初创公司尽调
尽调报告 Hardware / Silicon Photonics — Optical Interconnects for AI & HPC Series D (Late Venture) 2026-05-25

Ayar Labs

战略股东加持的独角兽踩上 AI 扩容浪潮,但商业化收入、CPO 采用时点和 NVIDIA 自研光子方案仍卡住投资判断

Ayar Labs 是最受资本支持的纯 in-package optical I/O 公司,战略投资人阵容一线,TeraPHY+SuperNova 已跑通,2026-2027 年放量窗口也可信;但商业收入尚未验证,NVIDIA 和 Broadcom 自研 CPO 项目带来真实竞争风险,在看不见收入的情况下,Series D 相对 Lightmatter 和 Celestial AI 的估值溢价也难以证明。应紧密跟踪;尚未去风险到可以激进承销。

封面要素

成立时间 01
2015 [CO001]
总部 02
Santa Clara, California [CO002]
最新轮次 03
$155M Series D (Dec 11, 2024) [CI003]
Series D 轮估值 04
Above $1B (analyst/press; exact undisclosed) [CI004, CV002]
已披露累计融资 05
370 USD M [CI004, CV003]
Series D 轮领投方 06
Advent Global Opportunities & Light Street Capital [CI003, CV004]
核心产品 07
TeraPHY optical I/O chiplet + SuperNova remote laser source [CO003, CE001]

公司概况

Ayar Labs 总部位于 California 州 Santa Clara,成立于 2015 年,主攻面向 AI 和 HPC 系统的封装内光学 I/O 商业化。公司的两段式产品栈由 TeraPHY 光学 I/O chiplet 和 SuperNova 多波长远端激光源组成,目标是在加速器封装内部和跨机架连接中替代电 SerDes,让 scale-up fabric 在长距离下获得显著高于铜缆或可插拔光模块的带宽、更低延迟和更低功耗。公司累计融资约 $370 million,最近一轮是 2024 年 12 月 $155 million Series D,将估值推到 $1 billion 以上;战略投资方包括 NVIDIA、AMD Ventures、Intel Capital、GlobalFoundries、HPE、Lockheed Martin Ventures 和 Advent。商业化收入、ARR、客户集中度、毛利率和烧钱速度没有公开披露;与超大规模云厂商以及 DoD/DOE 项目的接触更多由合作方和补助驱动,还没有在商业规模上被验证。核心投资分歧在于,Ayar 的先发 IP 和战略股东阵容,能否跑赢 NVIDIA 自研 Spectrum-X/Quantum-X CPO 项目、Broadcom 的 CPO 推进,以及 2026-2028 年仍不确定的 CPO 采用曲线。

官网
ayarlabs.com
成立时间
2015-01-01
创始人
Vladimir Stojanović, Mark Wade, Chen Sun
创立地点
Emeryville, California (Bay Area — DARPA POEM project from MIT/UC Berkeley)
总部
Santa Clara, California
产品
Ayar Labs 销售 TeraPHY 光学 I/O chiplet——一颗 CMOS 硅光子裸片,可通过标准 chiplet 接口与 CPU、GPU、交换机 ASIC 或 FPGA 一起集成进客户计算封装——并搭配 SuperNova,一种远端多波长 CW 激光光源,通过单根光纤为 TeraPHY 供光。组合系统在每颗 chiplet 上提供 Tbps 级带宽,功耗为个位数 pJ/bit,把高带宽互连距离从毫米级拉到公里级,并在 AI scale-up 和分解式计算 fabric 中替代电 SerDes 加可插拔收发器。
客户
AI 加速器厂商、超大规模数据中心运营商(目标包括 Microsoft/Google/Meta/AWS)、HPC 系统厂商(HPE Slingshot、防务 HPC)以及美国 DoD/DOE 高性能计算项目(DARPA PIPES、国家实验室)。
商业模式
面向计算封装集成商和系统 OEM,直接 B2B 销售 TeraPHY chiplet 与 SuperNova 激光模块;现阶段通过评估套件准入推进商业化,客户项目投产后再按高批量 chiplet / 激光器单价变现。DARPA/DOE 的非稀释性补助补充 R&D 资金。
阶段
Late-venture (Series D) private company
融资情况
从 Seed、Series A、B、C 到 D 轮,加上 DARPA/DOE 补助,累计融资约 $370 million;最近一轮为 2024 年 12 月 11 日宣布的 $155 million Series D,由 Advent Global Opportunities 和 Light Street Capital 领投,NVIDIA、AMD Ventures、Intel Capital、3M Ventures、Autopilot、GlobalFoundries、HPE Pathfinder、Lockheed Martin Ventures、Boardman Bay、Founders Fund 和 Playground Global 参投;公司仅将 Series D 轮估值描述为超过 $1 billion。
[CO001, CO002, CO003, CO005, CE001, CI003, CI004, CV001]

执行摘要

主要优势

  • 在 monolithic-CMOS in-package optical I/O 上先发且专利厚,技术脉络锚定十年 DARPA POEM/PIPES 研究。
  • 战略股东已包括 NVIDIA、AMD Ventures、Intel Capital、GlobalFoundries、HPE Pathfinder、Lockheed Martin Ventures——在私有 CPO 同行中,AI 硅战略方最集中。
  • 2024 年 12 月 $155M Series D 把累计融资推至约 $370M、估值抬到 $1B 以上,为 2026-2027 年 CPO 拐点提供多年资金跑道。
  • TeraPHY + SuperNova 产品栈已可运行,OCP 验证了 2 Tbps 级 chiplet 性能,公开路线图还指向更高带宽和更多波长。
  • GlobalFoundries 合作给出高产能 300mm CMOS 硅光工艺路径,少数同行能匹配。

主要风险

  • NVIDIA 已发布 Spectrum-X、Quantum-X Photonics 交换机,Broadcom 也在推进 CPO 项目;CPO TAM 刚要拐点时,交换机层可能绕过 Ayar chiplet。
  • CPO 商业化节奏仍有争议——LightCounting、Dell'Oro、SemiAnalysis 给出的 2026 TAM 估算相差 3 倍,分析师也质疑 hyperscaler 是否会按 Ayar 估值暗含的时间表,从可插拔 / LPO 光模块迁移。
  • 商业收入、ARR、毛利率、烧钱速度、现金、跑道和客户集中度都未披露,今天「$1B 以上」估值无法用收入或单位经济性锚点交叉验证。
  • NVIDIA、AMD、Intel、HPE、GlobalFoundries 等战略投资人与潜在客户重叠,既给生态期权,也在任何一方选择自研 optical I/O 而非采购时放大集中度风险。
  • DARPA / DoD 项目暴露,加上美国对 photonics 出口管制收紧,带来监管和客户结构风险,尤其针对任何面向中国的需求。
  • 单一依赖 GlobalFoundries Fotonix 和先进封装(CoWoS 级)产能,形成供应链集中度;即便需求存在,也可能卡住爬坡。

未决问题

  • 公司未披露 Series D 确切 post-money 估值;「above $1 billion」只是媒体 / 分析师表述,不是官方标记。
  • 没有公开审计收入、ARR、毛利率、EBITDA、现金、烧钱速度或资金跑道数据。
  • 管线覆盖、已签但未披露的 offtake 承诺,以及客户集中度百分比仍属保密。
  • hyperscaler 设计中标(Microsoft、Google、Meta、AWS)的确认量产时间未公开披露。
  • 相对 NVIDIA / Broadcom 自研 CPO 路线图的长期独立性,以及任何 M&A 可选性,仍是未解的战略问题。

目录

Chapter 01

01公司概况

1.1 业务身份与核心产品

Ayar Labs 把自己定位为一家共封装光学公司,切的是 AI 基础设施里的数据搬运问题。公司公开材料一直称其成立于 2015 年,核心是用可直接集成进加速器、CPU、GPU、交换芯片和其他先进封装的光学 I/O,替代铜缆和可插拔光模块造成的瓶颈。商业系统围绕两个紧耦合组件搭建:TeraPHY 光学引擎和 SuperNova 多波长远端光源。按当前产品定位,Ayar 声称这套光学 I/O 栈相对传统可插拔光模块加电 SerDes,可实现 5x-10x 更高带宽、约 10x 更低延迟和显著更好的能效,同时把通信距离从毫米级延伸到公里级。公司官网话术也已经从只讲组件,转向 AI scale-up 平台:当前材料强调“超越机架”的连接,宣传每个光学引擎超过 8 Tbps,并把数千块 GPU 像一个统一系统运行作为目标终态。技术页面进一步说明,这套方案更偏标准化,而不是完全定制:当前支持 AIB,下一代路径指向 UCIe,并与 CW-WDM MSA 对齐。这个定位很关键,因为 Ayar 不只是在卖一个光子器件;它想成为下一代 AI 封装、机架和分解式内存架构里的光互连层。 [CO001, CO002, CO003, CO004, CO005, CO006]

快照 KPI 表
指标数值 / 状态日期置信度缺口
成立时间20152026-03-03None
总部硅谷;公开材料同时使用加州 Santa Clara 和 San Jose2026-03-03各次发布对总部表述不完全一致
现任 CEOMark Wade(联合创始人)2026-05-25None
核心产品TeraPHY 光引擎,搭配 SuperNova 远程光源2026-05-25None
当前对外宣传性能>8 Tbps/光引擎;<25 ns 端到端光 I/O 延迟2026-05-25公司声称的当前方案指标
最新披露轮次Series E,$500M2026-03-03None
已披露累计融资$870M2026-03-03None
最新披露估值$3.75B2026-03-03私营公司估值;确切股权结构表未披露
办公布局San Jose 总部扩张,并开设 Hsinchu, Taiwan 办公室2025-07-17未披露确切办公室面积或员工数
机架级里程碑Wiwynn 系统设计目标为 1,024 个加速器,且每个加速器具备 >100 Tbps 光连接2026-03-11参考设计里程碑,尚非已披露客户部署
收入 / 运行率未公开披露2026-05-25需要管理层披露或原始财务记录
员工数未公开披露2026-05-25招聘和办公地点扩张是公开信息;确切员工数不是

这些是基于当前产品页面和融资公告整理的时点型公司概览指标。若干行有意保留披露缺口,因为 Ayar 是私营公司,不发布标准经营指标看板。

[CO001, CO002, CO003, CO008, CO009, CO023]
FO002: 公司快照逻辑

Ayar 的研究根基、产品栈、制造生态和机架级合作伙伴如何连成商业化投资逻辑。

这张图是商业化逻辑图,不是字面意义上的供应链图。它概括 Ayar 的公开叙事:光小芯片、远置激光器,以及合作伙伴主导的制造 / 系统集成如何拼成一套可部署的 AI 基础设施栈。

[CO003, CO004, CO016, CO017, CO030, CO032]

1.2 领导层与治理

Ayar Labs 的公开治理画像仍由创始人主导,Mark Wade 现在已经明确成为运营重心。公司在 2023 年 12 月宣布,联合创始人、时任 CTO Wade 接替 Charles Wuischpard 出任 CEO;Wuischpard 在交接期短暂担任顾问后离开。Ayar 自身公告和 Lightwave 报道都把这次变动描述为一次有意安排:公司从外部引入的规模化高管,重新切回技术创始人来带下一阶段商业化。公开领导层页面显示,公司的管理层厚度比早年披露更强,包括 CFO Lisa Cummins Dulchinos、CTO 兼联合创始人 Vladimir Stojanovic、首席科学家兼联合创始人 Chen Sun、首席战略官 Vivek Gupta、制造与台湾运营负责人 Scott Clark,以及总法律顾问兼公司秘书 Matthew Gloss。同一页面还列出了辨识度较高的董事会名单,包括 Craig Barratt、Pat Gelsinger、Will Graves、Jordan Katz、Ganesh Moorthy 和 Geoff Tate。仍然缺失的是后期投资者通常想看的治理细节:Ayar 没有公开拆分委员会架构、董事独立性、持股比例,或 Series E 董事会观察员披露之外的详细董事权利。因此,这是一套足以支撑生态信用的治理画像,但在硬治理机制上仍带有私营公司信息披露偏轻的特征。 [CO009, CO010, CO011, CO012, CO037, CO044]

领导层与创始人表
人员角色公开背景 / 职责覆盖范围 / 创始人-市场匹配关键人依赖
Mark Wade首席执行官兼联合创始人前 CTO;来自 MIT / UC Berkeley 光 I/O 工作的共同发明人,该研究催生 Ayar Labs把技术源头与当前商业领导力连起来
Vladimir Stojanovic首席技术官兼联合创始人产品路线图和 AI scale-up 架构的公开技术面孔掌握核心架构可信度和外部技术叙事
Chen Sun首席科学家、副总裁兼联合创始人公开领导层页面列名的联合创始人和科学负责人延续来自创始研究基础的器件和光子能力
Lisa Cummins Dulchinos首席财务官Ayar 进入后期规模化融资阶段的公开财务负责人在创始人运营团队之外增加机构化财务覆盖
Vivek Gupta首席战略官公开列名的战略负责人支撑超大规模云合作和生态定位
Scott Clark制造与运营副总裁;台湾区负责人在公开组织图中负责制造和台湾运营对高量产爬坡和台湾执行至关重要
Matthew Gloss总法律顾问兼公司秘书公开法律与治理负责人在后期扩张中增加合规和董事会流程能力
公开列名的董事会成员董事:Craig Barratt, Pat Gelsinger, Will Graves, Jordan Katz, Ganesh Moorthy, Geoff Tate只公开姓名;委员会和独立性未披露传递行业可信度,但治理透明度有限

仅包括公开列名的领导层和董事会成员。Ayar 不披露委员会结构、持股比例或完整治理矩阵,因此本表有意保持不完整。

[CO009, CO010, CO011, CO012, CO037, CO044]

1.3 融资历史与资本基础

Ayar Labs 的融资轨迹已经清楚地从深科技风险资本,走向大额 AI 基础设施成长资本。2022 年 4 月的 Series C 融资 $130 million,由 Boardman Bay Capital Management 领投,Hewlett Packard Enterprise 和 NVIDIA 等战略方参投,早期支持方 Applied Ventures、GlobalFoundries、Intel Capital 和 Lockheed Martin Ventures 也在其中。这轮融资重要,不只是因为扩充了股权结构表,更因为 Ayar 当时表示已有首批批量商业出货签约,并预计年底前出货数千件,把融资和商业化主张挂钩,而不只是 R&D。2024 年 12 月,公司宣布 $155 million Series D,由 Advent Global Opportunities 和 Light Street Capital 领投,并称累计融资达到 $370 million、估值升至 $1 billion 以上。到 2026 年 3 月,Ayar 披露了更大一轮 Series E:Neuberger Berman 领投 $500 million,使累计融资达到 $870 million,估值达到 $3.75 billion。Series E 投资人再次扩围,引入 ARK Invest、Insight Partners、Qatar Investment Authority、Sequoia Global Equities 和 1789 Capital 等机构资本,也新增 Alchip 和 MediaTek 作为半导体生态战略投资方。Neuberger Berman 还获得董事会观察员席位,意味着公司在未上市前已叠加更正式的成长阶段治理层。 [CO013, CO014, CO015, CO016, CO021, CO022]

利益相关方 / 投资人图谱
利益相关方角色 / 轮次战略重要性控制权 / 经济重要性尽调问题
Neuberger BermanSeries E 轮领投方(2026)锚定最新成长期融资,并获得董事会观察员角色本轮领投方;未披露持股比例澄清观察员权利、信息权和未来流动性预期
Boardman Bay Capital ManagementSeries C 轮领投方(2022)帮助融资商业化,并支撑首次批量出货主张较早的领投财务支持方,在规模化阶段仍有持续影响澄清当前持股,以及在 D/E 轮是否按比例跟投
投资人:Advent Global Opportunities + Light Street CapitalSeries D 轮联合领投方(2024)将 Ayar 推到独角兽以上,并资助高量制造准备商业化到 Series E 之间的重要桥接投资人澄清任一方在 Series E 后是否仍有治理权或特殊权利
AMD Ventures + NVIDIA后期轮次战略投资人验证其与未来加速器生态和 AI 工作负载的相关性战略投资人而非控制型投资人;确切经济权益未披露判断投资是否还包含技术路线图或商业承诺
HPE Pathfinder战略投资人和合作伙伴把 Ayar 接入 Slingshot / HPC 和 AI 架构工作战略关系看起来比已披露股权规模更重要测试 HPE 是否仍是活跃开发和设计导入伙伴
GlobalFoundries战略投资人和制造伙伴提供晶圆代工和硅光子工艺杠杆制造依赖集中,因此运营重要性高确认当前工艺节点、产能预留和单一供应商风险缓释
战略方:Lockheed Martin Ventures / Lockheed Martin战略投资人和防务应用伙伴支撑双用途可信度和非超大规模云应用可选性经济权益未披露;关系对应用证明可能更重要澄清防务项目是否已变成经常性设计项目,还是仍为试点
Alchip + MediaTekSeries E 战略投资人 / 生态伙伴强化面向超大规模云设计的先进 ASIC 和封装资源即便披露股权很小,战略对齐也重要确认联合参考设计范围,以及是否存在客户专属项目
Wiwynn机架级系统伙伴(2026)推动 Ayar 从封装级证明走向可部署机架系统未披露为股权投资人,但在商业上重要,关系到系统化判断该机架设计是展示样板、联合 SKU 路径,还是绑定客户项目

本图谱混合了融资利益相关方和商业生态利益相关方,因为 Ayar 的资本结构和市场进入路径高度交织。公开来源不披露持股比例、清算优先权或按比例跟投权。

[CO013, CO015, CO016, CO017, CO020, CO022]
FO003: 商业化就绪指标

2025-2026 年的关键指标显示,Ayar 正尝试从已验证的组件技术走向规模化部署。

这张图强调就绪指标,而非完整公司 KPI。它把资本、版图和技术部署标志放在一起,因为 Ayar 没有公开披露收入、具名客户数等后期 SaaS 或硬件公司的标准运营指标。

[CO006, CO007, CO023, CO025, CO027, CO028]

1.4 运营足迹、生态系统与商业化姿态

Ayar 的运营足迹和生态系统,现在更适合被理解为一套商业化堆栈,而不是一串标志。公开材料仍把公司锚在 Silicon Valley,但最新公告显示运营版图已经扩大:2025 年 7 月,Ayar 称其在 Taiwan Hsinchu 开设新办公室,并把 San Jose 总部面积扩大一倍,明确把扩张与高批量制造和共封装光学采用绑定。公司没有披露精确员工数,而是通过美国和台湾招聘计划、设施扩张以及专门的运营负责人来释放规模信号。围绕这套足迹的是一张密集的伙伴网络。HPE 通过多年 Slingshot 与 AI 架构合作接入;NVIDIA 通过光学 I/O 合作和投资接入;GlobalFoundries 通过工艺与制造合作接入;Lockheed Martin 通过防务应用开发接入;Intel 通过高曝光 FPGA 演示接入;Wiwynn 则通过机架级系统设计与制造接入。Gazettabyte 2026 年 1 月的访谈进一步补足细节:Ayar 正与 Alchip 和 GUC 合作,把光学引擎塞进先进 ASIC 封装,尤其面向超大规模云厂商部署路径。合在一起,这些关系说明 Ayar 的商业化逻辑不是以狭窄的独立商用模式卖一个光子模块,而是嵌入更广泛的 AI 封装与机架生态。 [CO015, CO016, CO017, CO019, CO020, CO023]

1.5 里程碑与当前扩容轨迹

公司的里程碑路径呈现出重复节奏:先拿 DARPA 相关技术验证,再拿生态证明,系统级 scale-up 到 2026 年才开始成为主线。Ayar 的历史叙事可追溯到 DARPA POEM 中关于处理器内部和处理器之间光通信的工作,随后进入 PIPES 项目,Ayar 与 Intel 在先进 FPGA 封装上演示了光学 I/O。后续里程碑更明确地转向商业化:2022 年 Series C 把融资与已签约的首批批量商业出货绑定;2023 年 Intel FPGA 展示公开了 4 Tbps 双向演示;2024 年 12 月 Series D 将光学 I/O 描述为已可商业化,并与客户路线图对齐;2025 年 7 月台湾扩张则被定位为高批量制造准备。2026 年 3 月的 Series E 和 Wiwynn 公告把故事又推高一层,从 chiplet 与封装就绪,转向面向 1,024 个加速器、每个加速器超过 100 Tbps 光连接的机架级参考架构。这一顺序很重要。Ayar 的公开证据已经不再局限于器件级或单封装性能主张;它现在给出了一条从封装集成走向可部署超大规模机架系统的可信路径。 [CO018, CO019, CO025, CO027, CO028, CO029]

里程碑表
日期事件类型金额 / 估值 / 状态参与方含义
2015Ayar Labs 从 MIT / UC Berkeley 光 I/O 工作中成立创立公司成立创始团队包括 Mark Wade、Vladimir Stojanovic、Chen Sun 及相关研究合作者确立深科技出身和 DARPA 相关研究根基
2020Intel 和 Ayar 在 DARPA PIPES 下演示 TeraPHY 光 I/O监管2 Tbps 演示里程碑Intel, Ayar Labs, DARPA显示在先进封装中用光 I/O 替代电 I/O 的技术可行性
2022-02-24HPE 宣布多年合作和投资合作战略合作 + 投资合作方:HPE, Hewlett Packard Pathfinder, Ayar Labs把光 I/O 延伸到 HPC / AI 系统架构讨论
2022-04-26宣布 Series C 融资融资$130MBoardman Bay, HPE, NVIDIA, 现有战略投资人为商业化提供资金,并与首次按合同批量商业出货的说法同步出现
2022-05-25宣布 NVIDIA 合作合作战略 AI 合作NVIDIA, Ayar Labs传递其与未来 AI scale-out 架构的相关性
2022-10-12宣布 Lockheed Martin 合作合作防务应用开发Lockheed Martin, Ayar Labs增加 DoD 导向应用路径和防务可信度
2023-11-08Intel Agilex 光 FPGA 在 Supercomputing 2023 展示产品4 Tbps 双向展示Intel, Ayar Labs把产品故事从概念推进到公开系统演示
2023-12-11Mark Wade 被任命为 CEO治理创始人领导层交接人员:Mark Wade, Charles Wuischpard, Playground Global让公司重新聚焦创始人带领的商业化
2024-12-11宣布 Series D 融资融资$155M;估值超过 $1B;累计融资 $370MAdvent Global Opportunities, Light Street, AMD Ventures, Intel Capital, NVIDIA 等赋予 Ayar 独角兽地位,并资助制造与客户路线图对齐
2025-07-17宣布领导层与办公布局扩张规模化新增 Hsinchu 办公室;San Jose 总部面积翻倍Ayar Labs表明公司围绕台湾和硅谷推进面向量产的建设
2026-03-03宣布 Series E 轮融资融资$500M;估值 $3.75B;累计融资 $870MNeuberger Berman 及机构和战略投资者为大规模量产和测试产能提供后期资金
2026-03-11宣布与 Wiwynn 开展机架级 AI 合作合作1,024 加速器参考设计;每个加速器 >100 TbpsAyar Labs、Wiwynn将 Ayar 从 chiplet 供应商推到机架级架构合作伙伴位置

本时间线重点列出公开披露的融资、合作、治理、技术演示和当前规模化基础设施里程碑。未公开的客户试点和未披露的出货里程碑, 除非明确宣布,否则不纳入。

[CO001, CO009, CO013, CO015, CO016, CO018]
FO001: 公司里程碑时间线

从 Ayar Labs 的研究根基到 2026 年 3 月,关键创立、融资、合作和商业化里程碑。

时间线只跟踪有公开日期的已披露里程碑。客户试点、设计定点和出货量大多仍未披露,因此没有作为有日期事件绘制。

[CO009, CO013, CO015, CO018, CO019, CO020]

1.6 采用风险与剩余披露缺口

公司概况里的主要风险不在于 Ayar 技术是否有意思,而在于这项技术能否足够快地穿过可靠性、可维护性和超大规模部署的完整关卡,从而支撑后期私营估值。Ayar 自有材料现在称方案已经验证、可投产,但独立报道仍把 2026 年描绘成过渡期,而不是超大规模部署已经完成的故事。The Register 2026 年 3 月的报道尤其适合作为反向视角,因为它点出了公司新闻稿弱化的运营细节:液冷交互、光纤走线、可制造性、遥测要求,以及与可热插拔光模块相比,共封装光学引擎一旦失效会带来更大的“故障影响半径”。DARPA 和 Laser Focus World 也提醒读者,行业长期目标——100 Tbps 以上、每 bit 能耗很低的光学 I/O——仍是一段开发旅程,不是已经完成的市场事实。披露层面,Ayar 仍不公开精确员工数、已点名客户部署、收入、股权拆分和委员会级治理细节。对尽调而言,这意味着规模只能从融资、设施扩张、技术里程碑和伙伴密度中推断,而不能直接读取标准后期经营指标。 [CO024, CO035, CO036, CO037, CO038, CO039]

1.7 图表与证据

Chapter 02

02市场分析

2.1 市场边界与定义

Ayar Labs 并不参与整个数据中心网络市场;它参与的是更窄的一块:共封装光学引擎和光学 I/O chiplet,用光而不是长距离电走线,把数据直接从加速器、交换芯片和先进封装中搬出来。更宽的可比集合是 AI 集群光互连市场,包括可插拔收发器、线性驱动可插拔模块、主动光缆、交换机侧 CPO,以及同时用于 scale-out 和 scale-up 网络的光链路。Ayar 的核心产品定位还要更窄:TeraPHY 光学 I/O chiplet 搭配 SuperNova 远端激光源。因此,头部光互连 TAM 数字会高估 Ayar 的直接收入池,因为大量支出落在收发器模块、系统级交换机、布线和传输设备上,而这些不是 Ayar 自己销售的产品。合理的市场边界应从嵌入封装内的光学引擎内容开始,再向外扩展到可插拔模块和 LPO 等相邻替代支出。这个区分很重要,因为买方决策不是简单的“买光模块”,而是“当带宽密度、功耗、延迟和传输距离足以抵消集成成本时,用封装级光学 I/O 替代铜缆或可插拔模块”。Ayar 公开材料一直把 scale-up AI fabric 作为最高价值用例,而独立分析师来源显示,可插拔模块、LPO 和 DCI 光互连仍是重要的相邻和替代市场,不会一夜消失。[CM001, CM002, CM003, CM004, CM005, CM041]

市场定义表
细分市场 / 类别纳入支出排除支出主要买方 / 付款方对 Ayar 的意义
AI 集群光学 TAMAI 计算节点和后端互连网络中使用的可插拔光模块、LPO、CPO、有源光缆和光链路仅电网络芯片、GPU/CPU、服务器、电力、散热、地产超大规模云厂商、neocloud 厂商、交换机 OEM、加速器供应商可作为标题级 TAM 背景,但明显宽于 Ayar 的直接产品范围
CPO 模块和交换机侧光学集成进 Ethernet 或 InfiniBand 交换机平台的共封装光引擎 / 模块独立可插拔光模块、多数长途电信光器件、非 AI 传输硬件交换机厂商和超大规模云厂商网络架构团队近期商业 SAM,因为分析师来源预计 CPO 会先在交换机上放量
光 I/O chiplet / 封装内光引擎与加速器或交换机共封装的 chiplet、retimer 和远程光源部件完整交换机机箱、机架系统、外部模块、通用传输设备加速器 ASIC 团队、先进封装合作伙伴、超大规模云厂商芯片团队Ayar 核心市场,也是最贴近产品的价值池
可插拔和线性驱动可插拔光模块AI 集群使用的 800G/1.6T 可插拔光模块和 LPO封装级光 chiplet 和大多数封装内互连网络平台团队和超大规模云基础设施买方主要替代品集合;既能延后 Ayar 采用,也受益于更广光学需求
横向扩展 / DCI 光学连接分布式 AI 园区的 ZR/ZR+、相干光模块、WDM 传输与 AI 工作负载无关的城域 / 长途运营商网络云网络架构师和传输团队重要邻近市场,因为分布式集群会扩大光学预算,但 Ayar 目前只能部分变现

这些行刻意把宽口径 AI 光学 TAM 与更贴近 Ayar 的窄口径光引擎市场拆开。 纳入和排除支出反映 Ayar 真正在卖什么,以及哪些支出只是定义替代品集合或邻近基础设施预算。

[CM001, CM002, CM005, CM041, CM043]

2.2 市场规模——TAM、SAM 和贴近产品的 SOM 代理

Ayar Labs 的公开市场规模口径差异很大,因为各方测量的东西并不相同。最宽的 TAM 口径是 AI 集群光互连:LightCounting 公开摘要称,该市场将从 2024 年约 $5 billion 增至 2026 年超过 $10 billion,并在光互连从 scale-out 扩散到 scale-up fabric 的过程中持续增长到 2030 年。更窄的 SAM 口径是独立 CPO 市场。在这里,公开 2026 年估计已经相差超过 3x:PW Consulting 基于 2025 年 $560 million 的基数,隐含 2026 年约 $1.247 billion;HDIN Research 则给出更宽的 2026 年 $2.2-$4.2 billion 区间。差异来自边界选择:只算模块还是算更广生态价值,先看交换机部署还是所有 CPO 架构,以及对超大规模云厂商采用时点的不同假设。对 Ayar 而言,由于公开来源没有披露定价、设计定点数量或已承诺出货量,一个更贴近产品的口径比经典企业 SOM 预测更有辩护力。若用 PW Consulting 自身 2025 年细分占比作为低情景代理,服务 AI/HPC 工作负载的光学引擎切片在 2026 年约为 $0.43 billion。这不是 Ayar 收入预测,但相较完整光互连 TAM,它更现实地代理了 CPO 栈中最接近 Ayar 当前产品足迹的部分。[CM006, CM007, CM008, CM009, CM010, CM011]

TAM / SAM / SOM 规模测算视角表
发布方年份地理范围市场定义数值 / 区间增长信号方法说明置信度关键限制
LightCounting / Optical Connections News2025-2026全球AI 集群光学(可插拔 + LPO + CPO)2026 年 >$10B,高于 2024 年约 $5B两年内大约翻倍;增长延续至 2030 年面向 AI 集群和云数据中心的自上而下光学需求模型宽于 Ayar 的直接产品范围,也包含替代光学品类
LightCounting2025全球云数据中心用 Ethernet 光收发器、LPO 和 CPO摘要未披露单一 2026 年美元数值2025-2026 年年增长 30-35%;2027-2030 年 15-20%2025 年 7 月云数据中心光学报告公开摘要中高增长预测偏方向性,且混合多个光学品类
Dell’Oro Group2026 年展望全球包含 CPO 贡献的 AI 网络堆栈CPO 可能为市场扩张贡献数十亿美元预计 2026 年 AI 网络支出实现强劲两位数增长分析师展望锚定超大规模云厂商资本开支和交换机市场切换不是独立 CPO 收入数字
HDIN Research2026全球共封装光学市场2026 年约 $2.2B-$4.2B至 2031 年 CAGR 为 25%-35%围绕 1.6T 拐点的厂商 / 供应链市场模型中低方法只做到部分透明
PW Consulting2025-2032全球CPO 模块市场2025 年:$560M;2026 年:约 $1.247B;2032 年:$32.68B至 2032 年 CAGR 为 76.45%披露组件和终端用户份额拆分,可据此推导更贴近产品的口径曲线激进,发布方层级较低
作者基于 PW Consulting 份额推导2026 年低位情景全球 AI/HPC 部分AI/HPC CPO 模块中的光引擎内容~$0.43B推导口径,不是增长预测56.7% AI/HPC 终端用户份额 × 60.7% 光引擎份额 × 2026 年低位市场规模产品近端切片代理,不是 Ayar 收入或市场份额预测
LightCounting2030 年展望全球CPO 引擎 / 端口2030 年市场价值约 $10B,端口约 100M2027 年后放量2025 年 12 月 CPO 说明公开摘要2030 年引擎市场不能与 2026 年仅模块口径直接比较

这些行刻意混用多种测算视角,因为没有公开来源能干净隔离 Ayar 的直接 SAM/SOM。 宽口径 AI 光学数据应作为 TAM 背景,2026 年 CPO 数据作为 SAM 式边界, 推导出的光引擎切片则更像贴近产品的 SOM 代理,而不是公司预测。

[CM006, CM007, CM009, CM010, CM011, CM012]
FM001: 市场规模视角

宽口径 AI 光学 TAM 一旦收窄到 CPO 层,再收窄到最接近 Ayar 当前产品足迹的光引擎内容,规模会明显变小。

TAM 采用 LightCounting 的 AI 集群光学口径,SAM 采用公开的 2026 年 CPO 市场区间,底层则是作者基于 PW Consulting 已发布的组件和终端用户份额推导出的光引擎代理指标。这是贴近产品的视角,不是 Ayar 收入预测。

[CM006, CM008, CM012, CM041]
FM002: 市场估计区间

公开估计的低位、基准和高位,衡量同一个指标:2026 年全球共封装光学市场规模(十亿美元)。

所有数值均为十亿美元。低位点采用 PW Consulting 对 2026 年模块市场的估计;基准和高位点采用 HDIN Research 公开的 2026 年区间。这些是公开区间边界,并非彼此校准后的模型。

[CM008, CM009, CM010]

2.3 买方、用户与预算结构

Ayar 式光学 I/O 的直接买方通常不是最终超大规模云厂商单独一方,而是加速器设计方、交换机厂商、先进封装伙伴和系统集成商的组合;这些角色把云厂商的带宽需求转化为封装级设计定点。在最有吸引力的细分市场——直接挂在加速器上的 scale-up fabric——运营用户是云厂商 AI 平台团队,但即时采购决策落在集成封装的硅片或子系统供应商手里。交换机侧 CPO 是第二个细分市场,买方是网络平台团队或交换机 OEM,超大规模云厂商的网络架构与基础设施财务团队则是实际付款方。ODM 和先进 ASIC 伙伴等系统集成商夹在中间,因为它们要判断基于 chiplet 的光学架构能否以可接受良率完成封装、测试、散热和制造。HPC 与 exascale 系统是更小但仍有意义的细分市场,因为它们重视低延迟、分解式架构和内存 / fabric 效率,即便超大规模云的出货量尚未完全释放。跨这些细分市场看,预算所有权通常在基础设施架构、网络或硅平台负责人手里,而不是采购团队购买大宗光模块。这有利于 Ayar,只要它能被写入下一代加速器或交换机平台规格;但这也意味着认证周期漫长且由伙伴牵引,不是快速更换模块。[CM026, CM027, CM028, CM029, CM030, CM031]

细分市场 / 买方地图
细分市场直接买方终端用户 / 付款方工作流预算所有者采用触发点Ayar 当前匹配度
加速器封装级 scale-up 互连加速器 / ASIC 供应商或超大规模云厂商芯片团队为集群吞吐和利用率买单的超大规模云 AI 平台负责人结合先进封装和协议验证的封装级导入设计芯片平台 VP;AI 基础设施架构负责人机架级和多机架 scale-up 互连网络内铜互连遭遇距离 / 功耗墙战略匹配度最高,但认证路径最长
AI 后端网络中的交换机侧 CPO交换机 OEM、交换机 ASIC 供应商或超大规模云网络团队超大规模云网络资本开支预算绑定 1.6T/3.2T 路线图和液冷机架的交换机平台换代网络架构;交换机芯片 GM1.6T+ 下的端口密度、功耗和韧性近期 SAM,但会被 NVIDIA/Broadcom 策略挤压
机架级系统和 ODM 平台ODM / 系统集成商,加上封装和板级设计合作伙伴超大规模云平台工程预算参考设计、子系统协同设计和可制造性爬坡平台工程;先进系统架构师需要可制造封装支撑多机架 scale-upAyar 生态的重要渠道放大器,包括 Alchip 和未来 ODM
HPC / exascale 系统系统 OEM 或国家实验室主承包商政府实验室、企业 HPC 买方或主权项目多年架构项目,要求严格的延迟和内存互连网络项目 CTO;首席架构师;政府项目办公室需要解耦、低延迟和高带宽 Slingshot / 互连网络演进有用的验证场景,但体量低于超大规模云
分布式 AI 园区 / scale-across DCI云网络架构和传输团队云基础设施财务和容量规划园区 / 站点间光学设计,与 AI 集群选址决策绑定网络架构 VP;容量规划受电力约束、分布在建筑或园区间的集群会扩大光学支出的邻近市场,但许多价值落在相干 / DCI 供应商手里,而不是 Ayar 直接拿到

买方和付款方在各行不同,因为 Ayar 通常被指定进更大的芯片或系统项目, 而不是像大宗光收发器那样被采购。因此预算归平台、网络和芯片团队掌握, 而不是普通组件采购。

[CM026, CM027, CM029, CM030, CM031, CM042]
FM003: 买方 / 细分市场图

Ayar 的采用路径穿过芯片和系统项目;超大规模云厂商通常买单,但封装、交换机和系统合作伙伴往往先拍板设计导入。

矩阵条目综合了分析师和公司证据中的直接买方与付款方角色。标签只表示方向,用于说明采用发生前谁必须先被说服。

[CM029, CM041, CM042, CM043]

2.4 增长驱动因素与采用时点

核心需求驱动很简单:AI 集群正在变得更大、更受电力约束,也更依赖通信。Ayar 自身围绕生成式 AI 的叙事认为,推理已经变成机架级问题,而微调和训练会扩展到数百、数千,最终数万块 GPU。独立分析师工作也指向同一方向。Dell’Oro 称,最大型超大规模云 AI 集群已经接近 100,000 个加速器,并可能在不久后达到 1 million,这让互连设计从事后补充变成一阶系统问题。电力是第二个主要驱动。IEA 现在预计,到 2030 年,全球数据中心用电需求将翻倍以上,其中 AI 优化数据中心的需求将增长超过四倍;Dell’Oro 也强调未来 AI 机架可能达到 600 kW 级,并且集群会分布在多个设施中。资本可得性是第三个驱动:Alphabet 和 Meta 最新文件显示基础设施支出仍然巨大,Amazon 文件也继续强调物业和设备投资中的服务器与网络设备。这些力量都利好每个封装、每瓦提供更高带宽的光互连架构。标准和生态进展同样有帮助:UCIe 降低封装边界的集成摩擦,CW-WDM 标准化则支持远端光源互操作。合在一起,这些趋势形成一条可信的多年采用跑道,即便准确斜率仍不确定。[CM020, CM021, CM022, CM023, CM024, CM030]

增长驱动因素和约束表
因素方向时间公开证据对 Ayar 的影响尽调追问
AI 模型和集群扩展到单机架之外正面即刻至 2030 年Ayar 对工作负载的表述,加上 Dell’Oro 对集群规模的评论提高封装级带宽密度和低延迟互连网络的价值索取 2026-2028 年设计中按客户拆分的带宽地图和 radix 假设
超大规模云基础设施预算维持高位正面2026-2028Alphabet、Meta 和 Amazon 文件,加上 Dell’Oro 与资本开支挂钩的网络展望支撑客户继续愿意为先进互连实验和平台换代投钱跟踪网络 / 光学支出是否随计算支出同步扩张
AI 数据中心的电力和供电约束正面2026-2030IEA 电力需求预测和 Dell’Oro 对 600 kW 机架的评论如果能降低每 bit 功耗和铜走线长度,光 I/O 价值主张会更强验证相对高端可插拔光模块和 LPO 的真实系统级节能
标准和生态进展(UCIe、CW-WDM、封装合作伙伴)正面2025-2028Ayar 产品页、OFC 摘要和合作公告降低集成摩擦,让多厂商堆栈中的设计赢单更可信确认当前头部客户要求哪些标准
可插拔光模块和 LPO 仍是可用替代品负面2026-2028LightCounting 和 Dell’Oro 均提到混合采用和可插拔光模块仍然有用即便整体光学市场增长,也可能延后 Ayar 拿到份额梳理哪些买方群体真正需要封装级光学,而不是升级模块即可
封装良率和已知良品 chiplet 要求负面2026-2028IEEE OFC 关于连接器化 chiplet 和 HVM 流程的论文给光 chiplet 带来商业化和毛利率风险在假设规模经济前,先索取良率、组装和测试指标
交换机优先的 CPO 放量混合 / 负面2026-2027分析师说明和 NVIDIA photonics 公告都指向交换机平台先行扩大 CPO 整体 SAM,但可能让 Ayar 的加速器封装逻辑排在后面每个预测都要拆开交换机层和加速器层采用假设
chiplet 侧光学的公开单位经济数据稀少负面持续公开来源披露区间,不披露出货 ASP 或 attach rate从 TAM 换算到 Ayar 可拿收入高度不确定把市场规模换算成估值前,先在 NDA 下取得客户、端口和定价数据

这些驱动和约束行刻意混合市场拉动、技术就绪度和商业化摩擦。 时间基于基准运行日期时可获得的公开证据,不应被解读为确定的部署排期。

[CM016, CM019, CM025, CM033, CM034, CM035]
FM004: 采用漏斗 / 价值链图

AI 基础设施中的 CPO 采用通常从架构瓶颈出发,经过封装验证再进入试点部署;最大摩擦集中在封装和平台集成阶段。

漏斗数值是相对就绪度的顺序评分,不是已观察到的管线转化率。它们反映公开证据:最大流失发生在封装、可制造性和全平台集成环节。

[CM018, CM019, CM025, CM038, CM040]

2.5 约束、替代方案与可插拔–LPO–CPO 之争

Ayar 市场的最强近端约束不是带宽需求不足,而是更容易部署的替代方案仍然可用。LightCounting 和 Dell’Oro 都暗示,采用大概率分阶段推进:可插拔模块继续稳固,LPO 作为低功耗过渡台阶增长,交换机侧 CPO 先起量,封装级光学 chiplet 则在封装和可维护性问题解决后再到来。这与 Ayar 直接相关,因为它最高价值的逻辑依赖光学 I/O 向加速器靠近,而不只是进入网络交换机。独立技术出版物也强化了制造挑战:已知良品光学 chiplet、无源光纤贴装、连接器化、封装良率和热集成,都是高批量部署的闸门。公开分析师评论还明确指出,当前 CPO 实现大多以交换机为中心,大规模 scale-up 出货更可能从 2027 年而非 2026 年开始。竞争压力进一步放大时点风险。NVIDIA 没有等待第三方生态成熟;它已经发布带有激进能效主张的硅光子交换平台,这可能把交换机层价值留在既有平台栈内部。净结果是,Ayar 明显受益于托举光互连市场的同一组宏观驱动,但其商业捕获窗口大概率比最乐观的头部 TAM 数字所暗示的更晚、更窄。[CM016, CM017, CM018, CM019, CM025, CM038]

2.6 图表与证据

Chapter 03

03竞争格局

3.1 市场地图与竞争边界

Ayar Labs 卖的不是面板收发器;它试图把光学 I/O 嵌入封装本身,替代电 I/O 以及部分可插拔光模块角色。这意味着它的竞争集合横跨多层。直接架构层包括 Lightmatter、Celestial AI 和 Ranovus 等私营 scale-up 光子创业公司,它们同样向更大 AI 集群推销光学 fabric。既有厂商层包括 Broadcom、Marvell、Intel、Cisco/Acacia、Coherent 和 Lumentum,它们各自已经掌握交换芯片、硅光子、光 DSP、激光器或收发器分销中的若干环节。替代层包括 Innolight、Eoptolink 和 Accelink 等厂商的可插拔与近封装光模块,以及线性驱动可插拔模块;这些方案在解决部分功耗问题的同时,保留了可维护性和多供应商采购。 这个框架很关键,因为 Ayar 并不需要每个对手都复制它完全相同的 chiplet 设计才会丢份额。NVIDIA 的 Spectrum-X Photonics 交换机证明,交换机 OEM 可以在 fabric 层把共封装光子内部化。Broadcom 和 Marvell 正借助超大规模云客户关系和已落地的硅光子经验,把 CPO 从实验室推向现场。同时,LightCounting 和 Avnet 都认为,LPO 和先进可插拔模块会在许多链路中继续成为默认选择多年;这意味着 Ayar 必须先赢下最难、最高密度的 AI scale-up 用例,否则行业可能先落在更易维护的中间架构上。[CP001, CP005, CP006, CP019, CP033, CP034]

Ayar Labs 竞争对手市场地图
竞争对手 / 类别竞争层级规模信号与 Ayar 的主要重叠重要性
Broadcom交换机层 CPO上市既有厂商;最新 10-K 于 2025-12-18 提交超大规模交换和 AI 互连网络中的 CPO最强证据:交换机厂商可在 Ayar 赢得广泛 socket 前先把 CPO 工业化
Marvell定制 XPU + 互连上市既有厂商;最新 10-K 于 2026-03-11 提交面向多机架定制 XPU scale-up 的共封装光学把硅光、光 DSP 和超大规模云厂商关系捆在一起
Cisco / Acacia客户端光学 + 光引擎上市既有厂商;最新 10-K 于 2025-09-03 提交AI scale-out 光学和光引擎即便封装深度重叠较弱,仍掌握光学客户关系
Intel Silicon PhotonicsOCI chiplet + 可插拔光模块上市既有厂商;最新 10-K 于 2026-01-23 提交共封装光 I/O chiplet 和大规模出货的可插拔基础在 chiplet 式光 I/O 上最接近 Ayar 的既有厂商类比对象
Coherent激光器 / 收发器 / 子系统上市既有厂商;最新 10-K 于 2025-08-19 提交向 AI 互连网络供应光模块和激光器掌握许多 CPO 系统仍需要的使能组件
Lumentum可插拔光模块 + 激光源上市既有厂商;OFC 2026 展示 1.6T 可插拔光模块和 CPO 激光源AI 数据中心可插拔光模块和 CPO 高功率光源同时在过渡架构和使能组件层竞争
Lightmatter端到端光子互连路线图累计融资 $850M;披露估值 $4.4B面向 AI scale-up 的 NPO、OBO、2D/3D CPO 和 3D interposer融资最充足的私营对手,产品面比 Ayar 更宽
Celestial AIPhotonic Fabric scale-up 互连累计融资 $515M;2025 年宣布被 Marvell 收购封装到机架的光学 scale-up 互连网络和 16 Tbps chiplet与创业公司路线直接重叠,也证明创业公司正被既有厂商整合
RanovusXPU 和交换机 CPO私营;推广 12.8 Tb/s scale-up CPO,并声称获得 Tier-1 验证面向多机架 scale-up 的 XPU 和交换机光引擎在明确 XPU / 交换机 CPO 叙事上,仅次于 Lightmatter/Celestial 的最接近创业公司同行
可插拔光模块 / LPO 生态现状方案和过渡替代品Innolight、Eoptolink 和 Accelink 已在出货 800G/1.6T 等级产品让运营方可以推迟封装集成最务实的近期替代品,因为保留热插拔可维护性

截至 2026-05-25 运行日期,这些行汇总 Ayar 在商业上最重要的竞争对手和替代品; 规模信号采用最新公开披露,而不是标准化收入或出货指标。

[CP006, CP008, CP010, CP014, CP019, CP022]
FP001: 竞争定位图

这是一张顺序评分图:横轴是商业化能力,纵轴是与 Ayar 封装内光 I/O 投资逻辑的直接架构重叠度。

分数是作者根据已披露的商业化信号、上市公司规模,以及各平台把光学能力插入计算或交换封装的深度所作的顺序判断。它们不是来源直接报告的数字指标。

[CP006, CP008, CP010, CP014, CP019, CP023]

3.2 上市既有厂商与平台供应商

最强竞争压力来自上市既有厂商,因为每家公司攻击的是 Ayar 价值主张的不同层。Broadcom 是最清晰的交换机层基准:到 2025 年底,它已展示 Meta 测试的 400G-equivalent CPO 端口累计 100 万个链路小时、无链路抖动,并声称光模块功耗比可插拔方案低 65%,这把 CPO 从概念推向生产就绪。Marvell 走的是更偏计算中心的路线,将自有定制 XPU 平台延伸到共封装光学,并主张集成光子可跨多个机架连接数百个 XPU。Intel 的威胁来自另一点:它的 OCI chiplet 同样瞄准共封装光学 I/O,但把激光器直接集成在 chiplet 上,而不是采用 Ayar 的外置激光模式。 Cisco/Acacia、Coherent 和 Lumentum 在封装内光学 I/O 上没有那么一一对应,但它们重要,因为它们已经拥有客户关系、光学制造能力和可转向最高价值光学层的路线图。Acacia 明确营销 AI scale-out 光互连和 3D siliconization。Coherent 和 Lumentum 供应许多 AI fabric 仍然需要的激光器、模块和光学子系统,即便最先进节点逐步迁移到 CPO。Broadcom、Marvell、Cisco、Coherent、Lumentum 和 Intel 的 SEC 文件节奏提醒投资者,Ayar 竞争的不是只有创业公司,还有拥有经审计上市公司资源的供应商。[CP008, CP009, CP010, CP011, CP012, CP013]

与上市既有厂商正面对比
公司相关光学资产市场就绪度证据与 Ayar 的重叠相对 Ayar 的弱点
Broadcom交换机层 CPO 平台Meta 测试过 100 万小时链路无 flap;声称光学功耗降低 65%在交换机层 CPO 上非常高对第三方封装内计算 die 光 I/O 的聚焦较弱
Marvell3D SiPho 引擎 + 定制 XPU CPO 架构CPO 架构可用于下一代定制 XPU 设计;硅光器件现场运行 >10B 小时在 scale-up 计算互连网络上高依赖定制芯片合作模式,而不是商用光 I/O chiplet
Cisco / AcaciaAI scale-out 光学、客户端光学组件、光引擎当前上市公司文件节奏和 Acacia 活跃产品定位作为邻近光学既有厂商为中等公开证据在 scale-out / 客户端光学上更强,封装内计算 I/O 证据较弱
IntelOCI chiplet 和大批量可插拔硅光4 Tbps OCI chiplet;PIC 出货 >8M在光 I/O chiplet 上高集成激光设计不同于 Ayar 可维护外置光源模式
Coherent收发器、AOC、组件、仪器高速光传输系统产品组合很宽作为使能组件和模块供应商为中等直接进入计算封装光 I/O 的证据较少
Lumentum1.6T 可插拔光模块、高功率激光器、DWDM 光源OFC 2026 展示 scale-out / scale-up 和 CPO 激光光源作为可插拔 / LPO 与激光光源对手,威胁中等仍更偏模块和光源,不是完整封装内光学 I/O 栈

本表比较最值得关注的上市在位厂商与 Ayar 最接近的当前产品或平台,而不是比较公司整体业务宽度。

[CP008, CP009, CP010, CP012, CP013, CP014]
FP002: 架构特征对比

Ayar 与最相关的既有厂商或初创对照方在功能覆盖上的对比。

单元格概括各平台在已审阅来源材料中的主导公开定位;"Unknown" 标记来源未明确披露的属性。

[CP008, CP010, CP014, CP015, CP019, CP023]

3.3 私营 CPO 创业公司与相邻挑战者

私营挑战者中,Lightmatter 和 Celestial AI 是两个最可比的战略威胁,因为两者都在向 Ayar 想要触达的同一批超大规模云厂商和加速器受众销售光学 scale-up 叙事。Lightmatter 的路线图覆盖 NPO、OBO、2D CPO、3D CPO 和 3D 中介层,Passage 产品从 12.8 Tbps 扩展到 32–64 Tbps 级系统。其官网披露累计融资 $850 million、估值 $4.4 billion,资本量超过 Ayar,也让它能销售更广的互连平台,而不只是一对 chiplet 加激光器。Celestial AI 同样资本充足,累计融资 $515 million,并把 Photonic Fabric 从封装推到机架级;Marvell 计划收购它,说明超大规模云光学 scale-up 创业公司已经有足够价值,被既有半导体平台吸收。 Ranovus、Scintil、Nubis/Ciena 和 POET 范围更窄,但仍然相关。Ranovus 明确营销面向 scale-up 的 12.8 Tb/s XPU 和交换机 CPO。Scintil 聚焦于 AI scale-up 链路所需的高密度多波长光源,能够攻击 Ayar 当前通过 SuperNova 外置化的激光源层。Ciena 通过 Nubis 域名呈现的共封装光学页面强调低功耗光学引擎和可现场维护的外置光源,这一可维护性叙事接近 Ayar 自身架构。POET 更偏相邻而非直接竞争,但其光学中介层平台瞄准 800G、1.6T 及以上,因此也会争夺 AI fabric 周边相同的封装和互连预算。[CP019, CP020, CP021, CP022, CP023, CP024]

私营 CPO 初创公司与邻近挑战者对比
公司已披露资本 / 状态产品范围与 Ayar 最直接的重叠限制因素
Ayar Labs$370M 累计融资;估值 >$1B封装内光学 I/O 与远程光源直接切入计算 / 交换芯片封装商业设计导入和量产时间披露仍很少
Lightmatter$850M 累计融资;估值 $4.4BNPO、OBO、2D/3D CPO、3D interposer、光引擎scale-up 光互连架构与封装内光子学路线图更宽,可能分散焦点;产品化复杂度高
Celestial AI$515M 累计融资;Marvell 已宣布收购从封装到机架的 Photonic Fabric;16 Tbps chipletscale-up 光互连架构与封装级光子 chiplet交易宣布后,后续大概率受在位平台优先级牵引
Ranovus私营;已审阅来源未见当前公开融资金额12.8 Tb/s XPU 与交换 CPO面向多机架 scale-up 的 XPU 与交换光引擎商业牵引力只做定性描述,未给出具名部署
Scintil Photonics私营;聚焦激光光源面向 AI scale-up 链路的单芯片 DWDM 激光光源切入 CPO 引擎周边的外置光源层范围比 Ayar 窄,不是完整光学 I/O 栈
Nubis / Ciena私营品牌;见于 Ciena 官方 CPO 页面面向 CPO/NPO 的光引擎,配可维护外置激光器低功耗光引擎与现场可维护光源模式更像平台中立的光引擎供应商,而非完整计算裸片 I/O 方案
POET Technologies上市小盘光学中介层供应商面向 800G、1.6T 及更高速率的光学中介层产品争夺 AI 光学封装预算更像相邻封装玩家,不是直接替代封装内光学 I/O
Rockley 既有 IP2023 年进入 Chapter 11 破产重组;IP 以 $20M 出售给 Celestial原硅光子 IP 组合说明投资人愿意买光子 IP,但无法证明独立公司能扛得住资本市场失败削弱纯光子同行估值情绪

资本字段只使用公开披露总额;若已审阅来源看不到融资金额,则只列清晰标注的状态。私营公司收入大多仍未披露。

[CP004, CP021, CP022, CP024, CP025, CP027]
FP003: 头部私营 CPO 初创披露资本规模

Ayar、Lightmatter 和 Celestial AI 披露的累计融资总额,单位为百万美元。

数值采用保留的官方或新闻来源中可见的最新累计融资披露;Ranovus、Scintil 等未披露同业被剔除,而不是估算。

[CP004, CP021, CP022]

3.4 可插拔、LPO 与过渡架构

最容易被低估的竞争对手,是那些并不彻底替代封装边界的架构。Innolight、Eoptolink 和 Accelink 显示,既有可插拔生态在 800G 和 1.6T 上已经很强。Lumentum 和 Coherent 继续推进更高密度模块和激光光源,改善前面板与机架级光互连的经济性。分析师和行业证据一致表明,这些架构不会消失:LightCounting 预计高批量 CPO 部署要到 2027 年,LPO 部署更早启动,800G 及以上可插拔出货仍在增长。Avnet 还明确认为,可插拔模块和 CPO 会共存五到十年,因为可维护性、可配置性和多供应商采购在运营上仍然重要。 对 Ayar 而言,近期任务不是逐端口击败每个可插拔厂商,而是证明在最高性能 AI scale-up 场景中,继续留在封装外的成本、功耗和密度惩罚已经不可接受。如果运营方可以用 LPO、NPO 或高密度可插拔模块推迟这次架构跃迁,Ayar 的采用曲线就会后移。因此,外置激光器的可维护性具有战略意义:它试图在把光学带入封装的同时,保留可插拔模块的一项关键优势。[CP016, CP017, CP018, CP028, CP030, CP031]

过渡架构与切换成本对比
架构代表厂商运营优势运营劣势对 Ayar 的含义
重定时可插拔模块模块厂:Coherent、Lumentum、Innolight、Eoptolink、Accelink可热插拔、多供应商、运维熟悉800G–1.6T 下功耗更高,热约束更紧最密集 fabric 场景的默认基线;Ayar 必须在 TCO 上打赢它
线性驱动可插拔模块(LPO)生态广;LightCounting 和 Avnet 重点提及保留可插拔可维护性,同时削掉部分 DSP 开销光学仍留在封装外早期 CPO 采用中最可信的近期替代方案
近封装光学(NPO)Ciena/Nubis、Lightmatter L-series电气走线更短,密度优于前面板可插拔模块拿不到完整封装级 I/O 收益能满足想降低风险、不愿直接上完整封装内光学的买家
交换层 CPOBroadcom、NVIDIA、Marvell核心 fabric 中密度和功耗收益最高集成复杂,维修负担重若交换机 OEM 自己内化光子能力,Ayar 会受威胁
封装内计算裸片光学 I/OCPO / NPO 方案:Ayar、Intel OCI、Celestial AI、Lightmatter M-series最大化缓解芯片边缘 I/O 压力,并提升带宽密度认证与封装负担最高只有更深封装切入值得承担复杂度,Ayar 才能赢
自建定制光学方案NVIDIA 和超大规模云厂商牵头的生态完全控制架构和供应链需要深厚的光子、封装和制造能力最大客户可能把最有价值的光学层内化

对比重点放在可维护性和封装深度,而非标价;私营 CPO 平台的具体合同价格未公开披露。

[CP028, CP030, CP031, CP032, CP033, CP034]

3.5 差异化、护城河与切换成本

Ayar 真正的差异化是光学插入深度。公司试图利用主要 XPU 厂商熟悉的制造和封装流程,把光学 I/O 直接放进计算或交换封装,同时让光源保持远端且可维护。这不同于大多数可插拔厂商,也不同于 Acacia 偏客户侧光模块的重点,更不同于 Intel 集成激光的 OCI chiplet。若 Ayar 赢下一个插槽,切换成本会很可观:封装协同设计、光纤拓扑、热认证和远端激光架构都会成为平台定义的一部分,之后再切回前面板光模块或换成另一家光学 I/O 供应商,成本都会很高。 这条护城河并非不可攻破。更大的既有厂商有足够资本和制造深度来补齐功能差距,NVIDIA 也已经证明战略买方可以在交换机层内部化光子能力。资本市场历史同样是警示:Rockley 的 Chapter 11 重组和后续 IP 出售说明,硅光子专门厂商可能在规模经济到来前失败。因此,Ayar 拥有一条狭窄但有价值的护城河:当客户今天就想要直接封装内光学 I/O 和可维护外置激光器时,它处在领先位置;但护城河能持续多久,取决于它能否在市场落到既有厂商自有平台,或更方便运营的 LPO 与可插拔混合方案之前,把技术认证转化为设计定点。[CP002, CP003, CP004, CP036, CP040, CP042]

护城河耐久性与威胁评分
威胁 / 护城河议题证据严重度重要性尽调追问
封装深度差异化Ayar 将光学 I/O 直接接入主要 XPU 和交换芯片厂商已使用的封装流程中等正向这是 Ayar 相比可插拔模块和许多 NPO 对手最清晰的技术护城河哪些客户已在量产封装设计中认证 TeraPHY?
外置激光器可维护性Ayar 的远程光源,以及 Ciena/Nubis 的外置光源叙事,都保留了一部分可维护性优势中等正向激光器可维护,可以削弱市场对 CPO 的一项主要反对意见外置激光器多久需要更换,由谁更换?
NVIDIA 交换层内化Spectrum-X Photonics 将在 2026 年把自研硅光子和 CPO 带入以太网交换高度负向交换机 OEM 可以在 fabric 层绕过商用光学 I/O 供应商如果交换机客户把光学内化,Ayar 还控制任何差异化层吗?
在位厂商资金实力挤压Broadcom、Marvell、Cisco、Intel、Coherent 和 Lumentum 都背靠上市公司资产负债表高度负向在位厂商能出钱熬更长认证周期,也能顶住更慢采用哪些现金和合作伙伴承诺能让 Ayar 熬过 CPO 爬坡延迟?
LPO 与可插拔模块延续分析师来源预计,LPO 会早于大批量 CPO 爬坡,可插拔模块仍将在许多链路占主导高度负向架构切换变慢,会拉长 Ayar 销售周期,并可能压缩早期 TAM哪些 AI 链路在功耗或密度上已经糟到必须现在采用封装内光学?
板块资本市场怀疑Rockley 通过 Chapter 11 破产重组,随后出售核心硅光子 IP中等负向纯光子公司可能在规模经济到来前就断掉融资如果不再做一轮大额融资,Ayar 路线图中有多少已筹到量产所需资金?

严重度标签是作者基于已审阅证据作出的判断,用于排列尽调优先级,并不暗示概率预测。

[CP002, CP006, CP033, CP036, CP038, CP042]
FP004: 护城河与切换成本快照

简要展示 Ayar 竞争位置周围的结构性优势与风险。

[CP002, CP033, CP038, CP042, CP048]

3.6 图表与证据

Chapter 04

04财务情况

4.1 收入模式与变现

Ayar 的公开变现信号仍更像硬件认证和早期商业采用,而不是已经披露的规模化收入引擎。公司产品页围绕 TeraPHY 光学 I/O chiplet 和 SuperNova 光源来描述变现;评估套件优先给已经承诺长期商业协议的客户;2022 年融资材料称 Ayar 已完成首批按合同批量商业出货。后续材料把重点转向不断增长的客户基础,以及与客户路线图对齐的高批量制造,但公开来源仍未给出标价、已签订单积压、单位出货量和收入确认政策。这意味着,公开信息下的投资判断者能看到需求面包屑——样品、设计定点、长期商业意向和首批出货——却无法把它们转化为可审计的收入、毛利或 CAC/回本周期模型。实际看,Ayar 仍像一家早期商业化硬件公司,证明点更多是伙伴采用和制造就绪,而不是已披露的财务产出。[CI001, CI020, CI021, CI022, CI023, CI024]

收入来源与变现
来源机制单位 / 合同基础当前公开状态收入质量尽调追问
TeraPHY 与 SuperNova 产品销售向计算和互连项目销售光学 I/O chiplet 与光源按 chiplet / 系统或项目合同计价;价格未披露公开页面将两款产品定位为商业化产品,但未披露 ASP 或已订收入商业化早期,可见度低提供按产品线、客户项目和季度拆分的经审计收入
评估套件与工程样品向潜在量产客户分配认证套件和样品硬件套件 / 样品价格未公开评估套件优先分配给签有长期商业协议的客户仍在认证阶段,难以预测披露样品到量产的转化率、支持负担和平均合同规模
联合开发与商业化工作与战略支持方开展伙伴牵头的集成、共同设计和生态建设里程碑或工程付款条款未披露HPE、NVIDIA、Intel/DoD 等合作已公开,但会计处理未公开质量混合;可能推动采用,但未必是经常性产品收入说明 NRE、里程碑收入,以及任何可报销工程工作
客户路线图下的批量商业供货随长期商业协议和客户制造计划扩大硬件出货多年供货协议,单价条款未披露Ayar 披露 2022 年首次批量商业出货,后来又围绕高批量制造来叙述 Series D潜在吸引力高,但时间表和 backlog 仍不透明提供 backlog、合同约定量、转化里程碑和收入确认政策

公开变现线索来自产品、评估套件和出货披露,但价格和已确认收入仍未公开。

[CI001, CI020, CI021, CI022, CI023, CI024]
FI003: 收入模型桥

公开可见的商业化路径从送样和评估走向潜在规模收入;桥的末端仍存在披露缺口。

这座桥是定性分析,因为 Ayar 披露了商业化里程碑,但没有披露里程碑之间的财务转化率。

[CI001, CI006, CI022, CI023, CI024, CI035]

4.2 融资轮次与资本结构

Ayar 的资本结构以股权为主,已披露主要轮次合计约 $369 million,覆盖 Series A、B、C/C1 和 D。2024 年 Series D 增加 $155 million,并公开使用估值超过 $1 billion 的表述,但没有给出精确投后数字。早期轮次呈阶梯式:2018 年 $24 million Series A、2020 年 $35 million Series B、2022 年 $130 million Series C,以及 2023 年 $25 million C1 延伸轮。股权之外,Ayar 还披露了一笔 $3 million SVB 定期贷款,以及与 DARPA 和 DoD 项目绑定的政府支持资金。投资判断含义是,Ayar 历史上通过反复风险融资加选择性非股权支持来资助商业化,而不是靠公开可见的经营现金生成。2026 年的关键财务问题因此不是 Ayar 能否融资——历史记录表明可以——而是商业化能否降低对下一轮融资周期的依赖。[CI003, CI004, CI005, CI006, CI007, CI008]

已披露主要股权融资轮次
轮次交割日期金额(USD M)领投 / 叙事重要参与方公开信息要点
Series A 轮2018-11-0724Playground Global 支持的 Series A 轮投资人:Founders Fund、GlobalFoundries、Intel Capital用于首个光学 chiplet 平台商业化的资金
Series B 轮2020-11-0535Downing Ventures 和 BlueSky Capital 联合领投投资人:Applied Ventures、Castor Ventures、SGInnovate、Founders Fund、GlobalFoundries、Intel Capital、Lockheed Martin Ventures、Playground Global延展商业化和国际布局
Series C 轮2022-04-26130Boardman Bay Capital Management 领投投资人:HPE Pathfinder、NVIDIA、Applied Ventures、GlobalFoundries、Intel Capital、Lockheed Martin Ventures、Founders Fund、Playground Global商业化、扩产和可靠性认证
Series C1 轮2023-05-2425Capital TEN 领投投资人:VentureTech Alliance、Boardman Bay、IAG Capital Partners、NVIDIA、Tyche Partners、Applied Ventures、 GlobalFoundries、HPE Pathfinder、Intel Capital、Lockheed Martin Ventures将 Series C 扩展至 $155M,并加速执行生成式 AI 路线图
Series D 轮2024-12-11155Advent Global Opportunities 和 Light Street Capital 领投投资人:AMD Ventures、Intel Capital、NVIDIA、3M Ventures、Autopilot、Applied Ventures、Boardman Bay、GlobalFoundries、 Lockheed Martin Ventures、Playground Global、VentureTech Alliance推动高批量制造;公司称估值升至 $1B 以上

本表覆盖本章审阅的主要公开披露定价融资轮,不包括未披露种子资本和非股权工具。

[CI003, CI004, CI005, CI007, CI009, CI010]
FI001: 融资时间线

从 Series A 到 Series D 的关键公开融资和政府支持资本事件。

时间线聚焦本章的规范资本事件,并把股权融资与已披露的政府 / 债务里程碑放在一起,展示融资节奏。

[CI003, CI007, CI009, CI010, CI012, CI013]
FI002: 资本强度 / 现金流图

公开披露的主要定价股权轮次,构成 Ayar Labs 截至 Series D 约 $369M 的公开股权基础。

瀑布图只汇总主要已披露定价轮次;未披露的种子资本和单独披露的非股权支持不计入。

[CI003, CI007, CI009, CI010, CI012, CI045]

4.3 投资者基础与战略资本

Ayar 的股权结构表值得注意,不只是因为规模,也因为构成。半导体和系统生态支持方——Intel Capital、GlobalFoundries、NVIDIA、AMD Ventures、HPE Pathfinder 和 Lockheed Martin Ventures——与 Advent Global Opportunities、Light Street Capital、Boardman Bay、Capital TEN、Playground Global、Founders Fund、BlueSky Capital 和 Downing Ventures 等财务投资人并列。这种组合有正面意义:战略投资人可以验证产品相关性、联合开发、制造准入和设计导入潜力。HPE 把投资与联合硅光子工作绑定,GlobalFoundries 把资本与制造合作配对,NVIDIA 则从合作扩展到跟投。反面是,战略资本可能掩盖公开单位经济的缺位,因为投资人买的也许是生态期权,而不是近期财务效率。公开证据因此支持把 Ayar 的投资者基础视为商业化资产,但不能把它当作硬经营披露的替代品。[CI005, CI011, CI029, CI030, CI031, CI032]

投资人基础与战略资本角色
投资人首次公开进入类型叙事中的角色公开证据
Playground GlobalSeries A(2018)财务 / 深科技 VC为光学 chiplet 扩产提供早期商业化资金和董事会层面支持在 Series A 新闻稿中被引用
Founders FundSeries A(2018)财务 / 深科技 VC与半导体战略投资方一道提供早期风险资本Series A 新闻稿中具名
Intel CapitalSeries A(2018),后续跟投至 Series D战略半导体投资人证明光互连切中计算生态需求在 Series A 和 Series D 新闻稿中具名
GlobalFoundries2020 年前已有战略关系,Series B 时已披露为投资人战略晶圆代工伙伴制造合作加未披露投资GF 合作新闻稿和融资公告
Lockheed Martin Ventures战略投资(2020)战略防务投资人为 AI/HPC/波束成形和防务用例提供商业化支持Lockheed 投资新闻稿
HPE PathfinderSeries C / 合作(2022)战略系统投资人联合 HPC/AI 架构工作和客户接触支持HPE 合作新闻稿
NVIDIASeries C 合作与投资(2022)战略 AI 平台投资人未来规模 AI 架构合作;2023 年增持,并参与 2024 年融资NVIDIA 合作、Series C 和 Series C1 新闻稿
AMD VenturesSeries D(2024)战略半导体投资人显示 GPU 生态对 AI 基础设施光学 I/O 的兴趣Series D 新闻稿和后续报道
Advent Global Opportunities 投资人Series D(2024)财务成长股权投资人后期 scale-up 资金和董事席位Series D 新闻稿和 Washington Technology 报道
Light Street CapitalSeries D(2024)财务 crossover 投资人为批量制造阶段提供大额融资支持Series D 新闻稿和 Crunchbase News 报道

各行突出具名的主要战略投资人和领投财务投资人,而非 Ayar Labs 完整法律股权结构表或每一个 SPV。

[CI005, CI011, CI029, CI030, CI031, CI032]

4.4 单位经济、资本充足性与披露缺口

Ayar 公开财务档案中最弱的一块是经营披露。已审阅来源没有提供收入、ARR、毛利率、EBITDA、现金余额、月度烧钱、现金跑道或管理层确认的员工数,因此无法支撑效率比率。公开的是方向性单位经济信号:光学 I/O 的营销重点是带宽、延迟和功耗节省;公司披露过长期商业协议、首批商业出货和招聘计划;Series D 资金用途也围绕高批量制造展开。这些线索暗示商业化支出重、资本密集度高,但不足以让外部投资人把 Ayar 转成可靠的收入桥或现金跑道模型。结果是,关键投资判断指标本该出现的地方,全是 null。对一家后期风险投资硬件公司而言,这不致命,但确实使下一次融资触发点无法仅靠公开证据精确建模。[CI006, CI035, CI036, CI037, CI038, CI039]

公开财务画像与单位经济可见度
指标公开数值披露状态重要性尽调追问
年收入 / ARR已审阅来源未公开披露无法核验规模与销售产能提供季度收入、backlog,以及按产品和客户阶段拆分的结构
毛利率未公开披露硬件毛利率路径决定风险投资级回报账能否成立提供按产品、工程项目与量产拆分的毛利率
EBITDA / 经营利润率未公开披露外部看不到经营杠杆或固定成本吸收提供调整后 EBITDA 调节表和按职能拆分的 opex
现金余额未公开披露流动性无法与制造计划或合同爬坡挂钩提供月末现金、受限现金和债务余额
月度烧钱速度未公开披露没有 burn,就无法建模下一轮融资时间和下行情景提供按 capex 和 opex 拆分的月度现金消耗
现金跑道未公开披露投资人无法测试 Series D 是否足以支撑公司穿过量产爬坡提供基准、下行和拉伸情景下的跑道假设
与财务模型挂钩的员工数招聘计划公开,但财务语境下没有管理层确认的员工数挡住人均收入和人均 burn 校验提供按职能拆分的当前 FTE 数,以及按季度招聘计划

未知指标有意显示为 null,因为 Ayar Labs 是私营公司,已审阅公开来源未披露这些指标。

[CI035, CI036, CI037, CI038, CI039, CI040]
资本投放与充足性信号
维度公开信息置信度承销解读尽调追问
当前现金余额公开资料无法判断 Series D 后的流动性要求提供当前现金、受限现金和最低现金契约细节
月度烧钱速度与现金跑道无法仅凭公开证据推演下一轮融资时间要求提供月度烧钱速度、分情景现金跑道,以及产能爬坡期间的资本开支节奏
Series D 轮披露资金用途配合客户路线图的大批量制造资金看起来主要投向规模化,而不是保守地补强资产负债表要求提供董事会批准的资金用途桥表和逐季度制造计划
Series C 和 C1 轮披露资金用途商业化、供应链认证、产能扩大、路线图提速和招聘早期轮次同样在资助规模化,而不是靠自有资金扩张要求提供历史资金用途桥表和前几轮剩余里程碑
已披露的非股权资金抵消$15M KANAGAWA OTA,加 $3M SVB 定期贷款;PIPES 金额未披露能部分抵消 R&D 和制造支出,但不足以摆脱创投资金依赖要求提供补助提款时间表、成本分担义务和债务摊还安排
可能触发下一轮融资的节点从送样和早期出货转向可重复的大批量部署下一轮融资叙事似乎由经营里程碑主导,而不是公开盈利指标要求提供客户量产里程碑、内部收入计划和董事会融资触发条件

本表强调公开可见的充足性信号,而不是推断现金余额;未披露指标按设计保留为 null。

[CI006, CI008, CI013, CI015, CI036, CI045]

4.5 非稀释性资金与资本效率结论

Ayar 的非稀释性资金故事重要到不能忽略,但还不足以补上投资判断缺口。公司把源头追溯到 DARPA 资助研究,2019 年披露了 DARPA PIPES 补助,后来又获得 $15 million DoD Project KANAGAWA OTA。这些项目很可能帮助技术成熟降风险,但并没有消除对风险资本的依赖。围绕共封装光学的公开反向评论也有财务意义:行业分析师和基础设施观察者仍在质疑可维护性、供应商锁定、维修复杂度以及大规模部署时点,常把 CPO 描述为必然发生,但并非迫在眉睫。这很重要,因为 Ayar 的资本基础推进速度快于公开经营披露。财务结论因此是:战略相关性和融资能力为正,但投资判断仍需谨慎;Ayar 看起来能融到钱,但财务透明度还不够。[CI014, CI015, CI016, CI017, CI018, CI019]

赠款、OTA 与非股权资本
项目 / 工具资助方公开金额(USD M)状态财务意义
DARPA 支持的创始研究脉络DARPA公司称,创始人的突破来自 DARPA 资助的十年研究解释了非稀释研发根基,但没有给出公司层面的现金金额
PIPES 赠款 / Intel PIPES 项目入选DARPA,通过 Intel 牵头的 PIPES 项目Ayar 2019 年披露 DARPA PIPES 赠款,另披露入选 Intel 该项目支持产品化和互连标准采用,但金额未披露
Project KANAGAWA OTANSWC Crane / OUSD(R&E) Trusted & Assured Microelectronics 计划152022 年获授,形式为多年期 prototype OTA为防务转化和本土制造提供已披露的重要非稀释资本
SVB 灵活提款定期贷款Silicon Valley Bank32019 年宣布,用于资本支出和制造支出非股权制造资本,但更像债务,不是经营收入

本表覆盖已审阅来源中识别出的公开披露非股权资本工具;若干 DARPA 资金项的金额仍未披露。

[CI013, CI014, CI015, CI016, CI017, CI018]
Chapter 05

05产品与技术

5.1 产品组合与客户要完成的任务

Ayar 的产品线异常集中:公司卖的是两段式光学 I/O 栈,而不是一大菜单的光模块。TeraPHY 是封装内光学 I/O chiplet,贴着计算硅片放;SuperNova 是远端多波长激光源,通过光纤给这些 chiplet 供光。从客户任务看,这对组合要替代耗电的电链路和可插拔光链路,解决 scale-up AI 系统的瓶颈,之后再延伸到池化内存和其他可组合架构。公开材料一直把 AI scale-up 定位为第一个商业落点,把远端或分解式内存定位为第二个落点。这让 Ayar 看起来不太像通用光模块供应商,更像未来 xPU 封装和机架 fabric 的组件供应商。[CE001, CE002, CE003, CE004, CE005, CE006]

产品模块 / 资产矩阵
模块 / 资产用户 / 工作流公开状态 / 成熟度关键差异点尽调缺口
TeraPHY 现有光 I/O chipletGPU / CPU / ASIC 封装间或封装到机架 I/O当前公开世代;约 2 Tbps 级双向,现用 AIB、下一步 UCIechiplet 形态,光互连覆盖毫米到公里距离付费 SKU 结构、ASP 和出货量未披露
TeraPHY 下一代光重定时器AI scale-up 互连架构和封装间光链路2025 OFC / 验证阶段;论文给出 8.192 Tbps 汇总结果,验证博客给出 8 Tbps首个 UCIe 光重定时器主张,以及 16 波长微环链路需要客户认证、良率和流片证据
SuperNova 远程光源为多个 chiplet 和端口供光的共享激光源产品页已上线;16 个波长 / 16 个端口 / 256 个通道 / 16 Tbps 双向外置、现场可更换,符合 CW-WDM MSA 的激光源需要 MTBF、替换成本和模块 BOM 明细
带连接器的 V-groove 封装选项大批量封装的组装 / 测试流程2025 IEEE/OFC 封装论文无源光纤贴装,加已知良品的带连接器 chiplet 流程未公开组装良率、插入损耗或返工指标
机架级 Wiwynn / ELSFP 参考系统超大规模 AI 机架集成与热管理OFC 2026 演示 / 可量产营销表述把 Ayar 从组件演示推向机架级液冷架构没有公开具名量产部署或运行出货量数据

Ayar 公开技术栈仍然较窄,但围绕光引擎 + 远程激光源组合做了纵向整合;多数部署证明由生态伙伴牵引,而不是面向所有客户的正式可用性证明。

[CE001, CE002, CE004, CE021, CE030, CE031]
FE001: 从封装到机架的光 I/O 系统流

Ayar 的 TeraPHY 小芯片和 SuperNova 光源如何连接计算封装、远程内存和机架级互连网络。

[CE001, CE005, CE011, CE013, CE019, CE030]

5.2 架构与集成模式

这套架构是以 chiplet 为中心的光学 fabric。电侧,当前公开代际仍使用 AIB;Ayar 的下一代叙事则明确绑定 UCIe 和更广泛的多供应商 chiplet 生态。光侧,系统依赖 CW-WDM 式多波长光源,以及能扛住真实封装流程的光纤贴装方法。Ayar 自身材料称,UCIe 加光学 I/O 是让内存离开计算封装、同时不牺牲带宽的关键;OCP 和 IEEE/OFC 材料则强调 V-groove 无源贴装和已知良品 chiplet 流程,这是从实验室演示走向制造的实际桥梁。重要的尽调结论是,Ayar 的差异化不只在光子裸片,而在跨标准、封装设计、光纤贴装和外置光源的完整集成配方。[CE011, CE012, CE013, CE014, CE015, CE018]

集成、标准与依赖表
层级 / 接口当前公开实现路线图 / 方向依赖技术风险
电气 die-to-die 接口当前公开世代采用 AIB迁移至兼容 UCIe 的光重定时器客户 chiplet 生态和 PHY 互操作性均衡、验证负担和标准节奏
光源接口CW-WDM MSA 光源模型;当前一代参考 8 波长,SuperNova 产品页为 16 波长更高波长数远程光源路线图激光 die 供应商和光功率预算共享故障域和插入损耗管理
封装模式多芯片模块 / 中介层 / 先进封装集成通过 UCIe 2.0+ 支持 2.5D 和 3D chiplet 封装晶圆厂、OSAT 和基板伙伴良率损失和热耦合可能吃掉性能收益
光纤贴装V-groove 无源贴装和带连接器 chiplet 流程用于 HVM 的已知良品带连接器 chiplet组装工艺控制和测试覆盖光学对准容差和维修复杂度
系统用例先 AI scale-up,其次远程内存,机架互连架构开始出现机架外互连架构和多机架 scale-up交换机、内存封装和机架系统伙伴软件、拓扑和维护模式仍不成熟

Ayar 的差异化取决于完整集成配方;任何单层失败都可能主导封装级经济性。

[CE011, CE012, CE018, CE019, CE021, CE030]
FE004: 互连选项矩阵

Ayar 的小芯片光互连相对于既有电互连 / 可插拔链路和交换机中心 CPO 设计的定位。

定性矩阵综合了官方产品主张和反向分析师评论;语气表示相对位置,不是绝对分数。

[CE008, CE031, CE040, CE041, CE042, CE043]

5.3 性能边界与验证状态

Ayar 的公开性能叙事现在横跨三层:当前公开产品描述中的 2 Tbps 级能力、4 Tbps 级 scale-up 示例,以及仍在 DVT / 认证中的 8 Tbps 级验证引擎。头部优势仍是熟悉的几项——更高带宽密度、更低延迟、更低每 bit 功耗、比重铜方案更长的传输距离——但更有价值的信号是,Ayar 现在发布验证细节,而不只给顶层营销倍数。2024 年验证文章描述了五天 UCIe 链路测试、光学裕量结果、热循环结果、高瞬态温度仿真,以及低于 25 ns 的端到端 UCIe-over-optics 延迟。这不能证明已经具备广泛量产成熟度,但确实把产品从“科研项目”推进到在类似封装的热和链路条件下具备可运行证据的阶段。[CE002, CE008, CE014, CE016, CE017, CE020]

与电互连的公开性能对比
指标Ayar 光 I/O电 SerDes / 可插拔模块公开优势限制
带宽密度按 Ayar 营销口径高 5x–10x;现有 chiplet 为 2 Tbps 级,下一代演示为 4–8 Tbps长距离下,铜线 / 可插拔架构很快吃掉板边资源和重定时器预算同等封装预算下,可用的封装外带宽更高比例来自公司声明,不能跨所有工作负载做同口径比较
每 bit 能耗SC23 级公开演示中 <5 pJ/bit;效率提升声明为 4x–8x传统长距离电互连和可插拔路径在可比距离下通常高得多更低 I/O 功耗是 AI scale-up 经济性的核心Ayar 未公布完整系统级功耗拆分或冷却开销
时延产品声明约每个 chiplet 5 ns + TOF;验证中端到端 <25 ns50 Gbps 以上电链路通常需要 FEC,并累积串行化 / retimer 延迟光链路能守住更低的封装间时延公开数据多来自实验室和演示配置
覆盖距离公开材料显示从毫米到 2 km铜互连通常从封装 / 板上英寸级延伸到电缆形态下米级一套架构可覆盖 die、板、机架和远程内存距离长距离仍取决于激光预算、光学余量和封装质量
维护模式外置远程激光器可在封装外更换可插拔模块仍最易更换;纯 CPO 引擎更难维护Ayar 通过把激光源外置,保留了一部分可插拔性一个远程光源故障可能同时影响多个通道

对比混合了公司声明、验证结果和分析师观察,应按方向性判断阅读,而不是当作统一基准测试套件。

[CE002, CE003, CE008, CE017, CE027, CE040]
FE002: 公开性能边界

选取的公开上限指标,覆盖当前世代、下一代演示和路线图披露。

这张图混合了当前世代公开数值、2025 年技术论文结果,以及 Ayar 已验证的下一代引擎披露。它不是单一基准配置。

[CE002, CE004, CE008, CE017, CE020, CE027]

5.4 制造工艺与生态系统

制造能力是 Ayar 公开叙事里最强的一块。公司长期把 TeraPHY 与 GlobalFoundries 的单片硅光子工艺绑定在一起, 外部技术资料也强化了这一点:GF Fotonix / 45SPCLO 原本就是为了把电子与光子集成在同一个 45 nm SOI 平台上, 并纳入光纤贴附等封装能力。同样重要的是,Ayar 围绕该工艺拼出了一个可见生态:HPE 面向未来 Slingshot 网络, NVIDIA 面向 AI 互连开发,Wiwynn 负责机架级系统框架,Alchip/TSMC 方向的参考设计则指向先进 AI 插槽。 这个生态是重要背书,但也是一张依赖关系图。Ayar 卖的不是可直接替换的可插拔模块;它要求客户信任一整套堆栈, 从晶圆代工工艺、封装组装、激光器来源,到互操作标准和系统集成商。[CE018, CE019, CE028, CE029, CE030, CE031]

FE003: 产品和成熟度时间线

从 DARPA 源头研究到机架级产品化叙事的公开里程碑。

最后一个节点代表 Ayar 称其必须赶上的认证窗口,用于支持后续超大规模爬坡;它不是有保证的收入转化日期。

[CE016, CE028, CE030, CE033, CE034, CE035]

5.5 IP 护城河、路线图与 TRL

Ayar 的公开轨迹把 DARPA 资助的学术起点、现有专利和最新会议论文串在一起,因此 IP 和成熟度都强过典型光子初创公司。 POEM 和 PIPES 勾勒了 WDM、低能耗光链路和封装级光 I/O 的研究脉络。当前专利清单显示,Ayar 保护的不只是基础光引擎概念, 也包括池化内存、远程内存、面向交换的光 I/O chiplet 等相邻架构。近期技术论文同时覆盖 UCIe 光重定时器路径, 以及如何把已知良品的带连接器 chiplet 做进封装。按 TRL 看,Ayar 已经越过纯实验室研究,但还没到大规模超大规模云部署: 演示加验证阶段可信,也足以支撑生态协同设计;可到 2027 年并进入 2028 年,仍要看客户认证、良率学习和供应链执行。[CE020, CE021, CE033, CE034, CE035, CE036]

IP、专利与发表表
资产 / 发表公开记录日期 / 状态具名发明人 / 作者重要性
由单片封装内光 I/O 支持的池化内存系统美国专利 125679102026 年授权发明人:Roy Meade, Vladimir Stojanović, Chen Sun, Mark Wade, Hugo Saleh, Charles Wuischpard保护光扇出和内存池化架构,而不只是基础链路
面向以太网交换机的低功耗光 I/O chiplet(TeraPHYe)美国专利 125679202026 年授权发明人:Roy Meade, Vladimir Stojanović, Chen Sun, Mark Wade, Hugo Saleh, Charles Wuischpard表明 Ayar 将光 I/O 延伸进交换芯片,而不只做计算封装
由单片封装内光 I/O 支持的远程内存架构美国专利申请 20210258078 / 17/175677已公开申请;核心专利族仍有相关性发明人:Meade, Stojanović, Sun, Wade, Saleh, Wuischpard将路线图锚定到池化 / 远程内存和 HBM 扇出
面向 AI scale-up 互连架构的 UCIe 光 I/O 重定时器 + 带 V-groove 的连接器化光 I/O chipletOFC 2025 摘要加 IEEE/OFC 发表2025 年发表Ayar 工程团队和共同作者显示 Ayar 持续发表技术成果,重点放在互操作性和大批量封装

IP 线索最强的地方,是 Ayar 把产品主张连到系统架构:内存扇出、交换机变体和封装方法。

[CE036, CE037, CE038, CE039]
路线图与 TRL 表
里程碑公开证据阶段 / TRL时间含义
DARPA POEM / 早期光子学项目根基DARPA POEM 页面TRL 2–32020 年前WDM、低能耗光链路和 CMOS 光子集成的科学基础
PIPES 和 Stratix 10 光互连演示DARPA + Laser Focus 报道TRL 4–52020首个 2.56 Tbit/s 级封装内 chiplet 证明,并从研究跃迁到封装演示
GF / HPE / NVIDIA 生态成形官方伙伴公告TRL 52022表明晶圆厂、互连和 AI 生态玩家愿意围绕 Ayar 共同设计
SC23 4 Tbps Agilex 演示Ayar 官方新闻稿TRL 62023验证了共封装 FPGA 路径和公开的低于 5 pJ/bit 目标
8 Tbps 验证 + OFC UCIe 重定时器 + 连接器化 chiplet 论文Ayar 验证博客 + OFC / IEEE 论文TRL 6–72024–2025将叙事从演示带宽推进到验证、互操作性和可制造性
OFC 2026 机架级 Wiwynn 和 ELSFP 演示Ayar OFC 2026 页面TRL 72026展示可部署的机架集成和热管理框架
支撑 2028 年客户爬坡的认证关卡媒体:The Next PlatformTRL 7–8 目标2027–2028商业规模化仍取决于认证、良率学习和供应链执行

公开证据支持 TRL 稳步推进,但还没有形成广泛具名的量产部署节奏。

[CE016, CE028, CE030, CE033, CE034, CE035]

5.6 技术风险与尽调优先级

最难的问题已经不是“Ayar 能否演示光 I/O?”,而是“客户能否维护、认证并规模化采购?”反向分析师观点在这里很有用: LightCounting 仍认为 LPO 会早于 CPO 落地,Lightwave/CIR 则指出,非超大规模运营商仍担心良率、热表现、维修模式和供应商锁定。 Ayar 的外置激光器策略有利于可维护性,因为激光器可以在封装外冷却和更换;但它也集中故障域,并带来新的插入损耗和供应依赖。 与此同时,公开披露仍缺少投资判断最需要的数字——封装良率、激光器 MTBF、已定价 SKU、积压订单和具名量产部署。 因此,本章给出有利的技术方向判断,但在这些私下尽调项开放之前,对近期商业化时间点只给中等置信度。[CE040, CE041, CE042, CE043, CE045, CE048]

Chapter 06

06客户情况

6.1 客户画像与理想客户分层

Ayar Labs 的公开客户图谱很特殊:它仍更像设计导入生态,而不是成熟的已签收入账户清单。公司把光 I/O 定位在 AI 训练和推理基础设施里; 直接“买方”往往不是传统采购部门,而是 ASIC、GPU、网络或机架级架构团队。公开具名的交易对手因此混合了几类角色: DARPA 和防务项目是验证资金方;HPE、NVIDIA、Intel、GlobalFoundries、Lockheed Martin 是战略合作方和投资人; GUC、Alchip、Wiwynn 则为未具名的超大规模云或云买方打开市场路径。Ayar 自己在 2024 年融资稿里称, 公司正把大规模制造与客户路线图对齐,并与 Tier 1 客户开展令人兴奋的接触,但仍未点名这些账户,也没有披露多少在付费、多少还在评估。 这意味着本章的核心投资判断不是客户兴趣是否存在——显然存在——而是公开证据能否证明可复制的量产部署。今天还不能。[CU001, CU002, CU025, CU026, CU036]

公开客户画像与理想客户细分
细分买方 / 用户 / 付款方公开证据公开接触阶段尽调含义
DARPA 与国防 R&D 项目项目经理 / 研究团队 / 政府 R&D 预算POEM、PERFECT 和 PIPES 勾勒创立问题和受资助验证路径受资助验证和技术牵引验证锚点强,直接经常性收入证据弱
GPU / 加速器平台厂商架构团队 / 数据中心产品团队 / 平台厂商NVIDIA 合作;Intel FPGA 演示和 Intel Capital 支持设计合作和集成证明证明与计算路线图相关,不证明采购量
HPC / AI 网络 OEM互连架构师 / 系统设计师 / HPEHPE Slingshot 合作及联合客户接触多年共同开发公开资料中最强的具名 OEM / 渠道路径
晶圆厂与制造生态晶圆厂 BD / 封装团队 / GF 和供应商GF 在 45 nm 硅光子上的共同开发和投资制造生态锚点降低供应风险,但形成单一伙伴依赖
国防系统集成商国防架构师 / 平台工程师 / Lockheed 和 DoD 项目Lockheed 投资和传感系统合作受资助设计导入和方案塑形重要的军民两用渠道,但采购时间不透明
ASIC / 先进封装服务封装架构师 / ASIC 交付团队 / 间接触达终端超大规模云厂商GUC 和 Alchip 伙伴关系,面向 >100 Tbps/XPU 和一线 AI 设计进入市场路径赋能间接路径意味着终端客户名称继续隐藏
机架 / 云基础设施建设方机架架构师 / 云基础设施工程 / 间接触达超大规模云运营商Wiwynn 机架级 AI 系统合作和私下客户简报预览和早期方案营销公开信息中最接近云部署的一步,但仍处于 GA 前

公开细分混合了客户、伙伴和资助方角色,因为 Ayar 没有公布清晰的客户名单。本表反映 Ayar 被拉入系统的公开需求入口,而不是已披露的经常性收入账户清单。

[CU001, CU006, CU010, CU012, CU015, CU020]
FU001: Ayar 公开商业化流程——从研究验证到超大规模云厂商设计导入

公开进入市场路径从 DARPA 支持的验证和战略投资者触达开始,再通过送样和演示进入封装伙伴、机架集成商,以及仍未具名的终端客户。

该流程综合了 Ayar 似乎遵循的公开渠道顺序;它不是公司披露的销售漏斗,也不暗示各阶段之间存在已知转化率。

[CU005, CU010, CU020, CU022, CU023, CU026]

6.2 已公开具名交易对手与证据质量

最可信的具名公开证据,是那些把 Ayar chiplet 拉进真实系统语境的合作。HPE 是最清晰的 OEM / 渠道伙伴, 因为双方宣布围绕未来 Slingshot 世代和联合客户接触开展多年合作。Intel 是最清晰的技术验证伙伴, 因为 Ayar 与 Intel 在 DARPA PIPES 下合作,之后又在 Supercomputing 2023 展示了 4 Tbps 光互连 FPGA。 Lockheed Martin 是最清晰的防务侧集成商,因为它既投资了 Ayar,也宣布了面向 DoD 应用的感知系统合作。 GlobalFoundries 与其说是“客户”,不如说是关键制造锚点;但它愿意投资,并公开围绕封装内客户 ASIC 共同开发,仍然是有意义的客户就绪度证据。 NVIDIA 的角色很强,但更模糊:公司参与未来 AI 架构合作并两次投资,可公开记录仍没有走到具名部署。综合来看,这些证据强过标识墙, 但弱于已披露的经常性量产合同。[CU003, CU004, CU006, CU008, CU010, CU012]

具名客户证明表
交易对手客户旅程中的角色已披露用例公开阶段证据质量限制
NVIDIA战略投资方和架构合作方未来 AI 和 HPC 光 I/O 架构路线图合作官方证据强,且有行业媒体印证无具名部署或采购承诺
Intel演示伙伴、PIPES 参与方、投资方封装内光 FPGA 和 PIPES 验证集成演示 / 验证官方证据强,且有 DARPA 印证未披露生产合同
HPEOEM / 渠道合作方和投资方未来 Slingshot 光网络和联合客户接触多年共同开发官方证据强,且有伙伴证明未公开采购量
GlobalFoundries制造伙伴和投资方面向客户 ASIC 集成的 45 nm CMOS / 硅光子制造生态和共同开发官方证据强,且有 GF 证明本身不能证明终端客户收入
Lockheed Martin国防集成商和投资方面向 DoD 应用的下一代传感系统国防设计导入合作官方证据强,且有伙伴证明采购时间和金额未披露
GUCASIC 设计服务进入市场路径面向超大规模云厂商的 >100 Tbps XPU 封装概念封装和 ASIC 集成伙伴关系官方证据中等,且有行业媒体印证无具名超大规模云厂商或出货数据
Alchip封装和 AI-ASIC 进入市场路径采用先进封装的一线超大规模 AI 设计生态 / 设计伙伴关系官方证据中等无具名终端客户或合同金额
Wiwynn机架级系统集成商面向云和超大规模工作负载的光互连机架级 AI 系统预览和私下客户简报公司证据强,且有伙伴印证仍处于正式上市前,客户未具名

这是 Ayar 具名交易对手中可公开验证的子集。多数行体现合作方、投资方或进入市场角色,而不是明确披露的经常性收入客户合同。

[CU006, CU008, CU009, CU010, CU012, CU014]
FU002: 客户类型 × 合作阶段矩阵

政府和战略平台关系在验证和设计导入上最强,但公开的量产证据几乎处处偏弱。

矩阵评分描述的是公开证据强度,不是每段关系的内在价值。「弱」往往反映披露缺失,而不代表经营表现为负。

[CU006, CU008, CU010, CU012, CU014, CU020]

6.3 政府与防务渠道

在 Ayar 的公开记录里,政府和防务是最早、也最具体有文档支撑的准客户渠道。DARPA 的 POEM 和 PERFECT 项目界定了 Ayar 创始人最初要解决的性能和功耗问题, PIPES 则提供了最清晰的正式桥梁,把研究脉络接到与 Intel 合作的封装光 I/O。DARPA 自己的 PIPES 页面明确把该技术描述为既适用于专用 DoD 产品, 也适用于双用途 CPU、GPU、AI 和 HPC 系统。随后,Lockheed Martin 把这条脉络延伸到实际防务用例,将 Ayar 的光 I/O 定位于下一代感知系统和 DoD 频谱处理工作负载。 HPE 又增加了一条相邻的公开 HPC 渠道,把 Slingshot 合作与未来 DOE exascale 系统挂钩。缺口同样重要: 没有任何已审阅来源点名 LLNL、ORNL、Sandia 或其他国家实验室部署,也没有任何防务或政府关系披露公开合同金额。 因此,这条渠道真实且战略重要,但更像验证锚点,还不是透明收入基础。[CU015, CU016, CU017, CU018, CU019]

政府、国防与 HPC 渠道枚举
项目 / 渠道公开交易对手公开证据内容runDate 时点状态商业解读
POEM 研究脉络DARPAAyar 称该光子微处理器项目帮助产出首款 TeraPHY chiplet历史研究脉络说明技术问题源自国防 / HPC,而非直接收入
PIPES 项目DARPA、Intel、Ayar Labs与 Intel FPGA 集成光 I/O;DARPA 目标是每封装 100 Tbps、低于 1 pJ/bit官方项目加 2022 年演示验证对封装光 I/O 最强的政府背书验证
PERFECT 项目背景DARPA在严苛 SWaP 限制下,嵌入式和超算系统的能效目标历史需求信号强化 Ayar 架构面向军事和嵌入式场景的适配性
国防传感系统Lockheed Martin / DoD 应用围绕光谱传感和相控阵架构的战略合作设计导入合作可能形成国防客户渠道,但采购金额未披露
DOE E 级超算相邻机会HPE Slingshot 和未来 DOE 超级计算机HPE 将光 I/O 工作与未来 DOE E 级系统的 Slingshot 需求相连仅属 HPC 相邻机会通向实验室的间接路径;未点名 LLNL、ORNL 或 Sandia 部署

本表穷尽了截至报告运行日,已审阅公开来源明确识别出的政府、国防和 HPC 渠道。

[CU015, CU016, CU017, CU018, CU019]

6.4 超大规模云厂商的进入路径与采用节奏

2025-2026 年的合作浪潮显示,Ayar 正从组件证明进入超大规模云厂商实际采购所依赖的供应链层。GUC 给 Ayar 打开了面向未来超大规模云 XPU 的 ASIC 设计服务路径; Alchip 给它提供了先进封装和 TSMC 对齐的路径,通向一线 AI 加速器;Wiwynn 则通过机架级基础设施和 OFC 2026 的私下客户说明会, 把公司推到最接近可部署云系统的语境里。合在一起看,这些公告意味着 Ayar 近期的市场进入路径不是以自己名义直接卖给 Microsoft、Google、Meta 或 AWS, 而是嵌入服务这些客户的封装、ASIC 和机架厂商。对 chiplet 公司来说,这在商业上合理;但也意味着那个标题级问题——谁真正承诺了批量购买——在公开层面仍无答案。 采用轨迹已经可见;轨迹尽头的客户名单还不可见。[CU020, CU021, CU022, CU023, CU024, CU027]

采用、留存与商业证据轨迹
信号公开证据最新披露状态指向什么缺失分母
客户名单披露Ayar 未公布具名客户名单截至 runDate 仍缺失公开证明仍由生态伙伴牵引,且披露有选择性付费账户数和细分构成
客户增长披露Series D 轮提到客户群增长和 Tier 1 接洽仅定性有势头,但公开资料无法量化活跃客户数,以及从评估到生产的转化
首批商业出货Series C 轮称 Ayar 已按合同完成首批批量商业出货2022 年历史披露2024 年前已有部分商业变现客户名称和当前出货节奏
单位规模披露Series C 轮称 Ayar 预计到 2022 年底出货数千个单元未发现后续公开跟进可能已发生送样 / 早期生产该目标是否达成并延续
具名生产部署公开记录集中在演示、合作和预览未发现具名超大规模云厂商生产订单商业就绪度提升快于公开账户披露已签生产账户和上线日期
留存 / 耐久性指标未披露 NRR、GRR、流失率或续约计划截至 runDate 仍缺失公开证据还撑不起耐久性判断队列、续约、流失率和合同期限
定价 / ASP 可见度未发现 TeraPHY 或 SuperNova 的公开定价截至 runDate 仍缺失经济性和客户回本周期仍不透明价格区间、BOM 影响和毛利率

类 null 的缺失是有意保留。该表把 Ayar 已公开披露的采用证据,与把这些证据转成持久客户判断所需的关键分母分开。

[CU002, CU003, CU004, CU028, CU029, CU030]
FU003: 已披露客户和伙伴里程碑(按年)

2020–2021 年,公开披露的客户相邻里程碑很少;2022 年集中出现;2025–2026 年随着 Ayar 从演示转向通往超大规模云厂商的伙伴关系,又重新加速。

计数来自保留下来的公开公告,这些公告直接推进了客户、伙伴或市场进入路径证据。它们不代表所有内部交易、试验或私下设计定点。

[CU020, CU022, CU023, CU037]

6.5 合作经济性、持久性与集中度

Ayar 公开客户证据的主要商业弱点不是缺少兴趣,而是缺少经济披露。TeraPHY 或 SuperNova 没有公开定价,NRR、GRR、流失或续约数据没有披露, 客户数分母没有披露,最大客户收入占比也没有披露。公开具名证据还集中在少数战略投资人与政府相关渠道里,这会真实放大外界高估多元化的风险。 Next Platform 明确提醒读者,不要因为 NVIDIA、AMD 或 Intel 投资就推断其已进入量产部署;它还单独指出,硅光子在生产系统里的存在感仍然很低, 因为买方仍担心规模经济性和可靠性。LightCounting 的 SC23 评论指向类似方向:Ayar 的演示值得关注,但 Nvidia 自己的 NVLink 网络仍非光学。 Ayar 自己的 2026 年网络研讨会同样承认,制造流程、封装和供应链就绪度仍要在 2026-2028 年继续成熟。换句话说,公开客户集合在战略上很亮眼, 但商业细节仍不足。[CU028, CU029, CU030, CU031, CU032, CU033]

公开集中度与收入路径估计
公开交易对手类别具名示例重要性集中度信号关键限制
政府和国防项目DARPA POEM、PIPES、PERFECT最早且最清晰的公开验证 / 资金渠道具名证据占比高不等同于经常性产品收入
战略投资方中的计算厂商战略投资人:NVIDIA、Intel Capital、HPE Pathfinder资本加路线图入口,能撬动强设计导入投资不能证明生产部署
制造与封装锚点GlobalFoundries、GUC、Alchip将光引擎做成客户可用封装的必要环节中-高运营依赖可能被误读成需求多元
国防集成商Lockheed Martin为 Ayar 提供具体的 DoD 用例桥梁采购时间和项目规模未披露
机架与系统集成商Wiwynn公开证据中最接近可部署云基础设施的一步中且在上升少数客户预览不等于已入账的超大规模云厂商订单
直接超大规模云客户公开资料未点名最终价值可能取决于云规模采用Unknown已审阅来源未公开确认 Microsoft、Google、Meta 和 AWS

集中度信号为定性判断,因为账户数和收入结构都不公开。因此,该表估计的是具名证据集中在哪里,而不是已实现收入必然来自哪里。

[CU019, CU020, CU022, CU023, CU027, CU028]
FU004: Ayar 客户采用旅程

典型的公开 Ayar 客户旅程,先从识别瓶颈和获得资助验证开始,再进入送样和演示;到量产和留存阶段,仍止于披露缺口。

这张旅程图描述的是保留来源中可见的披露模式;它是综合出的采用路径,不是 Ayar 导出的 CRM 阶段模型。

[CU005, CU010, CU024, CU028, CU029, CU030]

6.6 证据质量结论

本章的证据质量最适合概括为“生态验证强,已闭合商业证明弱”。可以较有把握纳入投资判断的是,Ayar 已经获得计算、封装、防务和机架集成领域一流交易对手的严肃接触; 公司早已越过纯概念科学,进入集成演示、验证测试和选择性客户预览;公共资金与伙伴支持也都围绕同一个核心问题对齐, 即把数据传得比铜互连更远、更快。尚不能判断的是客户规模化持久性:公开来源没有显示具名超大规模云采购承诺、量产单位数量、价格区间、集中度表或留存指标。 已审阅来源也没有找到公开量子计算客户公告,尽管公司在更广泛的行业讨论里与这一主题相邻。投资人因此应把 Ayar 视为高质量设计导入故事: 商业转化有可能发生,但公开透明度还不够。[CU031, CU032, CU034, CU035, CU036, CU038]

6.7 展项

Chapter 07

07风险

7.1 风险图谱与严重程度排序

Ayar Labs 进入 2026 年决策窗口时,风险画像在时间上异常压缩。公司有可信的产品信号、强大的战略投资人名单和内部报告的验证数据, 但决定风投回报的闸门项彼此紧扣:可制造性必须成熟,同时市场还在判断从可插拔模块和 LPO 转向真正共封装光学(CPO)的速度。 因此,近期最重要的风险不是抽象的技术怀疑,而是采用时间点、制造集中度、大平台的竞争性内化、可见需求集中在战略参与者手里, 以及爬坡后期才浮现的出口管制或 IP 摩擦。这些风险会叠加,而不是相互抵消。哪怕市场采用只推迟一年,而先进封装仍受约束、客户继续要求可运维性, Ayar 也可能在 NVIDIA、Broadcom 和 Marvell 强化自身光子布局时,面对更窄的商业化窗口。本章排序把这种时间压缩——不是单纯器件可行性——视为核心投资判断问题。[CR004, CR009, CR017, CR026, CR040, CR043]

综合风险清单
风险类别可能性影响缓释成熟度剩余敞口投资含义
CPO 采用推迟到 2027 年之后市场时点低-中压缩收入爬坡,并延长融资需求
GF、封装或激光器供应瓶颈制造推迟出货,并让毛利率继续不确定
NVIDIA/Broadcom/Marvell 把该层内化竞争Ayar 放量前,商用外供 TAM 收窄
早期客户群仍集中客户低-中试点转成耐久收入的结果偏二元
出口管制或国防合规摩擦监管中-高中国相关机会和伙伴导入放缓
IP 或 FTO 挑战后期出现IP低-中中-高错过窗口时,被迫授权、重设计或延期
创始人与扩张期高管流失治理低-中路线图和伙伴信心受扰
AI 资本开支放缓或周期下行宏观 / 资本中-高商业化窗口后移,现金需求还在
技术不错但出现 Rockley 式资本市场失败资本市场中-高降价轮、被迫出售或战略依赖上升

基于公开证据按严重性排序的综合判断;缓释成熟度为定性判断,反映外部能观察到多少缓释动作。

[CR009, CR017, CR024, CR026, CR039, CR043]
FR001: 风险热力图

根据当前公开证据,按可能性与影响定位 Ayar 的主要投资判断风险。

单元格是定性判断,不是概率模型输出。

[CR017, CR022, CR039, CR043, CR047]
FR002: 各类风险严重度

0–10 顺序严重度评分,显示哪些风险类别最限制当前投资逻辑。

评分是定性严重度排序,综合了可能性、影响和缓释可见性。

[CR024, CR026, CR029, CR039, CR043, CR049]

7.2 技术、集成与供应链风险

Ayar 的技术论据强到值得重视,但还不足以抹掉制造风险。公开公司材料展示了多代 TeraPHY 路线图,以及罕见详细的内部热测试和 BER 测试, 包括 30–80°C 循环和模拟 800°C/s 爬升。这些结果重要,因为它们回应了外界对热计算封装中微环型 CPO 的标准质疑。 但同一篇博客也强化了核心风险:Ayar 仍在谈 EVT、DVT、晶圆测试和验证关口,这些都必须关闭,制造才真正进入规模化。 公开来源没有披露良率曲线、光学对准缺陷率、封装漏检数据或长期现场返修。供应链集中度进一步放大不确定性。 Ayar 的故事仍依赖 GlobalFoundries 的硅光子工艺、台湾封装伙伴的封装集成,以及 Sivers 等外部激光器供应商提供 SuperNova 光源。 Laser Focus World 和 CIR 都清楚说明了远程激光器架构存在的原因:激光器对温度敏感,可维护性仍是现实问题。 这套架构缓解了一个问题,同时创造了另一个问题——外部激光器认证、供应保障和更换模式都变成量产关键堆栈的一部分。[CR007, CR008, CR009, CR010, CR011, CR012]

技术 / 集成 / 供应风险清单
失效模式公开证据可能性严重性现有缓释剩余缺口
PIC 良率或封装缺陷率未公开Ayar 披露 EVT/DVT 和验证进展,但未披露量产良率曲线多代验证和晶圆分选工作未披露独立良率或漏检率
热耦合扰动高温 XPU 旁的光链路Ayar 报告在高强度模拟升温热测试中 BER 稳定远程激光架构和调谐子系统没有客户集群可靠性或现场退货数据
外部激光阵列认证延后Sivers 合作聚焦认证和制造就绪中-高预购和伙伴协同供应商集中和配额风险仍在
GF 单一晶圆厂集中拖慢爬坡在 Ayar 公开材料中,GF 是可见的硅光子锚点与 GF 战略协同未公开第二来源或工艺转移路径
先进封装成为瓶颈需要 TSMC 生态封装,而 CoWoS 全行业仍紧张中-高Alchip 与 GUC 集成路径Ayar 的封装产能预留未披露
可维护性模型对运营商缺乏吸引力CIR 与 APNIC 强调维修焦虑和供应商锁定中-高远程激光器保留部分可替换性具体现场更换模型未公开

该表把 Ayar 已公开验证的内容,与客户集群或量产数据里仍未证实的内容分开。

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

7.3 市场时点与采用风险

Ayar 最大的风险,可能是市场认同它的方向,却不认同它的时间表。Ayar 相关来源和 Gazettabyte 把首次商业使用框定在 GPU scale-up 场景, 并指向 2026–2028 年商业化窗口。独立分析师和行业媒体更谨慎。LightCounting 认为高量 CPO 部署大概率在 2027 年, 而 LPO 部署已经开始,重新定时的可插拔模块在本十年剩余时间仍会有韧性。Optical Connections 还补充说, 第一波热度降温后,AI 光学增长应在 2026–2027 年放缓。CIR 的 2026 年分析更进一步:除少数超大规模环境外, 可见的 CPO 部署仍然很少,许多运营商对可维护性和供应商锁定不舒服,企业可能把 CPO 视为下一个十年的技术。 这不意味着 Ayar 的长期方向错了。它意味着收入窗口可能从“很快必然发生”变成“以后必然发生”,而深科技硬件公司常常按前一种叙事融资, 却扛不住后一种叙事。在这个语境下,LPO、NPO 和更好的可插拔模块不是失败替代品;它们是真实的桥接架构, 能在不推翻光 I/O 投资逻辑的情况下推迟 Ayar 的收入爬坡。[CR015, CR016, CR017, CR018, CR019, CR020]

FR003: CPO 商业化时点区间

Ayar 相关来源、独立分析师和更广泛的非超大规模云采用评论,指向的时间窗口并不相同;本图用区间方式归纳。

年份区间是从引用来源推断出的方向性时间窗口,不是约束性预测。

[CR016, CR017, CR018, CR021, CR023]

7.4 竞争与 IP 风险

Ayar 并不是在一条无人竞争的架构赛道里。NVIDIA 已经把硅光子纳入自己的交换机组合。Broadcom 有清晰的 XPU-CPO 和光学路线图, 还拿到了来自 Meta 的客户可靠性统计。Marvell 正把内部硅光子经验与收购 Celestial AI 结合。Lightmatter 仍在面向 AI 互连网络推广 Passage, 这意味着即便不计入更专门的挑战者,初创公司赛场也已经拥挤。这个竞争集合重要,因为每多一个既有巨头或资本充足的初创公司, 光 I/O 保持为可对外销售、且独立经济性丰厚的品类的概率就会下降。IP 图景也提高了犯错成本。Justia 的受让人页面显示, Ayar、Lightmatter 和 Rockley 之间存在多组相邻专利组合;Rockley 破产以及后续专利组合出售说明,困境中的光子 IP 可能回到更有资金实力的竞争对手手里。 如果冲突出现,PTAB 提供了现成挑战场所。这些都不能证明 Ayar 今天存在 FTO 问题;但确实说明,公司身处密集且活跃可交易的专利环境, 后期争议即便没有法庭胜负,也可能造成战略伤害。[CR026, CR027, CR028, CR029, CR030, CR031]

竞争威胁评分
竞争对手攻击层当前证明点威胁等级重要性反向观点
NVIDIA交换机层光子技术Spectrum-X / Quantum-X Photonics,2026 年以太网可用平台所有者可把光网络内化Ayar 仍更直接瞄准计算封装 I/O
Broadcom交换机和 XPU 光互连200T 路线图,加上 Meta 可靠性小时数路线图宽度叠加已验证的超大规模部署证据封装级外供机会可能仍在
Marvell / Celestial AIScale-up XPU 互连网络10B+ 设备小时和收购 Celestial AI兼具在位优势、资本和 M&A架构并不等同于 Ayar 的 chiplet 形态
Lightmatter私营初创公司的 scale-up 互连网络Passage 3D 光子平台中-高瞄准同一波 AI 集群升级窗口生态证据看起来弱于在位厂商
LPO / NPO / 先进可插拔模块替代架构谨慎买家偏好的过渡技术可推迟全面 CPO 需求,但不否定光互连在极端 scale-up 覆盖距离或功耗上可能仍会输
困境 IP 整合者IP 和资产市场Rockley 破产后专利组合低价售出竞争对手无需自建即可买到可用 IP单靠买资产不能保证采用

威胁等级反映未来 24–36 个月内,各竞争对手或替代方案能在多大程度上直接压窄 Ayar 的可寻址窗口。

[CR026, CR027, CR028, CR029, CR030, CR031]

7.5 客户集中、资本与治理风险

Ayar 可见的支持基础质量很高,但也偏窄。公开记录里充满战略投资人、合作方和政府相关项目:GlobalFoundries、HPE、NVIDIA、Intel 相关 DARPA 工作, 以及 DoD 相邻应用。更难找到的是具名量产客户、多元化对外销售收入或批量采购承诺的广泛证据。评估套件页面进一步显示, 公司仍在把准入机会分配给签有长期商业协议的客户;从商业化角度看这合理,但也说明早期需求可能由少数账户主导。 这种集中度叠加在资本强度之上。公司披露了大量融资和超过独角兽水平的估值,但本次审阅材料里没有公开烧钱速度或现金跑道数据。 如果采用时间点拉长到分析师窗口的后段,Ayar 可能在规模经济性得到证明之前还需要额外资本。治理风险同样集中。 Ayar 的公开领导层页面比早期初创公司的组织图深得多,但公司似乎仍异常依赖 Mark Wade、Vladimir Stojanovic 和 Chen Sun, 来维持商业、架构和科学连续性。在它变成单点依赖之前,这是一项优势。[CR003, CR004, CR006, CR038, CR039, CR040]

依赖与关键人物集中度表
依赖项证据集中原因若失去的影响当前缓释待尽调问题
Mark Wade公开领导层页面显示其为 CEO 兼联合创始人商业战略和生态叙事围绕他展开可见更广的高管团队要求披露继任深度和留任条款
Vladimir StojanovicCTO 兼联合创始人;公开技术代表架构可信度很大程度压在创始技术团队身上已有硅工程副总裁和更广团队要求提供已下放的技术责任图
Chen Sun首席科学家兼联合创始人科学连续性和器件诀窍仍与创始人绑定中高工程梯队已扩充审查联合创始人以下的梯队深度
战略生态支持方NVIDIA、HPE、GF、Intel/DARPA 背景突出可见支持集中在少数战略参与方产业栈上有多个参与方披露战略投资者和项目以外的收入结构
早期商业访问评估套件配额偏向长期协议少数共同设计伙伴可能主导首轮量产配额管控可提高转化质量披露面向量产的客户数量和集中度
规模化职能已具名的制造、法务、财务和硅工程负责人仍相对有限执行和合规仍取决于梯队厚度中高近期领导层扩充提供规模化职能的组织架构和招聘 / 留任指标

本表聚焦公开来源可见的集中度,而非未披露的股权结构表或组织架构动态。

[CR003, CR038, CR039, CR040, CR041, CR042]

7.6 监管、出口管制与防务相邻风险

Ayar 的出口管制风险不在于光 I/O 今天显然被禁,而在于 AI 硬件周边的合规边界扩张速度,快过年轻硬件公司通常建设法务基础设施的速度。 Cleary 2025 年 3 月对 BIS 会议的总结称,围绕中国、AI、半导体和超级计算的 EAR 执法大幅增加,Entity List 也在扩张。 2025 年 1 月 BIS 临时最终规则尤其相关,因为它收紧了对处理先进计算 IC 的晶圆厂和 OSAT 的尽调预期——这正是光互连供应商依赖的生态系统。 Ayar 与 DARPA PIPES 和 POEM 的接近度会提高敏感性,而不是降低敏感性:即便产品面向商业 AI 基础设施,双用途和防务相邻的性能目标也会强化筛查、安全和客户接受义务。 本次审阅的公开来源没有披露 Ayar 的 ECCN 立场、出口管制意见,也没有说明公司是否在商业 AI、中国相关机会和防务敏感项目之间划出严格边界。 这种不确定性眼下可管理,但它不是文书脚注;如果处理太晚,会直接压缩 TAM 或延迟出货。[CR043, CR044, CR045, CR046]

监管 / 法律风险清单
规则或场域范围当前信号可能性Ayar 敞口尽调需要剩余风险
EAR / BIS 执法姿态中国、AI、半导体、超级计算执法强度上升中-高交易对手或地域敏感时,筛查和许可负担上升取得产品分类和律师备忘录中-高
2025 年 1 月 BIS IFR面向先进计算 IC 的晶圆厂和 OSAT 尽调2025 年 1 月生效合规摩擦可能上移到供应链伙伴梳理各伙伴承担哪些义务
实体清单 / 拒绝政策受限交易对手Cleary 提到新增 80+ 个实体项目流程后期可能失去渠道或伙伴建立受拒方和最终用途筛查纪律
国防项目敏感性DARPA / DoD 关联PIPES 和 POEM 仍是相关参照点部署放量后,军民两用审查可能升温厘清商业工作与国防敏感工作边界
密集的硅光子专利版图Ayar、Lightmatter、Rockley 及同业可见多组重叠专利组合FTO 与许可成本风险可能到后期才抬升委托外部律师做专利格局审查中高
PTAB 挑战路径有效性与授权后争议USPTO/PTAB 仍是活跃争议场所低中商业化跑通后,争议可能从谈判升级准备挑战应对方案,并预留准备金假设

残余风险仍偏高,因为公开来源没有披露 Ayar 的 ECCN 备忘录、FTO 意见或挑战历史。

[CR036, CR037, CR043, CR044, CR045, CR046]

7.7 缓释因素、监测指标与投资逻辑破裂触发点

Ayar 确实有可信的缓释因素。远程激光器架构正是为了改善热表现和可维护性,公司也在台湾 ASIC 生态里推进多个封装伙伴, 并拥有不只会出资的战略投资人。但这些缓释因素必须变得可被外部观察,才值得给满分。信号最强的正面指标包括: 具名量产部署,而不是又一个架构合作;公开证据显示封装和激光器供应商从就绪走向持续出货;以及至少一个大型平台客户在 LPO 或先进可插拔桥接方案之前选择光 I/O。 负面指标同样清楚。如果 2027 年后高量 CPO 仍看起来被推迟,如果桥接架构主导早期 AI scale-up,如果出口管制复杂度封堵一个重要地域市场, 或如果 Ayar 在客户集中度下降之前再次融资,投资逻辑破裂就会开始成形。因此,投资人不应把公司当作纯组件供应商来盯, 而应把它看成依赖系统生态的平台押注;每一次披露的伙伴、合规事件或客户认证里程碑,都可能实质改变其价值。[CR013, CR016, CR024, CR043, CR047, CR048]

情景概率与监控触发因素
情景概率信号可监控触发因素哪些变化会改变判断含义
乐观:2026-2027 年生产跨越制造准备度合作和内部验证开始转化已具名生产发布,或披露重复订单公开客户证据,并带有可重复的可靠性数据风险溢价明显压缩
基准:2027-2028 年选择性爬坡符合 LightCounting 和 Laser Focus 的时间判断更多评估套件配额和封装伙伴胜出到 2027 年至少一个已具名、面向量产的账户跟踪观察,而非按超高速增长下注
悲观:过渡技术拖慢 CPOLPO/NPO 和可插拔方案仍有吸引力客户先标准化到过渡架构过渡技术在实际部署中输在功耗或延迟商业化窗口后移
合规冲击EAR 执法和尽调负担加重新的筛查事件、分类问题或伙伴限制律师确认分类干净且渠道合规TAM 下修,导入延迟
资本挤压AI 资本开支红旗早于广泛收入出现融资条款转弱,或超大规模云厂商资本开支放缓客户支持订单或重大战略融资完成稀释风险或被迫并购上升

情景概率是定性信号,不是精算预测;重点是让触发因素可观察。

[CR016, CR017, CR024, CR025, CR043, CR047]
FR004: 风险传导图

展示 Ayar 根本风险如何传导到收入延后、融资收紧和投资逻辑失效。

连线表达的是方向性逻辑,不是测得的系数。

[CR009, CR024, CR042, CR043, CR047, CR049]
Chapter 08

08估值

8.1 上一轮定价是锚点,不是完整估值证明

Ayar Labs 2024 年 12 月 Series D 是当前资本结构里唯一公开的硬定价事件,因此应优先于其他所有视角。该轮融资 $155M, 将已披露累计融资推到约 $370M,并明确把公司估值推到 $1B 以上。它还带来了异常强的战略投资背书: AMD Ventures、Intel Capital、NVIDIA、3M Ventures 和 Autopilot 加入 Advent Global Opportunities、Light Street Capital, 以及一个既有投资团,后者已包括 GlobalFoundries、Lockheed Martin Ventures、Playground Global、Boardman Bay 和 Applied Ventures。 这张股权结构表重要,因为它显示多个半导体、系统和防务参与者相信,Ayar 的光 I/O 堆栈可能在 AI 基础设施内部变得具有战略重要性。 但同一组证据也暴露了核心投资判断上限:没有确切投后估值,没有已发布收入,没有毛利率,也没有公开优先股堆叠。 换句话说,上一轮是可靠标记,但单独不足以证明公允价值。[CV001, CV002, CV003, CV004, CV005, CV006]

建议摘要表
建议置信度风险评级估值立场决策含义
观察合理以 >$1B 的上一轮估值标记作为持有 / 监控锚点;但没有收入、股权结构表和客户转化证据,不应按溢价入场。

建议对价格敏感:Ayar 在战略上看起来真实,但公开证据仍太薄,不足以给出买入判断或按收入倍数承销。

[CV002, CV037, CV047, CV048, CV049]
估值视角摘要
视角锚点信号对 Ayar 的含义主要限制
上一轮估值标记2024 年 12 月 Series D> $1B 估值;融资 $155M主要锚点,也是最好的公开起点准确投后估值和优先股堆叠仍未披露
私有同业轮次Lightmatter / Celestial / Scintil$4.4B / >$2.5B / $133.6M显示 Ayar 处在跨度很大的私有光子公司估值区间中部同业在产品范围、融资规模和时间点上不同
公开市场光学可比公司Rockley / POET / Lumentum / Coherent从困境到规模化的分散区间可作为边界条件和情绪检查噪音太大,业务太分散,无法给出干净的直接倍数
战略并购先例Cisco-Acacia / Marvell-Inphi / Ciena-Nubis / II-VI-Coherent$270M 至 $10B 的结果设计定点成熟后,支持战略退出可选性多数交易涉及规模更大或更多元化的资产
收入倍数NM没有公开收入或利润率基础不应主导当前建议Ayar 尚未披露该方法所需的财务输入

本表是决策层级,不是公式。第一个视角权重最高,因为它是唯一与 Ayar 自身绑定的当前定价事件。

[CV002, CV018, CV021, CV022, CV023, CV026]
FV003: 围绕上一轮估值下限的估值驱动瀑布图

战略和生态正向因素,大体被收入不透明与采用时点风险抵消,当前公允价值中枢仍落在上一轮估值下限附近。

示意性估值桥,单位为百万美元。正负调整是判断视角下的权重,不是管理层数字。

[CV002, CV010, CV012, CV037, CV041, CV047]

8.2 私有光子同业界定上行语境

最干净的私有同业信号说明,Ayar 今天并没有按类别龙头估值。Lightmatter 2024 年 10 月 Series D 用 $400M 新资金定出 $4.4B 估值, 累计融资 $850M。Celestial AI 2025 年 3 月 Series C1 融资 $250M,把累计融资推到 $515M 以上,并将估值抬到 $2.5B 以上。 即便在类别低端,Scintil Photonics 在 2025 年 8 月后期轮后也接近 $133.6M 估值。换句话说,私有光学基础设施市场已经横跨从低于 $200M 的专用资产, 到数十亿美元级 AI 光子平台。Ayar 自身略高于 $1B 的位置,因此有意义,但相对于融资最激进的光子公司仍有折价。 这个折价可以理解:Ayar 有战略投资人、早期出货历史和强产品主张,但公开证据仍没有显示广泛商业转化或财务披露, 足以支撑按 Lightmatter 或 Celestial 水平溢价。[CV018, CV019, CV020, CV021, CV022, CV038]

可比估值表
可比对象日期 / 阶段估值或金额重要性关键限制
Ayar Labs Series D2024 年 12 月 / Series D> $1,000M 投后估值;融资 $155M标的公司上一轮锚点准确投后估值未公开披露
Ayar Labs Series C2022 年 4 月 / Series C融资 $130M显示独角兽轮之前,投资人质量和商业化叙事已上台阶该轮没有披露干净的公开估值标记
Lightmatter2024 年 10 月 / Series D$4,400M 估值;融资 $400M估值最高、最接近的光子基础设施同业AI 光子平台更宽,资本基数大得多
Celestial AI2025 年 3 月 / Series C1> $2,500M 估值;融资 $250MAI 基础设施里的直接光互连同业产品架构和融资日期不同
Scintil Photonics2025 年 8 月 / 后期 VC$133.6M 估值;累计融资 $82.5M低端私有光子参考点硅 PIC 公司,不是直接的机架级光 I/O 竞争对手

估值标记和融资额混用,因为同业披露细节不均。真正的信号是相对位置,而不是公式化平均。

[CV002, CV007, CV018, CV020, CV021, CV022]
FV001: 精选估值与交易标记

私募融资定价和近期战略交易,把 Ayar 夹在尚未放量的光子初创公司与数十亿美元级类别龙头之间。

所有数字均为百万美元。Ayar 采用公开披露的 >$1B 下限,而不是虚构精确的投后估值。

[CV002, CV018, CV021, CV022, CV030, CV032]

8.3 公开光学可比公司给出边界,而非直接倍数

公开市场光学公司差异太大,无法为 Ayar 支撑干净的直接倍数,但它们仍可作为边界条件。区间悲观端是 Rockley Photonics, 该公司 2023 年 1 月申请 Chapter 11,并通过重组消除了超过 $120M 的有担保债务。当光子叙事无法转化为持久资本市场支持时, 这就是警示性的公开结果。POET Technologies 处在早期阶段光谱的另一端:这是一家上市光学公司,虽然过去收入几乎可以忽略、分析师覆盖很薄且共识为卖出, 却拥有数十亿美元公开市值。与此同时,Lumentum 和 Coherent 是大型多元化公开光子供应商,披露收入达到数十亿美元, 可作为规模参照,却不是阶段匹配的风险投资可比公司。分析师也继续给这个类别注入现实感:Lightwave、Cignal AI 和 LightCounting 都认为, 广泛高量 CPO 部署仍处早期,并且很可能明显推到 2027 年或更晚。因此,公开可比公司不否定 Ayar 的估值标记,但反对虚假精确。[CV023, CV024, CV025, CV026, CV027, CV028]

公开市场光子可比公司
公司公开市场状态公开快照重要性关键限制
Rockley Photonics曾上市的光子公司;Chapter 11 案件于 2024 年 7 月结案2023 年 1 月申请 Chapter 11;债务压力导致失败后重组光子纯公司的最佳下行情景警示破产结果不是仍可使用的估值倍数
POET TechnologiesNASDAQ 上市的外国私人发行人~$2.24B-$2.52B 市值,但过去 12 个月收入仅 ~$1.07M-$1.41M显示光学公司在公开市场上的可选性可能跑在基本面前面覆盖薄、股价波动大,也不是纯粹的 Ayar 类比
LumentumNASDAQ 上市通信组件供应商~$1.65B 年销售额 / ~$2.49B 过去 12 个月收入成熟光学供应商的规模和披露参考业务多元、已盈利,且远超风险投资阶段
CoherentNYSE 上市光子平台~$5.81B 年销售额 / ~$6.6B 过去 12 个月收入可参考广义光子平台在整合后的估值方式激光、材料和仪器敞口比 Ayar 宽得多

公开市场光学公司横跨困境重组、投机性可选性和规模化多元龙头。这个分散度有信息量,恰恰因为它不是干净的一对一可比组。

[CV023, CV024, CV025, CV026, CV027, CV028]
FV004: 可比公司定位矩阵

从战略契合和产品就绪度看,Ayar 表现不错;但相对任何需要传统估值锚的投资人,财务可见性很弱。

仅为顺序型定性评分。矩阵用于说明:Ayar 看起来具备战略吸引力,但仍缺少干净的估值支撑基础。

[CV018, CV021, CV023, CV026, CV037, CV041]

8.4 M&A 先例显示战略胃口,但规模仍主导价格

光互连和光子 M&A 先例是支持性的,但前提是记住区间有多宽。Cisco 在 2021 年以约 $4.5B 现金收购 Acacia, 以强化其 Internet-for-the-Future 战略中的相干光学。Marvell 同意以 $10B 现金加股票收购 Inphi, 以加深其数据中心和 5G 光连接版图。更近一些,Ciena 同意以 $270M 现金收购 Nubis Communications, 以强化其面向 AI 工作负载的数据中心内部战略;II-VI 则用大型现金加股票结构完成了与 Coherent 的合并。 共同主线很清楚:只要光学资产能解决关键基础设施内的带宽、延迟或封装瓶颈,战略买家愿意支付实质价格。 但价格离散同样清楚:市场给有规模、有收入、平台属性强的资产,与给更窄能力补强型收购的定价完全不同。Ayar 属于这场战略对话, 但今天的公开证据还不能有把握地把它放在 Acacia 或 Inphi 那一端。[CV030, CV031, CV032, CV033, CV042]

光互连并购先例
收购方 / 标的日期交易价值战略理由对 Ayar 的启示
Cisco / Acacia Communications2021 年 3 月完成~$4.5B 现金将相干光技术纳入 Cisco 的 Internet-for-the-Future 战略显示战略方会为绑定核心路线图的差异化光学资产支付高价
Marvell / Inphi2020 年 10 月宣布;2021 年 4 月完成~$10B 现金加股票拓宽 Marvell 的数据中心和 5G 光连接组合显示光学资产已经嵌入主要数据通路时,价值上限有多高
Ciena / Nubis Communications2025 年 9 月宣布$270M 现金用 CPO/NPO 和 ACC 技术扩展数据中心内 AI 互连战略当前 AI 互连小型收购的低端基准
II-VI / Coherent2022 年 7 月完成$220 现金 + 每股 Coherent 换 0.91 股 II-VI打造更宽的材料、网络和激光平台显示光子平台整合可以有战略意义,但规模和广度很重要

启示是战略买方愿意收购,不是直接倍数。交易横跨补强型收购、平台组合和规模化网络押注。

[CV030, CV031, CV032, CV033, CV042]

8.5 情景区间围绕上一轮,因为争论在时间点,不只在技术

估值区间最好围绕时间点和商业化置信度来框定,而不是围绕收入倍数。乐观情景下,如果光 I/O 能把战略关系转化为可见设计定点, 如果客户路线图能变成更广泛的量产部署,并且分析师对 CPO 时间点的怀疑开始消退,Ayar 有可能明显高于上一轮地板。 基准情景下,公司仍接近上一轮标记,因为战略投资团质量很高,类别也仍然重要;但公开证据仍无法展示收入、单位经济性或股权结构清晰度。 悲观情景下,如果 CPO 部署继续慢于预期,如果光学买方更久地停留在桥接技术上,或下一轮融资暴露投资人不友好条款、需要折价清场, 估值标记会压缩到上一轮以下。因此,这个情景差更多反映商业化时间点、战略可选性和披露质量,而不是传统财务模型。[CV034, CV035, CV036, CV043, CV044, CV045]

乐观、基准与悲观估值情景
情景假设隐含估值区间(USD M)概率信号决策含义
乐观Ayar 将战略关系转化为更广泛的量产设计定点,分析师对 CPO 时间点的怀疑缓和,商业规模开始在外部可见。1500-2200需要 2026-2027 年部署证据出现实质变化。若有真实客户和经济性数据支撑,可以支持高于上一轮估值标记付款。
基准战略投资团继续支持,产品准备度推进,也没有出现重大的负面商业化意外,但收入披露仍然有限。900-1300最符合当前证据。支持围绕上一轮平价持有 / 跟踪,而不是激进上调估值。
悲观CPO 采用慢于预期,过渡技术主导更久,或下一轮融资在当前估值标记下方完成,并带有沉重优先权包袱。500-800若公开部署证据到 2027 年仍稀少,风险上升。意味着上一轮估值偏撑,且下行保护比可选性更重要。

所有区间都是以百万美元表示的判断性公开市场情景。Ayar 没有提供收入或利润率基础,无法做更收窄的模型,因此区间刻意保持宽。

[CV034, CV035, CV036, CV043, CV044, CV045]
FV002: 隐含估值区间

中枢情景贴近上一轮估值下限;上行需要更快证明商业规模,下行则来自时点继续后移。

单一单位图,单位为百万美元。这些是情景区间,不是模型输出或管理层指引。

[CV002, CV037, CV043, CV044, CV045, CV047]

8.6 建议:观察,按面值合理,溢价前继续研究

从公开证据出发,最可防守的判断是中等置信度观察、估值立场合理;但前提是,未披露的优先股堆叠如果对投资人不友好,同一估值标记可能显得偏高。 Ayar 值得获得真实信用:罕见的战略投资团,来自 HPE 和 GlobalFoundries 的生态验证,DARPA 支撑的类别相关性, 以及明显对齐 AI scale-up 痛点的产品故事。但反向逻辑同样具体:最好的独立类别研究仍认为共封装光学不可避免但不会马上到来, 公开市场纯光子标的在该板块已经出现严重下行,而 Ayar 既不披露收入,也不披露单位经济性。正因如此,一个对价格敏感的建议必须止步于买入之前。 能推动结论变化的工作很直接:确切投后估值和优先权条款、当前出货和收入规模、设计定点正在转化为量产项目的证据, 以及光学堆栈经济性能支撑持久回报的证据。在这些事实公开或经尽调确认之前,上一轮平价合理,但支付有意义的溢价不合理。[CV037, CV041, CV047, CV048, CV049, CV050]

投资逻辑破裂与叫停触发因素表
触发因素阈值或事件对估值的传导行动含义
CPO 采用再次滑坡独立分析师观点仍认为主流大规模部署被推迟到 2027 年之后拉长规模化时间,削弱为战略可选性支付的溢价将立场从合理转向偏高,并要求较上一轮平价有更大折扣。
没有有意义的商业披露下一次融资事件前,没有公开收入、积压订单或已具名量产规模客户证据挡住基于收入的估值,只剩可选性支撑投资逻辑不支付溢价;建议维持观察 / 继续研究。
下一轮暴露对投资人不友好的条款优先股堆叠、参与分配条款或反稀释保护让新资金实质上处于劣后表面估值可保持不变,但实际回报经济性被摧毁在完成股权结构表尽调前,把当前估值标记视为偏高。
光学市场下行再次显现更多类似 Rockley 的困境结果,或超大规模云厂商明确拒绝 CPO推高行业折现率,压低退出胃口视严重程度,从合理转向偏高 / 昂贵。

每个触发因素都可从未来融资公告、分析师评论或公司 / 客户披露中外部监控。

[CV023, CV024, CV034, CV035, CV041, CV045]
最终尽调问题表
主题缺失证据重要性尽调路径
准确投后估值和优先股堆叠没有公开的投后估值数字、清算优先权安排或反稀释条款回报测算可能与名义独角兽估值大幅背离在 NDA 下要求投资条款清单摘要或股权结构表视图。
当前收入和出货规模没有公开收入、ARR 或出货金额披露没有这些,收入倍数估值和现金消耗评估都无从下手索取过去 12 个月收入、已签积压订单和出货运行率。
客户转化证据战略和 Tier-1 参与话术已公开,但已具名量产项目没有公开设计参与到规模化部署之间的缺口决定估值时间点索取客户阶段地图、量产时间和附着率。
单位经济性没有公开毛利率、模块成本或系统级节省披露溢价估值取决于经济性,而不只是技术优势要求产品毛利率桥和客户 ROI 数据。
制造准备度公开来源有准备度说法,但没有经审计的良率或产能利用率数据从合理估值到偏高估值,核心差别仍是规模化风险要求制造 KPI、良率轨迹和认证里程碑。

把观察名单仓位变成有承销依据的溢价入场前,至少需要这些尽调项。

[CV012, CV014, CV047, CV048, CV049, CV050]

8.7 展项

免责声明

本报告仅为尽调目的撰写,所用资料均为截至 2026-05-25 可获得的公开来源。 本报告不构成投资建议。未披露的私营公司指标保留为 null,或标记为分析师估计,而非自行推算。 所有论断均可追溯至各章节引用的 localEvidence claim id。

证据索引

结论
编号陈述可信度来源
CO001 Ayar Labs was founded in 2015. SO006, SO007, SO008, SO009
CO002 Current public materials place Ayar Labs in Silicon Valley and use both Santa Clara and San Jose, California, as headquarters datelines. SO006, SO009, SO016
CO003 Ayar Labs’ core commercial offering combines the TeraPHY optical engine with the SuperNova remote light source. SO001, SO003, SO004
CO004 Ayar positions optical I/O as a way to connect accelerators over distances ranging from millimeters to kilometers and to let many GPUs behave as a unified system. SO001, SO003, SO021
CO005 Current Ayar product materials claim optical I/O delivers roughly 5x-10x higher bandwidth, about 10x lower latency, and materially better power efficiency than pluggable optics plus electrical SerDes. SO003, SO004, SO021
CO006 The SuperNova light source is marketed as delivering up to 16 wavelengths into 16 fibers, supporting 256 data channels and 16 Tbps of bidirectional bandwidth. SO004
CO007 Open Compute’s technical listing says the current-generation Ayar solution provides up to 2.048 Tbps full-duplex throughput, up to 2 km reach, current AIB compliance, and a next-generation UCIe path. SO021
CO008 Ayar’s current home-page positioning advertises greater than 8 Tbps per optical engine and less than 25 ns end-to-end optical I/O latency. SO001
CO009 Mark Wade has served as Ayar Labs’ CEO since the company announced on December 11, 2023 that he would succeed Charles Wuischpard. SO002, SO009, SO019
CO010 Ayar describes Wade as a co-founder and photonics pioneer whose MIT and UC Berkeley work led to the company’s formation. SO009, SO019
CO011 Ayar’s public leadership page names Lisa Cummins Dulchinos, Vladimir Stojanovic, Amelia Thornton, Vivek Gupta, Scott Clark, Matthew Gloss, Chen Sun, and other senior leaders. SO002
CO012 Ayar’s public leadership page lists Craig Barratt, Pat Gelsinger, Will Graves, Jordan Katz, Ganesh Moorthy, and Geoff Tate among board members. SO002
CO013 Ayar’s April 2022 Series C announcement said the company raised $130 million led by Boardman Bay Capital Management with HPE and NVIDIA entering the round. SO008, SO019
CO014 In the Series C release, Ayar said it had made first volume commercial shipments under contract and expected to ship thousands of units by the end of 2022. SO008
CO015 HPE and Ayar announced a multi-year strategic collaboration in February 2022 focused on optical-I/O solutions for future HPE Slingshot and HPC / AI architectures, alongside a strategic investment from Hewlett Packard Pathfinder. SO010
CO016 Ayar’s NVIDIA collaboration materials frame optical I/O as a future enabler for AI and ML scale-out architectures and also show NVIDIA as a strategic investor. SO008, SO012
CO017 Ayar’s GlobalFoundries collaboration announcement says GF would co-develop and manufacture Ayar optical I/O using GF’s 45 nm CMOS process and that GF also invested in the company. SO011, SO020
CO018 Ayar and Intel demonstrated TeraPHY optical I/O under DARPA PIPES with a co-packaged FPGA solution capable of 2 Tbps of bandwidth on GlobalFoundries’ photonics process. SO013, SO020, SO024
CO019 Ayar’s November 2023 Supercomputing release said its optically enabled Intel Agilex FPGA used two TeraPHY chiplets each capable of 4 Tbps bi-directional bandwidth and operated at less than 5 pJ/bit with roughly 5 ns per chiplet latency plus time-of-flight. SO014, SO019
CO020 Lockheed Martin and Ayar announced in October 2022 that they would develop future sensory platforms using Ayar’s TeraPHY and SuperNova optical I/O technology, extending the company into defense use cases. SO015
CO021 Ayar’s December 2024 Series D announcement said the company raised $155 million, bringing total funding to $370 million and valuation above $1 billion. SO007
CO022 Ayar’s Series D investor list included AMD Ventures, Intel Capital, NVIDIA, 3M Ventures, and Autopilot alongside earlier investors such as Applied Ventures, GlobalFoundries, Lockheed Martin Ventures, Playground Global, and VentureTech Alliance. SO007
CO023 Ayar’s July 2025 leadership-expansion release said the company opened a new office in Hsinchu, Taiwan and doubled the size of its San Jose headquarters. SO016, SO006, SO022
CO024 The same July 2025 release said Ayar would continue hiring in the U.S. and Taiwan but did not disclose an exact employee count. SO016
CO025 Ayar’s March 2026 Series E materials say the company raised $500 million, bringing total funding to $870 million and valuation to $3.75 billion. SO006, SO022, SO026
CO026 Series E added institutional investors including ARK Invest, Insight Partners, Qatar Investment Authority, Sequoia Global Equities, and 1789 Capital, while giving Neuberger Berman a board observer role. SO006, SO022
CO027 Ayar said Series E capital would be used to scale high-volume production and test capacity, strengthen ecosystem partnerships, and expand global operations including Hsinchu, Taiwan. SO006, SO022
CO028 Ayar and Wiwynn said in March 2026 that their rack-scale reference path is designed to scale to 1,024 accelerators and more than 100 Tbps of optical connectivity per accelerator. SO017, SO018, SO023
CO029 Ayar’s Wiwynn materials and The Register’s reporting both frame rack-scale CPO deployment around practical issues such as fiber management, liquid cooling, manufacturability, telemetry, and serviceability. SO017, SO023
CO030 DARPA’s public PIPES page sets a target of 100 Tbps per package at below 1 pJ/bit, and Ayar’s own PIPES write-up says the demonstration leveraged prior work under DARPA’s POEM program. SO013, SO024, SO025
CO031 Ayar’s founder narrative ties the company to MIT and UC Berkeley work on processors communicating with light, supported by DARPA-funded research origins. SO009, SO011, SO025
CO032 Gazettabyte’s January 2026 interview says Ayar sees AI GPU scale-up as the first commercial use case for optical I/O, with extended memory as the next step. SO027
CO033 Independent 2026 coverage shows Ayar working with Alchip and GUC so optical engines can be integrated into advanced ASIC packages for hyperscaler deployment paths. SO023, SO027
CO034 Across Ayar’s current company and partnership materials, the central commercial message is co-packaged optics for AI scale-up beyond a single rack. SO001, SO006, SO017
CO035 Ayar’s public company-overview and financing materials do not disclose revenue or run-rate. SO006, SO007, SO022
CO036 Ayar’s company-overview sources discuss a growing customer base and customer milestones, but they do not publicly identify named customers or provide a verified customer count in this chapter’s source set. SO006, SO007, SO019
CO037 Ayar publicly discloses executives and board names but does not publish board committees, independence designations, or ownership percentages in the materials reviewed for this chapter. SO002
CO038 The Register argues that one barrier to adoption for co-packaged optics is the increased blast radius versus pluggables, because a failed optical engine can take the entire chip with it. SO023
CO039 Independent reporting also identifies liquid-cooling design, fiber routing, monitoring, and rack-level mechanical integration as unresolved complexity for rack-scale CPO deployment. SO017, SO023
CO040 Ayar’s own current materials present the optical-I/O stack as proven or production-ready rather than as a purely experimental laboratory technology. SO001, SO006, SO022
CO041 DARPA and Laser Focus World materials show that even after early successes, the path from demonstration to 100 Tbps-plus embedded optical I/O remains a long maturity curve. SO020, SO024
CO042 The Wiwynn partnership shifts Ayar’s public story from component validation toward system-level architecture and rack delivery. SO017, SO018, SO023
CO043 Ayar emphasizes that TeraPHY is built on standard form factor, manufacturing, and packaging flows used by major accelerator and switch vendors. SO001, SO006
CO044 Current Ayar technical materials point to a standards stack spanning AIB today, UCIe next, CW-WDM MSA for optics, and broader ecosystem alignment with CXL, OCP, OIF, and UALink. SO005, SO021
CO045 Wiwynn says it has shipped general and AI servers to more than 750 data centers worldwide, giving Ayar a partner with meaningful rack-level delivery experience. SO017, SO018
CO046 Ayar’s Series C release and later Lightwave coverage together indicate that commercialization had progressed beyond research demos by 2022-2023, including first volume shipments under contract and a growing customer base narrative. SO008, SO019
CM001 The defensible market boundary for Ayar Labs is co-packaged optical engines and optical I/O chiplets used in AI infrastructure, not the full data-center networking hardware stack. SM001, SM004, SM009
CM002 Ayar Labs’ solution combines the TeraPHY optical I/O chiplet with the SuperNova remote light source to move data optically between packages and systems. SM001, SM002, SM003
CM003 Ayar says TeraPHY is built to fit standard packaging and integration flows already used by major XPU and switch vendors. SM001, SM002
CM004 Ayar argues that scale-up is the larger AI-fabric challenge and says it needs at least 10x more bandwidth and 10x lower latency than current approaches. SM004
CM005 Most pluggable-module, switch-chassis, cable, and transport spend in AI optics is adjacent or substitute spend rather than Ayar’s direct monetizable pool. SM001, SM004, SM009, SM014
CM006 LightCounting’s public optics-for-AI-clusters lens rises from about $5 billion in 2024 to more than $10 billion in 2026. SM009, SM012
CM007 LightCounting’s public cloud-data-center-optics summary projects 30%-35% annual growth in 2025 and 2026, then 15%-20% annual growth in 2027-2030. SM009
CM008 Public 2026 CPO market estimates span more than 3x, ranging from roughly $1.247 billion to $4.2 billion depending on boundary assumptions. SM027, SM028
CM009 PW Consulting says the CPO module market was $560 million in 2025 and implies roughly $1.247 billion in 2026, with AI clusters/HPC at 56.7% of 2025 end-user demand and optical engines at 60.7% of component share. SM027
CM010 HDIN Research publishes a 2026 CPO market range of roughly $2.2-$4.2 billion and a 25%-35% CAGR through 2031. SM028
CM011 LightCounting’s public December 2025 note says the CPO engine market could reach about $10 billion and close to 100 million ports in 2030. SM011
CM012 Applying PW Consulting’s published AI/HPC and optical-engine shares to its 2026 market value yields a low-case product-near optical-engine slice of roughly $0.43 billion. SM027
CM013 LightCounting describes optics used in AI scale-up networks as a new market segment that expands in 2026-2030 as clusters spill beyond single-building footprints. SM009, SM012
CM014 Dell’Oro expects another strong year of AI-related investment in 2026, with strong double-digit growth in AI networking spend and CPO potentially adding multi-billions to market size. SM014
CM015 Dell’Oro expects 1.6 Tbps switches to ship in volume in 2026 and says the ramp should be faster than the earlier 800G transition. SM014
CM016 LightCounting expects both LPO and CPO to be deployed in AI scale-up networks starting in 2026-2027 and to reach high volumes by 2028. SM010, SM012
CM017 LightCounting says re-timed pluggables are not going away and that shipments of 800G-and-higher transceivers will triple from 2025 to 2030. SM010
CM018 LightCounting says current CPO implementations are still limited to switch-centric scale-out networks and that scale-up products are expected to start shipping in 2027. SM010, SM011
CM019 Dell’Oro’s public commentary implies a hybrid adoption path in which CPO appears first on networking switches while pluggable optics remain part of deployed AI-network SKUs. SM014, SM015
CM020 Ayar’s generative-AI market framing says inference may use 10-100 GPUs, fine-tuning 100-1,000 GPUs, and training thousands to tens of thousands of GPUs. SM008
CM021 Ayar says pluggable GPU-to-GPU optical links consume about 30 pJ/bit versus less than 5 pJ/bit for in-package optical I/O. SM008
CM022 Ayar says pluggables have more than 10x lower edge bandwidth density, more than 100x lower area density, and historical costs around $1-$2/Gbps versus in-package optical I/O targets. SM008
CM023 Ayar’s homepage says TeraPHY optical engines provide more than 8 Tbps per engine and less than 25 ns end-to-end optical-I/O latency. SM001, SM013
CM024 An OFC 2025 Optica abstract reports a UCIe optical I/O retimer that delivers 1.024 Tbps of bidirectional bandwidth per optical port and 8.192 Tbps aggregate bandwidth. SM021, SM013
CM025 An IEEE OFC 2025 paper says connectorized optical I/O chiplets, passive fiber attach, and known-good-chiplet process flow are important prerequisites for scalable high-volume manufacturing. SM022
CM026 The HPE-Ayar collaboration explicitly targeted future Slingshot interconnect generations and disaggregated servers for HPC and AI. SM005
CM027 The GlobalFoundries-Ayar collaboration said Ayar’s optical I/O on GF’s 45nm process could offer up to 10x higher bandwidth, up to 5x lower power, and data capacity of 10 Tbps and beyond. SM006
CM028 The NVIDIA-Ayar collaboration framed optical I/O as necessary for future AI and HPC architectures and cited model scales exceeding 100 trillion connections. SM007
CM029 Ayar and Alchip’s 2025 public materials describe a scale-up CPO solution with more than 100 Tbps of bandwidth per accelerator and more than 256 optical scale-up ports per device. SM018, SM019
CM030 Dell’Oro says the largest hyperscaler AI clusters are already approaching 100,000 accelerators and could reach one million in the near future. SM015, SM016
CM031 Dell’Oro forecasts cumulative coherent optical transceiver shipments to exceed five million over the next five years, with nearly half shipping on routers and Ethernet switches. SM016
CM032 Dell’Oro says the optical transport market grew 10% in 2025, direct purchases of WDM equipment for DCI grew nearly 40%, and direct cloud-provider purchases grew around 50%. SM017
CM033 The IEA says global data-center electricity demand is set to more than double by 2030 to about 945 TWh and that AI-optimized data centers will more than quadruple their electricity demand. SM029, SM031
CM034 Dell’Oro says future Rubin Ultra racks could consume about 600 kW and that clusters approaching one million GPUs are pushing AI fabrics to become highly distributed. SM015
CM035 Alphabet’s Q1 2026 10-Q says purchases of property and equipment were $35.674 billion and primarily reflected investments in technical infrastructure. SM023
CM036 Meta’s Q1 2026 10-Q says capital expenditures, including principal payments on finance leases, were $19.84 billion, and construction in progress mostly relates to data centers, network infrastructure, and servers. SM024
CM037 Amazon’s Q1 2026 10-Q shows very large ongoing property-and-equipment investment and explicitly lists server and networking equipment among major property-and-equipment categories. SM025
CM038 Mitsui’s 2026 technology outlook says 2026 is likely the year CPO enters full-scale mass production and practical rollout as AI-data-center power consumption surges. SM026
CM039 NVIDIA’s Spectrum-X and Quantum-X photonics announcement claims 4x fewer lasers, 3.5x greater power efficiency, 10x better resiliency, and up to 400 Tbps throughput, highlighting strong in-house competition at the switch tier. SM030
CM040 The most defensible public adoption sequence is pluggables first, LPO next, then switch-side CPO, with accelerator-package optical chiplets following once packaging and ecosystem maturity improve. SM010, SM011, SM014, SM015, SM022
CM041 Ayar’s realistic monetizable opportunity is materially smaller than the headline AI-optics TAM because the company sells chiplets and optical engines rather than entire transceiver, transport, or system stacks. SM001, SM002, SM009, SM027
CM042 Ayar’s UCIe and CW-WDM framing implies that standardization lowers integration friction for multi-vendor chiplet ecosystems and remote-light-source adoption. SM003, SM008, SM013
CM043 Ayar positions SuperNova as the first CW-WDM MSA-compliant 16-wavelength light source, capable of up to 16 Tbps bidirectional bandwidth across 256 data channels. SM003, SM013
CP001 Ayar positions itself as an in-package optical I/O and co-packaged-optics provider for large-scale AI workloads. SP001, SP002
CP002 Ayar says TeraPHY integrates directly into existing package architectures and delivers more than 8 Tbps per optical engine with under 25 ns end-to-end latency. SP001
CP003 Ayar says its optical engine uses standard form factors, manufacturing flows, and packaging flows already used by major XPU and switch vendors. SP001
CP004 Ayar announced a $155 million Series D in December 2024 that lifted total funding to $370 million and valuation above $1 billion. SP002
CP005 Ayar frames its optical I/O as a replacement for copper interconnects and pluggable optics in AI infrastructure. SP002
CP006 NVIDIA announced Spectrum-X and Quantum-X Photonics switches that integrate silicon photonics and make Spectrum-X available in 2026. SP003
CP007 NVIDIA named Coherent, Lumentum, Eoptolink and Innolight as ecosystem participants around its photonics program while still supporting pluggable transceiver technologies. SP003
CP008 Broadcom says Meta testing accumulated one million 400G-equivalent CPO port device hours without a link flap. SP004
CP009 Broadcom says its CPO data showed 65% lower optics power than pluggable-module solutions and that the platform is ready to scale. SP004
CP010 Marvell announced a custom XPU architecture with co-packaged optics that scales from tens of XPUs within a rack to hundreds across multiple racks. SP005
CP011 Marvell says integrated optics in its XPU architecture allow XPU-to-XPU connections over distances 100 times longer than electrical cabling. SP005
CP012 Marvell says its silicon photonics devices have logged more than 10 billion field hours and have shipped for more than eight years. SP005
CP013 Acacia positions itself around AI scale-out optics, 3D siliconization, and client-optics components for higher-bandwidth, lower-power AI links. SP027
CP014 Intel says its first-generation OCI chiplet supports 4 Tbps bidirectionally, requires no external laser source, and is designed to be co-packaged with CPUs, GPUs and other SoCs. SP007
CP015 Intel says it has shipped more than 8 million silicon-photonics PICs and more than 32 million on-chip lasers in pluggable transceivers since 2016. SP007
CP016 Coherent markets a broad portfolio of transceivers, active optical cables, components and instruments for high-speed optical transmission systems and data centers. SP006
CP017 Lumentum says its datacom portfolio includes 800 Gbps and 1.6 Tbps OSFP modules for AI and cloud data centers. SP025
CP018 Lumentum showcased a 16-channel DWDM ultra-high-power laser source at OFC 2026 to support next-generation co-packaged-optics architectures. SP026
CP019 Lightmatter markets a complete photonics roadmap from NPO and OBO to 2D and 3D CPO and 3D interposers. SP008, SP009
CP020 Lightmatter says Passage L200 supports 32 to 64 Tbps of aggregate bandwidth and Passage M1000 delivers 114 Tbps across a 4,000 mm² footprint. SP008, SP009
CP021 Lightmatter discloses $850 million of total funding and a $4.4 billion October 2024 valuation on its homepage. SP008
CP022 Celestial AI closed a $250 million Series C1 funding round that brought total raised to $515 million. SP010
CP023 Celestial AI's Photonic Fabric is described as an optical interconnect platform for package, system and rack-level AI connectivity. SP010, SP011
CP024 Marvell says Celestial AI's first-generation Photonic Fabric chiplet delivers 16 Tbps in a single chiplet and is intended for co-packaging with custom XPUs and scale-up switches. SP011
CP025 Ranovus explicitly markets 12.8 Tb/s XPU and switch co-packaged optics for scale-up. SP012
CP026 Ranovus says its multi-terabit optical engines are validated by Tier-1 customers but does not name those customers on the reviewed source. SP012
CP027 Scintil markets a single-chip dense multi-wavelength laser source for AI scale-up links and claims more than 1 Tbps in a single fiber. SP013
CP028 Ciena's co-packaged and near-package optics page says linear architectures such as CPO and NPO eliminate local DSPs or retimers and use external light-source modules that can be serviced in the field. SP014
CP029 POET says its optical interposer products target 800G, 1.6T and beyond for AI and data-center optical networks. SP015, SP016
CP030 Innolight markets 1.6T OSFP-XD DR8+ and low-power 800G pluggable optical transceivers. SP030
CP031 Eoptolink says it is a market leader in 800G, 400G, 200G and 100G optical transceivers with high-volume production capability. SP031
CP032 Accelink markets datacom, DWDM and data-center interconnect optical systems with high-volume production messaging. SP032
CP033 LightCounting said in July 2025 that high-volume CPO deployments are likely in 2027 while LPO deployments start earlier and reach millions of units the following year. SP017
CP034 Avnet says pluggables are expected to remain the default choice for most links over the next five to ten years and that CPO is deployed selectively where density and power constraints are extreme. SP018
CP035 Avnet says CPO reduces serviceability and configurability because failed optical engines may require replacing the entire CPO assembly or line card. SP018
CP036 Rockley emerged from Chapter 11 in 2023 after about 46 days and received roughly $35 million of additional stakeholder funding. SP028
CP037 Celestial AI acquired Rockley's silicon-photonics IP for $20 million in October 2024, increasing Celestial's patent portfolio to more than 200 patents globally. SP029
CP038 Broadcom, Marvell, Cisco, Coherent, Lumentum and Intel all had current 10-K filing trails on the SEC by early 2026. SP019, SP020, SP021, SP022, SP023, SP024
CP039 NVIDIA's 2025 photonics announcement indicates that public incumbents and pluggable suppliers remain intertwined rather than fully displaced by CPO. SP003, SP030, SP031
CP040 Ayar's main architectural differentiator is pushing optics directly inside the compute or switch package while many rivals still emphasize pluggables, NPO, or broader photonics platforms. SP001, SP009, SP025, SP030, SP031
CP041 CPO and pluggables are likely to coexist for years, making LPO and advanced pluggables a transition competitor rather than an immediately obsolete architecture. SP017, SP018
CP042 NVIDIA's move into silicon-photonics switching internalizes part of the optical stack and therefore bypasses merchant vendors at the switch tier. SP003
CP043 Lumentum's reviewed 2026 messaging is focused on pluggable modules and laser sources rather than Ayar-style in-package optical I/O. SP025, SP026
CP044 Intel's OCI differs from Ayar's architecture because Intel integrates lasers on the chiplet while Ayar relies on a remote light source. SP001, SP007
CP045 Lightmatter and Celestial both pursue scale-up optical fabrics that extend beyond a single rack, directly overlapping Ayar's AI-cluster scale-up narrative. SP008, SP010, SP011
CP046 Cisco/Acacia's reviewed optics positioning is stronger in AI scale-out and client-optics components than in explicit compute-die optical I/O. SP021, SP027
CP047 Recent market history shows optical-interconnect startups increasingly either consolidate into larger semiconductor platforms or monetize IP before long-term standalone scale is proven. SP011, SP029
CP048 Once Ayar is qualified, switching costs can be meaningful because package co-design, fiber topology, remote-laser architecture and thermal validation become part of the platform definition. SP001, SP014, SP018
CI001 Ayar Labs prioritizes evaluation-kit allocations to customers that have committed to long-term commercial agreements. SI002
CI002 Ayar Labs said in April 2022 that it had made first volume commercial shipments under contract. SI004
CI003 Ayar Labs announced a $155 million Series D on 2024-12-11 led by Advent Global Opportunities and Light Street Capital. SI003, SI016, SI017
CI004 Ayar Labs said its total funding reached $370 million and its valuation moved above $1 billion after Series D. SI003, SI016, SI018
CI005 Series D publicly named AMD Ventures, Intel Capital, NVIDIA, 3M Ventures, Autopilot, and several existing investors as participants. SI003, SI016, SI017
CI006 Series D was framed as capital to support high-volume manufacturing aligned to customer roadmaps. SI003
CI007 Ayar Labs announced a $130 million Series C on 2022-04-26 led by Boardman Bay Capital Management. SI004
CI008 Ayar Labs said Series C financing would fund reliability qualification and production scaling beginning in 2022. SI004
CI009 Ayar Labs announced an additional $25 million Series C1 in 2023, bringing total Series C funding to $155 million. SI005
CI010 Ayar Labs announced a $35 million Series B on 2020-11-05 co-led by Downing Ventures and BlueSky Capital. SI006
CI011 Series B publicly listed Applied Ventures, Castor Ventures, SGInnovate, Founders Fund, GlobalFoundries, Intel Capital, Lockheed Martin Ventures, and Playground Global among the investors. SI006
CI012 Ayar Labs announced a $24 million Series A in 2018 with Playground Global, Founders Fund, GlobalFoundries, and Intel Capital participating. SI007
CI013 Ayar Labs disclosed a $3 million flexible draw term loan from Silicon Valley Bank in 2019 for capital and manufacturing expenditures. SI008
CI014 Ayar Labs said in 2019 that it had been awarded a DARPA grant as part of the PIPES program. SI008, SI009
CI015 Ayar Labs disclosed a $15 million multi-year prototype OTA under Project KANAGAWA in 2022. SI010
CI016 Project KANAGAWA pairs Ayar Labs with Intel, Lockheed Martin, and Qorvo to push optical I/O into defense-oriented domestic manufacturing. SI010
CI017 Ayar Labs says its founders' breakthrough was the result of a decade of DARPA-funded research collaboration. SI007, SI012
CI018 DARPA's POEM program targeted low-energy, high-capacity photonic communications within and between microprocessors and DRAM. SI020
CI019 DARPA's PIPES program targets embedded optical signaling for advanced IC packages with goals of 100 Tbps per package at less than 1 picojoule per bit. SI019
CI020 Ayar Labs' public monetization story centers on selling optical I/O chiplets and light sources rather than a software or subscription model. SI001, SI015, SI025
CI021 Ayar Labs markets TeraPHY and SuperNova as a combined hardware system whose value proposition is bandwidth, latency, and power efficiency. SI015, SI025
CI022 Evaluation kits and long-term commercial agreements imply that public monetization begins with qualification activity before broader production revenue. SI002
CI023 Ayar Labs said in 2020 that it was working with select semiconductor manufacturers, OEM system builders, and end users on sampling and co-design partnerships. SI014
CI024 Ayar Labs publicly disclosed first commercial shipments, but it did not disclose shipment revenue or per-unit pricing. SI004
CI025 Series D materials referenced a growing customer base and Tier 1 customer engagements without providing booked backlog or volume commitments. SI003, SI016
CI026 Reviewed public Ayar Labs sources do not disclose list pricing, realized ASPs, or revenue-recognition terms for hardware, samples, or partner programs. SI002, SI003, SI004
CI027 HPE framed its collaboration and investment around future silicon-photonics architectures rather than current public product pricing. SI011
CI028 NVIDIA collaboration materials focused on future AI infrastructure development rather than disclosed current Ayar Labs revenue. SI013
CI029 HPE Pathfinder made a strategic investment in Ayar Labs as part of a multi-year collaboration on future HPC and AI architectures. SI011, SI004
CI030 GlobalFoundries invested an undisclosed amount in Ayar Labs as part of a manufacturing collaboration. SI012
CI031 Lockheed Martin Ventures disclosed a strategic investment in Ayar Labs in March 2020 to accelerate commercialization for AI, HPC, and digital beamforming applications. SI014
CI032 NVIDIA entered Ayar Labs' Series C in 2022 and increased its investment in the 2023 Series C1 extension. SI004, SI005
CI033 Intel Capital has backed Ayar Labs since Series A and publicly reaffirmed its commitment in the 2024 Series D financing. SI007, SI003
CI034 Ayar Labs' investor base mixes strategic ecosystem partners with financial growth investors, which supports commercialization but does not replace operating disclosure. SI003, SI011, SI012, SI013, SI014
CI035 Reviewed public sources do not disclose an Ayar Labs revenue or ARR figure as of the run date. SI001, SI003, SI004
CI036 Reviewed public sources do not disclose Ayar Labs' gross margin, EBITDA, cash balance, monthly burn, or runway. SI003, SI004, SI008, SI021
CI037 An SEC EDGAR company search for Ayar Labs returned no matching companies. SI021
CI038 Series D materials said Ayar Labs planned to increase hiring, but reviewed public financial sources did not provide a management-confirmed headcount suitable for efficiency modeling. SI003, SI001
CI039 Public unit-economics evidence is limited to value-proposition claims such as higher bandwidth, lower latency, and better power efficiency versus traditional interconnects. SI001, SI015, SI025
CI040 Ayar Labs argues that lower power consumption and system economics drive adoption, but it does not publish product gross margins or customer payback periods. SI001, SI013, SI015
CI041 Lightwave reported in 2026 that many users still question whether co-packaged optics is mature enough and whether it creates more operational risk than performance benefit. SI022
CI042 APNIC argued that co-packaged optics introduces serviceability, repair, and vendor-lock-in concerns relative to pluggables. SI023
CI043 Gazettabyte quoted LightCounting's Vladimir Kozlov warning against irrational decisions amid enthusiasm for co-packaged and near-packaged optics. SI024
CI044 Public adverse commentary frequently characterizes co-packaged optics as inevitable but not imminent, implying a slower mainstream revenue ramp than funding headlines suggest. SI022, SI024
CI045 Major disclosed Ayar Labs equity rounds sum to about $369 million, which is large relative to the absence of public operating metrics and therefore heightens financing-dependency risk for outside underwriters. SI003, SI004, SI005, SI006, SI007
CI046 Non-equity support is meaningful but incomplete in public view: KANAGAWA disclosed $15 million and SVB disclosed $3 million of debt-like capital, while the PIPES grant amount was not published. SI008, SI009, SI010
CE001 Ayar Labs’ public product stack is a two-part optical I/O system pairing the TeraPHY optical I/O chiplet with the SuperNova remote multi-wavelength light source. SE001, SE005
CE002 The current public TeraPHY implementation is still a roughly 2 Tbps-class bidirectional chiplet, with OCP materials describing eight 256 Gbps full-duplex optical ports and up to 2 km reach. SE014, SE025
CE003 Ayar markets optical I/O as extending accelerator connectivity from millimeters to kilometers so separate chips or racks can behave like one larger compute domain. SE001, SE005
CE004 The current SuperNova product page says one light source can deliver up to 16 wavelengths into 16 ports, enough for 256 data channels or 16 Tbps of bidirectional bandwidth. SE002, SE006
CE005 SuperNova’s external laser architecture is explicitly marketed as platform-flexible and field-replaceable rather than permanently buried inside the compute package. SE002, SE023
CE006 Ayar says SuperNova is CW-WDM MSA compliant and meets GR-468 reliability requirements for optoelectronic devices and pluggable optics. SE002, SE006
CE007 Ayar has also published a 64-addressable-wavelength SuperNova variant, indicating a roadmap beyond the 16-wavelength commercial product page. SE006, SE035
CE008 Across its product pages and architecture resources, Ayar claims optical I/O delivers 5x–10x higher bandwidth, 10x lower latency, and 4x–8x better power efficiency than pluggable optics plus electrical SerDes. SE001, SE005, SE014
CE009 Ayar’s first named product workflows are AI scale-up and memory disaggregation rather than generic switch replacement. SE007, SE024
CE010 Ayar frames optical I/O as the fabric that lets accelerators operate as “one giant GPU” across package, board, rack, and data-center distances. SE001, SE007, SE035
CE011 The current generation electrical interface is AIB, while Ayar’s next generation is described as UCIe compatible. SE014, SE007
CE012 The UCIe specifications provide an open die-to-die physical layer, protocol stack, and software model, and version 2.0 adds explicit support for 3D packaging and manageability. SE015, SE007
CE013 Ayar says UCIe plus optical I/O enables remote memory located up to 2 km away while keeping latency to less than 2 x 5 ns plus fiber time-of-flight. SE007, SE034
CE014 In Ayar’s public AI scale-up example, TeraPHY plus SuperNova provide 4 Tbps of bidirectional bandwidth per optical engine package connection. SE007, SE009
CE015 Ayar’s UCIe blog projects future optical I/O chiplets above 100 Tbps of off-package bandwidth and up to 128 ports per package. SE007, SE016
CE016 At SC23, Ayar showcased an Intel Agilex-based optical FPGA that integrated two TeraPHY chiplets and two SuperNova light sources around a 10 nm FPGA fabric die. SE009, SE025
CE017 That SC23 optical FPGA demo claimed sub-5 pJ/bit energy and about 5 ns per chiplet plus time-of-flight latency. SE009, SE008
CE018 Ayar’s optical I/O is manufactured on GlobalFoundries’ 45 nm monolithic silicon-photonics CMOS platform rather than a separate photonics die bonded to logic later. SE010, SE028
CE019 GF Fotonix / 45SPCLO combines monolithic RF, analog, and silicon-photonics blocks and includes V-groove-style fiber-attach support needed for dense package integration. SE028, SE029
CE020 A 2025 OFC abstract describes a UCIe optical I/O retimer chiplet for AI scale-up fabrics that delivers 1.024 Tbps per optical port and 8.192 Tbps aggregate package-to-package bandwidth. SE019, SE008
CE021 A 2025 IEEE/OFC paper on connectorized optical I/O chiplets says passive V-groove fiber attach and known-good connectorized chiplets are central to high-volume manufacturability. SE020, SE028
CE022 Ayar’s 2024 validation blog describes its latest TeraPHY as an 8 Tbps optical engine, the first UCIe optical interconnect chiplet, and the first demonstration of 16-wavelength microring-based links. SE008, SE019
CE023 Ayar reports that a five-day validation run kept all 16 UCIe modules error-free in a dual-die package. SE008
CE024 Ayar reports optical links with more than 5 dB of margin across hundreds of links even when SuperNova power per wavelength was lowered to 2.5 dBm. SE008
CE025 Ayar says thermal cycling between 30 °C and 80 °C produced no BER dependence and kept all links under the 1e-12 BER spec. SE008
CE026 Ayar says its emulated thermal-ramp testing kept links stable at conditions equivalent to 160 °C/s through 800 °C/s. SE008
CE027 Ayar reports full-duplex end-to-end UCIe-over-optics latency below 25 ns across more than 10 hours of testing. SE008
CE028 HPE’s multi-year collaboration with Ayar targets future Slingshot interconnect generations and disaggregated server architectures. SE011, SE029
CE029 Ayar’s NVIDIA collaboration is framed around accelerating optical interconnect deployment in AI and machine-learning architectures. SE012, SE029
CE030 Ayar’s OFC 2026 materials show rack-scale Wiwynn systems and liquid-cooled high-power ELSFP demonstrations, moving the message from package demos to deployable rack architectures. SE035, SE013
CE031 As of OFC 2026, Ayar is publicly positioning its co-packaged optics stack as production-ready for AI scale-up and high-volume manufacturing. SE035, SE031
CE032 EE Times and The Next Platform both describe Ayar working with Alchip and TSMC-class packaging flows to embed optical engines directly into advanced AI ASIC sockets. SE030, SE031
CE033 The Next Platform says Ayar aimed to have products selected, validated, and qualified by the second half of 2027 so customer ramps could begin in 2028. SE031
CE034 DARPA’s PIPES program target of 100 Tbps per package at less than 1 pJ/bit remains far beyond Ayar’s current public product metrics, defining the residual technical gap. SE016, SE025
CE035 DARPA’s POEM program centered on WDM-based, low-energy photonic links integrated with CMOS and DRAM-class interfaces, forming the public research lineage behind Ayar’s architecture. SE017, SE016
CE036 Justia’s 2026 Ayar Labs patent list includes “Pooled memory system enabled by monolithic in-package optical I/O” as US patent 12567910. SE033
CE037 That same 2026 Justia list includes “Low-power optical input/output chiplet for ethernet switches (TeraPHYe)” as US patent 12567920 with Wade, Stojanović, and Sun among the inventors. SE033
CE038 USPTO application 20210258078 describes remote memory architectures that connect HBM-class interfaces to optical links through electro-optical chips and optical macros. SE034
CE039 Ayar’s public technical trail in 2025 includes both an OFC abstract on UCIe optical retimers and an IEEE/OFC publication on connectorized chiplet packaging for AI and HPC. SE019, SE020
CE040 LightCounting expects LPO deployments to start before CPO and says high-volume CPO deployments are more likely in 2027 than 2026. SE022, SE023
CE041 Lightwave/CIR says non-hyperscale operators still want proof on reliability, field serviceability, and multi-vendor interoperability before adopting CPO broadly. SE023
CE042 Lightwave/CIR argues that external laser architectures reduce repair anxiety because lasers fail and are better kept cool, but they also create insertion-loss and shared-failure risks. SE023
CE043 Across Lightwave/CIR and the connectorized-chiplet paper, packaging yield, thermal behavior, maintenance models, and vendor lock-in emerge as the central barriers to broad CPO adoption. SE023, SE020
CE044 Ayar’s remote-memory patent filing and UCIe blog together show the roadmap extends beyond GPU-to-GPU scale-up into pooled and remote memory topologies. SE007, SE034
CE045 The Next Platform says Ayar’s volume ramp will depend on every weak link in the production flow, from materials and laser dies to package assembly and test. SE031
CE046 Gazettabyte reports Ayar expects the first commercial use to be GPU scale-up across and within racks, with extended memory following later. SE024, SE035
CE047 Gazettabyte says Ayar believes winning factors will be ecosystem access, manufacturable chiplet form factor, and proven reliability plus a multi-generation roadmap. SE024, SE008
CE048 Public proof still centers on demos, validation campaigns, conference papers, and reference designs rather than named broad hyperscale production deployments. SE023, SE031, SE024
CE049 Public sources still do not disclose package yields, laser MTBF, product ASPs, backlog, or named production deployment counts, leaving core underwriting questions unresolved. SE023, SE031, SE024
CU001 Public Ayar customer evidence is concentrated in collaborators, strategic investors, and government-backed programs rather than a disclosed commercial account list. SU001, SU003, SU004, SU010, SU011
CU002 Ayar's 2024 Series D release said the company was aligning high-volume manufacturing to customer roadmaps and had exciting engagements with Tier 1 customers without naming those accounts or disclosing a customer count. SU001, SU003
CU003 Ayar's 2022 Series C release said the company had made its first volume commercial shipments under contract. SU004
CU004 The same 2022 Series C release said Ayar expected to ship thousands of units of its in-package optical interconnect by the end of that year. SU004
CU005 Ayar's 2020 Lockheed investment release said the company was working with select semiconductor manufacturers, OEM systems builders, and end users on sampling and co-design partnerships. SU012
CU006 HPE and Ayar announced a multi-year collaboration to develop optical I/O for future HPE Slingshot generations and joint customer engagements. SU005, SU006
CU007 HPE Pathfinder invested in Ayar as part of the 2022 Slingshot collaboration announcement. SU005, SU006
CU008 NVIDIA and Ayar announced a 2022 collaboration focused on future AI and HPC optical-I/O architectures. SU007, SU008
CU009 NVIDIA participated in Ayar's 2022 Series C round and 2024 Series D round, making it both a strategic investor and public ecosystem collaborator. SU001, SU004, SU002
CU010 Intel and Ayar jointly demonstrated optical connectivity under DARPA PIPES and later showcased a 4 Tbps optically enabled Intel FPGA at SC23. SU009, SU010, SU011
CU011 Ayar's SC23 optical FPGA combined two TeraPHY optical I/O chiplets with Intel Agilex FPGA fabric and two SuperNova light sources. SU009
CU012 Ayar and GlobalFoundries said their collaboration would develop and manufacture optical I/O on GF's 45 nm CMOS process and integrate it with customer ASICs as a multi-chip module. SU014, SU015
CU013 GlobalFoundries invested in Ayar and remained a named existing investor in later financing rounds. SU014, SU001, SU002
CU014 Lockheed invested in Ayar in 2020 and framed the relationship as more than funding because it could help design Ayar into future complex systems. SU012, SU013
CU015 Lockheed and Ayar's 2022 collaboration targeted next-generation sensory platforms that could be used across Department of Defense applications. SU013, SU012
CU016 DARPA PIPES aims to embed photonic signaling inside ASIC and FPGA packages with a target of 100 Tbps per package at less than 1 pJ per bit for DoD and dual-use systems. SU011, SU010
CU017 Ayar said its work under DARPA POEM helped generate the first TeraPHY optical I/O chiplet. SU010, SU025
CU018 PERFECT shows the defense lineage of Ayar's founding problem set was power-efficient embedded and supercomputing systems rather than hyperscale cloud alone. SU025, SU026
CU019 HPE tied the Slingshot collaboration to future DOE exascale supercomputers, but no national-lab deployment was named in reviewed sources. SU005, SU006
CU020 GUC and Ayar partnered in 2025 to integrate co-packaged optics into advanced ASIC design services for future hyperscaler customers. SU020, SU021
CU021 GUC said the joint XPU package design enables more than 100 Tbps full-duplex optical interface from the package. SU020
CU022 Alchip and Ayar partnered in 2025 to bring optical I/O into tier-1 hyperscale AI designs using advanced TSMC packaging technologies. SU022
CU023 Ayar and Wiwynn announced a 2026 rack-scale AI system partnership aimed at cloud and hyperscale customers. SU023, SU024
CU024 The Wiwynn partnership said the joint architecture is designed to scale to 1,024 AI accelerators and would be previewed to select customers, press, and analysts at OFC 2026. SU023, SU024
CU025 Ayar's current website positions the company around AI training and inference infrastructure where optical I/O improves bandwidth, power efficiency, and latency for scale-up networks. SU016, SU017
CU026 Ayar's public go-to-market is increasingly indirect, combining chiplets with packaging partners and system integrators rather than selling directly to named hyperscalers in public. SU017, SU020, SU022, SU023
CU027 No reviewed public source named Microsoft, Google, Meta, or AWS as contracted Ayar customers as of 2026-05-25. SU003, SU017, SU023
CU028 No reviewed public source disclosed a public customer count, ARR-linked account count, or named top-customer share, even though Ayar referenced a growing customer base. SU001, SU003, SU016
CU029 No reviewed source disclosed public pricing for TeraPHY chiplets or SuperNova light sources. SU001, SU016, SU017, SU018
CU030 No reviewed source disclosed public NRR, GRR, churn, or renewal rates for Ayar customer relationships. SU001, SU003, SU016
CU031 The Next Platform cautioned that investor participation by NVIDIA, AMD, and Intel should not be read as proof those firms will deploy Ayar technology. SU003
CU032 The Next Platform also said little silicon photonics had yet reached production and that volume economics and reliability remained unresolved for buyers. SU003
CU033 LightCounting's SC23 note highlighted Ayar's Intel FPGA demo but noted Nvidia's own NVLink networks were still not using optical connectivity, underscoring transitional market timing. SU019, SU009
CU034 Ayar's 2026 webinar said commercial CPO deployment still depends on packaging, supply-chain, and integration progress across 2026-2028. SU027
CU035 Ayar's 2026 validation blog said its 8 Tbps TeraPHY optical engine had completed EVT and was moving from demos toward manufacturing at scale. SU018
CU036 Public evidence suggests strategic-investor customers and DARPA or DoD channels dominate the named Ayar customer set while direct hyperscaler wins remain undisclosed. SU001, SU003, SU011, SU013, SU024
CU037 Ayar's cadence of disclosed customer-adjacent milestones accelerated in 2022 and then re-accelerated in 2025-2026 as the story shifted from demos to hyperscaler-route partnerships. SU004, SU005, SU007, SU010, SU009, SU001, SU020, SU022, SU023
CU038 Reviewed sources did not surface any public quantum-computing customer announcement or named quantum sampling cohort for Ayar as of the run date. SU016, SU017, SU027
CR001 Ayar Labs publicly positions itself as a co-packaged optics supplier for AI scale-up beyond the rack rather than as a conventional pluggable-optics vendor. SR001, SR002
CR002 Ayar’s public product stack is a combined architecture of TeraPHY optical I/O chiplets plus the remote SuperNova multi-wavelength light source. SR003, SR004
CR003 Ayar says evaluation-kit access is being prioritized for customers with long-term commercial agreements, indicating concentrated early commercial access rather than broad self-serve availability. SR011
CR004 Ayar’s December 2024 Series D raised $155 million, brought total disclosed funding to about $370 million, and marked a valuation above $1 billion. SR006, SR007
CR005 DARPA’s PIPES program targets embedded optical signaling inside packages at 100 Tbps per package and under 1 picojoule per bit, keeping Ayar’s core architecture close to defense-grade performance goals. SR014, SR015
CR006 Ayar’s own DARPA and DoD-linked announcements show that government programs remain part of the company’s public credibility and funding narrative. SR013, SR014, SR015
CR007 Ayar’s collaboration with GlobalFoundries makes GF both a strategic investor and a manufacturing partner in Ayar’s public supply chain. SR008
CR008 GlobalFoundries markets its silicon photonics platform as the industry’s only high-volume 300mm CMOS manufacturing foundry for silicon photonics. SR016
CR009 Because Ayar’s public manufacturing story centers on GF and no alternative fab path is publicly disclosed in these sources, single-foundry concentration remains a material risk. SR008, SR016
CR010 Gazettabyte reports that Ayar is relying on Alchip and GUC within the TSMC ecosystem to package optical engines into compute or switch packages for hyperscalers. SR018
CR011 Sivers says its expanded relationship with Ayar is focused on product qualification and manufacturing readiness of high-precision laser arrays for high-volume optical I/O deployment. SR017
CR012 Laser Focus World says Ayar’s SuperNova light source comes from multiple suppliers and is intentionally kept remote because lasers degrade under the hotter thermal regime of compute packages. SR019
CR013 Ayar’s 2025 validation blog says wafer sort, EVT, and DVT are critical before optical I/O products can ramp into manufacturing at scale. SR005
CR014 Ayar reports that its latest TeraPHY optical engine maintained BER within spec during thermal cycling and during emulated ramps up to 800 degrees C per second. SR005
CR015 Gazettabyte says the first commercial use of Ayar’s technology is expected in GPU scale-up architectures that connect accelerators within and across racks. SR018
CR016 Laser Focus World places Ayar-style commercial optical I/O offerings in a 2026–2028 window rather than as an immediate 2026 step-function. SR019
CR017 LightCounting said in July 2025 that high-volume CPO deployments were likely in 2027. SR026
CR018 LightCounting also said LPO deployments were starting immediately and that millions of LPO units would be deployed next year. SR026
CR019 LightCounting said re-timed pluggables are not going away and that shipments of 800G and higher-speed transceivers will triple from 2025 to 2030. SR026
CR020 Optical Connections News summarized LightCounting’s view that AI-optics growth should moderate in 2026–2027 as the first wave of AI excitement subsides. SR027
CR021 CIR’s 2026 Lightwave analysis says there are still few visible CPO deployments outside a small number of hyperscale facilities and that the technology transition may take a decade to play out. SR028
CR022 CIR says cautious operators are likely to use NPO and LPO as bridge architectures because they do not trust early-generation CPO yields, thermal behavior, or repair models. SR028
CR023 CIR says many enterprise operators may view CPO as next-decade technology rather than a near-term mainstream networking choice. SR028
CR024 Lightwave’s 2026 capex analysis says LightCounting expects market growth to moderate in 2027–2031 and that supply-chain shortages remain an important constraint. SR029
CR025 CNBC reported in February 2026 that AI infrastructure spending could approach $700 billion and that the resulting cash impact was already raising red flags. SR034
CR026 NVIDIA’s March 2025 photonics launch integrated silicon photonics directly into switching products and put Spectrum-X Photonics into 2026 availability. SR020
CR027 Broadcom’s OFC 2025 roadmap highlighted a path to 200T optical interconnects, including XPU-CPO, LPO, and 200G-per-lane optics. SR021
CR028 Broadcom later claimed one million flap-free 400G-equivalent port device hours at Meta, showing reliability evidence that Ayar has not publicly matched with customer fleet data in these sources. SR022
CR029 Marvell says its CPO architecture is already available to custom XPU customers and is backed by more than eight years of silicon photonics shipments and over 10 billion field device hours. SR023
CR030 Marvell’s announced acquisition of Celestial AI shows that large semiconductor vendors are willing to internalize optical scale-up technology instead of leaving the layer to independent startups. SR024
CR031 Lightmatter continues to market Passage as a 3D photonics platform for AI applications, confirming that private startup competition for the same scale-up design window remains active. SR025
CR032 Light Reading’s OFC 2026 coverage describes the optics debate as fragmented across CPO, linear pluggables, and near-packaged optics rather than settled around one architecture. SR030
CR033 Gazettabyte’s OFC 2026 coverage says initial CPO deployment volumes are likely to be modest and characterizes integrated optics as an investment bubble. SR031
CR034 Rockley Photonics emerged from Chapter 11 only after a fast restructuring, underscoring that photonics specialists can hit capital-market failure before durable scale is reached. SR042
CR035 Celestial AI’s $20 million purchase of Rockley’s patent portfolio shows that distressed photonics IP can be reallocated to competitors at relatively low prices. SR043
CR036 Justia assignee listings show Ayar, Lightmatter, and Rockley all hold assigned patents in adjacent optical and photonics domains, indicating a dense IP landscape rather than clear whitespace. SR038, SR039, SR040
CR037 The USPTO PTAB remains the formal route for post-grant patent challenges, so any future optical-interconnect dispute can escalate beyond bilateral licensing talks. SR041
CR038 Ayar’s public leadership page concentrates the company’s visible operating authority in Mark Wade, Vladimir Stojanovic, Chen Sun, and a relatively small disclosed bench for manufacturing, finance, legal, and silicon engineering. SR012
CR039 Because those same people anchor strategy, core architecture, and scientific continuity, Ayar’s key-person risk is material even if the company now shows a broader executive roster than in earlier years. SR012, SR013
CR040 Ayar’s visible ecosystem of supporters is concentrated in a small set of strategic actors including GlobalFoundries, HPE, NVIDIA, Intel and DARPA-linked programs. SR008, SR009, SR010, SR013, SR014
CR041 Ayar’s public HPE and NVIDIA releases describe joint development and future-product work but do not disclose production purchase volumes or broad merchant demand. SR009, SR010
CR042 Ayar’s evaluation-kit policy implies that early customer conversion can be dominated by a small number of committed co-design partners before broader market access emerges. SR011, SR006
CR043 Cleary reported that BIS officials signaled a dramatic increase in EAR enforcement focused on China, AI, semiconductors, and supercomputing, with more fines and Entity List activity. SR035
CR044 The January 2025 BIS interim final rule added due-diligence expectations for foundries and OSATs handling advanced-computing integrated circuits and related supply-chain transactions. SR035, SR036
CR045 BIS’s EAR portal emphasizes that classification, screening, and compliance duties are ongoing operational requirements rather than one-time approvals. SR037
CR046 Ayar’s DARPA and DoD-linked materials keep part of the roadmap close to defense-sensitive applications, increasing export-control and program-security sensitivity even if the company is primarily a commercial AI-infrastructure supplier. SR013, SR014
CR047 TrendForce reporting summarized by TechPowerUp says CoWoS capacity kept selling out even while doubling and that 2026 equipment and delivery schedules were effectively booked, leaving advanced packaging a continuing AI hardware bottleneck. SR033
CR048 APNIC argues that CPO increases vendor lock-in and can turn an optical failure into board or system replacement rather than a simple module swap, which is especially relevant for remote-laser and serviceability decisions. SR032, SR028
CR049 Taken together, the public evidence suggests Ayar must prove volume manufacturability, convert a concentrated early customer set, and clear compliance hurdles before incumbents or bridge architectures absorb the market window. SR016, SR017, SR019, SR026, SR028, SR033, SR035
CV001 Ayar Labs raised $155 million in its December 11, 2024 Series D financing. SV001, SV002, SV035
CV002 Public sources describe Ayar Labs' Series D valuation only as above $1 billion rather than as a precisely disclosed post-money figure. SV001, SV002
CV003 Ayar Labs' disclosed lifetime funding reached about $370 million after the Series D round. SV001, SV002
CV004 Advent Global Opportunities and Light Street Capital led Ayar Labs' Series D financing. SV001, SV002
CV005 New Series D investors included AMD Ventures, Intel Capital, NVIDIA, 3M Ventures, and Autopilot. SV001, SV002, SV035
CV006 Existing investors named in Series D materials included Boardman Bay Capital Management, GlobalFoundries, Lockheed Martin Ventures, Playground Global, VentureTech Alliance, Applied Ventures, and Axial Partners. SV001, SV002
CV007 Ayar Labs raised $130 million in Series C financing in April 2022, led by Boardman Bay Capital Management. SV003
CV008 The 2022 Series C round added HPE and NVIDIA and reaffirmed participation from investors such as GlobalFoundries, Intel Capital, and Lockheed Martin Ventures. SV003
CV009 Ayar Labs said in 2022 that it had first volume commercial shipments under contract and expected to ship thousands of units by year end. SV003
CV010 HPE and Ayar Labs entered a multi-year strategic collaboration in February 2022 and HPE Pathfinder invested in the company. SV004, SV005
CV011 The HPE-Ayar collaboration focused on future HPE Slingshot, disaggregated servers, and next-generation HPC and AI networking architectures. SV004, SV005
CV012 GlobalFoundries said Ayar had partnered on GF Fotonix since its early development and that the platform set the stage for substantial Ayar volume shipments by year end 2022. SV006
CV013 DARPA's PIPES program targets embedded optics at 100 Tbps per package and under 1 picojoule per bit, underscoring the strategic relevance of Ayar's category. SV007
CV014 Ayar currently markets its solution as production-ready hyperscale optical I/O with more than 8 Tbps per optical engine. SV008, SV009
CV015 Ayar claims its optical I/O stack delivers 5-10x higher bandwidth, 10x lower latency, and 4-8x more power efficiency than pluggable optics plus electrical SerDes. SV009, SV010, SV011
CV016 SuperNova is described as a field-replaceable external laser source with up to 16 wavelengths and up to 16 Tbps of bidirectional bandwidth. SV010
CV018 Lightmatter raised $400 million in an October 2024 Series D that valued the company at $4.4 billion. SV012, SV013
CV019 Lightmatter's total capital raised reached $850 million after that financing. SV012
CV020 Celestial AI raised $250 million in a March 2025 Series C1 and disclosed cumulative funding above $515 million. SV014, SV015
CV021 Bloomberg-reported coverage said Celestial AI's valuation rose to more than $2.5 billion after that financing. SV015
CV022 Scintil Photonics was valued at about $133.6 million as of August 2025 after raising $82.5 million in total funding. SV036
CV023 Rockley Photonics filed Chapter 11 in the Southern District of New York on January 23, 2023 under case 23-10081. SV016, SV018
CV024 Rockley's restructuring eliminated more than $120 million of secured debt and generated about $35 million of new funding. SV017
CV025 POET Technologies is a Nasdaq-listed foreign private issuer and maintains investor relations as a public optical company. SV019, SV020
CV026 POET Technologies' public market value was roughly $2.24 billion to $2.52 billion in late May 2026. SV021, SV022
CV027 POET traded at that multi-billion market value despite only about $1.07 million to $1.41 million of trailing revenue in the same period. SV021, SV022
CV028 Lumentum had roughly $1.65 billion of annual sales and about $2.49 billion of trailing revenue in May 2026, making it a scaled diversified optics vendor. SV023, SV024
CV029 Coherent had roughly $5.81 billion of annual sales and about $6.6 billion of trailing revenue in May 2026, making it a broad photonics platform rather than a clean Ayar comp. SV025, SV026
CV030 Cisco completed its acquisition of Acacia for about $4.5 billion in cash at $115 per share in March 2021. SV027
CV031 Marvell agreed to buy Inphi in a $10 billion cash-and-stock transaction in 2020 and completed the acquisition in 2021. SV028, SV029
CV032 Ciena agreed to acquire Nubis Communications for $270 million in cash in September 2025 to expand its AI interconnect capabilities inside the data center. SV033
CV033 II-VI completed its acquisition of Coherent in July 2022 with consideration of $220 in cash plus 0.91 II-VI shares for each Coherent share. SV034
CV034 Lightwave reported in March 2026 that most data centers still had few CPO deployments outside a small number of hyperscale facilities. SV030
CV035 Cignal AI said in February 2025 that large-scale CPO deployment was still 3-5 years away, with only initial deployments starting in 2026. SV031
CV036 LightCounting said in July 2025 that high-volume CPO deployments were likely in 2027 while LPO was starting earlier. SV032
CV037 Ayar's last round is the most reliable valuation anchor because it is the only current price-setting event directly attached to the company. SV001, SV002, SV035
CV038 Ayar's current public mark sits below Lightmatter's $4.4 billion valuation and below Celestial AI's valuation above $2.5 billion. SV012, SV015, SV002
CV039 Public evidence supports valuing Ayar primarily on milestone progress, strategic optionality, and IP-backed category position rather than on a disclosed financial base. SV001, SV003, SV011
CV040 Ayar's strategic cap table and ecosystem support justify a premium to smaller private photonics startups such as Scintil. SV001, SV004, SV006, SV036
CV041 Rockley's bankruptcy plus continuing analyst skepticism on CPO timing show that the downside case for photonics ventures remains real. SV016, SV017, SV030, SV031
CV042 Optical M&A precedents show that strategics will pay for differentiated interconnect assets once the technology fits major infrastructure roadmaps. SV027, SV028, SV033, SV034
CV043 A bull-case valuation materially above the last-round floor requires visible design-win conversion, broader deployment proof, and better disclosure than is public today. SV001, SV030, SV031, SV032
CV044 A base-case valuation near the last-round floor assumes no major commercialization setback but also no new public financial proof. SV001, SV002, SV035
CV045 A bear-case markdown below the last-round floor becomes plausible if CPO timing slips, financing terms worsen, or bridge technologies delay optical scale-out longer than expected. SV030, SV031, SV032
CV046 Public optics comparables are too noisy or too diversified to provide a precise direct multiple for Ayar. SV021, SV023, SV025, SV016
CV047 The most defensible public-market stance today is track with medium confidence and a fair valuation view, which can shade stretched if diligenced terms prove investor-unfriendly. SV001, SV030, SV031
CV048 No public source reviewed for this chapter disclosed Ayar revenue, ARR, or gross margin, so revenue-multiple underwriting is not grounded in public evidence. SV001, SV011
CV049 Ayar's exact post-money valuation and liquidation-preference stack remain undisclosed in public materials, leaving return math dependent on diligence rather than on the headline unicorn label. SV001, SV002
CV050 The diligence answers most likely to change the recommendation are exact post-money and preferences, current revenue and shipment scale, customer conversion evidence, and unit economics. SV001, SV003, SV030, SV031
来源
编号出版方标题引文
SO001 Ayar Labs Ayar Labs: AI Scale-up Beyond the Rack Ayar Labs’ co-packaged optics (CPO) solution unlocks AI performance and profitability by enabling thousands of GPUs to operate as a single unified system.
SO002 Ayar Labs Leadership Team | Ayar Labs
SO003 Ayar Labs TeraPHY Optical I/O Chiplet | Silicon Photonics | Ayar Labs Optical I/O delivers 5x-10x higher bandwidth at 10x lower latency and 4x-8x more power efficiency compared to traditional interconnects.
SO004 Ayar Labs SuperNova Light Source | Ayar Labs The SuperNova™ external light source ... can supply light for 256 channels of data, or 16 Tbps bidirectional.
SO005 Ayar Labs Optical I/O Products | Ayar Labs
SO006 Ayar Labs Ayar Labs Closes $500M Series E, Accelerates Volume Production of Co-Packaged Optics This brings the company’s total funding to $870 million and raises the company’s valuation to $3.75 billion.
SO007 Ayar Labs Ayar Labs $155M Series D to Address AI Infrastructure Includes AMD, Intel Capital, NVIDIA This brings the company’s total funding to $370 million and raises the company’s valuation to above $1 billion.
SO008 Ayar Labs Ayar Labs Raises 130 Million in Series C Funding
SO009 Ayar Labs Ayar Labs Names Mark Wade CEO Wade succeeds Charles Wuischpard who will continue to serve in an advisory capacity over the coming weeks before he transitions out of the company in mid-January.
SO010 Ayar Labs Hewlett Packard Enterprise and Ayar Labs Announce Strategic Collaboration and Investment
SO011 Ayar Labs GlobalFoundries and Ayar Labs Establish Strategic Collaboration to Speed Up Data Center Applications
SO012 Ayar Labs Ayar Labs to Accelerate Development and Application of Optical Interconnects in Artificial Intelligence / Machine Learning Architectures with NVIDIA
SO013 Ayar Labs PIPES Researchers Demonstrate Optical Interconnects to Improve Performance of Digital Microelectronics
SO014 Ayar Labs Ayar Labs Showcases 4 Tbps Optically-Enabled Intel FPGA at Supercomputing 2023
SO015 Ayar Labs Lockheed Martin, Ayar Labs Partner to Advance Microchip Connectivity for Next-Generation Sensory Systems
SO016 Ayar Labs Ayar Labs Strengthens Leadership Team The company also announced the opening of a new office in Taiwan and that it has doubled the size of its San Jose headquarters.
SO017 Ayar Labs Ayar Labs and Wiwynn Partner to Bring Co-Packaged Optics to Rack-Scale AI Systems The new optically-connected rack-scale AI infrastructure is designed to scale to 1,024 AI accelerators and beyond.
SO018 Wiwynn Ayar Labs and Wiwynn Partner to Bring Co-Packaged Optics to Rack-Scale AI Systems
SO019 Lightwave Online Ayar Labs hands CEO baton to Mark Wade
SO020 Laser Focus World DARPA PIPES Program demonstrates 2 Tbit/s optical interconnects at the chip level
SO021 Open Compute Project Ayar Labs In-Package Optical I/O Solution
SO022 Business Wire Ayar Labs Closes $500M Series E, Accelerates Volume Production of Co-Packaged Optics
SO023 The Register Ayar Labs, Wiwynn to cram 1,024 GPUs into photonic system One of the barriers to adoption for co-packaged optics has been the increased blast radius.
SO024 DARPA PIPES | DARPA
SO025 DARPA POEM: Photonically Optimized Embedded Microprocessors
SO026 Data Center Dynamics Optical interconnect startup Ayar Labs closes $500m funding round backed by Nvidia and AMD
SO027 Gazettabyte Ayar Labs prepares to fulfil its optical input-output (I/O) vision
SO028 IEEE Connectorized Optical I/O Chiplet with V-groove for AI and High Performance Computing
SM001 Ayar Labs Ayar Labs: AI Scale-up Beyond the Rack
SM002 Ayar Labs TeraPHY™ Optical I/O Chiplet
SM003 Ayar Labs SuperNova™ light source
SM004 Ayar Labs AI Data Center Scale-Up with Co-Packaged Optics | Ayar Labs
SM005 Ayar Labs Hewlett Packard Enterprise and Ayar Labs Announce Strategic Collaboration and Investment to Develop Next-Generation Data Center Architectures and Networking with Optical I/O
SM006 Ayar Labs GlobalFoundries and Ayar Labs Establish Strategic Collaboration to Speed Up Data Center Applications
SM007 Ayar Labs Ayar Labs to Accelerate Development and Application of Optical Interconnects in Artificial Intelligence/Machine Learning Architectures with NVIDIA
SM008 Ayar Labs How Optical I/O is Enabling the Future of Generative AI: A Q&A with Vladimir Stojanovic
SM009 LightCounting Scale-up networks in AI Clusters is a new market for optical interconnects
SM010 LightCounting Highlights from the 1st virtual conference on Co-Packaged Optics (CPO)
SM011 LightCounting 2025 was the year of CPO
SM012 Optical Connections News AI drives demand for optical transceivers, LPO, CPO - report
SM013 Optical Connections News OFC 2025: Ayar Labs first UCIe with optical chiplet for AI scale-up architectures
SM014 Dell'Oro Group Data Center Networking in 2025–2026: Milestones and Opportunities Amid Supply Risk - Dell'Oro Group
SM015 Dell'Oro Group Insights from GTC25: Networking Could Tip the Balance in the AI Race - Dell'Oro Group
SM016 Dell'Oro Group Large Scale AI Clusters to Fuel More Growth in Coherent Optical Transceiver Shipments, According to Dell’Oro Group
SM017 Dell'Oro Group Optical Transport Market Grew to $16 Billion in 2025, According to Dell’Oro Group
SM018 EEJournal Alchip and Ayar Labs Unveil Co-Packaged Optics for AI Datacenter Scale-Up
SM019 Business Wire Ayar Labs and Alchip to Scale AI Infrastructure With Co-Packaged Optics
SM020 Electro Optics Ayar Labs raises $130m for silicon photonics optical I/O
SM021 Optica Publishing Group A UCIe Optical I/O Retimer Chiplet for AI Scale-up Fabrics
SM022 IEEE Connectorized Optical I/O Chiplet with V-groove for AI and High Performance Computing
SM023 U.S. Securities and Exchange Commission Alphabet Inc. Form 10-Q for quarter ended March 31, 2026
SM024 U.S. Securities and Exchange Commission Meta Platforms, Inc. Form 10-Q for quarter ended March 31, 2026
SM025 U.S. Securities and Exchange Commission Amazon.com, Inc. Form 10-Q for quarter ended March 31, 2026
SM026 Mitsui & Co. Global Strategic Studies Institute Top 5 Technologies to Watch in 2026: Co-Packaged Optics
SM027 PW Consulting Worldwide Co-Packaged Optics Module (CPO) Market 2026 - PW Consulting
SM028 HDIN Research Co-Packaged Optics (CPO) Market Analysis: 1.6T Transition & AI Interconnect
SM029 International Energy Agency AI is set to drive surging electricity demand from data centres while offering the potential to transform how the energy sector works
SM030 NVIDIA NVIDIA Announces Spectrum-X Photonics, Co-Packaged Optics Networking Switches to Scale AI Factories to Millions of GPUs
SM031 International Energy Agency Energy demand from AI – Energy and AI – Analysis - IEA
SP001 Ayar Labs Ayar Labs: AI Scale-up Beyond the Rack Ayar Labs’ co-packaged optics (CPO) solution unlocks AI performance and profitability by enabling thousands of GPUs to operate as a single unified system.
SP002 Ayar Labs Ayar Labs, with Investments from AMD, Intel Capital, and NVIDIA, Secures $155 million to Address Urgent Need for Scalable, Cost-Effective AI Infrastructure This brings the company's total funding to $370 million and raises the company's valuation to above $1 billion.
SP003 NVIDIA NVIDIA Announces Spectrum-X Photonics, Co-Packaged Optics Networking Switches to Scale AI Factories to Millions of GPUs By integrating silicon photonics directly into switches, NVIDIA is shattering the old limitations of hyperscale and enterprise networks.
SP004 Broadcom Broadcom Showcases Industry-Leading Quality and Reliability of Co-Packaged Optics Compared to pluggable module solutions, the data highlights that CPO reduces optics power by 65 percent and also demonstrates higher link reliability.
SP005 Marvell Marvell Announces Breakthrough Co-Packaged Optics Architecture for Custom AI Accelerators Marvell is now extending its custom silicon leadership by enabling customers to seamlessly integrate CPO into their next-generation custom XPUs.
SP006 Coherent Networking | Coherent Create next-gen high-speed optical transmission systems, networks, and data centers from our unmatched portfolio.
SP007 Intel Intel Silicon Photonics The first-generation chiplet supports 4 Tbps bidirectionally, with a roadmap to tens of Terabits per second per device.
SP008 Lightmatter Lightmatter Lightmatter’s photonic chips form a complete interconnect platform. Passage interconnects and Guide light engines scale networking for AI supercomputers.
SP009 Lightmatter 3D Photonics for AI Applications | Passage Passage L200 supports 32 to 64 Tbps of aggregate bandwidth through co-packaged optics.
SP010 Data Center Dynamics Optical interconnect startup Celestial AI raises $250m Celestial AI has closed a $250 million Series C1 funding round, bringing the total raised to $515m.
SP011 Marvell Marvell to Acquire Celestial AI, Accelerating Scale-up Connectivity for Next-Generation Data Centers Celestial AI’s first-generation Photonic Fabric chiplet for scale-up interconnect is the industry’s first scale-up optical solution delivering 16 terabits per second of bandwidth in a single chiplet.
SP012 Ranovus Architecting Optical Infrastructure For AI - Ranovus XPU and Switch CPO for Scale-up.
SP013 Scintil Photonics SCINTIL Photonics | Dense Multi-Wavelength Laser Source For AI Datacenters Introducing the world's first single chip dense multi-wavelength laser source to revolutionize AI scale-up photonic interconnects.
SP014 Ciena Optics for co-packaged applications Linear architectures such as CPO/NPO eliminate the need for local signal conditioning digital signal processors (DSPs) or retimers.
SP015 POET Technologies Home | POET Technologies Achieve lightning-fast transmission rates of 800G, 1.6T and beyond.
SP016 POET Technologies Products | POET Technologies POET is powering the age of photonics with innovative products that seamlessly integrate into existing networking infrastructure.
SP017 LightCounting Highlights from the 1st virtual conference on Co-Packaged Optics (CPO) The level of development activity on CPO is at an all time high now with high volume deployments likely in 2027. Deployments of LPO are starting now and millions of them will be deployed next year.
SP018 Avnet Pluggable vs. co-packaged optics in AI data centers: Power, scale and design trade-offs Pluggable optics are still projected to dominate deployed port counts because they are flexible, mature and straightforward to operate.
SP019 U.S. Securities and Exchange Commission EDGAR search results for Broadcom 10-K filings
SP020 U.S. Securities and Exchange Commission EDGAR search results for Marvell 10-K filings
SP021 U.S. Securities and Exchange Commission EDGAR search results for Cisco 10-K filings
SP022 U.S. Securities and Exchange Commission EDGAR search results for Coherent 10-K filings
SP023 U.S. Securities and Exchange Commission EDGAR search results for Lumentum 10-K filings
SP024 U.S. Securities and Exchange Commission EDGAR search results for Intel 10-K filings
SP025 Lumentum Datacom Transceivers | Lumentum Our portfolio includes 800 Gbps and 1.6 Tbps OSFP modules supporting DR4 reaches.
SP026 Lumentum Lumentum Demonstrates Industry-Leading Technologies and Products for Scale-Out, Scale-Up and Scale-Across AI Infrastructure at OFC 2026 The company is featuring a 16-channel DWDM laser source, illustrating Lumentum's readiness to support next-generation CPO architectures.
SP027 Acacia Acacia Client Optics Components enable higher bandwidth, lower power, and smaller footprint with high-quality components and optical engines designed to meet the demand for AI scale out.
SP028 Rockley Photonics Rockley Photonics Completes Financial Restructure, Emerges from Chapter 11 The company emerged with a strengthened capital structure having received approximately $35 million of additional funding from its stakeholders.
SP029 Data Center Dynamics Celestial AI acquires Rockley Photonics patent portfolio for $20m Optical interconnect company Celestial AI has acquired the silicon photonics IP from Rockley Photonics ... for $20 million.
SP030 Innolight The World's Leading Developer and Manufacturer of High-speed Optical Transceivers InnoLight demonstrates pluggable 1.6T OSFP-XD DR8+ and low power 800G optical transceivers.
SP031 Eoptolink Eoptolink - market leader in high speed optical transceivers This investment allowed us to become an 800G, 400G, 200G and 100G Market Leader.
SP032 Accelink 光迅科技 | 让光引领梦想 Accelink can provide end-to-end DWDM system with almost all optical devices ... and optical communication subsystem solutions suitable for data center interconnection.
SI001 Ayar Labs Innovators in Co-Packaged Optics for LLM Data Transfer | Ayar Labs Ayar Labs is transforming AI infrastructure by accelerating data movement.
SI002 Ayar Labs Optical I/O Evaluation Kit | Ayar Labs We are currently prioritizing our evaluation kit allocations to customers who have committed to long-term commercial agreements.
SI003 Ayar Labs Ayar Labs, with Investments from AMD, Intel Capital, and NVIDIA, Secures $155 million to Address Urgent Need for Scalable, Cost-Effective AI Infrastructure This brings the company's total funding to $370 million and raises the company's valuation to above $1 billion.
SI004 Ayar Labs Ayar Labs Raises $130 Million in Series C Funding, Accelerating Commercialization of Industry's First In-Package Optical I/O Products Ayar Labs also announced that it made its first volume commercial shipments under contract and expects to ship thousands of units of its in-package optical interconnect by end of year.
SI005 Ayar Labs Ayar Labs Adds $25 Million in Expansion of its $130 Million Series C Ayar Labs has raised an additional $25 million in Series C1 funding, bringing its total Series C raise to $155 million.
SI006 Ayar Labs Ayar Labs Secures $35 Million in Series B Funding and Adds New Strategic and Global Investors as Interest in Optical I/O Grows Ayar Labs is pleased to announce the completion of a $35M Series B financing co-led by Downing Ventures and BlueSky Capital.
SI007 Ayar Labs Ayar Labs Raises $24M Series A Ayar Labs, a company that improves the speed and energy efficiency of high performance computing systems with silicon chips that transmit data using light, has raised a $24M Series A.
SI008 Ayar Labs New Financing from Silicon Valley Bank, DARPA Research Contract, and Strategic Technical Hire from UC Berkeley The funding, which totals $3M, comes in the form of a flexible draw down term loan and will be used on specific capital and manufacturing related expenditures as Ayar Labs ramps sampling and production.
SI009 Ayar Labs Intel selects Ayar Labs as Optical Solution for DARPA PIPES Project Ayar Labs is pleased to announce that it has been selected by Intel to provide the Optical I/O solution for the company's recently awarded DARPA PIPES research project.
SI010 Ayar Labs Ayar Labs Partners with Department of Defense to Accelerate Transition to Optical I/O in Next-Gen Defense Applications Ayar Labs, a leader in chip-to-chip optical connectivity, was awarded a $15 million multi-year prototype Other Transaction Agreement (OTA) in support of Project KANAGAWA.
SI011 Ayar Labs Hewlett Packard Enterprise and Ayar Labs Announce Strategic Collaboration and Investment to Develop Next-Generation Data Center Architectures and Networking with Optical I/O HPE's venture arm, Hewlett Packard Pathfinder, has made a strategic investment in Ayar Labs to accelerate the development and adoption of joint future technologies.
SI012 Ayar Labs GlobalFoundries and Ayar Labs Establish Strategic Collaboration to Speed Up Data Center Applications As part of the agreement, GF has also invested an undisclosed amount in Ayar Labs.
SI013 Ayar Labs Ayar Labs to Accelerate Development and Application of Optical Interconnects in Artificial Intelligence/Machine Learning Architectures with NVIDIA The collaboration will focus on integrating Ayar Labs' technology to develop scale-out architectures enabled by high-bandwidth, low-latency and ultra-low-power optical-based interconnects for future NVIDIA products.
SI014 Ayar Labs Strategic Investment from Lockheed Martin Ayar Labs publicly demonstrated its monolithic electronic photonic TeraPHY chiplet at the Supercomputing 2019 conference and is now working with select semiconductor manufacturers, OEM systems builders, and end users on sampling and co-design partnerships in 2020.
SI015 Ayar Labs SuperNova™ light source Combined with Ayar Labs TeraPHY optical I/O chiplet, the solution provides 5x-10x higher bandwidth, 10x lower latency, and is 4x-8x more power efficient compared to traditional interconnects.
SI016 Crunchbase News Ayar Labs, Latest Optical Interconnect Startup To Raise Big, Hits Unicorn Status Ayar Labs locked up a $155 million Series D led by Advent International and Light Street Capital at a valuation of more than $1 billion.
SI017 Washington Technology Ayar Labs captures $155M to bring light into chip manufacturing Advent International and Light Street Capital led the Series D round, which brings Ayar Labs' total amount of investment raised since inception to $370 million and touted valuation to $1 billion.
SI018 Electronics Weekly Ayar Labs raises $155m Ayar Labs, the optical I/O specialist, has raised $155 million which brings the company's total funding to $370 million and gives it a valuation of over $1 billion.
SI019 DARPA PIPES | DARPA PIPES seeks to embed optical signaling technologies within the package of application-specific integrated circuits and field-programmable gate arrays to enable data links with unprecedented bandwidth density, efficiency, and reach.
SI020 DARPA POEM: Photonically Optimized Embedded Microprocessors The POEM program aims to address electrical communications link limitations by developing chip-scale, integrated photonic technology to enable seamless intrachip and offchip photonic communications that provide the required bandwidth with low energy/bit.
SI021 U.S. Securities and Exchange Commission EDGAR Search Results for Ayar Labs company search No matching companies.
SI022 Lightwave Online Co-Packaged Optics and the AI data center: From skepticism to strategic adoption Many users remain in a quandary, wondering whether they need CPO; whether it is mature enough, and whether adopting CPO will create more operational risk than performance benefit.
SI023 APNIC Blog Co-Packaged Optics — a deep dive Many will interpret CPO as “vendor lock-in disguised as innovation.”
SI024 Gazettabyte OFC Reflections - Part 2 Co-packaged optics integration is inevitable but not imminent.
SI025 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.
SI026 University of Illinois Urbana-Champaign The Perfect Project Parameter Variation at Near Threshold Voltage: The Power Efficiency versus Resilience Tradeoff.
SE001 Ayar Labs TeraPHY™ Optical I/O Chiplet TeraPHY™ optical I/O chiplets deliver up to 1000x bandwidth density improvements at 1/10th the power compared to electrical I/O, enabling ASICS to communicate with each other across a wide range of distances, from millimeters up to two kilometers.
SE002 Ayar Labs SuperNova™ light source Capable of supplying light for 256 channels of data, or 16 Tbps bidirectional.
SE005 Ayar Labs In-Package Optical I/O: Unleashing Innovation | Ayar Labs Watch this video and discover how Ayar Labs’ optical I/O solution is the key to unleashing innovation in AI, scaling cloud and HPC, launching new aerospace systems, enabling the next wave of 5G, and more.
SE006 Ayar Labs SuperNova™: The Industry’s First Multi-wavelength, Multi-port Optical Source with 64 Wavelengths | Ayar Labs Ayar Labs demonstrates the industry’s first multi-wavelength, multi-port optical source with 64 addressable wavelengths.
SE007 Ayar Labs AI Scale-Up and Memory Disaggregation: Two Use Cases Enabled by UCIe and Optical I/O UCIe combined with optical I/O enable AI scale-up and memory disaggregation by providing reduced latency, increased bandwidth, and interoperability across different vendors’ chiplets.
SE008 Ayar Labs Let’s Get Serious: TeraPHY™ Optical Engine Passes the Test for AI Scale-Up at Volume The TeraPHY optical engine is the industry’s first Universal Chiplet Interconnect Express (UCIe) optical interconnect chiplet that can deliver 8 Tbps of bandwidth.
SE009 Ayar Labs Ayar Labs Showcases 4 Tbps Optically-enabled Intel FPGA at Supercomputing 2023 Explore Ayar Labs' integration of optical I/O with FPGA, offering unprecedented HPC and AI performance, unveiled at Supercomputing 2023.
SE010 Ayar Labs GlobalFoundries and Ayar Labs Establish Strategic Collaboration to Speed Up Data Center Applications Learn how the Ayar Labs and GlobalFoundries partnership will improve data center application speeds with Ayar Labs' CMOS optical I/O technology.
SE011 Ayar Labs Hewlett Packard Enterprise and Ayar Labs Announce Strategic Collaboration and Investment to Develop Next-Generation Data Center Architectures and Networking with Optical I/O | Ayar Labs Ayar Labs is solving the I/O bandwidth and power bottlenecks by moving data using light. See how we are enabling the next phase of Moore’s Law. Read our latest post "Hewlett Packard Enterprise and Ayar Labs Announce Strategic Collaboration and Investment to Develop Next-Generation Data Center Architectures and Networking with Optical I/O".
SE012 Ayar Labs Ayar Labs to Accelerate Development and Application of Optical Interconnects in Artificial Intelligence/Machine Learning Architectures with NVIDIA Ayar Labs is solving the I/O bandwidth and power bottlenecks by moving data using light. See how we are enabling the next phase of Moore’s Law. Read our latest post "Ayar Labs to Accelerate Development and Application of Optical Interconnects in Artificial Intelligence/Machine Learning Architectures with NVIDIA".
SE013 Ayar Labs Ayar Labs and Wiwynn Partner to Bring Co‑Packaged Optics to Rack‑Scale AI Systems Collaboration moves CPO beyond component-level innovation and into deployable AI infrastructure for hyperscale workloads
SE014 Open Compute Project Open Compute Project The 4 Tbps optical I/O solution from Ayar Labs combines a TeraPHY electro-optical chiplet and a SuperNova remote light source.
SE015 UCIe Consortium Specifications | UCIe Consortium UCIe 2.0 supports 3D packaging, offering higher bandwidth density and improved power efficiency compared to 2D and 2.5D architectures.
SE016 DARPA PIPES | DARPA PIPES aims to integrate photonic modules that deliver 100 Tbps per package at energies less than 1 picojoule per bit.
SE017 DARPA POEM: Photonically Optimized Embedded Microprocessors The POEM program aims to address electrical communications link limitations by developing chip-scale, integrated photonic technology to enable seamless intrachip and offchip photonic communications that provide the required bandwidth with low energy/bit.
SE019 Optica Publishing Group A UCIe Optical I/O Retimer Chiplet for AI Scale-up Fabrics We demonstrate a UCIe Optical I/O Retimer for scale-up AI fabrics... delivering an aggregate bandwidth of 8.192 Tbps.
SE020 IEEE Connectorized Optical I/O Chiplet with V-groove for AI and High Performance Computing This paper presents connectorized in-package optical I/O chiplets with V-groove for passive fiber attach... scalable for high-volume manufacturing.
SE022 LightCounting Highlights from the 1st virtual conference on Co-Packaged Optics (CPO) The level of development activity on CPO is at an all time high now with high volume deployments likely in 2027.
SE023 Lightwave Online / CIR Co-Packaged Optics and the AI data center: From skepticism to strategic adoption Operators currently move gradually because they don’t trust early-generation CPO manufacturing yields, thermal behavior, or repair/maintenance models.
SE024 Gazettabyte Ayar Labs prepares to fulfil its optical input-output (I/O) vision - Gazettabyte The first commercial use of the technology will be for GPU scale-up architectures that connect accelerators within and across racks.
SE025 Laser Focus World DARPA PIPES Program demonstrates 2 Tbit/s optical interconnects at the chip level In production, the DARPA team expects the capacity of a single chiplet to reach up to 2.56 Tbit/s I/O over 24 channels.
SE028 CMC Microsystems GlobalFoundries® Silicon Photonics - GF Fotonix™ (45SPCLO) – CMC Microsystems 45SPCLO, built on a 45 nm SOI platform, enables monolithic integration of RF, analog and Si-Photonic circuits.
SE029 Electronic Design Is It Finally Time for Silicon Photonics to Shine? GF Fotonix is a completely monolithic process... The platform also supports 2.5D packaging technology to glue chiplets together.
SE030 EE Times Ayar Labs and Alchip Unveil Optical I/O Reference Design Partnering with both TSMC and Alchip allows us to place our optical engine right into this AI socket.
SE031 The Next Platform Ayar Labs Gets $500 Million To Ramp Photonics Into 2028 AI Systems We have to get our end to end products selected, validated, and qualified by the second half of 2027... to support those ramps in 2028.
SE033 Justia Patents Patents Assigned to Ayar Labs, Inc. Pooled memory system enabled by monolithic in-package optical I/O... Patent number: 12567910.
SE034 USPTO.report Remote Memory Architectures Enabled by Monolithic In-Package Optical I/O Patent Application A remote memory system includes... an electro-optical chip... configured to optically connect with an optical link.
SE035 Ayar Labs OFC 2026 From package to rack-scale fabrics, Ayar Labs’ co-packaged optics solution is purpose-built for AI scale-up and high volume manufacturing.
SU001 Ayar Labs Ayar Labs, with Investments from AMD, Intel Capital, and NVIDIA, Secures $155 million to Address Urgent Need for Scalable, Cost-Effective AI Infrastructure We believe that Ayar Labs has industry-leading technology and exciting engagements with Tier 1 customers.
SU002 Crunchbase News Ayar Labs, Latest Optical Interconnect Startup To Raise Big, Hits Unicorn Status
SU003 The Next Platform Nvidia, AMD, And Intel Help Stuff The Coffers At Ayar Labs We know that HPE made a strategic investment and collaboration agreement with Ayar Labs, but don't jump to any conclusions based on funding.
SU004 Ayar Labs Ayar Labs Raises $130 Million in Series C Funding, Accelerating Commercialization of Industry-First Optical I/O Solution
SU005 Ayar Labs Hewlett Packard Enterprise and Ayar Labs Announce Strategic Collaboration and Investment to Develop Next-Generation Data Center Architectures and Networking with Optical I/O
SU006 Hewlett Packard Enterprise Hewlett Packard Enterprise and Ayar Labs announce strategic collaboration and investment to develop next-generation data center architectures and networking with optical I/O HPE and Ayar Labs will partner on photonics research and commercial development, building a joint ecosystem of solution providers, and customer engagements.
SU007 Ayar Labs Ayar Labs to Accelerate Development and Application of Optical Interconnects in Artificial Intelligence/Machine Learning Architectures with NVIDIA
SU008 EE Times Ayar Labs Partners with Nvidia for Optical I/O Chiplets
SU009 Ayar Labs Ayar Labs Showcases 4 Tbps Optically-enabled Intel FPGA at Supercomputing 2023
SU010 Ayar Labs PIPES Researchers Demonstrate Optical Interconnects to Improve Performance
SU011 DARPA PIPES | DARPA Beyond FPGAs and specialty DoD products, photonic connectivity for CPUs, GPUs, and accelerators will impact dual-use applications including AI and high performance computing.
SU012 Ayar Labs Strategic Investment from Lockheed Martin | Ayar Labs
SU013 Lockheed Martin Lockheed Martin, Ayar Labs Partner to Advance Microchip Connectivity for Next Generation Sensory Systems The new platforms could be used across Department of Defense applications to capture, digitize, transport, and process spectral information.
SU014 Ayar Labs GlobalFoundries and Ayar Labs Establish Strategic Collaboration to Speed Up Data Center Applications
SU015 GlobalFoundries GlobalFoundries and Silicon Catalyst Partner to Accelerate Differentiated Technology Solutions for Semiconductor Startups
SU016 Ayar Labs Innovators in Co-Packaged Optics for LLM Data Transfer | Ayar Labs
SU017 Ayar Labs AI Data Center Scale-Up with Co-Packaged Optics | Ayar Labs
SU018 Ayar Labs Let's Get Serious: TeraPHY Optical Engine Passes the Test for AI Scale-Up at Volume
SU019 LightCounting Photonics-enabled disaggregated computing
SU020 Ayar Labs GUC and Ayar Labs Partner to Advance Co-Packaged Optics for Hyperscalers
SU021 EE Times GUC and Ayar Labs Partner to Advance Co-Packaged Optics for Hyperscalers
SU022 Ayar Labs Ayar Labs and Alchip to Scale AI Infrastructure with Co-Packaged Optics
SU023 Ayar Labs Ayar Labs and Wiwynn Partner to Bring Co-Packaged Optics to Rack-Scale AI Systems
SU024 Wiwynn Ayar Labs and Wiwynn Partner to Bring Co-Packaged Optics to Rack-Scale AI Systems Together, we are enabling advanced optical I/O that delivers greater scalability and energy efficiency for cloud and hyperscale customers.
SU025 DARPA Photonically Optimized Embedded Microprocessors
SU026 DARPA Power Efficiency Revolution For Embedded Computing Technologies
SU027 Ayar Labs Webinar: Next-Gen AI Architecture Through Co-Packaged Optics
SR001 Ayar Labs Ayar Labs: AI Scale-up Beyond the Rack Ayar Labs’ co-packaged optics solution unlocks AI performance and profitability by enabling thousands of GPUs to operate as a single unified system.
SR002 Ayar Labs Innovators in Co-Packaged Optics for LLM Data Transfer | Ayar Labs Ayar Labs is transforming AI infrastructure by accelerating data movement.
SR003 Ayar Labs TeraPHY™ Optical I/O Chiplet Ayar Labs’ optical I/O solution combines TeraPHY optical I/O chiplets with Ayar Labs’ SuperNova light source.
SR004 Ayar Labs SuperNova™ light source The SuperNova external light source provides up to 16 wavelengths of light and powers optical I/O links across the system.
SR005 Ayar Labs Let’s Get Serious: TeraPHY™ Optical Engine Passes the Test for AI Scale-Up at Volume Recent testing found it performs as expected at temperature change rates as high as 800 degrees C/s, with bit error rate staying under spec.
SR006 Ayar Labs Ayar Labs, with Investments from AMD, Intel Capital, and NVIDIA, Secures $155 million to Address Urgent Need for Scalable, Cost-Effective AI Infrastructure This brings the company's total funding to $370 million and raises the company's valuation to above $1 billion.
SR007 Business Wire ADDING MULTIMEDIA Ayar Labs, with Investments from AMD, Intel Capital, and NVIDIA, Secures $155 Million to Address Urgent Need for Scalable, Cost-Effective AI Infrastructure Advent Global Opportunities and Light Street Capital lead Series D to accelerate high volume manufacturing of Ayar Labs’ in-package optical interconnects.
SR008 Ayar Labs GlobalFoundries and Ayar Labs Establish Strategic Collaboration to Speed Up Data Center Applications As part of the agreement, GF has also invested an undisclosed amount in Ayar Labs.
SR009 Ayar Labs Hewlett Packard Enterprise and Ayar Labs Announce Strategic Collaboration and Investment to Develop Next-Generation Data Center Architectures and Networking with Optical I/O | Ayar Labs HPE and Ayar Labs team up to design future silicon photonics solutions for HPE Slingshot interconnect and advanced disaggregated servers.
SR010 Ayar Labs Ayar Labs to Accelerate Development and Application of Optical Interconnects in Artificial Intelligence/Machine Learning Architectures with NVIDIA The collaboration will focus on integrating Ayar Labs’ technology for future NVIDIA products.
SR011 Ayar Labs Optical I/O Evaluation Kit | Ayar Labs We are currently prioritizing our evaluation kit allocations to customers who have committed to long-term commercial agreements.
SR012 Ayar Labs Leadership Team | Ayar Labs Mark Wade is listed as Chief Executive Officer and Co-Founder, with Vladimir Stojanovic as CTO and Chen Sun as Chief Scientist.
SR013 Ayar Labs Intel selects Ayar Labs as Optical Solution for DARPA PIPES Project | Ayar Labs Ayar Labs has been selected by Intel to provide the Optical I/O solution for the company’s DARPA PIPES research project.
SR014 DARPA PIPES | DARPA PIPES seeks to embed optical signaling technologies within the package of ASICs and FPGAs to enable data links with unprecedented bandwidth density, efficiency, and reach.
SR015 DARPA POEM: Photonically Optimized Embedded Microprocessors The POEM program aims to address electrical communications link limitations by developing chip-scale integrated photonic technology.
SR016 GlobalFoundries Silicon Photonics | GlobalFoundries GF’s silicon photonics platform offers supply confidence with the industry’s only high-volume 300mm CMOS manufacturing foundry for silicon photonics.
SR017 Sivers Semiconductors Sivers Semiconductors and Ayar Labs to expand their partnership on enabling high volume manufacturing of optical I/O solutions for scalable cost-effective AI infrastructure - Sivers Semiconductors The partnership framework will focus on product qualification and manufacturing readiness of Sivers high precision laser arrays for high volume deployment.
SR018 Gazettabyte Ayar Labs prepares to fulfil its optical input-output (I/O) vision - Gazettabyte Ayar has partnered with Alchip Technologies and Global Unichip to integrate optical engines directly into compute or switch packages for hyperscalers.
SR019 Laser Focus World A shift to photonics for data center networks and AI is underway Ayar Labs’ chiplets are fabricated by GlobalFoundries on their Fotonix process, while SuperNova comes from multiple suppliers.
SR020 NVIDIA NVIDIA Announces Spectrum-X Photonics, Co-Packaged Optics Networking Switches to Scale AI Factories to Millions of GPUs By integrating silicon photonics directly into switches, NVIDIA is opening the gate to million-GPU AI factories.
SR021 Broadcom Broadcom Advances Optical Connectivity for AI Infrastructure with Industry-Leading Solutions at OFC 2025 Broadcom’s OFC 2025 portfolio highlights a roadmap toward 200T optical interconnect solutions, including XPU-CPO and LPO.
SR022 Broadcom Broadcom Showcases Industry-Leading Quality and Reliability of Co-Packaged Optics Broadcom announced one million cumulative 400G equivalent port device hours of flap-free CPO operation at Meta.
SR023 Marvell Marvell Announces Breakthrough Co-Packaged Optics Architecture for Custom AI Accelerators Marvell says its silicon photonics technology has shipped for over eight years with more than 10 billion device hours of field operation.
SR024 Marvell Marvell to Acquire Celestial AI, Accelerating Scale-up Connectivity for Next-Generation Data Centers Marvell is acquiring Celestial AI to accelerate scale-up connectivity for next-generation data centers.
SR025 Lightmatter 3D Photonics for AI Applications | Passage™ Lightmatter markets Passage as 3D photonics for AI applications with scalable wavelength and switching architecture.
SR026 LightCounting Highlights from the 1st virtual conference on Co-Packaged Optics (CPO) High volume deployments are likely in 2027, LPO deployments are starting now, and re-timed pluggables are not going away.
SR027 Optical Connections News AI drives demand for optical transceivers, LPO, CPO - report The analyst expects rapid growth to moderate in 2026-2027 as the first wave of excitement about AI subsides.
SR028 Lightwave Online Co-Packaged Optics and the AI data center: From skepticism to strategic adoption Many users remain in a quandary, wondering whether they need CPO and whether it will create more operational risk than performance benefit.
SR029 Lightwave Online AI scale-up raises the stakes for transceivers and CPO adoption LightCounting forecasts moderation in market growth for 2027-2031 and says supply-chain shortages restore some sanity.
SR030 Light Reading OFC 2026: Optical networking's AI reckoning arrives The optics debate now centers on a fragmented set of approaches, with co-packaged optics, linear pluggables and near-packaged optics competing for position.
SR031 Gazettabyte OFC Reflections - Part 2 - Gazettabyte Co-packaged optics integration is inevitable but not imminent, and the excess hype around integrated optics confirmed an investment bubble.
SR032 APNIC Blog Co-Packaged Optics — a deep dive | APNIC Blog Many will interpret CPO as vendor lock-in disguised as innovation.
SR033 TechPowerUp TSMC CoWoS Capacity Doubles for Two Years, Still Insufficient: TrendForce TSMC CoWoS capacity is doubling but demand continues to outpace supply, and 2026 delivery schedules are effectively booked.
SR034 CNBC Tech AI spending may approach $700 billion this year, but the blow to cash raises red flags Tech AI spending may approach $700 billion this year, but the blow to cash raises red flags.
SR035 Cleary Foreign Investment and International Trade Watch U.S. Government Signals Intent to Increase Enforcement of U.S. Export Controls Commerce officials signaled a dramatic increase in BIS enforcement focused on China, AI, semiconductors and supercomputing.
SR036 Federal Register / BIS Implementation of Additional Due Diligence Measures for Advanced Computing Integrated Circuits; Amendments and Clarifications; and Extension of Comment Period The interim final rule provides additional due diligence procedures regarding advanced computing integrated circuits and assists foundries and OSATs in complying with the EAR.
SR037 Bureau of Industry and Security EAR | Bureau of Industry and Security BIS says the EAR resource is intended to help users read, search and download specific sections of the regulations and compliance materials.
SR038 Justia Patents Patents Assigned to Ayar Labs, Inc. The assignee page lists multiple Ayar Labs patents across optical waveguides, lasers and photonic integration.
SR039 Justia Patents Patents Assigned to Lightmatter, Inc The assignee page lists Lightmatter patents and applications related to photonics-based processors and interconnects.
SR040 Justia Patents Patents Assigned to Rockley Photonics Limited The assignee page lists Rockley Photonics patents in photonic sensing and integrated photonics.
SR041 USPTO PTAB Patent Trial and Appeal Board The Patent Trial and Appeal Board remains the venue for reviewing patentability challenges and other post-grant proceedings.
SR042 Business Wire Rockley Photonics Completes Financial Restructure, Emerges from Chapter 11 Rockley Photonics completed a comprehensive financial restructuring and emerged from Chapter 11 after filing for bankruptcy protection in Q1.
SR043 Data Center Dynamics Celestial AI acquires Rockley Photonics patent portfolio for $20m Celestial AI acquired Rockley Photonics’ patent portfolio for $20 million, showing how distressed photonics IP can move into competitors’ hands.
SV001 BusinessWire (Ayar Labs) ADDING MULTIMEDIA Ayar Labs, with Investments from AMD, Intel Capital, and NVIDIA, Secures $155 Million to Address Urgent Need for Scalable, Cost-Effective AI Infrastructure This brings the company’s total funding to $370 million and raises the company’s valuation to above $1 billion.
SV002 Crunchbase News Ayar Labs, Latest Optical Interconnect Startup To Raise Big, Hits Unicorn Status Ayar Labs locked up a $155 million Series D led by Advent International and Light Street Capital at a valuation of more than $1 billion.
SV003 BusinessWire (Ayar Labs) Ayar Labs Raises $130 Million in Series C Funding, Accelerating Commercialization of Industry’s First In-Package Optical I/O Products Ayar Labs also announced that it made its first volume commercial shipments under contract and expects to ship thousands of units of its in-package optical interconnect by end of year.
SV004 Hewlett Packard Enterprise Hewlett Packard Enterprise and Ayar Labs announce strategic collaboration and investment to develop next-generation data center architectures and networking with optical I/O HPE’s venture arm, Hewlett Packard Pathfinder, has made a strategic investment in Ayar Labs.
SV005 Ayar Labs Hewlett Packard Enterprise and Ayar Labs Announce Strategic Collaboration and Investment to Develop Next-Generation Data Center Architectures and Networking with Optical I/O
SV006 PR Newswire (GlobalFoundries) GLOBALFOUNDRIES ANNOUNCES NEXT GENERATION IN SILICON PHOTONICS SOLUTIONS AND COLLABORATES WITH INDUSTRY LEADERS TO ADVANCE A NEW ERA OF MORE IN THE DATA CENTER Since our earliest days, Ayar Labs and GlobalFoundries have partnered on the development of GF Fotonix.
SV007 DARPA PIPES | DARPA PIPES seeks to embed optical signaling technologies within the package of ASICs and FPGAs to enable data links with unprecedented bandwidth density, efficiency, and reach.
SV008 Ayar Labs Ayar Labs: AI Scale-up Beyond the Rack
SV009 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.
SV010 Ayar Labs SuperNova™ light source The 16 wavelength SuperNova light source ... can drive 256 optical carriers for 16 Tbps of bi-directional bandwidth.
SV011 Ayar Labs Innovators in Co-Packaged Optics for LLM Data Transfer | Ayar Labs Ayar Labs’ solution combines two industry-first technologies — the TeraPHY™ optical I/O chiplet and SuperNova™ multi-wavelength remote light source — to maximize data transfer and compute efficiency while reducing costs, latency, and power consumption.
SV012 Gunderson Dettmer Lightmatter Announces $400M Series D and $4.4B Valuation This new capital values the company at $4.4 billion and brings the total capital raised to date to $850 million.
SV013 Lightmatter Lightmatter Raises $400M Series D; Quadruples Valuation to $4.4B as Photonics Leader for Next-Gen AI Data Centers
SV014 BusinessWire (Celestial AI) Celestial AI Secures $250 Million Funding to Revolutionize AI Infrastructure with Its Photonic Fabric™ Celestial AI ... has raised $250 million in its Series C1 funding round ... bringing the total capital raised to date to more than $515 million.
SV015 Yahoo Finance (Bloomberg) Fidelity Backs Chip Startup Celestial AI at $2.5 Billion Value Celestial AI’s valuation rose to more than $2.5 billion following the latest financing, according to its CEO David Lazovsky.
SV016 Kroll Restructuring Administration Kroll Restructuring Administration On January 23, 2023, Rockley Photonics Holdings Limited ... filed a voluntary petition for relief under Chapter 11.
SV017 Pillsbury Law Pillsbury Team Confirms Prepackaged Chapter 11 Plan for Rockley Photonics Holdings Limited Rockley ... restructure[d] its business through transactions that eliminated more than $120 million of secured debt and generated approximately $35 million in new funding.
SV018 Bankruptcy Observer ROCKLEY PHOTONICS HOLDINGS LIMITED Bankruptcy Case in Southern District of New York ROCKLEY PHOTONICS HOLDINGS LIMITED filed a 11 chapter bankruptcy in the Southern District of New York bankruptcy court on January 23, 2023.
SV019 POET Technologies Investors | POET Technologies
SV020 Securities and Exchange Commission EDGAR Entity Landing Page
SV021 MarketBeat POET Technologies (POET) Stock Price, News & Analysis $POET
SV022 Yahoo Finance POET Technologies Inc. (POET) Valuation Measures & Financial Statistics
SV023 MarketBeat Lumentum (LITE) Stock Price, News & Analysis $LITE
SV024 Yahoo Finance Lumentum Holdings Inc. (LITE) Valuation Measures & Financial Statistics
SV025 MarketBeat Coherent (COHR) Stock Price, News & Analysis $COHR
SV026 Yahoo Finance Coherent Corp. (COHR) Valuation Measures & Financial Statistics
SV027 Cisco Cisco Completes Acquisition of Acacia Communications, Inc. Cisco has agreed to acquire Acacia for $115.00 per share in cash, or approximately $4.5 billion on a fully diluted basis.
SV028 Yahoo Finance (Reuters) Marvell to buy Inphi in $10 billion chip deal to bolster data center, 5G business Marvell ... has agreed to buy peer Inphi Corp in a $10 billion cash-and-stock deal.
SV029 MarketScreener Marvell : Completes Acquisition of Inphi
SV030 Lightwave Online Co-Packaged Optics and the AI data center: From skepticism to strategic adoption Outside hyperscale environments, the average data center manager may know little about CPO.
SV031 Cignal AI Co-Packaged Optics: Inevitable but Not Imminent - Cignal AI Large-scale CPO deployment is not expected for 3-5 years, but some initial deployments will start in 2026.
SV032 LightCounting Highlights from the 1st virtual conference on Co-Packaged Optics (CPO) The level of development activity on CPO is at an all time high now with high volume deployments likely in 2027.
SV033 Ciena Ciena to Acquire Nubis Communications to Expand its Inside the Data Center Strategy and Further Address Growing AI Workloads Under the terms of the agreement, Ciena will acquire Nubis in an all-cash transaction for $270 million.
SV034 Coherent II-VI Incorporated Completes the Acquisition of Coherent Each share of Coherent common stock was converted into the right to receive $220.00 in cash and 0.91 of a share of II-VI common stock.
SV035 TechPowerUp Ayar Labs, with Investments from AMD, Intel Capital, and NVIDIA, Secures $155 Million Series D Funding
SV036 Premier Alternatives Scintil Photonics Valuation: $133.6M (2026) Scintil Photonics is currently valued at $133.6M as of August 29, 2025. The company has raised a total of $82.5M in funding.