Velaura AI
量产验证过的能效 IP,估值 $1B+,但 AI 芯片客户一个未点名,收入也未披露
Velaura AI 有真正经过量产验证的高能效硅平台,也有可信董事会;但其 $1B+ 估值目前建立在没有任何具名 AI 硅客户、也未披露收入或 royalty rate 的基础上。
封面要素
公司概况
Velaura AI, Inc. 是一家硅谷半导体公司(前身为美国比特币挖矿 ASIC 厂商 Auradine)。公司在 March 2026 转向授权超低功耗硅设计 IP,品牌为 Titan Core,面向超大规模数据中心和 Physical AI(机器人、无人机、自主机器)所用的 AI 加速器。August 18, 2026,公司完成由 Seligman Ventures 领投的 $110 million Series A,估值超过 $1 billion,成为独角兽;底层低功耗设计技术已在历史比特币挖矿硬件业务的 30 million 多颗量产 ASIC 中验证。
- 创始人
- Rajiv Khemani, Manu Gulati
- 创立地点
- Santa Clara, California, USA
- 总部
- Santa Clara, California, USA
- 产品
- Titan Core 是一套硅设计与 IP 平台,授权客户接入其自有 AI 加速器 RTL 设计,据称可将每瓦性能提升 2-4x;另有历史 Teraflux 风冷 / 浸没式 / 水冷比特币挖矿 ASIC 硬件。
- 客户
- 超大规模云数据中心运营商、Physical/具身 AI(机器人、无人机、自主机器)OEM、边缘 AI 设备厂商;历史客户群是工业比特币挖矿运营商。
- 商业模式
- 前置 IP 授权费,加上与客户实测节电量挂钩的版税(类似 Arm 的授权模式);同时保留 Teraflux 挖矿系统的历史直接硬件销售。
- 阶段
- Series A (post-AI-pivot); unicorn valuation
- 融资情况
- $110M Series A 于 August 18, 2026 完成,估值超过 $1 billion;2023-2026 四轮累计披露融资约 $424 million。
执行摘要
主要优势
- 超低功耗硅设计方法已在 30M+ 出货 ASIC 上通过量产验证;在早期 AI 硅竞争者中,这种规模证据很少见。
- 管理层和董事会履历可信且经验充足:团队有 Apple、NVIDIA、Google、Qualcomm、Marvell 背景,Intel CEO Lip-Bu Tan 与 MARA CEO Fred Thiel 也在董事会。
- 宏观顺风明确:AI 数据中心电力需求到 2030 年几乎翻倍,美国存在 9.3 GW 电力缺口,直接抬高 performance-per-watt 技术的价值。
- MARA Holdings 在 SEC 10-Q 中确认公允价值收益,为估值上调提供独立且经审计的佐证,不只是依赖公司新闻稿。
主要风险
- 支撑 $1B+ 估值的 Titan Core AI 硅业务没有任何具名客户;所有牵引力说法都可追溯到公司表述,再由媒体转述。
- 公司任何业务线都未披露收入、royalty rate、毛利率、现金余额、烧钱速度或现金跑道,估值无法与可比公司校准。
- 商业 EDA 工具带来商品化风险:Cadence Cerebrus、Synopsys DSO.ai 可能让超大规模云厂商在内部复刻 Velaura 的效率提升。
- Velaura 证据最充分的客户关系 MARA 同时也是关联方投资人;其可触达的传统硬件市场(Bitcoin 矿机)又在行业转向 AI/HPC colocation 时结构性收缩。
- AI 硅可比公司在一年内波动极大:Groq 估值约下调 50%,Untether AI 破产,Cerebras 计划约 $56B IPO;今天的估值标记未必能守住。
未决问题
- Velaura AI 硅业务的实际收入、royalty rate 与 royalty 计量方法。
- 至少一个具名且经独立确认的 Titan Core 超大规模云厂商、机器人或边缘设备客户。
- 2026 年 8 月 Series A 后的当前现金余额、月度烧钱速度与现金跑道。
- 可调和的员工数(公开估计从 62 人到 100-150 人不等)。
- 对所称 2-4x MATMUL performance-per-watt 提升的独立同行评议基准。
- 股权结构表细节:清算优先权堆叠、反稀释条款与期权池规模。
目录
01公司概况
1.1 身份、产品与商业模式
Velaura AI, Inc. 是一家总部位于硅谷的半导体公司,为 AI 计算基础设施开发超低功耗硅和软件技术。公司此前名为 Auradine,是美国比特币挖矿 ASIC 厂商;March 2026 更名为 Velaura AI,作为转向 AI 计算的战略转型的一部分。Velaura 的旗舰产品 Titan Core 不是独立芯片,而是一套硅设计和 IP 平台:客户提供自有 RTL 处理器设计,Velaura 则套用自研超高能效电路和物理设计库,在 3nm、2nm 等先进工艺节点交付优化后的、可直接接入的物理版图。公司称,Titan Core 可让 AI 训练和推理中最耗能的矩阵乘法运算每瓦性能提升 2-4x,折算到典型 1000W 级 GPU 或 XPU 上,最多可节省 500W。Velaura 通过前置授权费加按实际节电量计提的版税变现这套 IP;CEO Rajiv Khemani 明确把这种结构类比为 Arm 的处理器授权模式。底层低功耗设计方法已经在领先半导体节点的 30 million 多颗量产 ASIC 中得到商业规模验证,给 Velaura 提供了大多数首次进入 AI 芯片领域的公司没有的真实可靠性和良率记录。公司瞄准三个相邻市场:希望降低电力和冷却成本的超大规模 AI 数据中心、受严格功耗和热预算约束的 Physical 与具身 AI 系统(机器人、无人机、自主机器),以及需要高能效本地处理的边缘 AI 部署。[CO001, CO002, CO003, CO004, CO005, CO006]
| 指标 | 数值 / 状态 | 截至 | 置信度 | 缺口 |
|---|---|---|---|---|
| 最新估值 | $1.0B+(Series A 投后) | 2026-08-18 | 高 | |
| 累计披露融资额(所有轮次) | $424M+ | 2026-08-18 | 中 | 精确累计金额未获独立审计 |
| 最新轮次规模 | $110M Series A | 2026-08-18 | 高 | |
| 已部署芯片(累计 ASIC) | 30,000,000+ | 2026-08-18 | 高 | 公司披露数字;无法获得独立计数 |
| 已披露超大规模云厂商接触 | 四大云服务商中的 3 家(未具名) | 2026-08-18 | 中 | 客户身份未披露 |
| 员工数 | 2026-08-19 | 低 | 未找到公开员工数;尽调路径:通过管理层访谈或 LinkedIn 员工数三角验证获取 | |
| 年经常性收入 / 运行率 | 2026-08-19 | 低 | 私营公司;未披露收入数字 | |
| 总部 | 美国加利福尼亚州硅谷 | 2026-08-18 | 高 |
封面指标混合了公司披露(估值、融资、部署规模)与未披露私营指标(员工数、收入)的显式 null;置信度和缺口列标出仍未验证的部分。
[CO001, CO007, CO008, CO006, CO017, CO024]Velaura AI 的身份、产品、客户、资本与关键依赖如何连在一起。
[CO003, CO016, CO017, CO008, CO019]1.2 领导层、创始人与治理
联合创始人兼 CEO Rajiv Khemani,是本章审阅的所有融资和产品公告中 Velaura AI 的公开面孔;Mayfield 的 Navin Chaddha 称,2026 轮融资是其公司与 Khemani 的第四次合作。Manu Gulati 担任首席开发官,并与 Khemani 一起被点名为领导团队的技术核心;Mayfield 表示这是其第二次与 Gulati 合作。除创始二人外,Velaura 的团队页面列出 Sanjay Gupta(战略与 GTM 总裁)、YJ Kim(产品总裁)、Tim Vehling(产品副总裁)、Brian Campbell(工程运营副总裁)和 Atul Dhablania(首席系统运营官),组成一支公司和媒体都称由 Apple、NVIDIA、Google、Qualcomm、Marvell 校友构成的管理梯队。治理上,董事会混合了风险资本和战略利益:Navin Chaddha(Mayfield)、Umesh Padval(Seligman Ventures,持有 Series A 领投方董事席位)、Dipender Saluja(Capricorn Investment Group)、Fred Thiel(MARA CEO,既是投资人,也是前 Teraflux 硬件客户)、Sriram Viswanathan(Celesta Capital),以及 Intel CEO Lip-Bu Tan;后者通过 Walden International 进入董事会,为公司背书了重要的半导体行业信誉。两位学术 / 技术顾问 Aditya Grover(Inception AI CTO;UCLA 教授)和 Magnus Egerstedt(UNC Chapel Hill 教务长;机器人学者)补强了公司的 Physical AI 与机器学习能力。公开评论高度集中在 Khemani 和 Gulati 身上,而其余高管几乎没有独立传记细节,这显示公司对创始二人存在实质性的关键人依赖。[CO012, CO013, CO014, CO015, CO025, CO029]
| 人物 | 角色 | 过往背景 | 创始人-市场匹配 / 职能覆盖 | 关键人物依赖 |
|---|---|---|---|---|
| Rajiv Khemani | 联合创始人兼 CEO | 曾任半导体与系统领域高管 | 公司叙事与战略;推动 Auradine 转向 Velaura | 高——所有融资 / 产品公告中唯一公开 CEO 声音 |
| Manu Gulati | 首席开发官 | 据 Mayfield 评论,曾在 Apple、Google 及 Broadcom 级别公司担任芯片架构角色 | 硅架构领导;分别与 Khemani / Gulati 建立第四次和第二次 Mayfield 合作 | 高——与 Khemani 一同被点名为核心技术二人组 |
| Sanjay Gupta | 战略与商业化总裁 | 战略和商业化领导 | 版税模式的商业战略与授权商业化 | 中 |
| YJ Kim | 产品总裁 | 产品领导 | 负责 Titan Core 产品路线图 | 中 |
| Tim Vehling | 产品副总裁 | 产品管理 | 支持产品战略执行 | 低-中 |
| Brian Campbell | 工程运营副总裁 | 工程运营 | 制造 / 工程运营规模化 | 低-中 |
| Atul Dhablania | 首席系统运营官 | 系统运营 | 系统级运营监督 | 低-中 |
| Navin Chaddha | 董事会成员;Mayfield 董事总经理 | 风险投资人,曾多次与 Khemani 合作 | 治理与资本市场关系 | 中——投资人董事,而非运营高管 |
汇总自 Velaura AI 官方团队页面;在独立履历细节未获另行印证处,背景 / 职能匹配表述改写自公司和投资人博客描述。
[CO012, CO013, CO014, CO035, CO029]| 利益相关方 | 角色 | 控制权 / 经济重要性 | 尽调问题 |
|---|---|---|---|
| Seligman Ventures | Series A 领投方 | 设定领投条款;Umesh Padval 持有董事会席位 | 确认董事会席位条款及任何信息权 / 否决权 |
| Capricorn Investment Group | 新 Series A 投资方 | 有意义的少数股权;Dipender Saluja 持有董事会席位 | 确认持股比例和共同投资历史 |
| Prosperity7 Ventures | 新 Series A 投资方 | 少数股权(规模未披露) | 确认分配规模及任何战略权利(与 Aramco 相邻的 LP 出资人基础) |
| Mayfield | 横跨 Series A/B/C 和 2026 轮次的重复投资方 | Navin Chaddha 董事会席位;第四次 Khemani 合作 | 确认所有轮次累计持股 |
| Maverick Silicon | 重复投资方(Series C 和 2026 轮) | 少数股权 | 确认行业专属董事会 / 顾问权利 |
| MARA (Marathon Digital Holdings) | 重复投资方和前 Teraflux 硬件客户;Fred Thiel 董事会席位 | 投资方 / 客户双重关系;若传统挖矿收入重要,则存在集中度风险 | 评估剩余 Teraflux 商业依赖与 AI 转型的关系 |
| Premji Invest | 重复投资方(Series C 和 2026 轮) | 少数股权 | 确认分配规模 |
| Samsung Catalyst Fund | 重复投资方(Series C 和 2026 轮) | 少数股权;潜在战略代工 / 客户关系 | 评估与 Samsung Foundry 制程节点的任何商业联系 |
| StepStone Group | Series C 领投方;2026 轮重复参投 | 大型私募市场配置方;跨轮次资本承诺规模可观 | 确认总投资资本和任何清算优先权条款 |
私营公司未公开披露持股比例;经济重要性表述为定性判断,来自已公告的领投 / 重复投资方角色和董事会席位。
[CO009, CO010, CO011, CO015, CO041]1.3 融资历史、估值与投资人
Velaura AI 的资本化历史横跨两个公司身份下的四轮已披露机构融资。作为 Auradine,公司在 May 2023 完成 $81 million Series A(Celesta Capital 与 Mayfield 领投),用于推出 Teraflux 比特币挖矿 ASIC 产品线;April 2024 完成超额认购的 $80 million Series B,同时产生 $80 million 客户订单;April 2025 完成由 StepStone Group 领投的 $153 million Series C,Maverick Silicon、Premji Invest、Samsung Catalyst Fund、Qualcomm Ventures、Mayfield 和 MARA 参与,明确用于从比特币挖矿扩展到区块链和 AI 基础设施。August 18, 2026,刚更名的 Velaura AI 宣布完成由 Seligman Ventures 领投的 $110 million Series A,新投资人 Capricorn Investment Group 和 Prosperity7 Ventures 加入,Mayfield、Maverick Silicon、MARA、Premji Invest、Samsung Catalyst Fund 和 StepStone Group 等老股东继续跟投;该轮把 Velaura 估值推至 $1 billion 以上,使其成为独角兽。四轮累计披露融资超过 $424 million,但没有公开来源审计或调和过这个总额。MARA 既是重复财务投资人,又曾是 Teraflux 挖矿硬件客户,这种双重身份是治理和独立性上需要单独尽调的事项。Seligman Ventures 本身是 Seligman Investments 的新设 $500 million 风险投资载体(成立于 November 2025),Managing Partner Umesh Padval 称,本次 Series A 是该机构首笔披露的 Physical AI 投资。[CO007, CO008, CO009, CO010, CO011, CO021]
| 日期 | 事件 | 类型 | 金额 / 估值 / 状态 | 参与方 | 含义 |
|---|---|---|---|---|---|
| 2023-05 | Auradine Series A 融资 | 融资 | $81M | Celesta Capital、Mayfield(领投) | 为 Teraflux 比特币挖矿 ASIC 产品线提供初始资本 |
| 2024-04 | Auradine Series B 融资 | 融资 | $80M,超额认购,$80M 订单 | 投资方:StepStone Group、Top Tier Capital Partners、MVP Ventures、Maverick Capital、Celesta Capital、Mayfield、MARA | 在比特币减半需求前扩大 Teraflux 生产 |
| 2025-04-16 | Auradine Series C 融资 | 融资 | $153M | StepStone Group(领投)、Maverick Silicon、Premji Invest、Samsung Catalyst Fund、Qualcomm Ventures、Mayfield、MARA | 桥接资本支持从比特币挖矿扩展到区块链 / AI 基础设施 |
| 2026-03-24 | Titan Core 硅设计和 IP 平台发布 | 产品 | 声称单位功耗性能提升 2-4x | Velaura AI(前 Auradine) | 首次公开披露以 AI 为重点的产品战略 |
| 2026-03-25 | Auradine 更名为 Velaura AI | 治理 | 公司更名;Teraflux 库存转向自营挖矿 | Velaura AI | 标志着公司正式从比特币挖矿硬件供应商转向 AI 计算基础设施公司 |
| 2026-08-18 | Velaura AI Series A 融资(AI 时代) | 融资 | $110M;估值超过 $1B | Seligman Ventures(领投)、Capricorn Investment Group、Prosperity7 Ventures、Mayfield、Maverick Silicon、MARA、Premji Invest、Samsung Catalyst Fund、StepStone Group | 独角兽估值里程碑;资本用于工程和商业规模化 |
| 2026-08-18 | 披露超大规模云厂商接触 | 规模 | 据报道已接触四大云服务商中的 3 家(未具名) | 未披露的超大规模云厂商 | 显示商业牵引力,但客户集中度仍无法验证 |
| 2026-08-18 | 30M+ ASIC 部署里程碑被重申 | 规模 | 30,000,000+ 台在生产中 | Velaura AI | 用此前挖矿芯片产量支撑 Titan Core 的制造良率和可靠性主张 |
| 2026(进行中) | Intel CEO Lip-Bu Tan 加入董事会(经 Walden International) | 治理 | 董事会席位 | Lip-Bu Tan、Walden International | 增加高知名度半导体行业治理可信度 |
| 2025(更名前) | MARA 作为投资方和 Teraflux 客户,加深合作 | 合作 | 参与 Series B/C 和 2026 轮投资 | MARA、Fred Thiel | 形成投资方-客户双重关系,关系到独立性和集中度分析 |
这是 Velaura AI/Auradine 唯一记录时间线;2026 年董事会新增成员和更名前 MARA 关系加深的日期,在精确日期未独立披露处为近似(仅月 / 年)。
[CO021, CO022, CO023, CO002, CO007, CO008]从 Auradine 2023 年 Series A 到 Velaura AI 2026 年独角兽 Series A 的有日期里程碑。
[CO021, CO022, CO023, CO003, CO002, CO007]1.4 牵引力、部署规模与反向信号
Velaura AI 最核心的牵引力主张,是其超低功耗设计方法已经过规模验证:公司在 Titan Core 发布稿和 Series A 公告中都引用了领先工艺节点上 30 million 多颗 ASIC 的制造量,并用这项规模来支撑新 AI 产品的制造良率和可靠性主张。客户侧,公司称正在与四大云厂商中的三家推进合作;据 Heise Online 报道,CEO Rajiv Khemani 向 Reuters 确认了这一点,但他拒绝披露涉及哪些超大规模云厂商,也没有任何客户独立确认与 Velaura 或 Titan Core 存在关系。没有具名客户验证,再叠加 2-4x 每瓦性能提升缺乏任何独立、同行评审基准,这就是本章识别出的最关键证据缺口。行业媒体对底层电力约束逻辑总体正面,但也提示真实执行风险:TechBooky 明确提醒,「估值超过 $1 billion 并不保证商业采用」,并指出芯片初创公司需要深厚工程能力、制造伙伴和客户验证,才能把设计定点转成经常性版税收入。Moor Insights & Strategy 的独立分析师 Patrick Moorhead 给出了更建设性的判断,称 Velaura 的路径「有潜力」降低客户总拥有成本并缓解热约束;这是带保留的背书,不是独立验证。[CO006, CO017, CO018, CO019, CO026, CO033]
1.5 封面指标与尽调缺口
后续章节会依赖的几项封面指标,截至运行日期仍未由公司披露,公开来源也查不到:当前员工数、AI 计算业务的收入或收入运行率(有别于历史 Teraflux 订单),以及具名客户名单。本章把这些指标明确记为空值而非猜测,并给出具体尽调路径——例如用职业社交网络员工数交叉推算员工规模,或在管理层访谈中直接要求披露;在估值工作采用 2-4x 主张之前,要求提供经审计或第三方验证的节电基准。本章证据台账中尚未解决或仅部分完成的研究问题(客户名称、员工数、AI 专属收入、独立基准验证)应视为开放事项,后续章节(财务、客户、风险、估值)必须用新证据关闭,或作为具名缺口继续保留。[CO038, CO019, CO034]
主要成熟度和牵引力指标,以及关键未解决缺口。
[CO008, CO024, CO006, CO034, CO019]1.6 图表
02市场分析
2.1 市场边界与定义
Velaura AI 不销售可直接上市的芯片;它授权的是超低功耗硅设计 IP,由其他公司集成进自有 AI 加速器。因此,相关市场边界不只是笼统的「AI 芯片市场」,而是那些可能引入第三方能效 IP 厂商的 AI 芯片设计决策子集:按照 Velaura 自己的市场口径,包括超大规模数据中心加速器、Physical 与具身 AI 系统(机器人、无人机、自主机器)以及边缘 AI 设备。数据中心 AI 芯片市场本身,一些发布方用窄口径跟踪(SemiconductorInsight 估计 2026 为 $13.8 billion,到 2034 达到 $45.2 billion);另一些把商用 GPU、定制 ASIC、存储和网络芯片纳入后,口径要宽得多(Deloitte 与 Gartner 均把 2026 整体 AI 半导体市场放在 $300-500 billion,而 IDC 较窄的「intelligent datacenter」分部为 $281 billion,包含在 $477.1 billion 数据中心半导体总量内)。仅 NVIDIA 一家就在 NVIDIA、AMD、Intel 之间拿下 Q1 2026 商用数据中心 AI 芯片收入约 87.4%,说明这个标题市场的大部分价值目前由 GPU 厂商捕获,而不是 Velaura 这样的 IP 授权方。明确不属于 Velaura 可服务市场的,是通用 CPU / 服务器芯片、非 AI 网络芯片,以及超大规模云厂商设计自有定制 ASIC(Google TPU、AWS Trainium、Microsoft Maia、Meta MTIA)时内部捕获的成品芯片收入——不过这些内部项目本身又可能成为 Velaura 底层能效技术的 IP 授权客户,而不是它的直接替代品。[CM025, CM001, CM002, CM004, CM019]
| 分部 / 类别 | 纳入支出 | 排除支出 | 买方 / 付款方 | 与 Velaura 的相关性 |
|---|---|---|---|---|
| 数据中心 AI 芯片市场 | 用于超大规模 / 企业数据中心训练和推理的 AI 加速器 / GPU / ASIC 硅采购 | 通用 CPU、仅联网硅、非 AI 服务器 | 超大规模云厂商基础设施 / 资本开支组织 | 主要授权目标:Titan Core IP 集成进这些加速器 |
| 物理 / 具身 AI 硬件市场 | 用于机器人、无人机和自主机器的计算、传感器和执行器硅 | 纯软件机器人平台、非 AI 工业自动化 | 机器人 / 设备 OEM 工程和采购 | 面向受电池寿命约束的具身 AI 的次级授权目标 |
| 边缘 AI 设备市场 | 用于智能手机、可穿戴设备、IoT、智能摄像头的本地推理芯片 | 仅云端推理工作负载 | 消费电子和 IoT 设备 OEM | 按 Velaura 自身三分部框架,为第三级目标 |
| 整体 AI 半导体市场 | 所有 AI 相关芯片收入,包括商用 GPU、定制 ASIC、存储、联网硅 | 非 AI 半导体收入 | 混合(超大规模云厂商、OEM、企业) | 最宽参考框架;Velaura 的 IP 授权模式不能直接覆盖 |
| 超大规模云厂商 AI 基础设施资本开支 | 数据中心建设、电力、冷却和硬件采购预算 | 与 AI 无关的企业运营支出、非数据中心资本开支 | 超大规模云厂商财务 / 基础设施领导层 | 最终为嵌入 Velaura IP 的芯片付款的资金来源 |
分部来自 Velaura 自身市场框架(数据中心、物理 / 具身 AI、边缘 AI),另加两个相邻参考市场(整体 AI 半导体、超大规模云厂商资本开支)用于圈定分析边界;边界由分析师定义,不同发布方口径会变化。
[CM025, CM001, CM005, CM012]2.2 TAM/SAM/SOM 与互相矛盾的测算口径
对 Velaura 特定的超低功耗 AI 芯片 IP 授权细分市场,没有一个受证据约束的单一 TAM/SAM/SOM 数字;所有公开数字都描述相邻且更宽的类别。数据中心侧,即使不考虑口径差异,三套独立的 2026 方法论也已经相差近 2x:IDC 分部层面的 $281 billion「intelligent datacenter」数字、由 Gartner $1.3 trillion 半导体总预测隐含出的约 $390 billion AI 份额,以及 Deloitte 修订后的约 $500 billion 全年估计。Physical AI 侧——这关系到 Velaura 的机器人 / 具身 AI 野心——三家发布方的 2026 估计跨度超过 7x(MarketsandMarkets 较窄 2025 基数对应 $0.89 billion、SNS Insider 为 $6.93 billion、Mordor Intelligence 为 $7.11 billion),反映的不是共识趋同,而是各方对「physical AI」硬件和软件边界的定义差异很大。由于 Velaura 的收入模式是版税制(授权费加实测节电量分成),而不是直接卖芯片,即便底层加速器市场的 SAM 定义清楚,也会高估 Velaura 自己能拿到的收入;分析上正确的 SOM 应按每颗出货 XPU 对应的版税美元计量,但没有公开来源给出这个数字的自下而上估计。因此,本章明确保留规模测算矛盾,而不是压成一个标题数字,并把缺失的 SOM 估计记录为具名证据缺口。[CM002, CM008, CM005, CM006, CM007, CM027]
| 发布方 | 年份 | 地理范围 | 数值(USD) | CAGR | 方法 | 置信度 | 限制 |
|---|---|---|---|---|---|---|---|
| SemiconductorInsight | 2026 | 全球 | $13.8B(数据中心 AI 芯片市场) | 到 2034 年 14.1% | 自下而上的供应商 / 分部跟踪 | 中 | 范围较窄,排除更广的 AI 半导体类别 |
| Deloitte(经 Axis Intelligence) | 2026 | 全球 | ~$500B(整体 AI 芯片市场,修订后) | n/a(点估计) | 基于初始 $300B 估算的自上而下需求信号修订 | 中 | 年中修订;方法未完全公开 |
| Gartner(经 Axis Intelligence / VoxBooster) | 2026 | 全球 | ~$390B($1.3T 半导体总收入中约 30% 的隐含 AI 份额) | n/a | 自上而下的半导体总市场份额估算 | 中 | 派生 / 隐含数字,并非直接发布的仅 AI 条目 |
| IDC(经 Axis Intelligence) | 2026 | 全球 | $281B(智能数据中心细分市场)/ $477.1B(数据中心半导体) | n/a | 细分市场层面的自下而上预测 | 中 | 与 Deloitte/Gartner 数据的细分市场边界不同 |
| Mordor Intelligence | 2026 | 全球 | $7.11B(物理 AI 市场) | 37.46% (2026-2031) | 自下而上的机器人 / 硬件 + 软件市场模型 | 中 | 宽泛的物理 AI 定义覆盖工业机器人和服务机器人 |
| MarketsandMarkets | 2025 | 全球 | $0.89B(物理 AI 市场,定义更窄) | 47.2% (2026-2032) | 物理 AI 口径比 Mordor Intelligence 窄 | 中 | 基准年和口径与 Mordor/SNS Insider 不同,限制直接可比性 |
| SNS Insider | 2026E | 全球 | $6.93B(物理 AI 市场) | 32.53% (2026-2033) | 自下而上的市场模型 | 中 | 独立估算与 Mordor Intelligence 大体一致,尽管方法不同 |
各数值沿用发布方披露的方法和口径;由于发布方的范围边界(数据中心 AI 芯片、整体 AI 半导体、物理 AI)实质不同,各行不能直接相加。置信度均为中,因为这些数据都不是经过审计的一手数据。
[CM001, CM002, CM005, CM006, CM007, CM008]在证据约束下,为 Velaura 能效 IP 授权机会拆分 TAM / SAM / SOM。
SAM 混合了两家发布方对 2026 年单点估计(数据中心 AI 芯片市场 + 物理 AI 市场),只应视为示意,不是经验证的单一来源数字;SOM 没有公开自下而上估计,因此标为证据缺口,而不是给出数字。
[CM002, CM001, CM005, CM027, CM028]三种独立方法给出的 2026 年全球 AI 半导体市场规模低 / 基准 / 高估计(十亿美元)。
三项数据均描述 2026 日历年,但口径不同(细分市场子集 vs. 推算份额 vs. 自上而下需求修订);应把它们视为一组方法区间,而不是同一数字的置信边界。
[CM002, CM003]2.3 买方分层与采用路径
Velaura 的三类目标买方在预算负责人和采用触发点上结构不同。超大规模数据中心里,买方是云基础设施 / 硅团队,付款方是超大规模云厂商 capex 预算——Big Five 超大规模云厂商 2026 capex 预计为 $600-725 billion,约 75% 投向 AI 基础设施——采用触发点则是数据中心扩张遇到电力 / 冷却约束。Physical 与具身 AI 里,买方是机器人 / 无人机硬件工程团队,付款方是 OEM R&D 和物料清单预算,采用触发点是常开自主系统对电池续航和热包络的要求;端侧计算已经占 physical-AI 部署的 71.43%,与 Velaura 的片上效率叙事直接对齐。边缘 AI 里,设备 OEM 采购团队购买芯片,是为了延长电池寿命并启用本地推理。第四个横向子板块,是 Google TPU、AWS Trainium、Microsoft Maia、Meta MTIA 等超大规模云厂商内部定制 ASIC 项目;它们动用的是更广义数据中心板块的同一 capex 预算,但采用路径不同:这些团队已经自研硅,只会为了补上相对商用 GPU 竞争者的每瓦性能差距而授权 Velaura IP,而不是替代既有芯片采购决定。Broadcom 与 Marvell 合计支撑了 80% 以上超大规模云厂商定制硅背后的 ASIC 设计服务,这又带来一层复杂性:它们可以作为渠道伙伴,把 Velaura IP 嵌入设计服务,也可以成为直接竞争者,提供自己的功耗优化能力。[CM012, CM011, CM019, CM031, CM035]
| 细分市场 | 买方 | 用户 | 付费方 | 工作流 | 预算负责人 | 采用触发因素 |
|---|---|---|---|---|---|---|
| 超大规模数据中心 | 云基础设施 / 芯片团队 | 云租户运行的 AI 训练 / 推理工作负载 | 超大规模云厂商资本开支预算 | RTL 交付 -> Velaura 物理设计优化 -> 流片 -> 量产 | 超大规模云厂商基础设施 / 资本开支负责人 | 数据中心扩张受电力 / 散热约束 |
| 物理 / 具身 AI OEM | 机器人和无人机硬件工程团队 | 在电池 / 热约束下运行的自主机器 | OEM R&D 和 BOM(物料清单)预算 | 平台设计阶段选芯片 -> IP 授权谈判 -> 集成 | 机器人 OEM 工程 / 采购负责人 | 常开式具身智能对续航和热包络的要求 |
| 边缘 AI 设备厂商 | 消费电子 / IoT 芯片团队 | 智能手机、可穿戴设备、智能摄像头终端用户 | 设备 OEM 组件预算 | SoC 设计 -> 低功耗 IP 集成 -> 量产 | 设备 OEM 采购 / 工程负责人 | 更长续航和本地(非云端)推理需求 |
| 超大规模云厂商自研 ASIC 项目(TPU、Trainium、Maia、MTIA) | 超大规模云厂商内部 ASIC 设计团队 | 内部云端 AI 工作负载 | 与数据中心细分市场相同的超大规模云厂商资本开支预算 | 自研 RTL 设计 -> 可能为功耗优化引入第三方 IP 授权 -> 内部晶圆厂 / 代工厂认证 | 超大规模云厂商芯片 VP/GM 预算 | 拉大相对商用 GPU 竞争对手的每瓦性能差距的压力 |
采用触发因素和预算负责人由公开资本开支、市场结构和 Velaura 产品定位来源推断;各超大规模云厂商内部的确切采购流程未被独立披露。
[CM012, CM019, CM025, CM035]Velaura 三个目标细分市场及定制 ASIC 超大规模云厂商子细分中的买方、用户、付款方关系。
[CM019, CM031]对 Velaura 这类硅 IP 授权厂商,从初始超大规模云厂商接触到经常性版税收入的典型步骤。
阶段间转化率是基于引用来源所述典型半导体 IP 授权周期的示意估计,并非 Velaura 披露的具体数据。
[CM032, CM027]2.4 增长驱动与采用约束
支撑 Velaura 投资逻辑的主导增长驱动,是 AI 数据中心用电需求正走在近乎翻倍的轨道上:从 2025 约 485 TWh 增至 2030 的 950 TWh,而仅美国在 2026 就面临估计 9.3 GW 结构性电力缺口。这个物理约束抬高了任何能从每瓦电力中榨出更多有效算力的技术价值。超大规模云厂商 capex 增长(2026 预计 $600-725 billion,同比约增 75%)扩大了未来可支付 IP 授权费的预算池;定制 ASIC 出货增速(44.6% CAGR)快于商用 GPU(16.1% CAGR),也扩大了天然适合授权的自研硅项目集合。反过来,几项真实采用约束也在压制这些驱动:美国对华出口管制和中国稀土反制增加合规成本和供应链不确定性;先进节点芯片设计和制造的高资本强度提高了竞争者进入门槛,但并不消除 Velaura 自身执行风险;典型 12-24 个月 SoC 集成和晶圆厂认证周期意味着,即便签下设计定点,版税现金流也会延后一到两年。估值时还应提示另一项约束:更广泛行业正在同时从多个角度解决同一个电力问题——AI 服务器中液冷采用率从 2024 的 15% 升至 2026 预计 76%——这意味着系统级冷却改善可能随时间部分替代芯片级效率提升,而不完全依赖 Velaura 这类厂商。[CM015, CM016, CM012, CM017, CM022, CM034]
| 驱动因素 / 约束 | 方向 | 时间 | 影响 | 尽调追问 |
|---|---|---|---|---|
| AI 数据中心用电需求到 2030 年接近翻倍 | 驱动因素 | 现在至 2030 年 | 直接提高买方为每瓦性能提升付费的意愿 | 确认 Velaura 签约节省额是否基于一致基线测算 |
| 美国 9.3 GW 结构性电力缺口 | 约束 | 2026 年起并继续恶化 | 无论芯片效率多高,新增 AI 容量建设都会受限,但每一瓦可用电力的效率价值更高 | 评估 Velaura 的 TCO 叙事是否会改变超大规模云厂商的站点选择决策 |
| 超大规模云厂商资本开支在 2026 年增至 $600-725B(同比 +约 75%) | 驱动因素 | 2026 | 扩大可寻址预算池,未来可能为 Velaura 授权费买单 | 确认这部分资本开支中有多少可自由支配,多少已通过合同锁定给现有供应商 |
| 自研 ASIC 出货增速超过商用 GPU(44.6% vs 16.1% CAGR) | 驱动因素 | 至 2033 年 | 扩大超大规模云厂商内部芯片项目池,这些项目天然可能采购 Velaura 的 IP 授权 | 核实 2026 年四个具名超大规模云厂商 ASIC 项目中,是否有任何一个就是 Velaura 未披露的合作 |
| 美中出口管制与稀土反制 | 约束 | 持续,2026 年 | 给 AI 芯片价值链增加合规成本和供应链不确定性,间接影响 Velaura 的代工厂 / 客户基础 | 评估 Velaura 对中国相关客户或代工产能的敞口 |
| 资本强度高 / 晶圆厂和验证成本达数十亿美元 | 约束 | 结构性,持续 | 推高切换成本和进入壁垒,有利于 Velaura 这类已有生产出货规模(30M+ ASIC)的现有玩家 | 确认 Velaura 实际 R&D 和认证支出,与其 Series A 融资金额相比如何 |
| 新硅 IP 需要 12-24 个月 SoC 集成 / 认证周期 | 约束 | 持续 | 即便拿下设计导入,收入确认也会变慢,版税现金流被推迟 | 索取 Velaura 披露的设计导入至版税收入时间线 |
| 液冷采用率从 15%(2024)升至 76%(2026E) | 驱动因素 | 2024-2026 | 表明行业正从整条技术栈解决电力 / 散热约束,可能部分替代芯片级效率提升 | 评估冷却改善是否会随着时间削弱 Velaura 芯片级叙事的紧迫性 |
方向(驱动因素 / 约束)和时间为作者对底层引用市场数据的综合判断,并非任何单一发布方的原始框架。
[CM015, CM016, CM012, CM017, CM022, CM034]2.5 尽调缺口与保留的矛盾
本章刻意保留两类开放问题,而不是用一个自信数字把它们抹平。第一,没有公开来源测算 Velaura 特定版税授权 SOM;所有可用数字都指向更宽的相邻市场(数据中心 AI 芯片、physical AI 硬件或整体 AI 半导体),把它们混合使用(如本章市场规模金字塔中的示意 SAM)已明确标记为近似值,而非经验证估计。第二,三家独立且声誉较高的发布方,对实质上同一时间窗口的 physical-AI 市场规模估计相差超过 7x;这更应被解读为「physical AI」尚未成为标准化市场类别的信号,而不是某一个估计必然错误的证据。在获得自下而上、跨来源三角验证的估计前,尽调应对 Velaura 自有材料或任何单一第三方报告中的单点 TAM 主张保持相应怀疑。[CM028, CM008, CM027]
2.6 图表
03竞争格局
3.1 竞争格局:直接、在位、相邻与替代
Velaura AI 面对的不是一个边界清晰的同业集合,而是四类不同竞争者。在直接低功耗 AI 芯片同业中,过去一年行业急剧整合:近存算竞争者 Untether AI 于 June 2025 关闭,并在 October 2025 申请破产,负债超过 $128 million,其工程团队被 AMD 收购式招揽;拥有约 100 家客户的边缘 AI 芯片厂商 Hailo 于 July 2026 被 Microchip Technology 收购;Mythic 则在 December 2025 完成 $125 million Series D 前,经历了一次完整的架构和领导层重组。Groq 展示了可比公司重估速度有多快:December 2025,一笔据称 $20 billion 的 NVIDIA 交易吸收了其创始技术团队和 IP 后,Groq 完全转向新的商业模式,并以 $3.5 billion 估值融入新资本,低于一年前的 $6.9 billion。相比之下,Cerebras 于 May 2026 以约 $56 billion 完全稀释估值上市,验证了该板块的上行空间。NVIDIA 和 AMD 仍是主导在位者,并与 Intel 一起拿走商用数据中心 AI 芯片收入的绝大多数(仅 NVIDIA 约 87.4%)。但结构上最重要的替代品,是超大规模云厂商内部自建:Google TPU、AWS Trainium、Microsoft Maia、Meta MTIA 等自有定制硅项目出货增速接近商用 GPU 的三倍;这些团队原则上可以用 Cadence 或 Synopsys 的商业 EDA 工具在内部实现类似能效提升,而不是授权 Velaura IP。[CP001, CP002, CP004, CP005, CP006, CP007]
| 竞争对手 | 类别 | 规模 / 融资 | 目标细分市场 | 差异化 | 局限性 |
|---|---|---|---|---|---|
| Velaura AI | 直接(自身) | 累计融资 $424M+,4 轮;估值 $1B+;已出货 30M+ ASIC | 数据中心、物理 / 具身 AI、边缘 AI | 将能效 IP 授权嵌入客户 RTL;大规模量产已验证 | 未披露具名超大规模云客户;没有独立基准 |
| Untether AI | 直接(已倒闭) | 2025 年前获得风投融资;破产时负债 $128M+ | 边缘到云端 AI 推理 | 近存计算架构,SpeedAI 在 66W 下达到 2 PFLOPS | 2025 年 6 月关闭;2025 年 10 月破产;产品停产 |
| Mythic | 直接 | 2025 年 12 月 $125M Series D;累计融资 $260M+ | 机器人、汽车、国防、边缘 / 数据中心推理 | 模拟存内计算,声称 GPU 能效最高提升 100x | 早期财务困境迫使公司全面重构架构和团队 |
| Hailo | 直接(已被收购) | 收购前融资 $344M;约 100 家客户;2026 年 7 月被 Microchip 收购 | 工业、汽车、嵌入式视觉边缘 AI | Hailo-8/10/15 NPU;开发者社区强(10,000+、Raspberry Pi) | 不再独立;路线图现由 Microchip 控制 |
| Groq | 直接 | $20B NVIDIA IP / 人才交易(2025 年 12 月);2026 年 8 月以 $3.5B 重置估值融资 $350M;此前 2025 年 9 月估值 $6.9B 融资 $750M | AI 推理(LPU 架构) | 确定性低延迟推理;转向「AI 推理新云」模式 | 创始技术团队 / IP 流向 NVIDIA;估值重置约低 50% |
| Cerebras Systems | 直接 | 2026 年 5 月 14 日 IPO,完全摊薄估值约 $56B | 大模型训练(晶圆级) | Wafer-Scale Engine 通过减少数据移动提升训练效率 | 为训练规模优化,并非主要面向边缘 / 具身能效 |
| Tenstorrent | 直接 | 截至 2026 年中仍独立;未披露独角兽估值重置 | 开放、可适配 AI 加速器;RISC-V | 开放架构 Tensix 核;IP 授权加硬件 | 头部估值 / 能见度弱于 Groq 或 Cerebras |
| NVIDIA | 现有厂商 | 商用数据中心 AI 芯片收入约 87.4%(2026 年 Q1) | 所有 AI 计算细分市场 | 主导生态(CUDA)、全栈机架 / 系统、NVLink 开放 | 在推理市场加速被超大规模云厂商自研 ASIC 夺走份额 |
| AMD | 现有厂商 | 商用数据中心 AI 芯片收入约 6-7% | 数据中心与边缘 AI 加速 | 收购式雇佣 Untether AI 工程团队,补强 AI 硬件 / 软件 | 在商用 AI 加速器份额上远落后于 NVIDIA |
| Google TPU / AWS Trainium / Microsoft Maia / Meta MTIA(超大规模云厂商自研 ASIC) | 内部自研 / 替代 | 由超大规模云厂商资本开支资助(2026 年 $600-725B) | 内部云端 AI 工作负载 | 完整垂直控制芯片;没有外部授权费 | 不向第三方商业销售;也可能成为 Velaura 授权客户 |
| Arm(相邻 - IP 授权模式) | 相邻 | 成熟上市公司;2026 年推出新 AI 芯片部门 | 横跨移动、云、嵌入式的广泛处理器 IP 授权 | Velaura CEO 直接引用的“版税 + 授权”先例 | 2026 年开始直接销售成品芯片,削弱自身纯授权先例 |
| Synopsys / Cadence(EDA 工具 - 现状) | 现状 / 替代 | 成熟上市 EDA 厂商 | 任何推进内部功耗优化的芯片设计团队 | AI 驱动的低功耗设计工具(DSO.ai、Cerebrus),无需第三方 IP 授权即可使用 | 需要内部设计能力;不提供 Velaura 这种已预验证、经量产证明的库 |
| Broadcom / Marvell(ASIC 设计服务) | 潜在进入者 / 渠道 | 为 80%+ 超大规模云厂商自研芯片背后提供设计服务 | 超大规模云厂商自研 ASIC 设计 | 深厚的超大规模云客户关系和制造规模 | 可能内部开发竞争性功耗优化 IP,而不是与 Velaura 合作 |
融资和估值数据均为各公司或独立媒体截至运行日期公开披露;Velaura 行综合前文数据以便直接比较。
[CP001, CP002, CP004, CP005, CP006, CP007]Velaura 与主要竞争对手在所称每瓦性能提升和已证明生产 / 部署规模上的序数定位。
坐标轴是作者基于各公司披露主张和产量数据综合出的、有证据支撑的序数评分(1-10),不是标准化第三方基准测试;Velaura 的 y 轴评分把传统 Teraflux 出货量与更新、未验证的 AI 芯片专属部署混合在一起。
[CP018, CP004, CP005, CP001, CP010, CP011]3.2 能力、定价与商业模式对比
仅看每瓦性能主张,Velaura 的 2-4x 提升不是该板块最大的标题数字——Mythic 声称其模拟存内计算架构最高可带来 100x 效率提升——但 Velaura 的特殊之处在于,它把效率主张与 30 million+ 已出货量产 ASIC 绑定在一起;同阶段的 Mythic 或已倒闭的 Untether AI 都拿不出这种规模证明。商业模式在同业之间差异很大:Velaura 及其最接近的结构类比 Arm 历史处理器授权业务,都是收前置费用再加按使用计提的版税;NVIDIA、AMD、Hailo 和 Mythic 则都直接销售成品硅,捕获完整单颗经济性(也承担完整库存 / 制造风险),而不是收取版税流。值得注意的是,Arm 自身在 2026 打破 30 年纯 IP 授权路径,开始直接销售成品 AI 芯片;这说明,即便成熟且成功的版税模式在位者,也可能面临与自己授权客户直接竞争的压力——这对 Velaura 模式是一个值得跟踪的先例风险。披露和独立验证方面,Velaura 相比部分同业处于劣势:Hailo 在被收购前公开点名约 100 家客户,而 Velaura 没有点名其声称正在合作的三家超大规模云厂商;本章审阅来源中,也不存在 Titan Core 与 Mythic 或 Hailo 架构的独立正面对比基准。[CP018, CP014, CP015, CP016, CP029, CP020]
| 采购标准 | Velaura AI | 商用 GPU(NVIDIA/AMD) | 超大规模云厂商自研 ASIC | Mythic / Hailo(边缘推理) | Arm(授权先例) |
|---|---|---|---|---|---|
| 每瓦性能提升主张 | MATMUL 运算提升 2-4x(公司声称) | 代际渐进提升 | 未对外披露 | 声称最高 100x(Mythic,架构特定) | 未知——不是芯片性能指标 |
| 量产 / 部署规模证明 | 已出货 30M+ ASIC | 数亿台 | 数百万台(仅内部) | Hailo:约 100 家客户;Mythic:披露的出货有限 | 历史累计授权数十亿颗内核 |
| 商业模式 | 授权费 + 与节电收益挂钩的版税 | 硬件直销 | 内部成本中心,不对外销售 | 硬件直销 | 授权费 + 单台版税(历史模式) |
| 具名客户披露 | 未披露(4 家超大规模云厂商中的 3 家,未具名) | 公开客户基础广泛 | N/A(内部) | Hailo 具名约 100 家客户;Mythic 部分披露 | 授权客户基础披露较广 |
| 是否有独立第三方基准 | 未知——审阅来源未识别 | 广泛参与基准测试(MLPerf 等) | 通过超大规模云厂商披露部分可见 | Mythic 的 100x 主张未知 | N/A |
| 出口管制 / 监管敞口 | 当前出口管制报道未直接点名 | 被美中出口管制规则直接点名(如 H200 上限) | 因超大规模云厂商全球布局产生间接敞口 | unknown | unknown |
标注“未知”的单元格代表明确证据缺口,而非假定为负面;对应尽调路径见 evidenceGaps。
[CP018, CP010, CP013, CP020, CP022]| 竞争对手 | 价格 / 单位 / 合同模式 | 包含能力 | 折扣或未知项 | 影响 |
|---|---|---|---|---|
| Velaura AI | 前置授权费 + 与每颗 XPU 实测节电额挂钩的版税 | 物理设计 IP、库、集成支持 | 未披露具体版税率和测算方法 | 收入随客户实现的节省额增长,而不只是随出货量增长;激励更一致,但收入预测更复杂 |
| Arm(历史先例) | 前置授权费 + 单位版税(公开估计历史上约为芯片 ASP 的 1-2%) | 处理器内核 IP、架构、工具生态 | 当前具体版税率因授权层级而异,未完全公开 | 评估 Velaura 自身版税经济性时最接近的结构性参照 |
| NVIDIA / AMD(商用 GPU) | 硬件单元直销(每颗加速器数千美元) | 完整芯片、参考系统、CUDA / 软件生态(NVIDIA) | 给超大规模云厂商的批量折扣未公开 | 单位收入最高,但需要投入完整制造和分销 |
| 超大规模云厂商自研 ASIC | 无外部价格;完全计入超大规模云厂商资本开支 | 仅为内部工作负载定制 | 完整 R&D 成本未公开拆分 | 对外部供应商没有收入模式,除非它们授权 Velaura 这类赋能 IP |
| Hailo / Mythic(边缘推理芯片) | 硬件单元直销(单芯片定价大多未公开) | 芯片加 SDK / 开发者工具 | 具体价位未公开 | 标准无晶圆厂半导体经济模型;不同于 Velaura 轻资产版税模式,它完全暴露于单位成本和库存风险 |
具体版税百分比和单价具有商业敏感性,受访公司均未完全披露;表内数字是目前最佳公开估计或结构性描述。
[CP014, CP029, CP017]Velaura 及最接近的直接 / 相邻竞争对手在监管和披露透明度上的姿态——这是不同于更广能力矩阵表的独立视角。
[CP020, CP022, CP023]3.3 切换成本、锁定效应与分发权力
一旦超大规模云厂商决定把 Velaura 的 Titan Core IP 集成进芯片设计,切换成本可能相当高:先进工艺节点上新硅 IP 模块的 SoC 集成和晶圆厂认证通常需要 12 到 24 个月,设计定点一旦落地就会形成真实锁定,但也意味着 Velaura 自己的收入确认会比任何已宣布合作晚一年或更久。更广泛 AI 芯片价值链中的分发权力高度集中:Broadcom 与 Marvell 合计支撑 80% 以上超大规模云厂商定制硅背后的 ASIC 设计服务,因此它们对 Velaura 这类第三方 IP 能否进入某个设计拥有重要影响力——它们可以作为渠道伙伴,把 Velaura 的库嵌入设计服务,也可以在内部搭建等价功耗优化能力,成为直接竞争者。NVIDIA 应对更广泛定制硅迁移的方式,不是单纯比拼独立芯片性能,而是销售更完整的一体化机柜和系统,并向第三方 ASIC 开放 NVLink 互连;这条策略如何演化,将决定它是为 Velaura IP 创造新集成机会,还是进一步加深 NVIDIA 生态锁定。[CP026, CP027, CP028]
3.4 护城河耐久性与反向竞争证据
Velaura 最清晰的护城河,是经过量产验证的制造历史:30 million+ 已出货 ASIC 给超大规模云厂商伙伴提供了真实良率和可靠性记录,白手起家的竞争者很难快速复制。不过,这条护城河也有前提:这些出货量主要来自 Auradine 的比特币挖矿芯片,而不是 AI 专属的 Titan Core 硅,所以它降低的是制造执行风险,而不是完全验证 AI 专属性能主张。耐久性面临的最清晰威胁,是商品化风险:Cadence 的 Cerebrus 和 Synopsys 的 DSO.ai 已经提供 AI 驱动的低功耗设计自动化,任何竞争者或客户都可以用它们在内部追求类似节电;如果晶圆厂伙伴(Samsung、TSMC)选择把等价库直接嵌入标准工艺设计套件,Velaura 的很大一部分差异化可能被侵蚀。行业竞争密度变化对这项评估有双重含义:Untether AI 关闭、Hailo 被收购,减少了 Velaura 必须正面竞争的独立低功耗 AI 芯片厂商数量;但也同样可以被解读为,该板块可能还无法支撑太多独立厂商——这个风险同样适用于 Velaura,而不只是已经退出的同业。[CP035, CP034, CP030, CP031, CP017, CP033]
| 护城河主张 | 威胁 | 严重性 | 缓释 / 尽调追问 |
|---|---|---|---|
| 30M+ 量产 ASIC 证明制造良率 / 可靠性 | 既往规模来自比特币挖矿芯片,不是 AI 专用 Titan Core 芯片 | 中 | 索取嵌入 Titan Core 的芯片具体量产规模,不只看传统 Teraflux 出货量 |
| 与 4 家最大超大规模云厂商中的 3 家建立先发关系 | 名称未披露;无法独立验证关系是否持久或排他 | 高 | 寻求客户确认或合同排他条款 |
| “版税 + 授权”模式创造类似 Arm 的经常性收入 | Arm 自身在 2026 年正从纯授权转向直接卖芯片 | 中 | 评估集成诀窍转移后,Velaura 的超大规模云客户是否也可能绕开 Velaura |
| 能效 IP 难以快速复制 | Cadence 和 Synopsys 出售 AI 驱动低功耗 EDA 工具,任何竞争对手或客户都能内部使用 | 高 | 用当代 EDA 工具可实现的最佳内部结果,对比 Titan Core 声称的收益 |
| 董事 / 投资人网络(Intel CEO、MARA CEO、多家 VC)释放可信度信号 | 投资人董事席位不保证商业排他,也不能阻止投资人支持竞争对手 | 低 | 确认是否有任何投资人持有竞争性低功耗 AI 芯片公司的股份 |
| 轻资产授权模式避开晶圆厂资本开支风险 | 代工合作伙伴(如 Samsung、TSMC)可能把等效功耗库直接嵌入自家的工艺设计套件 | 高 | 评估 Velaura 与代工 / 工艺合作伙伴之间是否有排他或优先合作条款 |
| 直接竞争对手已倒闭(Untether AI)或被收购(Hailo),独立竞争密度下降 | 整合也可能说明该细分市场无法支撑很多独立供应商,Velaura 也包括在内 | 中 | 跟踪未来 12 个月是否还有其他独立低功耗 AI 芯片厂商退出市场 |
| 物理 AI 聚焦的是比数据中心 AI 芯片更新、竞争更少的利基市场 | 资金更厚、体量更大的机器人 / 芯片厂商(Qualcomm、NVIDIA Jetson 产品线)可能直接切入 Physical AI 能效赛道 | 中 | 跟踪既有机器人芯片厂商路线图,看是否发布专门针对能效的动态 |
严重性为作者基于每行所引证据作出的定性评估,并非公司披露的风险评级。
[CP035, CP034, CP013, CP015, CP016, CP030]将 Velaura 规模证明与市场结构信号结合的紧凑竞争耐久性摘要。
[CP018, CP001, CP005, CP008]3.5 图表
04财务情况
4.1 收入模式、定价与确认
Velaura AI 核心 AI 芯片收入机制是两段式结构:前置 IP 授权费,加上与客户使用 Titan Core 后可衡量节电量挂钩的版税;公司明确把这一模式类比为 Arm 历史处理器授权业务。公司材料把客户侧节省基础示意为三年每颗 XPU 约 $1,300,但本章审阅的任何来源都没有披露版税比例、「已实现节电量」的测量方法,或最低授权费。鉴于整套版税机制都依赖双方从未公开说明的节省测量,这是一个实质性缺口。由于版税收入绑定的是客户实际节省,而不只是单颗出货,收入确认很可能落后于实体芯片生产;滞后时间取决于客户自身认证和部署周期,合理上可能与竞争格局章节中识别的 12-24 个月 SoC 集成窗口相同。相比之下,Velaura 历史 Teraflux 比特币挖矿硬件业务采用传统直接硬件销售模式;MARA 自己的 SEC 文件显示,历史业务按预付款运行,在截至 March 31, 2025 的季度内,MARA 针对产品采购预付 $22.3 million,对应 $57.2 million 未结余额;到最近一个申报期已全额结清,且没有新增预付款。没有来源披露 Velaura 当前收入组合中,有多少仍来自历史硬件业务,又有多少来自新的 Titan Core 授权模式。[CI007, CI008, CI009, CI028, CI030, CI005]
| 收入来源 | 机制 | 计量单位 | 当前数值 / 状态 | 质量 | 尽调要求 |
|---|---|---|---|---|---|
| Titan Core IP 授权(AI 芯片) | 预付授权费 + 与每颗 XPU 实测节电额挂钩的版税 | 每份授权按 USD 计价 + 实测节省额百分比 | 未披露;据报道,4 家最大超大规模云厂商中有 3 家已接洽,未具名 | 公司声称,未验证 | 要求披露版税率、最低费用,并提供至少一个已签约客户参考 |
| 旧 Teraflux 比特币挖矿硬件销售 | 直接销售 ASIC 矿机系统,历史上由客户预付 | 每套矿机按 USD 计价 | 根据 SEC 文件,MARA 在 2025 年 Q1 预付 $22.3M 用于产品采购;2026 年 Q1 未披露新的预付款 | 监管文件佐证(仅投资方一侧) | 若仍有剩余 Teraflux 积压订单,要求 Velaura 提供自身收入确认时间表 |
| 旧 Auradine 客户订单额(区块链 / 挖矿,2026 年前) | 交付前订单额 | 按 USD 计的订单额 | Series B(2024 年 4 月)披露订单额 $80M | 公司声称,数据已陈旧 | 如有,要求提供转向 AI 后的最新订单额 |
| Physical AI / Edge AI 授权(未来业务) | 将同样的授权费 + 版税机制用于机器人 / 具身 AI 和边缘设备厂商 | 每份授权按 USD 计价 + 节省额百分比 | 未披露客户或收入;该业务仅被描述为目标市场 | 仍属愿景,暂无收入证据 | 要求提供 Physical AI / Edge AI 业务的已签约或在谈交易 |
Velaura 未披露新 AI 芯片授权业务收入,任何独立文件也未佐证;唯一来自文件的数字(MARA 的产品采购预付款)对应旧 Teraflux 硬件业务,而非 Titan Core 授权。
[CI007, CI005, CI006, CI017, CI020]| 价格 / 单位 / 合同 | 标价与实际成交价 | 折扣 / 未知项 | 来源 |
|---|---|---|---|
| Titan Core 预付授权费 | 未公开披露(标价或实际成交价均无) | 费用可能随客户规模和制程节点变化;未公开价目表 | 公司新闻稿(SI003);无独立确认 |
| Titan Core 基于节电额的版税 | 按实测节省额分成,而非固定费率;实际费率未知 | “已实现节电额”的测算方法未披露 | 来源:Heise Online(SI011)、GamesBeat(SI021) |
| 示例性客户节省额基础:3 年每颗 XPU 约 $1,300 | 公司模型估算,不是客户实际数字 | 假设公司自称的 MATMUL 能效提升 2-4x 成立;缺少独立验证 | Velaura Titan Core 新闻稿(SI003)、Semiconductor Online 转载(SI020) |
| Arm 可比版税率(历史上约为芯片平均售价(ASP)的 ~1-2%) | 公开估计的成熟授权龙头区间,不是 Velaura 自身费率 | Velaura 实际版税比例可能与 Arm 类比明显不同 | 来源:Stock Dividend Screener(SI025)、Welcome.ai(SI026) |
没有来源披露 Velaura 的实际标价、成交价或折扣做法;所有定价相关数字要么来自公司自己的示例性节省额估算,要么是外部类比(Arm),并非 Velaura 特定的实际经济性。
[CI008, CI009, CI025]超大规模云厂商 / OEM 客户活动如何转化为 Velaura 授权和版税收入。
阶段名称和顺序根据 Velaura 自身产品描述及行业通行的硅 IP 授权周期推断;Velaura 未披露其内部收入确认流程。
[CI007, CI030, CI028]4.2 成本结构与单元经济
本章审阅来源没有披露 Velaura 的实际毛利率、销货成本结构、R&D 支出或获客成本。行业基准只能提供粗略可比框架:无晶圆厂半导体公司通常报告 50-65% 毛利率(NVIDIA 等领先者超过 70%),早期无晶圆厂初创公司通常把收入的 20-40% 花在 R&D(收入前阶段可能冲到 50% 以上),同时每月烧掉 $500,000 到 $2,000,000;12-18 个月现金跑道被视为审慎水平。这些数字都不是 Velaura 专属,直接套用可能低估基于版税的 IP 授权方与按颗销售的无晶圆厂芯片公司之间独特经济性:Velaura 模式所需制造 capex 和库存风险远低于典型无晶圆厂可比公司,因为它不是为晶圆开工付费的一方;但这也意味着,每个设计定点带来的收入只是底层芯片总价值的一小部分,类似 Arm 历史估计 1-2% of ASP 的版税率。对于 Velaura 任一业务部分,唯一可独立交叉验证且接近现金流的数据点,是 MARA 在 SEC 文件中披露的产品采购预付款和投资账面价值;这些数据对应历史 Teraflux 硬件业务,而不是本估值逻辑主要依赖的新 Titan Core IP 授权模式。[CI011, CI012, CI013, CI025, CI027, CI029]
| 指标 | 数值 / 空缺 | 置信度 | 为什么重要 | 尽调要求 |
|---|---|---|---|---|
| 毛利率(Velaura 特定) | 低 | 决定一次设计定点有多少能转成现金利润;无晶圆厂可比公司毛利率为 50-65% | 要求 Velaura 提供实际销货成本(COGS)结构或可比毛利披露 | |
| 获客成本 / 销售周期长度 | 低 | 用来评估面向超大规模云厂商的企业销售落地效率 | 要求提供授权交易平均签约周期和专职销售人数 | |
| 月度烧钱速度 | 低 | 结合已披露的 Series A 融资额,判断现金跑道 | 要求提供最近月度烧钱数字或现金流量表节选 | |
| 研发支出占收入比例 | 低 | 可比无晶圆厂初创公司通常为 20-40%,收入前可能冲到 50%+;反映资本效率 | 要求提供研发人员数和支出拆分 | |
| 设计定点到版税收入的滞后期 | 低 | 根据第 2/3 章发现,12-24 个月 SoC 认证周期会拖慢任何已宣布接洽后的现金转化 | 要求披露最早超大规模云厂商接洽进入量产的时间表 | |
| MARA 产品采购预付款(旧 Teraflux,2025 年 Q1,来自监管文件) | 已预付 $22.3M;未结余额 $57.2M | 高 | 目前可得、唯一来自独立监管文件的 Velaura 相关现金流代理指标 | 跟踪 MARA 是否在未来 10-Q 中披露进一步预付款或余额 |
所有数值为 null 的单位经济字段都是真实的公开披露缺口,不代表隐含为零;只有来自 MARA 文件的一行获得独立佐证。
[CI011, CI012, CI013, CI027, CI005]在已披露输入缺失情况下,从一次设计定点到版税现金流的定性单位经济桥接。
这张桥接图除 $1,300/XPU 节电估算外,其余美元数字均为定性或未披露;它展示机制,不是计算出的单位经济模型。
[CI008, CI009, CI029]4.3 公开牵引力与员工数
Velaura AI 唯一公开披露且带具体金额的牵引力数据,是 Auradine 在 April 2024 Series B 融资时同步披露的 $80 million 客户订单;这个数字比 AI 转型早约两年,无法说明 Titan Core 专属牵引力。员工数估计在来源之间差异明显:Tracxn 跟踪的员工趋势显示,截至 March 26, 2026 为 62 名员工;其他分析师数据库(PitchBook、AI Market Watch)给出更宽的 100-150 名区间,LinkedIn 式公司规模区间则显示 101-500 名员工。这些数字没有一个得到公司确认,区间本身(低高估计约差 2-3x)已经是有用信号:一家刚更名的私营公司,如果没有直接披露,员工数跟踪并不可靠。没有来源披露当前 ARR、收入运行率、带名称的活跃客户数,或转型后的任何 AI 专属订单数字。[CI014, CI015, CI016, CI017, CI018]
以低 / 基准 / 高情景框定 Velaura/Auradine 四轮累计已披露股权融资额。
$424M 是四次单独公告轮次规模的简单加总,未经过独立审计,也未与单一披露总额核对;MARA 的 $85.4M 账面价值是该总额中一位投资人的份额,不是额外金额。
[CI033, CI001, CI017]4.4 资本充足性与融资依赖
Velaura AI/Auradine 四轮已披露股权融资累计为约 $424 million:$81 million Series A(2023)、$80 million Series B(2024)、$153 million Series C(2025)和 $110 million 2026 Series A;不过这只是把分别公告的数字简单相加,并非经审计或公司确认的总额。公司自有公告和 Mayfield 领投方评论均称,August 2026 融资用途是扩充工程团队和面向客户的团队,并深化战略伙伴关系;但没有来源披露当前账上现金、月度烧钱速度或由此得到的现金跑道。没有来源披露任何独立于已披露股权融资的债务额度或项目融资安排。对于 Velaura 任一业务部分,唯一独立申报(SEC 来源)的财务数据点,是 MARA Holdings 的关联方披露:截至 March 31, 2026,投资账面价值为 $85.4 million,一个董事席位,以及后续 Velaura 融资轮确定更高可观察价格后,MARA 对现有优先股和普通股持仓按 ASC 321 记录的 $11.9 million 加 $2.7 million 公允价值收益。这本身是间接证据,说明 Velaura 估值随融资轮上升;也从独立、经审计来源印证了公司自己的独角兽估值主张,而不只是媒体报道。[CI033, CI010, CI024, CI034, CI001, CI002]
| 账上现金 | 月度烧钱 | 现金跑道(月) | 募资用途 | 下一轮触发因素 | 债务 / 项目融资义务 |
|---|---|---|---|---|---|
| 扩大工程和面向客户团队;深化战略伙伴关系(据 Series A 公告) | 未披露;可能取决于超大规模云厂商设计定点能否转化为已签署版税合同 | 已审阅来源均未披露 |
本表刻意只保留一行,因为 Velaura 仅披露募资用途;账上现金、烧钱速度、现金跑道以及任何债务 / 项目融资义务都没有公开信息。完整融资轮时间线(四轮分别为 $81M/$80M/$153M/$110M)见“公司概况”;本表只关注前瞻资本充足性输入。
[CI010, CI013, CI033, CI034]基于监管文件的 MARA 与 Velaura 在两个报告期内的现金流关系,是目前唯一可独立验证的资本强度数据点。
这张图只反映 MARA 投资人侧的会计披露;它不是 Velaura 自身的资本开支、库存或项目融资数据,这些仍未披露。
[CI001, CI005, CI006]4.5 财务结论:收入质量、利润率路径与尽调阻塞点
截至运行日期,Velaura AI 的财务画像最准确的描述是证据稀薄、但方向上有交叉印证:唯一独立申报数据点(MARA 的 SEC 披露投资及相关公允价值收益)支持公司估值上行叙事;但所有能让尽调团队评估收入质量、利润率路径或资本效率的经营指标——版税率、毛利率、烧钱速度、现金跑道、ARR、具名客户集中度——都没有披露。行业媒体也独立提示了与该结论直接相关的两项保留:TechBooky 警告,独角兽估值并不保证商业采用;Heise Online 指出,Velaura 拒绝披露芯片或客户名称,使未来任何版税流需要参照的收入基础本身变得模糊。考虑到同一板块 Groq 需要第二笔大额融资所释放的资本密集信号,也考虑到 Velaura 的版税经济性更可能类似 Arm 历史上较薄的单颗抽成(约 1-2% of ASP),而不是完整硬件销售利润率,核心财务尽调阻塞点不是 Velaura 技术是否可行,而是其授权收入能否足够快、也足够透明地规模化,在下一轮融资到来前支撑 $1 billion+ 估值。[CI031, CI032, CI035, CI026, CI025, CI021]
| 缺失的私有指标 | 影响 | 精确尽调路径 |
|---|---|---|
| 当前 ARR / 分业务收入运行率(AI 授权与旧硬件) | 无法评估新 Titan Core 业务的增长轨迹或收入质量 | 要求在保密协议(NDA)下披露分业务收入,或等待未来融资轮的 S-1 / 新闻稿披露 |
| 毛利率和销货成本(COGS)结构 | 无法建模盈利能力,也无法直接与无晶圆厂半导体可比公司对比 | 要求提供单位成本拆分或经审计财务报表节选 |
| Series A 后账上现金和月度烧钱 | 无法评估现金跑道,或公司在盈利前再融资的概率 | 要求提供最新现金流量表或银行余额证明 |
| 具名客户清单和单客户收入集中度 | 无法评估版税收入基数中的客户集中风险 | 要求提供客户参考或超大规模云厂商一侧确认 |
| 节电额版税率 / 测算方法 | 无法验证“已实现节电额”如何转换为实际应付美元金额,而这是整个收入模型的核心机制 | 要求提供标准授权合同条款清单(必要时隐藏客户身份) |
| 债务或项目融资设施 | 无法评估已披露股权轮之外的总资本结构或杠杆风险 | 要求提供股权结构表或债务明细表 |
本表列出本章研究识别出的高影响私有指标缺口;每一项都交叉引用到对应的证据缺口条目。
[CI018, CI013, CI027, CI021, CI009, CI034]4.6 图表
05产品与技术
5.1 产品定义与架构
Velaura AI 不卖成品芯片;它通过自有网站描述的三阶段合作模式交付物理设计 IP 和工具流产品:客户提交 RTL 设计,以及对面积、功耗和其他因素的优先级;Velaura 套用自研低电压 IP 库和为良率、可靠性构建的自研工具流;客户最终收到优化后的、特定节点 GDS 版图或 chiplet。架构上,Titan Core 瞄准 AI 训练和推理能耗占主导的矩阵乘法(MATMUL)运算,声称通过已在先进 3nm、2nm 工艺节点验证的自研电路和库技术,将这些运算所需能量降低 2-4x。集成设计目标是在较低工作电压下保持完整功能等价,不要求软件修改,因此 Titan Core 更像是可直接接入的物理设计优化,而不是客户必须全盘采用的新架构。Velaura 保留两条反映公司历史的产品 / 解决方案线:Ultra Low Power AI Compute(面向 AI 加速器的 Titan Core IP 平台)和 Blockchain(历史风冷、浸没式、水冷比特币挖矿硬件),底层低电压工程能力在两者之间共用。[CE001, CE005, CE006, CE007, CE008, CE030]
| 模块 / 资产 / 产品线 | 用户 | 状态 / 成熟度 | 差异化 | 尽调缺口 |
|---|---|---|---|---|
| Titan Core IP(数据中心) | 超大规模云厂商 XPU / 加速器设计团队 | 2026 年 3 月发布;与 4 家最大云厂商中的 3 家持续接洽(未具名) | 声称 MATMUL 能效提升 2-4x;方法论在 30M+ 旧 ASIC 上验证 | 仍无具名客户或独立基准测试 |
| Titan Core IP(物理 / 具身 AI) | 机器人、无人机和自主机器 OEM | 作为目标业务营销;未披露具名客户或已部署设备数量 | 以电池续航 / 热包络能效为卖点,区别于数据中心叙事 | 没有任何已签署 Physical AI 设计定点证据 |
| Titan Core IP(边缘 AI) | 消费电子 / IoT 设备 OEM | 公司网站将其列为目标业务;三项中细节最少 | 定位于更长电池续航和本地推理 | 三个业务中证据最少;未披露产品细节 |
| Teraflux 比特币挖矿硬件(旧业务) | 比特币挖矿运营商(如 MARA) | 成熟且已规模出货(累计 30M+ ASIC);根据 2026 年品牌重塑,该业务正在收缩 / 转向自营挖矿 | 9.8 J/TH 能效,最高 600 TH/s,风冷 / 浸没式 / 水冷版本,4nm 制程 | 该产品线仍贡献多少持续收入或支持不清楚 |
| 低电压 IP 库和工具流(底层平台) | Velaura 内部工程团队使用;通过接洽模式交付给客户 | 核心且经量产验证的技术,复用于所有产品线 | 自研工具流服务于良率 / 可靠性;团队有低电压定制电路经验 | 没有公开技术白皮书详述工具流内部方法论 |
数据中心牵引力主张证据最强(公司新闻稿 + 独立行业媒体);截至运行日期,Physical AI 和 Edge AI 仍处营销阶段,未披露客户或部署设备。
[CE001, CE004, CE005, CE018, CE032]分层展示 Velaura 技术栈:从客户 RTL 输入到交付 chiplet/GDS 输出。
[CE001, CE007, CE008, CE030]5.2 部署、工作流与关键依赖
超大规模云厂商硅团队若不授权 Titan Core,也可以用 Cadence 的 Cerebrus AI Studio 或 Synopsys 的 DSO.ai 等商业 EDA 工具在内部追求类似能效提升;截至 2026,这两者都提供 AI 驱动的低功耗设计自动化。因此,Velaura 的价值主张取决于其预构建库和工程专业能力,能否胜过资源充足的内部团队用通用工具能做到的水平。客户一旦启动合作,Velaura 交付还依赖几项不在其直接控制内的外部条件:客户愿意在 NDA 下分享自有 RTL;先进节点晶圆厂产能(如 TSMC 或 Samsung Foundry 等伙伴)可用于实体实现超低电压设计;Velaura 自己专门的低电压电路工程团队——规模相对小、难以复制——能够跑通物理设计工具流。值得注意的是,Samsung Catalyst Fund 既是 Velaura 投资人,又是 Samsung 的晶圆厂关联风险投资臂,这创造了一个合理但未披露的战略晶圆厂关系,值得单独尽调。客户侧,Velaura 自有材料估计典型数据中心每年最多可节省 $100 million 电费;这个数字明确建立在若干假设之上(10 cents/kWh 电价,以及 100,000 到 250,000 颗芯片下每颗每年约 $433.30),而不是客户特定的实测结果。相比财务章节中完全未披露的版税率方法,公司对这些假设的透明度明显更高。[CE003, CE023, CE026, CE022]
| 用户任务 | 当前工作流(现状) | 公司方案 | 可量化收益 | 限制 |
|---|---|---|---|---|
| 超大规模云厂商芯片团队想降低 AI 加速器功耗 | 用商用 EDA 工具(Cadence Cerebrus、Synopsys DSO.ai)做内部低功耗设计,或接受现有 GPU 能效路线图 | 授权 Titan Core IP;提交 RTL 设计和优先级;获得优化后的 GDS / 小芯片 | MATMUL 能效最高提升 2-4x;3 年每颗 XPU 约节省 $1,300(公司估算) | 需要 12-24 个月集成 / 认证周期;版税测算方法未披露 |
| 数据中心运营商想在机群规模降低电力 / 冷却总支出 | 承受上升电力成本,或投资设施级冷却 / 供电升级 | 部署内嵌 Titan Core IP 的加速器,在全机群提高每瓦性能 | 公司称典型数据中心每年最高可省 $100M(公司估算,依赖假设) | 节省额估算依赖公司列明假设(10c/kWh、设备量),未独立验证 |
| 机器人 / 无人机 OEM 希望自主机器电池续航更长 | 从现有边缘 AI 芯片中选择,这些芯片主要针对成本或原始性能优化 | 集成基于 Titan Core 的 Physical AI 芯片,提高每瓦性能 | 公司声称其能效与数据中心产品相当;未披露客户特定收益数字 | 仍未披露具名 Physical AI 客户或实测收益 |
| 比特币挖矿运营商希望获得高效、美国制造的挖矿硬件 | 从亚洲供应商进口 Bitmain / MicroBT ASIC 矿机 | 从 Velaura 旧硬件线购买 Teraflux 风冷 / 浸没式 / 水冷矿机 | 9.8 J/TH 能效,最高 600 TH/s,美国本地支持,并享有海关 / 关税优势 | AI 转向后,该产品线的持续战略优先级不清楚 |
可量化收益来自公司披露估算,除非注明独立来源;具体美元或能效数字均未获第三方基准确认。
[CE001, CE003, CE006, CE019, CE023]| 层级 / 流程 / 组件 | 角色 | 依赖 | 风险 |
|---|---|---|---|
| 客户 RTL 设计输入 | 定义基准芯片逻辑和设计优先级(面积、功耗、性能) | 需要客户在保密协议(NDA)下共享自有 RTL | 客户不愿共享敏感 IP,可能限制接洽管线 |
| Velaura 低电压 IP 库 | 核心自研技术,可将 MATMUL 运算能耗降低 2-4x | 基于专利化的电路设计诀窍(具体内容未确认) | 没有独立基准验证其声称的能效倍数 |
| 自研工具流(物理设计方法论) | 将客户 RTL 和 Velaura 库转成优化后的物理版图 | 依赖 Velaura 经验丰富的低电压电路设计团队 | 对小型专门工程团队存在关键人依赖 |
| 先进制程节点准入(3nm/2nm) | 支撑超低电压设计的物理实现 | 依赖晶圆厂伙伴(如 TSMC、Samsung Foundry)支持这些节点 | 晶圆厂产能限制或制程节点延迟可能卡住交付 |
| 输出:优化后的 GDS 或小芯片 | 最终交付物由客户集成进自有芯片或系统 | 需要成功流片并通过认证 | 12-24 个月认证周期拉长客户量产时间 |
| 旧 Teraflux 硬件制造(4nm) | 比特币挖矿 ASIC 系统的实体生产 | 根据此前规避关税策略,采用美国本地制造 / 组装伙伴 | 旧硬件与新 AI IP 业务之间的持续产能分配不清楚 |
这张架构图综合了 Velaura 自身对接洽模式的描述,以及来自公司和独立媒体的制程节点、制造细节。
[CE001, CE007, CE008, CE020, CE021]超大规模云厂商或 OEM 客户如何接入 Velaura 并集成 Titan Core 的端到端流程。
[CE001, CE006, CE025]Velaura 产品交付依赖的关键供应商、平台和合作伙伴。
[CE001, CE007, CE029]5.3 成熟度、证据强度与跨板块差异化
Velaura 三个目标板块的证据强度差异很大。数据中心产品线证据最充分:底层低电压 IP 和工具流已在 30 million 多颗量产 ASIC 中验证(主要来自历史 Teraflux 比特币挖矿业务),给 Titan Core 提供了真实制造良率和可靠性基础;Moor Insights & Strategy 的独立分析师 Patrick Moorhead 对此给出正面评价,称该路径有潜力降低客户总拥有成本并缓解热约束。相比之下,Physical AI 和边缘 AI 板块仍停留在营销阶段:Velaura 的 Physical AI 页面和 Solutions 页面都没有点名具体机器人或边缘设备客户,没有披露已部署设备数量,也没有提供区别于数据中心叙事的技术规格。历史 Teraflux 硬件线则在三类业务中拥有最强的具名客户证据(MARA),并披露了技术规格(9.8 J/TH 效率、最高 600 TH/s、4nm 工艺、多冷却方式版本);不过,它比 Titan Core 当前目标的 3nm/2nm 节点落后一代工艺,2026 AI 转型后这条线的战略优先级也不清楚。本章审阅的任何来源中,都不存在对具体 2-4x MATMUL 效率主张或 $100 million/year 节省估计的独立、同行评审基准。[CE004, CE014, CE018, CE019, CE020, CE024]
| 日期 / 阶段 | 功能 / 里程碑 | 状态 | 含义 | 来源 |
|---|---|---|---|---|
| 2025 年及以前 | Teraflux 比特币挖矿 ASIC 产品线(风冷 / 浸没式 / 水冷,4nm) | 已规模出货(产品历史累计 30M+ ASIC) | 在 Titan Core 之前已建立制造和良率记录 | 来源:PR Newswire(SE021)、Hashrate Index(SE023)、ForkLog(SE024) |
| 2026-03-24 | Titan Core 芯片设计与 IP 平台发布 | 已发布;与超大规模云厂商 XPU 伙伴的接洽被称为“持续进行” | 首次公开披露聚焦 AI 的产品及其 3nm/2nm 验证 | 来源:PR Newswire(SE009)、GamesBeat(SE010) |
| 2026-03-25 | 公司从 Auradine 更名为 Velaura AI | 已完成 | 显示公司将 AI 芯片 IP 置于旧比特币挖矿硬件之上 | 公司网站(SE005)、前文来源 |
| 2026-08-18 | Series A 融资($110M),用于工程和面向客户团队扩张 | 已完成 | 获得资金,可扩大 Titan Core 工程能力和超大规模云厂商接洽能力 | Business Wire/Morningstar(SE020) |
| 2026-08(进行中)及以后 | 扩大超大规模云厂商和 Physical AI 接洽;未披露 2nm 之后带日期的功能路线图 | 进行中 / 细节未披露 | 投资者和客户无法验证具体未来发布时间表 | 公司网站(SE005)、Business Wire(SE020) |
没有来源披露当前 Titan Core 代际之后带日期、到功能层级的产品路线图;2026 年 3 月之后各行描述的是融资和扩张里程碑,而非具体技术发布。
[CE018, CE019, CE020, CE025, CE007]Velaura 三个目标细分市场及其传统硬件线的成熟度和证据强度。
[CE004, CE005, CE018, CE035]5.4 信任、合规与开发者信号缺口
Velaura 的公开信任面很薄:公司发布了一份通用隐私声明(版本日期为 2024 年 8 月,审阅版本仍使用此前 Auradine, Inc. 的法定名称),但审阅过的资料里没有披露专门的安全认证(如 ISO 27001 或 SOC 2),也没有披露安全 / 质量认证(如 AEC-Q100 或 ISO 26262;考虑到其物理 AI / 机器人布局,这类认证尤其相关)。作为一家硬件 IP 公司,Velaura 也没有软件公司常见的公开开发者表面:未找到 GitHub 仓库、Hacker News 讨论、Stack Overflow 标签,或专门围绕 Titan Core 的公开开发者论坛讨论。公司确实在 Lever 上保持活跃招聘,说明工程岗位仍在招人,但抓取到的内容没有披露具体岗位层面的技术栈细节。按照本技能对缺少公开开发者表面公司的处理指引,最接近的从业者社区代理信号来自更广泛的半导体行业场域:SEMI 2026 年 8 月的「AI Techniques in Semiconductor Manufacturing」研讨会,以及 2026 年更广的半导体会议日程(DAC、SEMICON WEST),显示 Velaura 所处的 AI 与硅结合这一通用主题有活跃从业者参与;但两个场域都没有提供 Velaura 专属的技术讨论。公司只在营销语言里把 IP 保护描述为「已获专利」技术;本章研究未能通过 USPTO Patent Public Search 或 Google Patents 确认任何具体专利号。[CE009, CE010, CE011, CE012, CE013, CE027]
| 控制 / 认证 / 质量指标 | 状态 | 范围 | 缺口 |
|---|---|---|---|
| 通用隐私声明 | 已发布(使用原 Auradine, Inc. 名称,2024 年 8 月) | 网站和产品数据处理做法 | 审阅版本截至运行日期尚未更新以反映 Velaura AI 品牌重塑 |
| 安全认证(如 ISO 27001、SOC 2) | 未披露 | N/A | 没有任何正式安全认证公开证据 |
| 汽车 / 功能安全认证(如 AEC-Q100、ISO 26262) | 未披露 | N/A | 鉴于 Physical AI / 机器人业务野心,该认证相关性高,但未发现认证 |
| 制造良率 / 可靠性记录 | 公司声称:30M+ ASIC 已以“世界级”良率和可靠性投产 | 主要是旧 Teraflux 硬件,未确认适用于 AI 特定 Titan Core 芯片 | 未披露经独立审计的良率或缺陷率数字 |
| 专利保护 | 公司称技术“已获专利” | 范围未说明 | 本章通过 USPTO 或 Google Patents 搜索未确认具体专利号 |
多数信任 / 质量控制要么未披露,要么只有公司说法支撑;本表更应被视作合规缺口图,而非认证清单。
[CE009, CE010, CE018, CE002, CE027]5.5 展示材料
06客户情况
6.1 客户基础分层
Velaura AI 已披露的客户基础可以清晰分成两个时代,证据质量差异很大。传统 Teraflux 比特币挖矿硬件业务声称累计服务 30 家或更多工业挖矿运营商、40 家或更多数据中心运营商;其中只有两家——MARA Holdings 和 Genesis Digital Assets Limited(GDA)——被单独点名,并能在公开资料中得到独立印证。相比之下,新的 Titan Core AI 芯片业务声称正在接触四大云服务商中的三家,但没有披露任何名称;另两个目标分部——物理 AI / 机器人 OEM 和边缘 AI 设备厂商——没有披露任何客户、已部署单元或案例研究。各分部的买方、用户和付款方角色差异明显:比特币挖矿客户由挖矿运营财务团队同时承担买方和付款方;而超大规模云厂商的 Titan Core 接触则会涉及基础设施 / 资本开支预算所有者与云硅设计团队用户分离。这个分层模式——传统且战略上被弱化的业务有强具名证据,新业务才是真正支撑估值逻辑却没有任何具名证据——是本章客户尽调的核心发现。[CU007, CU008, CU009, CU019, CU028, CU034]
| 细分市场 | 买方 / 用户 / 付款方 | 用例 | 规模 | 收入 / 战略价值 | 缺口 |
|---|---|---|---|---|---|
| 比特币挖矿运营商(旧 Teraflux) | 买方 / 付款方:挖矿运营财务;用户:矿机机群运营团队 | 部署实体 ASIC 矿机,用于生产比特币算力 | 累计 30+ 工业矿商、40+ 数据中心运营商(公司声称) | 历史硬件销售收入;Series B(2024)披露订单额 $80M | 声称的 30+ 客户中,只有 2 家(MARA、GDA)被单独具名且获独立佐证 |
| 超大规模云厂商(Titan Core) | 买方 / 付款方:超大规模云厂商基础设施 / 资本开支负责人;用户:云芯片设计团队 | 将能效 IP 授权用于 AI 加速器设计 | 据报道,4 家最大云厂商中有 3 家已接洽(未具名) | 是 $1B+ Series A 估值逻辑的核心,但完全未具名 | 具名客户为零;除公司 / 媒体复述外,没有独立确认 |
| Physical AI / 机器人 OEM | 买方 / 付款方:机器人研发和物料清单(BOM)预算负责人;用户:自主机器运营者 | 受电池续航和散热约束的具身 AI 计算 | 未披露客户或已部署设备 | 愿景型细分市场;没有收入证据 | 未识别到具名客户、已部署设备数量或案例研究 |
| Edge AI 设备厂商 | 买方 / 付款方:设备 OEM 部件预算负责人;用户:终端设备用户 | 消费 / IoT 设备上的本地节能 AI 推理 | 未披露客户或已部署设备 | 愿景型细分市场;四项中细节最少 | 证据最少的细分市场;未识别到产品细节或客户 |
细分市场来自 Velaura 自身产品线披露(旧 Teraflux 与覆盖数据中心、Physical AI 和 Edge AI 的 Titan Core);规模和具名客户证据质量差异很大,只有旧硬件业务拥有获独立佐证的具名客户。
[CU007, CU009, CU019, CU028, CU034]Velaura 四类客户的采用触点和扩张路径。
[CU030, CU001, CU023]6.2 具名客户证据:量产与试点
MARA Holdings 提供了 Velaura 最强的客户证据:一份 SEC 10-Q 文件独立确认,在截至 2025 年 3 月 31 日的季度,MARA 向 Velaura(当时名为 Auradine)预付 $22.3 million 用于产品采购,对应 $57.2 million 的未结余额;独立媒体报道还分别确认,MARA 2025 年上半年的 Teraflux 总采购达到 $73.3 million,并已在 2025 年 6 月 30 日前全部交付——这是监管文件印证的真实量产部署,而不只是新闻稿。问题在于,MARA 同时也是 Velaura 投资方,持有董事会席位和 $85.4 million 的投资账面价值,因此这条证据最强的关系并不是独立的、保持距离的客户背书。相比之下,Genesis Digital Assets(GDA)是真正独立的客户,未披露与 Velaura 存在投资方关系:2025 年 6 月,GDA 同意为其 Texas Glasscock County 的 40 MW 设施采购 1,000 台 Teraflux AT2880-277 风冷矿机,并有执行总裁 Abdumalik Mirakhmedov 具名、公开背书,且确认参与 ERCOT 的需求响应项目;但截至运行日期,该具体部署是否已在该设施进入全面生产,仍没有独立确认。专门看 AI 芯片业务,则完全没有具名客户:Velaura 自有材料和 GamesBeat 访谈报道都把三家未具名超大规模云厂商接触描述为处于技术评估和 RTL 基准测试阶段,而非已确认量产部署;Heise Online 的独立报道还明确指出,Velaura CEO 即使被 Reuters 直接追问,也拒绝透露任何一家名称。[CU001, CU002, CU003, CU004, CU005, CU006]
| 指标 | 数值 | 日期 | 来源 | 置信度 | 含义 | 缺失分母 |
|---|---|---|---|---|---|---|
| 工业挖矿客户汇总(Teraflux) | 30+ | 2025 | Hashrate Index (SU009) | 中 | 显示客户分布较广,但大多未具名 | 未披露可触达挖矿运营商总数 |
| 数据中心运营商汇总(Teraflux,水冷) | 40+ | 2025 | PR Newswire (SU010) | 中 | 暗示部署范围比两个具名账户更广 | 未按部署规模或地域拆分 |
| Series B 轮订单额 | $80 million | 2024-04 | Yahoo Finance (SU011) | 中 | 迄今唯一披露的汇总性近收入指标 | 未更新至 2025-2026;没有 AI 专属订单额数据 |
| MARA 2025 年上半年采购 | $73.3 million | 2025-06-30 | 来源:The Energy Mag(SU005)、AInvest(SU006) | 中 | 已披露的最大单笔具名交易金额 | 该交易占 Auradine 2025 年总收入比例未披露 |
| GDA 采购协议 | 1,000 units | 2025-06-23 | PR Newswire (SU001) | 高 | 按单位数计的第二大具名交易 | GDA 交易金额未披露 |
| 超大规模云厂商接洽(Titan Core) | 四大云服务商中的 3 家(未具名) | 2026-08-18 | Business Wire/Morningstar (SU012) | 中 | AI 芯片估值逻辑的核心 | 未披露单位数、金额或名称 |
所有汇总数字均为公司披露或新闻稿转述的估计值;只有 GDA 和 MARA 两个条目有具体、独立报道的交易细节。
[CU007, CU008, CU001, CU004, CU009]| 客户 | 细分市场 | 部署 / 用例 | 生产部署 / 试点 | 结果 | 局限 |
|---|---|---|---|---|---|
| MARA Holdings(Marathon Digital Holdings,客户 / 投资方) | 比特币挖矿运营商(传统 Teraflux)+ 投资方 | 2025 年上半年采购 $73.3M Teraflux 矿机;SEC 文件披露 $22.3M 预付款 / $57.2M 余额 | 生产部署(媒体称订单已于 2025 年 6 月 30 日前全部交付;SEC 文件确认真实现金流活动) | 独立佐证的最大交易金额;唯一有 SEC 文件支撑的客户关系 | MARA 同时是 Velaura 投资方并占董事会席位,作为客户背书的独立性受限 |
| Genesis Digital Assets Limited(GDA,客户) | 比特币挖矿运营商(传统 Teraflux) | 1,000 台 AT2880-277 风冷矿机,用于 Texas 州 Glasscock 县 40 MW 数据中心,2025 年 6 月 | 公告为采购协议;截至报告运行日,特定设施的生产状态尚未得到独立确认 | 具名高管证言(Abdumalik Mirakhmedov);参与 ERCOT 需求响应计划 | 独立、保持距离交易的客户(无投资关系),但部署完成情况尚未独立核验 |
| 四大云服务商中的三家(未具名) | 超大规模数据中心(Titan Core / AI 芯片) | 据公司及媒体报道,双方持续技术接洽并开展 RTL 基准评估 | 根据 GamesBeat 采访报道,处于试点 / 评估阶段;尚未确认生产部署 | 未披露;完全依赖公司和媒体表述 | 无名称、无单位数、无任何独立确认 |
只有两个客户(MARA、GDA)被单独具名并获独立佐证,且都属于传统 Teraflux 硬件业务;未具名超大规模云厂商这一行用于呈现 AI 芯片业务当前(弱得多)的证据状态,标记为 coverage=partial,因为 Velaura 关于「30+ 家矿商、40+ 个数据中心」的汇总说法没有在这些条目之外逐项拆分。
[CU001, CU002, CU004, CU005, CU009, CU010]从发现到扩张的漏斗,展示每个已披露客户关系目前所处阶段。
百分比是示意性漏斗阶段估计,反映 MARA、GDA 及未具名超大规模云厂商接洽中披露的大致模式;不是公司披露的转化率。
[CU018, CU023, CU017]Velaura 三个最具体客户关系的证据质量、独立性与生产成熟度。
[CU006, CU011, CU023, CU036]6.3 留存、耐久性与重复使用
没有公开资料披露任何 Velaura 客户关系的净留存率、总留存率、流失率、续约率或客户满意度指标,无论传统业务还是新业务都没有。唯一可用的重复接触信号是定性的:MARA 与 Velaura 的关系从 Series B/C 财务投资方,发展到 2026 年 Series A 投资方,再到披露的 $73.3 million 硬件采购方,呈现出先落地再扩张的关系演进;但这不是正式留存指标,而且本身被投资方关系所干扰。2026 年 3 月更名后,Velaura 自身表述显示,即使公司把未售库存转向自营挖矿、不再继续对外销售,既有 Teraflux 客户仍保留保修支持——这说明公司继续履行对既有客户的义务,但也说明这条产品线的新客户获取正在下降。没有任何资料为超大规模云厂商 Titan Core 接触提供耐久性信号(队列、续约或留存数据);这些接触仍在评估阶段,客户关系还不够久,无法衡量留存。[CU016, CU030, CU033, CU013, CU029]
| 指标 | 数值 / 空值 | 细分市场 | 置信度 | 尽调请求 |
|---|---|---|---|---|
| 净 / 总收入留存率 | 全部细分市场 | 低 | 若内部跟踪,请直接向 Velaura AI 索取留存指标 | |
| 合同续约率 | 全部细分市场 | 低 | 索取传统 Teraflux 硬件客户群的续约率数据 | |
| 重复采购证据(MARA) | MARA 从 Series B/C 轮投资方,延伸为 2026 年 Series A 轮投资方,再到 2025 年上半年 $73.3M 硬件采购方 | 比特币挖矿(传统业务) | 中 | 确认 MARA 自 2025 年上半年以来是否追加订单 |
| 客户满意度 / NPS | 全部细分市场 | 低 | 索取任何客户满意度调查数据或第三方评价网站痕迹 | |
| 更名后的保修 / 支持连续性 | 公司称,尽管业务转向自营挖矿,现有 Teraflux 客户仍保留保修支持 | 比特币挖矿(传统业务) | 中 | 确认仍向传统客户提供的具体保修条款和支持 SLA |
留存和满意度指标几乎完全未披露;唯一的定性重复接洽信号(MARA 关系扩大)不是正式留存指标。
[CU016, CU030, CU033, CU013]无法制作示意性时间序列留存视图;这张队列图用唯一有多期交易数据的 MARA 关系记录披露缺口。
这不是真正的留存队列(百分比只是占位,用来表示关系跨年延续,不是测得的收入留存);没有公司披露实际留存指标,也没有其他客户有多期公开数据可填充更多行。
[CU030, CU006]6.4 扩张、集中度与反向市场信号
本章识别出的最大集中度风险是:Velaura 价值最高、证据最强的客户交易(MARA 2025 年上半年的 $73.3 million 采购,由 SEC 文件印证)来自关联方,而非独立买家;唯一真正独立的具名客户(GDA)则只代表一笔交易,没有确认重复业务。更不利的是,Velaura 唯一目前有收入证据的业务——传统 Teraflux 硬件——所在可服务市场正在结构性收缩:Cipher Mining、Core Scientific、TeraWulf、Bitdeer 等主要上市比特币矿企正在把资本和电力容量转向 AI/HPC 托管交易,而不是扩大比特币挖矿硬件机群;随着 AI 收入在全行业加速,上市矿企在 2026 年合计流失 21% 的比特币算力。这给 Velaura 造成艰难的转型窗口:传统客户基础正在收缩,而新的 AI 芯片客户基础(三家未具名超大规模云厂商)仍完全未获确认。在审阅资料中,未发现任何业务线存在部署失败、客户投诉或负面评价的证据,这是一个轻微正面信号;但也要放在更大的背景里看:Velaura 完全没有任何独立客户评价平台(G2、Gartner Peer Insights)覆盖。[CU006, CU020, CU014, CU015, CU035, CU026]
| 扩张驱动因素 | 集中度风险 | 影响 | 尽调路径 |
|---|---|---|---|
| MARA 关系从投资方扩展为大型硬件采购方,并继续参与 2026 年 Series A 轮 | MARA 同时是 Velaura 已披露价值最高的客户交易方,也是关联方投资人 / 董事会成员 | 缺乏独立性削弱了背书质量;真实独立需求信号不清楚 | 寻找与 Velaura 无投资关系的独立客户背书 |
| GDA 是真正独立(非投资方)的具名客户 | 单笔交易关系(1,000 台);无重复订单证据 | 正向但证据薄;尚不能证明持久重复需求 | 跟踪 GDA 是否追加订单或提供后续案例研究 |
| 若转化,Titan Core 与超大规模云厂商的接洽可能扩展为 Velaura 最大收入板块 | 三项接洽均未具名,且未确认为生产部署 | 影响最高的扩张机会也是证据最少的一项 | 索取具名客户确认或超大规模云厂商侧的独立披露 |
| 更广泛的比特币挖矿行业转向 AI/HPC 托管 | 挤压剩余 Teraflux 硬件销售的可触达市场;这是 Velaura 目前唯一产生收入且有具名证据的客户板块 | 在 AI 芯片业务仍未证实时,削弱传统业务收入的持久性 | 关注 Velaura 是否披露 AI 专属订单额,以抵消传统业务下滑 |
本表凸显的核心集中度风险是:Velaura 证据最强的客户关系(MARA)并不独立;最独立的具名客户(GDA)只是单笔交易;价值最高的未来板块(超大规模云厂商)没有任何具名客户。
[CU006, CU020, CU030, CU014, CU015, CU035]6.5 展示材料
07风险
7.1 按严重度排序的风险概览
本章识别出五类会实质影响 Velaura AI 投资判断的风险。按发生可能性和严重度综合排序,最高优先级风险是客户 / 收入集中:$1B+ Series A 估值几乎完全依赖一个没有任何具名客户的 AI 芯片授权业务,而唯一获得独立印证的客户关系(MARA)同时又是关联方投资人。紧随其后的是来自商业 EDA 工具(Cadence、Synopsys)的商品化风险,这些工具可能让超大规模云厂商客户在内部复现 Velaura 的能效收益。美国—中国半导体出口管制持续演变,监管风险重要但间接;风险通过 Velaura 未披露的代工厂和超大规模云厂商关系传导,而不是来自直接、已确认的敞口。运营风险集中在 Velaura 传统比特币挖矿硬件客户基础的结构性收缩:行业转向 AI/HPC 托管,正在拿走公司唯一目前有收入证据的业务,而新业务仍未被证明。财务 / 模型风险来自现金余额、烧钱速度、现金跑道和版税计量方法完全未披露。每一类风险都对应不同投资含义:考虑到底层证据大量回溯到公司说法而非独立确认,尽调不应把 Velaura 的独角兽估值当成商业执行已降险的证明。[CR030, CR016, CR002, CR013, CR019, CR009]
7.2 监管与法律风险
截至 2026 年 1 月 15 日,美国针对先进计算芯片的出口管制政策,从对出口至中国和澳门预设拒绝,转向逐案牌照审查,并为受覆盖产品增加认证和第三方测试要求。Velaura 自身披露没有说明其对该制度的敞口,但传导渠道真实存在:如果三家未披露的超大规模云厂商接触中有任何一家涉及服务中国的基础设施,或涉及与中国相关的代工产能,这些规则会直接影响交易经济性和时间表。另一个维度是,AI 相关专利诉讼已在全行业激增;法律行业分析统计,过去五年全球 AI 专利诉讼超过一千起。截至运行日期,没有诉讼或专利纠纷专门点名 Auradine 或 Velaura AI;但公司自己未经验证地声称拥有「已获专利」技术,却没有通过 USPTO 或 Google Patents 确认具体专利号,这本身就是一个尽调缺口,使 IP 强度无法验证。本章未完成对 Velaura AI, Inc. 的 Delaware 法律实体状态的独立确认(州注册系统需要付费直接查询),也未发现公司或其具名管理层存在制裁名单或被拒方敞口。[CR001, CR002, CR003, CR004, CR005, CR006]
| 规则 / 许可 / 案件 | 管辖区 | 状态 | 可能性 | 严重性 | 缓释措施 | 剩余风险敞口 | 尽调路径 |
|---|---|---|---|---|---|---|---|
| 美中先进芯片出口管制许可审查政策(BIS,2026 年 1 月 15 日生效) | 美国 / 中国 / 澳门 | 已生效;逐案许可审查取代推定拒绝 | 中 | 高 | Velaura 未披露具体措施;假设采用行业通用合规做法 | 高 - Velaura 的代工厂以及任何中国相关超大规模云厂商敞口未量化 | 索取 Velaura 的出口管制合规立场和中国敞口评估 |
| AI 相关专利诉讼行业激增(5 年内 1,000+ 起诉讼) | 美国(为主) | 行业趋势持续;没有案件具体点名 Velaura | 中 | 中 | 公司声称拥有专利技术,但未披露具体授权专利 | 中 - 未核实的 IP 保护本身可能成为目标或未来纠纷来源 | 通过 USPTO / Google Patents 搜索确认具体专利申请 |
| Velaura 关于「专利技术」的说法缺少已确认的具体专利号 | 美国 | 截至报告运行日未核实 | 低 | 中 | 未发现 | 中 - 没有已确认专利,无法完成 IP 强度尽调 | 直接向 Velaura AI 索取专利号 |
| 法律实体状态核验(Delaware 注册) | 美国 Delaware 州 | 隐私声明称公司在 Delaware 注册;本章未完成独立登记确认 | 低 | 低 | 可查询标准公开登记库(完整详情需付费) | 低 - 常规核验缺口,并非已识别问题 | 完成一次付费 Delaware Division of Corporations 实体搜索 |
| Physical AI / 机器人功能安全认证(如 ISO 26262、AEC-Q100) | 美国 / 全球 | 已审阅的 Velaura 材料均未提及 | 中 | 中 | 未披露 | 中 - 如果客户要求且缺失,可能阻碍 Physical AI 设计中标 | 索取 Velaura 面向 Physical AI 板块的认证路线图 |
| 制裁 / 拒绝交易方名单敞口 | 美国 / 全球 | 已审阅来源未发现敞口 | 低 | 低 | N/A | 低 - 缺少反向证据,不代表确认清白 | 直接对照 OFAC/BIS 拒绝交易方名单筛查 Velaura AI 和具名领导层 |
各行按严重性(高到低)排序。截至报告运行日,没有诉讼、执法行动或制裁匹配明确点名 Velaura AI 或 Auradine;风险主要是公司尚未公开回应的行业层面敞口。
[CR001, CR002, CR003, CR004, CR005, CR008]7.3 运营与合作伙伴 / 依赖风险
Velaura 最紧迫的运营风险是商品化:截至 2026 年,Cadence 的 Cerebrus AI Studio 和 Synopsys 的 DSO.ai 都已提供 AI 驱动的低功耗设计自动化;任何超大规模云厂商客户原则上都可以尝试用商品化工具在内部复现 Titan Core 的效率收益,而不是向 Velaura 授权 IP。第二项运营风险是 Velaura 传统 Teraflux 硬件客户基础的结构性收缩:随着 AI 收入加速,上市比特币矿企在 2026 年合计流失 21% 的算力;据报道,约 70% 的头部矿企正靠 AI 相关收入撑过当前熊市,这直接压缩了新增挖矿硬件采购需求。Velaura 更名后把未售库存转向自营挖矿,也说明公司意识到了这种下滑。在合作伙伴依赖侧,有两组关系带来异常高的集中度风险:MARA Holdings 既是 Velaura 最大、可独立印证的客户交易来源,又是持有董事会席位的关联方投资人;Samsung Catalyst Fund 既是投资方,也隶属于 Samsung,其代工业务可能是 Velaura 能效 IP 的制造伙伴,也可能成为竞争对手。整条超大规模云厂商接触管线——未来任何 AI 芯片收入的来源——仍未具名、未确认,是本章识别出的单一最高严重度依赖。[CR016, CR013, CR014, CR015, CR011, CR012]
| 失效模式 | 可能性 | 严重性 | 缓释成熟度 | 剩余风险敞口 | 未解决缺口 |
|---|---|---|---|---|---|
| 商业 EDA 工具(Cadence、Synopsys)把 Velaura 核心低功耗 IP 差异化变成更容易替代的能力 | 中 | 高 | 低 - 除声称的 IP 保护外,未披露防御策略 | 高 | Velaura 的库是否能明显优于用当前 EDA 工具可达到的最佳内部结果 |
| 传统 Teraflux 硬件业务可触达市场继续收缩,因为比特币矿商转向 AI/HPC 托管 | 高 | 中 | 中 - 公司已将库存转向自营挖矿 | 中 - 传统业务下滑是已知且被管理的转型,不是意外 | 新 AI 芯片收入能否在相近时间表上抵消传统业务下滑 |
| 可比低功耗 AI 芯片初创公司 Untether AI 于 2025 年关闭并申请破产 | n/a(已发生在同行身上) | 高 | 低 - 没有 Velaura 专属缓释措施直接回应这一行业先例 | 中 - 说明即便技术可信,公司仍面临全行业商业化风险 | Velaura 更强的量产记录(30M+ 颗 ASIC)是否足以与 Untether AI 的失败模式拉开差异 |
| Velaura 公开员工数估计差异很大(62 vs. 100-150) | n/a(持续存在的数据质量问题) | 低 | N/A - 不是公司造成的风险,而是披露 / 跟踪限制 | 低 | 真实当前员工数和组织规模 |
| 先进 3nm/2nm 节点的代工产能约束或工艺节点延迟 | 中 | 中 | Velaura 未具体披露 | 中 - 瓶颈不在 Velaura 直接控制范围内 | Velaura 的具体代工伙伴关系和任何产能承诺 |
| 未发现环境、安全或产品召回事件 | 低 | 低 | N/A - 缺少反向证据 | 低 | 是否存在未在媒体报道中浮出的非公开事件 |
各行按严重性(高到低)排序;多行记录的风险,其可能性 / 严重性判断受限于缺少 Velaura 专属披露,而不是已确认风险较低。
[CR016, CR013, CR014, CR015, CR011, CR012]| 依赖项 | 交易对手 | 角色 | 集中度 | 失效情景 | 严重性 | 缓释措施 | 剩余风险敞口 |
|---|---|---|---|---|---|---|---|
| 具名客户关系 | MARA Holdings | 投资方(董事会席位、$85.4M 账面价值)和客户(2025 年上半年 $73.3M 硬件采购) | 高 - MARA 同时是 Velaura 单笔最大且获独立佐证的交易,也是关联方 | MARA 缩减或退出其投资 / 采购关系 | 高 | 未具体披露;MARA 继续参与 2026 年 Series A 轮是正向连续性信号 | 高 - 失去 MARA 会拿掉 Velaura 唯一由 SEC 文件佐证的收入数据点 |
| 代工产能 | 先进节点代工厂(如 TSMC、Samsung Foundry) | 在 3nm/2nm 上物理实现超低电压设计 | 未知 - 具体代工关系未披露 | 代工产能约束、涨价或地缘政治扰动推迟 Titan Core 交付 | 中 | 未具体披露 | 中 - 不在 Velaura 直接控制范围内 |
| 具有代工关联的战略投资方 | Samsung Catalyst Fund(Samsung Electronics 风险投资部门) | 投资方;母公司运营竞争 / 合作型代工业务 | 中 - Velaura 未回应潜在利益冲突 | Samsung 的代工业务可能与 Velaura 竞争,或限制其独立性 | 中 | 未披露 | 中 - 关系条款未披露 |
| 超大规模云厂商接洽管线 | 四大云服务商中的 3 家(未具名) | 潜在 Titan Core 授权客户 | 高 - 整个 AI 芯片收入逻辑依赖未具名、未确认关系 | 三项接洽中的任意或全部未能转化为签署许可 | 高 | 未披露;仅称为「持续接洽」 | 高 - 核心估值逻辑没有具名 / 已确认客户支撑 |
| EDA 工具体系 | Cadence、Synopsys | 为潜在 Velaura 客户提供替代性的内部低功耗设计能力 | 中 - 不是直接供应商依赖,而是替代技术风险 | 超大规模云厂商选择基于内部 EDA 的设计,而不是授权 Velaura IP | 中 | 未披露 | 中 - 如运营风险登记表所记录,存在商品化风险 |
各行按严重性(高到低)排序;MARA 的投资方 / 客户双重角色,以及完全未具名的超大规模云厂商管线,是已识别的两项最严重依赖风险。
[CR007, CR026, CR033, CR030, CR016]支撑 Velaura 风险画像的关键伙伴、供应商和融资依赖。
[CR033, CR026, CR007, CR016, CR001]7.4 财务 / 模型风险与人员 / 执行风险
没有公开资料披露 Velaura 在 2026 年 8 月 Series A 后的现金余额、月度烧钱速度或现金跑道,也没有披露版税收入模型背后的计量方法——这是财务 / 模型风险;Groq 这家资金充足的独立 AI 芯片竞争者,在此前完成 $750 million 融资后,同月仍需要再融一笔大额资金,说明该赛道持续资本密集,进一步放大这一风险。本章没有发现 Velaura 存在欺诈、重述或会计异常的证据,也没有发现 MARA 披露其 Velaura 投资时存在此类问题;MARA 采用的 ASC 321 公允价值处理符合标准股权投资会计实践,专门在财务诚信风险上构成缓释信号。人员 / 执行风险方面,Velaura 的公开叙事高度集中在 CEO Rajiv Khemani 和 CDO Manu Gulati 身上,对其他高管团队成员的独立履历细节很少,形成显著关键人物依赖;2025-2026 年未发现高管离职或辞任,这是连续性的缓释信号;董事会监督(包括 Intel CEO Lip-Bu Tan 和 MARA CEO Fred Thiel)提供治理可信度,部分抵消了运营层面的关键人物集中。[CR019, CR027, CR035, CR038, CR024, CR025]
| 角色 / 职能 | 依赖或缺口 | 可能性 | 严重性 | 缓释措施 | 尽调路径 |
|---|---|---|---|---|---|
| CEO(Rajiv Khemani) | 所有融资 / 产品公告中唯一的公开发声者;公司叙事高度依赖个人 | 低(未发现离职信号) | 高 | 多轮融资均在稳定领导层下完成;董事会包括资深风投和行业人士 | 索取关键人物保险条款及任何雇佣 / 留任协议细节 |
| CDO(Manu Gulati) | 与 Khemani 并列为具名联合技术负责人;曾有 NUVIA 联合创始人经历 | 低(未发现离职信号) | 高 | Mayfield 分别与 Khemani / Gulati 第四次、第二次合作,显示关系受信任且持续 | 索取 Gulati 专属留任 / 归属条款 |
| 更广泛高管梯队(Gupta、Kim、Vehling、Campbell、Dhablania) | 头衔之外,独立传记细节很少 | 中(数据可得性缺口,并非已确认风险) | 中 | 未具体披露 | 索取非创始人高管的组织厚度和继任规划细节 |
| 董事会监督 | 包括 Intel CEO Lip-Bu Tan、MARA CEO Fred Thiel,以及多名风投合伙人 | 低 | 低 | 强治理可信度在监督层面部分缓释创始人关键人物风险 | 确认董事会会议频率和正式监督机制 |
| 人员规模 / 组织规模 | 公开估计分歧(62 人 vs. 100-150 人) | n/a(数据质量缺口) | 低 | N/A | 直接向 Velaura AI 索取当前员工数 |
行按严重程度从高到低排序;已审阅来源中未发现高管离职或辞任,这是缓释信号,但除创始双人组外,公开履历深度仍薄弱。
[CR024, CR025, CR037, CR018]7.5 缓释因素、监控指标与打破投资逻辑的触发点
Velaura 用来抵御纯执行风险质疑的最强缓释因素,是其 30 million+ ASIC 量产历史;它能回应制造良率担忧,但不能解决本章另行识别的客户集中、披露或商品化风险。本章定义了六个尽调团队应跟踪的、可监控的投资逻辑破裂触发点:Series A 交割后 12-18 个月内仍没有具名超大规模云厂商设计定点;主要 EDA 厂商或代工伙伴发布任何竞争性低功耗 IP 库;传统硬件订单继续同比下滑且没有 AI 芯片收入抵消;MARA 披露的投资或采购关系出现任何恶化;12-18 个月内再次融资且未披露收入拐点;任一创始联合创始人离职。截至运行日期,这些触发点都尚未触发;但每一项都是具体、由来源支撑的事件,尽调应持续监控,而不是把当前估值视为已定局。[CR029, CR031, CR032, CR030, CR028]
| 风险 | 可监测触发项 | 阈值 / 事件 | 行动含义 |
|---|---|---|---|
| 无具名 AI 芯片客户 | 具名超大规模云厂商设计定点的公告(或确认缺席) | 2026 年 8 月 Series A 完成后 12-18 个月内没有具名设计定点 | 重新评估估值论点;将持续不透明视为商业转化的负面信号 |
| EDA 厂商 / 晶圆厂使能力商品化 | Cadence、Synopsys 或晶圆厂伙伴发布直接竞争的低功耗 IP 库或设计套件 | 任何等效、可自由授权低功耗库的公开发布 | 重新评估 Velaura 的差异化和版税率防御力 |
| 传统硬件业务下滑 | 季度 Teraflux / 传统硬件订单趋势 | 已披露订单持续同比下滑,且未披露可抵消的 AI 芯片收入 | 视为艰难且缺乏收入支撑的转型期得到证实 |
| MARA 关系恶化 | MARA 未来 10-Q/10-K 中关于其 Velaura 投资和采购活动的披露 | MARA 降低持股、退出董事席位,或披露对 Velaura 投资减值 | 鉴于 MARA 在现有客户 / 投资人证据中权重过大,将其视为重大反向信号 |
| 资本充足性 | 后续融资公告或披露现金头寸 | Velaura 在未披露收入拐点的情况下,于 12-18 个月内再融资 | 表明版税收入尚未按预期兑现;重新评估资本效率 |
| 关键人物连续性 | 任何高管离职公告(尤其是 Khemani 或 Gulati) | 任一联合创始人离职 | 鉴于关键人物依赖集中,将其视为高严重性负面触发项 |
阈值和事件是作者基于可监测、来源支撑触发项的综合判断;均非公司披露的正式里程碑。
[CR031, CR032, CR030, CR019, CR027]本章识别的最高优先级风险在发生可能性、严重度和缓释成熟度上的分布。
[CR030, CR007, CR016, CR013, CR002, CR024]关键风险如何传导到收入、客户转化、融资和估值。
[CR009, CR016, CR013, CR019, CR028]7.6 展示材料
08估值
8.1 投资逻辑与反向逻辑
Velaura AI 的投资逻辑建立在三根支柱上:经量产验证的超低功耗硅 IP(已出货 30 million+ ASIC)、可信的领导层与董事会阵容(包括 Intel CEO Lip-Bu Tan 和 MARA CEO Fred Thiel),以及有利的宏观顺风——美国 9.3 GW 电力缺口和 2026 年超大规模云厂商 $600-725 billion AI 基础设施资本开支——这些因素提高了单位功耗性能技术的理论价值。反向逻辑同样具体:真正支撑估值的 Titan Core AI 芯片业务没有任何具名客户,没有披露任何版税率或收入数字;Heise Online 的独立报道确认,即使被直接追问,Velaura CEO 也拒绝透露三家声称接触的超大规模云厂商名称。更复杂的是,Arm——Velaura 自己明确拿来对标的商业模式可比公司——在 2026 年打破三十年纯 IP 授权模式,开始直接销售芯片;这个先例让基于版税的经济模型作为独立模式的长期耐久性受到质疑。本章总体判断是,投资逻辑和反向逻辑都有证据支撑;也正因如此,在现有信息下,不应给出笃定的买入或回避建议。[CV001, CV016, CV023, CV020, CV018]
| 论点 | 什么会改变判断 |
|---|---|
| 正方论点:Velaura 已量产验证的低功耗 IP(30M+ 颗 ASIC)叠加电力约束宏观顺风,足以支持继续跟踪其走向独角兽级 AI 芯片授权业务。 | 具名超大规模云厂商设计定点或已披露 AI 专项收入,会把这个论点从合理叙事推进为有证据支撑的牵引力。 |
| 正方论点:版税 + 授权模式虽然单台经济性弱于硬件销售,但资产更轻、毛利更高,具备类似 Arm 历史路径的扩张潜力。 | 披露实际版税率和至少一份已签授权协议后,才能做量化建模,而不是靠类比假设。 |
| 反方论点:估值所依托的 AI 芯片业务没有任何具名客户;所有牵引力说法都来自媒体转述的公司表述。 | 独立确认超大规模云厂商关系(客户推荐、云厂商侧披露或基准测试)会显著缓解这个担忧。 |
| 反方论点:可比公司波动极大(Groq ~50% 重置、Untether AI 破产、Cerebras IPO 带来估值暴涨),今天 $1B+ 的估值标记 12 个月后可能完全变样。 | 若下一轮融资估值稳定或更高,且没有降价轮调整,可缓解(但不能消除)这个担忧。 |
| 反方论点:Cadence、Synopsys 等商业 EDA 工具可能让超大规模云厂商在内部复制 Velaura 的效率提升,从而把核心 IP 商品化。 | 如果有证据显示 Velaura 的库显著优于内部 EDA 工具可达到的最佳结果(例如独立基准测试),可回应这个问题。 |
每条反方论点都配有一个具体、可监测事件,用于解决或缓解该问题,符合本章以证据变化为核心的判断方式。
[CV018, CV016, CV017, CV023, CV031]8.2 当前融资与估值背景
Velaura AI 于 2026 年 8 月 18 日完成 $110 million Series A,估值超过 $1 billion;本轮由 Seligman Ventures 领投,Capricorn Investment Group、Prosperity7 Ventures、Mayfield、Maverick Silicon、MARA、Premji Invest、Samsung Catalyst Fund 和 StepStone Group 参投。除「超过 $1 billion」这一标题外,没有资料披露准确投后估值、股价或优先权结构,也没有披露任何清算优先权、反稀释或期权池细节,足以让尽调区分普通股价值和优先股结构推高后的价值。估值标记上升的唯一独立申报印证来自 MARA Holdings 的 SEC 10-Q:当后续融资轮确立更高可观察价格时,MARA 对既有 Velaura 持股分别记录 $11.9 million 和 $2.7 million ASC 321 公允价值收益——这是真实、经审计的证据,说明 Velaura 估值在各轮之间上升;但它不能独立确认具体 $1B+ 数字。Velaura 的股权结构表混合了财务型风险投资方、战略投资方和关联方投资人(Samsung Catalyst Fund、MARA),增加了解读复杂度:MARA 上升的公允价值标记同时是 Velaura 估值提升的最佳证据,也是一项不应被视为完全独立的关联方数据点。[CV001, CV002, CV007, CV014, CV015]
8.3 可比公司集合与乐观 / 基准 / 悲观情景
本章可比集合里没有任何公司在阶段和商业模式上与 Velaura 精准匹配。最接近版税模式的上市可比公司 Arm Holdings,FY2026 收入 $4.92 billion、市值 $270.57 billion,对应约 52-55x 市销率;这反映的是 Arm 成熟、多元的规模,而不是可直接迁移到早期授权公司的基准。直接 AI 芯片赛道可比公司在一年内波动极大:Cerebras 2026 年 5 月以约 $56 billion IPO 验证了该分部的上行空间;Groq 则在一笔据报道 $20 billion NVIDIA 交易吸收其创始团队后,估值从 $6.9 billion 重置到 $3.5 billion;Untether AI 这家可直接对比的低功耗 AI 芯片创业公司,在仍在运营后约十六个月内破产。即使用典型半导体 IP 授权倍数的保守端(8x revenue)套到 Velaura 的 $1B+ 估值,也意味着需要约 $125 million 年收入才能支撑这个价格——这一数字没有任何公开证据支持。在这个背景下,乐观情景要求至少一家超大规模云厂商接触在 18 个月内转为具名设计定点;基准情景假设按典型 12-24 个月认证周期逐步、部分转化;悲观情景——考虑到可比集合已展示的波动,这是真实尾部风险——假设没有转化且传统业务继续下滑,可能迫使公司折价融资,甚至更糟。[CV003, CV004, CV005, CV009, CV008, CV010]
| 情景 | 假设 | 估值 / 回报逻辑 | 关键风险 | 概率信号 |
|---|---|---|---|---|
| 乐观 | 三个超大规模云厂商合作中至少一个在 18 个月内转为具名、已签约设计定点;传统硬件下滑企稳;电力约束宏观顺风持续利好效率型 IP。 | 收入拐点可能支撑下一轮加价,或走向类似 Cerebras ~$56B 结果的战略收购 / IPO 路径;但鉴于 Velaura 阶段更早,规模会小得多。 | 晶圆厂认证时间表执行风险;EDA 工具商品化;若只有一个设计定点落地,客户集中度会偏高。 | 已量产验证技术和强宏观顺风让该情景具备合理性,但目前没有任何具名客户证据支撑。 |
| 基准 | 逐步、部分转化(例如一个设计定点并伴随多年量产爬坡),同时传统业务继续下滑;Velaura 至少再融一轮,才能达到可自我维持的版税收入。 | 下一轮估值大概率大致持平到小幅上行,取决于可见管线进展,而非已确认收入。 | 如果收入可见度改善前就需要下一轮融资,会有融资依赖风险;持续披露不透明让估值证据仍然薄弱。 | 符合半导体 IP 通常 12-24 个月认证周期,也符合当前缺少已披露财务指标的状态。 |
| 悲观 | 三个超大规模云厂商合作在 18 个月内均未转化;随着比特币矿工转向 AI/HPC 托管,传统 Teraflux 业务继续萎缩;EDA 工具商品化削弱第三方 IP 紧迫性。 | 估值可能遭遇降价轮或减记,类似 Groq 在 NVIDIA 交易后约 50% 重置;最不利情况下,也可能走向 Untether AI 破产式更严重结果。 | 收入拐点前资本耗尽;关键人物离职;无法与内部 EDA 方案拉开差距。 | 鉴于 Untether AI、Groq 已显示板块波动,这是真实尾部风险;不过截至运行日,Velaura 本身尚未出现反向信号。 |
概率信号是作者基于引用证据作出的定性判断,不是统计模型概率。
[CV024, CV025, CV026, CV020, CV021, CV018]| 可比对象 | 指标 | 倍数 / 估值 / 状态 | 相关性 | 局限 |
|---|---|---|---|---|
| Arm Holdings(上市) | 市值 / FY2026 收入 | $270.57B 市值;$4.92B FY2026 收入;~52-55x P/S | 最接近的上市版税 + 授权商业模式可比对象 | Arm 是成熟、多元、已有数十年历史的上市公司;与 Velaura 的 Series A 阶段和规模不可比 |
| Cerebras Systems(上市,2026 年 5 月 IPO) | IPO 首日估值 | 完全稀释后约 $56B | 直接 AI 芯片板块可比对象;近期美国最大科技 IPO | Cerebras 销售成品晶圆级芯片(硬件销售模式),不是 Velaura 这种版税授权业务 |
| Groq(私营,NVIDIA 交易 + 重置融资) | 估值历史 | $6.9B(2025 年 9 月)-> ~$20B NVIDIA IP / 人才交易(2025 年 12 月)-> $3.5B 重置融资(2026 年 8 月) | 直接 AI 芯片板块可比对象,显示估值波动性 | Groq 商业模式已完全转向(“AI 推理新云”),与 Velaura 授权模式的直接可比性下降 |
| Tenstorrent(私营) | 估值状态 | 截至 2026 年中,未披露独角兽级估值重置 | 直接 AI 芯片加速器 / IP 板块可比对象 | 公开估值透明度有限;可能低估 Tenstorrent 的实际私募估值 |
| Untether AI(已停业) | 结果 | 破产(2025 年 10 月);负债 $128M+ | 对技术可信的低功耗 AI 芯片创业公司来说,是下行尾部可比案例 | Untether AI 销售成品芯片,而非授权 IP;商业模式不同于 Velaura |
| 半导体 IP 授权板块(汇总) | 典型 EV / Revenue 倍数 | 高质量、成长型公司 8-12x;增长较慢 / 商品化产品组合 4-8x(2026) | 基于版税的 IP 业务通用估值方法基准 | 板块汇总数据会掩盖公司层面的质量 / 增长差异;并非 Velaura 专属 |
列出的可比对象都不是 Velaura 在阶段和模式上的精确匹配;纳入它们是为了从不同角度框定合理估值区间(上市版税模式同业、IPO 同业、私营 AI 芯片同业、板块倍数基准,以及下行失败案例)。
[CV003, CV004, CV009, CV008, CV010, CV018]在一组可比收入倍数下,为支撑 Velaura $1B+ 估值所需的示意收入。
数值是示意性隐含收入(USD millions),用 $1 billion 估值除以各倍数得出;这些收入数字均未由 Velaura 披露或确认,Arm 的极端倍数反映其独特规模 / 增长画像,而不是典型可比公司。
[CV013, CV004, CV005]围绕 AI 芯片可比公司组已经展现的波动,框定 Velaura 下一轮融资或退出事件的乐观 / 基准 / 悲观估值区间。
这些区间是作者基于所引用可比公司组的历史估值摆动(Groq 下调、相对 Cerebras 的假设上调尺度)构建的示意情景边界,不是 Velaura 已披露的特定预测或模型。
[CV008, CV009, CV024, CV025, CV026]8.4 建议、退出就绪度与最终尽调问题
本章建议为「观察」(继续研究),置信度为中,风险评级为高,估值立场仍未解决——这不是泛泛的公司质量打分,而是因为支撑有把握、对价格敏感判断所需的收入、版税率和具名客户证据几乎完全缺失。核心估值输入缺失时,选择继续研究而不是虚假精确,符合本章逻辑。退出就绪度完全未经证明:没有披露任何老股交易、要约收购或 IPO/M&A 时间表;不过,Velaura 的董事会构成(Intel 的 Lip-Bu Tan、MARA 的 Fred Thiel,以及多位经验丰富的风险投资合伙人)提供了治理可信度,水平可比于曾帮助 Cerebras IPO 和 Groq 走向类似被收购结果的治理基础。按优先级排列,五个尽调问题最能高效化解本章不确定性:(1) 实际收入和版税率;(2) 至少一个具名超大规模云厂商客户引用;(3) 股权结构表和优先权条款;(4) 当前现金状况和现金跑道;(5) 对所称 2-4x 效率提升的独立技术基准测试。只要解决其中两三项,本章建议很可能就会明确转向买入或回避。[CV033, CV028, CV036, CV037, CV038, CV039]
| 触发项 | 阈值 | 对论点的传导 | 行动含义 |
|---|---|---|---|
| 无具名超大规模云厂商设计定点 | Series A 后 18 个月(即大约到 2028 年 2 月)仍无具名或独立确认客户 | 直接削弱估值所依托的核心 AI 芯片收入论点 | 若无其他新证据,将估值立场从“未解决”下调为“昂贵 / 回避” |
| 披露降价轮或估值减值 | 任何后续融资轮或 MARA SEC 文件显示 Velaura 估值标记低于 2026 年 8 月的 $1B+ | 直接表明市场已下调公司定价 | 视为本章提示的高估风险被证实 |
| EDA 厂商或晶圆厂商品化事件 | Cadence、Synopsys、TSMC 或 Samsung Foundry 公开发布直接竞争的低功耗 IP 库 | 削弱 Velaura 核心差异化和版税率防御力 | 重新评估护城河耐久性,并提高熊情景概率权重 |
| 传统业务继续下滑且无 AI 芯片抵消 | 已披露(或推断)传统硬件连续两个或以上季度下滑,且未披露 AI 芯片收入 | 证实基准 / 熊情景中的转型路径正在发生,且没有牛情景收入抵消 | 提高对熊情景资本耗尽风险的权重 |
| 关键人物离职 | CEO Rajiv Khemani 或 CDO Manu Gulati 离职 | 移除市场所押注的核心执行能力 | 视为高严重性负面估值触发项,需要立即重新评估 |
阈值是作者基于来源支撑、可监测事件综合设定;均非公司披露的正式契约或里程碑。
[CV034, CV035, CV023, CV025]| 主题 | 缺失证据 | 重要性 | 负责人 / 尽调路径 |
|---|---|---|---|
| 收入与版税率 | 任何已披露 AI 芯片收入数字或版税百分比 | 没有它,就无法用任何上市或私营可比对象计算估值倍数 | 在 NDA 下直接向 Velaura AI 管理层 / IR 索取 |
| 具名超大规模云厂商客户 | 至少一个已确认、具名的超大规模云厂商设计定点或客户推荐 | 这是整份报告中最关键的缺失证据 | 向 Velaura 索取客户推荐,或寻求超大规模云厂商侧独立确认 |
| 股权结构表与优先权条款 | 清算优先权层级、反稀释条款、期权池规模 | 用来判断 $1B+ 标题估值反映的是普通股价值,还是优先股结构安排 | 在 NDA 下索取资本结构表 |
| 现金头寸与资金跑道 | 当前账面现金、月度烧钱额,以及 Series A 后资金跑道 | 决定融资依赖度和近期追加融资概率 | 索取近期现金流量表或银行余额证明 |
| 独立技术基准测试 | 任何第三方对 2-4x MATMUL 效率提升主张的验证 | 决定核心技术差异化能否顶住 EDA 工具商品化风险 | 委托或跟踪独立半导体基准测试发布 |
这五项是杠杆最高的尽调追问;任何一项得到解决,都会实质改变本章建议的置信度。
[CV038, CV012, CV015, CV028, CV037]从量产规模与宏观顺风,到未解决的客户 / 收入证明,再到最终「观察」建议的逻辑链。
[CV041, CV033, CV020, CV012, CV031]IC 可直接使用的评分快照,覆盖市场、证明、护城河、经济性、风险、估值和证据质量。
[CV020, CV016, CV033]8.5 展示材料
免责声明
本报告是基于截至生成日期(2026-08-19)公开来源的尽调研究综合,不构成投资建议。Velaura AI 是一家公开财务披露有限的私营公司;多项关键输入(收入、版税率、现金状况、具名 AI 芯片客户)仍是未解决的证据缺口,已在报告各处说明。
证据索引
| 编号 | 陈述 | 可信度 | 来源 |
|---|---|---|---|
| CO001 | Velaura AI is headquartered in Silicon Valley, California. | 高 | SO003, SO025 |
| CO002 | Velaura AI was previously known as Auradine and rebranded to Velaura AI in March 2026. | 中 | SO015 |
| CO003 | Velaura AI's flagship product is Titan Core, a silicon design and IP platform rather than a standalone chip. | 高 | SO004, SO001 |
| CO004 | Titan Core delivers a 2-4x improvement in performance per watt for the matrix-multiplication (MATMUL) operations central to AI training and inference. | 高 | SO003, SO004 |
| CO005 | Titan Core can cut overall AI accelerator chip power by up to 2x, saving up to 500W on a typical 1000W-class GPU or XPU. | 高 | SO004, SO023 |
| CO006 | Velaura AI's underlying ultra-low-power design technology has been deployed in more than 30 million ASICs in production. | 高 | SO003, SO004 |
| CO007 | Velaura AI raised $110 million in Series A financing, announced on August 18, 2026. | 高 | SO003, SO005 |
| CO008 | The August 2026 Series A brought Velaura AI's valuation to more than $1 billion, giving it unicorn status. | 高 | SO003, SO007 |
| CO009 | Seligman Ventures led Velaura AI's $110 million Series A round. | 高 | SO003, SO005 |
| CO010 | Capricorn Investment Group and Prosperity7 Ventures joined Velaura AI's Series A as new investors. | 高 | SO003, SO013 |
| CO011 | Existing investors Mayfield, Maverick Silicon, MARA, Premji Invest, Samsung Catalyst Fund, and StepStone Group also participated in the Series A. | 高 | SO003, SO006 |
| CO012 | Rajiv Khemani is co-founder and CEO of Velaura AI. | 高 | SO002, SO003 |
| CO013 | Manu Gulati serves as Chief Development Officer of Velaura AI. | 中 | SO002 |
| CO014 | Velaura AI's leadership team includes executives and engineers from Apple, NVIDIA, Google, Qualcomm, and Marvell. | 高 | SO003, SO002 |
| CO015 | Velaura AI's board of directors includes Navin Chaddha (Mayfield), Umesh Padval (Seligman Ventures), Dipender Saluja (Capricorn), Lip-Bu Tan (Intel CEO / Walden International), Fred Thiel (CEO, MARA), and Sriram Viswanathan (Celesta Capital). | 中 | SO002 |
| CO016 | Velaura AI's business model charges an upfront license fee plus royalties tied to the power savings customers achieve, which Velaura's CEO compares to Arm's licensing model. | 中 | SO019 |
| CO017 | Velaura AI has ongoing engagements with three of the four largest cloud providers (hyperscalers), though it has not disclosed their names. | 高 | SO019, SO004 |
| CO018 | Velaura's technology has been validated for advanced 3nm and 2nm process nodes in partnership with hyperscaler XPU teams. | 高 | SO004, SO023 |
| CO019 | Velaura AI does not publicly disclose which specific chips or customers embed its Titan Core IP. | 中 | SO019 |
| CO020 | As Auradine, Velaura AI previously designed Teraflux bitcoin-mining ASIC systems used by MARA (Marathon Digital Holdings) and other mining operators. | 中 | SO015, SO016 |
| CO021 | Auradine raised $81 million in a Series A round in May 2023 led by Celesta Capital and Mayfield. | 中 | SO026, SO027 |
| CO022 | Auradine raised $80 million in an oversubscribed Series B round in April 2024. | 中 | SO026 |
| CO023 | Auradine raised $153 million in a Series C round in April 2025 led by StepStone Group. | 中 | SO026, SO028 |
| CO024 | Cumulative disclosed funding across Auradine/Velaura AI's Series A, B, C, and 2026 AI Series A rounds exceeds $424 million. | 中 | SO026, SO003 |
| CO025 | Intel CEO Lip-Bu Tan sits on Velaura AI's board of directors via Walden International. | 高 | SO002, SO019 |
| CO026 | Patrick Moorhead, founder and chief analyst of Moor Insights & Strategy, said Velaura's approach could reduce total cost of ownership and ease thermal limitations for AI infrastructure customers. | 中 | SO003 |
| CO027 | Umesh Padval of Seligman Ventures described Velaura AI as the firm's first investment in Physical AI. | 中 | SO003 |
| CO028 | Dipender Saluja of Capricorn Investment Group said Velaura is addressing AI's energy footprint challenge at the silicon level. | 中 | SO003 |
| CO029 | Navin Chaddha of Mayfield said the Series A is Mayfield's fourth partnership with Rajiv Khemani and second with Manu Gulati. | 高 | SO014, SO003 |
| CO030 | Velaura AI cites an IEA projection that global AI-related electricity demand will more than double to over 945 TWh/year by 2030. | 中 | SO004 |
| CO031 | Titan Core integration is designed as a drop-in retrofit that preserves a customer's existing RTL design and SoC architecture without requiring software changes. | 高 | SO004, SO023 |
| CO032 | Velaura AI estimates Titan Core can save customers approximately $1,300 in electricity costs over three years per XPU. | 中 | SO004 |
| CO033 | Trade press coverage frames Velaura's unicorn valuation and rapid Series A close as creating pressure to convert licensing engagements into commercial revenue quickly. | 中 | SO020 |
| CO034 | Independent, peer-reviewed benchmarks of Velaura's claimed power-savings figures are not yet widely available in public sources as of the run date. | 低 | SO019, SO020 |
| CO035 | Velaura AI's leadership also includes Sanjay Gupta (President, Strategy & GTM), YJ Kim (President, Products), Tim Vehling (VP, Product), Brian Campbell (VP of Engineering Operations), and Atul Dhablania (Chief Systems Operations Officer). | 中 | SO002 |
| CO036 | Velaura AI's advisors include Aditya Grover (CTO, Inception AI; UCLA professor) and Magnus Egerstedt (Provost, UNC Chapel Hill; roboticist). | 中 | SO002 |
| CO037 | Velaura AI markets its technology across three segments: hyperscale data centers, Physical & Embodied AI (robots, drones, autonomous machines), and Edge AI. | 中 | SO001 |
| CO038 | Velaura AI does not publicly disclose current headcount, revenue, or a detailed customer list. | 中 | SO019, SO012 |
| CO039 | Heise Online reported that Velaura AI CEO Rajiv Khemani told Reuters the company is in talks with three of the four largest cloud providers but declined to name them. | 中 | SO019 |
| CO040 | TechBooky's coverage cautions that chip startups face high execution risk and that a unicorn valuation does not guarantee commercial adoption. | 中 | SO020 |
| CO041 | Auradine's board previously included Fred Thiel (CEO, MARA) reflecting MARA's role as both an investor and a Teraflux mining-hardware customer. | 中 | SO002, SO016 |
| CO042 | Velaura AI's Series A financing is earmarked to expand engineering and customer-facing teams and deepen collaborations with strategic partners developing AI infrastructure and Physical AI solutions. | 中 | SO003 |
| CM001 | The data center AI chip market (specialized inference/training processors) was valued at $13.8 billion in 2026, projected to reach $45.2 billion by 2034 at a 14.1% CAGR. | 中 | SM001 |
| CM002 | The broader global AI semiconductor market is estimated at $300-500 billion in 2026 across Deloitte and Gartner methodologies, versus IDC's narrower $477.1 billion data-center-semiconductor figure and $281 billion 'intelligent datacenter' subset. | 中 | SM002, SM003 |
| CM003 | Global semiconductor revenue overall is forecast by Gartner to exceed $1.3 trillion in 2026, a 64% year-over-year increase. | 中 | SM002, SM003 |
| CM004 | NVIDIA held approximately 87.4% of combined merchant data-center AI chip revenue among NVIDIA, AMD, and Intel in Q1 2026. | 中 | SM002 |
| CM005 | The physical AI market (Mordor Intelligence definition, spanning industrial/service robotics hardware and software) is projected at $7.11 billion in 2026, growing to $34.89 billion by 2031 at a 37.46% CAGR. | 中 | SM007 |
| CM006 | MarketsandMarkets estimates a narrower-scope physical AI market at $0.89 billion in 2025, reaching $15.28 billion by 2032 at a 47.2% CAGR. | 中 | SM008 |
| CM007 | SNS Insider estimates the physical AI market at $6.93 billion in 2026, reaching $49.73 billion by 2035 at a 32.53% CAGR. | 中 | SM009 |
| CM008 | Physical AI market-sizing estimates for the same 2025-2026 window vary by more than 7x across publishers ($0.89B to $7.11B), reflecting materially different scope definitions rather than a single agreed TAM. | 中 | SM007, SM008, SM009 |
| CM009 | By component, hardware held 52.42% of the physical AI market in 2025 while software is projected to grow faster, at a 40.43% CAGR through 2031. | 中 | SM007 |
| CM010 | Industrial robots held 58.23% of the physical AI market by robot type in 2025, while professional service robots are projected to grow faster, at 39.72% CAGR through 2031. | 中 | SM007 |
| CM011 | On-device compute deployment accounted for 71.43% of the physical AI market in 2025, directly relevant to Velaura's on-chip power-efficiency value proposition for edge/embodied AI. | 中 | SM007 |
| CM012 | Big Five hyperscalers (Amazon, Microsoft, Alphabet, Meta, Oracle) are projected to collectively spend $600-725 billion on capital expenditures in 2026, with roughly 75% (about $450-500 billion) directed at AI infrastructure. | 中 | SM018, SM021, SM022 |
| CM013 | 2026 hyperscaler AI capex represents a 45-57% share of revenue for these companies, a capital-intensity level more typical of industrial sectors than software. | 中 | SM021 |
| CM014 | Hyperscalers raised over $100 billion in new debt in 2025 partly to fund AI infrastructure capex, signaling capital-market limits on unconstrained AI data-center spending. | 中 | SM021, SM023 |
| CM015 | Global AI data center electricity consumption surged to approximately 485 TWh in 2025 and is projected to nearly double to 950 TWh by 2030. | 高 | SM018, SM020, SM019 |
| CM016 | The U.S. faces an estimated structural power shortfall of 9.3 GW as of 2026, a physical grid-capacity constraint on new AI data-center capacity that is distinct from chip supply. | 中 | SM018 |
| CM017 | Custom silicon (chiplet-based ASICs designed in-house by hyperscalers) is projected to grow shipments at a 44.6% CAGR through 2033, nearly three times faster than merchant GPU shipment growth at 16.1% CAGR. | 中 | SM024, SM025 |
| CM018 | Custom AI ASICs can deliver 40-65% total-cost-of-ownership savings versus multi-purpose merchant GPUs for sustained inference workloads, per industry analysis. | 中 | SM024 |
| CM019 | Major hyperscaler custom-silicon programs in 2026 include Google TPU v7 'Ironwood', AWS Trainium 3, Microsoft Maia 200, and Meta MTIA 500, all of which represent potential direct customers or substitute-technology owners for Velaura's licensable IP. | 中 | SM025, SM026 |
| CM020 | NVIDIA's share of the AI inference market (as distinct from training) is projected by some analysts to fall to 20-30% by 2028 as hyperscalers shift inference workloads to in-house custom accelerators. | 低 | SM025 |
| CM021 | Most hyperscaler custom AI accelerators are used internally and are not commercially available to third parties, leaving NVIDIA and AMD merchant GPUs as the default choice for buyers outside the largest five hyperscalers. | 中 | SM024, SM025 |
| CM022 | US export controls on advanced AI chips to China (including licensing regimes and volume caps on chips such as NVIDIA's H200) create regulatory uncertainty and added compliance cost across the AI-chip supply chain in 2026. | 中 | SM016, SM017 |
| CM023 | China has responded to US chip export controls with rare-earth export restrictions, fragmenting global semiconductor supply chains and encouraging parallel US/China AI-chip ecosystems. | 中 | SM016 |
| CM024 | Liquid-cooled AI server deployment share rose from 15% in 2024 to 54% in 2025 and is projected to reach 76% in 2026, reflecting the broader industry shift toward addressing power/thermal constraints across the stack, not solely at the chip level. | 中 | SM020 |
| CM025 | Velaura AI frames its addressable market as spanning three buyer segments: hyperscale data center operators, Physical/Embodied AI system makers (robotics, drones, autonomous machines), and Edge AI device makers. | 高 | SM028, SM029 |
| CM026 | Velaura AI's own market framing cites an IEA projection that global AI-related electricity demand will more than double to over 945 TWh/year by 2030, aligning its pitch with the independently reported ~950 TWh figure from other analysts. | 中 | SM028, SM018 |
| CM027 | Velaura's royalty-plus-license business model means its own revenue is a small fraction of the underlying chip's value, so its serviceable-obtainable-market is best measured in royalty dollars per XPU shipped rather than total chip revenue. | 中 | SM028 |
| CM028 | No public source provides an independently verified SAM or SOM figure specific to ultra-low-power silicon IP licensing (Velaura's specific niche); all available sizing lenses describe adjacent broader categories (data-center AI chips, physical AI hardware, overall AI semiconductors). | 低 | SM001, SM007 |
| CM029 | Groq's roughly $20 billion IP/talent deal with NVIDIA in December 2025 and subsequent $350M raise at a reset $3.5B valuation illustrate how quickly comparable-company valuations can re-rate in the independent AI-chip sector. | 中 | SM011, SM012, SM013 |
| CM030 | Cerebras went public on May 14, 2026 at approximately a $56 billion fully diluted valuation, the largest US tech IPO since Snowflake in 2020, resetting comparables for independent AI silicon companies. | 中 | SM013 |
| CM031 | Broadcom and Marvell together provide ASIC design services and networking IP behind more than 80% of hyperscaler custom AI silicon, making them potential channel partners or competitors to Velaura's IP-licensing approach. | 中 | SM013 |
| CM032 | Switching costs for a hyperscaler to integrate a new silicon-IP vendor's low-power libraries into an existing SoC design flow are non-trivial, since RTL, verification, and manufacturing qualification cycles for advanced nodes typically span 12-24 months. | 低 | SM026, SM027 |
| CM033 | Regulatory/thermal/power constraints, rather than chip compute availability alone, are now the primary gating factor for new AI data-center capacity, directly increasing the relevance of Velaura's power-efficiency value proposition. | 高 | SM018, SM019, SM020 |
| CM034 | High capital expenditure requirements for advanced-node chip design and manufacturing (multi-billion-dollar fab and validation costs) are a structural barrier to new entrants in the AI-chip and AI-chip-IP markets. | 中 | SM001 |
| CM035 | Buyer budget ownership for AI accelerator purchases sits primarily with hyperscaler infrastructure/capex organizations, while Physical AI chip buying sits with robotics/device OEM engineering and procurement teams — two structurally different buyer paths for the same underlying Velaura IP. | 中 | SM021, SM007 |
| CM036 | TrendForce reports NVIDIA is responding to the custom-ASIC shift by selling more integrated racks/systems and opening NVLink for third-party ASIC interconnection, rather than only competing on raw chip performance. | 中 | SM027 |
| CM037 | The 2026 HBM (high-bandwidth memory) market is estimated at $54.6 billion, up 58% year over year and sold out through 2026, indicating memory bandwidth is a co-constraint alongside power efficiency for AI accelerator scaling. | 中 | SM003 |
| CP001 | Untether AI, a Toronto-based low-power AI inference chip startup, shut down operations in June 2025 with its engineering team acqui-hired by AMD. | 高 | SP001, SP002 |
| CP002 | Untether AI filed for bankruptcy in October 2025 with liabilities exceeding $128 million, mostly owed to early and institutional investors. | 中 | SP002 |
| CP003 | AMD's acquisition of Untether AI's engineering team did not include Untether's products (SpeedAI accelerator, imAIgine SDK), which are now discontinued and unsupported. | 中 | SP001 |
| CP004 | Mythic, an analog compute-in-memory AI chip competitor, raised a $125 million Series D in December 2025 after earlier financial struggles, backed by DCVC, Honda, and Lockheed Martin among others. | 中 | SP003 |
| CP005 | Hailo, an Israeli edge AI chip maker with roughly 100 customers and $344 million raised, was acquired by Microchip Technology in July 2026. | 中 | SP004 |
| CP006 | Groq's founding technical team and IP were effectively absorbed via a reported $20 billion NVIDIA deal in December 2025, after which Groq pivoted to an 'AI inference neocloud' model and raised $350 million at a reset $3.5 billion valuation in August 2026. | 高 | SP009, SP011 |
| CP007 | Groq had previously raised $750 million at a $6.9 billion valuation in September 2025, before the NVIDIA deal reset its valuation sharply lower. | 高 | SP010, SP009 |
| CP008 | Cerebras went public on May 14, 2026, closing its first trading day at approximately a $56 billion fully diluted valuation, the largest US tech IPO since Snowflake in 2020. | 中 | SP011 |
| CP009 | Tenstorrent remains an independent, revenue-generating AI accelerator company as of mid-2026 without a disclosed unicorn-level valuation reset in public sources. | 中 | SP008 |
| CP010 | NVIDIA held approximately 87.4% of combined merchant data-center AI chip revenue among NVIDIA, AMD, and Intel in Q1 2026, making it the dominant incumbent competitor. | 中 | SP026 |
| CP011 | Hyperscalers' in-house custom silicon programs (Google TPU, AWS Trainium, Microsoft Maia, Meta MTIA) represent an 'internal build' substitute to licensing outside IP such as Velaura's Titan Core. | 中 | SP013, SP014, SP015 |
| CP012 | Custom ASIC shipments are projected to grow at a 44.6% CAGR through 2033, nearly three times faster than merchant GPU shipment growth of 16.1% CAGR, expanding the internal-build substitute pool. | 中 | SP013, SP014 |
| CP013 | EDA vendors Cadence and Synopsys sell low-power design, verification, and AI-driven optimization tools (Cerebrus, DSO.ai) that let any chip designer pursue power efficiency in-house without licensing a third-party IP platform like Titan Core. | 高 | SP005, SP006 |
| CP014 | Arm's historical processor-licensing-plus-royalty model is the explicit business-model comparator Velaura's CEO cites for Titan Core's own upfront-fee-plus-royalty structure. | 中 | SP017, SP020 |
| CP015 | In 2026 Arm began selling its own finished AI chips directly, reorganizing into Edge, Physical AI, and Cloud AI business units and breaking with three decades of pure IP licensing. | 中 | SP018, SP019 |
| CP016 | Arm's move into direct chip sales illustrates a risk that even a successful IP licensor may eventually face pressure to compete directly with its own customers or licensees. | 中 | SP018 |
| CP017 | Samsung Catalyst Fund, a Velaura investor, is the corporate venture arm of Samsung Electronics, whose foundry business could be both a partner and a potential vertically integrated competitor for power-efficient chip manufacturing. | 中 | SP027 |
| CP018 | Velaura AI's core competitive claim is a 2-4x performance-per-watt improvement validated across 30 million-plus production ASICs, a scale-proof point most early-stage low-power AI chip competitors (Mythic, Untether AI at shutdown) could not demonstrate. | 高 | SP021, SP022 |
| CP019 | Unlike Groq, Cerebras, Tenstorrent, and Untether AI, which design and sell (or sold) finished chips under their own brand, Velaura licenses IP into customers' existing chip designs rather than competing as a merchant silicon vendor. | 中 | SP021 |
| CP020 | Velaura does not publicly disclose which specific hyperscaler chips embed its IP, an opacity that independent press flags as a disclosure gap relative to competitors like Hailo or Mythic that name reference customers and design wins. | 中 | SP023 |
| CP021 | Trade press coverage frames chip-startup execution risk broadly, noting deep engineering, manufacturing partnerships, and customer validation are required to convert any design win into recurring revenue - a risk equally applicable to Velaura and its direct competitors. | 中 | SP024 |
| CP022 | Independent analyst Patrick Moorhead said Velaura's approach has the potential to reduce total cost of ownership and ease thermal limitations versus status-quo GPU-only approaches. | 中 | SP025 |
| CP023 | A 2026 industry roundup groups Velaura AI alongside Untether AI, Efficient Computer, Groq, Tenstorrent, Graphcore, SambaNova, Hailo, Cerebras, and Mythic as competitors in the low-power/inference-optimized AI chip landscape. | 中 | SP028 |
| CP024 | Velaura AI markets its technology across three segments (hyperscale data centers, Physical/Embodied AI, Edge AI), overlapping with Hailo (industrial/automotive edge), Mythic (edge/datacenter inference), and hyperscaler internal ASIC teams (data center). | 高 | SP029, SP021 |
| CP025 | No independent benchmark directly compares Velaura's Titan Core performance-per-watt claims against Mythic's compute-in-memory or Hailo's NPU architectures head-to-head. | 低 | SP021, SP004 |
| CP026 | Broadcom and Marvell provide ASIC design services behind more than 80% of hyperscaler custom AI silicon, positioning them as potential channel partners or competitors to Velaura's IP-licensing model depending on whether they build in-house power-optimization capability. | 中 | SP011 |
| CP027 | Switching from an incumbent merchant GPU to a licensed IP block like Titan Core requires a hyperscaler to commit to a 12-24 month RTL-to-tape-out integration cycle, creating meaningful switching costs and lock-in once a design win is secured. | 中 | SP015, SP013 |
| CP028 | NVIDIA is responding to the custom-silicon shift by selling more integrated racks/systems and opening its NVLink interconnect to third-party ASICs, rather than competing solely on standalone chip performance. | 中 | SP012 |
| CP029 | Velaura's royalty-based revenue model, tied to measured power savings, is structurally distinct from competitors like Hailo and Mythic, which sell finished chips directly and capture full unit economics rather than a licensing royalty. | 中 | SP021, SP004, SP003 |
| CP030 | Untether AI's shutdown was attributed in part to fundraising difficulty amid NVIDIA's market dominance and broader economic headwinds including US tech tariffs, an adverse signal for the sustainability of small independent low-power AI chip vendors. | 高 | SP001, SP002 |
| CP031 | Hailo's acquisition by Microchip Technology illustrates a pattern of established semiconductor players acquiring edge-AI chip startups rather than competing via in-house development alone. | 中 | SP004 |
| CP032 | Mythic's post-restructuring Series D included strategic corporate investors Honda and Lockheed Martin, signalling automotive and defense-sector interest in energy-efficient AI inference chips as a competitive vector distinct from Velaura's hyperscaler/data-center focus. | 中 | SP003 |
| CP033 | Seligman Ventures' Managing Partner described Velaura as the firm's first investment in Physical AI, implying Velaura's board/investor network sees limited direct competitive overlap with existing portfolio companies. | 中 | SP022 |
| CP034 | A competitive risk for Velaura is commoditization if EDA vendors (Synopsys, Cadence) or foundries embed equivalent low-power libraries directly into standard design kits, reducing the need for a separate third-party IP licensor. | 高 | SP005, SP006 |
| CP035 | Velaura's 30 million-plus ASIC production track record, inherited from its Auradine bitcoin-mining chip business, is a differentiated moat versus newer entrants without comparable manufacturing history, though it does not by itself prove the newer Titan Core AI-specific IP performs as claimed. | 中 | SP021, SP023 |
| CP036 | Efficient Computer, one of the named low-power AI chip startup peers identified in 2026 market roundups, is included in Analytics Insight's startup list alongside Velaura, though public functional/technical detail on Efficient Computer remains limited in sources reviewed. | 低 | SP028 |
| CI001 | MARA Holdings' Q1 2026 10-Q reports the carrying amount of its investment in Velaura AI (formerly Auradine), a related party, at $85.4 million as of March 31, 2026. | 中 | SI001 |
| CI002 | MARA holds one seat on Velaura AI's board of directors, per its own SEC filing disclosure. | 中 | SI001 |
| CI003 | On February 19, 2025, MARA converted a $1.2 million SAFE investment in Velaura into preferred stock and purchased an additional $20.0 million of Velaura preferred stock. | 中 | SI001 |
| CI004 | MARA recorded $11.9 million and $2.7 million in fair-value gains on its Velaura investments (preferred and common stock respectively) under ASC 321, adjusting carrying value to an observable price from a later financing round. | 中 | SI001 |
| CI005 | During the three months ended March 31, 2025, MARA advanced $22.3 million to Velaura (then Auradine) for product purchases, with a $57.2 million outstanding balance owed as of that date. | 中 | SI001 |
| CI006 | As of March 31, 2026, MARA had no outstanding balance or payment commitment to Velaura, and made no product-purchase advances during the quarter, indicating the prior Teraflux hardware prepayment obligation had been fully settled. | 中 | SI001 |
| CI007 | Velaura AI's core AI-silicon revenue mechanism is an upfront IP license fee plus a royalty tied to the power savings a customer measurably achieves using Titan Core. | 高 | SI003, SI021 |
| CI008 | Velaura estimates its Titan Core technology can save customers approximately $1,300 in electricity costs over three years per XPU, a figure used to justify the royalty pricing basis though the exact royalty rate is not disclosed. | 中 | SI003 |
| CI009 | Velaura AI's official announcements do not disclose a specific royalty percentage, minimum license fee, or contract-term structure for Titan Core licensing. | 中 | SI003, SI011 |
| CI010 | Velaura's Series A proceeds (announced August 2026) are earmarked to expand engineering and customer-facing teams and deepen strategic partnerships, per the company's own announcement. | 中 | SI004 |
| CI011 | Typical fabless semiconductor companies report gross margins in the 50-65% range, with leading players like NVIDIA exceeding 70%, providing a rough comparable benchmark for evaluating Velaura's undisclosed margin profile. | 中 | SI006 |
| CI012 | Early-stage fabless semiconductor startups typically spend 20-40% of revenue on R&D (spiking to 50%+ pre-revenue) and burn $500,000-$2,000,000 per month, with 12-18 months of runway considered prudent. | 中 | SI007 |
| CI013 | No public source discloses Velaura AI's actual monthly cash burn, current cash-on-hand balance, or runway following its August 2026 Series A. | 低 | SI012 |
| CI014 | Auradine had approximately 62 employees as of March 26, 2026 per Tracxn's tracked employee-count trend, shortly before or around its rebrand to Velaura AI. | 中 | SI010 |
| CI015 | Other analyst-database estimates place Velaura AI/Auradine's headcount in a broader 100-150 employee range as of mid-2026, with LinkedIn-style company-size bands showing 101-500 employees. | 低 | SI008, SI009 |
| CI016 | Headcount estimates for Velaura AI diverge meaningfully across sources (62 per Tracxn vs. 100-150 per other analyst databases), and no single figure is independently verified or company-disclosed. | 低 | SI010, SI008, SI009 |
| CI017 | Velaura AI's predecessor Auradine achieved $80 million in customer bookings by April 2024 alongside its Series B raise, its only publicly disclosed bookings figure to date. | 中 | SI014 |
| CI018 | No public source discloses Velaura AI's current annual recurring revenue, run-rate revenue, or revenue mix between legacy Teraflux hardware sales and new AI-silicon licensing royalties. | 低 | SI015, SI028 |
| CI019 | Auradine's Series C financing (April 2025, $153 million) was explicitly earmarked to expand beyond bitcoin-mining hardware into blockchain and AI infrastructure, representing the capital bridge that funded the eventual Velaura AI pivot. | 中 | SI013 |
| CI020 | Velaura's legacy Teraflux bitcoin-mining hardware business historically sold physical ASIC miners directly to customers (a hardware unit-sale model), structurally distinct from Titan Core's royalty-based IP-licensing model now being scaled. | 中 | SI016, SI017 |
| CI021 | Velaura's Series A closed at a valuation exceeding $1 billion in August 2026, but no independent source discloses a revenue multiple, EBITDA, or other standard valuation-basis metric behind that figure. | 中 | SI018, SI019 |
| CI022 | Semiconductor Online's republication of Velaura's Titan Core announcement confirms the same $1,300-per-XPU three-year savings estimate and 2-4x performance-per-watt claim as the original company release, with no additional independent cost-structure detail. | 中 | SI020 |
| CI023 | GamesBeat's interview coverage describes Velaura's royalty mechanism as tied to measured power savings rather than a flat per-unit fee, meaning realized revenue per chip should vary with actual customer power-savings outcomes rather than being fixed. | 中 | SI021 |
| CI024 | Mayfield's investor commentary confirms Velaura's Series A proceeds are intended to fund team expansion and deepen strategic partner collaborations, consistent with the company's own use-of-proceeds statement. | 中 | SI022 |
| CI025 | Arm's historical royalty rate on licensed processor IP has been publicly estimated at roughly 1-2% of a chip's average selling price, providing an external analogy for how small a royalty-based IP licensor's per-unit take can be relative to total chip value. | 中 | SI025, SI026 |
| CI026 | Groq's need to raise $350 million in new capital in August 2026 despite a prior $750 million round illustrates how capital-intensive the AI-silicon sector remains even for well-funded independent competitors, a comparable capital-intensity signal relevant to assessing Velaura's own financing runway needs. | 中 | SI027 |
| CI027 | No public source discloses Velaura AI's customer acquisition cost, sales-cycle length, or channel economics for closing a hyperscaler licensing deal. | 低 | SI011, SI012 |
| CI028 | Velaura's go-to-market motion, based on available company and press descriptions, appears to be a direct, high-touch enterprise sales/engineering-partnership process with hyperscalers rather than a self-serve or channel-reseller motion. | 中 | SI003, SI021 |
| CI029 | Velaura's working-capital and capex profile is not independently disclosed; MARA's 10-Q shows only the investor-side accounting treatment (investment carrying value and prior product-purchase advances), not Velaura's own balance sheet. | 中 | SI001 |
| CI030 | Because Velaura's royalty revenue depends on measuring a customer's actual power savings, revenue recognition timing likely lags physical chip shipment by the length of a customer's own qualification and deployment cycle, a recognition-timing risk not addressed in any source reviewed. | 低 | SI003, SI021 |
| CI031 | TechBooky's coverage explicitly cautions that a unicorn valuation 'does not guarantee commercial adoption,' a directly relevant caveat for assessing whether Velaura's Series A capital will convert into recurring royalty revenue on the timeline investors may expect. | 中 | SI012 |
| CI032 | Heise Online's coverage notes Velaura does not disclose which chips or customers use its IP, which also obscures the revenue base against which any future royalty stream would be measured. | 中 | SI011 |
| CI033 | Cumulative disclosed equity capital raised by Auradine/Velaura AI across all four rounds ($81M + $80M + $153M + $110M) totals approximately $424 million, before accounting for any debt or credit facilities that may exist but are undisclosed. | 中 | SI013, SI015 |
| CI034 | No source reviewed discloses any debt facility, project-finance arrangement, or credit line held by Velaura AI, separate from its equity fundraising. | 低 | SI001, SI015 |
| CI035 | MARA's related-party disclosure of its Velaura investment and historical product-purchase advances is the only independently filed (SEC-sourced) financial data point available for Velaura AI as of the run date; all other financial figures originate from company press releases or third-party estimates. | 中 | SI001 |
| CI036 | The AI semiconductor industry overall is reporting extremely strong vendor revenue growth in 2026 (e.g., NVIDIA's $75.2B single-quarter data-center revenue), a favorable macro backdrop against which Velaura's own undisclosed revenue base cannot yet be benchmarked. | 中 | SI023, SI024 |
| CE001 | Velaura AI's engagement model has three stages: the customer provides its RTL design plus priorities for area, power, and other factors; Velaura applies proprietary IP, toolflow, and low-voltage libraries; the customer receives an optimized, node-specific GDS or chiplet output. | 高 | SE001, SE009 |
| CE002 | Velaura AI describes its low-power compute technology as patented, though no specific patent number or grant is independently confirmed in sources reviewed. | 中 | SE001 |
| CE003 | Velaura's own materials estimate up to $100 million per year in power savings for a typical data center, based on an explicit assumption of 10 cents/kWh and approximately $433.30 per chip per year across 100,000-250,000 units. | 中 | SE001 |
| CE004 | Velaura's Physical AI solution targets robots, drones, industrial automation, humanoids, and edge devices, leveraging the same underlying Titan Core technology as the data-center product line. | 高 | SE002, SE003 |
| CE005 | Velaura AI maintains two distinct product/solution lines: Ultra Low Power AI Compute (silicon IP for AI accelerators) and Blockchain (air, immersion, and hydro-cooled Bitcoin mining hardware), per its own Solutions page. | 高 | SE003, SE005 |
| CE006 | Titan Core's underlying design methodology integrates into a customer's existing SoC architecture and design flow without requiring software changes, preserving full functional equivalence at a lower operating voltage. | 高 | SE009, SE010 |
| CE007 | Velaura's technology is validated and available for advanced 3nm and 2nm semiconductor process nodes. | 高 | SE009, SE020 |
| CE008 | Titan Core specifically targets the matrix-multiplication (MATMUL) operations that dominate AI training and inference energy use, reducing their energy requirement by 2-4x using proprietary circuit and library technology. | 高 | SE009, SE011 |
| CE009 | Velaura AI, Inc. is incorporated in Delaware with headquarters at 3200 Coronado Dr., Santa Clara, CA 95054, per its own privacy notice (filed under the prior Auradine, Inc. legal name). | 中 | SE008 |
| CE010 | No public source discloses a specific security certification (e.g., ISO 27001, SOC 2) or safety/quality certification (e.g., AEC-Q100 for automotive, ISO 26262 functional safety) held by Velaura AI for its silicon IP. | 低 | SE004, SE018 |
| CE011 | Velaura's careers page (hosted on Lever) confirms active hiring but does not publicly disclose specific open engineering roles' technical stack details in the content retrieved. | 中 | SE007 |
| CE012 | No public GitHub repository, Hacker News thread, Stack Overflow tag, or developer forum discussion specific to Velaura AI's Titan Core IP was identified in the sources reviewed for this chapter. | 低 | SE007 |
| CE013 | Practitioner-community engagement with the broader 'AI in semiconductor design/manufacturing' topic is active in 2026, evidenced by SEMI's August 2026 'AI Techniques in Semiconductor Manufacturing' workshop and a broader industry events calendar (DAC, SEMICON WEST) covering AI hardware and chiplet architectures. | 中 | SE014, SE015 |
| CE014 | Moor Insights & Strategy analyst Patrick Moorhead reviewed Titan Core's technical approach and said it has the potential to improve performance per watt in ways that could reduce total cost of ownership and ease thermal limitations. | 中 | SE012 |
| CE015 | SiliconANGLE's independent coverage confirms the same core technical claims (2x lower power, 30M+ ASICs, 3nm/2nm validation) as Velaura's own press release, without providing additional independent verification. | 中 | SE013 |
| CE016 | Heise Online's independent reporting flags that Velaura does not disclose which chips or customers embed its IP, limiting independent verification of real-world deployment and reliability claims. | 中 | SE018 |
| CE017 | TechBooky's coverage cautions that chip startups require deep engineering, manufacturing partnerships, and customer validation to convert design wins into usable, reliable production silicon, a general execution-risk caveat applicable to Titan Core's maturity claims. | 中 | SE019 |
| CE018 | Velaura's underlying low-voltage circuit and library technology has been deployed and validated across more than 30 million production ASICs, primarily in its legacy Teraflux bitcoin-mining hardware line, providing a real manufacturing-yield and reliability track record predating the AI-specific Titan Core product. | 高 | SE020, SE009 |
| CE019 | Velaura's legacy Teraflux bitcoin-mining hardware line offered air, immersion, and hydro-cooled variants, with the newest models achieving as low as 9.8 J/TH energy efficiency and up to 600 TH/s hashrate, demonstrating the same low-voltage engineering discipline now applied to Titan Core. | 中 | SE021, SE023 |
| CE020 | Auradine's Teraflux miners were built on 4nm process nodes as of early 2025, one process generation behind the 3nm/2nm nodes now targeted by Titan Core for AI accelerators. | 中 | SE024 |
| CE021 | The Block's independent coverage confirms Auradine engineered a hydro-cooled bitcoin miner in the US specifically to help customers navigate customs and tariff issues, indicating a domestic-manufacturing/supply-chain differentiation strategy predating the AI pivot. | 中 | SE022 |
| CE022 | MARA's own commentary on its relationship with Auradine/Velaura highlights the value of US-based engineering support and supply-chain resilience as decision factors, an operating-model signal relevant to assessing Velaura's reliability/support posture with hyperscaler customers. | 中 | SE027 |
| CE023 | Cadence's Cerebrus AI Studio and Synopsys's DSO.ai provide AI-driven low-power design automation as of 2026, representing an architecturally different (EDA-tool-based, in-house) approach to the same power-optimization problem Velaura's pre-built IP libraries solve. | 高 | SE025, SE026 |
| CE024 | No independent, peer-reviewed benchmark of Titan Core's claimed 2-4x MATMUL efficiency improvement or $100M/year data-center savings estimate exists in the sources reviewed for this chapter. | 低 | SE001, SE018 |
| CE025 | Velaura's roadmap, as described in its own materials, moves from the March 2026 Titan Core silicon IP announcement toward expanded hyperscaler and Physical AI engagements funded by the August 2026 Series A, but no dated public roadmap of future feature releases or node transitions beyond 2nm is disclosed. | 中 | SE009, SE020 |
| CE026 | Velaura's own $100M/year power-savings estimate is explicitly labeled with its assumptions (electricity price, per-chip savings, unit-volume range), which is a more transparent disclosure practice than the company's undisclosed royalty-rate methodology described in the Financials chapter. | 中 | SE001 |
| CE027 | Velaura's website does not include a dedicated public trust, security, or compliance page beyond a general privacy notice, unlike some competitors that publish explicit security/compliance documentation. | 中 | SE004, SE008 |
| CE028 | The USPTO Patent Public Search and Google Patents tools are the recommended primary channels for independently verifying any specific Velaura/Auradine patent grant, but no specific patent number was identified as confirmed in the sources reviewed for this chapter. | 低 | SE016, SE017 |
| CE029 | Velaura AI's leadership team, including CDO Manu Gulati (a former NUVIA co-founder, later acquired by Qualcomm), brings direct prior experience in custom silicon architecture relevant to executing on the Titan Core roadmap. | 中 | SE006, SE010 |
| CE030 | Velaura's Titan Core output is described as either an optimized GDS (Graphic Design System, the standard chip-layout file format) or a chiplet, indicating the IP can be delivered as either a full physical-design layout or a discrete chiplet component depending on customer integration needs. | 中 | SE001 |
| CE031 | No source discloses a specific reliability metric (e.g., mean time between failures, defect rate in parts per million) for Titan Core-embedded chips specifically, as distinct from the legacy Teraflux hardware line's disclosed reliability claims. | 低 | SE009, SE023 |
| CE032 | Velaura's product positioning across data center, Physical AI, and Edge AI segments relies on a single underlying low-voltage IP platform (Titan Core) rather than separate, segment-specific architectures, which the company frames as a capital-efficient way to address three markets with one core technology. | 中 | SE002, SE003, SE005 |
| CE033 | Velaura does not publish a public API, SDK, or software-integration surface for Titan Core, consistent with its positioning as a physical-design/IP licensing product rather than a software product. | 中 | SE001, SE004 |
| CE034 | Axis Intelligence's independent AI chip market data shows NVIDIA and AMD's merchant GPU architectures rely on different power-efficiency approaches (generational process shrinks and architectural tuning) than Velaura's third-party IP-licensing model, providing an architecture-level comparable for evaluating Titan Core's differentiation. | 中 | SE028 |
| CE035 | Velaura's official Physical AI and Solutions pages emphasize 'purpose-built' and 'real-world' framing for its robotics/embodied-AI offering, but neither page discloses a specific named robotics customer or deployed unit count for this segment distinct from its data-center traction claims. | 中 | SE002, SE003 |
| CE036 | This chapter finds no evidence that Velaura AI's technical claims (2-4x efficiency, 30M+ ASICs, 3nm/2nm validation) have been independently retracted, disputed, or contradicted by any adverse or regulatory source reviewed; adverse coverage instead focuses on disclosure gaps (customer names, benchmark independence) rather than factual inaccuracy. | 中 | SE018, SE019 |
| CU001 | Genesis Digital Assets Limited (GDA), one of the world's largest Bitcoin mining companies by hash rate, entered a purchase agreement in June 2025 to acquire 1,000 Auradine Teraflux AT2880-277 air-cooled miners for its 40 MW Glasscock County, Texas data center. | 高 | SU001, SU002 |
| CU002 | GDA's Executive President Abdumalik Mirakhmedov publicly credited Auradine's mining systems with offering the performance and flexibility GDA needs to compete globally while supporting sustainability and grid reliability goals. | 中 | SU001 |
| CU003 | The AT2880-277 miners GDA purchased deliver up to 260 TH/s hash rate at energy efficiency as low as 16 J/TH, and will participate in ERCOT's Texas demand response program once operational. | 高 | SU001, SU004 |
| CU004 | MARA Holdings purchased $73.3 million worth of Auradine Teraflux mining machines in the first half of 2025, with the order fully delivered by June 30, 2025. | 中 | SU005, SU006 |
| CU005 | MARA's SEC 10-Q filing corroborates a $22.3 million product-purchase advance to Velaura (then Auradine) in Q1 2025 with a $57.2 million outstanding balance as of March 31, 2025, providing an independently filed (non-company, non-press) confirmation of real production deployment. | 高 | SU007, SU005 |
| CU006 | MARA is simultaneously a Velaura investor (holding a board seat and $85.4 million investment carrying value) and a customer purchasing Teraflux hardware, creating a related-party relationship that should be read with that dual role in mind when assessing reference quality. | 高 | SU007, SU008 |
| CU007 | Auradine's Teraflux hardware has shipped to more than 30 large-scale industrial Bitcoin mining operators and more than 40 data center operators as of 2025, though most of these customers are not individually named in public sources. | 中 | SU009, SU010 |
| CU008 | Auradine disclosed $80 million in customer bookings alongside its April 2024 Series B financing round, its only publicly disclosed aggregate bookings figure. | 中 | SU011 |
| CU009 | Velaura AI states it has ongoing engagements with three of the four largest cloud providers (hyperscalers) for its Titan Core AI-silicon IP, but has not named any of them. | 高 | SU012, SU013 |
| CU010 | Heise Online's independent reporting confirms Velaura's CEO told Reuters the company is in talks with three of the four largest cloud providers but declined to name them, an adverse disclosure-quality signal. | 中 | SU014 |
| CU011 | No named hyperscaler, robotics OEM, or edge-device customer has been identified for Velaura's new Titan Core AI-silicon product in any source reviewed; all named-customer evidence (GDA, MARA) relates to the legacy Teraflux bitcoin-mining hardware business. | 高 | SU001, SU005, SU012 |
| CU012 | The Block's independent coverage confirms Auradine's hydro-cooled miners were engineered specifically to help customers navigate US customs and tariff issues, an operational benefit valued by its Bitcoin-mining customer base. | 中 | SU015 |
| CU013 | Following the March 2026 rebrand to Velaura AI, the company shifted unsold Teraflux inventory to in-house Bitcoin mining rather than continuing external hardware sales, while existing customers retain warranty support. | 中 | SU017 |
| CU014 | Major public Bitcoin miners (Cipher Mining, Core Scientific, TeraWulf, Bitdeer) - the natural customer base for Auradine's legacy Teraflux hardware - are themselves pivoting toward AI/HPC colocation revenue and away from expanding Bitcoin-mining hardware fleets, reducing the addressable market for any remaining Teraflux sales. | 中 | SU018 |
| CU015 | Public Bitcoin miners collectively shed 21% of Bitcoin hashrate in 2026 as AI revenue accelerated, an industry-wide adverse trend directly relevant to demand for Velaura's legacy mining-hardware customer base. | 中 | SU020, SU019 |
| CU016 | No public source discloses a retention, renewal, or net-revenue-retention metric for any Velaura/Auradine customer relationship, whether legacy hardware or new AI-silicon licensing. | 低 | SU009, SU012 |
| CU017 | No public source discloses whether GDA's 1,000-miner purchase has reached full production deployment versus still being in installation/commissioning at the Glasscock County facility as of the run date. | 低 | SU001, SU004 |
| CU018 | MARA's $73.3 million H1 2025 purchase was fully delivered by June 30, 2025 per independent press coverage, indicating this specific deployment reached production status rather than remaining a pilot. | 中 | SU005, SU006 |
| CU019 | Velaura's own product pages (Physical AI, home page) describe target customer segments (robotics/drone/autonomous-machine OEMs, hyperscalers, edge-device makers) in aspirational terms without naming a single specific account in any of these segments. | 中 | SU024, SU025 |
| CU020 | MARA's board seat and repeat investment across Velaura's Series C and 2026 Series A rounds mean the single best-evidenced customer relationship (MARA) is also a related party and investor, limiting its value as an independent reference. | 中 | SU007, SU026 |
| CU021 | GDA operates 20 data centers across North America, South America, Europe, and Central Asia with over 600 MW of total power capacity and 150,000+ miners online, providing meaningful scale context for its 1,000-unit Auradine purchase (a small fraction of its total fleet). | 中 | SU001 |
| CU022 | TechBooky's coverage cautions that converting any chip-startup design win into recurring revenue requires customer validation that has not yet been demonstrated for Velaura's new AI-silicon business specifically. | 中 | SU022 |
| CU023 | GamesBeat's interview coverage describes Velaura's hyperscaler engagements as being in active technical evaluation and RTL-benchmark stages rather than confirmed volume-production deployment for the AI-silicon business. | 中 | SU023 |
| CU024 | Quartz's independent coverage repeats Velaura's own unnamed-hyperscaler-engagement claim without providing additional named-customer verification, consistent with the broader pattern of press coverage relying on company disclosure for this fact. | 中 | SU028 |
| CU025 | Hyperscaler AI infrastructure capex (the budget pool from which any future Velaura AI-silicon licensing revenue would be drawn) is projected at $600-725 billion in 2026, providing macro context for the addressable buyer base even though no specific hyperscaler account is named. | 中 | SU027 |
| CU026 | No evidence of a failed deployment, customer complaint, product recall, or negative customer review was identified for either Velaura's legacy Teraflux hardware or its new Titan Core IP in the sources reviewed for this chapter. | 中 | SU014, SU022 |
| CU027 | Velaura's customer base for its legacy Teraflux hardware is geographically concentrated in the United States, reflecting the company's explicit domestic-manufacturing and tariff-avoidance positioning. | 中 | SU015, SU004 |
| CU028 | No named customer or aggregate customer-count figure has been disclosed specifically for Velaura's Physical AI or Edge AI product segments, distinct from its data-center (hyperscaler) and legacy-hardware customer bases. | 低 | SU024, SU025 |
| CU029 | Velaura's transition of unsold Teraflux inventory to in-house mining (rather than external sale) after the March 2026 rebrand suggests declining new-customer acquisition in the legacy hardware business, consistent with the broader miner-to-AI-pivot trend documented industry-wide. | 中 | SU017, SU018 |
| CU030 | Land-and-expand potential exists structurally in Velaura's model: MARA's relationship progressed from Series B/C investor to Series A (2026) investor to disclosed $73.3M hardware purchaser, illustrating how an initial investor relationship expanded into a larger commercial purchase over time. | 中 | SU007, SU005, SU026 |
| CU031 | No public source discloses a customer concentration percentage (e.g., top-customer share of revenue) for Velaura's legacy hardware or new AI-silicon business. | 低 | SU009, SU007 |
| CU032 | Auradine's aggregate '30+ industrial miners and 40+ data center operators' customer-count claim has not been independently itemized or audited beyond the two named accounts (MARA, GDA) corroborated in this chapter. | 中 | SU009, SU010 |
| CU033 | MARA's fireside interview with Rajiv Khemani provides a qualitative customer-side account of the Auradine relationship (US-based support, supply-chain resilience) but does not include specific retention, renewal, or satisfaction metrics. | 中 | SU008 |
| CU034 | Distinct customer segments identified for Velaura across its history include: Bitcoin mining operators (legacy Teraflux, e.g. MARA, GDA), hyperscale cloud providers (Titan Core, unnamed), Physical AI/robotics OEMs (aspirational, unnamed), and Edge AI device makers (aspirational, unnamed). | 高 | SU001, SU007, SU012, SU024 |
| CU035 | The addressable buyer base for Velaura's legacy Teraflux hardware is structurally shrinking as major public Bitcoin miners redirect capital and power capacity toward AI/HPC colocation rather than new mining-hardware purchases, an adverse trend for any residual hardware-sale revenue. | 中 | SU018, SU019, SU020 |
| CU036 | This chapter finds no publicly disclosed named reference customer, case study, or independent review for Velaura's core AI-silicon Titan Core product line, in contrast to the well-documented named-customer evidence (GDA, MARA) available for its legacy Teraflux hardware business. | 高 | SU001, SU005, SU012, SU014 |
| CR001 | As of January 15, 2026, US export control policy for advanced computing chips shifted from a presumption of denial to case-by-case license review for exports to China and Macau, adding compliance conditions (US supply certification, foundry-diversion certification, third-party testing) relevant to any semiconductor IP company with foundry exposure. | 高 | SR001, SR002 |
| CR002 | Velaura AI's exposure to US-China semiconductor export control rules is not directly addressed in any company disclosure reviewed; the company's foundry partners (e.g., TSMC, Samsung Foundry) and any China-linked customers or supply chain would be the transmission channel for this regulatory risk. | 中 | SR001, SR002 |
| CR003 | AI-related patent litigation has surged industry-wide, with over a thousand AI-related patent lawsuits filed globally in the past five years per legal-industry analysis, indicating elevated IP litigation risk for any AI-silicon IP licensor including Velaura. | 高 | SR004, SR003 |
| CR004 | No public record identifies any lawsuit, litigation, or patent infringement claim specifically naming Auradine or Velaura AI as of the run date. | 中 | SR003, SR004 |
| CR005 | Velaura's own materials describe its low-power compute technology as 'patented' without disclosing a specific patent number, creating both an IP-strength verification gap and a converse risk that an unverified patent claim could itself invite scrutiny or a future dispute. | 中 | SR028 |
| CR006 | MARA Holdings' SEC 10-K risk-factors disclosure practice (Item 1A) provides a template for the kind of related-party and investment risk language that would be expected in any Velaura-linked public filing, though Velaura itself is private and not subject to equivalent disclosure requirements. | 中 | SR005 |
| CR007 | MARA's SEC 10-Q filing independently confirms a related-party relationship with Velaura (investment, board seat, and historical product-purchase advances), which is itself a disclosed governance and concentration risk factor for MARA and, by extension, a data point on Velaura's customer/investor concentration. | 中 | SR006 |
| CR008 | No independent registry confirmation (e.g., Delaware Division of Corporations entity search) of Velaura AI, Inc.'s current legal entity status was completed in this chapter's research beyond the company's own privacy notice; the Delaware search tool requires a direct, fee-gated query to confirm entity status definitively. | 低 | SR007, SR026 |
| CR009 | Heise Online's independent reporting confirms Velaura's CEO declined to name any of the three hyperscaler cloud providers the company claims to be engaged with, even when directly asked by Reuters, an adverse disclosure-transparency signal relevant to diligence confidence in the company's traction claims. | 中 | SR009 |
| CR010 | TechBooky's independent coverage explicitly cautions that a unicorn valuation does not guarantee commercial adoption, and that chip startups require deep engineering, manufacturing partnerships, and customer validation to convert design wins into revenue. | 中 | SR010 |
| CR011 | Untether AI, a directly comparable low-power AI inference chip startup, shut down in June 2025 and filed for bankruptcy in October 2025 with liabilities exceeding $128 million, an adverse sector precedent illustrating that even technically credible low-power AI chip startups can fail commercially. | 高 | SR011, SR012 |
| CR012 | Untether AI's failure was attributed in part to fundraising difficulty amid NVIDIA's market dominance and broader economic headwinds including US tech tariffs, risk factors that could similarly affect Velaura given its comparable market position. | 中 | SR011, SR012 |
| CR013 | Public Bitcoin miners collectively shed 21% of Bitcoin hashrate in 2026 as AI revenue accelerated, directly shrinking the addressable market for Velaura's legacy Teraflux hardware business, its only currently revenue-evidenced product line. | 中 | SR014, SR013 |
| CR014 | 70% of top Bitcoin miners are reportedly using AI-related income to survive the current bear market, indicating that Velaura's legacy hardware customer base is itself under financial pressure and may deprioritize new hardware purchases. | 中 | SR013 |
| CR015 | Velaura's decision to shift unsold Teraflux inventory to in-house mining rather than continuing external sales (announced with the March 2026 rebrand) signals declining new-customer acquisition in its only currently revenue-evidenced business line. | 中 | SR015 |
| CR016 | Cadence's Cerebrus AI Studio and Synopsys's DSO.ai provide AI-driven low-power design automation as of 2026, a commoditization risk that could allow hyperscaler customers to internalize Velaura's core differentiation using commercial EDA tools rather than continuing to license Titan Core. | 高 | SR016, SR017 |
| CR017 | Arm broke from three decades of pure IP licensing in 2026 to begin selling its own finished AI chips directly, an adverse precedent illustrating that even a dominant, successful royalty-model incumbent can face pressure to compete directly with its own licensees over time. | 中 | SR018, SR019 |
| CR018 | Headcount estimates for Velaura AI/Auradine diverge significantly across sources (62 per Tracxn vs. a broader 100-150 range per other analyst databases), indicating unreliable public visibility into the company's actual organizational scale. | 中 | SR020, SR021 |
| CR019 | No public source discloses Velaura's current cash-on-hand, monthly burn, or runway following its August 2026 Series A, creating a financing-dependency risk that cannot be independently assessed. | 低 | SR029 |
| CR020 | Custom-silicon shipment growth (44.6% CAGR) outpacing merchant GPU growth (16.1% CAGR) could just as easily displace Velaura's IP-licensing opportunity as expand it, if hyperscalers choose to build equivalent power-efficiency capability fully in-house rather than licensing external IP. | 中 | SR022, SR023 |
| CR021 | NVIDIA's strategic response to the custom-silicon shift (selling integrated racks/systems, opening NVLink to third-party ASICs) could further entrench its ecosystem lock-in in ways that leave less room for a third-party power-efficiency IP vendor like Velaura. | 中 | SR023 |
| CR022 | A US structural power/grid shortfall (estimated at 9.3 GW in 2026) could delay hyperscaler data-center buildouts broadly, which could either increase demand for Velaura's efficiency technology or delay the underlying capex cycle Velaura's revenue ultimately depends on. | 中 | SR024, SR025 |
| CR023 | Velaura's general privacy notice (dated August 2024, published under the prior Auradine, Inc. name) had not been updated to reflect the Velaura AI rebrand as of the version reviewed, a minor compliance/currency gap. | 中 | SR026 |
| CR024 | Velaura's public narrative and investor commentary concentrate heavily on CEO Rajiv Khemani and CDO Manu Gulati, creating meaningful key-person dependence risk given the near-total absence of independent biographical detail on the rest of the executive bench. | 中 | SR027 |
| CR025 | No public source reviewed identifies any executive departure, resignation, or leadership change at Velaura AI/Auradine in 2025 or 2026, a mild positive signal for management continuity risk as of the run date. | 中 | SR008, SR015 |
| CR026 | Samsung Catalyst Fund is simultaneously a Velaura investor and the venture arm of Samsung Electronics, whose foundry business is a plausible manufacturing partner or competitor for Velaura's power-efficiency IP, creating a potential (undisclosed) conflict-of-interest or dependency risk. | 中 | SR030 |
| CR027 | Groq's need to raise a second large capital round ($350M in August 2026) despite a prior $750M raise illustrates persistent capital intensity in the AI-silicon sector even for well-funded independent competitors, a comparable financial-risk signal for assessing Velaura's own runway needs. | 中 | SR032 |
| CR028 | Cerebras's IPO at approximately a $56 billion valuation and Groq's roughly $20 billion NVIDIA deal both illustrate that AI-silicon sector valuations can swing dramatically based on single events (M&A, IPO), a volatility precedent relevant to assessing the durability of Velaura's own $1B+ valuation. | 中 | SR031, SR032 |
| CR029 | This chapter's mitigations analysis finds that Velaura's 30M+ ASIC production history (from its legacy Teraflux business) is the strongest available mitigant against pure execution-risk concerns, though it does not address the customer-concentration, disclosure, or commoditization risks identified separately. | 中 | SR028 |
| CR030 | The most severe unresolved risk identified in this chapter is customer/revenue concentration: Velaura's only independently corroborated customer relationships (MARA, GDA) are both in its de-prioritized legacy hardware business, and MARA is also a related-party investor, while the AI-silicon business underpinning the valuation has zero named customers. | 高 | SR006, SR009 |
| CR031 | A monitorable thesis-break trigger for Velaura would be the announcement (or confirmed absence) of a named hyperscaler design win within 12-18 months of the August 2026 Series A, given the company's own stated use of proceeds to expand customer-facing engagement capacity. | 中 | SR029 |
| CR032 | A second monitorable thesis-break trigger would be any disclosure that a major EDA vendor (Synopsys, Cadence) or foundry partner (TSMC, Samsung) has released a directly competing low-power IP library or design kit, which would materially undermine Velaura's differentiation. | 中 | SR016, SR017 |
| CR033 | Velaura's foundry dependency (advanced 3nm/2nm process nodes) means any foundry capacity constraint, price increase, or geopolitical disruption (e.g., Taiwan-related tensions given TSMC's role) would directly bottleneck Titan Core's physical delivery, independent of Velaura's own execution. | 中 | SR001, SR002 |
| CR034 | No public source discloses any environmental, safety, or product-recall incident involving Velaura AI or its legacy Teraflux hardware, a mitigating (absence-of-adverse-evidence) signal for operational/safety risk as of the run date. | 中 | SR015, SR009 |
| CR035 | Velaura's royalty-based revenue model depends on a power-savings measurement methodology that is entirely undisclosed, creating a financial/model risk that revenue recognized could be disputed by customers or difficult to audit independently. | 中 | SR028, SR009 |
| CR036 | The broader AI patent litigation surge (over 1,000 lawsuits in five years) combined with Velaura's own unverified patent claims suggests IP-related legal risk should be weighted as material even though no specific case names Velaura as of the run date. | 中 | SR004, SR028 |
| CR037 | Velaura's board composition (including Intel CEO Lip-Bu Tan and MARA CEO Fred Thiel) provides governance credibility that partially mitigates key-person execution risk at the board oversight level, even though operating key-person risk remains concentrated in Khemani and Gulati. | 中 | SR027 |
| CR038 | This chapter finds no evidence of fraud, restatement, or accounting irregularity at Velaura AI or in MARA's disclosures about its Velaura investment; the ASC 321 fair-value gain accounting described in MARA's 10-Q follows standard equity-investment measurement practice. | 中 | SR006 |
| CR039 | Velaura's exposure to China-related export control risk is indirect but real: if any of its undisclosed 'three of four largest cloud providers' hyperscaler engagements involve China-serving infrastructure or China-linked foundry capacity, the January 2026 BIS rule changes would directly affect deal economics and timelines. | 中 | SR001, SR002, SR009 |
| CR040 | Velaura's Physical AI and Edge AI segments carry additional undisclosed regulatory risk (e.g., automotive/robotics functional-safety certification requirements such as ISO 26262) that the company has not addressed in any public material reviewed. | 低 | SR028 |
| CR041 | No public source discloses any sanctions-list exposure, denied-party-list match, or export-control enforcement action involving Velaura AI, Auradine, or its named leadership as of the run date. | 中 | SR033, SR001 |
| CV001 | Velaura AI closed a $110 million Series A on August 18, 2026, at a valuation exceeding $1 billion, led by Seligman Ventures. | 高 | SV007, SV015 |
| CV002 | No public source discloses the exact post-money valuation figure, share price, or liquidation-preference structure of Velaura's Series A beyond the 'more than $1 billion' headline figure. | 中 | SV007, SV013 |
| CV003 | Arm Holdings, the closest public comparable for a royalty-based semiconductor IP licensing business, had a market capitalization of $270.57 billion as of August 18, 2026, up 83.93% year over year. | 高 | SV001, SV002 |
| CV004 | Arm Holdings reported fiscal year 2026 revenue of $4.92 billion (trailing-twelve-month revenue of $5.16 billion), implying a price-to-sales ratio of roughly 52-55x at its August 2026 market capitalization. | 中 | SV003 |
| CV005 | Semiconductor IP licensing companies broadly traded at 8-12x EV/Revenue for high-quality, growing firms in 2026, with best-in-class names like Arm trading well above that range, per industry valuation-multiple analysis. | 中 | SV004 |
| CV006 | The global semiconductor IP licensing market itself was estimated at roughly $8-9 billion in 2025, forecast to grow at 10-12% CAGR toward $13-18 billion by 2030, providing a market-size ceiling context for any single IP licensor's realistic revenue scale. | 中 | SV005 |
| CV007 | MARA Holdings' SEC 10-Q filing independently corroborates a rising valuation mark for Velaura: MARA recorded $11.9 million and $2.7 million in ASC 321 fair-value gains on its preferred and common stock holdings when a later Velaura financing round established a higher observable price. | 高 | SV006, SV007 |
| CV008 | Groq, a directly comparable independent AI-silicon company, saw its valuation reset from $6.9 billion (September 2025) to $3.5 billion (August 2026) following a reported $20 billion NVIDIA deal that absorbed its founding technical team and IP, illustrating how quickly AI-silicon valuations can compress. | 高 | SV008, SV009 |
| CV009 | Cerebras Systems went public on May 14, 2026, closing its first trading day at approximately a $56 billion fully diluted valuation, the largest US tech IPO since Snowflake in 2020, setting a high-water mark for independent AI-silicon comparables. | 中 | SV010 |
| CV010 | Tenstorrent remains an independent, privately held AI accelerator company as of mid-2026 without a disclosed unicorn-level valuation reset comparable to Velaura's, Groq's, or Cerebras's headline figures. | 中 | SV011 |
| CV011 | Velaura's Series A valuation of $1B+ sits far below Arm's $270B+ public market cap and Cerebras's ~$56B IPO valuation, but far above Tenstorrent's undisclosed (likely sub-unicorn at last public data point) private valuation, positioning Velaura in the middle of the current AI-silicon comparable set by headline valuation. | 中 | SV001, SV010, SV011, SV007 |
| CV012 | No public source discloses Velaura's revenue, ARR, or any other financial metric that could be used to compute a revenue multiple implied by its $1B+ valuation, making direct multiple-based comparison to Arm or other public comparables impossible with current evidence. | 高 | SV013, SV007 |
| CV013 | Applying even a conservative 8x revenue multiple (the low end of semiconductor IP licensing comparables) to Velaura's $1B+ valuation would imply roughly $125 million in annual revenue would be needed to justify the price on a pure multiple basis - a figure with no public evidence of support. | 低 | SV004, SV007 |
| CV014 | Velaura's investor base includes both financial VCs (Seligman Ventures, Capricorn, Prosperity7, Mayfield, Maverick Silicon, Premji Invest, StepStone) and strategic/related-party investors (Samsung Catalyst Fund, MARA), a mixed cap-table structure common in late-stage private financings but one that adds related-party complexity to valuation interpretation. | 中 | SV021, SV022, SV023, SV006 |
| CV015 | No public source discloses Velaura's liquidation preference stack, anti-dilution terms, or option pool sizing, all standard inputs for assessing whether the headline $1B+ valuation reflects common-equity value or is inflated by preferred-stock structuring. | 低 | SV017, SV007 |
| CV016 | Heise Online's independent reporting flags that Velaura does not disclose which chips or customers use its IP, directly limiting the ability to independently verify the revenue base implied by its valuation. | 中 | SV013 |
| CV017 | TechBooky's coverage explicitly cautions that a unicorn valuation does not guarantee commercial adoption, a direct overvaluation-risk caveat given Velaura's lack of named AI-silicon customers. | 中 | SV014 |
| CV018 | Untether AI, a comparable low-power AI chip startup, went from operating to bankrupt within roughly 16 months (June 2025 shutdown, October 2025 bankruptcy filing with $128M+ liabilities), illustrating the downside tail-risk scenario for a technically credible but commercially unproven AI-silicon company. | 高 | SV029, SV030 |
| CV019 | Independent analyst Patrick Moorhead of Moor Insights & Strategy provided qualified, non-binding support for Velaura's technical approach, saying it 'has the potential' to reduce total cost of ownership - a hedged endorsement rather than an independent valuation opinion. | 中 | SV016 |
| CV020 | Macro tailwinds supporting Velaura's valuation thesis include a 9.3 GW US power shortfall and hyperscaler AI capex projected at $600-725 billion in 2026, both of which increase the theoretical value of performance-per-watt technology if Velaura can convert engagements into signed licenses. | 中 | SV027, SV028 |
| CV021 | Custom-silicon shipment growth (44.6% CAGR) outpacing merchant GPUs (16.1% CAGR) expands the theoretical addressable base of hyperscaler in-house ASIC programs that could license Velaura's IP, a supporting comparable-growth data point for the bull case. | 中 | SV020 |
| CV022 | Arm's historically estimated royalty rate of roughly 1-2% of chip ASP provides an external analogy suggesting Velaura's own per-unit royalty take, once disclosed, is likely to be a similarly small fraction of the underlying chip's total value rather than a large share. | 中 | SV018 |
| CV023 | Arm's 2026 pivot to selling finished chips directly (breaking a 35-year pure-licensing model) is a cautionary precedent suggesting that even Velaura's chosen business model comparable is evolving away from pure licensing, adding long-term business-model risk to any valuation built on the assumption of durable royalty economics. | 中 | SV019 |
| CV024 | This chapter's bull case assumes at least one of Velaura's three undisclosed hyperscaler engagements converts to a signed, named design win within 18 months, translating macro tailwinds and technical credibility into verifiable revenue. | 中 | SV007, SV020 |
| CV025 | This chapter's bear case assumes none of the three hyperscaler engagements converts within 18 months and the legacy Teraflux hardware business continues to decline as Bitcoin miners pivot to AI/HPC colocation, leaving Velaura dependent on further equity financing without a revenue inflection. | 中 | SV013, SV014 |
| CV026 | This chapter's base case assumes gradual, partial conversion of hyperscaler engagements (e.g., one design win with multi-year ramp to production) alongside continued legacy-business decline, consistent with typical 12-24 month semiconductor IP qualification cycles. | 中 | SV007, SV020 |
| CV027 | Velaura's own materials estimate up to $100 million per year in power savings for a typical data center customer, a company-modeled figure that (if realized and monetized via royalty) could support meaningful per-customer revenue, though no royalty rate is disclosed to convert this into Velaura's own revenue. | 中 | SV024 |
| CV028 | No public source discloses any secondary-market transaction, tender offer, or employee share sale involving Velaura AI equity, leaving liquidity and exit-readiness entirely unproven for existing shareholders. | 低 | SV017, SV025 |
| CV029 | Headcount estimates for Velaura diverge across sources (62 per Tracxn vs. 100-150 per other analyst databases), complicating any attempt to benchmark valuation-per-employee against comparable semiconductor startups. | 中 | SV026, SV025 |
| CV030 | This chapter finds no evidence of a down round, valuation impairment, or write-down affecting Velaura AI specifically; MARA's SEC filing instead shows an upward fair-value adjustment, a mitigating signal against near-term overvaluation-correction risk. | 中 | SV006 |
| CV031 | The AI-silicon comparable set shows extreme valuation volatility within a single year (Groq's ~50% valuation decline, Cerebras's IPO at ~$56B, Untether AI's total loss), indicating Velaura's own $1B+ valuation should be treated as a point-in-time mark rather than a stable, durable figure. | 高 | SV008, SV009, SV010, SV029, SV030 |
| CV032 | Velaura's asset-light, royalty-based business model requires less capital intensity than a merchant chip vendor (no wafer-start financing risk directly on its balance sheet), a structural feature that could support a premium valuation multiple relative to fabless comparables if royalty revenue materializes. | 中 | SV018, SV031 |
| CV033 | Given the near-total absence of disclosed revenue, royalty rate, and named customers, this chapter's recommendation leans toward 'research-more' rather than a confident buy or avoid call, consistent with the skill's guidance to prefer research-more over false precision when valuation inputs are missing. | 高 | SV013, SV007 |
| CV034 | A thesis-break trigger for Velaura's valuation would be confirmation that any of the three unnamed hyperscaler engagements has been terminated or has failed to progress past technical evaluation within 18 months of the Series A close. | 中 | SV007 |
| CV035 | A second thesis-break trigger would be a disclosed down round, valuation impairment, or MARA write-down of its Velaura investment in a future SEC filing. | 中 | SV006 |
| CV036 | Velaura's board composition (Intel CEO Lip-Bu Tan, MARA CEO Fred Thiel, multiple experienced VCs) provides governance and exit-network credibility that could support a future IPO or strategic acquisition path, similar to how Cerebras and Groq's boards facilitated their respective IPO and M&A outcomes. | 中 | SV010, SV008 |
| CV037 | No public source discloses any planned IPO timeline, target exit multiple, or acquisition interest specific to Velaura AI as of the run date. | 低 | SV007, SV015 |
| CV038 | The final diligence priority this chapter identifies is obtaining Velaura's actual revenue, royalty rate, and at least one named hyperscaler customer reference, since every valuation methodology attempted in this chapter (comparable multiple, macro-tailwind analysis, cap-table review) is constrained by the absence of these three inputs. | 高 | SV013, SV007, SV006 |
| CV039 | Velaura's valuation stance in this chapter is assessed as 'unresolved/research-more' rather than 'attractive' or 'expensive' because the comparable set (Arm's 52-55x P/S, Cerebras's $56B IPO, Groq's volatile reset) spans too wide a range to triangulate a specific verdict without Velaura-specific revenue data. | 高 | SV001, SV010, SV008 |
| CV040 | Samsung Catalyst Fund's dual role as investor and Samsung's foundry-linked venture arm could support Velaura's valuation if it translates into preferential foundry access, but this potential synergy is undisclosed and cannot be credited in this chapter's valuation analysis without confirmation. | 中 | SV022 |
| CV041 | This chapter's overall recommendation reflects strong qualitative support (production-proven technology, credible board, favorable macro tailwinds) offset by a near-total absence of quantitative revenue/customer evidence, yielding a 'track' rather than 'buy' or 'avoid' recommendation as the most evidence-consistent call. | 高 | SV024, SV020, SV013 |
| CV042 | Global unicorn count reached nearly 1,700 companies representing $8.2 trillion in aggregate value by Q2 2026, with late-stage venture capital increasingly concentrated in AI, semiconductor, and hard-tech sectors - the broader market context within which Velaura's own unicorn valuation was achieved. | 高 | SV032, SV033 |
| CV043 | Industry venture-capital benchmarking data (NVCA) provides a broader deal-volume and valuation-trend context for 2026 late-stage financings, though it does not break out semiconductor-specific or Velaura-specific figures. | 中 | SV033 |