerebras 和 SambaNova 走推理优先或紧密集成路线;许多买方的现状仍是基于现有 GPU 云或私有机柜内部自建。独立市场报道反复提到同一组结构性压力:推理工作负载增长速度超过买方吸收电力、散热、内存和 GPU 采购成本的能力,这也是 d-Matrix 这类挑战者首先获得关注的原因。d-Matrix 自己的主张是,Corsair 卡、JetStream NIC 和 Aviator 软件攻击让大模型推理变贵的内存搬运惩罚。因此,公司竞争时不太像通用 GPU 替代品,更像是针对推理需求中受内存约束、低延迟敏感切片的专用答案。[CP001, CP002, CP025, CP026, CP029, CP032]
| 竞争者 | 类别 | 规模 / 融资 | 目标细分市场 | 关键差异化 | 主要局限 |
|---|---|---|---|---|---|
| d-Matrix | 只做推理的挑战者 | 据 2025 年报道,累计融资 $450M | 企业和数据中心 LLM 推理 | 数字内存内计算;Corsair 加速卡 + JetStream NIC + Aviator 软件;可插入标准服务器 | 公开基准证明和生产客户深度仍有限 |
| NVIDIA H100/H200/GB200 | 既有 GPU 堆栈 | 企业和云端装机基础占主导;不同来源和年份估计份额约 75-92% | 覆盖企业到超大规模的训练与推理 | CUDA/TensorRT/NIM、NVLink、InfiniBand、OEM 和云渠道触达 | 资本开支高,电力和冷却负担重;买方在仅推理场景下仍可能过度配置 |
| AMD MI300X | 主流 GPU 替代方案 | 通过主要 OEM 服务器渠道规模化 | 需要更高单 GPU 内存的超大规模和企业推理 | 192GB HBM3、5.3TB/s 带宽、ROCm 软件,与 GPU 相同的服务器采购路径 | 软件既有优势弱于 CUDA,端到端平台心智也较弱 |
| Groq | 推理服务挑战者 | 私营公司;提供公有、私有和 co-cloud 方案,以及本地部署选项 | 交互式低延迟推理、受监管和物理隔离工作负载 | LPU 架构、可预测支出、GroqRack 本地部署连续性 | 编译器 / 运行时路径不同于标准 GPU 堆栈 |
| Cerebras | 集成式大模型系统 | 拥有晶圆级平台的大型私营 AI 芯片公司 | 超大模型训练与推理 | 晶圆级引擎、巨大的片上算力、系统级内存策略 | 专用设备模式不如标准机架易插即用 |
| SambaNova | 集成式企业平台 | 据独立指南,2026 年融资轮 $350M+ | 智能体 AI 和企业推理部署 | 硬件、软件和模型协同调优;SN50 在性价比上对标 Blackwell | 采用平台意味着买入一套专有全栈 |
| 内部自建 / 定制芯片 | 替代方案 / 现状方案 | 依靠超大规模云厂商和大型企业资本开支预算,而非供应商融资 | 最大买方出于数据治理、延迟或利用率原因自有堆栈 | 可围绕自有工作负载优化,并避开供应商依赖 | 需要庞大的工程、采购和运营能力 |
覆盖范围并不完整,而是聚焦 2026 年企业推理采购中最影响决策的选项:既有 GPU 堆栈、 推理优先挑战者和内部自建替代。规模字段只在公开披露时引用融资、份额或部署背景。
[CP001, CP013, CP014, CP018, CP021, CP023]按生态杠杆(x 轴)与推理专项化程度(y 轴)给 d-Matrix 和最重要替代方案做序数映射。
坐标轴是基于公开证据的序数估计,并非厂商发布的数值尺度。X 轴衡量生态、渠道和部署杠杆; Y 轴衡量产品对推理经济性或延迟敏感服务的专项化程度。
[CP018, CP021, CP023, CP029, CP032, CP040]3.2 现有 GPU 堆栈:NVIDIA 与 AMD
即使买方说想要「更好的推理经济性」而不是「最快的原始集群」,NVIDIA 仍是 d-Matrix 必须在实践中击败的基准。H100、H200 和 GB200 覆盖从单 GPU 部署到机柜级 NVLink 域的完整梯队,NVIDIA 又在芯片之上叠加软件、网络和企业支持。H100 以 NVLink、InfiniBand、Magnum IO 和 AI Enterprise 锚定已安装基础;H200 靠更高内存容量和带宽进一步加深大模型推理护城河;GB200 再次抬高竞争门槛,把 72 颗 GPU 收进液冷机柜级系统,并由 NVIDIA 营销成一台巨型推理机器。AMD 是最清晰的主流替代方案,因为 MI300X 提供更大单 GPU 内存、高带宽和 ROCm,而不是 CUDA。但 AMD 仍在与 NVIDIA 相同的 OEM 和数据中心采购动作中竞争,这意味着它会挤压现有厂商经济性,却不会彻底改变买方评估或部署推理基础设施的方式。[CP006, CP007, CP008, CP009, CP010, CP011]
| 采购标准 | d-Matrix | NVIDIA | AMD | Groq | Cerebras | SambaNova |
|---|---|---|---|---|---|---|
| 小批量交互延迟优化 | 高 | 中 | 中 | 高 | 中 | 中 |
| 单元 / 系统的大模型内存承载力 | 中 | 高 | 高 | 中 | 高 | 高 |
| 成熟软件生态和框架可移植性 | 起步 | 很高 | 中等 | 中等 | 中等 | 中等 |
| 机架级网络和 OEM 可获得性 | 起步 | 很高 | 高 | 中等 | 中等 | 中等 |
| 物理隔离 / 本地部署路径 | 伙伴主导 | 高 | 高 | 高 | 高 | 高 |
| 价格透明度 / 按用量打包 | 低 | 低 | 低 | 中 | 低 | 低 |
| 推理买方复用训练生态 | 低 | 很高 | 高 | 低 | 低 | 低 |
评级是作者基于官方产品页面和独立市场分析综合得出的评估。它们概括与买方相关的匹配度, 而不是标准化基准分数;“起步”或“低”通常反映公开证明缺口,并不代表技术已被证明缺失。
[CP013, CP014, CP017, CP018, CP020, CP022]3.3 专用推理挑战者
非 GPU 挑战者阵营很分散,这一点重要,因为 d-Matrix 并不是在对抗一个单一的「推理芯片」原型。Groq 把低延迟包装成服务,提供公有、私有、co-cloud 和本地部署选项;在部署模式和支出可预测性上,它是 d-Matrix 最清楚的对照。Cerebras 走向相反,使用晶圆级引擎和超大系统内存,让客户把更多模型留在专用设备本地。SambaNova 把硬件、软件和调优模型打包成一体化平台,并直接拿该组合在性价比上对标 NVIDIA Blackwell。因此,独立格局调研把市场框定为同一问题的不同答案:买方应当用广泛 GPU 生态、服务主导的低延迟专家、巨型一体化系统,还是更容易嵌入标准企业服务器的内存就近卡来解决推理?d-Matrix 确实坐在这张桌子上,但尚未拥有默认品类叙事。[CP018, CP019, CP020, CP021, CP022, CP023]
| 供应商 / 选项 | 价格 / 单位 / 合同模式 | 包含能力 | 折扣 / 未知项 | 竞争含义 |
|---|---|---|---|---|
| d-Matrix | 无公开标价;公司称成本比基于 GPU 的系统低 3x | Corsair 加速器、JetStream NIC、Aviator 软件、标准 PCIe / 服务器部署 | 实际成交价、利用率假设和支持条款未披露 | 若在生产中验证会很强,但公开 TCO 证明仍薄弱 |
| NVIDIA H100 / H200 堆栈 | 硬件叠加企业软件和 OEM / 云合同;公开标价整体不透明 | CUDA、TensorRT、Triton、AI Enterprise/NIM、NVLink、InfiniBand、广泛 OEM 可获得性 | 批量折扣和云端转嫁价格不透明 | 买方若把生态确定性看得高于推理专用优化,默认会走这条采购路径 |
| NVIDIA GB200 NVL72 | 按机架级系统销售,而不是采购商品化加速卡 | 72-GPU NVLink 域、液冷、Mission Control、AI 工厂工具 | 公开交易价格未披露;设施升级可能很重 | 抬高性能门槛,但也把买方池收窄到超大部署 |
| AMD MI300X | 通过 OEM 服务器项目和云合同销售,而非公开街价 | 192GB HBM3、ROCm 堆栈、OAM 模块封装、OEM 服务器可获得性 | 净价和支持打包不公开 | 在熟悉的 GPU 采购路径内改善内存经济性,以此竞争 |
| Groq | Free、Developer 和 Enterprise 方案;按用量计费和定制企业合同 | 公有 / 私有 / co-cloud 接入、区域端点、提示缓存、可选 GroqRack 本地部署 | 单模型企业经济性需谈判,未完全公开 | 挑战者中最透明的服务式包装 |
| Cerebras / SambaNova | 系统级企业合同;公开价格缺乏统一口径 | 集成硬件加软件堆栈、调优后的大模型服务、企业支持 | 实际 token 经济性和设备价格大多未公开 | 靠交钥匙集成价值竞争,而不是开放组件对比 |
| 基于既有 GPU 云的内部自建 | 历史上可能意味着极高月度账单,加上数据中心升级或巨额 GPU 预留 | 最大化工作流控制,并复用既有 ML 工具 | 真实成本取决于模型大小、流量、利用率和预留容量 | 这种替代方案仍然昂贵,但熟悉,因此黏性强 |
推理硬件的公开定价很少,因此本表强调包装方式、合同形态和可见打包内容。 凡涉及成本影响的行,都区分供应商自述主张和 GPU 时代运营支出的独立案例。
[CP001, CP018, CP019, CP027, CP028, CP035]从买方适配角度展示各类厂商在常见推理负载需求上的强项。
单元格总结公开材料和独立分析中的最佳适配取向;应把它们视为相对买方适配,而不是经基准测试的性能分数。
[CP020, CP022, CP024, CP029, CP030, CP031]3.4 切换成本、分销与内部自建
核心竞争问题不在于 d-Matrix 能否讲出强硅片论证,而在于它能否越过让企业留在现有平台上的软件、网络和渠道惯性。NVIDIA 的优势深植系统:买方已经知道如何购买 GPU 集群、给它们散热、连接它们、用 CUDA 和 TensorRT 编程,并通过熟悉的 OEM 和云渠道采购。AMD 用 ROCm 和更大单 GPU 内存在边际上削弱这道护城河,但仍活在同一套机柜和服务器生态里。Groq、Cerebras 和 SambaNova 通过改变包装模式来攻击问题——API 主导服务、巨型设备或全一体平台。d-Matrix 站在这些阵营之间。模块化 chiplet 和卡式路径可能降低插入摩擦,但公开记录仍缺客户参考、价格表和规范化基准,无法证明买方愿意切换、多栖或在生产规模上标准化采用它。内部自建也是一个真实替代方案,因为超大规模云厂商和大型企业越来越多地围绕自身工作负载设计或共同设计推理芯片。[CP027, CP028, CP030, CP033, CP034, CP035]
3.5 护城河耐久性与反向证据
d-Matrix 具备可信但仍暂定的护城河。最强支柱是技术匹配:多个来源把推理成本和电力痛点连接到内存搬运,而 d-Matrix 的数字内存内计算加 chiplet 互连策略正是围绕这个瓶颈打造。第二个支柱是模块化;Jayhawk、卡、NIC 和软件暗示一套异构部署叙事,可能比专有超级设备更容易插入。但反向证据也有分量。NVIDIA 仍拥有最成熟的基准记录、软件栈和分销版图。AMD 持续缩小内存论点。Groq 拥有最清晰的可预测低延迟服务公开包装。Cerebras 和 SambaNova 为超大模型销售更紧密集成的系统。最重要的是,d-Matrix 公开的性能和成本主张仍主要由供应商撰写,而不是由独立、同口径的基准套件规范化。护城河耐久性因此处于「中等」:差异化足以重要,但尚未坚固到可以忽视现有厂商或一体化系统反击风险。[CP037, CP038, CP039, CP040]
| 护城河主张 | 威胁 | 严重性 | 缓释措施 / 有帮助的证据 | 尽调问题 |
|---|---|---|---|---|
| 数字内存内计算更匹配推理的内存瓶颈 | H200/GB200/下一代 GPU 的内存和带宽提升会收窄效率差距 | 高 | 与 H100/H200/MI300X 对比的独立同模型能耗和延迟基准 | 哪个公开基准能在相同模型、批大小和延迟目标下证明 d-Matrix 声称的增益? |
| 模块化 chiplet 加卡 / NIC 架构降低部署摩擦 | 买方可能仍偏好已验证的 NVLink/NIM/OEM 堆栈,或挑战者的交钥匙设备 | 高 | 更多具名 OEM 系统和生产客户案例 | 目前哪些服务器 OEM 交付带受支持软件镜像、可投产的 d-Matrix 系统? |
| 只做推理让 d-Matrix 产品逻辑比通用 GPU 更尖锐 | 训练加推理平台仍会赢下采购,因为用途更宽 | 中 | 证明买方会把推理预算同通用 AI 基础设施预算拆开 | 企业有多常专门为推理采购 d-Matrix,而不是扩容现有 GPU 集群? |
| 声称的更低成本和能耗提升规模化推理 ROI | 这些说法大多仍来自厂商,同口径归一化后未必站得住 | 高 | 第三方基准测试,或公开客户 BOM 证据 | 3x 更低成本说法背后,对利用率、模型规模和软件作了哪些假设? |
| 异构系统策略可能适合受监管或私有化部署 | Groq 已经有更清晰的云到本地包装,Cerebras/SambaNova 卖的是更一体化的系统 | 中 | 展示可复用的本地部署参考架构和支持模型 | d-Matrix 是否有像 GroqRack 一样清晰的公开部署路径,或能对标交钥匙设备的方案? |
| 推理市场增长为挑战者留出空间 | 内部自研和定制芯片可能先于 d-Matrix 吃掉最高价值负载 | 中 | 聚焦缺乏超大规模云厂商级设计预算的客户 | 哪些客户细分最不可能用内部芯片替代第三方推理硬件? |
严重程度是基于 2026-05-26 可得公开记录的定性判断。这是一张耐久性风险登记表,不是概率预测,因此优先列出若不回应、最可能压缩 d-Matrix 差异化的威胁。
[CP028, CP033, CP035, CP037, CP038, CP039]紧凑展示塑造 d-Matrix 相对现有巨头和挑战者竞争耐久性的关键数字。
[CP001, CP003, CP008, CP014, CP032]3.6 图表
04财务情况
4.1 收入模式与变现面
公开记录指向硬件平台收入模式,而不是纯软件或纯 API 生意。d-Matrix 反复把自己讲成一套由 Corsair 加速器、JetStream 网络和 Aviator 软件构成的全栈推理平台,商业化经由超大规模、企业、主权、OEM 和系统集成商渠道推进。关键在于,收入质量大概率取决于硬件出货和多组件部署赢单,而不是自助式软件采用。GTM 证据也明显偏资格认证。Series B 被描述为让公司开始商业化 Corsair 的一轮融资,而 2024 和 2025 年产品公告仍依赖采样、早期访问客户、OEM 认证和伙伴配置系统。SquadRack、SuperNODE 和 Gimlet 云集成都暗示,合同价值可能高于单卡销售,但也意味着服务器、网络和云伙伴之间共享经济性。关键是,d-Matrix 不公布 Corsair 标价、JetStream 定价、软件许可证定价或已预订客户数。即便生态页面也提醒,伙伴 logo 只是示意,并非实际客户。因此,变现面真实且在变宽,但公开证据仍更清楚地显示渠道形成和试点转化,而不是已实现收入。[CI001, CI002, CI003, CI004, CI005, CI006]
| 收入流 | 机制 | 买方 / 渠道 | 当前公开状态 | 收入质量 | 尽调问题 |
|---|---|---|---|---|---|
| Corsair 加速器 | 销售 PCIe 加速器卡和双卡配置 | 企业、超大规模云厂商、主权云、OEM、系统集成商 | late 2024 开始送样;目标 Q2 2025 广泛供货 | 可能是高 ACV 硬件收入,但公开资料没有 ASP、出货量或毛利率 | 索取出货台数、ASP、毛利率,以及试点转生产转化率 |
| JetStream NIC | 销售用于扩展 Corsair 集群的专用网卡 | 随多节点部署和机架方案销售 | Sep 2025 可送样;完整量产目标为 year-end 2025 | 可能带来附加收入,但附加率和定价未披露 | 索取按部署规模拆分的 NIC 附加率、BOM 和实际定价 |
| Aviator 软件 | 与硬件和部署工具捆绑的推理软件 | 企业和 OEM 部署团队 | 反复作为全栈平台的一部分营销;未披露单独价格 | 可能形成经常性软件收入,但公开证据只支持捆绑层面的变现 | 索取授权模式、维护条款,以及续费或支持收入 |
| 集成机架和服务器方案 | 伙伴配置的系统,例如 SquadRack 和 GigaIO SuperNODE | OEM 和基础设施渠道伙伴 | 参考架构和联合方案已公开;已入账收入未公开 | 合同金额可能更高,但经济收益可能要同服务器和网络伙伴分成 | 索取谁向客户开票、系统 ASP 和伙伴收入分成 |
| 异构云访问 | 通过 Gimlet Cloud 向精选客户提供 Corsair 访问,与 GPU 并行 | AI 原生模型提供商和云买方 | 目标在 2H 2026 面向精选客户开放 | 可能成为按使用量计费的证明,但当前证据仍处早期访问和基准测试阶段 | 索取合同结构、收入分成和试点转生产转化率 |
这些行只限于公开产品、伙伴和部署材料中明确可见的变现场景。材料包没有任何公开来源披露实际定价、收入结构或合同期限。
[CI001, CI006, CI007, CI008, CI009, CI010]| 产品 | 公开标价 | 公开性能 / 成本说法 | 实际定价可见度 | 变现含义 | 来源状态 |
|---|---|---|---|---|---|
| Corsair 卡硬件 | 未公开披露 | 交互速度最高快 10x,成本性能比 GPU 高 3x | None | 没有价目表或客户交易数据,就无法检验定价权 | 有官方产品说法;公开定价不存在 |
| JetStream 网络 | 未公开披露 | 400 Gbps NIC;与 Corsair 搭配时成本性能比提升 3x | None | 多节点附加可能抬高 ACV,但网络 BOM 也会上升 | 有规格表;无价格表 |
| Aviator 软件 | 未公开披露 | 企业级软件栈,是平台叙事的一部分 | None | 可能有经常性软件上行空间,但独立授权公开不可见 | 仅有捆绑营销 |
| GigaIO SuperNODE + Corsair | 未公开披露 | 单节点容纳 64+ Corsair,部署复杂度更低,TCO 更优 | None | 暗示有系统级合同潜力,但没有公开伙伴定价或收入分成 | 伙伴推广材料 |
| Gimlet Cloud 访问 | 未公开披露 | 异构推理基准中,在同等能效下延迟改善 2x-10x | None | 可能形成按使用量计费收入,但客户定价和利润率仍未公布 | 仅限精选客户早期访问 |
本表有意把可用性和性能话术同实际变现拆开。审阅来源没有给出 Corsair、JetStream、软件或集成系统的公开标价。
[CI008, CI017, CI019, CI024, CI036]公开证据指向一条以资格验证为先导的收入链:先从负载适配出发,经伙伴支持的硬件部署推进, 之后才扩展到软件和更大系统收入。
该顺序来自公开商业化表述,而非管理层报告的转化数据。定价、附加率和合同时长仍未披露。
[CI001, CI006, CI007, CI008, CI010, CI011]4.2 单位经济模型与成本驱动
d-Matrix 的单位经济模型故事容易描述,但公开证据仍只覆盖一部分。官方材料持续声称,相比 GPU 替代方案,交互速度最高快 10x,成本性能好 3x,能效好 3x 至 5x;第三方 Gimlet 测试又给出更具体的工作负载代理:把 speculative decode 卸载到 Corsair,在相同能效下把端到端请求延迟降低 2x 至 10x。这是有意义的证据,说明产品能改善客户在受内存约束推理阶段的经济性。但它不同于证明 d-Matrix 自身毛利率或回本曲线。白皮书和 JetStream 简报明确把成本和功耗数字标为初步数据,两份文件也说明为什么系统经济性不能简化为芯片基准。JetStream 会增加专用 400 Gbps NIC、光模块或 DAC、功耗、交换机端口和机柜集成工作。Corsair 服务器和机柜规格暗示更大的平均订单额,但也暗示服务器、网络和支持成本会落在加速器裸片之外。外部市场代理强化了这一点的重要性:AI 推理可能吞掉巨额运营预算,而 GPU 基础设施仍昂贵、耗电且供应受限。d-Matrix 可能改善这些经济性,但公开证明仍停留在基准和架构层,而不是损益表层。[CI013, CI014, CI015, CI016, CI017, CI018]
| 指标 / 代理项 | 公开数值 / 状态 | 置信度 | 重要性 | 尽调问题 |
|---|---|---|---|---|
| 声称的成本性能改善 | 相较 GPU 替代方案,性能 / TCO 高 3x | 中 | 这是企业买方的核心 ROI 主张 | 用客户负载和完整系统 BOM 复现 |
| 声称的能效改善 | 相较 GPU 替代方案,能效高 3x 到 5x | 中 | 电力和散热是推理经济性的核心 | 比较机架级功耗,纳入交换机、光模块和空闲开销 |
| 声称的交互速度改善 | 交互速度最高快 10x;Llama 70B 达 30K tokens/s、2 ms/token | 中 | 支撑延迟敏感负载值得使用专用硬件的逻辑 | 索取基准测试方法、提示词组合和并发假设 |
| JetStream 网络开销 | 400 Gbps NIC、PCIe Gen5 x16、最高 150W TDP、辅助供电输入 | 高 | 网络是成本栈的一部分,不是免费附加项 | 索取节点级功耗预算和收发器成本 |
| 服务器和机架密度代理 | 8 卡服务器最高 2 TB 容量内存;100B 模型用 8 节点机架;GigaIO 称单节点可容纳 64+ Corsair | 中 | 潜在交易规模很大,但部署复杂度和机架成本也会上升 | 索取典型每单卡数、节点数和已安装机架配置 |
| 第三方异构基准代理 | Gimlet 报告,在同等能效下,推测解码卸载延迟降低 2x-10x;1.6B 草稿模型可放入 2 张卡 | 中 | 目前最好的公开证据显示,专用卸载可改善真实工作负载经济性 | 索取按工作负载拆分的生产吞吐、利用率和胜率数据 |
本表混合了厂商说法、已发布规格和第三方基准代理项。不能把它读成 d-Matrix 披露了毛利率、CAC 或回本周期数据。
[CI014, CI015, CI016, CI017, CI018, CI019]d-Matrix 的公开投资回报(ROI)逻辑是:把内存受限的推理阶段迁移到专用硬件,同时让系统其他部分兼容标准机架和网络。
这张图把公开架构和基准证据转成商业逻辑链。它不是已披露成本模型,也不应被解读为毛利率桥。
[CI017, CI018, CI024, CI025, CI033, CI050]固定披露值和公开基准区间框定了财务讨论:资本已知,但大多数经营指标未披露,因此可用的公开区间在声称效率上, 而不是公司损益。
前两行是固定披露值。后几行是基准测试或厂商主张区间,不是经审计财务指标。d-Matrix 收入、 烧钱速度或毛利率没有公开区间。
[CI002, CI017, CI024]4.3 资本强度、伙伴依赖与融资依赖
融资形成是财务图景中最清楚的部分。d-Matrix 已公开披露累计融资 $450M,包括 2023 年 $110M Series B,以及 2025 年以 $2B 估值完成的 $275M Series C。两轮融资都明确绑定商业化、路线图执行、招聘、全球扩张和大规模部署。这是正面因素,因为公司并不是试图用很少资金扩展一套复杂的推理硬件栈。但这也是警示信号:经营计划目前看起来更由融资支撑,而不是由客户资金支撑。路线图本身资本强度很高。Corsair 依赖 chiplet、高带宽内存集成、服务器认证和伙伴分销。JetStream 增加专用网络硬件。下一代 3DIMC 路线图又明确依赖 Alchip 做 ASIC 设计、先进封装和制造管理。公开同业文件显示,这通常在实践中意味着库存承诺、押金、长交期、板卡和内存成本,以及需求时点偏移时的利润率压力。d-Matrix 借助伙伴而不是垂直拥有每个系统组件,可能从中受益,但这只是把部分固定成本换成对 Supermicro、Arista、Broadcom、GigaIO、云伙伴和先进封装供应商的执行依赖。没有公开现金、烧钱速度或现金跑道,投资人无法判断当前资产负债表是足以舒适覆盖扩张要求,还是只是在把公司桥接到下一轮融资。[CI002, CI003, CI004, CI005, CI019, CI020]
| 项目 | 公开数值 / 状态 | 证据 | 承销推断 | 尽调问题 |
|---|---|---|---|---|
| 已披露累计融资 | $450M | Series C 官方公告、投资方公告和独立报道相互吻合 | 私有推理硬件公司能拿到这样的资金,资本实力强,但不能证明经营已能自我供血 | 核对股权结构表、新股与老股交易资金,以及清算优先权 |
| 最新轮次规模 / 估值 | Series C 轮 $275M,估值 $2B | 官方资料,加上 Bullhound、QIA 和新闻报道 | 显示 late 2025 仍能获得投资人支持和估值支撑 | 索取投后持股、董事会权利和优先股堆叠 |
| Series B 过桥资本 | 2023 年 $110M,用于开始商业化 Corsair | Series B 官方公告 | 确认在 2025 融资前,公司经历了多年收入前或早期收入阶段的建设 | 索取 Series B 到 Series C 的支出桥 |
| 披露资金用途 | 路线图、招聘、全球扩张和大规模部署 | Series B 和 Series C 官方材料 | 资金指定用于执行和规模化,而不是已披露盈利 | 索取 24 个月经营计划和部署资本开支计划 |
| 现金余额 / 烧钱速度 / 现金跑道 | 未公开披露 | 审阅的官方、投资方或独立来源没有给出数字 | 无法用公开材料检验现金跑道和下一轮融资时点 | 索取当前现金余额、月度烧钱速度,以及基准和下行情景下的现金跑道 |
| 债务 / 项目融资 / 采购承诺 | d-Matrix 未公开披露;公开硬件可比公司显示这些项目可能影响利润率 | D-Matrix 披露缺口,加上 Nvidia 申报文件作为代理参照 | 即使固定 capex 有伙伴协助,硬件规模化仍可能藏着营运资本需求 | 索取供应商定金、NRE、库存承诺、质保和债务工具 |
已披露的融资结构是事实,但资本充足性仍只露出一部分,因为公开来源没有提供现金、烧钱速度、现金跑道或供应商承诺细节。
[CI002, CI003, CI004, CI005, CI039, CI040]主要财务负担落在系统和路线图层面:芯片和封装、服务器与机架集成、网络以及客户验证,都需要资本或伙伴执行, 之后才会出现经常性运营证明。
这是有证据支撑的定性矩阵,不是已披露的现金流报表。它突出源材料里能看到的资本需求和执行依赖。
[CI011, CI014, CI020, CI022, CI041, CI042]4.4 财务结论与披露缺口
合适的财务结论是谨慎但不否定。公开证据支持一个严肃资本基础、一套连贯收入机制,以及客户可能购买产品的可信工作负载层理由。它不支持用传统方式承销当前收入质量、毛利率、销售效率或现金跑道。d-Matrix 披露了融资、估值、商业化里程碑、伙伴架构和基准式效率主张,但没有公开披露收入、ARR、毛利率、客户数、重复订单、标价、已实现 ASP、软件附加率、现金余额、烧钱速度、现金跑道、债务或供应商承诺。在这门生意里,披露缺口比纯软件创业公司更关键,因为业务依赖硬件制造、封装、网络、服务器集成和部署时点。因此,投资案例仍面向未来且依赖融资。如果管理层能证明已预订收入、重复部署、按硬件和软件层拆分的利润率,以及覆盖 3DIMC 与部署路线图的现金跑道,这套融资栈就是一个真正发射台。缺少这些披露时,融资历史主要证明公司能融到钱,而不是已经跨过可重复、自我维持经济性的门槛。[CI036, CI037, CI038, CI039, CI040, CI041]
| 缺失指标 | 公开证据显示什么 | 重要性 | 精确尽调路径 |
|---|---|---|---|
| 收入 / ARR | 官方、投资方或独立来源均未找到公开数字 | 无法用当前商业规模或估值倍数锚定承销 | 索取按产品线、地区和客户队列拆分的月度和年度收入 |
| 毛利率 / 贡献利润率 | 未找到公开数字 | 无法检验硬件、网络和软件组合能否支撑持久盈利 | 索取 Corsair、JetStream、Aviator 和集成系统的毛利率桥 |
| 实际 ASP / 折扣 | 未找到公开 Corsair 或 JetStream 价格表 | 无法把性能说法转成交易经济性或折扣策略 | 索取价格手册、折扣区间,以及试点与生产 ASP |
| 软件附加 / 经常性收入 | Aviator 被作为平台的一部分营销,但授权模式未披露 | 经常性收入占比可能显著改变收入质量和倍数 | 索取带付费软件交易占比、年度合同价值和续约条款 |
| 客户数 / 重复订单 | 伙伴和基准测试可见,但广泛生产客户证据不可见 | 无法判断需求深度、可重复性或集中度风险 | 索取具名生产客户、重复订单、已部署卡或节点,以及积压订单 |
| 现金 / 烧钱速度 / 现金跑道 / 承诺 | 未找到公开现金余额、烧钱速度、现金跑道、债务或供应商承诺数据 | 无法从公开记录承销融资依赖和破产时点 | 索取现金瀑布、月度烧钱速度、采购承诺和下一轮触发条件 |
本表是主要承销阻塞清单。这些缺失指标对私营公司很正常,但在依赖制造、网络和部署的硬件业务里尤其关键。
[CI036, CI037, CI038, CI039, CI040, CI041]4.5 图表
05产品与技术
5.1 产品表面与工作负载匹配
d-Matrix 已不再只是兜售抽象芯片架构。当前面向客户的产品表面是一套完整推理栈:Corsair 加速卡、JetStream 网络卡、Aviator 软件,以及作为机柜级参考蓝图的 SquadRack。放到客户工作流里,公司瞄准的是交互式聊天、编码、智能体、推理和多模态工作负载,这些场景低延迟比蛮力基准吞吐更重要。一家创业硬件供应商把产品页面写到这种具体程度并不常见:它公布了单卡、双卡、服务器和机柜配置,并给出具体内存带宽和容量数字。这种具体性是真实的产品技术优势,因为它让买方和尽调团队看到可测试对象,而不是模糊的平台承诺。成熟度上的保留是,大多数公开证明仍停留在官方简报、伙伴公告和参考架构页面,而不是具名生产客户部署。架构信号很强,客户证明信号仍相对薄。公司近期文章还进一步锐化工作负载命题,认为更长上下文窗口、推理链和智能体循环会把推理变成上下文和延迟问题,而不只是原始吞吐竞赛。[CE001, CE002, CE003, CE004, CE005, CE036]
| 模块 / 产品线 | 主要用户 | 状态 / 成熟度 | 差异化 | 尽调缺口 |
|---|---|---|---|---|
| Corsair 单卡 / 双卡加速器 | 平台架构师和企业 AI 基础设施团队 | 当前旗舰产品,公开了卡、服务器和机架规格 | DIMC 加片上 SRAM,做成 PCIe 卡,面向低延迟解码和交互式推理 | 具名生产客户和现场可靠性数据未公开 |
| JetStream 透明 NIC | 跨节点和机架扩展 Corsair 的运营团队 | 当前网络产品,已宣布样品并发布详细规格 | 标准 Ethernet 上跑设备发起的加速器到加速器传输,而不是走主机中介的 RDMA 流程 | 广泛生产证明和兼容性矩阵未公开 |
| Aviator 软件栈 | ML 工程师、基础设施工程师和平台团队 | 当前软件层,文档覆盖构建和执行流程 | 共同设计的栈覆盖模型工厂、压缩、编译器、运行时、Kubernetes 插件和可观测工具 | 支持模型矩阵、SLA 姿态和客户支持范围未公开 |
| SquadRack 参考架构 | 云、主权云和企业数据中心运营商 | 参考蓝图,不是完全封闭的设备 | 采用 Corsair、JetStream、标准 PCIe 服务器和标准 Ethernet 网络结构的风冷机架设计 | 参考架构不等于广泛 GA 部署 |
| Pavehawk 测试硅和 Raptor 路线图 | 需要更大内存占用推理的未来买方 | 路线图阶段;公开验证偏架构,不是商业 | 3DIMC 路线图加入 3D DRAM,把低延迟内存中心架构从 Corsair 往外延展 | 独立基准、投产时间、封装良率和客户承诺未公开 |
该矩阵把已产品化内容同仍在路线图中的内容拆开。状态用语反映截至 2026-05-26 的公开语料,而不是管理层愿望。
[CE001, CE002, CE005, CE008, CE013, CE023]从目标工作负载出发,分层展示软件、网络、计算-内存硬件,以及未来内存扩展路线图。
层级边界采用分析口径,不是公司官方划分。该图旨在展示产品、软件、网络和路线图如何拼成一个平台。
[CE001, CE008, CE010, CE013, CE023, CE026]5.2 DIMC、JetStream 与 Aviator 架构
核心技术命题是,现代推理受内存搬运主导,尤其在 decode 阶段,因此 d-Matrix 针对以内存为中心的执行优化,而不是试图长得像另一块 GPU。Corsair 将 DIMC 计算与大规模片上 SRAM 和容量内存配对,再通过 chiplet、DMX Link 和 DMX Bridge 横向扩展。Aviator 与硅片同样重要,因为它负责模型转换、压缩、编译、运行时执行、集群编排和开发者工具。近期技术文章把这一框架从芯片本身向外扩展,认为上下文压力、KV-cache 降低,以及跨软件、网络和基础设施的优化,是 DIMC 的必要补充,而不是副业。JetStream 把同一套低延迟哲学延伸到多节点通信:分离数据平面和控制平面,并支持设备发起的加速器到加速器传输,底层仍走标准 Ethernet。技术故事最可信的地方在于,硬件、网络和软件层描述得足够具体,可以拼出一套连贯运营模型。主要尽调问题不是有没有架构,而是该架构在 d-Matrix 和少数技术伙伴之外被多广泛地验证过。[CE006, CE007, CE008, CE009, CE010, CE011]
| 用户任务 | 当前工作流痛点 | d-Matrix 方案 | 可衡量收益 | 限制 |
|---|---|---|---|---|
| 交互式聊天或编码助手 | 小批量场景下,纯 GPU 解码会对延迟和排队变得敏感 | 在 Corsair 上跑低延迟解码,同时把部署保留在标准 PCIe 服务器里 | 公司称,单台服务器中 Llama3 8B 最高可达 60,000 tokens/s、1 ms/token | 该性能数字由公司报告,且绑定特定工作负载 |
| 70B 级本地推理服务 | 大模型超出单卡能力,带来机架级内存和网络压力 | 使用八服务器 Corsair 机架,以及基于 JetStream 的节点间扩展 | 公司称 Llama3 70B 达 30,000 tokens/s、2 ms/token,并以 100B 参数规模为目标 | 公开证明最强的是参考架构和伙伴部署,不是具名客户大规模上线 |
| 搭配大型验证模型的异构流水线 | 纯 GPU 系统必须在快速解码和高质量大模型验证之间折中 | 用 Corsair 跑受内存约束的推测草稿阶段,把更大的验证器留在 GPU 上 | Gimlet 报告,在一个推测解码配置中,同等能耗下请求提速 2-5x | 独立验证范围较窄,并未覆盖每个推理阶段 |
| 企业现有数据中心改造 | 新 AI 系统往往需要专有 fabric 或重度基础设施改造 | 在现有数据中心空间里部署 PCIe 卡加 Ethernet 交换 | SquadRack 采用风冷,围绕标准服务器和 ToR Ethernet 交换机构建 | 部署质量仍取决于第三方服务器和交换机集成 |
| 多节点复杂度之前的单节点扩展 | 分布式推理会增加协调开销和运营复杂度 | 用 GigaIO SuperNODE 在单节点内容纳数十张 Corsair 卡 | 伙伴称多节点复杂度更低,并引用 Llama3 70B 达 30,000 tokens/s、2 ms/token | 单节点伙伴说法来自伙伴报告,而非广泛基准测试 |
收益混合了官方说法和一个独立从业者基准。本表讨论工作流适配度,而不是全面基准对决。
[CE005, CE012, CE015, CE016, CE017, CE018]| 层 / 流程 | 作用 | 关键依赖 | 关键风险 |
|---|---|---|---|
| 带片上 SRAM 的 DIMC 核 | 把内存贴近计算,用于低延迟 token 生成和交互式解码 | 成功把推理内核和数据移动映射到 DIMC 内存层级 | 架构有差异化,但广泛第三方工作负载验证仍有限 |
| 容量内存层 | 在不放弃内存中心设计的前提下,扩大模型和上下文占用 | Aviator 调度和运行时策略,决定哪些数据放快速内存、哪些放容量内存 | 性能取决于工作负载适配度,以及仍需在层级之间移动多少数据 |
| Chiplet 加 DMX Link 和 DMX Bridge | 从单个 chiplet 池扩展到双卡和机架级配置 | 跨 OEM 系统的封装、互连良率和板卡认证 | Chiplet 和桥接复杂度可能成为封装、信号完整性或认证瓶颈 |
| JetStream 透明 NIC | 提供低延迟节点间加速器通信,绕开主机中介 RDMA 开销 | 标准 Ethernet 光模块、交换机、PCIe Gen5 主机和 JetStream 固件 | 生产级网络证明晚于 Corsair 计算证明 |
| Aviator 软件栈 | 转换模型、压缩图、编译二进制文件、编排分布式执行并暴露工具 | 模型支持、编译器成熟度、运行时稳定性和开发者支持 | 软件质量是核心,因为硬件价值只有通过这套栈才能兑现 |
| 机架和服务器生态层 | 通过 Supermicro、GigaIO、Liqid、Arista 和 Broadcom,把卡和 NIC 变成可部署系统 | 伙伴认证、支持交接和标准服务器供应 | 开放标准降低锁定效应,但增加集成依赖 |
| Pavehawk 和 Raptor 上的 3DIMC | 用堆叠 DRAM 把架构推向更大内存推理 | Alchip 封装和 ASIC 工作、晶圆厂产能,以及测试硅成功转为产品硅 | 路线图价值很高,但商业时点和外部验证仍未证明 |
本表聚焦把 Corsair 变成可用平台所需的运营栈。依赖被明确列出,因为 d-Matrix 今天卖的不是封闭设备。
[CE006, CE007, CE008, CE010, CE011, CE020]模型如何从开发者输入进入异构 d-Matrix 部署,再提供低延迟推理服务。
该流程综合了已发布的 Aviator 构建与执行模型,以及 d-Matrix 自己的异构预填充 / 解码叙事;它对应本次审阅语料,不是官方产品图。
[CE008, CE010, CE017, CE018, CE045, CE046]5.3 部署、机柜架构与供应依赖
d-Matrix 试图借助标准 PCIe 服务器、标准 Ethernet 网络和生态伙伴来降低部署难度,而不是要求客户重建专有机柜。SquadRack 和 GigaIO SuperNODE 展示了预期服务器和机柜拓扑:尽可能在节点内纵向扩展,然后在模型尺寸或请求量增加时经 Ethernet 横向扩展。这应当降低企业和主权云环境的改造摩擦,让它们无需采用超大规模云式液冷也能获得 AI 推理。代价是依赖。d-Matrix 需要 Supermicro、Arista、Broadcom、GigaIO 和 Liqid 把卡变成可部署系统;路线图还依赖 Alchip、晶圆产能和先进封装,从 Corsair 推进到基于 3DIMC 的产品。近期公司和伙伴文章把这种相互依赖说得很清楚:d-Matrix 现在把开放标准框定为战略选择,其 GigaIO 公告把单节点纵向扩展纳入官方部署故事,Alchip 也独立确认了 3DIMC 背后的共同封装角色。更早的 Nighthawk 和 Jayhawk 披露显示,这种依赖模式并不新;封装和互连一直是公司命题的核心。因此,产品可以差异化,同时仍对执行敏感:架构对开放标准友好,但供应链并不自给自足。[CE013, CE014, CE015, CE021, CE022, CE023]
| 日期 / 阶段 | 里程碑 | 状态 | 含义 | 来源 |
|---|---|---|---|---|
| 2021 平台时代 | Nighthawk chiplet 平台 | 历史前身 | 说明 chiplet 和横向扩展在 Corsair 推出前已是设计命题的一部分 | EE News Europe / HPCwire |
| January 2023 | Jayhawk chiplet 平台发布 | 历史公开里程碑 | 表明商业化 Corsair 发布前,公司已有基于 TSMC 6nm 的封装和互连路线图 | EE News Europe / HPCwire |
| November 2024 | Corsair 发布,目标在 Q2 2025 广泛供货 | 当代产品里程碑 | 标志公司从平台命题转向具备明确卡规格、可以出货的计算产品 | d-Matrix 发布 / 白皮书 |
| May 2025 | GigaIO SuperNODE 纵向扩展合作 | 合作伙伴部署里程碑 | 展示了更广泛机架级部署前,单节点承载数十张 Corsair 卡的纵向扩展路径 | GigaIO / Data Center Dynamics |
| 2025 年 9 月至 10 月 | JetStream 发布与 SquadRack 蓝图 | 当代产品扩展 | 让 d-Matrix 从只讲加速卡,变成计算 + 网络 + 机架的叙事 | d-Matrix JetStream / SquadRack 产品 |
| November 2025 | 3DIMC、Pavehawk 验证与 Raptor 路线图 | 路线图阶段 | 把公司命题延伸到更大内存产品,但在本语料中仍处于商业化前 | d-Matrix Alchip 公告 / Going vertical |
时间线混合了历史架构里程碑、当代产品里程碑和未来路线图步骤。尽调关键要区分当前商业产品与未来内存扩展路线图。
[CE021, CE022, CE023, CE024, CE027, CE041]可部署的 d-Matrix 平台依赖硅片、封装、软件、网络和生态合作伙伴同步推进。
依赖图采用分析口径。它突出 d-Matrix 为什么既能拥有差异化架构,也仍会在合作伙伴和封装上承受明显执行敏感性。
[CE026, CE027, CE028, CE029, CE030, CE034]5.4 路线图、差异化与信任缺口
差异化逻辑已经足够清楚:d-Matrix 用以内存为中心的架构、开放标准机柜设计和未来 SRAM 加 3D-DRAM 混合路线图,攻击低延迟 decode 和推理时内存压力。这与现有 GPU 方向形成对比,后者强调 HBM 容量、专有 NVLink 域和越来越多的液冷机柜系统。公司近期围绕上下文增长、智能体延迟、开放标准和企业级推理的表述与该命题方向一致,但仍更像战略论证,而不是独立证明。弱点不是概念差异化,而是承销信心。本章语料显示,Gimlet 提供了一些独立实践者验证,也显示市场对专用推理硬件有更广泛兴趣,但没有展示广泛第三方基准覆盖、公开可靠性指标、具名生产客户或已发布合规凭证。3DIMC、Pavehawk 和 Raptor 故事尤其重要,因为它要把当前架构延展到更大内存部署;它也是路线图中外部验证最少的一部分。投资人今天可以承销一套具体架构,但仍需要尽调规模、信任和路线图执行。[CE031, CE032, CE033, CE034, CE035, CE037]
| 控制 / 信号 | 状态 | 范围 | 缺口 |
|---|---|---|---|
| JetStream 安全启动 | 公开披露 | 网卡级启动完整性,保护网络卡 | 尚未公开 Corsair、Aviator 与 JetStream 作为一个系统的更完整平台安全架构 |
| Kubernetes 插件、指标导出器、调试器和性能分析器 | 公开披露 | 支持部署、监控和开发者诊断的运营工具 | 有工具不等于公开了 uptime、MTBF 或支持性能数据 |
| 开放标准部署模式 | 公开披露 | PCIe 卡、以太网组网和标准机柜顶部交换机降低定制基础设施风险 | 开放标准不能消除集成、固件或合作伙伴认证风险 |
| Gimlet 的独立实践者基准 | 已公开披露但范围窄 | 异构 Corsair + GPU 配置上的一个推测解码工作流 | 验证面很窄,不能等同于广泛生产基准 |
| 公开合规资质 | 所审阅语料未披露 | 面向受监管、主权云和安全敏感买家的企业信任状态 | 在审阅的官方渠道中未发现公开的 SOC 2、ISO 27001、FedRAMP 或信任中心引用 |
| 公开机群可靠性指标 | 所审阅语料未披露 | 现场可靠性、服务耐久性和运营方信心 | 章节语料中未见公开的 MTBF、RMA、uptime 或故障率披露 |
本表区分已披露控制与缺失的信任材料。缺失仅指审阅语料,不应被假定不存在,而应通过尽调请求补齐。
[CE009, CE035, CE037, CE038, CE039, CE040]从当前产品栈到未来 3DIMC 路线图的成熟度视图,重点看外部验证和信任缺口。
这是作者基于公开证据深度作出的判断,不是公司计分卡。矩阵刻意对独立验证和信任披露从严。
[CE013, CE016, CE023, CE035, CE038, CE040]5.5 图表
06客户情况
6.1 买方细分与采购中心
即便公开材料没有透露多少账户已经转化,它们仍相当一致地描绘了 d-Matrix 想卖给谁。Series C 新闻稿把顶层目标客户锚定为超大规模、企业和主权客户;SquadRack 又把运营买方缩窄到云供应商、主权云,以及试图服务智能体 AI、推理和视频工作负载的企业。因此,反复出现的决策者画像不是应用团队,而是基础设施和平台领导者:数据中心运营商、AI 平台架构师,以及希望在既有设施中获得更低延迟推理的云建设者。围绕标准 PCIe 服务器、Ethernet 和风冷的销售叙事也强化了这一定位。d-Matrix 并不是在推销一个只适合超大规模云厂商、需要绿地基础设施的定制机柜;它推销的是一条改造路径,面向想增加推理容量、但不想承接专有网络或液冷强制要求的运营商。当前最强匹配看起来是企业和 neocloud 型运营商,加上重视基础设施可选性的主权云项目;公开材料仍没有披露超大规模云的生产引用。[CU001, CU002, CU003, CU004, CU005, CU006]
| 客群 | 买方 / 预算负责人 | 主要用户 | 公开证据 | 战略价值 | 关键缺口 |
|---|---|---|---|---|---|
| 超大规模云厂商和 AI 工厂 | 平台与基础设施负责人 | 服务前沿模型或高吞吐推理的 AI 基础设施团队 | Series C 公告把超大规模客户列为目标,但没有披露具体账户名称 | 若拿下,将验证超大规模生产采用和硅片可信度 | 尚无公开点名的超大规模生产客户 |
| 企业数据中心 | CIO / CTO / AI 平台预算 | 改造既有设施的基础设施与平台团队 | GigaIO、Supermicro 和 Liqid 的话术都围绕企业部署和 TCO 来包装 Corsair | 近期商业切入口,因为改造经济性重要 | 多数证据来自合作伙伴,而不是终端客户案例 |
| 主权云和国家项目 | 国家支持的云运营商和基础设施采购团队 | 服务受监管或国家级工作负载的运营方 | Series C 和 SquadRack 材料明确提到主权客户或主权云 | 可能是重视基础设施控制权的高价值账户 | 没有公开点名或量化任何主权客户 |
| 专业 AI 云 / neocloud | 云运营商管理层 | 模型提供商和 AI 原生公司 | Gimlet 是最清晰的具名运营方,使用包含 Corsair 的异构基础设施 | 不用先拿超大规模云客户标识,也能切入高使用量工作负载 | 公开可用性在 2H 2026 仍是选择性客户 |
| OEM / 渠道生态 | 系统集成商和 OEM 产品团队 | 服务器、机架和网络集成团队 | Supermicro、GigaIO、Arista、Broadcom 和 Liqid 出现在公开商业化材料中 | 是扩大触达、提升部署速度和覆盖支持的关键倍增器 | 渠道依赖会遮蔽真实终端客户数量和收入结构 |
本表把终端买方客群与帮助 d-Matrix 触达客户的 OEM / 渠道层拆开。公开证据在上市路径上最强,对各客群披露账户数的证明不足。
[CU001, CU002, CU004, CU005, CU007, CU025]d-Matrix 试图让买家从改造评估走向纵向扩展、机架级部署和异构云扩展,同时不强迫他们重建全新数据中心。
这些阶段是从公开商业化材料综合出的分析性旅程步骤,不是公司披露了规模的官方漏斗。
[CU003, CU019, CU020, CU021, CU022, CU041]6.2 具名证明与部署成熟度
把具名证明点与宽泛客户数叙事分开后,客户证据图景明显改善。公开层面,d-Matrix 可以指向一条真实商业化阶梯:2024 年末 Corsair 向早期访问客户采样,2025 年 JetStream 样品,2025 年 5 月 GigaIO 的集成 SuperNODE 报价,SquadRack 计划 2026 年 Q1 通过 Supermicro 渠道可用,以及 2026 年 3 月与 Gimlet Labs 合作做异构云部署。这些是有意义的里程碑,因为它们显示公司从硬件发布走向伙伴集成系统,再走向具名外部运营商。但同一语料也界定了证明边界。生态页面明确说展示的 logo 不是实际客户,公开的 GigaIO、Supermicro 和 Liqid 声明仍是伙伴背书,而不是终端客户案例。Gimlet 是这组材料中最清楚的具名外部运营商,但即便如此,公开记录也说,面向选定客户可用的时间计划在 2026 年下半年,而不是已经大规模生产。换句话说,部署成熟度是真实的,但公开证据更偏渠道就绪和运营商试点,而不是披露的生产账户深度。[CU008, CU009, CU010, CU011, CU012, CU013]
| 里程碑 | 公开状态 | 日期 | 来源 | 置信度 | 含义 | 缺失的分母 |
|---|---|---|---|---|---|---|
| Corsair 早期访问样品 | 向早期访问客户提供样品 | 2024-11-19 | 来源:d-Matrix / eeNews Europe | 高 | 说明产品已不再只停留在实验室定位 | 没有早期访问账户数或转化率 |
| GigaIO 纵向扩展合作 | 发布集成式企业纵向扩展系统 | 2025-05-01 | 来源:d-Matrix / GigaIO / AIwire | 中 | 证明单节点企业部署有一条路径 | 没有装机基数或已预订客户数字 |
| JetStream 网卡 | 样品可用;预计年末全面投产 | 2025-09-08 | d-Matrix | 中 | 把平台从加速卡扩展到多节点网络 | 没有出货量或附加率披露 |
| SquadRack 商业化 | 配置可在 Q1 2026 通过 Supermicro 购买 | 2025-10-14 | 来源:PRNewswire / d-Matrix | 高 | 把证据从参考架构推向渠道可交易性 | 未披露积压订单或客户数 |
| Series C 客户扩张 | 公司称融资支持多个大规模部署和快速客户增长 | 2025-11-12 | d-Matrix / 独立报道 | 中 | 暗示销售漏斗已越过试点阶段 | 没有账户名称、部署数量或收入拆分 |
| Gimlet 云服务 | 计划在 2H 2026 面向选定客户推出方案 | 2026-03-12 | 来源:d-Matrix / Data Center Knowledge | 中 | 这是语料中最新的具名外部运营方证据 | 选定客户计划不等于广泛 GA 或续约数据 |
轨迹里程碑是运营商业化标志,不是客户数量指标。公开材料显示阶段推进,但没有披露每一步背后的分母。
[CU008, CU009, CU010, CU011, CU014, CU015]| 证据点 | 客群 | 部署 / 使用场景 | 生产还是试点 | 结果 / 信号 | 局限 |
|---|---|---|---|---|---|
| Gimlet Labs | 专业 AI 云 / 运营方 | Corsair 与 GPU 并行部署,用于智能体推理和推测解码 | 广泛供货前;计划在 2H 2026 面向选定客户 | 具名外部运营方,加上技术基准和计划中的云服务 | 仍不足以证明已披露的广泛生产客户数或续约 |
| GigaIO | 企业纵向扩展基础设施合作伙伴 | Corsair 集成进 SuperNODE,支持大型单节点推理部署 | 集成方案,并有早期访问意向路径 | 展示可信的企业部署载体和纵向扩展路线 | 这是合作伙伴证据,不是具名终端客户的生产部署证明 |
| Supermicro | OEM / 渠道伙伴 | 用于机架级推理的 SquadRack 和 X14 AI 服务器平台 | Q1 2026 有可购买配置路径 | 表明渠道可交易,且机架集成准备度提高 | 公开声明没有指出实际终端用户或结果 |
| Liqid | 可组合基础设施合作伙伴 | 集成 Corsair,用于灵活推理部署 | 仅为合作伙伴背书 | 支撑 Corsair 可插入其他企业平台的叙事 | 未披露具名客户、部署规模或使用指标 |
| 未披露早期访问客户 | 评估群体 | 2024 年末的 Corsair 样品 | 试点 / 评估 | 显示更广泛商业化前已有部分 GA 前外部测试 | 账户数、转化率和客群组合未公开 |
公开材料中最强的具名证据,是一个云运营方和多家渠道 / OEM 合作伙伴的组合。客户标识和背书能提高对部署成熟度的信心,但无法回答核心问题:已有多少终端客户进入生产。
[CU008, CU012, CU014, CU015, CU016, CU019]公开证据显示,d-Matrix 正从早期访问抽样推进到渠道可用,再到面向精选客户的云端推出,但还没有披露装机基数指标。
数值只是定性的阶段排序,不代表客户数量或转化率。
[CU008, CU009, CU010, CU015, CU041, CU044]验证基础在已命名合作伙伴和运营方验证上最强,在终端客户具体性、留存可见性和集中度披露上最弱。
矩阵标签是基于公开语料得出的定性判断,用来比较证据质量,而不是给账户经济性打分。
[CU015, CU016, CU033, CU040, CU045]6.3 扩张路径、渠道依赖与采购摩擦
扩张命题是连贯的,也可能是本章最大的正面因素。d-Matrix 试图先在低延迟敏感推理可插入既有数据中心的地方落地,然后从卡评估扩展到 单节点纵向扩展、机柜级部署和更广泛的异构云部署。GigaIO 代表单节点纵向扩展步骤,SquadRack 加 Supermicro 代表机柜级和渠道销售步骤,Gimlet 代表云服务商路径,可在 GPU 旁边做工作负载特定加速。挑战在于,每一步都依赖伙伴执行。公开材料反复显示,d-Matrix 需要 OEM、系统集成商、交换伙伴和编排层,才能把技术差异化转化为可重复客户采用。独立来源把摩擦说得更直白:企业采购仍围绕 Nvidia 可信的软件和系统生态优化,异构部署只有在抽象层隐藏硬件复杂度之后才会变得容易。因此,d-Matrix 的客户扩张路径可信,但转化风险仍卡在 OEM 认证、开发者工作流兼容性,以及买方是否足够信任平台、愿意从试点或评估走向重复生产支出之间。[CU021, CU022, CU023, CU024, CU025, CU026]
| 扩张驱动 | 如何扩张 | 集中度或渠道风险 | 影响 | 尽调路径 |
|---|---|---|---|---|
| 既有数据中心改造 | 靠 PCIe 卡、标准以太网和风冷部署切入 | 扩张取决于买家是否信任非 Nvidia 加速能嵌入既有工作流 | 若兼容性主张成立,可缩短首次部署时间 | 在两个真实账户上验证部署工作量和运营方培训负担 |
| 通过 GigaIO 做单节点纵向扩展 | 从评估卡数扩展到单节点数十张 Corsair | 路径由合作伙伴中介,不透露终端客户集中度 | 有助于从概念验证过渡到有意义的工作负载规模 | 要求按合作伙伴和阶段拆分管线,并提供附加率 |
| 通过 Supermicro / SquadRack 走向机架级 | 从节点扩展到机架,再到多机架集群 | OEM 可用不等于终端销售或支持质量得到证明 | 可能打开更大的企业、主权和云预算 | 要求提供通过 Supermicro 获得的预订额、部署和投产时间 |
| 通过 Gimlet 切入云运营方 | 切入异构推理云工作负载和 AI 原生运营方 | 选择性客户发布会让公司依赖少数早期运营方胜利 | 可能快速形成高使用量参考账户 | 要求提供 Gimlet 当前已承诺容量、目标客户和收入模式 |
| 直接赢得超大规模或主权客户 | 会显著拉大平均合同规模 | 公开来源没有点名任何此类账户,集中度上行和下行都不透明 | 这是最大上行路径,也是今天最难靠公开信息承保的路径 | 要求提供具名账户或采购阶段证据,并说明时间和范围 |
本表把集中度风险同时视为上行和下行,因为未披露的大客户既可能加速增长,也可能掩盖依赖。
[CU020, CU021, CU022, CU023, CU024, CU032]| 摩擦点 | 成因 | 当前缓解 | 公开证据 | 剩余请求 |
|---|---|---|---|---|
| 根深蒂固的 Nvidia 生态 | 客户已熟悉既有厂商的软件、工具和部署模式 | d-Matrix 主打开放标准、更低延迟和更低功耗 / TCO | Forbes、GFM Review 和 IntuitionLabs 都强调现有生态优势 | 要求提供相对 Nvidia 方案的输赢分析 |
| 异构部署复杂度 | 只有编排能藏住硬件复杂度,多种加速器才有帮助 | Gimlet 抽象工作负载放置,d-Matrix 主打 Aviator 加 JetStream | Data Center Knowledge 和 Gimlet 都把抽象层描述为核心 | 要求运营方证明部署工作量和代码改动 |
| OEM 与渠道依赖 | 客户买的是系统、机架和支持包,不是裸硅片 | Supermicro、GigaIO、Liqid、Arista 和 Broadcom 扩大商业触达面 | 官方商业化材料反复依托这些伙伴 | 要求支持分工、认证矩阵和合作伙伴收入集中度 |
| 证据深度与证据质量 | 合作伙伴引述和架构里程碑比生产群组数据更容易发布 | d-Matrix 现在有一个具名运营方 Gimlet,以及可购买的 SquadRack 配置 | 语料显示实际进展,但仍缺少留存和装机基数披露 | 在承保持久需求前,要求生产参考访谈和群组指标 |
本表刻意保持定性。公开证据更能说明什么会让采购变容易,而不是这些缓解措施多快转成持久生产收入。
[CU024, CU034, CU035, CU036, CU037, CU038]6.4 耐久性、集中度与披露缺口
公开记录中最弱的是耐久性证据。官方和独立来源称公司客户快速增长,并在准备多个大规模部署,但所有已审阅材料都没有量化活跃客户数、重复订单频率、合同期限、总留存率、净留存率或流失。集中度风险也是如此:没有公开披露头部客户敞口、伙伴与直销渠道组合,也没有说明收入中有多少来自某一个超大规模、企业或主权账户。因此,本章可以承销商业化正在推进,但不能承销它已经广泛、粘性强且多元化。2026 年最新证据偏架构和伙伴主导,而不是队列主导。投资人应把这个核心承销缺口放在最前面:d-Matrix 有足够具名证明显示采用并非纯概念,但披露不足以证明这种采用在生产规模上具备耐久性、可重复性和多元化。[CU028, CU029, CU030, CU031, CU032, CU033]
| 指标 | 数值 | 客群 | 置信度 | 为什么重要 | 尽调请求 |
|---|---|---|---|---|---|
| 客户数量 | 全部客群 | 低 | 没有活跃账户数,部署话术就无法转成市场份额或采用深度 | 要求按超大规模、企业、主权和云运营方客群提供当前活跃生产客户 | |
| 净收入留存 | 全部经常性账户 | 低 | NRR 是检验基础设施先落地再扩张主张的最干净方式 | 要求按直销和渠道来源群组提供过去 12 个月 NRR | |
| 总留存 / 流失 | 全部客户 | 低 | 总留存能看出部署是否撑过概念验证 | 要求提供客户流失、总留存和未续约原因 | |
| 合同期限 / 重复订单节奏 | 企业和主权账户 | 低 | 硬件和基础设施业务即便缺少重复订单,管线看起来也可能健康 | 要求平均期限、续约节奏和积压订单转化数据 | |
| 可背书满意度 / 用户倡导 | 公开材料只有合作伙伴引述和一个具名运营方基准 | 渠道和运营方证据 | 中 | 产品会改变基础设施设计选择,因此参考访谈和用户满意度很关键 | 要求至少 3 个生产客户参考访谈,覆盖部署年限、工作负载和实际结果 |
null 占位是有意保留的。本章审阅的公开记录没有披露留存、续约或客户数量指标,无法支撑定量承保。
[CU028, CU029, CU030, CU031, CU040, CU045]6.5 图表
07风险
7.1 路线图与系统集成风险
排名最高的风险在于,d-Matrix 卖的不是一颗孤立芯片,而是在把一套同步栈商业化:Corsair 卡、JetStream 网络、Aviator 软件、合作伙伴认证机架,以及下一代 3DIMC 路线图都要一起跑通。这套架构可以形成差异化,但仍然脆弱。JetStream 公开口径仍是早期商业化,而不是成熟的机群部署;Pavehawk 到 Raptor 的路径也明确停留在实验室验证口径,不是量产良率或广泛客户证明。公司在头五年还做过重大架构转向,说明团队能适应变化,也意味着买方和投资人要为一条仍在移动的路线图买单。缓释因素确实存在,体现在具体产品形态和合作伙伴披露上;但缓释成熟度仍只算中等,因为公开记录仍高度依赖公司自述里程碑,而不是广泛、独立的生产证据。[CR001, CR003, CR004, CR005, CR006, CR007]
| 故障模式 | 发生概率 | 严重性 | 缓释成熟度 | 剩余暴露 | 未解缺口 |
|---|---|---|---|---|---|
| JetStream 和多节点部署仍比成熟在位厂商网络栈更早期 | 高 | 高 | 中等 | 高 | 公开证据仍停留在早期商业化,而不是长期生产验证 |
| Pavehawk 和 Raptor 仍是路线图资产,下一代内存扩展尚未经过现场验证 | 高 | 阻断 | 低 | 高 | 未找到公开的生产良率、GA 时间或外部基准包 |
| 一旦纳入网络、编排和支持开销,系统效率主张可能被压缩 | 中 | 高 | 低 | 中 | 审阅材料缺少独立生产用户的机架级 TCO 或集群运营数据 |
| 缺少公开 MTBF、正常运行时间或 RMA 指标,现场可靠性证据不足 | 中 | 高 | 低 | 高 | 除公司自述部署材料外,耐久性仍未解决 |
| 解耦式推理增加了更多集成面,固件、交换机或软件协同都可能出错 | 中 | 高 | 中等 | 中 | 安装时间、运营负担或回滚行为几乎没有公开证据 |
| 如果模型再次转向,快速演进的工作负载可能削弱当前架构假设的耐久性 | 中 | 中 | 中等 | 中 | 路线图必须跟上未来推理型、多模态和智能体化推理模式 |
运营风险主要来自成熟度和集成,而不是某个已披露缺陷。公开缓释措施存在,但多数仍停留在概念或伙伴主导阶段。
[CR001, CR003, CR004, CR005, CR006, CR008]梳理核心依赖:它们让 d-Matrix 的差异化架构既友好于开放标准,又对执行高度敏感。
节点按分析口径分组,展示投资判断依赖,而不是复刻某张官方产品图。
[CR002, CR004, CR005, CR009, CR013, CR014]7.2 供应、制造与生态依赖
第二大风险是封装、代工选择、网络和系统集成上的集中依赖。下一代内存路线图要靠外部先进封装落地,公开第三方报道也把 d-Matrix 的 chiplet 路径与 TSMC 节点选择和演进中的接口标准绑在一起。与此同时,当前上市路径要靠 Arista、Broadcom、Supermicro、GigaIO 等合作伙伴把卡变成可部署系统。合作伙伴重的模式降低了客户的专有锁定,但把执行风险转移到第三方认证、支持交接和 BOM 供应上。宏观供应环境并不帮忙:先进封装瓶颈和邻近的 GPU 短缺已经证明,AI 硬件供应链收紧可以来得很快。今天的缓释因素是概念上的开放性和合作伙伴广度,不是自给自足或高度冗余供应链的证据。[CR002, CR009, CR010, CR011, CR012, CR013]
| 依赖 | 交易对手 / 集群 | 角色 | 集中度 | 失败情景 | 严重性 | 缓释措施 | 剩余暴露 |
|---|---|---|---|---|---|---|---|
| 先进封装和 3DIMC 落地 | Alchip 以及未来代工和 OSAT 链条 | 支撑下一代路线图 | 高 | 封装、节点或认证工作滞后,导致 Raptor 延期,或出货但经济性偏弱 | 阻断 | 将下一代估值上行与清晰制造里程碑绑定 | 高 |
| 参考机架和 OEM 栈 | Arista、Broadcom、Supermicro | 把卡和 NIC 变成可部署系统 | 高 | 即便芯片可用,认证或支持缺口也会拖慢客户铺开 | 高 | 要求经过验证的 BOM 和支持交接模型 | 高 |
| 单节点纵向扩展伙伴 | GigaIO | 为更大配置提供纵向扩展部署路径 | 中 | 伙伴主导的部署证明未必转化为广泛、可重复的现场采用 | 中 | 尽调中把伙伴演示与可重复客户需求分开看 | 中 |
| 早期外部客户验证渠道 | Gimlet Labs | 提供云部署和早期客户可见度 | 中 | 参考合作未能转化为多元生产客户证明 | 高 | 跟踪从精选客户可用到具名生产胜利的转化 | 高 |
| 在位厂商软件和运营生态 | CUDA 时代买方工作流和采购惯例 | 塑造企业默认购买行为 | 高 | 切换成本盖过单点性能上行,客户继续默认选择在位栈 | 高 | 要求证明工作负载层面的胜利经得起全系统比较 | 高 |
本登记表混合了供应商、集成商和市场结构依赖,因为即便芯片本身表现达标,三者都可能卡住商业化。
[CR002, CR009, CR010, CR013, CR014, CR028]7.3 采用验证与部署摩擦
第三大风险是采用证据仍落后于技术叙事。公开证据最强的地方,是早期使用合作伙伴、参考架构和部分客户可用窗口;证据还不是一组广泛具名的生产客户。Gimlet 和 GigaIO 公告有帮助,因为它们显示了真实交易对手和部署意图,但读起来仍像早期商业化界面,不像成熟装机基础。这一点很关键:即便芯片层面的经济性更好,企业采用 AI 仍会被网络、供电和运营复杂度卡住。d-Matrix 主打风冷、标准服务器,这是实打实的销售优势;但公开记录仍没有说明,客户能否在不投入大量解决方案工程、也不经历漫长认证周期的情况下部署整套栈。因此,这既是客户证明风险,也是落地摩擦风险。[CR014, CR015, CR016, CR017, CR018, CR035]
7.4 合规、安全与监管风险
第四大风险是,合规、出口管制和安全证明比许多企业和主权买家今天对推理基础设施供应商的预期更薄。容易核验的公开法律材料,是网站隐私政策和使用条款;这些材料必要,但不能替代产品安全架构、审计姿态或专门的信任中心。这个缺口很重要,因为受监管买家越来越重视数据控制、可预期治理和本地部署保障,竞争平台也会明确营销这些能力。叠加其上的还有出口管制风险。BIS 已经收紧先进计算半导体和规避路径的管制,这意味着先进 AI 硬件供应链处在活跃监管制度之下。d-Matrix 未必会被这些规则直接挡住,但投资人仍需要搞清楚封装伙伴、制造地域和目标客户如何与不断演进的贸易管制相互作用。[CR023, CR024, CR025, CR026, CR027, CR041]
| 风险 | 管辖区 / 范围 | 当前公开信号 | 可能性 | 严重性 | 缓解成熟度 | 剩余敞口 | 尽调路径 |
|---|---|---|---|---|---|---|---|
| 先进芯片出口管制和反规避规则可能打乱供应链规划或客户地域布局 | 美国和受关注国家 | BIS 收紧了对先进计算半导体及相关规避路径的管制 | 中 | 高 | 低 | 高 | 获取制造地理分布图,以及覆盖封装伙伴和目标账户的出口管制法律备忘录 |
| 相较受监管买家的预期,公开产品安全和合规披露偏薄 | 全球企业和主权买家 | 审阅的公开法律页面未显示信任中心或产品保障包 | 中 | 高 | 低 | 高 | 索取安全架构、审计准备度和受监管客户参考案例 |
| 网站隐私和条款覆盖范围不能回答产品层面的数据治理或责任问题 | 客户合同和部署治理 | 公开法律页面描述网站数据处理和免责声明,而不是模型或集群控制 | 中 | 中 | 低 | 中 | 审阅客户合同、DPA 条款、赔偿条款和安全承诺 |
| 即便现有产品已在出货,贸易政策变化也可能影响下一代封装或节点迁移 | 代工、封装和未来路线图 | 路线图显示,先进封装和节点敏感度会随时间上升 | 中 | 高 | 低 | 高 | 梳理哪些未来项目最容易受到规则变化或许可审查影响 |
| 公开合规姿态可能不足以支撑主权或高度监管部署竞标 | 政府、国防相关、金融和医疗买方 | 竞争对手对外传递的安全或本地部署保障表述比 d-Matrix 当前更强 | 中 | 高 | 低 | 中 | 索取产品安全路线图,以及受阻或延后的受监管机会清单 |
本登记表聚焦公开法律、出口管制和信任信号。现有材料更多讲架构,少有可审计的合规证据,因此剩余暴露仍然偏高。
[CR023, CR024, CR025, CR026, CR027]7.5 竞争、融资与叫停条件
最后一组风险来自既有巨头竞争、生态锁定和融资纪律的叠加。NVIDIA、AMD,以及 Groq、Cerebras 这类专门化替代者,持续抬高内存容量、机架级集成、部署模型和软件预期的门槛。独立分析也认为,切换成本存在于整个系统,而不只是芯片层面;市场到 2026 年仍会保持集中。这意味着 d-Matrix 必须赢下特定工作负载,并且在完整系统对比下保住延迟 / 功耗或存量环境部署优势。融资风险又放大了挑战。2025 年这一轮融资买来了时间,但也在广泛公开收入证明出现前立下了估值锚点。管理层自己的历史里有过濒死的现金时刻;公司又在扩大办公足迹和员工数,而公开客户证明仍然有限。因此,投资逻辑应继续按里程碑推进,并把路线图时点、生产客户转化和供应链稳定性设为明确的逻辑失效触发点。[CR019, CR020, CR021, CR022, CR028, CR029]
| 角色 / 职能 | 依赖或缺口 | 发生概率 | 严重性 | 缓释措施 | 尽调路径 |
|---|---|---|---|---|---|
| 全球工程和运营布局 | 团队已横跨多个地区,必须协调芯片、软件和商业化项目 | 中 | 高 | 项目管理纪律和更强发布治理 | 索取组织架构图、发布节奏和跨站点决策权模型 |
| 商业化领导力和现场支持 | 客户成功和支持运营能规模化的公开证据有限 | 中 | 高 | 大规模铺开前先建设客户工程和售后能力 | 索取支持团队人数、升级模型和部署人员假设 |
| 产品定义纪律 | 市场演进中,公司多次调整架构 | 中 | 中 | 用里程碑约束路线图,并设置明确的客户设计伙伴反馈闭环 | 审查路线图治理、阶段门,以及终止或新增项目的标准 |
| 扩张期资本配置 | 招聘和布局扩张可能让烧钱速度快于已验证收入证据 | 中 | 高 | 将招聘和全球扩张与已签需求、支持负荷绑定 | 索取烧钱桥接表、招聘计划和费用控制触发点 |
执行风险不在于创始人可信度,而在于公司能否足够快地放大商业化和发布纪律,跟上技术雄心。
[CR007, CR019, CR020, CR021, CR022, CR044]| 风险 | 可监控触发器 | 阈值 / 事件 | 行动含义 |
|---|---|---|---|
| 路线图成熟度 | JetStream 或 Raptor 时间线 | 相对于客户认证或采购窗口出现重大延期 | 投资假设从路线图上行转为仅按当前可部署产品面估值 |
| 客户证明 | 具名生产部署 | 到 2026-2027 年,除伙伴或早期访问参考外仍没有多元生产客户群 | 下调商业化信心和估值支撑 |
| 供应与封装 | 先进封装或出口管制冲击 | 规则变化、产能未达或成本飙升,实质改变路线图时间或单位经济性 | 将下一代收入假设视为延后,并重新审视现金需求 |
| 竞争差异化 | 全系统每瓦延迟比较 | 在位厂商或专业替代方案在没有相同集成负担的情况下缩小差距 | 降低为架构溢价和护城河假设付费的意愿 |
| 安全和受监管采用 | 信任中心和合规证据 | 受监管机会仍是叙事一部分时,却没有产品安全保障包 | 约束 TAM 假设,并推迟主权或受监管市场上行 |
| 融资纪律 | 烧钱与证据 | 预订或生产证据仍薄时,员工数和全球扩张继续增加 | 将投资逻辑转向保全资本和下行情景测算 |
这些触发器设计成能通过后续尽调或未来更新观察到,而不是泛泛的叙事担忧。
[CR037, CR038, CR039, CR040, CR041, CR045]剩余风险视图显示,路线图和客户验证风险集中在影响最高的格子,合规和融资风险紧随其后。
可能性和影响是基于审阅语料综合出的定性排序,不是统计损失估计。
[CR036, CR037, CR038, CR039, CR040, CR041]展示路线图、供应、合规和采用风险如何传导到订单、烧钱、融资和估值。
该图是分析视角,不是公司官方图;它可视化了投资判断需要关注的二阶影响。
[CR016, CR021, CR022, CR036, CR037, CR038]7.6 附录
08估值
8.1 已验证估值、价格纪律与核心判断
d-Matrix 不是纸面公司。2025 年 Series C 公开披露融资 $275 million、估值 $2 billion,投资人名单真实存在;围绕 Corsair 加速器、JetStream 网络和 Aviator 软件,公司也有可见的产品叙事。这一轮发生在 2023 年 $110 million Series B 之后,也发生在公司多年围绕面向推理的数字存内计算做技术定位之后。这些足以让公司值得认真跟踪,但不足以让投资人盲目接受上一轮估值。 估值问题出在分母。已审阅的公开材料仍未披露收入、年经常性收入(ARR)、毛利率、自由现金流、客户数或股权结构条款。Series C 公告声称客户快速增长、合作伙伴扩张,但没有告诉外部投资人,这个故事中有多少是试点活动、生产部署,或经济性足够好的经常性需求。没有这些数字,常规的后期收入倍数无法靠公开证据自洽。 因此,正确结论必须对价格敏感。公司或许在推理栈里有战略相关性,但在 $2 billion 估值上,公开记录仍显得不完整。所以当前时点的判断是继续研究,而不是买入:技术叙事质量足以支撑继续工作,但公开证明仍太薄,不能把最近一轮估值当作干净的公允价值锚。[CV001, CV002, CV003, CV005, CV006, CV007]
| 维度 | 评估 | 证据锚点 | 决策含义 |
|---|---|---|---|
| 建议 | 继续研究 | 融资和产品证明真实存在,但缺少运营分母指标 | 只有尽调权利或价格改善时才继续推进 |
| 信心 | 中 | 融资和技术事实真实,但收入 / ARR / 利润率仍未披露 | 使用宽情景区间,避免伪精确 |
| 风险评级 | 高 | 商业化、基准归一化和下一轮价格发现仍未解决 | 私有证据补上缺口前,先假设下行不对称 |
| 估值立场 | $2B 估值偏高 | Series C 轮价格高于本章公开证据中点 | 把上一轮视为里程碑估值,而不是已验证的便宜价格 |
| 参考入场纪律 | 低于约 $1.5B,或带硬性下行保护 | 基准情景中心低于上一轮,公开证据无法支撑 $2B 的目标回报 | 要求价格改善、结构安排或异常强的数据室证据 |
| 可能退出路径 | 战略出售,或 IPO 前较晚期老股交易 | 战略价值比公开市场披露准备度更清晰 | 不要基于公开记录假设近期 IPO 流动性 |
建议明确对价格敏感。它评估的是 2026 年证据集和最新披露轮次,而不只是架构质量。
[CV001, CV017, CV049, CV050, CV051, CV052]| 论点 | 乐观解读 | 悲观解读 | 什么会改变判断 |
|---|---|---|---|
| 推理时点 | 投资人正确预判了推理会成为 AI 的经济主战场 | 主题可能正确,但这个价格和这家公司能捕获的份额仍可能过于乐观 | 展示今天已经拿下推理预算的具名生产客户 |
| 架构 | 近内存设计和模块化插入可能解决真实的延迟与功耗问题 | 主张仍过多来自公司自述,未必经得起归一化测试 | 发布第三方同模型基准,同时给出吞吐、延迟、功耗和成本 |
| 商业化 | GigaIO、Gimlet 和主权 / 企业表述显示可能存在真实市场拉力 | 公开证据可能仍以试点、伙伴和路线图证明为主,而不是已规模化的经常性需求 | 披露已签收入、客户数和试点转生产转化 |
| 可比背景 | 稀缺推理资产相对泛硬件或软件标的可能应享有溢价 | NVIDIA 和 AMD 已披露巨大规模,私营挑战者仍高度异质且不透明 | 证明 d-Matrix 属于稀缺性溢价资产,而不是投机型创业公司 |
| 价格纪律 | 如果乐观情景里程碑很快兑现,$2B 可能合理 | 目前估值已在缺少公开财务证明的情况下计入成功 | 更低入场价或强力私有数据室证据会显著改善判断 |
反向逻辑不是 d-Matrix 缺少技术价值,而是最新价格已经跑在公开材料真正能支撑的范围前面。
[CV013, CV014, CV015, CV016, CV017, CV030]判断链条从真实融资和技术可信度出发,经过缺失的分母指标,落到对价格敏感的“继续研究 / 偏高”建议。
这是对本章决策逻辑的压缩,不是因果经营模型。
[CV001, CV013, CV014, CV017, CV032, CV044]8.2 公开记录支持什么,也不支持什么
最强的公开证据支持把 d-Matrix 视为一家以推理为先的硬件公司:融资真实、架构连贯、生态关系在增长。公司、投资人和合作伙伴材料反复讲述同一套运营叙事:推理正在成为 AI 的经济瓶颈,d-Matrix 围绕内存搬运问题设计产品,Corsair 等产品相对 GPU 重的替代方案可以改善吞吐、延迟、功耗和总成本。GigaIO 和 Gimlet 也至少给出了一些外部证据,显示硬件正在被集成进真实系统,并在异构推理工作流中测试。 但估值最需要证据变强的地方,证据反而变弱。多数数字化经济性主张来自公司自述或合作伙伴放大。公开材料里最好的独立技术支持,是 Gimlet 的推测解码分析;它有用,但仍比不上与既有 GPU 集群在同模型、同批量、同延迟下的基准对比。更重要的是,已审阅来源没有披露经常性收入基础、毛利结构、定价曲线或客户集中度,投资人无法把技术潜力换算成可辩护的企业价值。 这种不对称很关键。今天的公开证据支持“d-Matrix 可能变得重要”。它不支持“当前私募估值已经便宜”。从估值角度,本章可以承保战略可选性和里程碑进展;但不能像隐藏单位经济性已经公开证实一样去承保。[CV004, CV005, CV006, CV009, CV010, CV011]
KPI 仪表盘突出少数公开硬锚,以及驱动建议的主要分母事实缺口。
看板混合了已披露事实和推断得出的估值结果;每个推断项都会在 detail 字段中标明。
[CV001, CV004, CV017, CV049, CV050, CV053]8.3 可比公司集合与反推承保逻辑
可比集合有用,但前提是用法正确。NVIDIA 和 AMD 是最干净的公开参照点,可用于衡量规模、披露质量和既有巨头强度。它们的官方产品页定义了内存、带宽和机架级集成上的技术门槛;2026 年文件也展示了已经以巨大商业规模运行的 AI 基础设施业务该如何充分披露。因此,它们是不可缺少的规模边界,但不是一家没有公开收入分母的私营创业公司的直接定价可比。 私营和专门化挑战者也很不均质。Groq 因为披露服务层级和本地部署可选项,商业上更容易理解。Cerebras 通过晶圆级集成系统竞争。低置信度二级市场报道显示,SambaNova 又属于另一个桶:一体化企业 AI 基础设施,而不是模块化插卡。d-Matrix 又在别处,试图销售一套贴近内存的推理栈,可以插进更标准的服务器和机架环境。这些不是细小的封装差异;它们会改变客户预算、部署摩擦,以及公开可观察的估值支撑类型。 最合理的推断是,Series C 投资人承保的是品类时点和架构契合度,而不是已经披露的财务规模。换句话说,他们大概率是在为几件事付费:推理需求会继续上升,市场会奖励更低延迟、更低功耗的替代方案,d-Matrix 能在既有巨头栈吸收机会之前,把合作伙伴、主权或企业兴趣转化成一个耐久利基。这个逻辑可以理解,但它仍不是公开估值支撑。[CV018, CV019, CV020, CV021, CV022, CV023]
| 参照对象 | 状态 | 观测指标 | 倍数 / 估值背景 | 为什么重要 | 关键限制 |
|---|---|---|---|---|---|
| d-Matrix 2025 年 11 月 Series C 轮 | 已验证私营估值 | 融资 $275M,估值 $2B;累计融资 $450M | 最新披露定价锚 | 审阅材料中唯一干净的私营估值数据点 | 无公开收入、ARR 或利润率分母 |
| NVIDIA FY2026 + FY2027 Q1 | 公开在位厂商规模边界 | FY2026 收入 $215.9B;数据中心收入 $193.737B;最新季度 $81.615B | 可作为披露和规模边界,不是创业公司倍数可比对象 | 展示完整披露的 AI 基础设施规模长什么样 | 训练和推理混在一起;业务远比 d-Matrix 宽 |
| AMD FY2025 + 2026 Q1 | 公开替代厂商规模边界 | FY2025 收入 $34.639B;数据中心收入 $16.635B;最新季度 $10.253B | 又一个规模和披露边界,而不是直接价格可比对象 | 说明即便是挑战者型在位厂商,也会披露 d-Matrix 没有披露的分母指标 | 同样混入推理之外更广的组合暴露 |
| Groq | 私营推理服务参照 | 公有云、私有云和共建云计划,以及 GroqRack 本地部署选项可见 | 商业化打包比 d-Matrix 公开定价更清楚 | 有助于比较商业模式和部署路径清晰度 | 审阅材料无法为当前估值提供高置信支撑 |
| Cerebras | 私营集成系统参照 | 晶圆级推理和训练平台;系统级架构 | 可作为差异化架构参照,不是模块化板卡可比对象 | 展示另一条推理差异化路径 | 审阅材料无法为当前估值提供高置信支撑 |
| SambaNova / 其他一体化推理硬件 | 低可信度私营公司参照 | 二手报道描述了一体化企业平台和近期融资 | 仅作为品类背景纳入,不作为数字定价锚 | 证明私营可比公司样本分散,并未标准化 | 来源质量不足以支撑高可信估值迁移 |
这张表有意混合一个已验证私营估值标记、两个公开规模边界和三个商业模式参照。它服务于承销逻辑, 不是干净的倍数筛选。
[CV001, CV022, CV023, CV026, CV027, CV028]相对 $2B Series C 标记的方向性证据调整说明,为什么当前公开中点低于上一轮。
柱状条是相对最新私募估值标记的方向性百万美元级调整,只用来展示哪些证据把判断上推或下拉。它们不是管理层指引或交易数据。
[CV013, CV015, CV017, CV025, CV039, CV040]8.4 乐观、基准与悲观情景
由于收入和 ARR 未披露,情景分析必须按里程碑走,而不是按倍数走。正确问题不是“私营 AI 芯片创业公司到底该按多少 EV / 收入倍数交易?”公开记录回答不了这个问题。正确问题是:什么证据状态足以支撑低于、接近或高于最近披露的 $2 billion 估值。 悲观情景假设商业化仍主要由试点和合作伙伴牵引,独立基准归一化仍未出现,下一轮融资或老股交易把 d-Matrix 的价格定在更接近 IP 价值加有条件商业可选性的位置,而不是已验证的平台领导力。这会产生一个宽泛但明显低于上一轮的区间,约为 $700 million 到 $1.0 billion。基准情景假设部署继续推进、融资渠道仍在,但仍没有强到足以证明 Series C 估值溢价的公开财务披露;这支撑大约 $1.2 billion 到 $1.8 billion。乐观情景要求具名大规模部署、被验证的经济性,以及更清楚的证据表明 d-Matrix 能在生产环境中与既有 GPU 基础设施共存或替代它;这支撑大约 $2.5 billion 到 $3.5 billion。 这个框架很重要,因为最近披露的一轮估值高于本章中点。Series C 不是完全无法辩护,但它已经计入了一条更接近乐观情景、而不是基准情景的执行路径。所以当前立场是偏高,而不是有吸引力。[CV041, CV042, CV043, CV044, CV045, CV046]
| 情景 | 证据状态假设 | 估值逻辑 | 隐含估值区间 | 概率信号 | 关键触发器 |
|---|---|---|---|---|---|
| 悲观 | 试点没有清晰转化,基准测试证明仍薄,下一轮融资重新定价风险 | 资产 / IP 加有条件商业选择权 | ~$700M-$1.0B | 25% | 融资重置或部署证据停滞 |
| 基准 | 部署推进且融资仍可获得,但分母指标仍不公开 | 围绕上一轮的里程碑进展,但不足以支撑完整溢价背书 | ~$1.2B-$1.8B | 50% | 有商业进展,但公开证据不足以证明 $2B+ 合理 |
| 乐观 | 具名大型部署、已验证经济性,以及在在位 GPU 旁边站得住的利基证明 | 战略稀缺性加已降风险的商业化 | ~$2.5B-$3.5B | 25% | 独立验证和可重复生产胜利 |
| 概率加权中点 | 上述三种状态的组合 | 基于公开证据混合出的情景重心 | ~$1.5B-$2.0B | 100% | 相对于 $2B Series C 轮估值,有助于约束估值假设 |
这些区间是以百万美元计的里程碑情景判断。它们不是披露价值、公允价值评估或收入倍数输出,因为所需分母指标不公开。
[CV041, CV042, CV043, CV045, CV046, CV047]以百万美元计的里程碑式悲观、基准和乐观估值区间,对比上一轮披露估值。
除最后一行披露的 Series C 估值外,所有数字都是以百万美元计的推断情景值。由于收入和 ARR 未公开,该区间以里程碑为基础。
[CV001, CV045, CV046, CV047, CV048, CV049]8.5 退出就绪度、逻辑失效触发点与最终尽调要求
公开证据更支持战略相关性,而不是退出就绪度。公司正在建设 AI 栈中最拥挤但也最重要的层之一;如果平台成为必备推理组件,确实会创造战略出售或后续老股交易的可选性。但当前公开披露不像已经 IPO 就绪。公开可比公司披露收入、利润率、文件节奏和分部表现;d-Matrix 没有。 披露缺口决定了叫停条件。如果一次定价融资显著低于 $1 billion,就会强烈表明市场已经重估 Series C 叙事。如果合作伙伴公告和早期使用计划无法转化成具名生产部署,商业化故事就会被削弱。如果独立基准无法支持公司声称的成本或效率优势,支付溢价购买专门化推理架构的核心理由就会消失。 因此,所需尽调必须具体,而不是泛泛而谈:经审计收入和 ARR、毛利率历史、客户集中度、订单到部署的转化、实际定价和支持条款,以及优先股堆叠或近期 409A 或老股估值。没有这些,新投资人承保的是 d-Matrix 可能重要这个故事,而不是 $2 billion 已经是纪律性入场价这个命题。[CV017, CV050, CV053, CV054, CV055, CV056]
| 触发因素 | 阈值 / 事件 | 投资逻辑传导 | 行动含义 |
|---|---|---|---|
| 融资重置 | 任何定价轮或老股成交价明显低于 ~$1B | 证明 C 轮叙事没有站住,压缩后续上行空间 | 按新的市场出清价格重建投资案例 |
| 商业化停滞 | 到下一轮融资周期仍无具名生产部署,或试点转生产转化偏弱 | 削弱“架构正在变成真实收入引擎”的论点 | 立即转向悲观情景 |
| 基准测试失败 | 独立同一负载测试未能验证实质性总拥有成本(TCO)或效率优势 | 拿掉为专用推理硬件支付溢价的核心理由 | 剔除溢价假设,重新评估投资意愿 |
| 经营披露薄弱 | 私下尽调显示毛利质量差、客户集中或转化低 | 技术故事变成低质量硬件商业化故事 | 要求大幅降价,否则退出 |
| 被现有巨头吸收 | 客户标准化采用 NVIDIA/AMD 或一体化替代方案,而不是引入 d-Matrix | 压缩 C 轮看似押注的细分空间 | 假设增长更慢、退出选项更少 |
放弃标准可监测,并直接绑定估值传导,而不是泛泛的公司风险。
[CV039, CV041, CV043, CV053, CV054, CV055]| 主题 | 缺失证据 | 重要性 | 负责人 / 尽调路径 |
|---|---|---|---|
| 经审计收入 / 年经常性收入(ARR) | 2024-2026 年月度和年度收入,以及年经常性收入(ARR) | 所有公开估值方法都缺少核心分母 | 索取经审计财报和董事会关键指标包 |
| 毛利率与现金转化 | 毛利率、运营费用、烧钱速度和营运资本画像 | 用来区分有价值的基础设施软件杠杆与昂贵硬件部署 | 索取经审计损益表、现金流历史和毛利率桥 |
| 客户证据 | 具名生产客户、队列扩张和头部账户集中度 | 用来区分伙伴活动与可持续经常性需求 | 索取客户名单、集中度明细和试点转生产数据 |
| 定价与支持经济性 | 实际平均售价(ASP)、折扣、软件附加销售和支持义务 | 用来检验公开总拥有成本(TCO)主张能否经受商业现实 | 索取价格表、试点发票和支持条款摘要 |
| 股权结构表与优先权 | 完全摊薄所有权、优先股堆叠、附函,以及近期任何 409A 或要约估值标记 | 因为在估值下调情景下,企业价值和普通股价值可能大幅分叉 | 索取股权结构表、瀑布模型,以及最新 409A 或老股交易数据 |
| 基准测试归一化 | 与 H100/H200/MI300X 以及实际进入候选名单的一体化挑战者做同模型第三方基准测试 | 用来验证本章全部溢价经济性假设 | 承销价格前委托或取得归一化基准证据 |
这些要求是阻断项,不是可选项。缺少它们,投资人押注的是战略叙事,而不是可辩护估值。
[CV017, CV030, CV054, CV055, CV056]8.6 附录
免责声明
本报告是基于公开证据的尽调快照,不构成投资建议。重要的财务、法律、技术和合同事实仍未公开;作出任何投资决定前,应直接向管理层核验,并对照一手文件验证。
证据索引
| 编号 | 陈述 | 可信度 | 来源 |
|---|---|---|---|
| CO001 | d-Matrix was founded in 2019. | 高 | SO002, SO004, SO006 |
| CO002 | d-Matrix is headquartered in Santa Clara, California. | 高 | SO004, SO005, SO006 |
| CO003 | The latest public office list in the fetched pack names Toronto, Sydney, Bangalore, and Belgrade in addition to the Santa Clara headquarters. | 高 | SO004, SO011 |
| CO004 | The founders publicly identified by d-Matrix are Sid Sheth and Sudeep Bhoja, with Sheth as CEO and Bhoja as CTO. | 高 | SO002, SO004 |
| CO005 | The public executive bench on the about page includes software, product, finance, legal, and manufacturing leaders. | 中 | SO002 |
| CO006 | The public board roster names Per Roman, Sasha Ostojic, Jeff Huber, Michael Stewart, Connie Sheng, and Russell Tham. | 中 | SO002 |
| CO007 | The fetched public record does not disclose board committees, ownership percentages, or explicit protective provisions behind the named board slate. | 中 | SO002, SO004, SO008 |
| CO008 | d-Matrix positions itself as an accelerated computing company built for AI inference in data centers rather than as a general-purpose training hardware vendor. | 高 | SO001, SO004 |
| CO009 | The home page says d-Matrix's 3DIMC architecture uses a chiplet-based, PCIe-friendly design that can scale to models up to 100 billion parameters. | 高 | SO001, SO004 |
| CO010 | By late 2025 d-Matrix publicly described a three-part product stack of Corsair accelerators, JetStream networking, and Aviator software. | 高 | SO004, SO022, SO024 |
| CO011 | VentureBeat reported that d-Matrix raised $44 million in a Series A round in April 2022. | 中 | SO014 |
| CO012 | d-Matrix closed a $110 million Series B funding round on September 6, 2023, led by Temasek. | 高 | SO006, SO012 |
| CO013 | The Series B release said the new capital would help d-Matrix commercialize Corsair and invest in recruitment and commercialization. | 中 | SO006 |
| CO014 | The Series B release listed Playground Global, M12, SK Hynix, Nautilus Venture Partners, Marvell Technology, and Entrada Ventures among d-Matrix backers and strategic partners. | 高 | SO006, SO012 |
| CO015 | d-Matrix said its November 2025 Series C raised $275 million at a $2 billion valuation and brought total disclosed funding to $450 million. | 高 | SO004, SO007, SO008, SO009 |
| CO016 | The 2025 Series C was co-led by Bullhound Capital, Triatomic Capital, and Temasek, with new participation from QIA and EDBI plus follow-on participation from M12, Nautilus Venture Partners, Industry Ventures, and Mirae Asset. | 高 | SO004, SO007, SO008 |
| CO017 | Independent coverage corroborated the $275 million raise, $2 billion valuation, and $450 million total funding figures announced in Series C. | 高 | SO009, SO010, SO011 |
| CO018 | The latest disclosed employee count in the fetched pack is 250+ worldwide from the November 2025 Series C key facts. | 中 | SO004 |
| CO019 | The November 2025 materials identify Santa Clara as headquarters and Toronto, Sydney, Bangalore, and Belgrade as global office locations. | 高 | SO004, SO011 |
| CO020 | The public record supports treating d-Matrix as a private Series C infrastructure startup rather than a public company or early research project. | 高 | SO004, SO009 |
| CO021 | d-Matrix unveiled Corsair on November 19, 2024 as a new computing platform designed for AI inference in modern datacenters. | 中 | SO005 |
| CO022 | The Corsair launch materials claimed 60,000 tokens per second at 1 millisecond per token for Llama3 8B in a single server and 30,000 tokens per second at 2 milliseconds per token for Llama3 70B in a single rack. | 中 | SO005 |
| CO023 | Corsair was sampling to early-access customers in November 2024 and was expected to be broadly available in Q2 2025. | 中 | SO005 |
| CO024 | d-Matrix launched JetStream in September 2025 as a custom I/O card and said the NICs deliver up to 400Gbps bandwidth with full production expected by year-end. | 中 | SO022 |
| CO025 | JetStream, paired with Corsair and Aviator, pushed d-Matrix toward a full compute-networking-software platform rather than a single-chip offering. | 高 | SO022, SO023 |
| CO026 | SquadRack was announced in October 2025 with Arista, Broadcom, and Supermicro as an open standards-based rack-scale reference architecture for AI inference. | 高 | SO023, SO004 |
| CO027 | GigaIO said in May 2025 that integrating Corsair with SuperNODE would support enterprise-scale inference and cited 30,000 tokens per second at 2 milliseconds per token on Llama3 70B. | 中 | SO020 |
| CO028 | The November 2025 Alchip collaboration pointed to a 3D DRAM roadmap and said the commercial debut would come on the Raptor accelerator as Corsair’s successor. | 中 | SO024 |
| CO029 | Gimlet Labs reported in March 2026 that offloading speculative decoding to d-Matrix Corsair produced 2–10x end-to-end latency improvements at similar energy efficiency versus a homogeneous GPU setup. | 中 | SO021 |
| CO030 | The d-Matrix news index lists a March 2026 Gimlet Labs collaboration and an April 2026 acquisition of GigaIO’s data center business as post-Series C milestones. | 中 | SO003 |
| CO031 | CRN included d-Matrix in its list of the 10 hottest semiconductor startups of 2023 after the company’s $110 million Series B. | 中 | SO012 |
| CO032 | Silicon Republic described d-Matrix as an AI startup to watch in 2024 because major product launches were scheduled after its $110 million funding round backed by Temasek and M12. | 中 | SO013 |
| CO033 | VentureBeat linked Microsoft Project Bonsai support to d-Matrix before Corsair was public, giving the company early ecosystem validation. | 中 | SO014 |
| CO034 | CNN reported in 2023 that GPU and interposer shortages were constraining AI infrastructure, with Sid Sheth arguing companies would need more efficient approaches. | 中 | SO015 |
| CO035 | CNBC described LLM inference costs as structurally high and quoted Sheth arguing that GPUs were not built for spiky inference workloads. | 中 | SO016, SO018 |
| CO036 | TechTarget likewise quoted Sheth saying inference economics are about dollars per inference, power efficiency, and latency rather than only raw compute. | 中 | SO017 |
| CO037 | Bloomberg described investor interest in chips that process data where it is stored, a market tailwind that aligns with d-Matrix's in-memory positioning. | 中 | SO019 |
| CO038 | d-Matrix and Bullhound both claimed up to 10x faster performance, 3x lower cost, and 3-5x better energy efficiency than GPU-based systems, but those ratios remain vendor claims rather than independently audited field benchmarks in the fetched pack. | 中 | SO004, SO007, SO011 |
| CO039 | Named customer disclosure remains thin because Series C and product releases mention rapid customer growth and large-scale deployments without identifying a broad hyperscale or enterprise customer roster. | 中 | SO004, SO009 |
| CO040 | The clearest public commercialization proof in the fetched corpus comes from partner or customer-adjacent materials from GigaIO and Gimlet rather than from a broad set of named end customers. | 中 | SO020, SO021 |
| CO041 | GFM Review’s 2024 competitive overview argued that startups like d-Matrix still must overcome Nvidia’s entrenched ecosystem, scale production, and win broader adoption. | 中 | SO025 |
| CO042 | The public board affiliations mirror the capital base by linking Bullhound, Playground, Triatomic, M12, Nautilus, and Temasek to visible governance seats. | 中 | SO002, SO004, SO007, SO008 |
| CO043 | Series C materials said the new capital would accelerate global expansion and support multiple large-scale deployments for hyperscale, enterprise, and sovereign customers. | 高 | SO004, SO007, SO009 |
| CO044 | No public revenue or ARR figure appears anywhere in the fetched company, investor, or independent source pack. | 低 | SO001, SO004, SO009 |
| CO045 | No public companywide customer count appears in the fetched pack even though the company and partners refer to growth and deployments. | 低 | SO004, SO009, SO021 |
| CO046 | Current control economics remain opaque because no fetched source discloses ownership percentages, board committees, or investor-protection mechanics. | 低 | SO002, SO004, SO008 |
| CM001 | d-Matrix’s product and technical materials consistently place the company in the datacenter inference market rather than the training market. | 高 | SM002, SM003, SM014 |
| CM002 | d-Matrix’s commercial form factor is merchant infrastructure built from PCIe cards, servers, racks, and Ethernet networking rather than a bespoke end-to-end datacenter stack. | 高 | SM002, SM004, SM006, SM014 |
| CM003 | The most relevant market boundary for d-Matrix is merchant datacenter generative inference infrastructure, excluding model training clusters and broad software or services spend. | 中 | SM002, SM003, SM011 |
| CM004 | Public source material separates inference into compute-bound prefill and memory-bound decode, implying that the relevant hardware market cannot be analyzed with a training-style compute lens alone. | 高 | SM009, SM022 |
| CM005 | Reasoning and agentic workflows make end-to-end latency more decision-relevant than raw tokens-per-second because each extra inference step compounds user waiting time. | 中 | SM005, SM009, SM012 |
| CM006 | Distillation and better smaller models broaden the set of inference workloads that can be served by specialized merchant hardware rather than only frontier-scale GPU clusters. | 中 | SM005, SM012 |
| CM007 | d-Matrix’s own product envelope implies a narrower commercial wedge focused on low-latency and memory-efficient serving rather than universal coverage of every model size or all AI compute. | 中 | SM002, SM003, SM005 |
| CM008 | Independent coverage shows that AI demand has strained GPU supply and datacenter expansion plans, creating a category-level opening for alternative inference architectures. | 高 | SM017, SM018, SM020 |
| CM009 | Microsoft’s public disclosure that AI datacenters depend on predictable energy, networking supplies, and GPUs shows that inference scaling is constrained by infrastructure inputs beyond compute silicon alone. | 高 | SM018, SM020 |
| CM010 | Memory movement between processor and memory chips is itself a major source of AI power consumption, which strengthens the commercial importance of architectures that reduce that movement. | 高 | SM003, SM016 |
| CM011 | d-Matrix’s DIMC thesis is that HBM-centric GPU architectures are too power-hungry and too memory-bound for the best latency economics in interactive inference. | 中 | SM003, SM014 |
| CM012 | d-Matrix’s JetStream materials argue that host-mediated communication introduces avoidable latency in multi-node inference and that device-initiated communication is a meaningful scaling advantage. | 中 | SM004, SM007, SM008 |
| CM013 | Batch size acts as a direct trade-off knob between user experience and hardware ROI in interactive inference workloads. | 中 | SM010, SM020 |
| CM014 | Public examples from generative-AI applications show that inference can become a material operating-cost burden rather than a negligible variable cost. | 高 | SM019, SM020 |
| CM015 | The existence of API-based inference and incumbent GPU clouds means part of the real substitute set for d-Matrix is a service choice, not only a hardware choice. | 中 | SM019, SM026, SM027 |
| CM016 | The most realistic early buyer segments for d-Matrix are enterprise or private-cloud operators, neocloud or managed-inference providers, AI product companies, and OEM or integrator channels. | 中 | SM006, SM021, SM026, SM027 |
| CM017 | Hyperscalers are strategically important but structurally harder for d-Matrix to win because they already control large GPU estates and increasingly use captive silicon. | 中 | SM018, SM028, SM029 |
| CM018 | d-Matrix’s standard PCIe and Ethernet positioning is more naturally aligned with enterprises and partner-led deployments than with fully custom hyperscale rack architectures. | 中 | SM004, SM006, SM014, SM021 |
| CM019 | On-prem or private-cloud optionality matters because peer inference providers such as Groq explicitly sell predictable-spend cloud plus on-prem deployment choices. | 中 | SM026 |
| CM020 | Integrators and OEMs are part of the buying path because production inference deployments require server qualification, networking design, and operational support in addition to accelerator cards. | 高 | SM003, SM014, SM021 |
| CM021 | GigaIO’s SuperNODE positioning suggests that some enterprise buyers will prefer scale-up systems that avoid complex distributed inference before they adopt full rack-scale multi-node designs. | 中 | SM021, SM014 |
| CM022 | Broad generative-AI spending forecasts are too wide to use as d-Matrix’s TAM because they include software, services, and categories far beyond merchant inference hardware. | 中 | SM028 |
| CM023 | Forbes cites a MarketsandMarkets estimate that the AI inference market will grow from $106.15 billion in 2025 to $254.98 billion by 2030. | 低 | SM028 |
| CM024 | Silicon Analysts estimates the total AI accelerator market at roughly $160 billion in 2025E and $200 billion plus in 2026E, which is broader than inference-only merchant hardware. | 低 | SM029 |
| CM025 | CNBC’s estimate that Bing AI required at least $4 billion of infrastructure is a useful lower-bound deployment-cost lens even though it is not a market-size estimate. | 中 | SM019 |
| CM026 | d-Matrix’s accessible SAM is narrower than either the published AI inference market or the total accelerator market because it excludes training clusters, captive hyperscaler silicon, and much API-consumed inference. | 中 | SM003, SM006, SM018, SM029 |
| CM027 | No clean public standalone SAM exists for open-standards merchant inference systems optimized for low-latency multi-node serving, so any precise public SAM or SOM would be overstated. | 中 | SM027, SM028, SM029 |
| CM028 | d-Matrix’s official materials claim up to 10-20x advantages on latency, throughput, power, or TCO versus GPU comparisons, but those results are explicitly preliminary and workload-specific. | 中 | SM002, SM003, SM014 |
| CM029 | A third-party developer-signal benchmark from Gimlet reports that using d-Matrix for speculative decoding can improve end-to-end request speed by 2-10x versus a GPU-only speculative decoder at equal energy efficiency. | 中 | SM022 |
| CM030 | Forbes reports that MLPerf Inference 5.0 results still showed Nvidia winning every submitted benchmark and that GB200 met Nvidia’s rack-scale inference claims on the cited tests. | 中 | SM028, SM025 |
| CM031 | Nvidia’s public product pages show that the incumbent GPU stack continues to improve memory capacity, bandwidth, and rack-scale inference performance from H100 to H200 to GB200. | 高 | SM023, SM024, SM025 |
| CM032 | The substitute set for d-Matrix now includes not only Nvidia GPUs but also Groq’s inference cloud and partner-led scale-up systems built around different accelerator assumptions. | 中 | SM021, SM023, SM024, SM025, SM026 |
| CM033 | CNBC described Nvidia as having about 95% of the market for AI chips in early 2023. | 中 | SM019 |
| CM034 | CNN cited Nvidia as controlling 84% of the market for discrete GPUs during the 2023 AI chip shortage. | 中 | SM018 |
| CM035 | Silicon Analysts estimates Nvidia at roughly 80-90% of AI accelerator revenue in 2025 and about 75% by 2026 as custom silicon broadens. | 低 | SM029 |
| CM036 | The public market-share numbers disagree because they use different denominators—AI chips, discrete GPUs, or total accelerators including custom silicon—so raw share rhetoric is directionally useful but not directly comparable. | 中 | SM018, SM019, SM029 |
| CM037 | d-Matrix’s standard-form-factor story reduces some retrofit friction relative to bespoke liquid-cooled racks, but it does not remove the need to prove cluster design, memory capacity, and operating economics at production scale. | 中 | SM004, SM006, SM014, SM021, SM025 |
| CM038 | Hyperscaler custom silicon growth and Nvidia’s software moat cap the share of the broad inference market that is realistically contestable by merchant startups. | 中 | SM018, SM028, SM029 |
| CM039 | Reasoning and agentic AI increase the amount of inference-time compute buyers may need, which helps the category thesis while simultaneously increasing sensitivity to serving cost and latency. | 中 | SM011, SM012, SM028 |
| CM040 | The most investable wedge for d-Matrix is latency-sensitive, memory-bound, open-standards inference serving where buyers care about predictable TCO and do not already own enough optimized GPU capacity. | 中 | SM006, SM009, SM010, SM021, SM022, SM026 |
| CM041 | d-Matrix’s current deployment narrative is strongest in enterprise and partner-led channels because public materials emphasize OEMs, integrators, scale-up partners, and standard datacenter compatibility. | 中 | SM006, SM014, SM015, SM021 |
| CM042 | d-Matrix’s product materials explicitly frame standard servers, racks, and existing datacenter infrastructure as a feature, which aligns with buyers who want incremental deployment rather than facility redesign. | 中 | SM002, SM004, SM006 |
| CM043 | Because prefill remains compute-bound and GPUs still excel there, the strongest public third-party evidence for d-Matrix today supports heterogeneous co-deployment rather than universal GPU replacement. | 中 | SM009, SM022, SM028 |
| CM044 | The same GPU shortages and power constraints that support d-Matrix’s market narrative can also push buyers to reserve incumbent capacity or stay with existing vendors instead of adopting a startup accelerator. | 中 | SM017, SM018, SM029 |
| CM045 | The public source pack is strong enough to prove a large inference category but not strong enough to justify a precise public SOM for d-Matrix without management disclosures on deployments, cards per cluster, and workload mix. | 中 | SM015, SM028, SM029 |
| CM046 | Vendor-authored heterogeneous-stage comparisons and standardized full-system leaderboards measure different things, so the public benchmark record does not establish one universally dominant inference architecture. | 中 | SM022, SM023, SM025, SM028 |
| CP001 | D-Matrix says its Corsair, JetStream, and Aviator stack can deliver 10x faster performance, 3x lower cost, and 3-5x better energy efficiency than GPU-based systems. | 中 | SP001, SP003 |
| CP002 | Independent reporting frames the core AI inference energy problem as repeated movement of data between memory and processors. | 中 | SP004 |
| CP003 | D-Matrix’s Jayhawk chiplet platform was publicly described with 16 Gbit/s per-wire bandwidth and less than 0.5 pJ/bit energy efficiency. | 中 | SP007, SP008 |
| CP004 | Jayhawk was also presented as a modular, pre-validated chiplet architecture that can refresh compute platforms faster and accommodate third-party chiplets. | 中 | SP007, SP008 |
| CP005 | Historical H100 scarcity and silicon-interposer bottlenecks pushed the market to look for more efficient second-source inference hardware. | 中 | SP009, SP010, SP012 |
| CP006 | NVIDIA markets H100 as combining 900 GB/s NVLink, NDR InfiniBand, and Magnum IO software to scale inference across clustered systems. | 中 | SP014 |
| CP007 | NVIDIA markets H100 as up to 30x faster for inference on the largest models and bundles enterprise deployment software around the platform. | 高 | SP014, SP021 |
| CP008 | NVIDIA says H200 increases memory to 141 GB of HBM3e at 4.8 TB/s, which it describes as nearly double H100 capacity and 1.4x more bandwidth. | 中 | SP015 |
| CP009 | NVIDIA says H200 can deliver up to 2x H100 inference speed on Llama2-class workloads while improving energy efficiency and total cost of ownership. | 中 | SP015 |
| CP010 | NVIDIA markets GB200 NVL72 as a liquid-cooled rack with 72 Blackwell GPUs, 36 Grace CPUs, and a 130 TB/s NVLink domain. | 中 | SP016 |
| CP011 | NVIDIA markets GB200 NVL72 as delivering 30x faster real-time trillion-parameter inference and Forbes says the NVL72 matched NVIDIA’s promised ~30x gain over an 8-GPU H200 benchmark. | 高 | SP016, SP022 |
| CP012 | AMD markets MI300X as a 192 GB HBM3, 5.3 TB/s, eight-Infinity-Fabric-link accelerator with ROCm software support. | 中 | SP017 |
| CP013 | AMD explicitly positions MI300X around higher memory capacity and bandwidth than competing accelerators, making memory-per-GPU its clearest attack on NVIDIA and d-Matrix alternatives. | 中 | SP017 |
| CP014 | Independent benchmark coverage compares MI300X directly with H100, H200, and B200 under current LLM inference workloads, showing AMD is treated as a real production competitor. | 高 | SP017, SP020 |
| CP015 | ROCm gives AMD a less locked-down software position than CUDA, but framework compatibility remains a major practical selection criterion for enterprise buyers. | 中 | SP017, SP021 |
| CP016 | Groq publicly sells inference as public, private, or co-cloud service with Free, Developer, and Enterprise plans plus optional on-prem GroqRack. | 高 | SP018, SP021 |
| CP017 | Groq’s public packaging emphasizes low latency, predictable spend, regional endpoints, and no-batching-required operation rather than training versatility. | 中 | SP018 |
| CP018 | Independent analysis describes Groq as an inference specialist with its own compiler and runtime path, which can improve latency but still requires ecosystem switching away from standard GPU tooling. | 中 | SP021, SP023 |
| CP019 | Cerebras markets WSE-3 as a wafer-scale processor with four trillion transistors and 125 petaflops that survives defects through redundant compute and routing. | 高 | SP019, SP021 |
| CP020 | Independent analysis says Cerebras now competes by serving very large models on specialized systems rather than sharding those workloads across many smaller GPUs. | 中 | SP021 |
| CP021 | Independent analysis describes SambaNova’s SN50 as a full-stack platform with 1.6 PFLOPS FP16, 3.2 PFLOPS FP8, and very large model and context-window support. | 中 | SP021 |
| CP022 | SambaNova’s Intel collaboration and hardware-software co-design show that major inference challengers increasingly bundle system integration with silicon. | 中 | SP021 |
| CP023 | Independent landscape coverage places d-Matrix, Groq, and Cerebras inside the same inference-challenger set even though they attack different technical bottlenecks. | 中 | SP005, SP023 |
| CP024 | The AI inference market is growing around low-latency and efficient deployment needs, but high power demand and supply chain risk still constrain adoption. | 中 | SP004, SP025 |
| CP025 | Historical deployment cost evidence shows why inference economics matter: one startup described nearly $200,000 in monthly model-serving bills and analysts estimated ChatGPT-scale infrastructure could require 10,000 Nvidia GPUs plus costly upgrades. | 中 | SP011, SP012, SP013 |
| CP026 | D-Matrix’s lower-cost and lower-energy pitch is strategically relevant because buyers are already conditioned to think of inference as an operating-cost problem. | 中 | SP001, SP003, SP011, SP012, SP013 |
| CP027 | D-Matrix is not trying to beat GB200 on absolute rack-scale throughput; it is offering a purpose-built inference card and NIC stack optimized for memory locality in standard servers. | 中 | SP001, SP003, SP008 |
| CP028 | Relative to Groq, d-Matrix competes less on public-cloud API distribution and more on hardware insertion into enterprise or OEM infrastructure. | 中 | SP001, SP018, SP021 |
| CP029 | Relative to Cerebras and SambaNova, d-Matrix is the less integrated but potentially easier-to-insert option because it fits around cards, NICs, and partner servers instead of a giant proprietary appliance. | 中 | SP001, SP019, SP021 |
| CP030 | NVIDIA remains the default incumbent because its hardware is ubiquitous and its software plus channel ecosystem create switching costs far beyond chip specifications. | 高 | SP014, SP016, SP021, SP026 |
| CP031 | Intuition Labs cites NVIDIA at roughly 92% of the discrete GPU market in H1 2025 while Silicon Analysts still shows a 75-87% share band through 2026, which is adverse evidence for any challenger moat story. | 中 | SP021, SP024 |
| CP032 | D-Matrix’s core switching-cost problem is therefore software and networking, not just silicon performance, because buyers already know how to buy, cool, connect, and operate incumbent GPU clusters. | 中 | SP014, SP016, SP021 |
| CP033 | AMD pressures NVIDIA on memory-per-GPU and open software posture, but it still competes inside the same OEM server procurement motion rather than changing the buyer workflow as sharply as Groq or Cerebras. | 中 | SP017, SP020, SP021 |
| CP034 | Groq has a clearer public story than d-Matrix for regulated or air-gapped deployments because it explicitly advertises GroqRack and cloud-to-local continuity. | 中 | SP018, SP021 |
| CP035 | Internal build is a credible substitute because large buyers increasingly want inference hardware that fits existing ML pipelines and can be co-designed around proprietary workloads. | 中 | SP021, SP025 |
| CP036 | D-Matrix’s strongest moat candidate is technical fit to the memory bottleneck, because both independent reporting and its chiplet disclosures link inference efficiency to keeping compute close to memory. | 中 | SP004, SP007, SP008 |
| CP037 | A second moat candidate is modularity, because Jayhawk and the broader card-plus-NIC architecture imply d-Matrix can partner into heterogeneous systems rather than forcing a clean-sheet deployment. | 中 | SP001, SP007, SP008 |
| CP038 | Adverse evidence is material because d-Matrix still lacks a public MLCommons-style apples-to-apples benchmark against H100, H200, MI300X, Groq, or Cerebras. | 中 | SP001, SP003, SP020, SP022 |
| CP039 | Without standardized public benchmarking, buyers still have to underwrite d-Matrix’s cost and performance claims primarily through vendor-authored comparisons. | 中 | SP001, SP003, SP022 |
| CP040 | Overall moat durability looks moderate rather than strong because d-Matrix is differentiated on inference-specific memory locality, but NVIDIA and AMD own broader ecosystems, Groq has cleaner service packaging, and Cerebras plus SambaNova sell more integrated systems. | 中 | SP021, SP023, SP024 |
| CI001 | D-Matrix publicly frames its monetizable platform as Corsair accelerators, JetStream NICs, and Aviator software. | 中 | SI001, SI022 |
| CI002 | D-Matrix disclosed a $275 million Series C at a $2 billion valuation, bringing total disclosed funding to $450 million. | 高 | SI001, SI007, SI008, SI009 |
| CI003 | Series C proceeds were earmarked for roadmap execution, global expansion, and multiple large-scale deployments. | 高 | SI001, SI008 |
| CI004 | D-Matrix disclosed a $110 million Series B in 2023 to begin commercializing Corsair after prior chiplet launches. | 中 | SI004 |
| CI005 | The Series B announcement said the company planned to use proceeds for recruitment and commercialization to meet demand for lower-cost inference infrastructure. | 中 | SI004 |
| CI006 | Public financing materials identify hyperscale, enterprise, and sovereign customers as the intended buying base. | 中 | SI001, SI008 |
| CI007 | Corsair was sampling to early-access customers in November 2024 and was targeted for broad availability in Q2 2025. | 中 | SI003 |
| CI008 | JetStream samples were available in September 2025 with full production expected by year-end 2025. | 中 | SI023, SI006 |
| CI009 | The combined Gimlet Cloud offering was targeted for select-customer availability in the second half of 2026. | 中 | SI026 |
| CI010 | Official materials say d-Matrix is collaborating with OEMs and system integrators to qualify Corsair-based solutions. | 中 | SI003, SI005 |
| CI011 | Public partner materials name Supermicro, Arista, Broadcom, GigaIO, and Liqid as deployment or integration partners around the d-Matrix stack. | 中 | SI021, SI022 |
| CI012 | The d-Matrix ecosystem page says partner logos are illustrative and not actual customers. | 中 | SI021 |
| CI013 | JetStream is designed as a standard PCIe card that connects to off-the-shelf top-of-rack Ethernet switches. | 中 | SI006, SI022 |
| CI014 | JetStream's published specification includes 400 Gbps bandwidth, PCIe Gen5 x16, and 150W max TDP with transceivers. | 中 | SI006 |
| CI015 | The Corsair white paper says an 8-card inference server can provide up to 2 TB of capacity memory. | 中 | SI005 |
| CI016 | The same white paper says an 8-card inference server provides 16 GB of performance memory and 1,200 TB/s of performance-memory bandwidth. | 中 | SI005 |
| CI017 | Official and partner materials repeatedly claim up to 10x faster speed, 3x better cost-performance, and 3x to 5x better energy efficiency versus GPU-based systems. | 中 | SI001, SI003, SI006, SI018 |
| CI018 | Both the Corsair white paper and the JetStream brief say performance, cost, and power estimates are preliminary and subject to change. | 高 | SI005, SI006 |
| CI019 | GigaIO says its SuperNODE can support dozens or 64+ Corsair accelerators in a single node and simplify deployment relative to multi-node setups. | 中 | SI018, SI021 |
| CI020 | SquadRack publicly combines Corsair, JetStream, Supermicro servers, Broadcom switches, Arista Ethernet, and Aviator software in one rack-scale reference architecture. | 中 | SI022 |
| CI021 | D-Matrix says scaling beyond SRAM into 3D-stacked DRAM is necessary to support larger reasoning models and higher token consumption. | 中 | SI025, SI027 |
| CI022 | The Alchip collaboration makes advanced ASIC design, packaging, and manufacturing management explicit dependencies of the next-generation roadmap. | 中 | SI025 |
| CI023 | D-Matrix's public workload story complements GPUs in heterogeneous inference stacks rather than replacing GPU training or all inference phases outright. | 中 | SI026, SI028, SI029 |
| CI024 | Gimlet reported that running the speculative decoder on Corsair improved end-to-end request latency by 2x to 10x at equivalent energy efficiency. | 中 | SI019 |
| CI025 | Gimlet reported that a 1.6B speculative decoder fits on two Corsair cards and that Corsair is air-cooled and rack-compatible. | 中 | SI019 |
| CI026 | The Microsoft Project Bonsai partnership provided early compiler and toolchain validation before broad commercial availability. | 中 | SI012 |
| CI027 | CNBC reported that Latitude spent nearly $200,000 per month on OpenAI and AWS at peak usage, showing how inference costs can swamp a small startup. | 中 | SI015 |
| CI028 | CNBC cited analyst estimates that OpenAI could spend about $40 million for one month of ChatGPT inference and Bing AI would need at least $4 billion of infrastructure. | 中 | SI015 |
| CI029 | The Washington Post reported that leading AI chatbots lose money on every chat because operating costs are so high. | 中 | SI017 |
| CI030 | TechTarget reported that ChatGPT-class deployments require not just GPUs but meaningful networking and power-management upgrades. | 中 | SI016 |
| CI031 | VentureBeat reported that the perceived GPU shortage often reflects board-level component and systems bottlenecks rather than a lack of GPU dies alone. | 中 | SI013 |
| CI032 | CNN reported that AI chip shortages are shaped by supply-chain bottlenecks such as advanced packaging and interposer constraints. | 中 | SI014 |
| CI033 | Bloomberg explained that repeatedly moving data between memory and processors is a major electricity cost in AI systems. | 中 | SI011 |
| CI034 | The Corsair white paper says chiplets improve yields and lower costs while moving the architectural problem toward interconnect design. | 中 | SI005 |
| CI035 | AIMultiple reports that cloud GPU list prices for the same model can differ several times over across providers. | 中 | SI020 |
| CI036 | No public Corsair or JetStream price sheet appears in the reviewed sources, so list price, discounting, and realized ASP remain undisclosed. | 中 | SI003, SI006, SI021, SI023 |
| CI037 | Reviewed official, investor, and independent materials disclose funding and deployment plans but no revenue or ARR figure. | 中 | SI001, SI007, SI008, SI009 |
| CI038 | Reviewed sources do not disclose gross margin, contribution margin, or gross profit dollars for d-Matrix. | 中 | SI001, SI003, SI005, SI006 |
| CI039 | No reviewed public source discloses d-Matrix cash on hand, monthly burn, or runway months. | 中 | SI001, SI004, SI007, SI008 |
| CI040 | No reviewed public source discloses debt facilities or project-finance obligations for d-Matrix. | 中 | SI001, SI004, SI007, SI008 |
| CI041 | Public customer evidence remains partner and benchmark oriented; the source pack does not provide a broad named-customer roster or booked-customer count. | 中 | SI001, SI019, SI021, SI026 |
| CI042 | GFM Review argues that startups like d-Matrix still need substantial capital investment, brand building, and proof at scale to challenge Nvidia's ecosystem. | 低 | SI031 |
| CI043 | NVIDIA's 2026 annual report says competitive data-center AI platforms are sold as co-designed chips, networking, systems, software, and paid enterprise software licenses. | 中 | SI030 |
| CI044 | NVIDIA's 2026 annual report says the company may place non-cancellable component orders, pay premiums or deposits, and absorb inventory provisions when demand shifts. | 中 | SI030 |
| CI045 | NVIDIA's filing defines cost of revenue to include wafers, packaging, board and device costs, memory, shipping, warranty, and inventory charges. | 中 | SI030 |
| CI046 | NVIDIA reported fiscal 2026 capex of $6.1 billion and said gross margin fell to 71.1% partly because full-scale datacenter solutions and a $4.5 billion excess-inventory charge weighed on results. | 中 | SI030 |
| CI047 | D-Matrix's public financial case is strongest as a customer-economics thesis rather than a publicly proven d-Matrix margin profile. | 中 | SI017, SI018, SI019, SI030 |
| CI048 | With $450 million of disclosed financing but no public cash, burn, or revenue data, d-Matrix still appears financing-dependent for roadmap execution and deployment scaling. | 中 | SI001, SI004, SI007, SI008 |
| CI049 | The go-to-market motion appears account-driven and qualification-heavy because official materials emphasize sampling, OEM qualification, and select-customer deployment windows. | 中 | SI003, SI005, SI026 |
| CI050 | Because JetStream adds dedicated networking hardware, power draw, and switch infrastructure on top of accelerator cards, d-Matrix's TCO case depends on whole-cluster design rather than chip benchmarks alone. | 中 | SI006, SI022, SI029 |
| CI051 | In a 2024 retrospective, management said d-Matrix came within two weeks of running out of cash in 2020 before reframing its fundraising strategy, showing capital availability has been existential since the company's earliest phase. | 中 | SI032 |
| CE001 | d-Matrix's current product surface consists of Corsair accelerators, JetStream networking, Aviator software, and the SquadRack reference rack architecture. | 高 | SE003, SE005, SE006, SE023 |
| CE002 | Corsair is packaged as an industry-standard PCIe Gen5 full-height full-length card and is sold in single-card and dual-card configurations for standard datacenter servers. | 高 | SE003, SE007, SE009 |
| CE003 | A single Corsair card is specified with 2 GB of on-chip performance memory, up to 256 GB of off-chip capacity memory, 150 TB/s memory bandwidth, and 2400 TFLOPs of MXINT8 peak compute. | 高 | SE007, SE009 |
| CE004 | A dual-card Corsair configuration is specified with 4 GB of performance memory, 300 TB/s of bandwidth, up to 512 GB of capacity memory, and cross-card DMX Bridge connectivity. | 高 | SE003, SE009 |
| CE005 | The published rack configuration uses eight servers and 64 Corsair cards to reach 128 GB of performance memory and 9.6 PB/s aggregate bandwidth. | 高 | SE003, SE009 |
| CE006 | d-Matrix built Corsair around DIMC because it views generative inference as memory-bound during token generation rather than only compute-bound. | 高 | SE007, SE009, SE016 |
| CE007 | Within Corsair, DMX Link provides die-to-die connectivity inside the package and DMX Bridge extends that connectivity across two cards. | 高 | SE007, SE009 |
| CE008 | Aviator's build flow comprises Model Factory, Compressor, and Compiler, while execution runs through Inference Engine and Host Runtime. | 高 | SE003, SE009, SE021 |
| CE009 | Aviator integrates with PyTorch, MLIR, Triton DSL, Kubernetes device plugins, metrics export, debugging, and profiling workflows. | 中 | SE009 |
| CE010 | JetStream is a transparent NIC that extends device-initiated accelerator-to-accelerator communication across nodes by bypassing host-orchestrated transfers. | 高 | SE010, SE014, SE015 |
| CE011 | JetStream is specified as a PCIe Gen5 x16 Ethernet card with 400 Gbps maximum bandwidth, QSFP-DD optics, 150 W max TDP, and secure boot support. | 高 | SE006, SE010 |
| CE012 | d-Matrix claims that JetStream combined with Corsair and Aviator can scale beyond 100B-parameter models and deliver up to 10x speed, 3x cost-performance, and 3x energy-efficiency gains versus GPU-based alternatives. | 中 | SE006, SE010 |
| CE013 | SquadRack is an eight-node rack blueprint that combines Corsair accelerators, JetStream I/O, Supermicro servers, Broadcom PCIe switches, Arista leaf switches, and Aviator software. | 高 | SE005, SE013 |
| CE014 | SquadRack is designed to remain air-cooled and configurable by rack power or rack-height constraints rather than requiring special liquid-cooling infrastructure. | 中 | SE013 |
| CE015 | GigaIO's SuperNODE integration is positioned to host dozens of Corsair accelerators in one node and publicly cites 30,000 tokens per second at 2 milliseconds per token for Llama3 70B. | 中 | SE023, SE026 |
| CE016 | Gimlet Labs reported that using Corsair for a speculative-decoding draft model delivered 2-5x end-to-end request speedups at equal energy relative to a GPU-only draft-model configuration. | 中 | SE027 |
| CE017 | d-Matrix argues that prefill is compute-bound while decode is memory-bound, so heterogeneous pipelines should mix GPUs with memory-centric accelerators instead of forcing every phase onto one device class. | 中 | SE016, SE017, SE018 |
| CE018 | In d-Matrix's published deployment model, speculative decoding lets Corsair run the memory-bound speculator stage while a larger GPU-resident model performs verification. | 中 | SE017, SE027 |
| CE019 | Keyformer is d-Matrix's published KV-cache sparsification technique and its example reports baseline accuracy with 50% of prompt KV cache plus 2.1x latency improvement and up to 2.4x token-throughput improvement. | 中 | SE020 |
| CE020 | dmx.compressor is a torch.fx-based quantization toolkit for PTQ and QAT that is meant to feed compressed graphs into Aviator Compiler. | 中 | SE021, SE009 |
| CE021 | Before Corsair, d-Matrix publicly described Nighthawk and then Jayhawk as chiplet platforms for in-memory inference, with Jayhawk introduced in 2023. | 中 | SE024, SE025 |
| CE022 | Jayhawk was presented as a modular chiplet platform on TSMC 6nm with 16 Gbps per wire bandwidth and less than 0.5 pJ/bit energy efficiency over organic substrates. | 中 | SE024, SE025 |
| CE023 | d-Matrix says 3DIMC has been validated on Pavehawk test silicon and is planned to debut commercially in the Raptor accelerator, the successor to Corsair. | 中 | SE008, SE011 |
| CE024 | d-Matrix claims that the future Raptor 3DIMC product will use 3D-stacked DRAM and deliver up to 10x faster inference than HBM4-based solutions. | 中 | SE008, SE011 |
| CE025 | The strategic purpose of 3DIMC is to extend d-Matrix's low-latency SRAM-centric architecture to larger-memory models and more demanding agentic pipelines. | 中 | SE011, SE012 |
| CE026 | d-Matrix's deployment model depends on partner qualification across server, switch, and rack layers rather than shipping a fully closed vertical appliance on its own. | 中 | SE004, SE007, SE013, SE026 |
| CE027 | The 3DIMC roadmap depends specifically on Alchip for ASIC and packaging expertise, making manufacturing execution partly external to d-Matrix. | 中 | SE008 |
| CE028 | The earlier Jayhawk platform also depended on TSMC process technology and BoW or UCIe-style chiplet ecosystems, reinforcing foundry and packaging exposure. | 中 | SE024, SE025 |
| CE029 | d-Matrix's current deployment story is intentionally open-standards based, using PCIe cards, standard Ethernet, and off-the-shelf top-of-rack switches rather than proprietary interconnect domains. | 高 | SE003, SE010, SE013 |
| CE030 | That open approach should lower retrofit friction in existing datacenters, but it also makes system quality dependent on partners such as Supermicro, Arista, Broadcom, GigaIO, and Liqid. | 中 | SE004, SE005, SE013 |
| CE031 | The incumbent GPU roadmap emphasizes HBM, NVLink, and increasingly liquid-cooled rack-scale systems, which differs materially from d-Matrix's PCIe-plus-Ethernet design center. | 中 | SE028, SE029, SE030 |
| CE032 | NVIDIA publishes H100 with 188 GB HBM3, H200 with 141 GB HBM3e and 4.8 TB/s bandwidth, and GB200 NVL72 as a liquid-cooled 72-GPU rack with 130 TB/s rack communication. | 中 | SE028, SE029, SE030 |
| CE033 | Independent market commentary says enterprise buyers increasingly want specialized inference hardware for data control, latency, and predictable cost rather than only general-purpose GPUs. | 中 | SE031, SE032 |
| CE034 | Independent market commentary also says NVIDIA still dominates accelerator share and packaging access, which raises the bar for any challenger that depends on external foundry and packaging capacity. | 中 | SE032, SE034 |
| CE035 | d-Matrix has at least one external practitioner validation signal because Gimlet publicly benchmarked Corsair in a speculative-decoding workflow. | 中 | SE027 |
| CE036 | The strongest public proof of deployment maturity is still at the reference-architecture and partner-demo layer rather than in named production customer case studies. | 中 | SE004, SE005, SE023, SE026 |
| CE037 | JetStream's published specification lists secure boot support, which is the clearest explicit product-security control surfaced in the reviewed corpus. | 中 | SE010 |
| CE038 | The reviewed official homepage, about, technology, product, ecosystem, and news surfaces do not publish SOC 2, ISO 27001, FedRAMP, or a public trust-center or status-page reference. | 中 | SE001, SE002, SE003, SE004, SE005 |
| CE039 | The reviewed corpus does not provide public MTBF, uptime, RMA, or field-failure metrics for Corsair or JetStream. | 中 | SE001, SE003, SE005, SE009, SE010 |
| CE040 | Company performance claims for Corsair, JetStream, and 3DIMC remain either preliminary or self-reported, so the independent benchmark base is still narrow. | 中 | SE009, SE010, SE027, SE033 |
| CE041 | The Raptor and Pavehawk roadmap has public architecture claims but no independently verified production silicon, customer deployments, or broad benchmark suite in this chapter corpus as of runDate. | 中 | SE008, SE011, SE023, SE033 |
| CE042 | d-Matrix's public developer surface is improving because Aviator builds on open-source tools and dmx.compressor points users to a GitHub repo, but the ecosystem still looks thinner than NVIDIA's enterprise software stack. | 低 | SE009, SE021, SE028 |
| CE043 | d-Matrix positions its hardware for reasoning, agents, and video-generation inference workloads where token growth and KV-cache pressure amplify memory bottlenecks. | 中 | SE007, SE012, SE019 |
| CE044 | The hybrid SRAM-plus-DRAM roadmap gives d-Matrix a technical middle ground between tiny SRAM-only deployments and massive HBM-heavy GPU clusters. | 中 | SE011, SE012, SE029 |
| CE045 | JetStream carries accelerator-to-accelerator traffic over standard Ethernet and standard optics, so rack networking depends on conventional datacenter switching rather than a bespoke fabric. | 高 | SE010, SE014 |
| CE046 | Aviator's decoupled host and on-card runtime is intended to reduce launch latency by letting the host enqueue work independently of the card's current execution state. | 中 | SE009 |
| CE047 | d-Matrix says it built JetStream because traditional RDMA NICs would not satisfy the latency target or preserve decoupled execution across nodes. | 中 | SE010, SE014 |
| CE048 | d-Matrix explicitly pitches SquadRack to cloud providers, sovereign clouds, and enterprises that want high-performance inference without redesigning the entire datacenter around proprietary infrastructure. | 中 | SE005, SE013 |
| CE049 | d-Matrix's recent technical framing argues that rising context lengths make inference a context-management problem as much as a compute problem, increasing pressure on memory capacity, KV-cache efficiency, and data movement. | 中 | SE020, SE036, SE041 |
| CE050 | d-Matrix argues that agentic and reasoning workloads magnify latency sensitivity because chained tool use, long contexts, and verifier steps make per-token delay matter more than peak batch throughput. | 中 | SE037, SE041 |
| CE051 | d-Matrix explicitly frames open standards as a prerequisite for broader AI adoption, reinforcing its use of PCIe, Ethernet, and standard datacenter components instead of proprietary rack domains. | 中 | SE013, SE038 |
| CE052 | d-Matrix's recent product messaging argues that competitive inference now depends on optimizing every layer from model compression and runtime to networking and infrastructure, not just the accelerator die. | 中 | SE022, SE039 |
| CE053 | d-Matrix published its own announcement of the GigaIO partnership, showing that the SuperNODE scale-up design is part of the company-authored deployment strategy rather than only a partner press claim. | 中 | SE026, SE035 |
| CE054 | Alchip's own announcement corroborates that the 3DIMC roadmap depends on external ASIC and advanced-packaging collaboration, reinforcing packaging execution as a real dependency rather than just company positioning. | 中 | SE008, SE040 |
| CE055 | d-Matrix's recent reasoning-at-enterprise-scale messaging treats context management, low latency, software-hardware co-design, and open deployment as one system recipe, which is consistent with but does not independently validate the broader product thesis. | 中 | SE037, SE039, SE041 |
| CU001 | d-Matrix says the Series C capital is meant to support hyperscale, enterprise, and sovereign customers. | 高 | SU002, SU016 |
| CU002 | SquadRack materials explicitly target cloud providers, sovereign clouds, and enterprises that need low-latency inference capacity. | 高 | SU003, SU030 |
| CU003 | The public sales narrative targets infrastructure buyers who want low-latency inference inside standard PCIe servers and Ethernet-based datacenters rather than bespoke GPU racks. | 中 | SU009, SU010, SU023 |
| CU004 | GigaIO frames the first scale-up commercialization path as enterprise deployment inside a single SuperNODE rather than as a named hyperscale production cluster. | 中 | SU007, SU020, SU029 |
| CU005 | Gimlet frames the joint offer around model providers, frontier AI labs, and AI-native cloud workloads that benefit from heterogeneous inference infrastructure. | 中 | SU008, SU021, SU027 |
| CU006 | d-Matrix says customer conversations with large cloud service providers helped shape its inference-first product thesis. | 中 | SU013 |
| CU007 | The disclosed proof base fits enterprise, sovereign, and specialist-cloud operators more clearly than it fits named hyperscale production accounts. | 中 | SU002, SU022, SU023, SU027 |
| CU008 | Corsair was being sampled to early-access customers in November 2024 and was targeted for broad availability in the second quarter of 2025. | 高 | SU006, SU028 |
| CU009 | JetStream samples were available by September 2025 with full production expected by the end of that year. | 中 | SU004 |
| CU010 | SquadRack configurations were slated to be available for purchase through Supermicro in the first quarter of 2026. | 高 | SU003, SU030 |
| CU011 | Series C materials say the new capital will support multiple large-scale deployments, but they do not identify those deployments by customer name. | 中 | SU002, SU016, SU017 |
| CU012 | The ecosystem page's Supermicro, GigaIO, and Liqid statements are partner endorsements rather than disclosed end-customer case studies. | 中 | SU001, SU006 |
| CU013 | The d-Matrix ecosystem page explicitly says the showcased partner logos are not actual customers. | 中 | SU001 |
| CU014 | GigaIO public materials include an early-access call to action for SuperNODEs running Corsair, indicating an evaluation funnel rather than a disclosed installed base. | 中 | SU020, SU029 |
| CU015 | The d-Matrix and Gimlet combined solution was planned for select customers in the second half of 2026, so it remains pre-broad-availability at the run date. | 高 | SU008, SU026, SU027 |
| CU016 | Gimlet is a named external operator planning to deploy Corsair alongside GPUs in Gimlet Cloud. | 高 | SU008, SU027 |
| CU017 | Gimlet's technical writeup shows a concrete evaluated workload using gpt-oss-120b with a 1.6B speculative decoder on Corsair and GPU. | 中 | SU021 |
| CU018 | Gimlet says two Corsair cards can host the speculative decoder and that the cards are air-cooled and rack-compatible, which lowers pilot friction inside existing datacenters. | 中 | SU021 |
| CU019 | GigaIO says a single SuperNODE can support dozens of Corsair accelerators, making scale-up inside one node the first enterprise expansion path before broader scale-out. | 中 | SU020, SU029 |
| CU020 | SquadRack extends the expansion path from one node to an eight-server rack and then to hundreds of nodes across multiple racks. | 高 | SU003, SU030 |
| CU021 | d-Matrix repeatedly frames commercialization as land-and-expand inside existing datacenters using standard PCIe servers and standard Ethernet networking. | 中 | SU009, SU010, SU030 |
| CU022 | Existing-datacenter compatibility, air cooling, and standards-based networking are central to the company's appeal for enterprise and sovereign buyers with infrastructure constraints. | 中 | SU008, SU009, SU027 |
| CU023 | Public commercialization depends on OEM and integrator partners including Supermicro, GigaIO, Arista, Broadcom, and Liqid rather than on a closed turnkey appliance sold only by d-Matrix. | 中 | SU001, SU003, SU023 |
| CU024 | That partner-led route can widen reach but also makes expansion sensitive to OEM qualification, support handoffs, and channel execution. | 中 | SU023, SU027, SU030 |
| CU025 | Enterprise private-inference buyers often choose on-premises or near-edge deployments for data control, latency, and predictable cost, which matches d-Matrix's positioning. | 中 | SU023 |
| CU026 | Major system integrators package accelerators into turnkey servers, so d-Matrix's OEM relationships are a core part of customer conversion rather than a side channel. | 中 | SU023, SU030 |
| CU027 | AIMultiple's 2026 cloud-provider tiering suggests that d-Matrix's disclosed proof today speaks more clearly to specialist AI clouds and enterprise operators than to named hyperscalers. | 中 | SU022, SU027 |
| CU028 | No reviewed public source discloses d-Matrix's current active customer count by account or by segment as of 2026-05-26. | 中 | SU002, SU016, SU017, SU019 |
| CU029 | No reviewed public source discloses d-Matrix's net revenue retention. | 中 | SU002, SU014, SU016 |
| CU030 | No reviewed public source discloses gross retention, churn, or renewal rates for d-Matrix deployments. | 中 | SU002, SU016, SU017 |
| CU031 | No reviewed public source discloses average contract length, backlog conversion, or repeat-order cadence for d-Matrix customers. | 中 | SU002, SU016, SU019 |
| CU032 | No reviewed public source discloses top-customer concentration, channel mix, or revenue contribution by partner versus end customer. | 中 | SU002, SU015, SU016 |
| CU033 | Customer evidence is materially fresher on partner integrations and deployment architecture than on paid production-account counts or cohort durability. | 中 | SU015, SU016, SU027, SU030 |
| CU034 | GFM Review argues that challengers like d-Matrix must persuade customers to adopt new technologies that may require significant changes to existing workflows, and the GroqCloud and Cerebras product pages show that buyers can compare d-Matrix against other non-Nvidia inference alternatives during that evaluation. | 中 | SU025, SU034, SU035 |
| CU035 | GFM Review also argues that widespread adoption depends on proving consistent, reliable performance at scale against Nvidia's trusted ecosystem. | 中 | SU025 |
| CU036 | IntuitionLabs says Nvidia still dominates enterprise inference hardware while OEMs package alternative accelerators into deployable systems, and AMD's MI350 page separately markets drop-in compatibility plus Kubernetes-friendly deployment, underscoring how much deployment polish d-Matrix must match to reduce procurement inertia. | 中 | SU023, SU033 |
| CU037 | Forbes argues Nvidia keeps lowering deployment friction through software, solutions, and datacenter-scale offerings, which raises the bar for d-Matrix customer adoption. | 中 | SU024 |
| CU038 | Data Center Knowledge says the value of the Gimlet approach depends on an abstraction layer that lets developers use heterogeneous chips without rewriting code. | 中 | SU027 |
| CU039 | Data Center Knowledge reports that initial Gimlet targets include frontier AI labs and that cloud and enterprise deployment at scale remain the strategic goal. | 中 | SU027 |
| CU040 | d-Matrix's Series C release claims rapid customer growth, but the absence of counts, cohort metrics, or named production accounts means that growth cannot be underwritten quantitatively. | 中 | SU002, SU014, SU016 |
| CU041 | Supermicro purchase availability in Q1 2026 shifts proof from reference architecture toward channel transactability, but still not toward a disclosed installed-base metric. | 中 | SU003, SU030 |
| CU042 | GigaIO calls the 2025 announcement the next phase of a strategic partnership, suggesting d-Matrix moved from component collaboration toward a more integrated system sales story. | 中 | SU020, SU029 |
| CU043 | Supermicro, GigaIO, and Liqid all pitch TCO, deployment speed, or inference flexibility, but none of the public statements quantify live end-customer outcomes by logo. | 中 | SU001, SU003, SU029 |
| CU044 | Both the Gimlet partnership page and GigaIO's AIwire press carry explicit interest or early-access calls to action, indicating that the funnel still includes invitation-driven evaluation motions. | 中 | SU026, SU029 |
| CU045 | The central customer diligence issue is the gap between credible partner validation and the still-undisclosed number of production customers, retention, and concentration metrics. | 中 | SU001, SU002, SU027, SU029 |
| CR001 | JetStream was announced with sample availability and a year-end production target, so public proof still sits at an early commercialization stage rather than a mature installed base. | 中 | SR002 |
| CR002 | SquadRack is presented as a reference blueprint built with Arista, Broadcom, and Supermicro rather than a closed appliance, so deployment readiness depends on partner qualification. | 中 | SR001 |
| CR003 | d-Matrix says it built a transparent NIC because standard networking approaches would not hit the latency targets needed for distributed inference. | 高 | SR005, SR006 |
| CR004 | JetStream adds accelerator-to-accelerator transfers, dedicated switching paths, and standard-Ethernet coordination that increase integration complexity across NIC, switch, server, and software layers. | 高 | SR002, SR005, SR006 |
| CR005 | The 3DIMC roadmap is still pre-commercial because Pavehawk is described as lab validated and Raptor as the planned first product use rather than a shipping platform. | 高 | SR003, SR004 |
| CR006 | The going-vertical post describes Pavehawk as d-Matrix’s first crack at the next memory problem and emphasizes lab stress testing instead of customer deployment or production-yield evidence. | 中 | SR004 |
| CR007 | d-Matrix’s five-year retrospective says the company pivoted from analog IMC to digital IMC and later from early transformer workloads to generative inference, showing technical flexibility but also roadmap volatility. | 中 | SR026 |
| CR008 | d-Matrix argues that reasoning-heavy workloads after the DeepSeek moment increase inference-time compute and memory pressure, which makes the thesis more relevant but also forces the roadmap to keep pace with rapidly changing model behavior. | 中 | SR008 |
| CR009 | eeNews Europe reported Jayhawk as a second-generation chiplet platform tied to TSMC 6nm and open chiplet interfaces, reinforcing that packaging and interconnect execution have long been core dependencies. | 中 | SR012 |
| CR010 | SiliconANGLE reported that a future Raptor generation is expected to move from 6nm to 4nm, which would raise qualification and supply-chain execution sensitivity if timing slips. | 中 | SR010 |
| CR011 | CNN reported that silicon interposers and advanced packaging are major AI-chip bottlenecks, suggesting smaller accelerator vendors can face supply risk even when end demand is healthy. | 中 | SR014 |
| CR012 | VentureBeat reported acute Nvidia GPU shortages during the AI boom, showing how upstream component shortages can distort pricing and lead times across adjacent inference ecosystems. | 中 | SR013 |
| CR013 | SquadRack and the GigaIO partnership show that d-Matrix depends on an external ecosystem of servers, switches, interconnects, and scale-up platforms rather than a vertically integrated manufacturing stack. | 高 | SR001, SR024 |
| CR014 | The Gimlet partnership describes select-customer availability in 2H 2026, which is encouraging external validation but still early compared with broad production deployment evidence. | 高 | SR025, SR031 |
| CR015 | d-Matrix’s public materials still emphasize reference architectures, early access, and partner launches more than a named production customer roster, leaving adoption depth under-evidenced. | 中 | SR001, SR025, SR031 |
| CR016 | TechTarget argued that enterprises still need significant compute, networking, and power investment for generative AI, implying sales cycles can stall even when accelerator efficiency is attractive on paper. | 中 | SR015 |
| CR017 | The Washington Post argued that chatbot providers were losing money on every chat in the early LLM buildout, so customer unit economics must improve for buyers to accept switching and integration risk. | 中 | SR016 |
| CR018 | d-Matrix says Corsair ships as an air-cooled standard PCIe card, which lowers retrofit friction but does not itself prove field reliability or supportability at scale. | 高 | SR025, SR031 |
| CR019 | The careers page shows d-Matrix operating across Santa Clara, Sydney, Bengaluru, Toronto, and Belgrade, broadening talent access but increasing coordination demands for a hardware-software platform. | 中 | SR030 |
| CR020 | The retrospective says d-Matrix nearly ran out of cash in 2020 before raising follow-on funding, showing that the company has already experienced financing fragility during a long hardware gestation period. | 中 | SR026 |
| CR021 | Data Center Dynamics and SiliconANGLE reported that d-Matrix raised $275 million in late 2025 at a $2 billion valuation, bringing total funding to roughly $450 million before broad commercialization. | 中 | SR009, SR010 |
| CR022 | The same financing coverage says the new capital is meant to support commercialization, global expansion, and large-scale deployments, which implies continued capital intensity ahead of mature revenue proof. | 中 | SR009, SR010 |
| CR023 | BIS said in October 2023 that the United States was strengthening export controls on advanced computing semiconductors and circumvention pathways to countries of concern, making trade policy an active risk variable for advanced AI chip supply chains. | 中 | SR029 |
| CR024 | d-Matrix’s privacy policy is a website data-handling disclosure that mentions reasonable safeguards and compliance-by-law, but it is not a published product-security assurance package for datacenter buyers. | 中 | SR027 |
| CR025 | d-Matrix’s terms of use include broad warranty disclaimers for the site, while the reviewed public corpus did not surface a dedicated trust center, SOC 2, ISO 27001, or similar enterprise-control page for the product stack. | 中 | SR027, SR028 |
| CR026 | Intuition Labs says private inference buyers care about data control, predictable cost, and deployment support, so public compliance and security gaps can directly slow adoption in regulated or sovereign environments. | 中 | SR022 |
| CR027 | Groq publicly markets zero-data-retention and on-prem air-gapped options, and NVIDIA markets enterprise security features, which raises the disclosure bar for d-Matrix on security posture. | 中 | SR020, SR017 |
| CR028 | NVIDIA H200 increases incumbent memory and bandwidth for inference, and GB200 pushes the competitive bar to rack-scale systems, so d-Matrix competes against a moving full-stack target rather than one static GPU baseline. | 高 | SR017, SR018 |
| CR029 | AMD MI300X offers 192GB of HBM3 and 5.3 TB/s bandwidth, blunting the argument that memory-hungry inference buyers have no mainstream alternatives to d-Matrix. | 中 | SR019 |
| CR030 | Groq and Cerebras both market purpose-built inference platforms with public deployment narratives, so d-Matrix is competing inside a crowded specialist category rather than a greenfield niche. | 中 | SR020, SR021 |
| CR031 | Intuition Labs describes switching costs as stack-wide across software, networking, cooling, and operations, which means d-Matrix must overcome ecosystem lock-in as well as raw chip performance. | 中 | SR022 |
| CR032 | Silicon Analysts estimates NVIDIA keeps dominant AI accelerator share through 2026, implying challengers like d-Matrix still fight for a relatively narrow share-of-wallet unless they win very specific workloads or power-constrained segments. | 中 | SR023 |
| CR033 | Bloomberg described the AI energy crisis as a data-movement problem, which validates d-Matrix’s architecture thesis but also means the company must prove rack-level efficiency rather than only chip-level superiority. | 高 | SR011, SR007 |
| CR034 | d-Matrix’s own disaggregation materials argue that prefill, decode, memory, and networking must be co-optimized, which means system-level overhead can erode theoretical efficiency gains if integration is poor. | 高 | SR006, SR007 |
| CR035 | SquadRack, the Gimlet materials, and d-Matrix’s brownfield-retrofit blog all stress air-cooled standard-server deployment and reuse of spare data-center capacity, so a key diligence question is whether those benefits persist once networking, orchestration, and support overhead are included in real customer environments. | 高 | SR001, SR025, SR036 |
| CR036 | The highest residual execution risk is synchronized delivery across Corsair, JetStream, Aviator, OEM partners, and the Raptor roadmap because no single layer is sufficient on its own to win production deployments. | 高 | SR001, SR002, SR003, SR024 |
| CR037 | A practical thesis-break trigger is a material slip in JetStream or Raptor readiness relative to customer procurement cycles because current proof is still early and partner-led. | 高 | SR002, SR003, SR025 |
| CR038 | A second thesis-break trigger is failure to convert early partner and reference deployments into named production customers by 2026 or 2027, which would leave the $2 billion valuation resting mostly on architecture promise. | 中 | SR009, SR010, SR031 |
| CR039 | A third thesis-break trigger is evidence that incumbent GPU or specialist inference alternatives close the latency-per-watt gap without d-Matrix’s integration burden, removing the switching rationale. | 中 | SR017, SR018, SR020, SR021, SR022 |
| CR040 | A fourth thesis-break trigger is a material export-control or advanced-packaging disruption that stalls next-generation product availability or worsens unit economics. | 中 | SR029, SR014, SR013 |
| CR041 | Public risk mitigation today is more mature around architectural positioning and partner announcements than around disclosed reliability metrics, compliance credentials, or customer concentration data. | 中 | SR001, SR002, SR025, SR027, SR028 |
| CR042 | No public MTBF, uptime, or field-failure metrics were identified in the reviewed risk corpus, leaving durability under production conditions unresolved. | 低 | SR001, SR002, SR027, SR028 |
| CR043 | No public named production customer list or deployment count was surfaced in the reviewed risk corpus beyond partner, reference, and early-access announcements. | 中 | SR001, SR025, SR031 |
| CR044 | The 2026 careers footprint and global expansion narrative suggest d-Matrix is scaling organizationally before large public customer proof, which can raise burn and program-management pressure if bookings lag. | 中 | SR009, SR026, SR030 |
| CR045 | d-Matrix’s risk profile is dominated not by one binary technical flaw but by compounding dependencies across roadmap maturity, ecosystem integration, customer proof, compliance disclosure, and financing discipline. | 高 | SR001, SR003, SR009, SR022, SR029 |
| CV001 | The latest disclosed d-Matrix financing was a $275 million Series C at a $2 billion valuation, bringing total disclosed funding to $450 million. | 高 | SV001, SV002, SV003, SV004 |
| CV002 | The 2025 round was described as oversubscribed and co-led by Bullhound Capital, Triatomic Capital, and Temasek, with QIA, EDBI, M12, and other investors participating. | 高 | SV001, SV002, SV003, SV004 |
| CV003 | d-Matrix said the Series C proceeds would advance the roadmap, accelerate global expansion, and support large-scale deployments for hyperscale, enterprise, and sovereign customers. | 中 | SV001, SV002, SV003, SV004 |
| CV004 | The Series C key facts disclosed 250+ employees and offices in Toronto, Sydney, Bangalore, and Belgrade in addition to the Santa Clara headquarters. | 中 | SV001 |
| CV005 | The Series C materials described rapid customer growth but did not disclose revenue, ARR, or customer count. | 中 | SV001, SV002, SV003, SV004 |
| CV006 | The reviewed public materials also do not disclose gross margin, retention, free cash flow, or cap-table preference terms for d-Matrix. | 中 | SV001, SV008, SV009, SV010, SV011, SV012 |
| CV007 | d-Matrix's 2023 Series B raised $110 million to begin commercializing Corsair after earlier chiplet launches and a lower-TCO inference pitch. | 中 | SV008 |
| CV008 | The founder's retrospective says d-Matrix first raised a $40 million Series A after pivoting toward digital in-memory compute and later refocused Corsair around generative AI inference. | 中 | SV010 |
| CV009 | Official and partner GigaIO materials say the combined SuperNODE plus Corsair system targets around 30,000 tokens per second at 2 milliseconds per token while lowering TCO and power versus GPU-heavy approaches. | 中 | SV011, SV013 |
| CV010 | d-Matrix's March 2026 Gimlet announcement said select customers could access the joint solution in the second half of 2026. | 中 | SV012 |
| CV011 | Gimlet's technical writeup found a 2-10x end-to-end interactivity improvement versus a GPU-only speculative-decoding setup at comparable energy efficiency in the studied configuration. | 中 | SV014 |
| CV012 | d-Matrix's 2024 Cambrian or Forbes repost highlighted a company claim of 30,000 tokens per second at 2 milliseconds per token for Llama 70B in a single rack. | 中 | SV009 |
| CV013 | The investor and partner mix suggests the Series C underwrites inference demand growth, energy efficiency, and strategic infrastructure optionality more than disclosed financial scale. | 中 | SV001, SV002, SV003, SV011, SV012 |
| CV014 | The 2025 round looks more like milestone financing around architecture and deployment readiness than conventional ARR-backed late-stage software underwriting. | 中 | SV001, SV003, SV007, SV008, SV010, SV012, SV013, SV014 |
| CV015 | Public support for d-Matrix's economic claims remains heavily company-authored because 10x performance, 3x lower cost, and 3-5x better energy efficiency recur across company, investor, and partner materials. | 中 | SV001, SV002, SV003, SV004, SV007, SV011, SV012, SV013 |
| CV016 | Independent corroboration exists but is narrower because Gimlet evaluated a specific speculative-decoding configuration rather than a whole-platform same-model benchmark against incumbent clusters. | 中 | SV012, SV014 |
| CV017 | No source reviewed in this chapter discloses d-Matrix revenue, ARR, gross margin, or free cash flow, so an EV/revenue or EV/ARR underwriting model cannot be supported publicly. | 中 | SV001, SV004, SV005, SV006, SV007, SV008, SV009, SV010, SV011, SV012 |
| CV018 | NVIDIA markets H100 as an inference accelerator integrated with NVLink, InfiniBand, Magnum IO, and AI Enterprise. | 中 | SV018 |
| CV019 | NVIDIA says H200 adds 141 GB of HBM3e and 4.8 TB/s bandwidth for larger-model inference. | 中 | SV019 |
| CV020 | NVIDIA markets GB200 NVL72 as a liquid-cooled rack-scale system with 72 Blackwell GPUs, 36 Grace CPUs, and a 130 TB/s NVLink domain. | 中 | SV020 |
| CV021 | AMD markets MI300X as a 192 GB HBM3 accelerator with 5.3 TB/s bandwidth and ROCm software support for large-model AI. | 中 | SV021 |
| CV022 | Groq publicly exposes inference service tiers and an on-prem GroqRack option, making its commercial packaging more visible than d-Matrix's current public pricing. | 中 | SV022 |
| CV023 | Cerebras positions wafer-scale hardware for both training and inference, which is a very different deployment model from a modular PCIe-card strategy. | 中 | SV023 |
| CV024 | AIMultiple and Forbes both frame inference hardware as a market still led by NVIDIA but increasingly judged on cost, throughput, and energy efficiency. | 中 | SV024, SV025 |
| CV025 | Silicon Analysts estimates NVIDIA still holds about 80-90% of AI accelerator revenue in 2025, indicating a highly concentrated market. | 低 | SV026 |
| CV026 | NVIDIA's FY2026 10-K reports $215.9 billion of revenue and $193.737 billion of Data Center revenue. | 高 | SV029, SV032, SV033, SV035 |
| CV027 | NVIDIA's Q1 FY2027 10-Q reports $81.615 billion of quarterly revenue and 74.9% gross margin for the quarter ended April 26, 2026. | 中 | SV032, SV033, SV035 |
| CV028 | AMD's FY2025 10-K reports $34.639 billion of total revenue, $16.635 billion of Data Center revenue, and 50% gross margin. | 高 | SV030, SV031, SV034, SV036 |
| CV029 | AMD's Q1 2026 10-Q reports $10.253 billion of quarterly revenue and 53% gross margin for the quarter ended March 28, 2026. | 中 | SV031, SV034, SV036 |
| CV030 | Because NVIDIA and AMD disclose multi-billion revenue and filing-backed margins while d-Matrix discloses none, public GPU-company valuation logic cannot be transferred mechanically onto d-Matrix. | 中 | SV001, SV029, SV030, SV031, SV032 |
| CV031 | Low-confidence secondary coverage portrays SambaNova as another integrated enterprise platform, reinforcing that private challengers span multiple business models rather than one clean comparable set. | 低 | SV022, SV023, SV027 |
| CV032 | The best-supported use of NVIDIA and AMD in this chapter is as scale fences and disclosure benchmarks, not as direct pricing anchors for a private startup with hidden denominator metrics. | 中 | SV018, SV019, SV020, SV021, SV029, SV030 |
| CV033 | CRN and Silicon Republic showed d-Matrix had startup mindshare by 2023-2024, but press attention is not equivalent to named production-customer disclosure. | 中 | SV016, SV017 |
| CV034 | Partner references with GigaIO and Gimlet show ecosystem progress, but they still stop short of public recurring-revenue or customer-concentration disclosure. | 中 | SV011, SV012, SV013, SV014 |
| CV035 | Groq's packaging is more commercially legible than d-Matrix's because public plan tiers and on-prem optionality are visible while d-Matrix public pricing remains absent. | 中 | SV001, SV011, SV012, SV022 |
| CV036 | Cerebras and SambaNova-style integrated systems imply a different buyer motion from d-Matrix's modular PCIe-card and NIC approach. | 低 | SV013, SV023, SV027 |
| CV037 | Bloomberg and the market-report sources reinforce that energy, power, and infrastructure cost are central to inference buying and help explain why investors may pay for efficiency narratives. | 中 | SV015, SV024, SV028 |
| CV038 | The same sources also show that high power requirements, data-security concerns, and supply-chain stress remain structural adoption risks for inference hardware. | 低 | SV015, SV028 |
| CV039 | Concentrated incumbent share and stack lock-in argue against paying a startup scarcity premium unless d-Matrix proves repeatable production wins. | 中 | SV018, SV020, SV022, SV024, SV025, SV026 |
| CV040 | The Series C price can be rationalized only if investors believe inference demand will shift materially toward specialized hardware and d-Matrix will win a meaningful share of that spend. | 中 | SV001, SV002, SV003, SV015, SV024, SV025, SV026 |
| CV041 | A reasonable bull case requires named scale deployments, independent benchmark normalization, and evidence that customers choose d-Matrix for production inference rather than pilots. | 中 | SV003, SV010, SV011, SV012, SV014, SV017 |
| CV042 | The base case assumes deployments continue and the company avoids a down round, but public evidence remains insufficient to prove premium economics above the latest private mark. | 中 | SV003, SV009, SV010, SV011, SV012, SV017 |
| CV043 | The bear case is plausible because commercialization proof is still thin relative to the scale of NVIDIA and AMD incumbency and the fragmentation of private challengers. | 中 | SV017, SV024, SV025, SV026, SV029, SV030 |
| CV044 | Valuation therefore has to be milestone-based rather than multiple-based because the denominator is hidden and the comp set is structurally mismatched. | 中 | SV017, SV029, SV030, SV031, SV032 |
| CV045 | This chapter's valuation ranges are scenario judgments around the last disclosed round, not market-cleared fair values or implied revenue-multiple outputs. | 中 | SV001, SV017, SV029, SV030, SV031, SV032 |
| CV046 | A bear-case valuation range of roughly $700 million to $1.0 billion reflects IP and partner optionality but assumes pilots stall or the next financing resets materially below Series C. | 中 | SV009, SV011, SV012, SV013, SV017, SV024, SV026 |
| CV047 | A base-case valuation range of roughly $1.2 billion to $1.8 billion assumes commercialization progresses and financing remains available, but public disclosure still does not justify a clear premium above the latest round. | 中 | SV001, SV003, SV010, SV011, SV012, SV015, SV017 |
| CV048 | A bull-case valuation range of roughly $2.5 billion to $3.5 billion requires named large deployments, third-party validated economics, and evidence that d-Matrix can hold a differentiated niche beside incumbent GPU infrastructure. | 中 | SV003, SV011, SV012, SV013, SV014, SV018, SV020 |
| CV049 | A 25% bear, 50% base, and 25% bull weighting centers the public-evidence midpoint around roughly $1.5 billion to $2.0 billion, below the last disclosed $2 billion round. | 中 | SV001, SV017, SV029, SV030, SV031, SV032 |
| CV050 | Because the Series C price sits above the chapter's midpoint and still requires unproven bull milestones, the current public evidence supports a stretched valuation stance. | 中 | SV001, SV017, SV029, SV030, SV031, SV032 |
| CV051 | The recommendation is research-more rather than buy because the company may be strategically relevant, but the price is ahead of public proof. | 中 | SV001, SV017, SV029, SV030, SV031, SV032 |
| CV052 | Confidence is medium because financing and product facts are real, yet the most decision-critical operating metrics remain undisclosed. | 中 | SV001, SV003, SV017 |
| CV053 | Risk rating is high because the investment case still depends on commercialization proof, supply-chain execution, and a financing or exit path that public evidence does not yet de-risk. | 中 | SV013, SV014, SV015, SV017, SV026, SV027, SV028 |
| CV054 | Entry discipline should require a materially lower price or hard data-room evidence on revenue quality and cap-table terms. | 中 | SV017, SV030, SV031, SV032 |
| CV055 | Thesis-breakers include a down round below roughly $1 billion, failure to convert pilots into named deployments, or independent benchmarks that negate the claimed TCO edge. | 中 | SV011, SV012, SV013, SV014, SV017, SV028 |
| CV056 | Diligence should request audited revenue or ARR, gross margin, customer concentration, bookings-to-deployment conversion, realized pricing, and preference-stack details before underwriting the Series C mark. | 中 | SV017, SV030, SV031, SV032 |
| CV057 | The reviewed corpus does not provide high-confidence private-market valuation data for Groq, Cerebras, or SambaNova comparable to d-Matrix's disclosed $2 billion round, so they can inform business-model comparison but not numeric pricing. | 低 | SV022, SV023, SV027 |
| CV058 | Public evidence supports strategic sale or later secondary optionality more than a near-term IPO because d-Matrix has not reached public-company disclosure quality in the reviewed corpus. | 中 | SV017, SV029, SV030 |
| 编号 | 出版方 | 标题 | 引文 |
|---|---|---|---|
| SO001 | d-Matrix | d-Matrix - Ultra-low Latency Batched Inference for Generative AI | Chiplet-based design enables scaling SRAM-based architecture to power models up to 100B parameters. |
| SO002 | d-Matrix | About d-Matrix | Innovators in AI Compute Solutions & Technology | d-Matrix is led by a team of dedicated entrepreneurs with a 20+ year history in building businesses that have shipped over 100M chips and generated over $5B in revenue. |
| SO003 | d-Matrix | d-Matrix News | |
| SO004 | d-Matrix | d-Matrix Raises $275 Million to Power the Age of AI Inference | Founded: 2019 | HQ: Santa Clara, CA | Global Offices: Toronto (Canada); Sydney (Australia); Bangalore (India); Belgrade (Serbia) | Employees: 250+ worldwide. |
| SO005 | d-Matrix | d-Matrix Unveils Corsair, the World’s Most Efficient AI Computing Platform for Inference in Datacenters | Corsair is sampling to early-access customers and will be broadly available in Q2’2025. |
| SO006 | d-Matrix | d-Matrix Announces $110 Million in Series B Funding to Make Generative AI Commercially Viable with First-of-Its-Kind Inference Compute Platform | d-Matrix was founded in 2019 to solve the memory-compute integration problem, which is the final frontier in AI compute efficiency. |
| SO007 | Bullhound Capital | Bullhound Capital leads $275M investment into AI inference leader d-Matrix | d-Matrix ... has closed $275 million in Series C funding, valuing the company at $2 billion and bringing the total raised to date to $450 million. |
| SO008 | Qatar Investment Authority | QIA joins d-Matrix’s Series C USD 275m funding round | Valued at USD 2 billion and bringing the total raised to date to USD 450 million, d-Matrix will use the new capital to advance their roadmap and accelerate global expansion. |
| SO009 | Data Center Dynamics | Chip startup d-Matrix raises $275m against $2bn valuation | |
| SO010 | SiliconANGLE | Chip startup d-Matrix raises $275M to speed up inference with in-memory compute | |
| SO011 | Converge Digest | d-Matrix Raises $275 Million to Accelerate the Age of AI Inference | |
| SO012 | CRN | 10 Hottest Semiconductor Startups 2023 | |
| SO013 | Silicon Republic | 24 tech start-ups to watch in 2024 | |
| SO014 | VentureBeat | New Microsoft partnership accelerates generative AI development | |
| SO015 | CNN | AI chips supply chain bottlenecks | |
| SO016 | CNBC | ChatGPT and generative AI are booming, but the costs can be extraordinary | |
| SO017 | TechTarget | Infrastructure to support OpenAI’s ChatGPT could be costly | |
| SO018 | The Washington Post | The hidden cost of ChatGPT is all the computing power | |
| SO019 | Bloomberg | AI Energy Crisis Boosts Need for More Efficient Chips | |
| SO020 | GigaIO | GigaIO Partners with d-Matrix to Deliver Ultra-Efficient Scale-Up AI Inference Platform | The new GigaIO SuperNODE platform, capable of supporting dozens of d-Matrix Corsair accelerators in a single node, is now the industry’s most scalable AI inference platform. |
| SO021 | Gimlet Labs | Low-Latency Inference with Speculative Decoding on d-Matrix Corsair and GPU | Compared to the same speculative decoder on GPU and equivalent energy consumption, we've found that the Corsair-based solution delivers 2-5X end-to-end request speedup and up to 10X end-to-end speedup for energy-optimized configurations. |
| SO022 | d-Matrix | d-Matrix Announces JetStream I/O Accelerators Enabling Ultra-Low Latency for AI Inference at Scale | JetStream NICs are full-height PCIe Gen5 cards delivering a maximum 400Gpbs bandwidth. Samples are available now, with full production expected by year-end. |
| SO023 | d-Matrix | d-Matrix Announces SquadRack, Industry’s First Rack-Scale Solution Purpose-Built for AI Inference at Datacenter Scale | SquadRack™, the industry’s first blueprint for disaggregated standards-based rack-scale solutions for ultra-low latency batched inference. |
| SO024 | d-Matrix | d-Matrix and Alchip Announce Collaboration on World’s First 3D DRAM Solution to Supercharge AI Inference | 3DIMC will commercially debut on the d-Matrix Raptor™ inference accelerator, the successor to d-Matrix Corsair™. |
| SO025 | GFM Review | The New Wave of AI Chips: Can Cerebras, d-Matrix and Groq take on Nvidia? | These new entrants face several hurdles, including the need for substantial capital investment to scale production, challenges in building brand recognition, and the difficulty of competing against Nvidia’s established ecosystem. |
| SM001 | d-Matrix | Technology | |
| SM002 | d-Matrix | Product | Corsair is the first-of-its-kind AI compute platform. It offers unparalleled performance and efficiency for generative AI inference in the datacenter. |
| SM003 | d-Matrix | d-Matrix Corsair Redefines Performance and Efficiency for AI Inference at Scale | Traditional accelerators often use costly and power-hungry High Bandwidth Memory (HBM)... d-Matrix breaks this memory barrier by integrating a multiplier directly into memory bit cell using a logic process. |
| SM004 | d-Matrix | JetStream Product Brief | d-Matrix JetStream is a purpose-built network interface card (NIC) enabling efficient scaling of AI workloads and delivering ultra-low latency inference with Corsair clusters. |
| SM005 | d-Matrix | The Power of the Middle Lane: Why a Hybridized Approach to Memory Gives the Best of Both Worlds | The per-token compute for a given inference has become the “cheap” part of the inference process, while that movement to and from memory is a massive “tax” per token generation. |
| SM006 | d-Matrix | Scaling AI the Right Way: Introducing Our Rack-level Inference Solution | SquadRack is a reference rack blueprint ... based on rack power or rack height constraints. And it is air-cooled and doesn’t have any special cooling infrastructure requirements. |
| SM007 | d-Matrix | Why We Needed a New Transparent NIC Solution | We consider there to be three total barriers to fast, efficient, high-performance AI inference: compute, memory, and I/O. |
| SM008 | d-Matrix | Why We Decoupled Execution to Accelerate I/O | Rack-scale and multi-rack solutions with standard RoCE v2 systems accrue a large penalty scaling linearly due to the inherent costs of a system where the control plane and data plane both require a series of handshakes. |
| SM009 | d-Matrix | Why Modern AI Workloads Demand a Disaggregated Approach | During decode, however, each new token is dependent on information from the prior one... making the speed of the response governed by the speed at which those trips to and from memory happen. |
| SM010 | d-Matrix | Batching Just Right: How Interactive Apps Serve as a New Battleground | Lowering a batch size means a snappier, lower-latency user experience; while raising it increases overall throughput (and improves the economics). |
| SM011 | d-Matrix | What Is AI Inference and Why It Matters in the Age of Generative AI | Generative AI models are orders of magnitude bigger ... They need multiple cards or servers to perform inference, consuming significantly more computational resources and energy. |
| SM012 | d-Matrix | Impact of the DeepSeek Moment on Inference Compute | Reasoning models are highly memory bound and end up underutilizing the GPUs that are optimized for training. |
| SM013 | d-Matrix | Introducing dmx.compressor | Quantization plays a key role in reducing memory usage, speeding up inference, and lowering energy consumption at inference time. |
| SM014 | d-Matrix | d-Matrix Unveils Corsair, the World’s Most Efficient AI Computing Platform for Inference in Datacenters | Corsair delivers up to 10x faster interactive speed, 3x better performance per total cost of ownership (TCO), and 3x greater energy efficiency. |
| SM015 | Data Center Dynamics | Chip startup d-Matrix raises $275m against $2bn valuation | It brings the total raised by d-Matrix to $450 million, which the company said it will use to advance its roadmap, accelerate global expansion, and support multiple large-scale deployments of its data center inference platform. |
| SM016 | Bloomberg | AI Energy Crisis Boosts Need for More Efficient Chips | Every time the algorithm needs data, it goes to the library, known as a memory chip, checks out the data and takes it to another chip, known as a processor... a process that burns through lots of electricity. |
| SM017 | VentureBeat | Nvidia GPU shortage is top gossip of Silicon Valley | The problem with inference is if the workload spikes very rapidly ... There is no way your GPU capacity can keep up with that because it was not built for that. It was built for training. |
| SM018 | CNN | AI chips supply chain is straining under demand | Our datacenters depend on the availability of permitted and buildable land, predictable energy, networking supplies, and servers, including graphics processing units. |
| SM019 | CNBC | ChatGPT and generative AI are booming, but the costs can be extraordinary | Curran believes that it could have cost OpenAI $40 million to process the millions of prompts people fed into the software that month. |
| SM020 | TechTarget | Infrastructure to support OpenAI’s ChatGPT could be costly | Users have to build up not just compute, but their networking and power management. |
| SM021 | GigaIO | GigaIO Partners with d-Matrix to Deliver Ultra-Efficient Scale-Up AI Inference Platform | This integration enables enterprises to deploy ultra-low-latency batched inference workloads at scale without the complexity of traditional distributed computing approaches. |
| SM022 | Gimlet Labs | Low-Latency Inference with Speculative Decoding on d-Matrix Corsair and GPU | Compared to the same speculative decoder on GPU and equivalent energy consumption, we've found that the Corsair-based solution delivers 2-5X end-to-end request speedup on configurations optimized for interactivity, and up to 10X end-to-end speedup for energy-optimized configurations. |
| SM023 | NVIDIA | H100 Tensor Core GPU | H100 extends NVIDIA’s market-leading inference leadership with several advancements that accelerate inference by up to 30X and deliver the lowest latency. |
| SM024 | NVIDIA | H200 Tensor Core GPU | The H200 boosts inference speed by up to 2X compared to H100 GPUs when handling LLMs like Llama2. |
| SM025 | NVIDIA | GB200 NVL72 | The NVIDIA GB200 NVL72 ... delivers 30x faster real-time trillion-parameter large language model inference, with 10x greater performance for mixture-of-experts architectures. |
| SM026 | Groq | Products | The AI inference platform built for developers. Fast responses, scalable performance, and costs you can plan for. |
| SM027 | AIMultiple | AI Hardware | Hyperscalers run broad cloud platforms with GPU rental as one product among many. Specialist neoclouds focus on GPU and AI infrastructure as their core product. |
| SM028 | Forbes | AI Inference Is King: Do You Know Which Chip Is Best? | MarketsandMarkets projects that the AI inference market is expected to grow from $106.15 billion in 2025 to $254.98 billion by 2030. |
| SM029 | Silicon Analysts | Nvidia AI Accelerator Market Share 2024-2026 | NVIDIA commands approximately 80-90% of the AI accelerator market by revenue as of 2025. |
| SP001 | Data Center Dynamics | Chip startup d-Matrix raises $275m against $2bn valuation | Its full-stack platform is powered by the company’s Corsair inference accelerators, JetStream NICs, and Aviator software, which it claims can deliver 10X faster performance, 3X lower cost, and 3–5X better energy efficiency than GPU-based systems. |
| SP002 | SiliconANGLE | Chip startup d-Matrix raises $275M to speed up inference with in-memory compute | |
| SP003 | TechStartups | AI chip startup d-Matrix raises $275M at $2B valuation to challenge Nvidia’s grip on AI inference | Traditional GPUs, including Nvidia’s, move data constantly between memory and processing cores, consuming huge amounts of energy. |
| SP004 | Bloomberg | AI Energy Crisis Boosts Need for More Efficient Chips | AI requires massive amounts of data... a process that burns through lots of electricity. |
| SP005 | CRN | 10 Hottest Semiconductor Startups of 2023 | Nvidia may dominate the AI computing space... but that isn’t stopping a slew of semiconductor startup companies, such as Cerebras Systems, d-Matrix, Lightmatter and Tenstorrent. |
| SP006 | Silicon Republic | Start-ups to Watch in 2024 | d-Matrix claims its chips are optimised to support generative AI systems and is competing in the same space as Nvidia. |
| SP007 | eeNews Europe | Second-generation chiplet platform targets generative AI | The Jayhawk chiplet platform features... 16 Gbit/s/wire bandwidth and less than 0.5 pJ/bit energy efficiency. |
| SP008 | HPCwire | d-Matrix Launches New Chiplet Connectivity Platform to Address Growing Compute Demand for Generative AI | By using a modular chiplet-based approach, data center customers can refresh compute platforms on a much faster cadence using a pre-validated chiplet architecture. |
| SP009 | VentureBeat | Nvidia GPU shortage is top gossip of Silicon Valley | For LLM training... there is no substitute for Nvidia’s H100. |
| SP010 | CNN | AI chips supply chain bottlenecks | The shortage could force companies to find creative ways around the problem. |
| SP011 | CNBC | ChatGPT and generative AI are booming, but the costs can be extraordinary | At its peak in 2021, Latitude was spending nearly $200,000 a month on OpenAI’s software and Amazon Web Services. |
| SP012 | TechTarget | Infrastructure to support OpenAI’s ChatGPT could be costly | Reports surfacing earlier this month indicate that just to develop training models and inferencing alone for OpenAI’s ChatGPT can require 10,000 Nvidia GPUs. |
| SP013 | The Washington Post | ChatGPT’s hidden cost is GPU compute | The enormous cost of running today’s large language models is limiting their quality and threatening to throttle the global AI boom. |
| SP014 | NVIDIA | NVIDIA H100 | H100 extends NVIDIA’s market-leading inference leadership... with several advancements that accelerate inference by up to 30X and deliver the lowest latency. |
| SP015 | NVIDIA | NVIDIA H200 | The NVIDIA H200 is the first GPU to offer 141GB of HBM3e memory at 4.8TB/s. |
| SP016 | NVIDIA | NVIDIA GB200 NVL72 | The NVIDIA GB200 NVL72... delivers 30x faster real-time trillion-parameter LLM inference. |
| SP017 | AMD | AMD Instinct MI300X Accelerators | Dedicated Memory Size 192 GB... Peak Memory Bandwidth 5.3 TB/s... Supported Technologies AMD ROCm. |
| SP018 | Groq | Groq Products | Available in public, private, or co-cloud instances. |
| SP019 | Cerebras | Cerebras Product Chip | Four trillion transistors. 125 petaflops. One silicon wafer. |
| SP020 | AIMultiple | AI Hardware | We benchmarked NVIDIA’s B200, H200, H100, and AMD’s MI300X to assess how well they scale for LLM inference. |
| SP021 | Intuition Labs | LLM Inference Hardware Enterprise Guide | NVIDIA’s ecosystem advantages (CUDA software stack, TensorRT, etc.) create high switching costs. |
| SP022 | Forbes | AI Inference Is King: Do You Know Which Chip Is Best? | Now for the real test: is the NVL72 as fast as Nvidia promised at launch? Yes, it is thirty times faster than the 8-GPU H200. |
| SP023 | GFM Review | The New Wave of AI Chips: Can Cerebras, d-Matrix, and Groq Take on Nvidia | The specialized nature of Cerebras, d-Matrix, and Groq’s technologies allows these companies to carve out niches where their solutions can outperform Nvidia’s more generalized approach. |
| SP024 | Silicon Analysts | NVIDIA AI Accelerator Market Share 2024-2026 | NVIDIA commands approximately 80-90% of the AI accelerator market by revenue as of 2025... priority TSMC CoWoS allocation maintain dominance. |
| SP025 | Research and Markets via GlobeNewswire / Yahoo Finance | AI inference - Company Evaluation Report, 2025 | The market faces constraints such as the high power requirements of AI chips... and concerns around data security and supply chain disruptions. |
| SP026 | NVIDIA | NVIDIA L40S GPU | Experience breakthrough multi-workload performance with the NVIDIA L40S GPU. Combining powerful AI compute with best-in-class graphics and media acceleration, the L40S GPU is built to power the next generation of data center workloads—from generative AI and large language model (LLM) inference and training to 3D graphics, rendering, and video. |
| SI001 | d-Matrix | d-Matrix Raises $275 Million to Power the Age of AI Inference | D-Matrix has closed $275 million in Series C funding, valuing the company at $2 billion and bringing the total raised to date to $450 million. |
| SI002 | d-Matrix | d-Matrix launches Corsair for AI inference without GPUs / HBM | Corsair offers performance of 60,000 tokens/second at 1 ms/token for Llama3 8B in a single server and 30,000 tokens/second at 2 ms/token for Llama3 70B in a single rack. |
| SI003 | d-Matrix | d-Matrix Unveils Corsair, the World's Most Efficient AI Computing Platform for Inference in Datacenters | Corsair is sampling to early-access customers and will be broadly available in Q2'2025. |
| SI004 | d-Matrix | d-Matrix Announces $110 Million in Series B Funding to Make Generative AI Commercially Viable | The goal of the fundraise is to enable d-Matrix to begin commercializing Corsair. |
| SI005 | d-Matrix | d-Matrix Corsair Redefines Performance and Efficiency for AI Inference at Scale | Performance, cost and power estimates are preliminary and subject to change. Results may vary. |
| SI006 | d-Matrix | JetStream Product Brief | JetStream Product Specification ... Maximum bandwidth 400 Gbps ... Max TDP (w/ transceivers) 150 W. |
| SI007 | Bullhound Capital | Bullhound Capital leads $275M investment into AI inference leader d-Matrix | D-Matrix ... has closed $275 million in Series C funding, valuing the company at $2 billion and bringing the total raised to date to $450 million. |
| SI008 | Qatar Investment Authority | QIA joins d-Matrix's Series C USD 275m funding round | Valued at USD 2 billion and bringing the total raised to date to USD 450 million, d-Matrix will use the new capital to advance their roadmap, accelerate global expansion and support multiple large-scale deployments. |
| SI009 | Data Center Dynamics | Chip startup d-Matrix raises $275m against $2bn valuation | Its single-card configuration provides up to 256GB of off-chip capacity memory and 2GB of performance memory at 150Tbps. |
| SI010 | SiliconANGLE | Chip startup d-Matrix raises $275M to speed up inference with in-memory compute | |
| SI011 | Bloomberg | AI Energy Crisis Boosts Need for More Efficient Chips | AI requires massive amounts of data ... a process that burns through lots of electricity. |
| SI012 | VentureBeat | New Microsoft partnership accelerates generative AI development | Project Bonsai is a platform ... and the early results are very encouraging. |
| SI013 | VentureBeat | Nvidia GPU shortage is top gossip of Silicon Valley | When people use the word GPU shortage, they're really talking about a shortage of ... some component on the board, not the GPU itself. |
| SI014 | CNN | AI chips supply chain bottlenecks | The shortage could force companies to find creative ways around the problem. |
| SI015 | CNBC | ChatGPT and generative AI are booming, but the costs can be extraordinary | At its peak in 2021, Latitude was spending nearly $200,000 a month on OpenAI's generative AI software and Amazon Web Services. |
| SI016 | TechTarget | Infrastructure to support OpenAI's ChatGPT could be costly | Users have to build up not just compute, but their networking and power management. |
| SI017 | The Washington Post | The hidden cost of ChatGPT is all the computing power | AI chatbots have a problem: They lose money on every chat. |
| SI018 | GigaIO | GigaIO Partners with d-Matrix to Deliver Ultra-Efficient Scale-Up AI Inference Platform | Our single-node server eliminates complex multi-node configurations and simplifies deployment, enabling enterprises to quickly adapt to evolving AI workloads while significantly improving their TCO and operational efficiency. |
| SI019 | Gimlet Labs | Low-Latency Inference with Speculative Decoding on d-Matrix Corsair and GPU | Compared to the same speculative decoder on GPU and equivalent energy consumption, we've found that the Corsair-based solution delivers 2-5X end-to-end request speedup ... and up to 10X ... for energy-optimized configurations. |
| SI020 | AIMultiple | AI Hardware | Cloud GPU list prices for the same model can differ several times over from one provider to another. |
| SI021 | d-Matrix | d-Matrix Ecosystem | Not actual customers. |
| SI022 | d-Matrix | d-Matrix Announces SquadRack | Combined with Supermicro's AI servers, Arista's ethernet switches, and Broadcom's PCIe and ethernet switch chips, we're delivering an AI inference rack that speeds up time to deployment. |
| SI023 | d-Matrix | d-Matrix Announces JetStream I/O Accelerators | JetStream NICs are full-height PCIe Gen5 cards delivering a maximum 400Gpbs bandwidth. Samples are available now, with full production expected by year-end. |
| SI024 | d-Matrix | d-Matrix Emerges from Stealth with Strong AI Performance and Efficiency | d-Matrix uses a hybrid approach to memory ... using SRAM as Performance Memory and a larger DRAM store for Capacity Memory. |
| SI025 | d-Matrix | d-Matrix and Alchip Announce Collaboration on World's First 3D DRAM Solution | The collaboration has already enabled a key technology, d-Matrix 3DIMC, that is featured on d-Matrix Pavehawk test silicon and has been successfully validated in d-Matrix's labs. |
| SI026 | d-Matrix | Gimlet Cloud to Deploy d-Matrix Corsair Alongside GPUs | The companies plan to make their combined solution available to select customers through Gimlet Cloud in 2H 2026. |
| SI027 | d-Matrix | Why We Created a 3D DRAM Solution to Advance Low-Latency AI Inference | Our next task is meeting the incredible demand required by emerging AI workloads with high user expectations, and that starts with Pavehawk. |
| SI028 | d-Matrix | How Speculative Decoding Supercharged AI Inference in Disaggregated Pipelines | Rather than reducing the batch size on a GPU and wasting compute to generate a low-latency experience, it reduces the total number of sequential inferences the powerful model handles. |
| SI029 | d-Matrix | Democratizing AI Through Hardware-Software Codesign for LLM Inference | The team looks at the associated software design of modern systems including collective communication algorithms and the distributed inference serving stack. |
| SI030 | NVIDIA via Stocklight | NVIDIA Corporation Annual Report 2026 | These Data Center systems are extreme co-designed with the GPU, CPU, NVLink switch, DPU, NIC, and scale-out networking along with software stacks and algorithms to deliver data center-scale computing solutions. |
| SI031 | GFM Review | The New Wave of AI Chips: Can Cerebras, d-Matrix, and Groq Take on Nvidia? | These new entrants face several hurdles, including the need for substantial capital investment to scale production, challenges in building brand recognition, and the difficulty of competing against Nvidia's established ecosystem. |
| SI032 | d-Matrix | Transforming AI: d-Matrix's Pivotal Moments in Pursuit of GenAI Inference at Scale | After coming within two weeks of running out of cash, we had an a-ha moment ... why don't we go ask for $40 million and show a big vision to match the size of the opportunity. |
| SE001 | d-Matrix | d-Matrix homepage | |
| SE002 | d-Matrix | Technology | |
| SE003 | d-Matrix | Product | |
| SE004 | d-Matrix | Ecosystem | |
| SE005 | d-Matrix | SquadRack announcement | Corsair delivers the compute-memory acceleration, while JetStream delivers I/O acceleration. |
| SE006 | d-Matrix | JetStream announcement | JetStream NICs are full-height PCIe Gen5 cards delivering a maximum 400Gpbs bandwidth. |
| SE007 | d-Matrix | Corsair launch announcement | Each Corsair card is powered by DIMC compute cores with 2400 TFLOPs of 8-bit peak compute, 2 GB of integrated Performance Memory, and up to 256 GB of off-chip Capacity Memory. |
| SE008 | d-Matrix | d-Matrix and Alchip announce collaboration on 3D DRAM solution | 3DIMC will commercially debut on the d-Matrix Raptor inference accelerator, the successor to d-Matrix Corsair. |
| SE009 | d-Matrix | d-Matrix Corsair Redefines Performance and Efficiency for AI Inference at Scale | d-Matrix Aviator is an enterprise-grade software stack co-designed with d-Matrix hardware. |
| SE010 | d-Matrix | JetStream product brief | Maximum bandwidth 400 Gbps |
| SE011 | d-Matrix | Going vertical: Why we created a 3D DRAM solution to advance low-latency AI inference | |
| SE012 | d-Matrix | The power of the middle lane: why a hybridized approach to memory gives the best of both worlds | |
| SE013 | d-Matrix | Scaling AI the right way: introducing our rack-level inference solution | |
| SE014 | d-Matrix | Why we needed a new transparent NIC solution | |
| SE015 | d-Matrix | Why we decoupled execution to accelerate I/O | |
| SE016 | d-Matrix | Why modern AI workloads demand a disaggregated approach | |
| SE017 | d-Matrix | How speculative decoding supercharged AI inference in disaggregated pipelines | |
| SE018 | d-Matrix | Batching just right: how interactive apps serve as a new battleground | |
| SE019 | d-Matrix | Impact of the DeepSeek moment on inference compute | |
| SE020 | d-Matrix | Keyformer: KV cache reduction through attention sparsification for efficient generative inference | |
| SE021 | d-Matrix | Introducing dmx.compressor | |
| SE022 | d-Matrix | Democratizing AI through hardware-software codesign for LLM inference | |
| SE023 | Data Center Dynamics | Chip startup d-Matrix raises $275m against $2bn valuation | |
| SE024 | EE News Europe | Second generation chiplet platform targets generative AI | |
| SE025 | HPCwire | d-Matrix launches new chiplet connectivity platform to address growing compute demand for generative AI | |
| SE026 | GigaIO | GigaIO partners with d-Matrix to deliver ultra-efficient scale-up AI inference platform | |
| SE027 | Gimlet Labs | Low-Latency Inference with Speculative Decoding on d-Matrix Corsair and GPU | Compared to the same speculative decoder on GPU and equivalent energy consumption, we've found that the Corsair-based solution delivers 2-5X end-to-end request speedup. |
| SE028 | NVIDIA | NVIDIA H100 | |
| SE029 | NVIDIA | NVIDIA H200 | |
| SE030 | NVIDIA | NVIDIA GB200 NVL72 | |
| SE031 | AIMultiple | AI Hardware | |
| SE032 | Intuition Labs | Private LLM Inference: Key Hardware and Integrators for Enterprise | |
| SE033 | Forbes | AI Inference Is King: Do You Know Which Chip Is Best? | |
| SE034 | Silicon Analysts | NVIDIA AI accelerator market share 2024-2026 | |
| SE035 | d-Matrix | GigaIO partners with d-Matrix to deliver ultra-efficient scale-up AI inference platform | |
| SE036 | d-Matrix | AI is a context problem | |
| SE037 | d-Matrix | The fight for latency: why agents have changed the game | |
| SE038 | d-Matrix | Open standards are the path to the next AI breakthrough | |
| SE039 | d-Matrix | Why optimizing every layer of AI workloads from software to infrastructure is now critical as apps take off | |
| SE040 | Alchip | d-Matrix and Alchip announce collaboration on world's first 3D DRAM solution to supercharge AI inference | |
| SE041 | d-Matrix | The complete recipe to unlock AI reasoning at enterprise scale | |
| SU001 | d-Matrix | AI Computing Solutions for Cloud & Enterprise Markets | *Not actual customers. |
| SU002 | d-Matrix | d-Matrix Raises $275 Million to Power the Age of AI Inference | |
| SU003 | d-Matrix | SquadRack | |
| SU004 | d-Matrix | JetStream | |
| SU005 | d-Matrix | d-Matrix launches Corsair for AI inference without GPUs, HBM | |
| SU006 | d-Matrix | d-Matrix Unveils Corsair, the World's Most Efficient AI Computing Platform for Inference in Datacenters | |
| SU007 | d-Matrix | GigaIO Partners with d-Matrix to Deliver Ultra-Efficient Scale-Up AI Inference Platform | |
| SU008 | d-Matrix | Gimlet Cloud, built for running agentic AI inference, to deploy d-Matrix Corsair low latency, memory-optimized accelerators alongside GPUs | |
| SU009 | d-Matrix | Scaling AI the Right Way: Introducing Our Rack-Level Inference Solution | |
| SU010 | d-Matrix | Why We Needed a New Transparent NIC Solution | |
| SU011 | d-Matrix | How Speculative Decoding Supercharged AI Inference in Disaggregated Pipelines | |
| SU012 | d-Matrix | Impact of the DeepSeek Moment on Inference Compute | |
| SU013 | d-Matrix | Transforming AI: d-Matrix's Pivotal Moments in Pursuit of Gen AI Inference at Scale | |
| SU014 | Bullhound Capital | Bullhound Capital leads $275M investment into AI inference leader d-Matrix | |
| SU015 | QIA | QIA joins d-Matrix's Series C USD 275m funding round | |
| SU016 | Data Center Dynamics | Chip startup d-Matrix raises $275m against $2bn valuation | |
| SU017 | SiliconANGLE | Chip startup d-Matrix raises $275M to speed inference with memory-centric compute | |
| SU018 | TechStartups | AI chip startup d-Matrix raises $275M at $2B valuation to challenge Nvidia's grip on AI inference | |
| SU019 | Converge Digest | d-Matrix Raises $275 Million to Accelerate the Age of AI Inference | |
| SU020 | GigaIO | GigaIO Partners with d-Matrix to Deliver Ultra-Efficient Scale-Up AI Inference Platform | |
| SU021 | Gimlet Labs | Low-Latency Inference with Speculative Decoding on d-Matrix Corsair and GPU | |
| SU022 | AIMultiple | AI hardware | |
| SU023 | IntuitionLabs | Private LLM Inference: Key Hardware and Integrators for Enterprise | |
| SU024 | Forbes | AI Inference Is King. Do You Know Which Chip Is Best? | |
| SU025 | GFM Review | The New Wave of AI Chips: Can Cerebras, d-Matrix and Groq Take on Nvidia? | |
| SU026 | d-Matrix | d-Matrix and Gimlet Join Forces in Strategic Partnership | |
| SU027 | Data Center Knowledge | Gimlet Labs, d-Matrix Partner to Boost Agentic AI Inference | |
| SU028 | eeNews Europe | d-Matrix launches Corsair for AI inference without GPUs, HBM | |
| SU029 | AIwire | GigaIO Partners with d-Matrix to Deliver Ultra-Efficient Scale-Up AI Inference Platform | |
| SU030 | PR Newswire | d-Matrix Announces SquadRack, Industry's First Rack-Scale Solution Purpose-Built for AI Inference at Datacenter Scale | |
| SU031 | d-Matrix | Request Access or Buy | |
| SU032 | d-Matrix | The fight for latency: why agents have changed the game | |
| SU033 | AMD | AMD Instinct MI350 Series GPUs | AMD Instinct™ MI350 Series GPUs help enable frictionless adoption with drop-in compatibility, while the AMD GPU Operator simplifies deployment and workload configuration in Kubernetes. |
| SU034 | Groq | GroqCloud | The AI inference platform built for developers. Fast responses, scalable performance, and costs you can plan for. Available in public, private, or co-cloud instances. |
| SU035 | Cerebras | Chip | The world’s largest and most powerful processor for AI training and inference. |
| SR001 | d-Matrix | SquadRack announcement | |
| SR002 | d-Matrix | JetStream announcement | |
| SR003 | d-Matrix | d-Matrix and Alchip announce collaboration on 3D DRAM solution | |
| SR004 | d-Matrix | Going vertical: Why we created a 3D DRAM solution to advance low-latency AI inference | |
| SR005 | d-Matrix | Why We Needed a New Transparent NIC Solution | |
| SR006 | d-Matrix | Why We Decoupled Execution to Accelerate I/O | |
| SR007 | d-Matrix | Why Modern AI Workloads Demand a Disaggregated Approach | |
| SR008 | d-Matrix | Impact of the DeepSeek Moment on Inference Compute | |
| SR009 | Data Center Dynamics | Chip startup d-Matrix raises $275m against $2bn valuation | |
| SR010 | SiliconANGLE | Chip startup d-Matrix raises $275M to speed up inference with in-memory compute | |
| SR011 | Bloomberg | AI Energy Crisis Boosts Need for More Efficient Chips | |
| SR012 | eeNews Europe | Second-generation chiplet platform targets generative AI | |
| SR013 | VentureBeat | Nvidia GPU shortage is top gossip of Silicon Valley | |
| SR014 | CNN | AI chips supply chain bottlenecks | |
| SR015 | TechTarget | Infrastructure to support OpenAI’s ChatGPT could be costly | |
| SR016 | The Washington Post | ChatGPT’s hidden cost is GPU compute | |
| SR017 | NVIDIA | NVIDIA H200 | |
| SR018 | NVIDIA | NVIDIA GB200 NVL72 | |
| SR019 | AMD | AMD Instinct MI300X Accelerators | |
| SR020 | Groq | Groq Products | |
| SR021 | Cerebras | Cerebras Product Chip | |
| SR022 | Intuition Labs | LLM Inference Hardware Enterprise Guide | |
| SR023 | Silicon Analysts | NVIDIA AI Accelerator Market Share 2024-2026 | |
| SR024 | d-Matrix | GigaIO partners with d-Matrix to deliver ultra-efficient scale-up AI inference platform | |
| SR025 | d-Matrix | d-Matrix and Gimlet Labs accelerate agentic AI inference | |
| SR026 | d-Matrix | Transforming AI: d-Matrix’s pivotal moments in pursuit of Gen AI inference at scale | |
| SR027 | d-Matrix | Privacy Policy - d-Matrix | |
| SR028 | d-Matrix | Terms of Use - d-Matrix | |
| SR029 | Bureau of Industry and Security | Commerce Strengthens Restrictions on Advanced Computing Semiconductors, Semiconductor Manufacturing Equipment, and Supercomputing Items to Countries of Concern | |
| SR030 | d-Matrix | Join d-Matrix | Careers in AI Innovation and Technology | |
| SR031 | d-Matrix | d-Matrix and Gimlet Join Forces in Strategic Partnership - d-Matrix | |
| SR032 | Reuters | AI chip startup d-Matrix raises $110 mln with backing from Microsoft | |
| SR033 | Supermicro | Supermicro expands AI solutions and adds d-Matrix inference and rack-scale systems | |
| SR034 | Alchip | d-Matrix and Alchip announce collaboration on world’s first 3D DRAM solution to supercharge AI inference | |
| SR035 | NVIDIA | NVIDIA AI Enterprise | |
| SR036 | d-Matrix | Using what’s on hand: spare data center space is an untapped gold mine - d-Matrix | |
| SV001 | d-Matrix | d-Matrix raises $275 million to power the age of AI inference | d-Matrix has closed $275 million in Series C funding, valuing the company at $2 billion and bringing the total raised to date to $450 million. |
| SV002 | Bullhound Capital | Bullhound Capital leads $275M investment into AI inference leader d-Matrix | |
| SV003 | Qatar Investment Authority | QIA joins d-Matrix’s Series C USD 275m funding round | |
| SV004 | Data Center Dynamics | Chip startup d-Matrix raises $275m against $2bn valuation | It brings the total raised by d-Matrix to $450 million, which the company said it will use to advance its roadmap, accelerate global expansion, and support multiple large-scale deployments of its data center inference platform. |
| SV005 | SiliconANGLE | Chip startup d-Matrix raises $275M to speed up inference with in-memory compute | |
| SV006 | TechStartups | AI chip startup d-Matrix raises $275M at $2B valuation to challenge Nvidia’s grip on AI inference | |
| SV007 | Converge Digest | d-Matrix raises $275 million to accelerate the age of AI inference | |
| SV008 | d-Matrix | d-Matrix announces $110 million in Series B funding | |
| SV009 | d-Matrix | d-Matrix emerges from stealth with strong AI performance and efficiency | |
| SV010 | d-Matrix | Transforming AI: d-Matrix’s pivotal moments in pursuit of GenAI inference at scale | |
| SV011 | d-Matrix | GigaIO partners with d-Matrix to deliver ultra-efficient scale-up AI inference platform | |
| SV012 | d-Matrix | Gimlet Cloud to deploy d-Matrix Corsair accelerators alongside GPUs | |
| SV013 | GigaIO | GigaIO partners with d-Matrix to deliver ultra-efficient scale-up AI inference platform | |
| SV014 | Gimlet Labs | Low-Latency Inference with Speculative Decoding on d-Matrix Corsair and GPU | By offloading the speculative decoder to d-Matrix Corsair, we achieved 2-10X interactivity improvements over a homogeneous speculative decoding setup for the same energy efficiency. |
| SV015 | Bloomberg | AI Energy Crisis Boosts Need for More Efficient Chips | |
| SV016 | CRN | 10 Hottest Semiconductor Startups of 2023 | |
| SV017 | Silicon Republic | Start-ups to Watch in 2024 | |
| SV018 | NVIDIA | NVIDIA H100 Tensor Core GPU | |
| SV019 | NVIDIA | NVIDIA H200 Tensor Core GPU | |
| SV020 | NVIDIA | NVIDIA GB200 NVL72 | |
| SV021 | AMD | AMD Instinct MI300X | |
| SV022 | Groq | Groq products | |
| SV023 | Cerebras | The Future of AI is Wafer Scale | |
| SV024 | AIMultiple | Top 25+ AI Chip Makers: NVIDIA & Its Competitors | |
| SV025 | Forbes | AI Inference Is King: Do You Know Which Chip Is Best? | |
| SV026 | Silicon Analysts | NVIDIA AI accelerator market share 2024-2026 | |
| SV027 | Intuition Labs | Private LLM Inference: Key Hardware and Integrators for Enterprise | |
| SV028 | Yahoo Finance / GlobeNewswire | AI inference - Company Evaluation Report, 2025 | However, concerns around data security and supply chain disruptions remain persistent challenges for companies operating in the AI inference space. |
| SV029 | Securities and Exchange Commission | NVIDIA Corporation Form 10-K for fiscal year ended January 25, 2026 | |
| SV030 | Securities and Exchange Commission | Advanced Micro Devices, Inc. Form 10-K for fiscal year ended December 27, 2025 | |
| SV031 | Securities and Exchange Commission | Advanced Micro Devices, Inc. Form 10-Q for quarter ended March 28, 2026 | |
| SV032 | Securities and Exchange Commission | NVIDIA Corporation Form 10-Q for quarter ended April 26, 2026 | |
| SV033 | Securities and Exchange Commission | NVIDIA SEC submissions JSON | |
| SV034 | Securities and Exchange Commission | AMD SEC submissions JSON | |
| SV035 | Securities and Exchange Commission | NVIDIA SEC companyfacts JSON | |
| SV036 | Securities and Exchange Commission | AMD SEC companyfacts JSON |