Celero Communications
面向 AI 规模光互连的相干 DSP 芯片
Celero Communications 是一家技术路径可信的相干 DSP 初创公司,背后有 Alphabet 支持,也有顶尖半导体创始团队,但公司尚未放量、收入未经验证;若以约 $1B 估值进入,执行风险很高。
封面要素
公司概况
Celero Communications 是一家无晶圆厂半导体公司,由 Nariman Yousefi(CEO,曾任 Marvell/Inphi)和 Oscar Agazzi(CTO,曾任 Marvell/Broadcom/ClariPhy)于 2024 年创立。公司设计相干 DSP ASIC,处理复杂光调制格式(如 DP-16QAM/DP-64QAM),让 AI 数据中心内部的机架到机架光链路达到 400G–1.6T+。它的芯片瞄准 GPU 集群和分布式 AI 训练负载快速扩张带来的带宽瓶颈。Celero 在种子轮、A 轮及 2025 年 11 月由 CapitalG/Alphabet 领投的 B 轮中累计融资约 $140M,对应投后估值约 $1B。公司位于加州 Irvine,并在 Victoria(加拿大 BC 省)和 Buenos Aires(阿根廷)设有设计中心。
- 成立时间
- 2024-01-01
- 创始人
- Nariman Yousefi, Oscar Agazzi
- 创立地点
- Irvine, California, USA
- 总部
- Irvine, California, USA
- 产品
- 定制相干 DSP ASIC,在 400G、800G 和 1.6T+ 线速下完成光信号在数字域与模拟域之间的转换。芯片把前向纠错、自适应均衡、载波恢复和色散补偿放进硅片,去掉限制 AI 集群扩张的光传输距离和带宽约束,又不承担传统电信级相干模块的功耗和成本溢价。
- 客户
- 超大规模云服务商(Google/Alphabet、Microsoft、Amazon、Meta)以及大型 AI 基础设施运营商(CoreWeave、xAI、Lambda Labs)。这些机构在建设高密度 GPU 集群,需要超出传统直接探测光模块覆盖范围的芯片到芯片、机架到机架光连接。
- 商业模式
- 无晶圆厂半导体模式:Celero 设计相干 DSP 芯片,并把制造外包给领先代工厂(推定为 TSMC 先进节点)。收入按颗芯片销售产生,定价反映相干光器件的高性能、定制 ASIC 属性(约 $500–$5,000+ / 颗,取决于速度等级和出货量)。
- 阶段
- Series B
- 融资情况
- 累计融资 $140M:由 Sutter Hill Ventures 和 Valor Equity Partners 领投的 $40M 种子轮 / A 轮;2025 年 11 月由 CapitalG(Alphabet 旗下成长型风投)领投的 $100M B 轮,对应投后估值约 $1B。
执行摘要
主要优势
- 世界级创始团队:Nariman Yousefi(Marvell/Inphi)和 Oscar Agazzi(ClariPhy/Broadcom)是当今资历最硬的相干 DSP 架构师之一,手里有深厚 IP,也握有超大规模云厂商关系。
- 战略投资人信号:CapitalG(Alphabet)领投 Series B,强烈暗示 Google 是设计合作伙伴或早期商业化候选客户,需求侧风险大幅降低。
- 巨大市场拉力:AI 集群带宽需求的翻倍速度快于互连供给;相干 DSP 是 1.6T+ 机架级链路唯一可行路径,Celero 正卡在关键瓶颈上。
- 不对称 M&A 可选性:Marvell 2021 年以 $10B 收购 Inphi,加上 Alphabet 支持的硬件收购先例,说明 Celero 在达到可收购规模后是可信的战略收购标的。
主要风险
- ASIC 流片执行风险:首代先进制程相干 DSP 单次流片成本达 $10–50M+,失败率不低;只要进度滑 12–18 个月,入场估值就会明显受损。
- TSMC 单一晶圆厂集中:台湾供应中断、地缘政治限制或先进制程产能分配被挤压,都会让 Celero 制造管线停摆,短期没有替代方案。
- 架构替代:线性驱动光学(LPO)、共封装光学(CPO)或硅光子采用提速,可能在 Celero 放量前压缩独立相干 DSP 芯片的可服务市场。
- 资本强度与稀释:从工程样品爬坡到超大规模云厂商量产,需要继续投入大量资本,现有投资人面临显著稀释风险。
- 出口管制暴露:BIS 半导体法规(EAR Part 774)和 CHIPS Act 条款带来持续合规义务;任何失误都可能限制晶圆厂准入或客户关系。
未决问题
- 当前收入或 ARR(未公开披露;投资判断必需)
- 芯片级毛利率和单位经济模型(未披露)
- 具名量产客户或已确认设计定点(保密)
- Series B 条款清单中的准确投后估值(官方未确认)
- 股权结构表、清算优先权和老股交易价格(不可得)
- 具体 TSMC 制程节点和流片里程碑时间表(专有信息)
- 按职能和地点划分的员工数(未发布)
目录
01公司概览
1.1 身份、阶段与运营版图
Celero Communications 更适合被理解为一家已完成 B 轮融资的私营半导体初创公司,正在为 AI 网络连接打造相干数字信号处理芯片。公司官网首页和关于页面反复把使命落在同一件事上:AI 训练和推理集群一旦跨出单个机架、甚至单个园区,就会遇到距离、带宽和功耗瓶颈,Celero 要把这些瓶颈移走。公开市场数据资料也吻合这个定位:CB Insights 将公司列为 B 轮,并把产品描述为面向 AI 光传输的相干 DSP;Reuters 则更具体地解释为,把光纤中的光信号转换为计算网络使用的电域信号。地理分布在这个故事里也格外重要。Celero 官方总部位于加州 Irvine,招聘页和关于页则显示其工程团队分布在 Ottawa、San Jose,以及加拿大和阿根廷更广泛的设计中心。这一版图更像是一家公司仍在扩建,而不是已经披露营收、客户数或员工总数等商业规模指标。[CO001, CO002, CO003, CO004, CO005, CO006]
| 指标 | 数值 / 状态 | 日期 | 置信度 | 证据缺口 |
|---|---|---|---|---|
| 成立 / 设立 | 成立于 2024 年;加州备案足迹日期为 2024-02-20 | 2024-02-20 至 2024 | 高 | 正式备案前的运营启动日期未公开拆分 |
| 总部 | 加州 Irvine;街道地址多次显示为 5281 California Ave., Suite 200 | 当前公开资料访问于 2026-07-25 | 中 | 官网写明 Irvine,但未给出完整街道地址 |
| 运营地点 | Irvine 总部;Ottawa 和 San Jose 职位;加拿大和阿根廷设计中心 | 当前页面访问于 2026-07-25 | 中 | 加拿大和阿根廷未在官方页面拆到城市层级 |
| 当前阶段 | Series B 私营公司 | 当前第三方资料加 2025 年融资记录 | 中 | 没有公开股权结构或投后估值披露 |
| 产品 / 商业模式 | 面向 AI 光互连和 scale-across 网络的相干 DSP 平台 / 芯片 | 当前公司和第三方描述 | 高 | 客户细分信息仍偏泛 |
| 累计融资 | 公司称 $140M;CB Insights 列示 $138.4M | 2025-11-17 公告;2026 年资料访问 | 中 | 公开来源未调和这点小差异 |
| 最新轮次 | 100M USD Series B 轮,CapitalG 领投 | 2025-11-17 | 高 | 具体证券组合和条款未公开 |
| 估值状态 | 具体投后估值未公开披露 | 截至 2026-07-25 的当前公开记录 | 中 | 无法从已审阅来源确认独角兽状态 |
| 收入 / ARR | 未公开披露 | 截至 2026-07-25 的当前公开记录 | 中 | 没有可用的运营财务规模指标 |
| 客户 / 合作伙伴 | 没有公开披露客户数量或具名客户证明 | 截至 2026-07-25 的当前公开记录 | 中 | 商业牵引仍基本未公开 |
| 招聘 / 人数代理指标 | 可见 15 个开放职位;具体员工数未披露 | 2026-07-25 访问日期 | 中 | 招聘是扩张信号,但不是员工人数 |
这张快照混合了官方声明、备案和市场数据资料。未知财务和估值单元格是明确证据缺口,不是隐含负面结论。
[CO003, CO004, CO005, CO007, CO019, CO020]Celero 的身份叙事把经验丰富的相干 DSP 创始团队、分布式工程布局、AI 光互连产品和较大资本底座串起来,但商业和治理披露缺口仍未补上。
[CO001, CO002, CO009, CO010, CO011, CO019]1.2 领导层、治理与关键人物风险
创始人连续性是 Celero 公开记录中最清晰的正面信号之一。多项来源将 Nariman Yousefi 标注为联合创始人兼 CEO,将 Oscar Agazzi 标注为联合创始人兼 CTO,并称两人都曾在 Marvell、Inphi、Broadcom 和 ClariPhy 担任高层。这一背景契合产品命题:Celero 不是在发明一个新的终端市场,而是把成熟的相干光和网络芯片人才用到 AI 集群连接上。不过,治理可见度远低于融资新闻。2024 年 SEC Form D 比官网给出更多结构信息:Nariman 和 Oscar 是高管兼董事,Stefan Dyckerhoff 和 Robert Abbott 也列为董事。2025 年 11 月融资公告新增 CapitalG 的 James Luo,并再次提到 Stefan Dyckerhoff 是与董事会相关的投资人;加州注册信息镜像还新增 Avijit Roy 为 CFO。即便有这些碎片,Celero 仍未公开完整董事名单、治理文件或更厚的高管梯队,关键人物依赖仍明显集中在创始人身上。[CO009, CO010, CO011, CO012, CO013, CO014]
| 人物 | 职位 | 背景 | 创始人-市场匹配 | 关键人物依赖 |
|---|---|---|---|---|
| Nariman Yousefi | 联合创始人、CEO、董事 | 前 Marvell / Inphi / Broadcom / ClariPhy 高管;签署 Form D | 与相干网络芯片的商业和产品需求高度匹配 | 高:公开代表、CEO、备案签署人和注册代理人 |
| Oscar Agazzi | 联合创始人、CTO、董事 | 前 Marvell / Inphi / Broadcom / ClariPhy 高管 | 与相干 DSP 架构的技术需求高度匹配 | 高:产品可信度集中在创始 CTO 专长上 |
| James Luo | CapitalG GP;董事会成员 | 成长阶段投资人,在 Series B 加入 | 表明最新轮后引入机构监督 | 中:治理影响看起来更多来自投资人身份,而非运营角色 |
| Stefan Dyckerhoff | Sutter Hill MD;董事会成员 / 早期董事披露 | 早期投资人,在 SEC 和融资材料中有董事会可见度 | 为 Series B 前融资历史带来连续性 | 中:显示早期投资人影响力仍在 |
| Robert Abbott | 董事(2024 年 Form D 公开披露) | 已审阅公司面向外部的材料未解释其角色 | 至少为公开记录增加一名非创始人董事 | 中:职能和当前状态需要尽调跟进 |
| Avijit Roy | 2026 年注册信息镜像显示为 CFO | 财务高管出现在注册数据衍生资料中,但未出现在官网 | 表明创始人之外可能已有更宽团队 | 中:官方公司确认和职责范围都很薄 |
覆盖范围是部分的,因为公开材料只提供创始人和备案层面的治理碎片,而不是完整管理组织图或委员会材料包。
[CO009, CO010, CO011, CO012, CO013, CO014]1.3 资本结构、投资方与指标披露
资本形成是 Celero 公司概览中佐证最充分的部分。公司 2025 年 11 月 17 日发布的公告、Business Wire 分发稿和 Reuters 报道指向同一条主线:累计融资 $140M,其中包括 CapitalG 领投的 $100M B 轮,以及由 Sutter Hill Ventures 领投、Valor Equity Partners、Atreides Management、Maverick Silicon 等参与、合计 $40M 的种子轮与 A 轮。2024 年 SEC Form D 还给出更早的监管锚点:记录近 $30.0M 已售予 40 名投资者,首次销售日期为 2024 年 12 月 11 日;Intelligence360 后续报道了 2025 年末最高约 $100M 的另一项通知。薄弱处在经济条款细节。CB Insights 将 Celero 归为 B 轮并列示已融资 $138.4M,但本次审阅的公开记录没有披露精确投后估值、债务融资、老股转让或投资人控制权。因此,投资人可以验证融资轮次和联合投资方,却还无法只凭公开证据核算完整资本结构。[CO019, CO020, CO021, CO022, CO023, CO024]
| 利益相关方 | 角色 | 控制 / 经济重要性 | 尽调要求 |
|---|---|---|---|
| CapitalG | $100M Series B 领投方;通过 James Luo 与董事会相连 | 最新轮领投方可能拥有重要治理和信息权 | 索取董事会或观察员权利、按比例跟投条款,以及任何战略支持义务 |
| Sutter Hill Ventures | 早期种子轮和 Series A 领投方;通过 Stefan Dyckerhoff 与董事会相连 | 最早具名领投方;可能对公司设立和治理有影响 | 索取最早期条款清单、持股比例和保护性条款 |
| Valor Equity Partners | 早期轮次具名参与方 | 属于 Series B 前财团,也是潜在后续支持基础 | 澄清持股、后续参与以及任何特殊权利 |
| Atreides Management | 早期轮次具名参与方 | 体现跨界资金或主题型资金对 AI 基础设施论点的支持 | 澄清出资额,以及 Series B 后持股是否仍然活跃 |
| Maverick Silicon | 早期轮次具名参与方 | 半导体专门投资人即便经济权益较小,也可能有战略价值 | 澄清董事会访问权、技术顾问角色和当前持股 |
| Nariman Yousefi 和 Oscar Agazzi | 创始高管,可能是核心普通股持有人 | 创始人仍是运营重心和关键尽调依赖 | 索取股权结构集中度、归属、留任和继任规划细节 |
| Robert Abbott / 其他未具名董事 | 除投资人董事外公开披露的治理参与者 | 表明治理范围比官网呈现更宽,但披露仍不足 | 索取完整现任董事会名单、独立性状态和委员会结构 |
覆盖范围是部分的,因为公开证据识别了具名融资方和部分董事会相关利益相关方,但未披露完整股权结构、持股比例或控制权。
[CO009, CO010, CO012, CO013, CO016, CO019]公司资金充足、技术可信度高,但能否投资仍取决于估值、商业牵引和资本结构细节的私下尽调。
[CO007, CO019, CO020, CO027, CO028, CO030]1.4 里程碑、规模信号与负面提醒
Celero 带日期的公开记录显示,这家年轻公司从法律设立快速推进到大额融资,但披露密度没有达到硬件初创公司进入这一融资阶段后通常会看到的水平。最早的硬里程碑来自法律和监管:2024 年 2 月的加州备案痕迹、2024 年 12 月 Form D 中 40 名投资者和近 $30M 已售金额,以及年末前不久的签署日期。外部叙事在 2025 年 11 月加速:公司公布累计融资 $140M,Reuters 独立解释芯片架构和使用场景,James Luo 与 Stefan Dyckerhoff 的引述让投资人 / 董事会参与变得明确。后续一份豁免发行通知显示 B 轮期间还有额外融资文件,而 2026 年注册信息和招聘页快照显示公司仍在多地点扩张。最强的负面信号不是法律或监管麻烦,而是披露不完整:Tech in Asia 认为 Celero 尚未分享足够技术细节、流片状态或合作证明,无法证明它在拥挤的相干 DSP 领域拥有持久优势。这一点重要,因为客户、营收和估值证据仍然稀疏。[CO007, CO008, CO019, CO023, CO024, CO025]
| 日期 | 事件 | 类型 | 金额 / 估值 / 状态 | 参与方 | 含义 |
|---|---|---|---|---|---|
| 2024-02-20 | Celero Communications, Inc. 的加州备案足迹 | 创立 | 已备案 / 后续存续良好 | Nariman Yousefi;BizProfile 引用的 California Secretary of State 数据 | 公司概览时间线的硬法律起点 |
| 2024-12-11 | 2024 年 Form D 记录的首次出售日期 | 融资 | Reg D 销售进行中 | Celero Communications, Inc. | 显示机构融资在 2025 年公告潮前已经开始 |
| 2024-12-23 | CEO 签署了 2024 年 Form D 提交文件 | 治理 | 签署日期 | Nariman Yousefi | 确认创始人 CEO 亲自执行早期融资文件 |
| 2024-12-26 | SEC 发布 2024 年 Form D,显示已向 40 名投资人出售近 $30.0M | 融资 | 29,999,906 USD 已售 | SEC;Celero 投资人 | 为早期轮次叙事提供监管锚点 |
| 2025-11-17 | 公司宣布累计融资 $140M,其中包括 $100M Series B | 融资 | 140M 累计;100M Series B | 投资人:CapitalG;Sutter Hill;Valor;Atreides;Maverick | 让 Celero 从低调初创公司变成大额资本支持的私有硬件故事 |
| 2025-11-17 | 融资公告公开把 James Luo 和 Stefan Dyckerhoff 与 Celero 董事会角色相连 | 治理 | 投资人董事会相关引述 | CapitalG;Sutter Hill Ventures | 显示最新和最早领投方都参与董事会 |
| 2025-11-17 | 官方材料重申总部位于 Irvine,并在加拿大和阿根廷设有设计中心 | 规模 | 当前足迹已披露 | Celero Communications | 固定了后续章节使用的基本运营足迹 |
| 2025-11-18 | Reuters 独立描述了 Celero 的芯片架构和 AI 数据中心间使用场景 | 产品 | 独立报道 | Reuters;Nariman Yousefi | 在公司新闻稿之外增加第三方验证 |
| 2025-11-18 | Tech in Asia 发布了关于缺少规格、流片状态和合作证明的审慎说明 | 负面 | 披露警示 | Tech in Asia 报道 | 引入最清晰的公开负面尽调信号 |
| 2025-12-02 | Intelligence360 报道一则最高 $99,999,994 新融资通知 | 融资 | 最高 99,999,994 USD | Intelligence360;SEC 备案引用 | 提示 Series B 时期前后存在后期融资文件 |
| 2026-03-25 | 注册数据衍生资料显示活跃状态、Irvine 地址,以及 Avijit Roy 任 CFO | 规模 | 活跃 / 存续良好快照 | BizProfile;加州注册数据衍生资料 | 显示融资后组织继续成熟 |
| 2026-07-25 | 官网招聘页显示 Ottawa、Irvine 和 San Jose 共 15 个开放职位 | 规模 | 当前招聘快照 | Celero Communications | 当前招聘显示融资后扩张,尽管员工数未披露 |
这条时间线穷尽了第 1 章已审阅来源中恢复到的带日期里程碑,范围为 2024-02-20 至 2026-07-25。
[CO003, CO004, CO005, CO007, CO012, CO013]Celero 的公开记录显示,公司从 2024 年成立和融资备案,很快走到 2025 年大额 Series B;到 2026 年,公开信息仍以招聘和披露缺口为主,而不是商业指标。
[CO019, CO020, CO023, CO024, CO032, CO033]02市场分析
2.1 市场边界、纳入支出与现状替代方案
Celero 瞄准的市场比整个光网络更窄,却比传统长途相干电信更宽。核心纳入支出是光信号处理和互连层,它让 AI 集群从机架扩展到排、园区,并最终扩展到多站点计算织构:相干或轻量相干 DSP、相干可插拔模块、光收发器,以及未来取代电连接瓶颈的 CPO 或近封装光引擎。Cisco 的 DCI 材料和 Celero 自身定位指向同一项待完成任务:在分布式 AI 系统之间移动超大数据量,同时把时延、传输距离、功耗和可维护性控制在可接受范围内。应排除的支出包括通用 GPU 或服务器资本开支、内存、电力基础设施,以及多数传统运营商传输预算。今天最重要的替代方案仍是机架内短距 PAM4、铜缆或 AEC 链路,以及更长距离上的传统电信相干。Celero 的命题成立,是因为两种现状都无法完整覆盖 AI 跨域扩展问题:PAM4 和电连接在短距上胜在简单,传统相干则为高密度数据中心经济性配置过重、功耗过高。[CM001, CM002, CM003, CM004, CM005, CM006]
| 细分 / 类别 | 纳入支出 | 排除支出 | 买方 / 付款方 | 相关性 |
|---|---|---|---|---|
| AI scale-out 光互连核心 | 光 DSP、光收发器、相干可插拔模块,以及用于机架间和排间 AI 网络结构的交换机光学器件 | GPU 计算、服务器、内存、通用云资本开支 | 超大规模云和 AI 云基础设施预算,经由 OEM 和交换机厂商落地 | Celero 必须拿下设计赢单的核心市场 |
| AI scale-across / 园区相干链路 | 楼宇间和园区 DCI 光学、coherent-lite / ZR-class 模块,以及绑定 AI 集群的光线路系统 | 与 AI 园区无关的运营商骨干传输 | 云网络架构和 DCI 预算负责人 | 相关性很高的延伸,因为 Celero 围绕 scale-across 定位 |
| 未来 scale-up 光层 | CPO、近封装光学,以及把高带宽链路移近交换机和加速器的光引擎 | 当前距离下仍然足够的成熟纯铜机架内链路 | 平台架构团队、加速器系统设计方、超大规模云厂商 | 重要的未来选项,但时点仍很敏感 |
| 短距电连接替代品 | 当距离、成本和可维护性仍偏向电连接时使用的铜缆、AEC、重定时器和 PAM4 | 更长距离光 DCI 或相干应用 | 系统 OEM 和机架级平台团队 | 关键的排除且竞争性支出,因为光学尚未取代它 |
| 传统电信相干 | 长距 / 城域相干栈和嵌入式传输 DSP | 无法承受电信功耗和成本曲线的高密度 AI 数据中心部署 | 运营商和传输运营方 | 有用的技术祖先,但范围过宽且往往过度设计 |
| 通用 AI 基础设施资本开支 | 只包括在加速器、机架和站点之间移动数据所需的网络切片 | 模型训练软件、推理应用、设施和发电 | CIO / 基础设施领导层 | 背景预算池,本身不是 Celero 可服务收入 |
这张表定义可变现的互连层,并把它与计算、设施和运营商传输预算分开,避免扭曲 TAM 主张。
[CM001, CM002, CM003, CM006, CM007, CM008]2.2 规模测算视角:广义光器件 TAM、AI 网络织构需求与更窄的相干窗口
证据支持分层市场视图,而不是一个总括 TAM 标题。Yole 的 2026 年光收发器展望是最宽的可用视角:市场在 2025 年为 $23.4B,到 2031 年达到 $112.3B,预计到期末 AI 将贡献超过 90% 收入。这个外层天花板有用,但它包含的数据通信 光器件范围比 Celero 能直接获取的市场更宽。Dell'Oro 通过 AI 网络和光传输给出中间视角:与 AI 后端网络相关的以太网交换机销售已激增,2025-2029 年加速服务器后端网络的交换机支出预计超过 $100B,光传输市场在 2025 年反弹至 $16B,AI 驱动的 DCI 推动 2026 年增长预测高于 $18B。LightCounting 和 Cignal 再把孔径收窄到堆栈的相干边缘:PAM4 仍是 800G 和 1.6T AI 链路的近期出货赢家,而在跨楼宇距离、功耗和互操作性使其具备吸引力的场景中,相干可插拔和轻量相干正在出现。这意味着 Celero 可触达 SAM 位于广义光器件繁荣之内,但公开证据还无法把它干净切出来。[CM010, CM011, CM012, CM013, CM014, CM015]
| 发布方 | 年份 | 地理范围 | 数值 | 复合年增长率 | 方法论 | 置信度 | 限制 |
|---|---|---|---|---|---|---|---|
| Yole / 数据通信与电信光收发器 | 2025-2031 | 全球 | $23.4B(2025 年)至 $112.3B(2031 年) | 数据通信 CAGR 30% | 自上而下光子市场预测,汇总自 Yole 2026 报告 | 中 | 宽口径光学视角;不是 Celero 专属 SAM,且包含相干 DSP 之外的内容 |
| Dell'Oro / AI 后端交换网络结构 | 2025-2029 | 全球 | >$100B 累计交换机支出 | n/a | 针对加速服务器 AI 后端网络中部署交换机的预测 | 高 | 交换机支出并非全是光学,也没有单独隔离 DSP 捕获 |
| Dell'Oro / 光传输与 DCI 视角 | 2025-2026 | 全球 | $16B(2025 年);2026 年预测 >$18B | 2026 年增长 16% | 由 AI 驱动 DCI 拆分的分析师光传输预测 | 高 | 包含电信传输和 AI 园区之外更宽的光设备 |
| LightCounting / AI 光速率迁移视角 | 2025-2026 | 全球 | 2026 年 1.6T 芯片组 >$2B | n/a | 从收发器和有源线缆预测推导芯片组需求 | 高 | 跟踪的是芯片组需求,不是终端市场收入或 Celero 份额 |
| Cignal / 相干可插拔采用视角 | 2026-2027 | 全球 | 2026 年 800ZRx 端口 >200k;2027 年 1600ZRx | n/a | 相干 DSP 和可插拔出货跟踪 | 中 | 端口预测是窄口径相干视角,不是完整美元 TAM |
本章有意保留多个不可相加的视角,因为公开数据混合了收发器、交换、光传输和相干可插拔模块,而不是发布一个干净的短距相干 DSP 市场规模。
[CM010, CM011, CM013, CM014, CM016, CM019]Celero 位于三层市场口径之内:宽口径的数据通信光学繁荣,更窄的 AI fabric 传输层,以及更窄的相干 DSP 机会集。
各层是不可相加的分析口径。只有 TAM 和中间层美元数值由公开来源直接发布;窄层只做定性描述,因为公开来源没有单独拆出干净的短距相干 DSP 收入池。
[CM010, CM012, CM016, CM019, CM020, CM026]已发布数字横跨宽口径光学外层天花板和更窄的 DCI/传输层,说明单一 TAM 数字会夸大精度。
两行均使用十亿美元。中点值是可视化插值辅助,不是独立发布的市场估算。
[CM010, CM019, CM020, CM040]2.3 买方、用户、付款方与采用路径
买方图谱分多层,因为 DSP 厂商很少直接卖给单一终端用户。最直接的商业客户往往是模块、系统或交换机 OEM,它们把光引擎、可插拔模块或光子器件集成进出货硬件。但超大规模云和 AI 云架构师会重度影响规格,因为任何大规模部署之前,他们都要验证可靠性、功耗、互操作性和可维护性。Cisco、Acacia 和 Nvidia 都先围绕超大规模云定义市场,然后才是第二波新型云、Tier 2 提供商,以及最终运营私有 AI 园区的大型企业。日常用户是网络架构师、光学工程师和负责横向扩展及跨域扩展网络织构的平台团队;最终付款方是希望提升每瓦集群利用率的 AI 基础设施预算负责人。因此,采用路径通常从标准工作和 OEM 设计导入开始,进入有限生产部署;待可用性 和运营得到验证后,再扩展到园区或多站点。Celero 更像是在卖给一个生态体系认证流程,而不是完成一次常规芯片采购。[CM026, CM027, CM028, CM029, CM030, CM031]
| 细分 | 买方 | 用户 | 付款方 | 工作流 | 预算负责人 | 采用触发因素 |
|---|---|---|---|---|---|---|
| 一线超大规模云厂商 AI 工厂 | 交换机 / 模块 OEM,以及超大规模云厂商网络架构团队 | 光网络架构师、平台网络团队、数据中心运维 | 集中 AI 基础设施预算 | 先让光学方案通过横向扩展和园区扩展认证,再放量部署 | 基础设施 / 网络 VP,协同平台工程 | 需要提升每瓦带宽,并把超大集群跨机架、跨站点连起来 |
| 新云 / AI 基础设施提供商 | 具备云运营商直接认证的系统供应商 | 网络工程与容量规划团队 | GPU 集群增长资本开支 | 快速部署 Ethernet 或 InfiniBand 网络,把 GPU 供给转成收入 | COO / 基础设施财务 / 网络平台负责人 | 没有超大规模云厂商自建光学团队,也需要有竞争力的 AI 服务容量 |
| 模块 / 交换机 OEM | 光学组件与系统采购团队 | 产品、硬件、光学集成工程师 | OEM 产品线预算 | 导入符合超大规模云厂商规格的 DSP 和引擎 | GM 或硬件产品线负责人 | 需要低功耗或长距离光学方案,拿下下一代 AI 设计位点 |
| 企业或主权 AI 园区 | 集成商加企业网络架构团队 | 私有 AI 平台团队,以及安全 / 网络运维 | 数据中心现代化预算 | 连接承载训练和推理工作负载的多栋楼宇或托管机房 | CIO / CTO 基础设施组织 | 需要更低延迟的站点间 AI 网络,同时运维可控 |
| 传统运营商 / 广义传输采购方 | 光传输规划师和服务提供商工程师 | 传输运维 | 运营商资本开支 | 采购城域或长途相干系统 | 光传输业务部 | 任务和预算不同于 Celero 的 AI 原生用例 |
买方映射反映组件厂商如何沿认证链进入市场,而不是直接把一款独立企业应用卖给一个终端用户。
[CM009, CM018, CM027, CM029, CM030, CM037]市场首先由 OEM 和超大规模云厂商认证;企业和主权买方要等可靠性和可维护性被证明后才会跟进。
单元格为定性判断,因为公开证据更清楚点名参与者和触发因素,而不是披露预算份额。
[CM009, CM018, CM026, CM027, CM029, CM037]Celero 必须先从生态认证走到有限部署,再进入 scale-across 扩张;之后,更深层的 scale-up 光学迁移才有意义。
这是基于供应商、分析师和运营商证据合成的采用路径,不是 Celero 披露的销售漏斗。
[CM015, CM027, CM033, CM037, CM042]2.4 增长驱动、采用约束与尽调缺口
需求方向很强;难点在采用节奏。独立来源给出的增长驱动很清晰:超大规模云 AI 资本开支仍处高位,以太网正在 AI 后端网络中拿份额,端口速度从 800G 快速迈向 1.6T,Cisco 和 Nvidia 也量化了光器件取代独立传输设备或重度可插拔架构后带来的功耗或可用性 收益。但约束同样实质。LightCounting 显示 PAM4 仍赢得近期 AI 出货爬坡,意味着相干 DSP 厂商必须等待跨域扩展和轻量相干用例成熟。ADTEK、CNBC 和 Gazettabyte 都强调制造良率、可靠性、可维护性和标准成熟度是真正阻碍,而 SemiEngineering 给出更乐观的五年内全光视角。估值会受这些时间差影响,因为它们改变 Celero 从小众设计导入走向大规模部署的时间。公开数据也留下关键缺口:没有来源披露 AI 数据中心短距相干 DSP 的自下而上 SAM,OEM 与超大规模云之间的预算拆分仍是推断而非报告,轻量相干在 AI 园区内的挂接率或 ASP 仍属私有信息。[CM021, CM022, CM023, CM024, CM025, CM028]
| 驱动因素 / 约束 | 方向 | 时点 | 影响 | 尽调问题 |
|---|---|---|---|---|
| 超大规模云厂商 AI 资本开支和 Ethernet 后端扩容 | 正向 | 当前 | 支撑对更高速光学方案的需求,也提高试用新架构的意愿 | 询问 Celero 目前已进入哪些超大规模云厂商或新云项目的认证名单 |
| 从 800G 迁移到 1.6T,并继续走向 3.2T | 正向 | 当前至中期 | 加大电气传输距离、功耗和可维护性的压力 | 要求 Celero 按代际提供通道速率路线图和目标设计位点 |
| 楼宇间 / 跨域带宽增长 | 正向 | 当前至中期 | 相比很短的机架内链路,更自然地形成轻量相干用例 | 按距离层级索取设计目标:行内、园区、城域边缘 |
| CPO 和硅光成熟 | 正向 | 中期 | 可靠性一旦跑通,光学方案可向 AI 堆栈更深处扩展 | 询问 Celero 预计 CPO、NPO 或轻量相干配套何时具备商业意义 |
| 短距离里,铜缆、AEC 和 PAM4 仍更便宜或更简单 | 负向 | 当前 | 限制相干 DSP 今天能真正吃到多少广义光学繁荣 | 量化 Celero 在功耗、距离或系统总成本上何处胜过 PAM4 |
| 制造良率、可维护性和封装返工 | 负向 | 当前至中期 | 即便技术账面上有吸引力,也会拖慢光子方案大规模铺开 | 索取良率假设、故障维修流程和现场可更换模型 |
| 超大 GPU 集群可靠性风险 | 负向 | 当前 | 拉长认证周期,也抬高新光学供应商的举证门槛 | 索取 FIT/MTBF 数据、老化测试方法和超大规模云厂商认证状态 |
| 短距相干 DSP 没有公开 SAM / SOM 披露 | 负向 | 当前 | 估值仍高度受管理层对 ASP、搭载率和时点假设左右 | 按设计位点类别索取自下而上的销售管线、ASP 和搭载率模型 |
最强正向信号来自 AI 带宽增长和光学技术进步;最强负向因素来自时点、可靠性,以及缺少针对 Celero 细分领域的公开市场切片数据。
[CM014, CM015, CM021, CM022, CM025, CM033]03竞争对手
3.1 格局:直接同业、既有厂商、相邻玩家与替代方案
Celero 进入的不是空白品类。直接战场是已经卖进超大规模云和模块生态的第三方光 DSP 既有厂商,尤其是 Marvell 和 Broadcom,以及 Celestial AI、Ayar Labs、Lightmatter 等融资充足的光互连初创公司。格局还包括 Eliyan 等相邻 chiplet 互连厂商、Intel 和 Coherent 等既有硅光或组件供应商,以及 InnoLight 等现状模块厂商。合在一起,这些替代方案覆盖了买方在 AI 加速器、交换机和机架之间搬更多数据、同时不打爆功耗或时延预算的大多数解法。 替代集合同样重要。在最短距、耦合最紧密的纵向扩展领域,NVIDIA NVLink 和 UCIe 等 chiplet 标准可能在光链路进入考虑之前就吸走预算。对于成熟数据中心采购中心,既有运营模式仍是认证第三方 可插拔模块和收发器,而不是把风险押在一家新的相干 DSP 初创公司上。这意味着 Celero 不仅要赢在原始光学性能,还要对抗习惯:现有供应商、熟悉的认证流程,以及把光器件留在模块或封装边界、而不是把新的 DSP 厂商引入堆栈的架构。 实际结果是一片分层战场。Celero 的直接切入口是面向 AI 跨域扩展和光连接的相干 DSP,但真正的比较对象横跨第三方 DSP、模块供应商、共封装光学初创公司和电连接替代方案。买方可以多源采购模块、留在既有可插拔生态,或推进内部 chiplet 与纵向扩展织构。在 Celero 证明生产就绪和设计导入牵引力之前,这种广度会限制任何初创公司能期待的定价权或锁定能力。[CP001, CP004, CP009, CP014, CP017, CP021]
| 竞争对手 | 类别 | 规模 / 融资 | 目标细分市场 | 差异化 | 局限 |
|---|---|---|---|---|---|
| Celero Communications | 直接竞争初创公司 | 未上市;累计融资 $140M | AI 跨域扩展相干 DSP | 聚焦相干 DSP 定位,创始人有领域经验 | 未披露公开客户、价格或部署量 |
| Marvell | 直接现有厂商 | 上市半导体现有厂商;声称 1.6T 已大批量出货 | 超大规模云厂商、云、模块厂商、园区 / DCI | PAM4 + 相干 + 轻量相干覆盖广;装机基础;MACsec 选项 | 产品组合太宽,会让 Celero 在分销和路线图深度上显得偏窄 |
| Broadcom | 直接现有厂商 | 上市半导体现有厂商;声称组件已批量出货,早期访问 DSP 已开始送样 | AI 集群、收发器伙伴、交换机生态 | 光学栈覆盖广,伙伴生态强,DSP 路线图强调功耗 | 相干并非唯一重点,公开标价仍不透明 |
| Celestial AI | 直接 / 相邻初创公司 | 未上市;累计融资 >$515M | 从封装到多机架的光学纵向扩展网络 | Photonic Fabric 覆盖连接、交换和封装 | 平台范围更宽,可能拉长执行和认证负担 |
| Ayar Labs | 直接 / 相邻初创公司 | 未上市;累计融资约 $870M;报道称估值 $3.75B | 面向 GPU、交换机和机架级 CPO 的光学 I/O | TeraPHY + SuperNova 栈;标准和可靠性信号;融资雄厚 | 架构更偏向光学 I/O 和 CPO,而非商用相干 DSP |
| Lightmatter | 直接 / 相邻初创公司 | 未上市;检索来源未披露 2026 年融资 | 面向可插拔、NPO 和 CPO AI 链路的光子互连 | 带宽密度路线图很高,并绑定 NVIDIA 生态 | 商业化仍以早期访问伙伴表述 |
| Intel Silicon Photonics | 潜在进入者 / 相邻现有厂商 | 上市现有厂商;已出货 >8M PIC 和 >32M 激光器 | 共同封装 OCI 和可插拔数据中心光学 | 制造可信度很强,可靠性经过现场验证 | 产品叙事覆盖广义硅光,不是 Celero 式聚焦初创公司叙事 |
| Coherent / InnoLight | 现状模块供应商 | 上市 / 大型未上市模块生态 | 收发器、光学器件和商用模块部署 | 垂直整合,模块采购路径熟悉 | 通常依赖上游 DSP 选择,而不是掌握完整商用 DSP 叙事 |
综合上市公司规模、已披露初创公司融资和明确标记的未知项;各行比较的是买方面前的替代方案,而不是严格的类别纯度。
[CP002, CP006, CP012, CP014, CP017, CP019]基于公开证据的顺序视角,展示 AI 光互连中技术成熟度与分销触达之间的位置。
坐标轴使用证据支撑的 1-5 顺序尺度,因为多数供应商的准确市场份额或部署数量不公开。X 轴反映公开就绪度、出货量和标准成熟度;Y 轴反映生态触达、合作伙伴杠杆和公开部署渠道。
[CP001, CP006, CP009, CP014, CP017, CP021]3.2 直接竞争:既有厂商已覆盖全栈,初创公司则推进 CPO 与光子织构
Marvell 和 Broadcom 是最难的标杆,因为两家公司都能在客户的出货爬坡已经发生的地方对接需求。Marvell 现在面向短距、园区和更长距离的数据中心链路销售 PAM4、相干和轻量相干 DSP,并称其 3nm Ara 平台已经向全球超大规模云大批量出货。Broadcom 的产品组合同样宽:光 DSP、激光器、SerDes、CPO 和 PCIe-over-optics,公司还声称相邻光组件已经有出货规模,新 1.6T 平台也进入早期送样。换句话说,Celero 竞争的不是两个芯片 SKU,而是能够打包路线图可信度、标准参与和模块伙伴关系的既有厂商。 初创公司从另一个角度进攻。Celestial AI 讲的是从封装到机架的更广义 Photonic Fabric 故事。Ayar Labs 强调可在机架内外替代铜缆的光 I/O chiplet 和远程光源。Lightmatter 推进 3D 光子互连路线图、更大的参考系统,以及 NVIDIA NVLink Fusion 等生态关系。这些公司并非 Celero 的完全同类替代品,但都在争夺同一条战略预算线:用更靠近计算的更多光学内容,解决 AI 互连瓶颈。如果客户迁向共封装或近封装光学,Celero 的相干 DSP 切入口可能收窄,除非其架构也适配这些部署模式。 Celero 最强的相对差异化是聚焦。它公开讲述的是面向跨域扩展光连接的相干 DSP,而不是宽产品组合。如果客户想要专业化引擎,而不是平台型供应商,这会有帮助。风险在于,既有厂商和相邻初创公司已经展示出更宽的产品地图、更深的送样管线,以及比 Celero 迄今公开披露更可见的生态锚点。[CP005, CP006, CP007, CP009, CP010, CP011]
| 购买标准 | Celero | Marvell | Broadcom | Celestial AI | Ayar Labs | Lightmatter |
|---|---|---|---|---|---|---|
| 公开披露相干 DSP | 高 | 高 | 低 / 非相干主导 | 低 | 低 | 低 |
| 1.6T 时代公开产品节奏 | 中 | 高 | 高 | 中 | 中 | 高 |
| 封装到机架架构覆盖广度 | 中 | 中 | 高 | 高 | 高 | 高 |
| 公开出货 / 放量证明 | 低 | 高 | 高 | 低 | 中 | 低 |
| 标准 / 安全信任信号 | 低 | 高 | 高 | 中 | 高 | 中 |
| 已披露超大规模云厂商 / 生态杠杆 | 低 | 高 | 高 | 中 | 中 | 高 |
高 / 中 / 低标签是基于公开披露的证据判断;证据不足或未披露的领域有意下调,而不是凭空上调。
[CP001, CP006, CP008, CP009, CP011, CP014]竞品类型的能力覆盖显示,Celero 更聚焦,而对手越来越多地打包相邻光学功能。
[CP001, CP005, CP009, CP015, CP017, CP020]3.3 切换成本、分销力量与商业不透明度都有利于既有厂商
这个市场里的竞争耐久性不太取决于品牌,更取决于供应商处在认证和封装流程的哪个位置。Broadcom 和 Marvell 已经熟悉早期接入客户、伙伴和出货爬坡的语言。Broadcom 明确引导客户联系直销获取定价和样品;Futurum 则指出,收发器认证周期长,早期设计导入一旦拿下就很黏。Marvell 同样强调大批量出货和宽广装机基础。这些信号重要,因为一旦 DSP、光引擎或共封装设计被验证进模块或机架架构,买方切换供应商就要面对重新测试成本、固件和 FEC 调优工作,以及进度风险。 初创公司的位置不一。Ayar 和 Lightmatter 都在谈制造就绪、伙伴参与和早期接入计划,但公开证据仍显示其商业化路径 以设计伙伴关系为中心,而不是广泛目录销售。Celestial AI 声称与超大规模云和封装生态深度合作;如果这些项目落地,可能转化为集中但有力的渠道。Coherent 和 InnoLight 则提示 Celero 面临另一道难题:现状往往让买方继续使用熟悉的模块厂商,而 DSP 可由模块厂商在底层采购,上游新硅片品牌的可见度会被削弱。在这种世界里,多归属存在,但很大一部分权力掌握在模块厂商、超大规模云和平台所有者手中,而不是一家独立 DSP 初创公司手中。 最值得注意的未知项是商业透明度。审阅来源中没有任何一家公布 DSP 芯片的真实目录价。客户名称、实际 ASP 和量产承诺大多仍是私有信息。这使得证明 Celero 近期具备定价杠杆变得困难,也提高了商业权力流向那些已经控制认证窗口、伙伴生态和制造产能者的概率。[CP006, CP012, CP013, CP016, CP018, CP019]
| 供应商 | 价格 / 单位 / 合同信号 | 包含能力 | 折扣 / 未知项 | 影响 |
|---|---|---|---|---|
| Celero | 公开价格未披露;可能是直接设计定点的硅芯片销售 | 面向 AI 光链路的相干 DSP | 没有公开 ASP、具名客户或销量承诺 | 难以仅凭公开证据证明定价权或爬坡时点 |
| Marvell | 公开价格未披露;通过超大规模云厂商和模块生态销售 | PAM4、相干、轻量相干、遥测、模拟组件 | 没有公开标价;宽产品组合可能支撑打包经济性 | 因为掌握栈内相邻环节,竞争不只靠芯片价格 |
| Broadcom | 没有公开标价;来源要求买方联系销售代表获取样品和价格 | DSP、激光器、SerDes、CPO、光学承载 PCIe、铜链路 | 实际成交价、折扣和产品组合未披露 | 商业杠杆可能来自生态拉力和认证提前量 |
| Ayar Labs | 没有公开标价;方案看起来由伙伴驱动并集成进封装 | TeraPHY 小芯片加 SuperNova 光源 | 面向具体客户的集成经济性未披露 | 赢单取决于系统价值证明,而不是目录价 |
| Lightmatter | 没有公开标价;强调早期访问伙伴模式 | Passage 光子引擎、参考系统、Guide 光引擎 | 通用可用性和量产价格未披露 | 暗示大规模商用采购前会有很长评估周期 |
该市场公开信息里商业条款不透明。表格比较定价可见度和打包动作,而不是来源未披露的保密 ASP。
[CP013, CP018, CP022, CP024, CP043, CP044]3.4 护城河耐久性受路线图拥挤、CPO 迁移和内部自研风险约束
Celero 可能主张的护城河是技术聚焦、相干 DSP 专长和创始人履历。这些都真实存在,但耐久性好坏参半。创始人与 Marvell、Broadcom、Inphi 和 ClariPhy 的重叠给了 Celero 可信的执行基因,却也说明核心人才和知识并非公司独有。同时,2026 年市场证据指向的是路线图拥挤,而不是空白地带。LightCounting 已经把 Celero 与更大的既有厂商和其他新兴供应商放在一起跟踪;Futurum 描述的竞争是,1.6T 阶段的供应商选择可能把客户锁进 3.2T 路径;SemiEngineering 则认为,随着可靠性证据积累,CPO 可能压缩可插拔 DSP 架构的窗口。 最严重的替代风险不是纯竞争,而是架构性变化。如果 Broadcom、NVIDIA 或超大规模云把更多互连堆栈拉回内部,第三方 DSP 厂商的可服务池会变小。如果客户从可插拔跳向共封装或近封装光学的速度快于 Celero 预期,面向跨域扩展链路的相干 DSP 小众市场未必能拿到最高价值插槽。即便整体市场增长,对初创公司也可能是坏事,因为增长会吸引更多资本、更多既有厂商和更多基于标准的替代品。 这并不意味着 Celero 没有路径。如果轻量相干或跨域扩展光链路成为既有厂商服务不足的顽固瓶颈,或者客户想要一个聚焦的第三方替代方案,而不是捆绑平台,Celero 仍可能胜出。但今天的公开证据支持谨慎判断:Celero 的差异化存在,可耐久性要等公司拿出客户、制造伙伴,以及其方案能扛住更广义光平台和 AI 互连内部化迁移的理由后才能证明。[CP002, CP008, CP020, CP027, CP028, CP034]
| 护城河主张 | 威胁 | 严重性 | 缓解 / 尽调问题 |
|---|---|---|---|
| 聚焦相干 DSP 专业化 | Marvell 和 Broadcom 已覆盖相干或相邻 1.6T 栈,规模更可见 | 高 | 索取具名设计定点、量产时间线,并要求说明聚焦相干 DSP 为什么能打赢打包平台 |
| 创始人履历和领域经验 | 人才与 Marvell、Broadcom、Inphi 重叠,说明知识有价值但并不独占 | 中 | 询问 Celero 如何在团队履历之外守住架构、招聘和路线图优势 |
| 跨域光互连效率 | CPO 和光学 I/O 初创公司可能把光学更贴近计算,压缩商用 DSP 切入空间 | 高 | 询问 Celero 架构是否能延伸到 CPO、NPO 或封装级部署 |
| 相对平台厂商的中立商用替代方案 | 模块厂商和超大规模云厂商可通过多供和自研选项保留议价权 | 高 | 询问 Celero 目标设计位点中谁控制认证、固件和模块 BOM 决策 |
| 更快切入 AI 光学增长 | 现有厂商送样领先,可能锁定从 1.6T 到 3.2T 的多代设计路径 | 高 | 索取 Celero 相对 Broadcom 和 Marvell 认证窗口的精确时点 |
| 增长中的市场顺风 | 增长会吸引 NVIDIA 内部芯片、Intel 或相邻进入者,以及更多标准化替代方案 | 中 | 询问即便超大规模云厂商内化更多互连芯片,市场中哪一部分仍会保持商用外采 |
严重性反映 Celero 未来议价地位的下行风险,而不是整个光学市场的风险。多项缓解问题需要非公开尽调。
[CP006, CP012, CP017, CP022, CP024, CP037]浓缩列出 2026 年最影响 Celero 竞争耐久性的指标。
[CP006, CP013, CP030, CP034, CP038, CP039]3.5 展示材料
04财务
4.1 收入模式、定价与确认约束
Celero 是无晶圆厂半导体初创公司,因此经济引擎是硬件收入,而不是经常性软件订阅。公开材料一贯把产品描述为相干 DSP 芯片 / 平台,让超大规模云和云运营商以更低功耗、更高带宽效率在光纤上移动更多数据;这意味着变现来自芯片销售、参考平台设计导入,以及稳定量产前可能出现的工程样片或 NRE 款项。同样重要的是未公开部分:Celero 没有披露任何目录价、合同 ASP、出货单元数、已签未交付订单、具名客户, 或按产品阶段划分的收入贡献。行业基准显示,相干 DSP 是高端产品,由于复杂度更高、出货量更低,其 ASP 可维持在 PAM4 DSP 定价约 10x 的水平;但这些基准也显示,未来 1.6T 采用和新进入者可能压缩价格。承销时,核心收入确认问题是时间:投资人需要区分量产前 NRE 或认证收入与可重复量产出货,因为在芯片公司出现持久毛利率之前,两者都可能先出现在账上。[CI001, CI002, CI005, CI006, CI007, CI008]
| 收入来源 | 机制 | 单位 | 当前价值 / 状态 | 质量 | 尽调问题 |
|---|---|---|---|---|---|
| 量产相干 DSP 销售 | 通过认证并放量后,按芯片确认商用硅收入 | 芯片 / 设计位点 | 未公开披露;没有公开出货或在手订单 | 核心未来收入来源,但公开证据尚未证明 | 提供已签订单、按速率等级划分的 ASP,以及按客户群组划分的量产时点 |
| 工程样品 / 评估批次 | 发送预量产硅芯片,用于认证和互操作工作 | 样品批次 | 可能存在但未披露;商业化表述暗示仍处于放量前阶段 | 可验证设计定点,但本身不是耐久收入 | 披露样品状态、客户实验室,以及从评估到放量的转化 |
| NRE / 里程碑回款 | 与里程碑绑定的定制设计、调试拉起或认证付款 | 项目里程碑 | Celero 没有公开披露;可比 ASIC 模式使用里程碑 NRE | 若存在,是有用现金流入,但质量低于重复出货 | 拆分 NRE 收入、里程碑排期和毛利处理 |
| 参考平台 / 生态赋能 | 围绕相干引擎提供模块 / OEM 集成支持并获得收入 | 项目 / 合同 | 没有公开披露 | 可能具备战略意义,但经济性不清楚 | 给出模块或 OEM 伙伴名称及相关商业条款 |
| 未来软件 / 固件支持 | 围绕硅芯片提供可能的赋能、诊断或集成支持 | 支持合同 | 没有公开证据 | 只能作为增量,不应纳入基准测算 | 澄清是否存在任何经常性支持收入 |
Celero 未披露收入结构、出货或 ASP;因此各行把可能机制与已验证公开披露分开,并标出证据缺口。
[CI006, CI007, CI008, CI035]| 价格 / 合同概念 | 标价与成交价 | 折扣 / 未知项 | 来源 / 基准 |
|---|---|---|---|
| Celero 芯片 ASP | 未公开标价或实际 ASP | 按速率档位、出货层级或客户拆分均未知 | Celero 官方材料披露了产品和融资,但未披露价格 (CI006) |
| 相干 DSP 相对 PAM4 的市场溢价 | 仅为基准;相干 DSP ASP 预计仍约为 PAM4 ASP 的 10x | Celero 自身溢价未知;出货量上来后可能收窄 | LightCounting,经 Optical Connections 转引 (CI010) |
| NRE / 里程碑付款 | 可比 ASIC 合同常按里程碑回收 NRE | Celero 未确认是否单独收取 NRE | Marvell 10-K 基准 (CI009) |
| 未来 1.6T 定价曲线 | Celero 公开数据里尚不可观察;最高速率机会仍在形成 | 1.6T 相干以太网的出货量和定价仍不确定 | Optical Connections 与 Cignal 基准 (CI013, CI043) |
经济性空白占主导,因为 Celero 没有发布价格表、design-win 合同或实际 ASP 披露;这些基准行只是行业代理指标,不是公司事实。
[CI006, CI009, CI010, CI043]Celero 必须先把超大规模云厂商的认证活动转成高价相干 DSP 出货,品类 ASP 优势才会变成可持续毛利。
Celero 尚未披露自己处在这座桥的哪一段;这条流向结合了公司表述,以及同类无晶圆厂 ASIC 企业常见的硬件收入机制。
[CI003, CI007, CI008, CI009, CI021]4.2 GTM 动作、销售效率代理指标与公开牵引缺口
最可能的商业化动作是经典第三方硅片设计导入模式:Celero 必须说服少数超大规模云、云运营商、光模块厂商或系统 OEM 认证一套新的相干 DSP 架构,再把这些设计导入转化为生产爬坡。公开来源强调面向超大规模和云客户的商业化,而不是宽渠道软件分发,也强化了这一假设。因此,ARR、CAC 回收期或 NRR 等传统 SaaS 效率指标不是今天合适的公开视角;更相关的问题是认证周期长度、样片到设计导入转化率、NRE 回收、按速度等级划分的毛利率,以及客户集中度。公开牵引几乎完全不透明。没有披露营收、ARR、签约订单、活跃客户数、出货量、未交付订单 或利用率指标,也没有具名的超大规模云量产部署。指标缺失很重要,因为第三方网络芯片即便规模化,也可能高度集中在少数买方。成熟可比公司 Marvell 披露,2026 财年前十大客户贡献 82% 收入,这意味着更早期的相干 DSP 初创公司可能更暴露于少数账户。尽调结论是,Celero 可能具备战略 GTM 动能,但公开证据尚未证明可重复商业牵引。[CI003, CI006, CI017, CI018, CI024, CI032]
| 缺失的私有指标 | 对投资判断的影响 | 精确尽调路径 |
|---|---|---|
| 分阶段收入(样片、NRE、量产) | 无法区分高质量持续出货和一次性里程碑收入 | 按客户索取从样片 / NRE 到量产出货的月度收入桥 |
| 按速率档位的 ASP 和毛利率 | 无法检验相干产品溢价能否抵消 foundry 和 OSAT 成本 | 获取 SKU 级定价、已成本化 BOM、良率假设和毛利率瀑布图 |
| 客户集中度和 backlog | 无法判断对一两个超大规模客户的依赖 | 审阅 backlog、按设计阶段拆分的 pipeline 和头部客户敞口 |
| 现金 burn 和 runway | 无法评估资本充足性或融资依赖 | 索取月度 cash burn、已承诺支出和下行情景 runway 模型 |
| Foundry / 封装预留和重新流片预算 | 无法评估隐性营运资金和执行风险 | 审阅 foundry 协议、基板承诺和重新流片应急预留 |
这些是优先级最高的尽调缺口,因为它们决定 Celero 到底是早期收入、pre-revenue,还是在大额标题融资后仍受资本约束。
[CI006, CI035, CI037, CI046]4.3 成本结构、毛利率路径与单位经济基准
Celero 的成本结构应按研发重、依赖代工来建模。关于无晶圆厂流片和上市网络芯片公司的基准来源显示,营收放大前现金会流向哪里:设计团队薪酬、EDA 和 IP 许可、掩膜和晶圆支出、外包封装测试,以及在明确客户需求之前预留的库存或产能。对网络或光 ASIC 项目而言,7nm-28nm 总 NRE 约为 $5M-$30M,领先的 3nm-5nm 项目可达 $30M-$100M+,晶圆成本约 $16k-$22k,封装 / 测试还会继续增加单元成本。这些区间帮助界定 Celero 的挑战:即使相干 DSP ASP 在结构上有溢价,早期毛利率仍容易受低出货量制造、重流片和成本吸收缓慢影响。Marvell 的申报文件是有用基准,因为它把毛利率表现与规模、产品组合、库存纪律、代工可得性和外包封测经济性联系起来;文件还指出,ASIC 式合作可能带来较低毛利率和按里程碑回收的 NRE。Celero 没有公开毛利率、良率、每比特功耗或可变成本披露,因此毛利率承销只能保持暂定。最站得住脚的公开结论是定性的:这个品类规模化后可以很有吸引力,但从样片走向成熟毛利率的路径资本密集、执行敏感。[CI009, CI010, CI014, CI015, CI016, CI017]
| 指标 | 数值 / 状态 | 置信度 | 为何重要 | 尽调要求 |
|---|---|---|---|---|
| 7nm-28nm 项目 NRE 基准 | $5M-$30M 总 NRE | 中 | 框住主流网络 / 光芯片项目在规模收入前的成本区间 | 确认 Celero 的工艺节点、掩膜预算、IP 授权成本和重新流片预留 |
| 3nm-5nm 项目 NRE 基准 | $30M-$100M+ 总 NRE | 中 | 如果 Celero 走先进制程,资金需求会多快抬升 | 提供实际节点选择、代工伙伴和预计 full-mask 时间 |
| 先进制程晶圆成本基准 | 每片晶圆 $16k-$22k | 中 | 变动成本和现金预承诺会主导早期毛利测算 | 披露晶圆定价假设、良率目标和基板 / 封装 BOM |
| 相干 DSP ASP 相对 PAM4 的溢价 | 基准 ASP 约高 10x | 中 | 只要 Celero 产品过验证,溢价 ASP 可能支撑毛利上行 | 按速率档位展示 Celero 目标 ASP、每比特功耗和毛利率 |
| Celero 毛利率 / 良率 / 产能利用率 | 公开未披露 | 低 | 缺少这些指标,投资人无法核算贡献毛利或盈亏平衡点 | 开放毛利率桥、良率曲线和产能利用率假设的数据室 |
只有基准行有来源支持的数字;Celero 自身单位经济性仍未披露,是投资判断的核心卡点。
[CI010, CI015, CI016, CI017, CI046]公开基准显示,Celero 的经济性主要由制程节点、流片 / NRE、外包制造,以及高价相干 ASP 能否顶住认证和价格压力决定。
Celero 的毛利率、良率和产能利用率没有公开;因此这张桥接图使用基准节点,而非公司披露的数字输入。
[CI010, CI015, CI016, CI021, CI022, CI039]用于判断 Celero 资本强度的美元计价市场和成本基准;只有累计融资额是公司披露的专属数字。
所有行都使用百万美元,以满足渲染器契约。市场行把低 / 中 / 高重复为同一个点估计,因为来源只发布单年市场规模,而非区间。
[CI001, CI012, CI015]4.4 资本充足性、融资依赖与承销判断
从公开信息看,Celero 的融资故事强于运营披露。公司披露累计股权融资 $140M,并表示新资金将用于扩大面向超大规模和云客户的硅片开发及商业化。这足以支持下一阶段开发,但信息仍不足以让人有把握地承销流动性。没有公开现金余额、月度现金消耗、明确现金可支撑月数、债务或项目融资义务,也看不到晶圆预付款、基板预留,或是否已经为一次或多次先进节点 重流片编列预算。基准证据同时指向两件事:AI 互连初创公司只要解决明显瓶颈,仍能融到大额轮次;先进光学供应链仍暴露在激光器短缺、封装瓶颈和部署延迟之下,可能拉长收入到来的时间。实际看,Celero 下一轮融资触发点很可能不是另一个抽象技术里程碑,而是客户认证证据和可信量产时间。在管理层披露样片状态、收入结构、毛利率桥接和现金消耗计算之前,财务判断应保持谨慎:收入质量仍未证明,毛利率路径可能有吸引力但尚未展示,资本密集度高到足以让一次认证窗口错失在持久规模经济显现前迫使公司再融资。[CI001, CI003, CI026, CI027, CI028, CI029]
| 项目 | 当前数值 / 状态 | 含义 | 尽调要求 |
|---|---|---|---|
| 账面现金 | 2025 年 11 月融资后未公开披露 | 没有交割后余额和营运资金承诺,就无法估算真实 runway | 提供 Series B 交割后的月末现金和受限现金 |
| 月度烧钱 | 未披露;公开来源只能支持基准推断 | burn 敏感性比标题融资额更能决定下一轮时间 | 按季度提供月度 opex、capex、晶圆预付款和现金 burn |
| Runway 月数 | 未披露;仅靠公开证据无法支撑投资判断 | runway 取决于节点选择、重新流片和客户验证节奏 | 提供基准 / 下行情景 runway 和契约条款假设 |
| 计划资金用途 | 扩展面向超大规模 / 云客户的芯片开发和商业化 | 能支撑产品推进,但不足以证明单位经济性 | 拆解 R&D、tape-out、封装 / 测试和 GTM 的资金分配 |
| 下一轮触发点 | 最可能是客户验证通过,并拿出可信量产时间表 | 没有 design-win 证明,下一轮融资可能是在资助延期,而不是扩张 | 披露从样片到量产的里程碑计划 |
| 债务 / 项目融资义务 | 没有公开证据显示存在债务或项目融资 | 资本结构看似由股权融资支撑,但负债仍可能藏在表外 | 提供债务明细、设备租赁和 foundry 产能预留义务 |
唯一硬公开融资事实是总计 $140M 股权融资;runway 测算所需的其他充足性输入要么缺失,要么需要私下披露。
[CI001, CI003, CI036, CI037, CI042]Celero 的主要现金流压力发生在收入可见之前:流片、外包制造、库存承诺和认证延迟。
这是定性现金流图谱,因为 Celero 没有披露现金余额、月度烧钱额、资本开支或债务条款。
[CI019, CI020, CI021, CI027, CI028, CI037]4.5 展示材料
05产品与技术
5.1 产品定义与模块地图
Celero 在公开网站上卖的不是成品交换机、路由器或收发器模块;它的产品更适合描述为一套相干 DSP 半导体平台,嵌在连接 AI 加速器、服务器和数据中心集群的光互连硬件内部。从客户工作流看,买方面临的问题很直接:超大规模云和基础设施厂商需要把大流量传得比短距直接探测光学更远,同时守住功耗效率和带宽密度。Celero 的公开材料反复把公司定位为让这些光链路变得可用的信号处理层。官网强调加速器和集群之间的无缝光链路,CapitalG 投资笔记则把 DSP 解释为光收发器内部的芯片,把来自 GPU 或 XPU 的电信号转成光进行传输。这个定位重要,因为它把 Celero 放在计算网络织构 与光线路系统之间的控制点上;功耗、距离、调制和互操作性决策都会直接影响 AI 集群架构。[CE001, CE002, CE003, CE004, CE006, CE007]
| 模块 / 资产 | 主要用户 | 状态 / 成熟度 | 差异化 | 尽调缺口 |
|---|---|---|---|---|
| 相干 DSP ASIC 核心 | 光模块 OEM 或超大规模光网络团队 | 开发中 / 未公开披露 SKU | 专为 AI 规模相干链路而做,不是泛用电信堆叠 | 需要命名芯片家族、工艺节点、lane map 和功耗预算 |
| DSP + FEC 信号处理子系统 | 优化传输距离和带宽密度的系统架构师 | 从市场背景和岗位组合看,技术可信度较高 | 面向相干调制、均衡和纠错,切中每瓦带宽关键场景 | 需要公开 BER/FEC 阈值、支持的调制模式和延迟数据 |
| 高速 host I/O 和时钟模块 | 交换机、路由器和模块集成团队 | 从 SerDes、PLL、RTL 和验证招聘可推断 | 表明 Celero 在打完整电到光集成,不只是算法 IP | 需要 host 接口标准、lane 速率和 retimer 假设 |
| 固件 / 管理层 | 制造、验证和现场支持团队 | 可推断但未披露 | 可能需要用于 C-CMIS/CMIS 可视性、诊断和模块控制 | 需要遥测、升级和故障管理模型的公开文档 |
| 封装、验证和产品工程栈 | 运营和产品化团队 | 早期商业阶段 | 如果能缩短 bring-up 周期,foundry 和系统集成深度可形成护城河 | 需要封装策略、foundry/OSAT 披露、良率目标和验证里程碑 |
各行反映的是公开可见的产品层和招聘线索,而不是供应商发布的数据表层级。
[CE001, CE008, CE009, CE010, CE015, CE028]| 用户任务 | 当前工作流 | Celero 方案 | 可衡量收益 | 限制 |
|---|---|---|---|---|
| 在机架和楼宇之间扩展 AI 流量 | 在距离或功耗卡住前,先用短距电连接或直接探测光模块 | 在模块设计中加入由相干 DSP 驱动的光链路,支撑更长、更密的互连 | 公开主张是更高带宽效率和更远距离 scale-out | Celero 没有公开基准表量化增益 |
| 用光纤连接数据中心或园区 fabric | 依赖重传输型相干方案或碎片化光栈 | 使用为数据中心定制的相干 DSP 架构 | 承诺是以更低的数据中心功耗 / 成本,拿到电信级距离 | Celero 未公开传输距离或损耗预算规格 |
| 在 host 平台上验证可插拔相干模块 | 经过漫长验证周期,把光引擎、DSP 固件和 host 管理接起来 | 用 Celero silicon 加管理 / 诊断栈作为相干控制点 | 如果 host、模块和光学一起优化,系统瓶颈可能下降 | 互操作和 host 验证复杂度仍高 |
| 通向 800G 和 1.6T 相干 fabric 的路线图 | 从今天的 800ZR 推进到 1600ZR/CL 和 AI 专用 compute optics | 让 Celero silicon 对准下一波标准,而不是只服务当前链路 | AI fabric 推向更高密度和更远距离时,产品仍能贴住需求 | Celero 具体路线图里程碑的公开证据仍薄 |
| 为超大规模客户搭建制造和现场支持 | 协调设计中心、验证、产品工程和供应链 | 借经验丰富的相干 DSP 团队和 foundry 关系缩短 bring-up | 潜在护城河来自执行,不只是算法 | 未公开客户验证、RMA 条款或支持 SLA |
收益基于公开定位和行业基线,不是经审计的 Celero 现场结果。
[CE003, CE004, CE005, CE015, CE018, CE019]Celero 公开讲述的产品路径,把 AI 工作负载、相干 DSP 处理、模块管理和光传输串成一条互连链路。
Celero 没有公开区块图,因此这套栈由官方定位、招聘信息和标准语境重建。
[CE003, CE006, CE008, CE009, CE010, CE020]客户流程从 AI 集群带宽需求出发,走到相干模块认证和部署。
[CE003, CE004, CE005, CE016, CE018, CE024]5.2 架构、集成与部署模式
Celero 没有公开标准版框图,因此最佳架构重建来自招聘信号和外部标准工作。公开岗位覆盖 ASIC 设计、设计验证、RTL、DSP-FEC、光 DSP、SerDes、PLL、模拟版图、物理验证、物理设计、固件和产品工程。这种岗位组合强烈指向完整定制相干 DSP 项目,而不是狭窄的算法 IP 店。可能的运行栈从主机侧电 I/O 和时钟开始,经过 DSP 与前向纠错引擎,再进入相干光模块管理固件和封装级实现。OIF 技术工作提供了这套堆栈周围的标准语境:用于相干模块管理的 C-CMIS、CMIS 安全扩展、224G 和 448G 芯片到模块接口,以及 1600ZR、1600ZR+、1600CL 和面向 AI 纵向扩展链路的计算光学接口等路线图工作,协议包括 PCIe、NVLink 和 UALink。因此,部署很可能意味着 Celero 芯片进入模块厂商、交换机或路由器平台以及超大规模云光学团队的认证循环,而不是作为独立终端产品销售。[CE008, CE009, CE010, CE018, CE019, CE020]
| 层级 / 流程 / 组件 | 作用 | 依赖 | 风险 |
|---|---|---|---|
| Host 电 I/O(SerDes / PLL / 时钟) | 在 switch/GPU host 芯片和相干模块路径之间搬运流量 | 先进高速电气标准和封装完整性 | 信号完整性或功耗短板会抹掉光学优势 |
| 核心 DSP pipeline | 跑调制、均衡、载波恢复和其他相干处理步骤 | 算法设计、ASIC 实现和固件监测埋点 | 没有 Celero 公开数据表可验证吞吐、延迟或功耗 |
| FEC 引擎 | 提升高速相干流量的链路裕量和误码表现 | O-FEC 等标准对齐,以及模块互操作预期 | FEC 开销和延迟取舍会限制部署适配度 |
| 模块管理 / 诊断 | 通过 CMIS/C-CMIS 式接口暴露配置、遥测和控制 | 固件质量和标准合规 | 管理模型不透明,会拖慢客户验证和现场调试 |
| 光引擎和可插拔集成 | 把 DSP silicon 与调制器、激光器、接收器和封装选择配对 | 模块 OEM、光引擎供应商和热设计 | 互操作或热问题后期暴露,会延误产品爬坡 |
| 制造和产品工程 | 把设计从 tapeout 推到验证和客户出货 | foundry 准入、封装产能和良率学习 | 供应瓶颈或低良率会主导上市时间 |
该运营模型根据职业经历证据和公开相干标准材料重构,因为 Celero 没有发布公开架构图或模块数据表。
[CE008, CE009, CE010, CE019, CE020, CE021]Celero 的产品路径依赖标准组织、代工 / 封装执行、光模块伙伴和客户认证循环。
[CE019, CE020, CE021, CE022, CE024, CE027]5.3 成熟度、路线图与差异化
公开记录指向一家处于高级建设期、而非广泛商业成熟期的公司。Celero 2025 年 11 月融资公告称新资金将加速开发和部署,多份外部摘要也把商业化描述为下一步,而不是已经规模化的装机基础。这与官网证据相符:相干 DSP 叙事很强,但没有公开数据手册库、具名 SKU 组合或基准测试表。差异化论点仍然可信。创始人履历和投资人评论持续强调其此前在 Marvell、Inphi、Broadcom 和 ClariPhy 的相干 DSP 领导经验,投资人笔记还特别提到强代工关系和系统级集成。这些能力正是该产品类别所需要的:封装、光引擎适配、固件管理和互操作性可能和原始算法一样决定成败。外部市场和标准信号也显示 Celero 下一步可能想在哪里落子:商用 800G 相干可插拔模块今天已经存在,OIF 也已把路线图延伸到 1.6T 相干模块和面向 AI 优化的计算光学。因此,Celero 的产品成熟度应被看作技术上严肃,但公开披露仍不足。[CE011, CE012, CE013, CE014, CE015, CE018]
| 日期 / 阶段 | 功能 / 里程碑 | 状态 | 含义 | 来源 |
|---|---|---|---|---|
| 2024 年成立 | Celero 围绕来自 Marvell/Inphi/Broadcom/ClariPhy 的相干 DSP 经验组建 | 已完成的历史步骤 | 团队围绕已知光芯片问题组建,而不是从白纸创业题开始 | SE002 / SE013 |
| 2025-11 融资公告 | 公司融资 $140M,并称资金将加速开发和部署 | 已公告 | 信号指向产品化和 go-to-market 搭建,还不是广泛公开 GA | SE011 / SE016 |
| 2025 年 800ZR 市场基线 | OIF plugfest 和商用 800G 相干可插拔产品给出当前生态门槛 | 当前外部基线 | Celero 必须进入真实多供应商验证环境,而不是理论市场 | SE021 / SE022 |
| 2026 年标准路线图 | OIF 和 OFC 材料把路线图延伸到 1600ZR、1600ZR+、1600CL、COI 和 MACsec | 进行中 | 路线图压力正推向 1.6T 密度、更低功耗和 AI 专用光学 | SE020 / SE023 |
| 当前公开产品露出 | 网站和博客 / 新闻在 SKU 与性能材料上仍很薄 | 当前缺口 | 投资人可以判断 thesis,但客户尽调仍需要直接产品文件 | SE001 / SE006 / SE010 / SE024 |
路线图行混合了公司里程碑和外部生态里程碑,因为 Celero 没有发布详细公开发布日历。
[CE007, CE015, CE019, CE021, CE022, CE023]Celero 的公开成熟度,团队可信度最强,产品资料披露最弱。
[CE012, CE015, CE019, CE027, CE031, CE032]5.4 信任、质量控制与技术风险
Celero 公开的信任界面更偏法律和流程,而不是认证。隐私政策披露联系人和招聘数据收集、Cookie、反欺诈措辞,以及遵守法律或监管要求,并称公司使用物理和电子保护措施。条款页面管理网站使用,Cookie 政策确认跟踪披露。缺失项至少同样重要:留存的公开来源没有显示信任中心、已发布 SOC 2 或 ISO 27001 报告、可用性 历史、产品安全说明、RMA 政策,或面向硅片客户的正式支持 SLA。产品侧,OIF 材料说明了为什么这个缺口重要:相干可插拔模块仍面临真实的互操作和验证复杂性,800ZR 互通测试会记录了一些组合需要进一步调试或未能建立链路。由此得到的风险图很清楚。Celero 可能拥有合适团队和市场时点,但产品尽调仍需要直接证据:可靠性指标、模块互操作数据、制造和封装依赖、客户认证,以及超出网站政策的安全或质量治理。[CE019, CE023, CE024, CE031, CE032, CE033]
| 控制项 / 认证 / 质量指标 | 状态 | 范围 | 缺口 |
|---|---|---|---|
| 隐私政策和数据处理披露 | 当前公开证明 | 网站联系表单、招聘收集、cookies、反欺诈和法律合规语言 | 不能转化为 silicon 客户安全保证或产品安全深度 |
| 条款与条件 | 当前公开证明 | 网站 / 服务使用规则和法律边界 | 未公开硬件支持 SLA、保修或 RMA 工作流 |
| 生态中 CMIS / C-CMIS / CMIS-Security 对齐 | 外部标准仍在推进 | 管理、诊断、相干模块控制,以及演进中的模块安全实践 | Celero 未公开说明产品支持哪些标准或寄存器 |
| 互操作测试纪律 | 800ZR 生态层面已有行业证明 | 面向可插拔相干模块的 plugfest 式验证 | Celero 未公开互操作报告、BER 图或验证矩阵 |
| 运营透明度 | 薄 | 联系页面、新闻帖子和招聘足迹可见 | 未公开信任中心、SOC 2、ISO 27001、uptime 历史或客户支持指标 |
公开信任证据主要来自政策和标准周边材料,而不是产品认证 / 证明。
[CE023, CE024, CE033, CE034, CE035, CE036]5.5 展示材料
06客户
6.1 客户细分、买方与付款方
Celero 的公开材料显示,客户宇宙狭窄但价值很高。公司把产品描述为面向高带宽、低功耗光链路的相干 DSP,用来连接 AI 加速器、服务器和云数据中心,因此决策者位于基础设施侧,而不是应用团队内部。最清晰的直接客户细分是超大规模云和大型云运营商,它们必须用确定性、能效高的带宽连接机架、排、楼宇和园区。第二类是 AI 云专家和新型云运营商,它们聚合模型开发者需求,但仍以 GPU 集群所有者或运营者身份采购基础设施。第三类是光模块厂商和基础设施 OEM,它们能把 DSP 认证进可插拔或嵌入式系统,再把通往超大规模云的访问权中介化。公开招聘尤其有用:Celero 正在招聘人员,把超大规模云网络需求转化为产品规格,并争取超大规模云和模块厂商的早期设计导入。这意味着买方是网络架构师和基础设施高管,用户是光学、硬件和系统团队,付款方是 AI 网络资本开支预算。[CU001, CU002, CU003, CU004, CU011, CU014]
| 客群 | 买方 / 用户 / 付款方 | 用例 | 规模信号 | 收入 / 战略价值 | 缺口 |
|---|---|---|---|---|---|
| 超大规模云厂商和云平台 | 买方 = 网络 / 基础设施高管;用户 = 光网络、系统和硬件架构师;付款方 = AI 基础设施 capex | 跨加速器集群、楼宇和数据中心扩展链路 | CapitalG/Google 邻近关系;OIF 演讲者来自 Google Cloud、Microsoft、Meta;NVIDIA 和 AWS 描述百万芯片规模需求 | 战略价值极高;一个 socket 可扩展到多个站点和产品代际 | 未披露具名 Celero 超大规模客户或已验证项目 |
| AI 云专业服务商 / neo-cloud | 买方 = 云平台和生产工程负责人;用户 = 集群 / 网络运维团队;付款方 = GPU 云 capex | 面向以服务方式运营的训练和推理集群,提供低延迟互连 | CoreWeave 报告 49 个 AI 数据中心、>1GW 活跃功率和 >100,000-GPU 集群 | 战略价值高;即便 Celero 卖给一个平台方,也能聚合许多模型开发者的需求 | 未披露 Celero 与任何具名 AI 云运营商的直接部署 |
| 光模块厂商和 OEM | 买方 = 产品和平台负责人;用户 = 模块、SerDes、DSP 和验证工程师;付款方 = 平台 R&D 和组件采购预算 | 把 DSP 嵌入或验证到卖给超大规模客户的收发器或子系统中 | Celero 市场岗位明确瞄准模块厂商以拿下早期 design-win | 渠道杠杆高;一个通过验证的模块伙伴可打开多个终端客户 | 毛利捕获和路线图控制可能在渠道伙伴手里,不在 Celero 手里 |
| 系统集成商 / 生态伙伴 | 买方 = 平台合作负责人;用户 = 参考设计和互操作团队;付款方 = 解决方案开发预算 | 面向 AI fabric 的参考架构、验证和互操作 | CapitalG 和 OIF 都强调伙伴网络、标准和互操作工作 | 战略价值中等;改善上市路径和可信度 | 未披露公开伙伴名单或 attach-rate 数据 |
各行反映公开 buyer-fit 证据,而不是已披露收入组合。战略价值是方向性判断,因为 Celero 未公开客户数量、平均销售价格或账户集中度。
[CU001, CU003, CU011, CU014, CU024, CU025]Celero 可能的客户动作先从架构痛点和标准对齐开始,再进入送样和认证,最后才变成多站点扩张。
[CU004, CU017, CU018, CU035, CU038]6.2 采用轨迹与公开客户证明
采用故事仍更多是组织层面的,而不是收入层面的。Celero 2025 年 11 月融资公告和 CapitalG 投资笔记显示,公司正从技术建设进入商业认证;团队已约 100 人,积极招聘,并明确围绕锚定设计导入搭建商业化 动作。对早期半导体公司来说,这是可信牵引,但它不等同于已披露的量产部署。Celero 自己的网站、投资人页面和新闻页没有公布客户数、出货量、利用率或具名量产账户。唯一具名、且有直接战略价值的外部验证者是 CapitalG/Alphabet,其投资论点聚焦 Celero 试图解决的同一个 AI 规模网络瓶颈,其平台也宣传可引荐 Google 顾问、客户账户和渠道伙伴。除此之外,证据是群体层面的,而不是具名客户层面的;公司显然在追逐超大规模云、模块厂商和 AI 云运营商,但已送样账户的确切数量和每段关系所处阶段仍未披露。承销时,这意味着公开证据对买方匹配和商业动作的支持强于对已实现采用的支持。[CU005, CU006, CU007, CU008, CU009, CU010]
| 指标 | 数值 | 日期 | 来源 | 置信度 | 含义 | 缺失分母 |
|---|---|---|---|---|---|---|
| 总融资额 | 总计 $140M;CapitalG 领投 $100M Series B | 2025-11-17 | Celero 新闻稿 + CapitalG | 高 | 一家 2024 年成立的基础设施芯片初创公司拿到了有分量的商业验证 | 未披露 bookings、客户数量或 design-win 数量 |
| 商业组织搭建 | 设立 Marketing Director 岗位,目标是拿下超大规模客户和模块厂商的 anchor design-win | 2026-07-25 访问 | Celero Greenhouse 招聘页 | 高 | 表明公司正从纯 R&D 转向主动获客和产品化 | 不能证明任何客户赢单已经落地 |
| 招聘 / 组织规模 | 约 100 人团队,多个技术和商业岗位在招 | 2025-11 至 2026-07 | CapitalG + Celero 招聘页 | 中 | 说明公司有能力支撑送样、认证和路线图迭代 | 未披露 R&D、支持和销售之间的人力拆分 |
| 公开客户披露 | 复核过的官方页面未披露任何具名生产客户 | 2026-07-25 访问 | Celero 网站复核 | 高 | 公开牵引力仍不足以看到完整标杆客户 | 可能存在 NDA 下未公开的设计赢单 |
| 目标账户紧迫性 | AI 工厂规模被描述为数万到数百万颗芯片;互连路线图正迈向 448G/800G | 2025-2026 | NVIDIA, AWS, OIF, CoreWeave | 中 | 可触达买家的痛点已经发生,并非纸面假设 | 还需要目标账户中 Celero 专属转化数据 |
Celero 未披露账户、出货量或利用率指标,因此本表把公司进展指标与市场采用代理指标放在一起。没有具名客户本身就是重要的采用信号。
[CU005, CU007, CU008, CU012, CU017, CU027]| 客户 / 交易对手 | 细分 | 部署 / 用例 | 生产环境 vs 试点 | 结果 | 限制 |
|---|---|---|---|---|---|
| CapitalG / Alphabet 生态 | 拥有 Google 顾问和客户网络入口的投资方 | Celero $100M Series B 的领投方,并作为公开观点合作方讨论 AI 规模网络瓶颈 | 战略验证;并非披露的生产客户 | 最强的具名外部证据:贴近超大规模云厂商的利益相关方认为 Celero 具备商业相关性 | 没有公开声明称 Google 运营部门正在采购或部署 Celero 芯片 |
| 超大规模云厂商设计伙伴(未披露) | 直接终端客户目标群体 | 营销岗位明确要争取超大规模云厂商的锚定设计赢单 | 可能处于评估 / 送样 / 设计导入阶段 | 说明 Celero 有一条通向最高价值买家类别的活跃早期采用漏斗 | 未披露名称、数量、送样规模或认证里程碑 |
| 光模块厂商(未披露) | 渠道 / OEM 目标群体 | 招聘信息明确把模块厂商与超大规模云厂商放进早期设计赢单动作 | 可能处于认证 / 生态接洽阶段 | 显示 Celero 正在推进能打开多个终端账户的渠道层 | 未披露具名模块伙伴、插槽赢单或量产模块 |
| AI 云基础设施运营商(未披露) | 次级直接客户群体 | 产品简报契合 CoreWeave 等 AI 云运营商描述的横向扩展、低时延网络需求 | 仅为潜在客户匹配;未披露 Celero 公开部署 | 把合理客户基底从传统超大规模云厂商扩展出去 | 这是 ICP 证据,不是直接的 Celero 客户证明 |
Celero 尚未发布客户名单,因此覆盖不完整。本表把一个具名战略验证者,与公开但仍未指向具体客户标识的群体级设计赢单信号分开。
[CU003, CU004, CU005, CU006, CU009, CU010]公开证据从庞大目标市场很快收窄到一组小得多、且大多未披露的真实设计伙伴关系。
计数是基于本章审阅的市场结构和公开证据得出的方向性估算,不是公司披露的客户总数。
[CU003, CU008, CU010, CU024, CU029, CU030]证据质量在战略买家匹配上最强,在量产部署和留存可见度上最弱。
[CU010, CU014, CU015, CU016, CU017, CU030]6.3 耐久性、留存可见度与采购摩擦
Celero 没有公开 NRR、GRR、流失率、合同期限或客户满意度披露,因此耐久性只能从半导体和互连采购行为推断,而不能直接衡量。好消息是终端市场真实且紧迫。AWS、NVIDIA、CoreWeave 和 OIF 材料都描述 AI 基础设施正在迈向数万、甚至数百万颗芯片,确定性时延、互操作性和无阻塞网络正在成为设计约束,而非可选升级。这些条件会拉长认证周期;组件一旦获批进入量产架构,黏性也会提高,因为重新认证昂贵,供应连续性很重要。坏消息是走向量产缓慢且牵动全系统。CoreWeave 的 2026 年生产推理分析称,多数企业从购买 GPU 容量到服务第一个生产请求需要超过四个月,供应和基础设施可得性仍是硬约束。实际看,Celero 一旦赢得插槽,可能受益于长生命周期;但投资人还无法验证公司是否已有任何量产插槽,或评估是否正以有吸引力的速度转化。[CU012, CU016, CU017, CU018, CU032, CU033]
| 指标 | 数值 / 缺口 | 置信度 | 含义 | 尽调路径 |
|---|---|---|---|---|
| NRR | 未披露 | 高 | 没有公开证据证明扩张效率 | 索取按客户类型和客户批次年份拆分的 NRR |
| GRR / logo 流失 | 未披露 | 高 | 无法把毛留存粘性与未来扩张预期拆开 | 索取按头部账户拆分的 GRR、logo 流失和流失原因 |
| 合同期限 / 供应期限 | 未披露 | 高 | 难以判断锁定效应或未来收入可见度 | 索取标准评估、认证和供应协议期限 |
| 设计赢单持续性代理指标 | 一旦通过认证,可能较高;认证和互操作负担不轻 | 低 | 赢下的插槽可能有粘性,但结论是推断而非实测 | 向管理层验证重新认证成本、插槽生命周期和平台份额 |
| 客户满意度 / 参考客户质量 | 未披露 G2 式评价或 NPS;公开证据主要面向投资人和市场 | 高 | 参考客户深度仍是持续性最大的盲区 | 索取具名参考客户、续约叙事以及支持 / 升级处理指标 |
Celero 不披露留存经济性。因此,持续性判断依赖行业采购行为、互操作负担和推断出的认证成本,而不是已观察客户批次或续约数据。
[CU033, CU035, CU036, CU037]初始认证后的关系类型留存估算;Celero 没有发布真实队列数据,因此这里只作方向性代理。
这些百分比是示例性估算,来自公开信号中体现的认证负担、互操作要求,以及重新认证 AI 网络组件的成本;它们不是 Celero 披露,尽调时应替换为真实续约和插槽寿命数据。
[CU033, CU035, CU037, CU040]6.4 扩张路径与集中度风险
Celero 最理想的客户动态是典型先落地、再扩张:赢下一个超大规模云插槽、一次模块厂商认证或一个 AI 云设计伙伴关系,再从单一链路类别扩展到更多机架、楼宇、园区和产品世代。CapitalG 模式可通过引荐和联合销售路径帮助这一过程;OIF 等标准机构也显示,买方社区正围绕 400G、448G 和 800G 互连路线图聚拢。但这些事实也带来集中的下行风险。少数超大规模云和模块生态控制着大部分相关需求,因此一个锚定账户可能严重影响路线图优先级和早期收入结构。渠道依赖也很实质,因为模块厂商、OEM 和系统集成商可能控制认证入口和采购杠杆。最后,竞争压力并非假设;Lightwave、Coherent、NVIDIA、Broadcom 和其他生态既有厂商已经向同一批买方营销 AI 数据中心网络、光学和供应伙伴关系。因此,商业问题不是市场是否存在,而是 Celero 能否在既有厂商和伙伴守门人吸走预算之前,把少数战略关系转化为多元化生产收入。[CU019, CU020, CU021, CU022, CU023, CU038]
| 集中度风险 | 扩张路径 | 渠道风险 | 尽调问题 |
|---|---|---|---|
| 一个锚定超大规模云客户可能主导早期路线图和收入结构 | 在同一账户内,从一个已验证链路类别扩到更多集群、园区和产品代际 | 直客可能要求定制功能,拖慢更广产品组合的多元化 | 索取头部账户管线以及任何独家或定制功能承诺 |
| 对模块厂商的依赖可能卡住终端客户入口 | 一个通过认证的模块伙伴可以打开多个超大规模云项目 | 伙伴可能吃掉利润并控制设计席位 | 索取具名模块伙伴、利润拆分和伙伴来源管线占比 |
| AI 云客户赢单可以聚合多个下游用户,但仍集中在一个平台所有者身上 | 在扩展到超大规模云之前,把 neo-cloud 作为动作更快的参考客户 | 云服务商预算周期和 GPU 供应约束可能推迟部署 | 索取与 AI 云运营商签署的任何 POC 或送样协议 |
| 投资方与客户网络重叠可能加速触达,但模糊独立验证 | CapitalG / Google 网络可以缩短引荐和联合销售周期 | 战略投资方影响力可能把路线图偏向单一生态 | 询问任何 CapitalG 或 Google 关系是否包含商业权利、价格优惠或路线图约束 |
| 既有光通信 / 网络生态拥挤,并由供给侧主导 | 靠功耗、带宽效率和对数据中心相干链路的适配来差异化 | Broadcom、Coherent、NVIDIA、Molex 相关厂商和其他既有厂商已在争夺同一批买家 | 索取相对既有替代方案的胜负数据,以及潜在客户给出的认证理由 |
扩张逻辑说得通,但买家群体小、认证负担重,早期收入很可能集中。每一行都把可信上行路径和它带来的依赖拆开。
[CU019, CU020, CU023, CU038, CU039, CU040]6.5 展示材料
07风险
7.1 按严重程度排序的风险图谱
Celero 的公开故事很吸引人,但仍极早期。公司成立于 2024 年,到 2025 年末融资 $140M,并把相干 DSP 芯片定位为在 AI 基础设施中提供低功耗光连接的路径。同样这些事实构成本章最高的剩余风险。技术上,Celero 所处品类中,既有厂商已经把功耗、距离和光学集成定义为门槛;首版硅片或认证时间一旦失败,可能在营收可见之前抹掉宝贵日历时间。运营上,公司仍在 ASIC、固件、验证、模拟和产品等核心岗位招聘,说明团队纵深仍在搭建,而产品路线图已经被期待交付。商业上,Celero 公开材料尚未披露具名客户、代工伙伴或已出货产品世代,因此投资人是在市场证明出现前承销架构和团队质量。财务上,一家没有公开营收证据、但资本需求巨大的芯片初创公司,如果流片、性能或客户时间表滑坡,可能很快从资金充足变成资本受限。[CR001, CR003, CR004, CR005, CR006, CR009]
| 风险 | 类别 | 严重性 | 概率 | 触发因素 | 缓释措施 | 监控信号 |
|---|---|---|---|---|---|---|
| 首版芯片或 tape-out 未达标 | 技术 | 致命 | 中 | bring-up、良率、功耗或互操作目标延期 / 未达标 | 阶段门验证、外部 PHY / 模块伙伴,并为重新流片留足现金 | tape-out 日期、bring-up 公告、基准测试披露 |
| 功耗预算不达标 | 技术 / 市场 | 高 | 高 | 每 bit 功耗达不到超大规模云门槛 | 围绕单链路功耗优化架构,并用第三方演示证明 | 公开功耗指标、客户试用反馈、热设计变化 |
| TSMC 或台湾中断 | 供应链 / 地缘政治 | 致命 | 中 | 出口许可证问题、自然灾害或地缘政治升级冲击晶圆厂连续性 | 规划多个季度缓冲、准备封装备选方案,并让客户排期贴近现实 | 晶圆厂披露、政策变化、台湾事件风险、交期拉长 |
| 出口管制扩张 | 监管 | 高 | 中 | 新的 BIS 规则或实体筛查收窄可触达客户或伙伴 | 尽早完成 ECCN、所有权筛查和出口律师复核 | BIS 通知、D:5 指引变化、许可证处理时间 |
| 客户集中 / 设计导入丢失 | 商业 | 致命 | 中 | 一两个头部机会延期,或选择既有厂商 / 内部替代方案 | 拓宽漏斗,先排布试点再承诺大单,聚焦差异化用例 | 具名设计赢单、管线多样性、重复评估赢单 |
| 关键人员流失 | 执行 | 高 | 中 | 创始人或资深 DSP / 模拟负责人离职 | 留任激励、岗位冗余和更深的中层管理 | 高管招聘节奏、组织流失、猎头搜索强度 |
| IP 或商业秘密纠纷 | 法律 | 高 | 低到中 | 前雇主主张受保护专有技术被滥用 | 洁净室流程纪律、律师复核和文档来源卫生 | 律师函、诉讼、招聘放缓、尽调例外 |
严重性和概率是在可见公开缓释措施之后的剩余暴露排序;公开来源尚未披露 Celero 内部 tape-out、客户或合规控制。
[CR003, CR009, CR017, CR026, CR029, CR041]剩余风险最重的区域,是技术执行、供应集中和客户集中交叠处。
[CR017, CR026, CR029, CR041, CR042, CR043]7.2 法律、监管与地缘政治暴露
即使目前没有已知公开纠纷,法律和监管风险仍然重大。Celero 的领导班底来自前 Marvell、Inphi、Broadcom 和 ClariPhy 人才;如果前雇主认为受保护的 know-how 被带入新公司,美国商业秘密法为其提供了有力救济。Defend Trade Secrets Act 允许普通民事索赔,在特殊情况下还允许单方扣押令,这意味着纠纷可能在终局裁判前很久就破坏价值。 出口侧,Celero 所在的半导体类别正处在美国先进计算管制的活跃范围内。BIS 负责执行 EAR,维护 Entity List 和 MEU 管制等基于主体的限制,并在 2026 年 5 月重申,涉及 D:5 和总部在澳门实体的某些先进计算物项仍需许可证。 因此,任何后续出口分类或客户所有权问题,都可能影响 Celero 可触达市场、合作伙伴选择和交付节奏。CHIPS 护栏规则又加一层: 如果 Celero 日后申请联邦激励,在关注国家进行重大扩张或开展受限技术许可,可能触发追回或失格。最后,先进半导体最关键的制造基地仍高度地理集中, TSMC 也警告,出口管制、制裁、地震、网络攻击和地缘政治紧张都可能扰乱供应。[CR002, CR014, CR015, CR016, CR017, CR018]
| 法规 / 法律 | 司法辖区 | 暴露类型 | 当前状态 | 缓释措施 |
|---|---|---|---|---|
| EAR Parts 730-744 与 BIS 管辖 | 美国 | 先进半导体出口合规 | 现行有效,直接关系先进计算物项和受限终端用户 | 尽早完成产品分类,筛查交易对手,并预先谈妥许可证策略 |
| BIS 2026 年 5 月 D:5 / Macau 指引 | 美国 | 与受限总部 / 母公司所有权相关的先进计算物项许可证要求 | 现行澄清,重申持续许可证义务 | 开展最终母公司尽调,发货前升级处理所有权变化 |
| CHIPS 国家安全护栏 | 美国 | 若获资助方在受关注国家扩产或参与受限许可,可能触发追回或丧失资格 | 适用于申请方 / 获奖方 | 避免被禁止的海外扩张,并记录国家安全合规 |
| 《商业秘密保护法》 | 美国 | 前雇主或伙伴提出民事商业秘密索赔风险 | 州际 / 对外商业争议中,原告始终可以援引 | 采用洁净室做法,复核发明转让,并由律师主导尽调 |
| 隐藏雇佣 / 发明转让条款 | 私人合同 | 公开来源看不到的潜在竞业限制、保密或转让摩擦 | 基于已复核公开证据无法验证 | 尽调时收集创始人和关键工程师协议 |
本表覆盖公开法律和监管暴露,以及一个重要合同缺口;不能替代公司律师复核或面向具体客户的出口分析。
[CR014, CR016, CR017, CR019, CR020, CR022]| 依赖 | 风险类型 | 严重性 | 缓释措施 | 尽调问题 |
|---|---|---|---|---|
| TSMC 先进制程制造 | 晶圆厂集中 / 地缘政治 | 致命 | 保留排期缓冲,认证封装和模块伙伴,并按现实供应假设分阶段承诺客户 | 索取制程节点、晶圆厂、封装和备份制造计划 |
| 先进半导体出口许可证链条 | 贸易合规 | 高 | 客户接触前完成 ECCN 分析、筛查和外部律师复核 | 索取 ECCN、出口矩阵和受限方工作流 |
| 园区 / 长距光模块生态 | 集成 / 可靠性 | 高 | 采用经验证的模块伙伴和第三方互操作实验室 | 索取模块伙伴名单、测试结果和现场失效假设 |
| 内部验证和 bring-up 带宽 | 运营执行 | 高 | tape-out 前在验证、固件和产品工程上建立冗余 | 索取招聘计划、组织架构和里程碑负责人 |
运营风险主要来自晶圆厂、集成和合规依赖,而不是任何已披露现场事故。
[CR005, CR017, CR026, CR028, CR029, CR031]7.3 执行、竞争与客户风险
Celero 面对的不是静态细分市场。Marvell 和 Broadcom 已经把多代光 DSP、共封装光学、自定义 ASIC 和机架级 AI 网络平台卖给超大规模云厂商和大型系统厂商。Broadcom 的公开材料显示,其技术栈覆盖自定义 XPU、交换、PHY、光组件,甚至 102.4 Tb/s 共封装光学交换机;Marvell 也披露了客户集中度,以及数据中心市场设计赢单的战略重要性。关键在于,初创公司的第一批收入很可能集中在少数设计导入上, 而这正是既有厂商杠杆、关系和路线图可信度最强的地方。开放光标准可能扩大可服务生态,但也可能把部分价值栈商品化,压缩小厂商仅靠专有集成取胜的空间。 客户验证缺口仍然很大:Celero 的公开材料没有点名已部署买家、生产节点或代工关系,因此投资人还无法把真正的商业拉力与 AI 基础设施生态热情区分开。 落到执行上,Celero 必须扩团队、完成硅片、证明每瓦性能,并拿下少数超大规模云级项目;与此同时,资本更厚的既有厂商还在继续推高标准。[CR012, CR013, CR032, CR033, CR034, CR035]
| 人员 / 团队 | 风险类型 | 严重性 | 缓释措施 | 监控信号 |
|---|---|---|---|---|
| Nariman Yousefi 和 Oscar Agazzi / 创始人核心 | 战略和架构集中 | 高 | 在创始人之下补强有经验的产品、运营和客户侧负责人 | 董事会补员、VP 招聘、客户 / 产品职责下放 |
| DSP 和算法团队 | 稀缺专才 | 高 | 留任激励,并在固件 / DSP / 架构角色上做冗余 | 开放岗位、人员流失和咨询依赖 |
| 物理设计、验证和产品工程 | tape-out 准备度 | 高 | 前置招聘,必要时用按里程碑管理的外部支持 | 岗位填补周期、承包商比例和验证收敛速度 |
| 招聘和梯队建设引擎 | 扩张执行 | 中到高 | 保持强技术招聘和清晰排序纪律 | 资深招聘人员利用率、招聘周期和录用接受率 |
公开招聘证据显示,Celero 仍在同时搭建多个关键职能;本表衡量的是梯队深度风险,而不是个人能力。
[CR002, CR005, CR025, CR038, CR039, CR041]7.4 致命场景、缓释措施与监测信号
评估 Celero,最清楚的框架不是泛泛谈初创公司不确定性,而是看致命场景。第一条致命链是硅片失败:如果首版硅片没有达到功耗、良率或互操作目标, Celero 可能持续烧钱,却拿不到下一轮融资或批量承诺所需的客户证明。第二条是代工或地缘政治:TSMC 准入问题、台湾扰动,或出口管制进一步扩张, 即使设计技术上成立,也可能打破进度、出货或成本假设。第三条是商业集中:如果一两个超大规模云级机会失败,或客户改选自研方案、既有厂商的开放标准生态, 工程投入与收入之间的缺口可能撑不住。第四条是法律摩擦:即便今天没有公开案件,商业秘密或雇佣纠纷一旦出现在错误时间点,可能卡住招聘、拖慢尽调或收窄战略选项。 缓释因素确实存在——资本充足、创始人经验深、顺着光学 AI 趋势、市场明显需要低功耗长距互连——但在公开证据证明流片进展、可量化功耗性能、 客户认证以及超出创始核心的团队深度之前,剩余风险仍然高。[CR003, CR017, CR020, CR026, CR029, CR030]
| 情景 | 概率 | 触发事件 | 影响 | 时间窗口 |
|---|---|---|---|---|
| 首版芯片未达功耗 / 良率 / 互操作目标 | 中 | 需要重新流片或漫长 bring-up,且没有客户认证里程碑 | 现金消耗上升,下一轮融资更难,客户项目延期 | 6-18 个月 |
| TSMC / 台湾中断或出口许可证中断 | 中 | 晶圆厂中断、年度许可证问题或政策升级 | 供应切断、出货延迟、利润承压和融资压力 | 立即到 24 个月 |
| 没有超大规模云级设计导入,或早期客户流失 | 中 | 一两个领先机会选择既有厂商、开放标准路径或自研方案 | 收入证据迟迟没有出现,估值大幅压缩 | 12-24 个月 |
| 产品爬坡期发生商业秘密或雇佣纠纷 | 低到中 | 律师函、禁令申请或尽调红旗 | 管理层分心、招聘受阻和交易风险 | 立即到 18 个月 |
致命情景聚焦可能实质损害股权价值的事件,而不是常规创业波动。
[CR020, CR023, CR026, CR027, CR041, CR042]大多数打穿投资论点的路径,都经过芯片时点、功耗表现、代工集中或客户集中度过窄。
[CR013, CR017, CR023, CR041, CR043, CR044]接下来六个监控信号,决定 Celero 能否从“融到钱的承诺”推进到“还能融钱的执行”。
KPI 面板使用可监控里程碑,而非量化公司 KPI,因为 Celero 尚未公开披露收入、良率或客户指标。
[CR003, CR005, CR017, CR026, CR037, CR039]08估值
8.1 估值锚点与可比背景
Celero 的公开估值锚点清楚但不完整:公司宣布由 CapitalG 领投 $100M Series B,累计融资 $140M;多家独立报道把这一轮视为约 $1B 的独角兽标记。缺的是分母。已审阅的公开材料都没有披露收入、毛利率或具名客户量,因此这一轮无法像公开可比公司那样按简单收入倍数承销。 这很重要,因为公开市场的光通信与 AI 网络区间很宽,但仍有边界:Coherent 约为 9.4x EV/收入,Marvell 约为 21.2x,Broadcom 约为 25.3x;后两者由规模、软件或深嵌超大规模客户的能力支撑,而 Celero 尚未公开证明这些能力。 Celestial AI、Lightmatter、Ayar Labs 等私有光学同行表明,投资人愿意为稀缺互连资产支付战略溢价,但这些公司也披露了更大的融资体量, 有些还给出了更成熟的商业化叙事。因此,Celero 的 $1B 标记作为稀缺性溢价是可信的;作为由基本面支撑的估值,只能有条件地成立。[CV001, CV002, CV003, CV005, CV018, CV019]
| 公司 | 领域 / 产品 | 阶段 / 状态 | 估值或倍数 | 可比性 | 对 Celero 的调整 |
|---|---|---|---|---|---|
| Marvell | AI 网络 / 光 DSP / 定制芯片 | 上市 | 21.24x EV / Revenue;~$170-174B 市值 | 最接近的上市相干 DSP 邻近公司,且有超大规模云敞口 | 因上市公司规模、更宽产品组合和已披露收入而折价 |
| Broadcom | 定制 AI ASIC + 网络 + 基础设施软件 | 上市公司 | 25.34x EV / Revenue;市值约 $1.82-1.87T | 说明软件和超大规模云覆盖面如何拉高倍数 | 需大幅折价;Broadcom 的软件和规模无法迁移 |
| Coherent | 光学组件与光子供应链 | 上市公司 | 9.43x EV / Revenue;市值约 $55B | 最好的上市光学组件估值锚 | Coherent 已有规模且偏组件属性,可作有用下限 |
| Celestial AI | 光子 fabric 光互连 | 未上市 Series C1 轮 | 估值约 $2.5B;融资 $255M | 直接的未上市光学可比公司,战略投资者阵容强 | Celestial 资本底座更大,未上市融资动能披露更多 |
| Lightmatter | 光子计算 / 互连 | 未上市 Series D 轮 | 估值约 $4.4B;融资 $400M | 高溢价光子可比公司,可作战略上限 | Celero 阶段更早,商业规模的公开证据更少 |
| Ayar Labs | 面向 AI scale-up 的共封装光学 | 未上市 Series E 轮 | 估值约 $3.75B;融资 $500M | 说明投资人愿为稀缺光学瓶颈控制权付多少钱 | Ayar 资本更厚,强调生产就绪,生态背书更广 |
仅选择已披露的上市与未上市光学 / AI 基础设施可比公司;未上市样本不完整,因为很多创业公司轮次没有清晰披露估值或收入分母。
[CV010, CV012, CV015, CV018, CV019, CV020]| 收购方 | 标的 | 年份 | 交易金额 | 倍数 | 战略逻辑 | 对 Celero 的适用性 |
|---|---|---|---|---|---|---|
| Marvell | Inphi | 2020 | $10B | 为云和 5G 基础设施补强电光互连领导力 | 相关性最高,因为相干互连领导力就是明确收购理由 | |
| Cisco | Acacia Communications | 2021 | $4.5B | 强化光学在未来互联网中的关键积木地位,服务相干网络和数据中心 | 相关,因为 Cisco 为相干光学和 DSP 邻近的互连深度付费 | |
| Intel | Habana Labs | 2019 | $2B | 在市场完全成熟前补强数据中心 AI 芯片组合 | 可证明战略买家会提前收购稀缺 AI 芯片资产 |
交易金额在审阅的新闻稿直接披露时列示;倍数单元格留空,表示审阅来源未披露交易倍数。
[CV022, CV024, CV037]公开市场估值倍数划出一个区间;证据成熟后,Celero 要么证明自己配得上,要么以折价切入。
公开市场倍数是截至 July 2026 的当前市场快照;Celero 隐含倍数是用 $1B 上轮估值做的简单股权价值计算。
[CV010, CV012, CV015, CV026, CV028, CV029]8.2 方法论与情景分析
Celero 更适合用按证据加权的估值方法,而不是单点可比。公开市场倍数给出重力;私募轮次显示战略稀缺性能换来多高定价;并购先例说明, 当互连瓶颈威胁更大的平台路线图时,买方为什么愿意加价。基于这一点,牛市情景假设 Celero 把技术可信度转化为两三个超大规模云或模块平台插槽, 中期收入达到约 $180-220M;如果溢价倍数守住,可支撑 $2.0-2.5B 的结果。基准情景假设出现一波有意义部署, 收入转化较慢但真实,走向约 $75-100M;仍享有溢价的 10-12x 倍数,使价值维持在上轮附近,约 $0.8-1.2B。 熊市情景假设商业化延迟一个周期、既有厂商反击更强,或市场转向更低功耗的光学路径从而压缩 DSP 含量;该情景可把价值推向约 $0.4-0.7B。 由于公开披露很薄,情景表比经典 DCF 给出的伪精确更有决策价值。[CV006, CV007, CV022, CV023, CV024, CV028]
| 场景 | 收入假设 | 倍数 | 隐含估值 | 概率 | 关键驱动因素 |
|---|---|---|---|---|---|
| 乐观 | 2-3 个超大规模云厂商或模块平台客户落地后,中期收入 $180M-$220M | 12x-14x | $2.0B-$2.5B | 25% | 设计定点叠加战略稀缺性溢价仍能守住 |
| 基准 | 来自一次有意义部署波次和克制执行的收入 $75M-$100M | 10x-12x | $0.8B-$1.2B | 50% | 一轮真实设计定点能支撑估值标记,但不足以重估 |
| 悲观 | 延迟一个周期、现有厂商反击或架构切换后,收入低于 $50M | 较小收入基数上的 8x-10x | $0.4B-$0.7B | 25% | 商业化延迟和倍数压缩同时到来 |
所有情景均为示意,因为 Celero 未公开披露当前收入;概率权重带主观判断,绑定商业化节奏,而非审计财务报表。
[CV033, CV034, CV035, CV036, CV048]| 方法 | 输入 | 输出 | 权重 | 置信度 | 局限 |
|---|---|---|---|---|---|
| 上市公司收入倍数三角校验 | MRVL ~21.2x、AVGO ~25.3x、COHR ~9.4x EV / Revenue 倍数 | 创业公司折价后,标准化区间为 $0.7B-$1.2B | 35% | 中 | 上市可比公司规模更大、披露更充分;Broadcom 还受软件业务抬升 |
| 未上市轮次对比 | Celestial ~$2.5B、Lightmatter ~$4.4B、Ayar ~$3.75B 估值锚 | $1.0B 作为稀缺性溢价有可能成立,但并不明显便宜 | 25% | 中 | 未上市轮次偏好条款重,且很少披露干净的收入分母 |
| 战略 M&A 先例 | Inphi、Xilinx、Habana 显示买方会为瓶颈控制权付费 | 支持高于上市可比公司估值锚的上行可选性 | 20% | 低-中 | 交易较早,产品宽度和宏观环境不同 |
| 风险调整情景模型 / 类 DCF | 收入区间 <$50M、$75M-$100M 和 $180M-$220M,并做概率加权 | 期望值更集中在上一轮估值标记附近,而不是大幅高于它 | 20% | 低 | 未披露起始收入、利润率或股权结构数据 |
权重为启发式,用来约束投资建议,不表示公开证据之外还有模型精度。
[CV018, CV019, CV020, CV026, CV033, CV034]瀑布图展示:设计定点成功会把基准估值推高,延误和倍数压缩会把估值拉低。
数值是以百万美元计的情景示意,锚定公开可比公司的估值重心和私募可比溢价,而非公司披露收入。
[CV033, CV034, CV035, CV036, CV048]按概率加权后,当前轮估值已经接近模型区间的期望价值中枢。
最终分数是委员会的判断速记:5 意味着跟踪,而不是买入或回避。
[CV036, CV041, CV042, CV043, CV044, CV048]8.3 建议、置信度与尽调要求
建议立场是跟踪,置信度中等、风险评级高、估值偏紧。这个判断比直接回避更积极,因为市场顺风真实存在,创始团队背景高度相关, 私有可比公司也证明,当产品市场匹配出现时,光互连稀缺性能撑住大估值。它又不及买入,因为已审阅证据还没有证明商业化达到新投资人从 $1B 入场并承销强回报所需的水平。落到投委会,不需要再听更多光学需求故事;需要证明的是,Celero 正在把这些需求转化为可落地的设计卡位、 收入和可防守经济性。最快上调路径,是经验证的 run-rate 收入超过约 $100M、具名超大规模云或模块项目赢单, 以及部署层证据证明 Celero 的带宽和功耗主张在生产环境里站得住。最快下调路径,是流片延迟、客户转化弱、架构迁移离开偏 DSP 的链路, 或融资条款实质性压低新普通股价值。[CV003, CV038, CV039, CV041, CV042, CV043]
| 问题 | 所需证据 | 对估值立场的影响 | 优先级 |
|---|---|---|---|
| 当前年化收入和毛利率是多少? | KPI 材料包、审计或董事会口径收入桥、毛利率历史 | 如果规模已超过 $100M,估值可从偏高上调为合理或有吸引力 | 关键 |
| 哪些超大规模云厂商或光模块合作伙伴处于认证或量产阶段? | 具名客户清单、设计定点阶段、ASP 和放量时间 | 可验证乐观 / 基准情景,并支撑战略溢价 | 关键 |
| 2025 年 11 月轮次条款是什么? | 股权结构 waterfall、优先权、员工股权压力、任何二级交易定价 | 相比标题式投后估值,普通股价值可能被大幅改写 | 高 |
| 相比 LPO / CPO 替代方案,DSP 架构有多耐久? | 独立基准数据、客户部署功耗指标、路线图对比 | 可能守住长期倍数区间,也可能压缩它 | 高 |
| 现金消耗和后续融资计划是什么? | 预算、晶圆厂承诺、招聘计划、现金续航模型 | 决定稀释风险和投资人实际回报 | 高 |
清单按预期估值敏感度排序,而不是按一般公司质量好奇心排序。
[CV038, CV039, CV040, CV045, CV046]8.4 风险调整回报框架与监测指标
以上轮价格看,只有公司快速越过少数商业化关口,回报才呈现不对称。上行至约 $2.0-2.5B,需要同时抓住高速增长的 AI 光学市场, 并守住溢价估值环境;下行至约 $0.5B,只要错过一个采用周期或发生一次重大架构转向就可能出现。因此,正确的监测框架不是泛泛押注初创乐观, 而是一张与价值传导绑定的短驱动清单:设计赢单转收入、收入转毛利率、毛利率转融资杠杆、融资杠杆转最终所有权结果。投资人因此应跟踪具名设计赢单、 季度收入转化、毛利率轨迹、AI 光学需求增长,以及相对于 $140M 融资底座的现金消耗。战略先例在这里仍重要:Marvell-Inphi、Intel-Habana 和 AMD-Xilinx 都说明,瓶颈控制型资产即使还没有成熟的公开市场式披露,也可能因战略原因被收购。但只有 Celero 证明自己掌握了买方真正需要的瓶颈, 这些先例才有帮助。[CV022, CV023, CV024, CV037, CV040, CV047]
| 价值驱动因素 | 当前证据 | 乐观情形贡献 | 悲观情形拖累 | 尽调问题 |
|---|---|---|---|---|
| 超大规模云厂商设计定点 | 未公开披露具名定点 | 增加规模可信度,抬高退出可选性 | 没有定点,收入爬坡慢,溢价会坍塌 | 客户访谈和认证 pipeline |
| 收入转化 | 融资已披露;收入未披露 | 把稀缺性叙事变成可投资的倍数支撑 | 缺少证据会让 $1B 估值标记仍以叙事为主 | 董事会 KPI 包和订单转化桥 |
| 架构适配 | 市场增长大,但路线图正转向 LPO/CPO 效率 | 如果 Celero 仍不可或缺,战略溢价可延续 | 如果 DSP 含量被压缩,终局倍数会很快下滑 | 独立功耗 / 带宽基准数据 |
| 资本效率 | 迄今仅融资约 $140M;若执行够快,仍有杠杆空间 | 下一轮前拿出更多证据,可保住持股 | 制造或 GTM 扩张可能迫使稀释 | 现金续航模型和下一轮融资计划 |
| 战略买家可选性 | 光学和 AI 芯片领域都有 M&A 先例 | 可在上市可比公司估值锚上方形成出价底线 | 目前没有直接买方证据;可选性可能被夸大 | 投行反馈和主动接洽历史 |
这个框架把经营证据连接到投资人结果,强调哪些因素会改变股权价值,而不是只改善故事。
[CV037, CV040, CV047, CV048]价值创造取决于一条短链条:从产品验证到收入验证,再到战略可选性。
这是定性传导图,不是打分版 Monte Carlo 模型。
[CV003, CV037, CV040, CV041, CV043, CV047]8.5 附录图表
免责声明
本报告仅基于截至 2026-07-25 的公开信息。Celero Communications 是私营公司,尚未披露收入、利润率、客户名称或股权结构细节。所有财务估算均来自公开可比公司分析,不应作为投资决策依据。本报告不构成投资建议。
证据索引
| 编号 | 陈述 | 可信度 | 来源 |
|---|---|---|---|
| CO001 | Celero Communications officially describes itself as developing coherent DSP solutions for AI infrastructure. | 高 | SO001, SO002 |
| CO002 | Celero says its platform enables high-bandwidth, low-power optical connectivity for AI infrastructure. | 高 | SO002, SO005 |
| CO003 | Celero Communications says it was founded in 2024. | 高 | SO002, SO006 |
| CO004 | Celero Communications says it is headquartered in Irvine, California. | 高 | SO002, SO006 |
| CO005 | Celero says it has design centers in Canada and Argentina. | 高 | SO002, SO006 |
| CO006 | Celero’s homepage frames the problem as eliminating distance bottlenecks in AI infrastructure through optical transmission. | 中 | SO001 |
| CO007 | Celero’s official careers page listed 15 open roles across Ottawa, Irvine, and San Jose on the access date. | 中 | SO003 |
| CO008 | Celero’s official news page still centered Reuters coverage and the company’s own November 17, 2025 funding release as of 2026-07-25. | 中 | SO004 |
| CO009 | Nariman Yousefi is Celero’s co-founder and CEO. | 高 | SO005, SO006, SO008 |
| CO010 | Oscar Agazzi is Celero’s co-founder and CTO. | 中 | SO006, SO024 |
| CO011 | Celero’s founders are described as former senior executives from Marvell, Inphi, Broadcom, and ClariPhy. | 高 | SO005, SO008 |
| CO012 | James Luo of CapitalG was a member of the Celero board of directors by the November 2025 financing announcement. | 高 | SO005, SO007, SO008 |
| CO013 | Stefan Dyckerhoff of Sutter Hill Ventures was identified as a Celero board member in the November 2025 financing announcement. | 高 | SO005, SO007 |
| CO014 | Celero’s 2024 SEC Form D lists Nariman Yousefi as an executive officer and director. | 高 | SO019, SO020 |
| CO015 | Celero’s 2024 SEC Form D lists Oscar Agazzi as an executive officer and director. | 中 | SO020 |
| CO016 | Celero’s 2024 SEC Form D lists Stefan Dyckerhoff and Robert Abbott as directors. | 中 | SO020 |
| CO017 | A 2026-03-25 California-registry mirror lists Avijit Roy as CFO, Nariman Yousefi as Secretary, and Nariman Yousefi as registered agent. | 中 | SO021 |
| CO018 | Celero’s reviewed official website pages do not publish a full board roster, governance documents, or a complete executive bench beyond the founders. | 中 | SO001, SO002, SO003, SO004 |
| CO019 | Celero announced on 2025-11-17 that it had raised a total of $140 million. | 高 | SO005, SO007 |
| CO020 | Celero’s announced funding stack consisted of a $100 million Series B and earlier seed plus Series A rounds totaling $40 million. | 高 | SO005, SO008 |
| CO021 | CapitalG led Celero’s $100 million Series B round. | 高 | SO005, SO008 |
| CO022 | Sutter Hill Ventures led Celero’s earlier seed and Series A rounds, with participation from Valor Equity Partners, Atreides Management, Maverick Silicon, and others. | 高 | SO005, SO024 |
| CO023 | Celero’s 2024 SEC Form D recorded a first sale date of 2024-12-11. | 中 | SO020 |
| CO024 | Celero’s 2024 SEC Form D reported $29,999,906 sold to 40 investors in that offering. | 中 | SO020 |
| CO025 | Intelligence360 reported that Celero later filed a notice to raise up to $99,999,994 in new funding in late 2025. | 中 | SO022 |
| CO026 | CB Insights classifies Celero Communications as a Series B company that is alive. | 中 | SO017 |
| CO027 | CB Insights lists Celero’s total raised at $138.4 million and its last raise at $100 million. | 中 | SO017 |
| CO028 | The reviewed public sources did not disclose an exact post-money valuation or explicitly confirm unicorn status for Celero’s 2025 financing. | 中 | SO005, SO008, SO017 |
| CO029 | The reviewed public sources did not disclose Celero’s revenue or ARR. | 中 | SO005, SO008, SO017 |
| CO030 | The reviewed public sources did not disclose a customer count or name any customers for Celero. | 中 | SO004, SO005, SO008, SO024 |
| CO031 | Official and independent profiles consistently frame Celero’s business model as selling coherent DSP platforms or chips for AI and data-center optical interconnects. | 高 | SO001, SO002, SO017, SO023 |
| CO032 | Reuters reported that Celero is designing a chip that translates bursts of light sent over fiber-optic cables into electrical ones and zeros used in computing networks. | 中 | SO008 |
| CO033 | Reuters reported that Celero plans to use algorithms in its chips to send more information over existing fiber faster and with lower power. | 中 | SO008 |
| CO034 | Celero’s careers page and company about page together support a distributed North American operating footprint spanning Ottawa, Irvine, and San Jose, alongside the broader Canada and Argentina design-center framing. | 高 | SO002, SO003 |
| CO035 | CB Insights, BizProfile, and FormDs all point to 5281 California Avenue Suite 200 in Irvine as Celero’s headquarters address. | 中 | SO017, SO021, SO025 |
| CO036 | Celero’s 2024 SEC Form D used a Laguna Beach mailing address, implying that the publicly surfaced Irvine headquarters address was documented later. | 中 | SO019, SO020, SO021 |
| CO037 | Tech in Asia argued that Celero entered a crowded 800G coherent DSP field with an unclear edge because it had not publicly disclosed key specs such as baud rate, process node, or power. | 中 | SO024 |
| CO038 | Tech in Asia said Celero had not publicly shared tapeout status or partnerships such as working 800G systems at OFC 2025. | 中 | SO024 |
| CO039 | The official careers page listed 15 open roles, which is clear evidence of active team expansion even though exact headcount remains undisclosed. | 中 | SO003 |
| CO040 | The reviewed public sources did not disclose debt facilities, credit lines, or secondary share transactions. | 中 | SO005, SO008, SO017, SO024 |
| CO041 | No material leadership change was identified in the reviewed 2025-2026 public sources; the same founders remained the public face of the company. | 中 | SO004, SO005, SO006, SO008 |
| CM001 | Celero says it develops coherent DSP solutions for high-bandwidth, low-power optical connectivity in AI infrastructure. | 高 | SM001, SM002, SM003, SM004 |
| CM002 | Celero positions its product around connecting large AI networks and scale-across infrastructure rather than generic carrier transport. | 高 | SM001, SM002, SM003 |
| CM003 | CapitalG describes today's alternatives as PAM4, which becomes reach-limited and noisier at 1.6T and 3.2T-class AI fabrics, and legacy coherent, which remains over-engineered and power-hungry for dense data centers. | 中 | SM004 |
| CM004 | Celero and CapitalG both frame AI interconnect as a bottleneck created by clusters expanding from single racks toward campus-scale or multi-building fabrics. | 高 | SM003, SM004 |
| CM005 | SemiEngineering and ADTEK both describe OFC 2026 as evidence that optical interconnect has shifted from telecom support technology to core AI infrastructure. | 中 | SM012, SM013 |
| CM006 | A disciplined market definition includes optical DSPs, transceivers, coherent pluggables, and future CPO or NPO used for AI scale-out and scale-across, while excluding generic compute and most carrier transport capex. | 中 | SM001, SM017, SM018 |
| CM007 | Cisco says AI workloads, cloud-native applications, and distributed data are increasing demand for scalable, simple, high-bandwidth data-center interconnect. | 中 | SM018 |
| CM008 | Cisco says coherent pluggables inserted directly into switches and routers can extend 400G links from roughly 40 km to more than 1,000 km, depending on amplification. | 中 | SM018 |
| CM009 | Cisco cites network-operations demand for low-latency, secure inter-data-center links and says 79% of enterprise networks need lower latency or higher throughput to fully support AI workloads. | 中 | SM018 |
| CM010 | A Yole summary says the optical transceiver market is $23.4 billion in 2025 and could reach $112.3 billion by 2031. | 中 | SM023 |
| CM011 | The same Yole summary says AI infrastructure should account for more than 90% of total optical transceiver market revenue by the end of the forecast period. | 中 | SM023 |
| CM012 | Yole identifies silicon photonics as the dominant next-generation platform while EML lasers, DSPs, and photonic integration remain bottlenecks. | 中 | SM023 |
| CM013 | Dell'Oro says Ethernet data-center switch sales nearly doubled versus 2022 by 2025 because of rapid AI back-end deployment. | 中 | SM006 |
| CM014 | Dell'Oro expects strong double-digit AI networking spending in 2026 and says co-packaged optics could add multibillions to market size while demand remains supply-constrained. | 中 | SM006 |
| CM015 | Dell'Oro expects initial volume ramp of co-packaged optics on both InfiniBand and Ethernet switches in 2026, with major hyperscalers already trialing the technology. | 中 | SM006 |
| CM016 | Converge Digest says Dell'Oro forecasts spending on switches deployed in AI back-end networks for accelerated servers will surpass $100 billion over 2025-2029. | 高 | SM006, SM007 |
| CM017 | Dell'Oro expects most AI back-end switch ports to be 800 Gbps by 2025, 1,600 Gbps by 2027, and 3,200 Gbps by 2030. | 中 | SM007 |
| CM018 | Dell'Oro says demand is broadening beyond Tier 1 cloud providers to Tier 2 and Tier 3 providers and large enterprises. | 中 | SM007 |
| CM019 | Dell'Oro says the optical transport market reached about $16 billion in 2025 and DCI purchases grew nearly 40%, led by hyperscalers and neo-cloud buyers. | 中 | SM009, SM010 |
| CM020 | Converge Digest and RCR summarize Dell'Oro's view that optical transport equipment will grow 16% in 2026 to above $18 billion, with cloud-provider DCI purchases up about 50% in early 2026 because of AI data-center demand. | 中 | SM008, SM010 |
| CM021 | LightCounting says 800G PAM4 chipset shipments nearly tripled in 2025 and sales more than doubled year over year as hyperscalers increased AI infrastructure investment. | 高 | SM005, SM021, SM022 |
| CM022 | LightCounting now expects 800G optical-transceiver shipments to more than double again in 2026, 1.6T shipments to reach tens of millions of ports, and 1.6T chipsets to exceed $2 billion in 2026. | 高 | SM005, SM021, SM022 |
| CM023 | LightCounting says coherent DSP sales grew only 16% in 2025 because the first AI volume wave favored PAM4 optics more than coherent DSPs. | 中 | SM005 |
| CM024 | LightCounting expects coherent-lite to begin meaningfully increasing coherent DSP sales by 2027 and coherent DSP shipments to exceed 8 million units by 2031. | 中 | SM005 |
| CM025 | LightCounting says AI training spread across multiple facilities on one campus could create upside to coherent DSP demand because it requires greater inter-building bandwidth. | 中 | SM005 |
| CM026 | Cignal AI says pluggable coherent optics now dominate growth, with 800ZRx forecast to ship more than 200,000 ports in 2026 and 1600ZRx arriving in 2027. | 中 | SM014 |
| CM027 | Acacia says geographically distributed AI clusters are becoming critical because power and space constraints limit single-site growth, and it expects 800G ZRx pluggables to be the workhorse before 1.6T ZRx. | 中 | SM019 |
| CM028 | NVIDIA says its co-packaged silicon photonics delivers 5x better power efficiency and 5x sustained AI application runtime compared with pluggable transceivers. | 高 | SM015, SM016, SM024 |
| CM029 | NVIDIA names CoreWeave, Lambda, Meta, Microsoft, and Oracle Cloud Infrastructure as first adopters of Spectrum-X photonics. | 中 | SM015 |
| CM030 | NVIDIA says Spectrum-X Ethernet Photonics provides 409.6 Tb/s across 512 ports of 800 Gb/s and is slated for second-half-2026 availability. | 高 | SM015, SM016 |
| CM031 | Broadcom says AI networking needs a shift from electrical to optical-based scale-up architectures and is backing an open OCI MSA to enable multi-vendor interconnect. | 中 | SM017 |
| CM032 | Broadcom's OFC 2026 portfolio still includes active electrical cables with up to 6 meters of reach alongside optical DSPs and CPO, implying electrical links remain relevant at very short reach. | 中 | SM017 |
| CM033 | ADTEK says large-scale CPO deployment in scale-up architectures is more likely around 2028 and that current deployments remain hybrid: copper or AEC in-rack, pluggable optics for scale-out, and emerging CPO or NPO for future scale-up. | 中 | SM013 |
| CM034 | SemiEngineering argues all high-bandwidth AI data-center interconnects could become optical within five years, with CPO replacing pluggables in scale-out and copper in scale-up as reliability improves. | 中 | SM012 |
| CM035 | CNBC reports that copper remains the main connectivity standard today because it is cheaper and more reliable, while manufacturing yield and packaging challenges make broad photonics adoption likely from 2028 onward rather than immediately. | 中 | SM025 |
| CM036 | Gazettabyte cites SemiAnalysis and Meta evidence suggesting optics failures can be very costly at cluster scale, including a model where a 100,000-GPU cluster sees a first failure within 26 minutes and cluster compute can drop 40% from such failures. | 中 | SM026 |
| CM037 | The immediate buyer is often an OEM or switch vendor, but hyperscaler qualification teams heavily influence which DSP and optics platforms reach production. | 中 | SM018, SM019, SM015 |
| CM038 | The ultimate payer is usually the AI infrastructure budget owner at a hyperscaler, neo-cloud, or enterprise campus, with OEM product budgets funding earlier design stages. | 中 | SM006, SM007, SM018 |
| CM039 | A conservative SAM lens should exclude most carrier transport and short in-rack electrical spend today while including coherent pluggables, optical transceivers, and future optical engines used for AI scale-out and campus links. | 中 | SM017, SM018, SM015 |
| CM040 | An evidence-constrained sizing stack places Celero inside a broad datacom-optics outer ceiling, a middle DCI and optical-transport budget layer, and a narrow coherent-pluggable / coherent-DSP lens rather than a single clean TAM. | 中 | SM023, SM009, SM010, SM014, SM005 |
| CM041 | No public source in this evidence set isolates a clean 2026 SAM or SOM for short-reach coherent DSPs in AI data centers; published figures are overlapping optics, switching, transport, and pluggable lenses with different inclusion rules. | 中 | SM023, SM006, SM014, SM018 |
| CM042 | The main adoption constraints are reliability, manufacturing yield, serviceability, and standards maturity rather than weak AI bandwidth demand. | 中 | SM013, SM015, SM025, SM026 |
| CP001 | Celero positions its product as coherent DSP technology for AI optical connectivity and says it removes scale-across distance bottlenecks with much higher transmission efficiency. | 中 | SP001 |
| CP002 | Celero disclosed $140 million total funding, including a $100 million Series B led by CapitalG and prior seed and Series A funding led by Sutter Hill Ventures. | 中 | SP002 |
| CP003 | Celero says it was founded in 2024 by veterans from Marvell, Inphi, Broadcom, and ClariPhy and is headquartered in Irvine, California. | 中 | SP001, SP002 |
| CP004 | Marvell says copper links become inadequate beyond roughly 10 meters for AI data-center bandwidth and reach, positioning optical DSPs as the foundation from 800G toward 1.6T and beyond. | 中 | SP003 |
| CP005 | Marvell markets separate PAM4, coherent, and coherent-lite DSP families for intra-data-center, campus, and longer-distance data-center interconnect use cases. | 中 | SP003 |
| CP006 | Marvell said its 3nm Ara 1.6T platform was already shipping in mass volume to global customers and hyperscalers by March 2026. | 中 | SP004 |
| CP007 | Marvell expanded its 1.6T portfolio in 2026 with Ara T, Ara X, Petra, and Aquila M, indicating a multi-SKU response instead of a single flagship DSP. | 中 | SP004 |
| CP008 | Aquila M extends Marvell into coherent-lite campus links and adds integrated MACsec, giving Marvell a security and distance-range story that Celero does not yet disclose publicly. | 中 | SP004 |
| CP009 | Broadcom’s OFC 2025 portfolio spanned CPO, 200G/lane DSP and SerDes, 400G optics, and PCIe Gen6 over optics, showing a broad optical interconnect stack for AI clusters. | 中 | SP005 |
| CP010 | Broadcom’s Sian3 and Sian2M are purpose-built for 800G and 1.6T AI-cluster links over single-mode and short-reach multimode fiber, respectively. | 中 | SP006 |
| CP011 | Broadcom claims Sian3 cuts 1.6T optical-module power by more than 20% versus the prior generation, making power-per-bit a direct competitive battleground. | 中 | SP006 |
| CP012 | Broadcom says it has already deployed more than 50 million channels of 100G VCSELs and is shipping 200G EML and photodiode products in volume. | 中 | SP006 |
| CP013 | Broadcom told customers to contact sales for samples and pricing, which indicates commercial activity but no public list pricing at the DSP layer. | 中 | SP006 |
| CP014 | Celestial AI raised a $250 million Series C1 round in 2025, bringing total disclosed funding above $515 million. | 中 | SP007, SP008 |
| CP015 | Celestial AI says Photonic Fabric spans connectivity, switching, and packaging from within processor packages to servers across multiple racks. | 中 | SP007, SP008 |
| CP016 | Celestial AI says its platform is compatible with standard manufacturing and 2.5D packaging and that it already has deep engagements with multiple hyperscalers and silicon or packaging partners. | 中 | SP007 |
| CP017 | Ayar Labs markets an optical-I/O stack of TeraPHY plus SuperNova for GPUs and accelerators with 5x–10x higher bandwidth, 10x lower latency, and 4x–8x better power efficiency than pluggables plus electrical SerDes. | 中 | SP010, SP011 |
| CP018 | Ayar says its optical engines use standard package flows already used by major XPU and switch vendors and are production-ready for hyperscale deployments. | 中 | SP009, SP010 |
| CP019 | Ayar disclosed a 2026 $500 million Series E round, roughly $870 million total funding, and a $3.75 billion valuation, giving it a deeper capital base than Celero. | 中 | SP012 |
| CP020 | Ayar’s SuperNova is CW-WDM MSA compliant and states GR-468 reliability compliance, which are concrete trust signals for standards and component reliability. | 中 | SP011 |
| CP021 | Lightmatter positions Passage and Guide as a complete photonics roadmap spanning pluggable optics, near-package optics, and 3D co-packaged or interposer approaches. | 中 | SP013, SP014 |
| CP022 | Lightmatter’s public 2026 product materials emphasize early-access partners rather than general availability, suggesting it remains earlier in commercialization than Marvell or Broadcom. | 中 | SP014, SP015 |
| CP023 | Lightmatter disclosed a sampled Passage CPO chiplet delivering 1.6 Tbps per fiber and a Passage M1000 reference platform advertising 114 Tbps total bandwidth. | 中 | SP015, SP017 |
| CP024 | Lightmatter joined NVIDIA NVLink Fusion in 2026, giving it a distribution bridge into semi-custom hyperscaler and ASIC ecosystems. | 中 | SP016 |
| CP025 | Eliyan is adjacent rather than a like-for-like coherent-DSP rival because its NuLink and UMI products target chiplet and memory interconnect inside multi-die AI chips, not optical modules across fibers. | 中 | SP018, SP019 |
| CP026 | Eliyan closed a $60 million Series B after a $40 million Series A, which is materially smaller than the capital raised by Ayar or Celestial AI. | 中 | SP019 |
| CP027 | Intel’s OCI chiplet is designed for co-packaging with CPUs, GPUs, and IPUs, supports 4 Tbps bidirectional today, and has a roadmap to tens of terabits per device. | 中 | SP020 |
| CP028 | Intel says its silicon photonics platform has already shipped more than 8 million PICs and 32 million on-chip lasers, giving it stronger volume and reliability evidence than optical startups. | 中 | SP020 |
| CP029 | Coherent competes mainly as a vertically integrated transceiver and component supplier rather than a merchant DSP startup, and it highlights a 1.6T DR8 module using Marvell’s Ara DSP. | 中 | SP021, SP022 |
| CP030 | Coherent says its vertical integration reduces supply-chain uncertainty, which matters when module vendors choose partners for new optical ramps. | 中 | SP022 |
| CP031 | InnoLight markets 800G and 1.6T pluggable transceivers, representing the status-quo module path that can pressure DSP vendors on pricing and qualification timelines. | 中 | SP023 |
| CP032 | NVIDIA NVLink is a powerful substitute for some scale-up budgets because NVLink 6 offers 3.6 TB/s per GPU and 260 TB/s aggregate bandwidth in a 72-GPU rack-scale domain. | 中 | SP024 |
| CP033 | UCIe keeps internal-build and chiplet-based alternatives alive by promoting an open ecosystem for on-package innovation rather than a closed merchant-DSP stack. | 中 | SP025 |
| CP034 | SemiEngineering argues CPO is poised to replace pluggable optics for scale-out and eventually displace copper for scale-up once reliability and manufacturability are proven. | 中 | SP026 |
| CP035 | SemiEngineering identifies Broadcom and NVIDIA as early deployed CPO players while naming Ayar Labs, Intel Photonics, Marvell/Celestial, and Lightmatter as additional contenders, implying a dense field around Celero’s wedge. | 中 | SP026 |
| CP036 | SemiEngineering also notes copper likely persists in-rack before optical expands further, so substitute pressure is mixed by distance regime rather than binary. | 中 | SP026 |
| CP037 | Futurum describes a strategic contest between Broadcom’s IMDD roadmap and Marvell’s coherent path at 1.6T, with vendor choice likely influencing who owns the 3.2T upgrade path. | 中 | SP027 |
| CP038 | Futurum says Broadcom’s earlier sampling window could yield sticky design wins because transceiver qualification cycles are long and re-engineering is costly. | 中 | SP027 |
| CP039 | Futurum flags two adverse scenarios for merchant DSP vendors: NVIDIA developing more internal DSP silicon and CPO adoption shortening the useful life of pluggable-DSP transceivers. | 中 | SP027 |
| CP040 | LightCounting’s 2026 report profiles Broadcom, Marvell, Celero, Coherent, and multiple other PAM4 or coherent suppliers, confirming that Celero is entering an already mapped vendor landscape rather than a blank category. | 中 | SP028 |
| CP041 | LightCounting projects IC chipset sales for PAM4 and coherent DSP markets to surpass $20 billion by 2031, which signals growth but also attracts more capital and incumbents. | 中 | SP028 |
| CP042 | Celero’s founder pedigree overlaps with incumbent optical-DSP talent pools at Marvell, Broadcom, Inphi, and ClariPhy, which narrows execution-risk differentiation but also makes poaching and imitation plausible. | 中 | SP001, SP002 |
| CP043 | Across the public sources reviewed, neither Celero nor major rivals publish standard list prices for DSP silicon, so the buying motion remains design-win driven and opaque. | 中 | SP002, SP006, SP014 |
| CP044 | Public sources reviewed in this run do not identify named production customers for Celero, so actual deployment breadth remains an open diligence question. | 低 | |
| CI001 | Celero said it had raised a total of $140 million as of November 2025, including a $100 million Series B and $40 million across earlier seed and Series A rounds. | 高 | SI003, SI004, SI006 |
| CI002 | Celero develops coherent DSP solutions for AI infrastructure that are positioned as high-bandwidth, low-power optical connectivity products. | 高 | SI001, SI002, SI003 |
| CI003 | Celero said the new capital will be used to expand silicon development and commercialization across hyperscale and cloud customers. | 高 | SI003, SI005 |
| CI004 | Celero was founded in 2024 and is headquartered in Irvine, California, with design centers in Canada and Argentina. | 高 | SI002, SI003, SI005 |
| CI005 | Reuters reported that Celero is building a chip to link AI data centers across long distances by translating light over fiber into electrical signals used inside computing networks. | 中 | SI006 |
| CI006 | Celero has not publicly disclosed revenue, ARR, customer count, unit shipments, or backlog in the sources reviewed for this chapter. | 中 | SI001, SI002, SI003, SI005 |
| CI007 | Public evidence supports a hardware revenue model built around coherent DSP chip sales and related commercialization activity rather than recurring software subscriptions. | 中 | SI002, SI003, SI006 |
| CI008 | Comparable fabless ASIC models can recognize NRE revenue on milestone completion before full production shipments begin. | 中 | SI017 |
| CI009 | Marvell disclosed that certain ASIC engagements use a business model with significant NRE costs that customers pay based on milestones and that this model tends to have lower gross margins. | 中 | SI017 |
| CI010 | LightCounting forecast that coherent DSP average selling prices would remain about ten times higher than PAM4 DSP average selling prices because coherent products are more complex and lower volume. | 中 | SI013 |
| CI011 | LightCounting reported 2023 sales of about $0.8 billion for coherent DSP chipsets versus more than $1.1 billion for PAM4 chips. | 中 | SI013 |
| CI012 | PMarketResearch estimated the coherent optical DSP market at about $2.37 billion in 2025 and $2.85 billion in 2026. | 中 | SI021 |
| CI013 | Cignal AI forecast that 800ZRx pluggable ports would ship more than 200,000 units in 2026 and that 1600ZRx would arrive in 2027. | 中 | SI010 |
| CI014 | Cignal AI said the coherent pluggable market is expanding while higher baud rates require the latest generation process technologies and more complex DSP gate counts. | 中 | SI010, SI011 |
| CI015 | Silicon Analysts estimated total NRE of roughly $5 million to $30 million for 7nm-28nm programs and $30 million to $100 million or more for 3nm-5nm programs. | 中 | SI014 |
| CI016 | Silicon Analysts estimated leading-edge 3nm-5nm wafer costs of about $16,000 to $22,000 each and said those nodes are usually justified only at volumes above 100,000 units per year. | 中 | SI014 |
| CI017 | Silicon Analysts estimated that 7nm MPW shuttles cost about $100,000 to $500,000 versus $5 million to $10 million for a full-mask run. | 中 | SI014 |
| CI018 | Silicon Analysts described the typical startup path as MPW validation followed by engineering samples for qualification and then full-mask production. | 中 | SI014 |
| CI019 | Marvell said it can place firm supplier orders as far as 26 weeks ahead of customer delivery and may extend commitments to 52 weeks to secure capacity. | 中 | SI017 |
| CI020 | Marvell said it often maintains substantial inventories and capacity-reservation arrangements with foundries and substrate partners to support demand. | 中 | SI017 |
| CI021 | Marvell outsources integrated-circuit fabrication to foundries and product packaging and testing to external subcontractors. | 中 | SI017 |
| CI022 | Marvell said its fiscal 2026 gross margin increased by 9.7 percentage points year over year partly because higher revenue improved cost absorption. | 中 | SI017 |
| CI023 | Marvell reported fiscal 2026 gross margin of 75.5 percent, which is a mature-scale benchmark rather than a startup underwriting assumption. | 中 | SI017 |
| CI024 | Marvell said its ten largest customers represented 82 percent of fiscal 2026 revenue. | 中 | SI017 |
| CI025 | Marvell said fiscal 2026 working-capital outflow was driven primarily by increases in accounts receivable and inventories. | 中 | SI017 |
| CI026 | TrendForce forecast that worldwide shipments of optical transceivers at 800G and above would reach 24 million units in 2025 and nearly 63 million units in 2026. | 中 | SI019 |
| CI027 | TrendForce reported that Nvidia had secured key EML supplier capacity, extending lead times beyond 2027 and tightening the optical supply chain. | 中 | SI019 |
| CI028 | Adtek said large-scale CPO deployment is expected around 2028 and that packaging, testing, and reliability remain bottlenecks. | 中 | SI020 |
| CI029 | Adtek said even advanced prototypes still use copper within racks because of cost, reliability, and serviceability considerations. | 中 | SI020 |
| CI030 | NewMarketPitch said 2026 semiconductor funding was concentrated on companies removing visible AI bottlenecks such as interconnect, optical I/O, packaging, memory, and cooling. | 中 | SI022 |
| CI031 | NewMarketPitch said 80 semiconductor startups raised $8.4 billion in Q1 2026 and about $10.7 billion had been raised by June 2026. | 中 | SI022 |
| CI032 | CapitalG participation likely improves Celero’s strategic credibility with hyperscalers and cloud customers, but it does not substitute for public revenue or customer proof. | 中 | SI003, SI005, SI006 |
| CI033 | Celero appears pre-revenue or at most early-revenue from public evidence because sources discuss development, commercialization, and market entry without disclosing shipments or financial metrics. | 中 | SI003, SI008, SI009 |
| CI034 | For Celero, design-win conversion, qualification-cycle time, NRE recovery, and gross margin by speed grade are more decision-useful than ARR or SaaS payback metrics. | 中 | SI006, SI014, SI017 |
| CI035 | Revenue quality cannot be underwritten publicly because there is no disclosed mix between evaluation activity, NRE recovery, and production shipments. | 中 | SI003, SI006, SI017 |
| CI036 | Publicly disclosed use of proceeds is focused on silicon development and commercialization rather than debt reduction or acquisitions. | 高 | SI003, SI004, SI005 |
| CI037 | Capital adequacy cannot be proven from public sources because Celero has not disclosed cash on hand, monthly burn, runway months, or working-capital commitments. | 中 | SI003, SI005, SI006 |
| CI038 | Celero’s $140 million of disclosed funding should be sufficient for at least one sampling-and-qualification cycle, but it may not be sufficient for multiple leading-edge respins plus commercial ramp if production timing slips. | 中 | SI003, SI014 |
| CI039 | Optical Connections warned that new entrants, including future Chinese suppliers, could materially lower coherent DSP selling prices by 2027-2029. | 中 | SI013 |
| CI040 | PMarketResearch said data center interconnect represented the largest 2025 application share of the coherent optical DSP market at 44.9 percent. | 中 | SI021 |
| CI041 | Celero’s public positioning implies initial revenue will likely depend on a small number of hyperscalers, cloud operators, module vendors, or OEM partners rather than a broad diversified customer base. | 中 | SI003, SI006, SI024 |
| CI042 | There is no public evidence in the reviewed sources that Celero has debt financing or project-finance obligations, so the disclosed capital stack appears equity funded. | 中 | SI003, SI005, SI006 |
| CI043 | Optical Connections said the current forecast did not include coherent 1.6T Ethernet transceivers in 2024-2026, which tempers near-term volume assumptions for the highest-speed opportunity. | 中 | SI013 |
| CI044 | LightCounting said ZR and ZR+ shipments were expected to surpass on-board coherent transceivers in 2025, with Microsoft and Amazon driving 400ZR demand and Google and Meta primary consumers of 800ZR and ZR+ for metro and regional networks. | 中 | SI012 |
| CI045 | Celero’s homepage claims its coherent DSP technology enables 100x more efficient optical transmission, but the company provides no public methodology or customer validation for that figure in the reviewed sources. | 低 | SI001 |
| CI046 | Celero’s official materials emphasize cost advantages and power efficiency but provide no public gross margin, ASP, yield, utilization, or power-per-bit disclosures. | 中 | SI001, SI003 |
| CE001 | Celero publicly describes itself as a developer of advanced coherent DSP solutions for AI infrastructure. | 中 | SE002, SE011 |
| CE002 | Celero’s homepage positions the product around high-bandwidth, low-power optical connectivity and removing distance bottlenecks in AI infrastructure. | 中 | SE001, SE002 |
| CE003 | The customer workflow Celero describes is connectivity between AI accelerators, servers, clusters, and cloud data centers rather than a generic networking chip sale. | 中 | SE001, SE011, SE017 |
| CE004 | CapitalG says Celero is building a new class of coherent DSP platforms to address the networking bottleneck in AI-scale infrastructure. | 中 | SE012 |
| CE005 | CapitalG frames the technical gap as one between PAM4 that struggles at future 1.6T-plus data-center scales and legacy coherent designs that are too costly or power-hungry for dense data centers. | 中 | SE012 |
| CE006 | CapitalG describes the DSP as the chip inside an optical transceiver that converts electrical signals from GPUs or XPUs into optical signals for transmission over fiber. | 中 | SE012 |
| CE007 | Celero says it was founded in 2024, is headquartered in Irvine, California, and has design centers in Canada and Argentina. | 中 | SE002, SE013 |
| CE008 | Celero’s careers page lists active openings in ASIC design, verification, DSP-FEC, optical DSP, analog, PLL, SerDes, firmware, physical design, physical verification, and product engineering. | 中 | SE003 |
| CE009 | The breadth of technical roles implies Celero is building a full custom silicon-and-firmware program rather than only licensing coherent DSP algorithms. | 中 | SE003 |
| CE010 | The specific mix of PLL, SerDes, RTL, analog, and physical-design roles implies the product includes high-speed I/O and clocking blocks required for coherent-module integration. | 中 | SE003, SE020 |
| CE011 | Celero says its founders have shipped multiple generations of DSPs and networking IC solutions. | 中 | SE011, SE018 |
| CE012 | Investor and company materials consistently argue that Celero’s differentiation comes from rare coherent-DSP experience and system-level integration depth. | 中 | SE005, SE012, SE011 |
| CE013 | CapitalG wrote in late 2025 that Celero had built a roughly 100-person team and was growing quickly. | 中 | SE012 |
| CE014 | The careers page corroborates active organizational scaling across Irvine, Ottawa, and San Jose roles. | 中 | SE003 |
| CE015 | Celero’s November 2025 announcement says the new funding will accelerate development and deployment, which indicates an early commercial stage rather than broad public general availability. | 中 | SE011, SE016 |
| CE016 | Orange County Business Journal reports that Celero is developing a chip to help connect data centers across vast distances. | 中 | SE017 |
| CE017 | Orange County Business Journal describes Celero’s chip as helping convert light over fiber-optic cables into electrical signals to speed links between distant facilities. | 中 | SE017 |
| CE018 | ConvergeDigest says Celero plans to commercialize next-generation coherent engines for hyperscale and cloud customers. | 中 | SE014 |
| CE019 | OIF’s current-work page shows that 1600ZR, 1600ZR+, 1600CL, compute optics, C-CMIS, MACsec support, and related electrical interfaces are active coherent-optics workstreams. | 中 | SE020, SE023 |
| CE020 | OIF’s compute-optics work explicitly targets AI scale-up links over protocols including PCIe, NVLink, and UALink. | 中 | SE020 |
| CE021 | OIF says 1600ZR targets a multi-vendor coherent 1600 Gbit/s interface based on single-carrier DP-16QAM for up to 120 km data-center interconnect links. | 中 | SE020 |
| CE022 | OIF says 1600ZR+ uses O-FEC, DP-16QAM, and two digital subcarriers in a low-power DSP to support higher-performance long-reach modes. | 中 | SE020 |
| CE023 | The OIF 800ZR interoperability white paper says the application uses DP-16QAM at 118 Gbaud with an O-FEC-based structure. | 中 | SE021 |
| CE024 | The OIF 800ZR plugfest found interoperability combinations that required additional debugging and at least two missing or failed transmitter-receiver measurements, showing that coherent-module validation is still nontrivial. | 中 | SE021 |
| CE025 | Coherent says 800G ZR/ZR+ pluggables support direct integration into routers and Ethernet switches and can reduce the need for separate transport equipment. | 中 | SE022 |
| CE026 | Commercial 800G coherent pluggables already support more than 500 km in ZR mode and more than 1000 km in high-performance ZR+ modes, setting a real market baseline for future entrants. | 中 | SE022 |
| CE027 | OFC’s 2026 OIF session says 1600ZR, 1600ZR+, and 1600CL development involves explicit trade-offs in optical system performance, latency, power, and density. | 中 | SE023 |
| CE028 | Across the homepage and funding materials, Celero consistently differentiates around bandwidth efficiency, low power, interoperability, and scale across AI fabrics. | 中 | SE001, SE011 |
| CE029 | Celero says its coherent DSP approach includes breakthrough silicon and system-level integration. | 中 | SE011, SE018 |
| CE030 | CapitalG cites strong foundry relationships as part of Celero’s ability to bring a new coherent-DSP generation into production. | 中 | SE012 |
| CE031 | Celero’s public official surfaces do not expose named SKUs, a public datasheet library, or a live dedicated technology page, and the /technology URL currently returns 404. | 中 | SE001, SE006, SE010, SE024 |
| CE032 | The retained public company surfaces do not publish chip-level BER, latency, power, reach, or yield metrics for a Celero product. | 中 | SE006, SE010, SE011 |
| CE033 | Celero’s privacy policy discloses collection of contact and recruiting data, cookie use, fraud-prevention language, and reasonable physical and electronic safeguards. | 中 | SE007 |
| CE034 | No retained public Celero source advertises SOC 2, ISO 27001, a trust center, or a product-security assurance artifact. | 中 | SE007 |
| CE035 | Celero’s public terms do not publish hardware support SLAs, RMA terms, or field reliability commitments for silicon customers. | 中 | SE008 |
| CE036 | OIF’s current work includes CMIS-Security and MACsec support for 1.6T pluggable modules, which implies security requirements at the module layer are still evolving. | 中 | SE020 |
| CE037 | Celero’s distributed engineering footprint and product-engineering hiring imply multi-site validation and support workflows that will need strong release and qualification discipline. | 中 | SE002, SE003, SE014 |
| CE038 | Independent coverage in SiliconANGLE and Wanture presents Celero as using fresh funding to scale development, which partially answers maturity but not production-deployment depth. | 中 | SE015, SE019 |
| CE039 | No retained public source provides named production customers, deployment counts, yield data, or foundry-node disclosure for Celero’s coherent DSP products. | 中 | SE001, SE006, SE011, SE014, SE015 |
| CE040 | Celero’s public trust layer is policy-led and its public technical layer is recruiting-, standards-, and funding-led; direct product collateral remains thin. | 中 | SE003, SE007, SE008, SE020, SE024 |
| CU001 | Celero sells coherent DSP solutions for high-bandwidth, low-power optical connectivity in AI infrastructure. | 高 | SU001, SU002, SU006 |
| CU002 | Celero positions its technology around links between AI accelerators, servers, clusters, and cloud data centers rather than around end-user applications. | 中 | SU001, SU006 |
| CU003 | Celero public hiring and marketing language identifies hyperscalers as a direct early-adopter customer target. | 高 | SU003, SU010 |
| CU004 | Celero same marketing language identifies module makers as a parallel early design-win target. | 中 | SU010 |
| CU005 | Celero raised $140 million in total, including a $100 million Series B led by CapitalG. | 高 | SU006, SU007 |
| CU006 | CapitalG markets its Google advisor network, customer introductions, channel introductions, and co-selling support as part of its value-add to portfolio companies. | 高 | SU008, SU009 |
| CU007 | CapitalG described Celero as a roughly 100-person and quickly growing team at the time of its investment note. | 中 | SU007 |
| CU008 | None of the reviewed Celero official pages disclosed customer count, deployment count, utilization, or shipments. | 高 | SU001, SU004, SU005 |
| CU009 | Public customer proof on reviewed official surfaces is still mostly strategic signaling rather than production reference-account disclosure. | 中 | SU004, SU005, SU010 |
| CU010 | CapitalG and the broader Alphabet ecosystem are the strongest named external validators publicly attached to Celero. | 高 | SU006, SU007, SU008 |
| CU011 | CapitalG frames Celero opportunity around linking thousands of GPUs or XPUs across racks, rows, and entire buildings with minimal latency. | 中 | SU007, SU021 |
| CU012 | AWS describes AI infrastructure as a scale problem that moves from a single chip to millions of chips over a non-blocking network. | 中 | SU013 |
| CU013 | Google Compute markets large VM fleets, TPUs, dedicated servers, and placement controls that are consistent with hyperscale infrastructure operation. | 中 | SU014 |
| CU014 | CoreWeave public infrastructure profile confirms AI-cloud operators as a distinct buyer class with 49 data centers, over 1GW of active power, and clusters exceeding 100,000 GPUs. | 中 | SU015, SU016 |
| CU015 | CoreWeave says 9 of its 10 leading foundation-model-provider customers already leverage its AI infrastructure, indicating that AI-cloud operators can aggregate substantial downstream demand. | 中 | SU015 |
| CU016 | NVIDIA says AI factories already span tens of thousands of GPUs and soon millions, and require deterministic latency across multiple sites. | 高 | SU020, SU021 |
| CU017 | OIF 2025 AI signaling workshop included Google Cloud, Microsoft, Meta, and OpenAI, showing that Celero target buyer and user community is actively defining next-generation interconnect requirements. | 中 | SU024 |
| CU018 | OIF 800G coherent work is scoped to campus and data-center interconnect applications up to 800G aggregate bandwidth. | 中 | SU023 |
| CU019 | Lightwave reported that Molex and Prysmian signed a 10-year fiber supply agreement to support hyperscaler and AI-data-center buildouts. | 中 | SU018 |
| CU020 | The same Lightwave report says Prysmian is separately pursuing agreements with hyperscalers and data-center infrastructure providers. | 中 | SU018 |
| CU021 | DellOro commentary cited by Lightwave says AI builds and scale-across DCI are raising expected demand for high-end routers and aggregation switches through 2030. | 中 | SU019 |
| CU022 | Coherent markets AI-driven datacom growth, hyperscale data-center optics, and supply-chain support to the same end market Celero wants to serve. | 中 | SU025 |
| CU023 | Broadcom markets AI infrastructure, networking, and data-center connectivity at scale, reinforcing that Celero competes for budgets already courted by large incumbents. | 中 | SU026 |
| CU024 | Celero public customer map therefore likely spans direct hyperscaler accounts, AI-cloud operators, and module or OEM channels. | 中 | SU003, SU010, SU015, SU024 |
| CU025 | Within those accounts, buyers are infrastructure and network leaders, users are optics and system engineers, and payers are AI-network capex owners. | 中 | SU010, SU014, SU016 |
| CU026 | North America likely matters first because Celero is California-based, its lead investor is U.S.-based, and reviewed demand signals concentrate on North American hyperscaler and AI-data-center expansion. | 低 | SU002, SU015, SU019 |
| CU027 | Celero public adoption trajectory is visible mainly through funding, headcount scale, and commercial hiring rather than through customer metrics. | 中 | SU006, SU007, SU003, SU010 |
| CU028 | No reviewed source disclosed active accounts, chips shipped, wafers committed, or network utilization attributable to Celero. | 高 | SU001, SU004, SU005 |
| CU029 | The public evidence is more consistent with sampling, evaluation, or early design-in than with broad production deployment. | 中 | SU004, SU005, SU010 |
| CU030 | No named public production customer was disclosed on the reviewed official or partner pages. | 高 | SU001, SU005, SU007 |
| CU031 | Celero job posting language implies parallel pursuit of hyperscaler and module-maker design wins rather than a single-channel go-to-market. | 中 | SU010 |
| CU032 | CoreWeave says 68% of enterprises are already beyond experimentation and more than 40% run some form of agent in production. | 中 | SU017 |
| CU033 | CoreWeave says six in ten enterprises wait more than four months from buying GPU capacity to serving their first production request. | 中 | SU017 |
| CU034 | CoreWeave says a third of enterprises cite accelerator availability as their biggest barrier to expanding AI infrastructure, with supply, memory, storage, power, and cooling also constraining rollout. | 中 | SU017 |
| CU035 | Once a networking or optical component is qualified into a large AI architecture, requalification costs and interoperability requirements can make the design win relatively sticky. | 低 | SU018, SU021, SU023 |
| CU036 | Celero has not publicly disclosed NRR, GRR, churn, renewal rate, contract length, or customer satisfaction metrics. | 高 | SU001, SU004, SU005 |
| CU037 | Public retention analysis for Celero therefore depends on market structure and qualification burden rather than on measured cohorts. | 低 | SU023, SU024, SU018 |
| CU038 | A first qualified socket could expand from one deployment into more racks, buildings, campuses, or product generations inside the same account. | 中 | SU015, SU021, SU007 |
| CU039 | CapitalG stated ability to open customer and partner introductions could accelerate land-and-expand once Celero clears an initial qualification gate. | 中 | SU008, SU009 |
| CU040 | Early commercial concentration risk is likely high because the relevant buyer universe is small and one anchor hyperscaler or channel partner could dominate first revenue. | 中 | SU010, SU018, SU019 |
| CU041 | Channel dependence is likely material because module makers and OEMs can mediate qualification access, pricing leverage, and roadmap control. | 中 | SU010, SU024, SU025 |
| CU042 | Competitive pressure is intense because Broadcom, Coherent, NVIDIA, Molex-linked ecosystems, and other incumbents already market AI networking or optical solutions to the same customers. | 中 | SU018, SU025, SU026 |
| CU043 | Public evidence supports customer need and Celero buyer fit much more strongly than it supports realized adoption. | 中 | SU006, SU007, SU010, SU017 |
| CU044 | The key diligence asks are named design partners, qualification stage, sample volume, expected production timing, and top-account concentration. | 中 | SU008, SU010, SU017 |
| CR001 | Celero says it was founded in 2024, is headquartered in Irvine, California, and has design centers in Canada and Argentina. | 高 | SR002, SR005 |
| CR002 | Celero says its founders and engineering team come from Marvell, Inphi, Broadcom, and ClariPhy. | 高 | SR001, SR002, SR005 |
| CR003 | Celero announced total funding of $140 million, including a $100 million Series B and $40 million of earlier seed and Series A capital. | 高 | SR004, SR005 |
| CR004 | Celero positions its product as coherent DSP silicon for high-bandwidth, low-power optical connectivity across AI infrastructure. | 高 | SR001, SR002, SR005 |
| CR005 | Celero is still hiring across ASIC design, design verification, DSP, analog, PLL, physical design, firmware, recruiting, and product functions. | 中 | SR003 |
| CR006 | Celero’s public materials emphasize architecture, recruiting, and funding but do not disclose named production customers or shipped product milestones. | 中 | SR001, SR004, SR005 |
| CR007 | Marvell says AI infrastructure scale-across architectures emerge when single data centers hit power, cooling, and footprint limits. | 中 | SR021 |
| CR008 | Marvell says electrical signaling is insufficient for long-reach scale-across links and that optical connectivity with advanced DSP is required across tens to thousands of kilometers. | 中 | SR021 |
| CR009 | Broadcom says networking infrastructure is one of the largest power consumers in AI data centers after GPUs. | 中 | SR025 |
| CR010 | Broadcom says typical 800G pluggable transceivers consume about 16 watts per link versus about 5 watts for its CPO system, and 1.6T pluggables about 25 watts versus about 8 watts. | 中 | SR024 |
| CR011 | Marvell says coherent-lite DSPs are optimized for 2 km to 20 km campus data-center interconnects with lower power, cost, and latency than traditional coherent approaches. | 中 | SR020, SR022 |
| CR012 | The OCI consortium is led by AMD, Broadcom, Meta, Microsoft, NVIDIA, and OpenAI to create an open optical scale-up interconnect specification. | 中 | SR023 |
| CR013 | The OCI consortium says a standardized roadmap reduces integration risk and creates a multi-vendor supply chain for optical AI interconnects. | 中 | SR023 |
| CR014 | BIS administers the Export Administration Regulations in 15 CFR parts 730 to 774 for items subject to the EAR. | 高 | SR007, SR010 |
| CR015 | Section 734.3 of the EAR defines what items are and are not subject to BIS export jurisdiction. | 高 | SR007, SR011 |
| CR016 | BIS says the Entity List, Unverified List, and Military End-User controls can impose licensing constraints even when item classification or destination would not otherwise require a license. | 高 | SR008, SR012 |
| CR017 | BIS clarified in May 2026 that advanced computing items destined for entities headquartered in Country Group D:5 or Macau still require licenses even if those entities are located outside D:5 or Macau. | 高 | SR006, SR009 |
| CR018 | BIS states that this D:5 and Macau licensing requirement predates the AI Diffusion Rule and remains enforceable notwithstanding BIS’s 2025 non-enforcement statement on parts of that rule. | 中 | SR009 |
| CR019 | The October 2023 Federal Register rule implemented additional export controls on certain advanced computing items and semiconductor end uses. | 中 | SR013 |
| CR020 | NIST says CHIPS guardrails can prohibit recipients from materially expanding semiconductor capacity in countries of concern for 10 years and can trigger clawback of the award if violated. | 高 | SR014, SR017 |
| CR021 | The CHIPS commercial fabrication funding program targeted construction, expansion, or modernization projects with capital investment of at least $300 million. | 中 | SR016 |
| CR022 | The Defend Trade Secrets Act lets a trade-secret owner bring a civil action when the alleged trade secret relates to a product or service used in interstate or foreign commerce. | 高 | SR018, SR019 |
| CR023 | The Defend Trade Secrets Act permits courts, in extraordinary circumstances, to issue ex parte seizure orders to prevent propagation or dissemination of an allegedly misappropriated trade secret. | 中 | SR018 |
| CR024 | The Defend Trade Secrets Act defines trade secrets broadly to include technical, scientific, and engineering information and treats breach of a secrecy duty as an improper means of acquisition. | 中 | SR019 |
| CR025 | Celero’s public materials do not disclose employment agreements, invention-assignment terms, or freedom-to-operate opinions for founders and key engineers. | 中 | SR001, SR002, SR005 |
| CR026 | TSMC says expanded U.S. export controls and the January 2025 rules can delay or prohibit shipments of certain products using 16-nanometer-or-below processes to specified destinations. | 高 | SR009, SR031 |
| CR027 | TSMC says its Nanjing fab currently operates under an annual export license and there is no assurance that the license will continue or be renewed on time. | 中 | SR031 |
| CR028 | TSMC says export controls, sanctions, and bans on semiconductor sales to certain entities can change semiconductor supply chains and affect customer demand. | 中 | SR031 |
| CR029 | TSMC’s public materials and 2025 annual report show a manufacturing footprint concentrated in Taiwan science parks. | 高 | SR030, SR031 |
| CR030 | TSMC says semiconductor manufacturing facilities require substantial fixed investment and margins can decline when customer demand or capacity utilization falls. | 中 | SR031 |
| CR031 | TSMC lists earthquakes, cyberattacks, supply-chain disruption, sabotage, and geopolitical tensions as operational risks to fab performance. | 中 | SR031 |
| CR032 | Marvell warns that a few customers account for a significant portion of revenue, data-center sales concentration is rising, and gain or loss of key design wins matters materially. | 中 | SR032 |
| CR033 | Marvell warns that China trade restrictions and export-license needs can cause customers to develop their own solutions, vertically integrate, or buy third-party alternatives. | 中 | SR032 |
| CR034 | Marvell warns that it relies on third parties for production and that natural disasters or power outages at Taiwan and Pacific Rim manufacturing partners could disrupt supply. | 中 | SR032 |
| CR035 | Broadcom discloses that hyperscalers, frontier-model companies, OEMs, and system integrators are target customers for AI custom silicon, switching, and optical infrastructure. | 高 | SR028, SR033 |
| CR036 | Broadcom’s AI stack spans custom XPUs, Ethernet switching, PHYs, optical components, and rack-scale systems within and across AI data-center sites. | 高 | SR028, SR033 |
| CR037 | Broadcom markets a 102.4 Tb/s co-packaged-optics switch with 64 integrated 1.6 TbE ports and 3-nm silicon, highlighting how fast the incumbent roadmap is moving. | 中 | SR029 |
| CR038 | Broadcom says optical links and open, low-power networking are central to scaling AI clusters across multiple data centers and sites. | 高 | SR023, SR028 |
| CR039 | Celero’s hiring breadth shows bench-building in verification, physical design, firmware, analog, product, and recruiting is still underway in parallel. | 中 | SR003 |
| CR040 | Because Celero was founded in 2024 and public evidence is still marketing- and funding-centric, product-to-revenue timing remains uncertain. | 中 | SR002, SR004, SR005 |
| CR041 | A first-silicon miss or delayed qualification would likely consume capital and time before Celero has public proof of diversified revenue or shipped volume. | 中 | SR003, SR005, SR031 |
| CR042 | Power-budget underperformance could block adoption because network watts directly trade off against GPU power capacity in AI clusters. | 中 | SR024, SR025 |
| CR043 | Loss of TSMC access or a Taiwan disruption would threaten Celero’s supply, schedule, and financing because alternative advanced-node capacity is limited and geographically concentrated. | 中 | SR017, SR029, SR031 |
| CR044 | Failure to win or retain a small number of hyperscaler-scale design-ins could materially impair early revenue because adjacent vendors already warn that customer concentration and design wins dominate outcomes. | 中 | SR032, SR033 |
| CR045 | No public source reviewed here disclosed a Celero lawsuit, export-enforcement action, or IP dispute, so latent legal and compliance downside remains hard to underwrite from public evidence alone. | 低 | SR004, SR006, SR008, SR018 |
| CV001 | Celero announced that it has raised $140 million in total funding, including a $100 million Series B led by CapitalG. | 高 | SV002, SV003, SV004 |
| CV002 | Celero said its earlier seed and Series A rounds totaled $40 million and were led by Sutter Hill Ventures with participation from Valor Equity Partners, Atreides Management, and Maverick Silicon. | 高 | SV002, SV003 |
| CV003 | Celero positions its product as coherent DSP silicon for high-bandwidth, low-power optical connectivity across AI accelerator clusters and cloud data-center links. | 高 | SV001, SV002, SV004 |
| CV004 | Celero says it was founded in 2024 by former Marvell, Inphi, Broadcom, and ClariPhy executives. | 高 | SV002, SV004 |
| CV005 | None of the reviewed Celero company or funding announcements disclosed current revenue, gross margin, or shipment scale. | 中 | SV001, SV002, SV003 |
| CV006 | TrendForce projects the AI-focused optical transceiver market to expand from $16.5 billion in 2025 to $26 billion in 2026, or more than 57% year over year. | 中 | SV022 |
| CV007 | LightCounting separately estimates optics for AI clusters and CPO reached $16.5 billion in 2025 and will reach $26 billion in 2026, implying about 60% growth. | 中 | SV023 |
| CV008 | Marvell's public data points place FY2025/FY2026 revenue around $8.19 billion with 2026 trailing revenue above $8.7 billion. | 中 | SV008, SV011 |
| CV009 | Marvell's market capitalization in July 2026 sits around $170-174 billion. | 中 | SV009, SV010 |
| CV010 | Yahoo Finance shows Marvell at roughly 21.24x enterprise value to revenue as of July 2026. | 中 | SV009 |
| CV011 | Broadcom's market capitalization in July 2026 sits around $1.82-1.87 trillion. | 中 | SV014, SV015 |
| CV012 | Yahoo Finance shows Broadcom at roughly 25.34x enterprise value to revenue as of July 2026. | 中 | SV014 |
| CV013 | Yahoo Finance also shows Broadcom at roughly 21x forward P/E in July 2026, underscoring that investors pay for scale, profitability, and software mix rather than optics exposure alone. | 中 | SV014 |
| CV014 | Coherent's market capitalization is about $55.25 billion in July 2026. | 中 | SV017, SV018 |
| CV015 | Yahoo Finance shows Coherent at about 9.43x enterprise value to revenue and 8.55x price to sales in July 2026. | 中 | SV017 |
| CV016 | Trefis and Forbes both argue that the Broadcom-versus-Marvell valuation gap is mainly explained by Broadcom's larger scale, broader hyperscaler reach, and software economics. | 中 | SV019, SV020 |
| CV017 | AInvest presents a more adverse optic, arguing Marvell's optical thesis already assumes limited asymmetric upside when the stock trades near 31x sales. | 低 | SV021 |
| CV018 | Tech Company News reports Celestial AI's 2025 Series C1 closed at $255 million and approximately $2.5 billion post-money after an initial March 2025 company announcement of $250 million raised. | 高 | SV024, SV025 |
| CV019 | Lightmatter's October 2024 Series D raised $400 million at a $4.4 billion valuation and brought total capital raised to about $850 million. | 高 | SV026, SV027 |
| CV020 | Ayar Labs' March 2026 Series E raised $500 million at a $3.75 billion valuation and lifted total funding to about $870 million. | 高 | SV028, SV029 |
| CV021 | Celero's approximately $1 billion last-round mark sits well below the latest private valuations reported for Celestial AI, Lightmatter, and Ayar Labs. | 中 | SV002, SV018, SV019, SV020 |
| CV022 | Marvell's 2020 agreement to acquire Inphi for about $10 billion was explicitly framed as accelerating leadership in cloud, 5G, and electro-optics interconnect infrastructure. | 中 | SV030 |
| CV023 | AMD said the Xilinx combination created a high-performance and adaptive computing leader aimed at an approximately $135 billion market opportunity. | 中 | SV031 |
| CV024 | Intel said its approximately $2 billion acquisition of Habana Labs was meant to strengthen its data-center AI silicon portfolio and accelerate participation in a fast-growing AI chip market. | 中 | SV032 |
| CV025 | SEC EDGAR shows current 10-K filing trails for Marvell, Broadcom, and Coherent, giving the valuation chapter a primary filing anchor for all three public comparables. | 高 | SV007, SV012, SV016 |
| CV026 | The reviewed public-optics comparables span roughly 9.4x EV/revenue for Coherent, 21.2x for Marvell, and 25.3x for Broadcom. | 中 | SV009, SV014, SV017 |
| CV027 | Those public multiples cannot be transferred one-for-one to Celero because Broadcom and Marvell are scaled, disclosed, and already embedded in large hyperscaler programs, while Coherent is profitable public infrastructure rather than an undisclosed startup. | 中 | SV016, SV019, SV020 |
| CV028 | At a $1 billion equity value, Celero would imply about 20x revenue at $50 million of revenue and about 10x revenue at $100 million of revenue. | 中 | SV002, SV026 |
| CV029 | Because public sources do not disclose Celero revenue, the $1 billion mark reads primarily as a strategic scarcity premium rather than a revenue-proven valuation. | 中 | SV002, SV005, SV021 |
| CV030 | TrendForce notes the roadmap is moving toward low-power LPO and silicon photonics as buyers look to reduce the power and thermal burden of traditional high-power DSP-based architectures. | 中 | SV022 |
| CV031 | LightCounting expects supply shortages to ease by mid-2026 and warns that the optical market could see one or two flat quarters in late 2026 as the supply chain rebalances. | 中 | SV023 |
| CV032 | For Celero, a faster move toward linear or co-packaged optics could cap multiple expansion if DSP content compresses faster than coherent links scale into AI clusters. | 中 | SV003, SV022, SV023 |
| CV033 | A workable bull case assumes Celero wins two or three hyperscaler design slots and becomes a scarce DSP control point in 1.6T or AI-cluster optics, supporting $180-220 million of medium-term revenue and a premium exit multiple. | 中 | SV004, SV006, SV022 |
| CV034 | A base case assumes Celero lands one meaningful hyperscaler or networking design cycle, reaches roughly $75-100 million revenue, and deserves a lower but still premium 10-12x multiple. | 中 | SV002, SV003, SV026 |
| CV035 | A bear case assumes a 12-18 month commercialization delay, incumbent response, or architecture shift, resulting in less than $50 million revenue and a valuation below the 2025 round. | 中 | SV021, SV022, SV023 |
| CV036 | Probability-weighting the public multiple band and scenario revenue assumptions produces a base valuation range of about $0.8-1.2 billion and a bear range of about $0.4-0.7 billion. | 中 | SV026, SV028, SV031 |
| CV037 | Strategic precedent suggests buyers will pay up for interconnect control points when the asset shortens roadmap time or locks scarce photonics and accelerator talent. | 中 | SV030, SV031, SV032 |
| CV038 | The highest-impact diligence asks are current revenue, gross margin, production readiness, and named customer or design-win evidence because those are the variables that separate a strategic premium from speculative narrative. | 中 | SV001, SV002, SV005 |
| CV039 | Public sources do not disclose Celero's cap-table preferences, liquidation stack, or secondary pricing terms for the 2025 round. | 中 | SV002, SV003, SV005 |
| CV040 | Because Celero has raised only about $140 million so far, future manufacturing scale-up and go-to-market expansion could still require materially dilutive follow-on capital. | 中 | SV002, SV022 |
| CV041 | The chapter's recommendation is track rather than buy because the technology and market case are attractive but the current price already discounts successful commercialization. | 中 | SV002, SV026, SV029 |
| CV042 | Confidence is medium because funding history and market demand are well corroborated, but the underwriting anchors of revenue, customer concentration, and margin remain undisclosed. | 中 | SV002, SV005, SV022 |
| CV043 | Risk rating is high because execution risk, architecture risk, and dilution risk can each compress value before Celero reaches public-company scale. | 中 | SV021, SV022, SV023 |
| CV044 | The valuation stance is stretched rather than obviously absurd: private optical peers are richer, but Celero lacks their publicly evidenced scale and capital depth. | 中 | SV018, SV019, SV020, SV029 |
| CV045 | Evidence that would upgrade the stance would include a verified revenue run rate above roughly $100 million, named hyperscaler sockets, or production proof that Celero's power and bandwidth claims hold in deployment. | 中 | SV003, SV022, SV023 |
| CV046 | Evidence that would downgrade the stance would include delayed tape-out, loss of investor support, weak early customer conversion, or a market shift away from DSP-heavy optics. | 中 | SV021, SV022, SV023 |
| CV047 | Return monitoring should center on named design wins, quarterly revenue conversion, gross margin trajectory, market growth for AI optics, and burn relative to the $140 million funding base. | 中 | SV002, SV022, SV023 |
| CV048 | Relative to a $1 billion entry mark, upside to roughly $2.0-2.5 billion needs both market capture and a still-premium multiple, while downside to roughly $0.5 billion can occur on one missed adoption cycle. | 低 | SV018, SV020 |
| 编号 | 出版方 | 标题 | 引文 |
|---|---|---|---|
| SO001 | Celero Communications | Celero Inc. | Celero’s coherent DSP technology eliminates distance bottlenecks — enabling 100x more efficient optical transmission for scale-across AI infrastructure. |
| SO002 | Celero Communications | About - Celero Inc. | Founded in 2024 by industry leaders from Marvell, Inphi, Broadcom, and ClariPhy... The company is headquartered in Irvine, California, with design centers in Canada and Argentina. |
| SO003 | Celero Communications | Careers - Celero Inc. | Open Positions ... Ottawa, Canada ... Irvine, CA ... San Jose, CA. |
| SO004 | Celero Communications | News - Celero Inc. | Latest News ... Reuters ... Celero raises $140 million ... Press Release ... Celero Communications Secures $140 Million to Accelerate AI Infrastructure. |
| SO005 | Celero Communications | Celero Communications Secures $140 Million to Accelerate AI Infrastructure - Celero Inc. | Celero Communications... has raised a total of $140 million in funding. This includes a recent $100 million Series B investment led by CapitalG... and earlier Series A and seed rounds totaling $40 million led by Sutter Hill Ventures... |
| SO006 | CapitalG | Celero Communications | Nariman Yousefi, Chief Executive Officer... and Oscar Agazzi, Chief Technology Officer... The company is headquartered in Irvine, California, with design centers in Canada and Argentina. |
| SO007 | Business Wire | Celero Communications Secures $140 Million to Accelerate AI Infrastructure | James Luo, General Partner at CapitalG and member of the Celero board of directors... Stefan Dyckerhoff, Managing Director at Sutter Hill Ventures and Celero board member. |
| SO008 | Reuters / Yahoo Finance | Celero raises $140 million from Alphabet's CapitalG fund, others for AI networking chip | The funding announced on Monday includes a $100 million Series B round led by CapitalG... and a previously undisclosed earlier $40 million round led by Sutter Hill Ventures... |
| SO009 | SiliconANGLE | Celero Communications raises $140M to scale coherent DSP technology for AI infrastructure | The $140 million in new funding came across two rounds: a Series A round led by Sutter Hill Ventures... and Series B round that was led by Capital G LP. |
| SO010 | TechStory | Celero Communications Raises $140 Mn to Build the Next Frontier of AI Connectivity | CapitalG’s participation — and its new board seat — is a strong signal. |
| SO011 | Pulse 2.0 | Celero Communications: $140 Million Raised To Expand AI Infrastructure Connectivity | James Luo, General Partner, CapitalG and Celero Board Member... Stefan Dyckerhoff, Managing Director, Sutter Hill Ventures and Celero Board Member. |
| SO012 | Communications Today | Celero Communications raises $140M | Celero Communications has raised $140 million to accelerate AI infrastructure. |
| SO013 | Intelligence360 | Celero Communications Secures $140 Million to Accelerate AI Infrastructure | Celero Communications... has raised a total of $140 million in funding. |
| SO014 | Markets Financial Content | Celero Communications Secures $140 Million to Accelerate AI Infrastructure | This includes a recent $100 million Series B investment led by CapitalG... and earlier Series A and seed rounds totaling $40 million led by Sutter Hill Ventures. |
| SO015 | Venture Capital Access Online | Celero Communications Secures $140 Million to Accelerate AI Infrastructure | Founded in 2024... The company is headquartered in Irvine, California, with design centers in Canada and Argentina. |
| SO016 | FinancialContent | Celero Communications Secures $140 Million to Accelerate AI Infrastructure | James Luo, General Partner at CapitalG and member of the Celero board of directors... |
| SO017 | CB Insights | Celero Communications - Products, Competitors, Financials, Employees, Headquarters Locations | Celero Communications was founded in 2024... Stage: Series B | Alive... Total Raised: $138.4M... Headquarters: 5281 California Avenue Suite 200, Irvine, California, 92617, United States. |
| SO018 | Seedtable | Celero Communications Raises 140.0M USD in Series B Funding | Seedtable | Celero Communications is an Irvine, California-based company founded by Nariman Yousefi. |
| SO019 | U.S. Securities and Exchange Commission | EDGAR Filing Documents for 0002019351-24-000002 | Filing Date 2024-12-26 ... Form D - Notice of Exempt Offering of Securities. |
| SO020 | U.S. Securities and Exchange Commission | Celero Communications, Inc. Form D primary XML | totalAmountSold 29999906 ... totalNumberAlreadyInvested 40 ... related persons: Nariman Yousefi, Oscar Agazzi, Stefan Dyckerhoff, Robert Abbott. |
| SO021 | BizProfile | Celero Communications, Inc. Irvine, CA - filing information | Officially filed on February 20, 2024... principal and mailing address ... 5281 California Ave., Suite 200, Irvine, CA 92617... Avijit Roy... Chief Financial Officer. |
| SO022 | Intelligence360 | Celero Communications has filed a notice of an exempt offering of securities to raise $99,999,994.00 in New Funding. | According to filings with the U.S. Securities and Exchange Commission, Celero Communications, Inc. is raising up to $99,999,994.00 in new funding. |
| SO023 | Yahoo Finance / Benzinga | Google Parent Alphabet Drops $100 Million On Celero: The $140 Million Startup Ending AI Data Center Bottlenecks | The startup develops coherent digital signal processor platforms that address bandwidth constraints in AI data centers... The company opened for business in 2024 with headquarters in California and design facilities in Canada and Argentina. |
| SO024 | Tech in Asia | US-based Celero raises $140m led by Alphabet's CapitalG | Celero enters a crowded 800G coherent DSP field with unclear edge... Missing specs make the matchup a guess... Celero raised $140M, yet gave no tapeout status or partnerships like working 800G systems at OFC 2025. |
| SO025 | FormDs.com | Celero Communications, Inc. - fund raising filing | Address: 5281 California Ave., Suite 200, Irvine, CA, 92617 |
| SM001 | Celero Communications | Celero Inc. | |
| SM002 | Celero Communications | About - Celero Inc. | |
| SM003 | Celero Communications | Celero Communications Secures $140 Million to Accelerate AI Infrastructure | |
| SM004 | CapitalG | Celero Communications | |
| SM005 | LightCounting | AI Capex Flows Down the Supply Chain to DSP Vendors | In 2025, hyperscalers' massive investments in AI infrastructure caused 800G PAM4 chipset shipments to nearly triple. |
| SM006 | Dell'Oro Group | Data Center Networking in 2025–2026: Milestones and Opportunities Amid Supply Risk | |
| SM007 | Converge Digest | Dell'Oro: Data Center Switch Sales in AI Back-End Networks | |
| SM008 | Converge Digest | Dell’Oro: Optical Transport Market to Surge 16% in 2026 on AI Data Center Demand | |
| SM009 | PR Newswire | Optical Transport Market Grew to $16 Billion in 2025, According to Dell'Oro Group | |
| SM010 | RCR Wireless | Optical transport bounces back, driven by AI interconnect | |
| SM011 | IDTechEx | From Copper to CPO: The Next Shift in AI Interconnects | |
| SM012 | SemiEngineering | All AI Data Center Interconnects Will Be Optical Within 5 Years | |
| SM013 | ADTEK | AI Data Center Interconnect 2026: CPO, Optical Interconnect and Deployment Challenges | |
| SM014 | Cignal AI | Tracking the Coherent DSP Supply Chain - 2026 | |
| SM015 | NVIDIA | NVIDIA Silicon Photonics for Agentic AI Networking | |
| SM016 | NVIDIA Technical Blog | Scaling Power-Efficient AI Factories with NVIDIA Spectrum-X Ethernet Photonics | |
| SM017 | Broadcom | Broadcom Showcases Industry-Leading Solutions for Scaling AI Infrastructure at OFC 2026 | |
| SM018 | Cisco | Data Center Interconnect with Cisco Coherent Pluggable Optics Solution Overview | |
| SM019 | Acacia | Coherent Technology Market Leadership at OFC 2026 | |
| SM020 | Cisco Live | BRKOPT-2699 | |
| SM021 | Anycom Fiberoptics | LC: AI Capex flows to downstream DSP suppliers | |
| SM022 | Tarluz | Rising AI Capex Drives Downstream Demand for DSP Vendors | |
| SM023 | DQIndia | From $23 billion to $112 billion in six years, AI is driving a new growth cycle in optical connectivity | |
| SM024 | TMCnet | NVIDIA Announces Spectrum-X Photonics, Co-Packaged Optics Networking Switches to Scale AI Factories to Millions of GPUs | |
| SM025 | CNBC | Nvidia is investing billions into this emerging technology that could change the AI industry | Copper is the main connectivity standard today because of its lower cost and high reliability, while broad photonics adoption is still early. |
| SM026 | Gazettabyte | Cisco's photonic plans for coherent, co-packaged optics, and pluggables | In a 100,000-GPU cluster, even with optics having a five-year mean-time-to-failure, the first failure will occur within 26 minutes. |
| SP001 | Celero Communications | Celero Inc. | Celero’s coherent DSP technology eliminates distance bottlenecks — enabling 100x more efficient optical transmission for scale-across AI infrastructure. |
| SP002 | Celero Communications | Celero Communications Secures $140 Million to Accelerate AI Infrastructure | Celero Communications... has raised a total of $140 million in funding. |
| SP003 | Marvell Technology | The evolution of AI interconnects | |
| SP004 | Marvell Technology | Marvell Ushers In the 1.6T Era with Expanded Optical DSP Platform Portfolio, Redefining AI Data Center End-to-End Connectivity | Now shipping in mass volume to global customers, Ara is enabling the world’s hyperscalers and cloud providers to deploy 1.6T pluggable connectivity for AI data centers. |
| SP005 | Broadcom | Broadcom Advances Optical Connectivity for AI Infrastructure with Industry-Leading Solutions at OFC 2025 | |
| SP006 | Broadcom | Broadcom Extends 200G/lane DSP PHY Leadership for Next-Generation AI Infrastructure | Broadcom’s new Sian3 and Sian2M DSPs... provide unprecedented levels of power efficiency and cost optimization for next-generation AI infrastructure. |
| SP007 | Business Wire | Celestial AI Secures $250 Million Funding to Revolutionize AI Infrastructure with Its Photonic Fabric | ...raised $250 million in its Series C1 funding round... bringing the total capital raised to date to more than $515 million. |
| SP008 | Data Center Dynamics | Optical interconnect startup Celestial AI raises $250m | |
| SP009 | Ayar Labs | Ayar Labs: AI Scale-up Beyond the Rack | Ayar Labs’ co-packaged optics (CPO) connect thousands of GPUs across racks into a single unified cluster via optical fabric. |
| SP010 | Ayar Labs | TeraPHY Optical I/O Chiplet | Optical I/O delivers 5x–10x higher bandwidth at 10x lower latency and 4x–8x more power efficiency compared to traditional interconnects. |
| SP011 | Ayar Labs | SuperNova light source | Ayar Labs has made the next big leap... the industry’s first CW-WDM MSA-compliant 16 wavelength light source. |
| SP012 | Converge Digest | Ayar Labs Raises $500M to Scale CPO | The financing brings Ayar Labs’ total funding to approximately $870 million and values the company at $3.75 billion. |
| SP013 | Lightmatter | Lightmatter home | |
| SP014 | Lightmatter | 3D Photonics for AI Applications | Passage | Passage is available for early access partners. |
| SP015 | Lightmatter | Lightmatter Achieves Record 1.6 Tbps Per Fiber to Accelerate AI Optical Interconnect | ...successful sampling of Passage Co-Packaged Optics chiplet, demonstrating a record-breaking 1.6 Tbps throughput per fiber. |
| SP016 | Lightmatter | Lightmatter Joins NVIDIA NVLink Fusion and Powers Next-Generation AI Infrastructure with Photonic Interconnects | ...creating a unified platform for semi-custom AI factories while reducing fiber and connector requirements by 50%. |
| SP017 | Lightmatter | Passage M-Series Photonic Superchip | The Passage M1000 3D Photonic Superchip reference platform demonstrates record-breaking 114 Tbps total bandwidth. |
| SP018 | Eliyan | The Ultimate Chiplet Interconnect | |
| SP019 | Eliyan | Chiplet Interconnect Pioneer Eliyan Closes $60 Million Series B Funding Round | ...announced the closing of its latest funding round, totaling $60 million. |
| SP020 | Intel | Intel Silicon Photonics | Our field-proven Intel Silicon Photonics platform has already shipped more than 8 million PICs with over 32 million on-chip lasers. |
| SP021 | Coherent | Global Leader in Materials, Networking, and Lasers | |
| SP022 | Coherent | Networking | Coherent | From silicon photonics... Coherent is a vertically integrated supplier alleviating supply chain uncertainty to better support our customers. |
| SP023 | InnoLight | The World’s Leading Developer and Manufacturer of High-speed Optical Transceivers | |
| SP024 | NVIDIA | NVLink and NVLink Switch for Advanced Multi-GPU Communication | The sixth-generation NVLink enables 3.6 TB/s of bandwidth per GPU for the NVIDIA Rubin platform. |
| SP025 | UCIe Consortium | Home | UCIe Consortium | Building an open ecosystem of chiplets for on-package innovations. |
| SP026 | SemiEngineering | All AI Data Center Interconnects Will Be Optical Within 5 Years | CPO will replace pluggable transceivers for scale out; CPO will enable all links within the rack to move from copper to optical for scale up. |
| SP027 | Futurum Group | Broadcom’s DSP Launch Intensifies the AI Optics Race with Marvell | The most immediately consequential detail... is that the BCM83640 is already sampling to early access customers and partners. |
| SP028 | LightCounting | PAM4 and Coherent DSPs | IC chipset sales will surpass $20 billion in 2031. |
| SI001 | Celero Communications | Celero homepage | |
| SI002 | Celero Communications | About Celero | |
| SI003 | Celero Communications | Celero Communications secures $140 million to accelerate AI infrastructure | The company said it has raised a total of $140 million, including a recent $100 million Series B, and that the new capital will support silicon development and commercialization across hyperscale and cloud customers. |
| SI004 | Business Wire | Celero Communications secures $140 million to accelerate AI infrastructure | |
| SI005 | CapitalG | CapitalG portfolio profile: Celero Communications | |
| SI006 | Communications Today / Reuters | Celero Communications raises $140M | Celero plans to use new algorithms embedded in its chips to process light over existing fiber-optic cables faster and using less power. |
| SI007 | Converge Digest | Celero raises $140M to scale coherent DSPs for AI networks | |
| SI008 | Benzinga | Google Parent Alphabet drops $100 million on Celero | |
| SI009 | TechStory | Celero Communications raises $140 Mn to build the next frontier of AI connectivity | |
| SI010 | Cignal AI | Tracking the coherent DSP supply chain - 2026 | |
| SI011 | Cignal AI | Tracking the coherent DSP supply chain - 2025 | |
| SI012 | LightCounting | PAM4 and Coherent DSPs update | |
| SI013 | Optical Connections News | Coherent DSPs could get cheaper – report | The forecast assumes average selling prices of coherent DSPs will remain 10 times higher than PAM4 chips, but new entrants could push prices down materially. |
| SI014 | Silicon Analysts | How much does a tapeout cost? A practical guide for fabless startups | |
| SI015 | Deloitte | 2026 Semiconductor Industry Outlook | |
| SI016 | GSA | Global Semiconductor Financial Tracker | |
| SI017 | Marvell Technology | Marvell Technology 2026 Annual Report (10-K) | Marvell said it places firm orders up to 26 weeks ahead, often maintains substantial inventories, outsources fabrication and packaging/test, and that ASIC models can include significant NRE costs. |
| SI018 | U.S. Securities and Exchange Commission | NVIDIA 2026 Q1 10-Q landing page | |
| SI019 | TrendForce | AI data centers ignite a laser shortage wave | TrendForce said Nvidia secured key EML supplier capacity, extending lead times beyond 2027 and tightening the optical supply chain. |
| SI020 | Adtek Fiber | AI data center interconnect 2026: CPO, optical interconnect and deployment challenges | |
| SI021 | PMarketResearch | Worldwide coherent optical DSPs market research | |
| SI022 | NewMarketPitch | How is the funding in the semiconductor industry now? | |
| SI023 | TechInsights | Market Model: Semiconductor Company Quarterly Financials IDM, Fabless, Foundry, OSAT 2025 Q2 | |
| SI024 | SiliconANGLE | Celero Communications raises $140M to scale coherent DSP technology for AI infrastructure | |
| SI025 | Ken Research | Fabless semiconductor market: future risks and opportunity map | |
| SE001 | Celero Inc. | Celero Inc. homepage | |
| SE002 | Celero Inc. | About - Celero Inc. | |
| SE003 | Celero Inc. | Careers - Celero Inc. | |
| SE004 | Celero Inc. | Contact Us - Celero Inc. | |
| SE005 | Celero Inc. | Investors - Celero Inc. | |
| SE006 | Celero Inc. | News - Celero Inc. | |
| SE007 | Celero Communications, Inc. | Privacy Policy - Celero Inc. | |
| SE008 | Celero Inc. | Terms and Conditions - Celero Inc. | |
| SE009 | Celero Inc. | Cookie Policy - Celero Inc. | |
| SE010 | Celero Inc. | Blog – Celero Inc. | |
| SE011 | Celero Communications | Celero Communications Secures $140 Million to Accelerate AI Infrastructure | |
| SE012 | CapitalG | Celero: Bringing Coherence to AI-Scale Networking | |
| SE013 | CapitalG | Celero Communications portfolio page | |
| SE014 | Converge Digest | Celero Raises $140M to Scale Coherent DSPs for AI Networks | |
| SE015 | SiliconANGLE | Celero Communications raises $140M to scale coherent DSP technology for AI infrastructure | |
| SE016 | Pulse 2.0 | Celero Communications: $140 Million Raised To Expand AI Infrastructure Connectivity | |
| SE017 | Orange County Business Journal | Celero Gets $140M to Speed Data Center Connections | |
| SE018 | iConnect007 | Celero Communications Secures $140 Million to Accelerate AI Infrastructure | |
| SE019 | Wanture | Celero raises $140M to power AI with light | |
| SE020 | OIF | Current OIF Work | |
| SE021 | OIF | OIF 800ZR Interoperability White Paper OFC 2025 Plugfest | |
| SE022 | Coherent Corp. | Coherent Announces General Availability of 800G ZR/ZR+ QSFP-DD Transceiver | |
| SE023 | OFC Conference | OIF Presents 800ZR/LR and 1600ZR/ZR+/CL - Changing the Game...Again | |
| SE024 | Celero Inc. | Technology page request returned 404 | |
| SE025 | Business Wire | Celero Communications Secures $140 Million to Accelerate AI Infrastructure | |
| SU001 | Celero Inc. | Celero Inc. | Celero’s coherent DSP technology eliminates distance bottlenecks — enabling 100x more efficient optical transmission for scale-across AI infrastructure. |
| SU002 | Celero Inc. | About - Celero Inc. | Celero Communications develops advanced coherent DSP solutions that enable high-bandwidth, low-power optical connectivity for AI infrastructure. |
| SU003 | Celero Inc. | Careers - Celero Inc. | |
| SU004 | Celero Inc. | Investors - Celero Inc. | |
| SU005 | Celero Inc. | News - Celero Inc. | We work with partners, customers and engineers who share our vision for faster, smarter AI infrastructure. |
| SU006 | Celero Inc. | Celero Communications Secures $140 Million to Accelerate AI Infrastructure | This includes a recent $100 million Series B investment led by CapitalG. |
| SU007 | CapitalG | Celero: Bringing Coherence to AI-Scale Networking | We’re thrilled to announce our lead investment in Celero. |
| SU008 | CapitalG | CapitalG is Alphabet’s independent growth fund. | Google's network of Global 2000 IT decision makers deliver highly qualified, warm introductions to potential channel partners and multimillion-dollar customer accounts. |
| SU009 | CapitalG | About us - CapitalG | Typical engagements include business model development, qualified customer introductions, executive or technical recruiting support, security reviews, marketing channel revamps, and leadership and engineering training. |
| SU010 | Greenhouse / Celero Communications | Marketing Director | Partner closely with executive leadership and sales to identify, pitch, and secure critical early-stage design wins with Hyperscalers and module makers. |
| SU011 | Greenhouse / Celero Communications | Principal Product Engineer | |
| SU012 | Greenhouse / Celero Communications | ASIC Design Engineer | |
| SU013 | Amazon Web Services | AWS Trainium | Trainium scales from a single chip to millions. The data center is now the new AI accelerator. |
| SU014 | Google Cloud | Compute Engine | |
| SU015 | CoreWeave | AI Data Centers | CoreWeave | 9 of the 10 leading foundation model providers are leveraging CoreWeave to stay at the frontier of AI. |
| SU016 | CoreWeave | Networking Services That Enhance Performance | CoreWeave | Create 300k+ GPU megaclusters via interconnected 100,000-strong GPU superclusters to unlock a new scale of compute. |
| SU017 | CoreWeave | Why Inference Is the Defining Layer of AI | Six in ten enterprises wait more than four months from buying GPU capacity to serving their first production request. |
| SU018 | Lightwave | Molex and Prysmian enter fiber AI data center supply pact | Access to expanded fiber cable capacity will enable hyperscalers and data center architects to secure supply through Molex. |
| SU019 | Lightwave | AI builds to drive router, aggregation switch market to $19B by 2030 | The largest upward revision is for Core Routers, driven by their use in scale-across DCI. |
| SU020 | NVIDIA | NVIDIA Data Centers for the Era of AI Reasoning | |
| SU021 | NVIDIA | Accelerated Networking Solutions from NVIDIA | AI factories now span tens of thousands of GPUs—and soon, millions—operating as a single distributed computing engine. |
| SU022 | OIF | OIF | |
| SU023 | OIF | 800G Coherent – OIF | In scope for the 800G Coherent project is to define interoperable 800G coherent line specifications for campus and DCI applications. |
| SU024 | OIF | 448Gbps Signaling for AI Workshop – April 15-16, 2025 – OIF | Google Cloud, Microsoft, Meta Platforms, and OpenAI all participated in OIF's 448Gbps Signaling for AI Workshop. |
| SU025 | Coherent | Global Leader in Materials, Networking, and Lasers | Coherent | The rapid growth in AI/ML is driving a demand for faster datacom enabled by innovative Coherent solutions. |
| SU026 | Broadcom | Broadcom Inc. | Connecting Everything | The best AI networking, connectivity solutions and custom XPUs. |
| SR001 | Celero Inc. | Celero Inc. | |
| SR002 | Celero Inc. | About - Celero Inc. | |
| SR003 | Celero Inc. | Careers - Celero Inc. | |
| SR004 | Celero Inc. | News - Celero Inc. | |
| SR005 | Celero Inc. | Celero Communications Secures $140 Million to Accelerate AI Infrastructure | |
| SR006 | Bureau of Industry and Security | Homepage | Bureau of Industry and Security | |
| SR007 | Bureau of Industry and Security | Determine what is subject to the EAR | |
| SR008 | Bureau of Industry and Security | Guidance on end-user and end-use controls and U.S. person controls | |
| SR009 | Bureau of Industry and Security | Guidance Regarding Enforcement of License Requirements for Advanced Computing Items for Entities Headquartered in Country Group D:5 and Macau | |
| SR010 | Electronic Code of Federal Regulations | 15 CFR Part 730 -- General Information | |
| SR011 | Electronic Code of Federal Regulations | 15 CFR Part 734 -- Scope of the Export Administration Regulations | |
| SR012 | Electronic Code of Federal Regulations | 15 CFR Part 744 -- Control Policy: End-User and End-Use Based | |
| SR013 | Federal Register | Implementation of Additional Export Controls: Certain Advanced Computing Items; Supercomputer and Semiconductor End Use; Updates and Corrections | |
| SR014 | Federal Register | Preventing the Improper Use of CHIPS Act Funding | |
| SR015 | NIST CHIPS for America | CHIPS FOR AMERICA | |
| SR016 | NIST CHIPS for America | Notice of Funding Opportunity: Commercial Fabrication Facilities | |
| SR017 | NIST CHIPS for America | National Security | |
| SR018 | Cornell Legal Information Institute | 18 U.S. Code § 1836 - Civil proceedings | |
| SR019 | Cornell Legal Information Institute | 18 U.S. Code § 1839 - Definitions | |
| SR020 | Marvell Technology | Marvell Unveils Industry’s First Coherent-lite 1.6 Tbps O-band-optimized DSP for Data Center Campus Connectivity | |
| SR021 | Marvell Technology | Scale Across Using Coherent ZR and ZR+ Explained | |
| SR022 | Marvell Technology | Coherent-lite Optical DSPs | Enabling intra-campus optical connectivity for AI cloud data centers | |
| SR023 | Broadcom | Optical Scale-up Consortium Established to Create an Open Specification for AI Infrastructure Led by Founding Members AMD, Broadcom, Meta, Microsoft, NVIDIA and OpenAI | |
| SR024 | Broadcom | Optical Interconnects: enabling large-scale AI clusters | |
| SR025 | Broadcom | Tomahawk® 5: Revolutionizing Power Efficiency for AI Infrastructure | |
| SR026 | Broadcom | How to scale the optical interconnect using Co-Packaged Optics (CPO) | |
| SR027 | Broadcom | OFC 2025: Broadcom shines a light on the future of AI and optical networking | |
| SR028 | Broadcom | AI Clusters | AI Servers | AI Infrastructure | |
| SR029 | Broadcom | 102.4 Tb/s | Co-Packaged Optics Switch | AI/ML | |
| SR030 | Taiwan Semiconductor Manufacturing Company | Annual Reports - Taiwan Semiconductor Manufacturing Company Limited | |
| SR031 | Taiwan Semiconductor Manufacturing Company | TSMC 2025 Annual Report / Form 20-F | |
| SR032 | Marvell Technology | Marvell Technology Form 10-K for fiscal 2026 | |
| SR033 | Broadcom | Broadcom Form 10-K for fiscal 2025 | |
| SV001 | Celero Communications | Celero Inc. | Celero’s coherent DSP technology eliminates distance bottlenecks — enabling 100x more efficient optical transmission for scaling AI infrastructure. |
| SV002 | Business Wire | Celero Communications Secures $140 Million to Accelerate AI Infrastructure | This includes a recent $100 million Series B investment led by CapitalG ... and earlier Series A and seed rounds totaling $40 million led by Sutter Hill Ventures. |
| SV003 | Yahoo Finance | Google Parent Alphabet Drops $100 Million On Celero: The $140 Million Startup Ending AI Data Center Bottlenecks | |
| SV004 | SiliconANGLE | Celero Communications raises $140M to scale coherent DSP technology for AI infrastructure | |
| SV005 | Market Analysis | Celero Communications Secures $140M to Push the Optical Frontier of AI Infrastructure | |
| SV006 | Semiconductor Engineering | Startup Funding: Q4 2025 | |
| SV007 | Securities and Exchange Commission | EDGAR Search Results for Marvell 10-K filings | |
| SV008 | Marvell Technology | Financial Results | |
| SV009 | Yahoo Finance | Marvell Technology, Inc. (MRVL) Valuation Measures & Financial Statistics | |
| SV010 | CompaniesMarketCap | Marvell Technology (MRVL) - Market capitalization | |
| SV011 | CompaniesMarketCap | Marvell Technology (MRVL) - Revenue | |
| SV012 | Securities and Exchange Commission | EDGAR Search Results for Broadcom 10-K filings | |
| SV013 | Broadcom | Annual Reports | Broadcom Inc. | |
| SV014 | Yahoo Finance | Broadcom Inc. (AVGO) Valuation Measures & Financial Statistics | |
| SV015 | CompaniesMarketCap | Broadcom (AVGO) - Market capitalization | |
| SV016 | Securities and Exchange Commission | EDGAR Search Results for Coherent 10-K filings | |
| SV017 | Yahoo Finance | Coherent Corp. (COHR) Stock Price, News, Quote & History | |
| SV018 | CompaniesMarketCap | Coherent Corp. (COHR) - Market capitalization | |
| SV019 | Trefis | Marvell vs. Broadcom: Same AI Stack, 20x Valuation Disparity | |
| SV020 | Forbes | Marvell Stock Or Broadcom -- A Look At Valuation | |
| SV021 | AInvest | Marvell Optical Upside is Real. The Chokepoint Isn't at Their Layer. | At $253B market cap, with $8.2B in revenue, the stock trades at roughly 31x sales. |
| SV022 | TrendForce | Global AI Optical Transceiver Market to Reach US$26 Billion in 2026; Component Shortages Identified as Primary Capacity Expansion Bottleneck, Says TrendForce | |
| SV023 | LightCounting | AI creates a new wave in demand for optical transceivers and accelerates CPO adoption | |
| SV024 | Business Wire | Celestial AI Secures $250 Million Funding to Revolutionize AI Infrastructure with Its Photonic Fabric™ | |
| SV025 | Tech Company News | Celestial AI Closes $255 Million Series C1 Funding Round | |
| SV026 | Lightmatter | Lightmatter Raises $400M Series D; Quadruples Valuation to $4.4B as Photonics Leader for Next-Gen AI Data Centers | |
| SV027 | Data Center Dynamics | Photonic computing company Lightmatter valued at $4.4bn in $400m funding round | |
| SV028 | Ayar Labs | Ayar Labs Closes $500M Series E, Accelerates Volume Production of Co‑Packaged Optics | |
| SV029 | Latham & Watkins | Latham & Watkins Advises Ayar Labs in US$500 Million Series E Financing | |
| SV030 | PR Newswire | Marvell to Acquire Inphi - Accelerating Growth and Leadership in Cloud and 5G Infrastructure | |
| SV031 | AMD | AMD Completes Acquisition of Xilinx | |
| SV032 | Intel | Intel Acquires Artificial Intelligence Chipmaker Habana Labs | Intel Newsroom |