Olix
产品尚未自证,估值已按赢家定价的高确信光子推理赌注
Olix 已具备成为严肃光子推理玩家的几块材料,但当前独角兽估值已经提前消化了仍未公开证实的技术和商业里程碑。
封面要素
公司概况
Olix 是一家总部位于伦敦的光子 AI 硬件初创公司,2024 年 3 月注册成立,2026 年 1 月由 Flux Corp Ltd 更名而来。公司在打造以 SRAM 加光子推理为核心的 DX-1 / OTPU 架构,主张传统围绕 HBM 的加速器无法在规模化场景下同时经济地提供吞吐和交互性。公开证据支持其在 2026 年 2 月完成由 Hummingbird Ventures 领投的 $220M Series A,估值超过 $1B,但公司尚未公开披露收入、设计导入、tapeout 完成情况或代工伙伴。因此,Olix 的资本形成和技术野心在同龄公司中异常突出,但商业就绪度的公开证据仍然有限。
- 官网
- olix.com
- 成立时间
- 2024-03-22
- 创始人
- James Dacombe
- 创立地点
- London, United Kingdom
- 总部
- London, United Kingdom
- 产品
- Olix 在开发 DX-1 / OTPU 平台,一个面向推理的加速器系统,采用以 SRAM 为中心的内存架构并整合光子技术。公开材料将其描述为机架级、跨层系统,覆盖逻辑、数据搬运、光学、封装,以及面向 vLLM、SGLang、NVIDIA Dynamo 等框架的服务软件集成。
- 客户
- 具备大规模推理负载的 hyperscaler、AI 云服务商和前沿模型运营方
- 商业模式
- 硬件加速器销售、配套系统软件集成,以及潜在平台授权
- 阶段
- Series A, pre-revenue / undisclosed-revenue
- 融资情况
- 2026 年 2 月完成 $220M Series A,投后估值 $1B+;披露总融资约 $250M
执行摘要
主要优势
- 早期融资能力强:$220M Series A、估值 $1B+,董事会周围聚集了高信号投资人与运营者网络
- 架构假设清晰差异化:SRAM 加光子、面向解码的系统设计,打的是推理真实瓶颈,而不是 GPU 小修小补
- 招聘证据显示,公司在光学、ASIC、分布式推理软件和制造运营上都在搭建严肃的跨职能团队
主要风险
- 尚未有收入,也尚未证明:截至 2026-06-22,公开信息没有设计胜出、tapeout 里程碑、晶圆厂披露或具名客户
- 光子 / CPO 类系统的制造和生态风险仍很尖锐,尤其是良率、热管理、测试和互操作性
- 相比已经出货或披露客户证据的同行,$1B+ 估值显得偏满
- 创始人 James Dacombe 是关键人物依赖点,治理和经济结构仍不透明
未决问题
- 具名设计胜出、客户 LOI 或任何付费试点证据
- Tapeout 日期、制程节点、foundry / OSAT 合作伙伴和硅片验证状态
- Series A 后股权结构、优先权栈和所有权集中度
- 当前经审计员工人数、现金消耗、收入或积压订单指标
目录
01公司概览
1.1 身份、成立与产品论点
Olix 的运营主体是 OLIX Computing Limited,这家英国私人公司于 2024 年 3 月 22 日注册,2026 年 1 月 19 日由 Flux Corp Ltd 更名。注册地址在 St Albans,但融资轮前后所有面向公众的运营描述都称公司总部在伦敦;招聘网站显示,公司在英国的伦敦、Bristol,以及北美的 Austin、San Francisco、Toronto 招人。这个差异在尽调中很重要:法律实体年轻、披露很轻,而市场身份已经是伦敦 AI 硬件独角兽。 公司选择的产品论点不同于标准 GPU 挑战者。官方材料称,Olix 希望用 SRAM 中心内存设计叠加光子技术,为「frontier AI」建设基础设施,避开对 HBM 重型加速器架构的依赖。公开材料同时提到 Optical Tensor Processing Unit(OTPU)和 OLIX Decode Accelerator 1(DX-1);宣言和职位页面都称系统专门面向推理和 decode-heavy 工作负载。官方与第三方来源传递的核心信息一致:Olix 认为现有加速器迫使用户在吞吐和交互性之间付出高成本取舍,而光子、机架级架构有机会打破这个约束。 因此,Olix 的高确信身份可信,但也暴露公司仍处在很早期。公司把论点、目标产品族和招聘议程讲得很清楚,却很少给出生产指标、客户胜利或已出货硅片。对一家 2024 年成立的 deep-tech 初创公司来说,这种不对称并不罕见,但后文评估商业化风险时必须保留这一点。[CO001, CO002, CO003, CO004, CO005, CO006]
| 指标 | 数值 / 状态 | 日期 / 期间 | 置信度 | 缺口 / 备注 |
|---|---|---|---|---|
| 法定名称 | OLIX COMPUTING LIMITED | 2026-06-22 | 高 | Companies House 公司 15585099 |
| 品牌 / 网站 | OLIX / olix.com(官网) | 2026-06-22 | 高 | 运营品牌与法定名称格式不同 |
| 运营总部视角 | 伦敦创业公司 | 2026-02 至 2026-06 | 中 | 媒体描述;注册地址在 St Albans |
| 注册地址 | Arquen House, 4-6 Spicer Street, St. Albans, AL3 4PQ(注册地址) | 2026-06-22 | 高 | 法定地址,不一定是运营总部 |
| 注册成立 | 2024 年 3 月 22 日 | 历史 | 高 | 最初注册为 Flux Corp Ltd |
| 当前阶段 | Series A 轮 | 2026-02 | 高 | Cooley 与多篇新闻报道相互印证本轮融资 |
| 最近一轮规模 | $220M 融资 | 2026-02-11 | 高 | 领投方 Hummingbird Ventures |
| 隐含估值 | 略高于 $1B | 2026-02-11 | 高 | Cooley 表述;媒体四舍五入为 $1B+ |
| 总融资 | 媒体报道约 ~$250M | 2026-02 | 中 | 媒体汇总;未在备案中核验 |
| 产品族 | OTPU / DX-1 推理加速器 | 2026-06 | 中 | 官方页面同时提到两个术语 |
| 首批出货目标 | 2027 | 2026-02 至 2026-06 | 中 | 多家媒体报道,并非合同承诺 |
| 员工人数 | >70 人(报道);目标 200+;2025-08 法定账目为 13 人 | 2025-08 至 2026-02 | 低 | 当前人数未经公开审计 |
| 收入 / ARR | 未公开披露 | 2026-06-22 | 低 | 私营公司 |
| 客户 | 未公开披露 | 2026-06-22 | 低 | 公开未确认任何具名量产客户 |
该表合并监管备案、公司官方页面和媒体报道。若当前运营指标未披露,表格保留缺口,不推断数字。
[CO001, CO003, CO004, CO013, CO015, CO019]Olix 成熟度、资本位置、披露画像和执行状态的压缩视图。
[CO003, CO013, CO014, CO018, CO019, CO023]1.2 创始人、董事与治理信号
核心关键人事实很直接:James Robert Dacombe 创立公司,自注册以来一直是 Companies House 记录中唯一连续任职的董事,也是融资叙事中公开出现的创始人。Companies House 记录显示,他于 2024 年 3 月 22 日获任;媒体报道反复称他是 25 岁创始人,也是推动 Olix 反共识计算战略的 CEO。这个组合让 Dacombe 既是募资资产,也是执行上的单点依赖。 2025-2026 年融资窗口前后,治理结构迅速成熟。Companies House 的高管名单和申报历史显示,Jonathan Heiliger 于 2025 年 6 月加入,Taavet Hinrikus 于 2025 年 10 月加入,Firat Ileri 和 Professor Nicholas McKeown 于 2026 年 2 月加入;2026 年 2 月 10 日的董事变更移除了 David Helgason 和 Krishna Visvanathan。Saul Klein 也出现在现任高管名单中。结合外部履历看,这套董事和董事会阵容把 Olix 连接到基础设施、欧洲创投和网络领域的强资源:Heiliger 曾领导 Facebook 基础设施,Klein 是长期欧洲种子投资人和运营者,McKeown 是 Stanford 网络教授,Hinrikus 是知名欧洲创始人投资人。 公司自己的招聘与文化材料又给出第二个治理信号:Olix 刻意打造高强度、线下办公文化,提供有意义的股权、每年 £24,000 的本地生活补贴,甚至为不匹配的员工在三个月节点提供 18 周带薪退出。这可能增强艰难 deep-tech 搭建期的团队对齐,但也说明公司高度依赖在异常苛刻的运营规范下招聘并留住顶尖技术人才。[CO007, CO008, CO009, CO010, CO011, CO012]
| 人物 | 职务 / 关联 | 公开背景 | 加入 / 状态 | 关键人或治理含义 |
|---|---|---|---|---|
| James Dacombe | 创始人、董事、公开 CEO 形象 | 英国创始人;公开信息也显示与 CoMind 有关联 | 任命于 2024-03-22;在任 | 主要关键人依赖,也是公司身份锚点 |
| Jonathan Heiliger | 董事 / 投资人型运营者 | 前 Facebook 基础设施负责人;Vertex/VVUS 投资人 | 任命于 2025-06-24;在任 | 补入超大规模基础设施运营视角 |
| Taavet Hinrikus | 董事 / 投资人 | 公开履历显示为 Wise 联合创始人、Plural 合伙人 | 任命于 2025-10-13;在任 | 释放高信号欧洲风投支持 |
| Firat Ileri | 董事 / 投资人 | Hummingbird Ventures 投资人 | 任命于 2026-02-10;在任 | 可能代表领投轮治理影响力 |
| Saul Klein | 董事 | Phoenix Court/LocalGlobe 创始合伙人、资深欧洲运营者 | 高管名单中显示在任 | 扩大网络与融资可信度 |
| Nicholas McKeown | 董事 | Stanford 网络教授,聚焦高性能交换机和路由器 | 任命于 2026-02-27;在任 | 增加深度网络架构可信度 |
董事会构成由 Companies House 高管名单和外部履历重建。公司没有发布正式董事会页面,因此除“董事”外的头衔均作保守推断。
[CO007, CO008, CO009, CO010, CO011, CO039]1.3 融资历史、投资人组合与披露边界
最强的已验证融资事实来自 Cooley 2026 年 2 月 11 日公告:Cooley 为 Olix 一轮 $220 million 融资提供法律服务,该轮由总部位于伦敦的 Hummingbird Ventures 领投,对公司的估值略高于 $1 billion。SiliconANGLE、Jon Peddie Research、IDCNova、Tech Funding News 等报道复述了大致相同条款,使 2026 年 2 月这一轮成为后续所有估值和 runway 讨论的锚点。媒体还称,公司此前获得 Plural、Vertex Ventures、LocalGlobe、Entrepreneurs First 支持,并将总融资额放在约 $250 million。 同样重要的是看不见的部分。Olix 仍是一家英国私人公司,没有公开股票研究,没有公开债务披露,没有披露二级交易,也没有随融资发布收入或客户指标。Companies House 申报显示,2026 年 1 月至 5 月围绕资本和治理出现多份文件,包括股份配发、章程采纳、股份权利变更和优先购买权变更,但公开表面仍无法拼出可用 cap table 或投资人持股拆分。 不过,投资人名单方向上很强。Hummingbird 将自己定位为愿意写大额 follow-on 支票的全球种子投资人;Olix 周边与董事会相连的参与者还包括 Plural、Phoenix Court/LocalGlobe、Vertex 和 Stanford 生态人物。这个组合意味着,Olix 比许多欧洲硬件初创公司更早获得了确信资本和运营者监督。但更深财务披露缺位,意味着 syndicate 质量不能替代商业牵引证据。[CO013, CO014, CO015, CO016, CO017, CO018]
| 利益相关方 | 角色 | 参与证据 | 重要性 | 尽调问题 |
|---|---|---|---|---|
| Hummingbird Ventures | 2026 年融资领投方 | Cooley 点名 Hummingbird 为本轮领投;Firat Ileri 于 2026 年 2 月进入董事会 | 可能在最新一轮中拥有最强治理和跟投影响力 | 确认持股、董事会权利和按比例跟投能力 |
| Plural | 早期投资人 | 多篇媒体报道点名;Taavet Hinrikus 出现在高管名单 | 高信号欧洲风投支持 | 确认支票规模和持续董事会角色 |
| Vertex Ventures / Jonathan Heiliger | 早期投资人 / 董事关联 | 媒体点名 Vertex;Heiliger 2025 年 6 月加入董事会 | 带来超大规模基础设施运营者的模式识别 | 澄清 Heiliger 是否正式代表 Vertex |
| LocalGlobe / Phoenix Court / Saul Klein(投资人网络) | 早期投资人 / 董事关联 | 媒体点名 LocalGlobe;Saul Klein 出现在高管名单 | 扩展欧洲创始人网络和政策触达 | 确认法定投资实体和治理权利 |
| Entrepreneurs First | 早期支持者 | 多篇融资报道点名 | 支撑公司创立阶段的起源故事可信度 | 确认时间和证券类别 |
| James Dacombe | 创始人运营者 | 自注册成立起担任创始人 / 董事 | 可能对战略和招聘有最大个人影响力 | 确认股权集中度和归属安排 |
| Nicholas McKeown | 独立技术董事视角 | Companies House 加 Stanford 资料 | 可能对网络和系统可信度很重要 | 澄清范围:战略顾问还是活跃董事会运营者 |
公开来源能识别参与方和部分董事,但不能给出精确证券类别、持股比例或清算优先权。这仍是核心私募尽调问题。
[CO013, CO015, CO016, CO017, CO018, CO030]1.4 运营足迹、招聘信号与里程碑记录
Olix 的公开运营足迹更多通过招聘显现,而不是客户披露。招聘中心列出的开放岗位覆盖光学、FPGA、ASIC、平台软件、法务、封装、制造运营和工艺集成;一家试图把光子计算论点变成可出货半导体系统的公司,本来就需要这样的职能宽度。具体职位页面还显示,Olix 在设计芯片到芯片、tray 内部和机架级光链路,同时为 vLLM、SGLang、NVIDIA Dynamo 等分布式推理栈搭建软件集成。这是一个有意义的信号:公司在架构硅片和系统软件,而不只是做实验室原型。 公开记录中的里程碑集中在三个阶段。第一阶段是 2024 年,公司以 Flux Corp Ltd 名义在 James Dacombe 旗下成立。第二阶段是 2025 年,公司扩张时完成治理扩充和地址变更。第三阶段是 2026 年 1 月至 5 月,公司密集完成更名、融资、资本配发、新董事任命和招聘提速,并以 AI 基础设施独角兽身份进入公众视野。公开报道还反复指向 2027 年为首批客户出货目标。 规模上仍有真实的测量缺口。Jon Peddie 和 Tech Funding News 称 Olix 已增长到 70 多名员工,并计划超过 200 人;但 Companies House 可获得的最新法定账目只覆盖到 2025 年 8 月 31 日。因此,投资人有证据证明公司在扩张,但没有一份干净、经审计的运营仪表盘,能够把历史法定披露与当下招聘说法对齐。[CO019, CO020, CO021, CO024, CO025, CO026]
| 日期 | 事件 | 类型 | 金额 / 状态 | 参与方 | 含义 |
|---|---|---|---|---|---|
| 2024-03-22 | Flux Corp Ltd 注册成立 | 创立 | James Dacombe | Olix 实体完成法定设立 | |
| 2025-06-24 | Jonathan Heiliger 被任命为董事 | 治理 | Heiliger / 公司 | 董事会开始加入外部基础设施运营者 | |
| 2025-10-13 | Taavet Hinrikus、David Helgason 和 Krishna Visvanathan 的董事变更完成备案 | 治理 | 公司和投资人关联人士 | 重大融资前治理扩容 | |
| 2025-11-05 | 注册地址从 Oxford 迁至 St Albans | 扩张 | 公司 | 公司扩张期间的行政调整 | |
| 2026-01-19 | 公司由 Flux Corp Ltd 更名为 OLIX COMPUTING LIMITED | 治理 | 公司 | 公开品牌重塑让法定实体与 Olix 品牌对齐 | |
| 2026-02-10 | Firat Ileri 被任命;Helgason 和 Visvanathan 离任 | 治理 | 公司 / Hummingbird 关联董事 | 领投轮治理重置 | |
| 2026-02-11 | $220M 融资公开,估值 >$1B | 融资 | $220M / $1B+ | Hummingbird Ventures 和辛迪加 | 阶跃式融资确立独角兽地位 |
| 2026-02-18 | 章程、股份权利和优先购买权决议完成备案 | 治理 | 公司 / 股东 | 融资前后更新资本结构 | |
| 2026-02-27 | Nicholas McKeown 教授被任命为董事 | 治理 | McKeown / 公司 | 董事会补入网络与系统可信度 | |
| 2026-03 至 2026-05 | 多项股份配发和资本备案完成记录 | 融资 | 已提交资本声明 | 公司 / 股东 | 完成融资后资本化机制 |
| 2026-06 | 官方网站和招聘页面围绕 DX-1、光学和分布式推理 | 产品 | 公司 | 释放从阐述投资论点转向执行搭建的信号 | |
| 2027 年目标 | 媒体报道首批客户出货目标 | 产品 | 仅为目标 | 公司 / 潜在客户 | 后续尽调的商业时间锚 |
日期取自 Companies House;缺失处使用带日期的媒体报道。2027 年条目是管理层目标,而不是已验证的商业交付。
[CO003, CO013, CO019, CO025, CO026, CO027]2024 年至 2027 年出货目标期间,Olix 的法律设立、治理搭建、融资、改名和公开执行里程碑。
[CO013, CO019, CO025, CO026, CO027, CO028]1.5 负面信号与开放尽调问题
Olix 的承诺位于一个战略上很有吸引力、但运营上极不宽容的技术栈中。官方材料自己强调,一家初创公司无法现实地按巨头条件争夺 HBM 和先进封装产能,所以 Olix 试图绕开这些卡点。这个论证在逻辑上成立,但也等于承认:主流 AI 硬件供应链仍有结构性约束,Olix 选择新架构,部分原因是标准路径难以进入。 独立市场评论强化了这种谨慎。Siemens 2026 年 co-packaged optics 说明称,功耗、热管理、测试、良率和互操作性在 2026-2027 年仍是生态瓶颈。Edgen 援引供应链报道称,由于良率仍然过低,CPO 大规模量产不太可能在 2026 年之后很快实现。CNBC 2026 年 5 月下旬的光子专题用更宽的方式给出同一判断:用光缓解 AI 瓶颈是行业最有意思的方向之一,但商业化相对既有 GPU 生态仍处早期。即便 Olix 的具体架构不同于以交换芯片为中心的 CPO 路线图,制造和生态预警仍然重要,因为 Olix 正在从同一个不成熟供应链中招聘光互连、封装和系统集成人才。 第二个尽调问题是治理透明度。Companies House 揭示了桌边坐着谁、何时加入,却不揭示职位背后的经济权益。资本结构、股权集中度、客户承诺、晶圆 / fab 安排,以及 2026 年融资所绑定的里程碑仍是私人信息。对早期 deep-tech 公司来说,这可以接受,但也意味着后续投资判断必须显式处理不确定性,而不是假装精确。[CO032, CO033, CO034, CO035, CO036, CO042]
Olix 如何把前沿 AI 使命连接到产品架构、人才扩张、资本,以及它试图绕开的外部瓶颈。
[CO005, CO006, CO013, CO019, CO024, CO032]1.6 展示项
02市场分析
2.1 市场边界、替代品与相邻市场
Olix 位于三个同心市场的交叉点。最外层是全球 AI 加速器芯片市场,覆盖数据中心、云平台和边缘部署中用于 AI 训练和推理的所有定制硅片与 GPU。第二层是 AI 推理加速器子市场,排除面向训练优化的硬件(NVIDIA H100/H200 类),聚焦为推理延迟、吞吐和 token 解码效率设计的芯片与系统。Olix 的 DX-1 Decode Accelerator 正处在这一推理层。第三层是更具体的技术市场:硅光子和光子集成电路,这是 Olix 试图商业化的使能技术层;该市场也服务电信收发器、lidar 和生物医疗传感器,意味着 AI 只占总量中不断增长但仍然部分的一块。 Olix 必须替代的现状方案包括:(1)在 HBM 耦合的 Blackwell 硬件上运行 vLLM 或 SGLang 式推理服务的 NVIDIA GPU;(2)Google TPU、Meta MTIA、Amazon Trainium/Inferentia、Microsoft Maia 等 hyperscaler 自研 ASIC;(3)可插拔光模块叠加传统铜质机架级互连,Broadcom CEO 曾公开称这会是硅光子变得必要前的中间步骤。Olix 的主要 SAM 不包括:AI 训练负载、边缘 / 移动端 AI 推理,以及 LLM 式架构之外的一般 HPC 工作负载。 值得跟踪的相邻市场包括 co-packaged optics 供应链(Ayar Labs、Broadcom photonics、Intel、Marvell 等 optical engine 供应商)、高带宽内存市场(Olix 明确希望用 SRAM 中心设计降低依赖),以及服务机架级和 pod 内连接的光网络 / 收发器市场。AI 数据中心电力与冷却市场是需求驱动,而非直接相邻:IEA 预计全球数据中心用电量到 2030 年将增加一倍以上至 945 TWh,AI 优化设施的份额将扩大四倍,这会形成强烈动机,推动市场采用能效更高的推理硬件。[CM001, CM002, CM015, CM025, CM026, CM035]
| 细分 / 类别 | 纳入支出 | 排除支出 | 主要买方 / 付款方 | 与 Olix 的相关性 |
|---|---|---|---|---|
| AI 推理加速器(商用芯片) | 用于推理服务和解码工作负载的定制 ASIC 与光子芯片 | 训练芯片、用于模型训练的 GPU 集群 | 超大规模云厂商基础设施团队;云业务 P&L | 直接 TAM;Olix DX-1 在这里竞争 |
| 硅光子市场(AI 份额) | 部署在 AI 推理和 CPO 中的光子 IC 与光引擎 | 电信收发器、lidar、生物传感 | 超大规模云厂商光学基础设施;CPO 集成商 | Olix OTPU 的技术层 SAM |
| 共封装光学(CPO)生态 | 与加速器和网络 ASIC 共封装的光引擎 | 可插拔光学器件、铜 DAC/AEC 线缆 | 超大规模云厂商、OEM、ODM | 邻近市场;CPO 采用打开机架级光子架构 |
| 超大规模云厂商定制 AI 加速器 | Google TPU、Meta MTIA、Amazon Trainium/Inferentia、Microsoft Maia 等自研 ASIC | 商用 GPU 市场(NVIDIA Blackwell) | 超大规模云厂商 AI 平台组织(内部采购) | 现状替代品;每一颗内部 ASIC 都会压缩 Olix 可服务基数 |
| AI 推理即服务(云) | 云托管推理 API、推理算力租赁 | 边缘 / 移动推理;消费设备 | 科技公司的 CTO / AI 产品团队 | 次级买方;若每 token 成本下降,可能加速 Olix 采用 |
纳入 / 排除支出基于公开产品范围和买方描述作定性划分。截至运行日期,未找到把 AI 推理与训练按金额拆分的独立市场边界数据;TM002 中估计值除非另有说明,均反映总加速器或硅光子市场。null 单元格表示不适用。
[CM035, CM036, CM025, CM039]2.2 市场规模——TAM、SAM 与受约束的 SOM
Olix 最宽口径的 TAM 是全球 AI 加速器市场:Grand View Research 将 2024 年规模估为 $25.56 billion,并预计到 2033 年增至 $256.84 billion,CAGR 为 29.3%。在这个范围内,推理负载是增长最快的部分。NVIDIA 自身财报披露暗示,归因于推理的年度数据中心计算支出已经达到每年数百亿美元量级,并仍在增长:仅 FY2026 Q3 一个季度,数据中心收入就达到 $51.2 billion;Jensen Huang 提到,过去十二个月 AI 推理 token 生成量增长了十倍。 硅光子技术市场为 Olix 的 SAM 提供了更窄代理。MarketsandMarkets 估计,全球硅光子市场将从 $2.65 billion(2025)增长到 2030 年 $9.65 billion,CAGR 为 29.5%;北美在此区间内从 $1.16 billion 增至 $4.35 billion。The Research Insights 给出趋同但略低的估计:2030 年 $8.13 billion。两组估计都把电信和 datacom 应用与 AI 推理一并纳入,因此纯推理切片更小——但按三家分析来源的说法,AI / 数据中心份额都是主要增长引擎。 2027–2029 年 Olix 受约束的 SOM 取决于尚未公开的变量:客户认证时间线、光子代工良率,以及 hyperscaler 是把推理拉回自研,还是采用商用硅片。分析师估计显示,在 CPO 和光子推理获得有意义的 hyperscaler 牵引时,到 2029–2030 年,可服务的年度推理光子收入机会可能达到 $1–3 billion。不过,这一 SOM 估计明确受证据约束,应视为方向性区间,而不是硬预测。[CM005, CM006, CM007, CM009, CM010, CM011]
| 发布方 | 年份 | 地域 | 数值 | CAGR | 方法论 | 置信度 | 局限 |
|---|---|---|---|---|---|---|---|
| Grand View Research | 2025–2033 | 全球 | $25.6B(2024 基准)→ $256.8B (2033) | 29.3% | 自下而上的行业模型 | 中 | 覆盖训练 + 推理;没有光子专属拆分 |
| MarketsandMarkets | 2025–2030 | 全球 | $2.65B (2025) → $9.65B (2030) | 29.5% | 自下而上;仅硅光子产品 | 中 | 包含电信 / 数据通信;未隔离 AI 推理份额 |
| MarketsandMarkets | 2025–2030 | 北美 | $1.16B (2025) → $4.35B (2030) | 30.2% | 硅光子报告中的区域子细分 | 中 | 与全球数字有相同范围限制 |
| The Research Insights | 2025–2030 | 全球 | $8.13B (2030) | ~25.8%(隐含) | 独立分析师估计 | 低 | 新闻稿未披露方法论;下限估计 |
| IDTechEx | 2024–2034 | 全球 | 10 年出货预测(公开摘要未披露金额) | 未披露 | 一手研究,AI 与计算用光子 IC | 低 | 完整报告付费;公开仅有章节标题 |
| 隐含 SOM(作者估计) | 2027–2030 | 超大规模云厂商推理(全球) | 每年 $1–3B(方向性区间) | n/a | 受约束测算:硅光子 SAM 中的光子推理切片 | 低 | 推导值;没有公开一手来源;Olix 未披露客户管线 |
数值除非另有说明,均按“数值”列所示年份表达。CAGR 按来源原文列示。SOM 行是作者推导的方向性区间,不是有来源支撑的数字。置信度反映来源方法论透明度,以及市场边界与 Olix 实际可服务范围的匹配程度。
[CM009, CM010, CM011, CM013]从全球 AI 加速器 TAM 到 Olix 受限推理光子 SOM 的嵌套规模层。
TAM 值以 GVR 2024 年基数按所述 CAGR 线性插值至 2026 年。SAM 是作者估计,将约 45–50% 的 AI/DC 收入占比应用于 MarketsandMarkets 对 2026 年硅光市场的预测。SOM 是方向性区间,没有一手来源。
[CM009, CM013, CM035]多个独立机构对 2030 年全球硅光市场价值的估计,单位为十亿美元。
所有数值均为十亿美元,预测至 2030 年。Research Insights 和 MarketsandMarkets 的数值来自已发布新闻稿。北美数字是全球 MarketsandMarkets 估计中的子区域。AI/DC 子细分行由作者估计,将约 45–50% 的 AI 占比应用于全球 MarketsandMarkets 数字;它不是单独来源估计。所有估计均包括非 AI 应用(电信、激光雷达、生物医疗)。
[CM009, CM010, CM011]2.3 买方、用户与付款方拆分
Olix 初始商业化阶段的潜在买方,很可能集中在 hyperscaler 和大型云服务商。这些客户的推理资本强度足够高,才有理由承受非标准硅片认证周期。Meta 2026 年 Q1 单季度资本开支为 $19.8 billion,并将 2026 全年指引上调至 $125–145 billion。Alphabet 2026 年 Q1 capex 约 $35.7 billion,全年指引为 $180–190 billion。Microsoft FY2026 总 capex 预计约 $190 billion。这些承诺说明,早期采用者的主要约束不是预算,而是认证风险和生态就绪度。 在 hyperscaler 推理硅片采购周期里,买方通常是向基础设施 VP 或 CTO 职能汇报的基础设施或 AI 平台组织。用户是 ML 平台工程团队,他们要把加速器接入 vLLM、SGLang、NVIDIA Dynamo 等推理服务框架——值得注意的是,Olix 自己的招聘信息确认公司在为这些框架建设软件集成,说明它理解用户集成要求。付款方则是管理推理 cost-per-token 和总拥有成本的 Cloud and AI 损益职能。 次级买方包括 inference-as-a-service 云服务商(CoreWeave、Lambda Labs、Together AI 及同类公司),这些公司已经表现出较早采用非 NVIDIA 硅片的意愿;还包括明确建设本国计算能力的国家 AI 基础设施项目。NVIDIA FY2026 Q3 财报提到,「世界各国都在把 AI 视为关键基础设施」;Olix 总部所在的英国,也被点名为 NVIDIA 及合作伙伴投资 £2 billion 建设下一代 AI 基础设施的市场。这个 national-champion 角度可能降低 Olix 在欧洲主权算力采购中的采用摩擦。[CM020, CM021, CM022, CM023, CM024, CM038]
| 细分 | 买方角色 | 用户角色 | 付款方 / 预算负责人 | 工作流 / 使用场景 | 主要采用触发因素 |
|---|---|---|---|---|---|
| 一级 hyperscaler | 基础设施 VP / CTO | ML 平台工程 | 云与 AI P&L | LLM 推理服务(vLLM、SGLang、Dynamo) | 单 token 推理成本超过 GPU 阈值;能源约束 |
| 云推理提供商 | 基础设施负责人 | GPU / 加速器软件团队 | 业务单元 P&L | API 推理、批处理 | 相对 NVIDIA 的算力成本套利;差异化 |
| 主权 / 国家级 AI 计划 | 政府 CTO / 国家 AI 办公室 | 研究与公共部门 ML 团队 | 政府 AI 基金或刺激资金 | 国家算力容量;战略自主 | 政策要求;本土 AI 基础设施目标 |
| 企业私有云 | CTO / 工程 VP | AI/ML 运营团队 | 技术资本开支预算 | 为合规或 IP 保护在本地部署 LLM 推理 | 降低电费;实时推理延迟 SLA |
各行根据公开信号估算(hyperscaler 财报、Olix 招聘页面、NVIDIA 合作伙伴公告)。Olix 尚未披露客户名单。预算负责人行反映公开 hyperscaler 披露中常见的组织模式,并非 Olix 的特定证据。
[CM020, CM021, CM022, CM008, CM038]Olix 主要可寻址客户细分中,买方、用户和付款方角色的映射。
决策周期按定制芯片采购的行业惯例估算。运行日期没有可用的 Olix 特定客户数据。
[CM020, CM021, CM022, CM038]2.4 增长驱动、采用约束与证据缺口
Olix 所在市场的首要需求驱动,是 AI 数据中心的电力与能源危机。IEA 的 Energy and AI 专题报告预计,数据中心用电需求到 2030 年将增加一倍以上至 945 TWh;AI 优化设施的用电量将扩大四倍。在美国,数据中心预计将贡献 2030 年前近一半的电力需求增长。这些约束制造了结构性动机:市场需要每 token 能效显著更好的推理硬件——这正是光子架构的核心价值主张。Data Center Frontier 对 IEA 数据的分析指出,到 2026 年,AI 行业用电量预计将达到 2023 年的十倍;仅 NVIDIA 在 2024 年生产周期出货的 AI 服务器,年耗电量就约 7.3 TWh。 最大的采用约束,是行业巨头已经描述过的成熟度顺序。Broadcom CEO Hock Tan 明确提出两波模型:第一,铜基机架级互连必须先被推到极限;第二,可插拔光模块要成熟;只有到那之后,硅光子和 CPO 的经济性才会显现。Broadcom 已拥有 $73 billion 的定制 AI 加速器订单 backlog,其中来自 hyperscaler 的定制 XPU 订单超过 $50 billion,说明巨头并没有等待光子方案。第二个约束是制造就绪度:edgen.tech 报道,CPO 因良率问题量产推迟到 2026 年之后;Siemens CPO 分析也确认,硅光子制造波动高于成熟 CMOS 工艺,良率偏离会放大成本冲击。DARPA 于 2026 年启动、资助 $35 million 的 PICASSO 计划明确承认,光子电路由于光衰减、噪声和杂散波干扰,很难在系统层面展现相对电子系统的性能优势——这是基础物理约束,不只是工程缺口。 尽管存在这些约束,结构性需求驱动仍有利于最终采用:hyperscaler 资本承诺的规模(仅 Alphabet 和 Microsoft 每年就 $180–190 billion)、推理 token 量的指数级增长,以及能效压力,都会形成长期拉力。关键尽调问题是,Olix 能否及时跨过客户认证和制造就绪门槛,赶上 2027–2028 年 hyperscaler 预算周期。[CM001, CM002, CM003, CM004, CM016, CM017]
| 驱动因素 / 约束 | 方向 | 时间 | 对 Olix 的含义 | 尽调问题 |
|---|---|---|---|---|
| AI 数据中心用电需求翻倍 | 驱动因素 | 2025–2030(持续) | 能效压力上升,有利于光子推理 | 验证 IEA 预测是否成立;确认 hyperscaler 的可持续发展承诺 |
| Hyperscaler 资本开支加速(每家公司 $150–190B/年) | 驱动因素 | 2026(当前) | 显示预算可用;但资本开支主要投向 NVIDIA Blackwell | 确认推理预算是否独立于训练预算 |
| AI 推理 token 量同比激增 10× | 驱动因素 | 2025(NVIDIA CEO 在 FY2026 Q1 表述) | 推理市场快速扩张,放大 Olix 的 SAM | 用独立需求侧证据验证增速 |
| Broadcom CEO:硅光子「短期内不会重要」 | 约束 | 近期(2025–2027) | 暗示买方可能推迟采购光子硅 | 判断 hyperscaler 基础设施买方是否普遍持有该观点 |
| CPO 因良率问题推迟到 2026 年之后量产 | 约束 | 2026–2027 | 光子 I/O 供应链可能还无法支撑 Olix 的 2027 年目标 | 向晶圆厂 / OSAT 获取 CPO 认证时间线 |
| DARPA PICASSO:光子电路的物理约束 | 约束 | 研究阶段(2026–2028) | 行业层面的根本扩展难题仍未解决 | 评估 Olix 架构如何处理信号衰减与干扰 |
| CUDA / NVIDIA 软件生态锁定 | 约束 | 2026–2030 | 推理栈迁移成本高;必须兼容 vLLM / Dynamo | 确认 Olix 面向标准服务框架的软件集成成熟度 |
方向和时间为基于公开来源的定性判断。并非所有约束都相互独立;良率问题与标准成熟度会相互影响。含义行反映分析师解读,并非公司指引。
[CM001, CM002, CM007, CM016, CM025, CM027]从 hyperscaler 识别需求,到光子推理大规模部署的各阶段。
漏斗阶段由公开芯片采购模式概括而来。Olix 尚未产生收入,也未披露客户管线;该漏斗代表预期路径,不是已确认交易流。
[CM035, CM025, CM039]2.5 展示项
03竞争对手
3.1 竞争格局与四层分类
Olix 进入的是一个 AI 推理硅片尚无单一主导形态的市场, 这既是机会,也说明光子计算板块仍然很早。尽调时可按技术物理 和商业位置,把竞争者清晰分成四层。 第一层是直接光子计算同行:这些公司在打造把光用于计算或 AI 工作负载内存带宽管理的硅片。Lightmatter 是最突出的名字, 但其 Passage 产品主要是光子互连 chiplet,Envise 计算平台 尚未以 hyperscaler 规模出货。收购线索包括 Celestial AI, 它已被 Marvell 吸收,其光学 fabric 知识产权现在落在半导体 巨头内部,而不是独立挑战者手中。 第二层是光互连和 co-packaged optics(CPO)供应商。Ayar Labs 领跑这一组,为 AI scale-up 集群出货 TeraPHY 光 I/O chiplet, 并让 NVIDIA 和 AMD 同时成为投资人。Intel Silicon Photonics 部门已向数据中心收发器出货超过 800 万颗光子集成电路,并在开发 4-Tbps Optical Compute Interconnect(OCI)chiplet。这些公司 供应 Olix OTPU 这类系统所依赖的光子管线;它们可能是伙伴,也 代表 hyperscaler 的另一种采购策略:把光子改装到现有 GPU 集群上, 而不是采用新的计算平台。 第三层是电学推理 ASIC:Groq(Language Processing Unit,已云端 部署)、Etched(transformer 专用硬编码 ASIC,尚未商业化),以及 NVIDIA 和 AMD 这些 GPU 巨头,其 Blackwell 和 MI 系列芯片主导当下 推理服务。第四层是 hyperscaler 自用硅片:AWS Inferentia、Google TPU 和 Meta MTIA。它们按每个 hyperscaler 削减可寻址市场,而不是 争夺商用硅片份额。 [CP001, CP002, CP003, CP005, CP017, CP018]
| 公司 | 类别 | 总部 | 关键产品 / 技术 | 最新融资 | 估值 / 规模 | 目标客群 | 商业阶段 |
|---|---|---|---|---|---|---|---|
| Olix(标的) | 直接光子同业 | 英国伦敦 | OTPU + DX-1(光子 + SRAM 推理) | $220M Series B,2026 年 2 月 | 隐含 $1B+(Series B) | Hyperscaler 推理、主权 AI | 收入前;目标 2027 年首批出货 |
| Lightmatter | 直接光子同业 | 加州山景城 | Passage(互连)、Envise(计算)、Idiom(ML 编译器) | $400M Series C,2024 | 未披露 | 数据中心 AI 基础设施 | 有限部署;Passage 进入客户试点 |
| Ayar Labs | 光互连供应商 | 加州圣何塞 | TeraPHY 光 I/O chiplet + SuperNova 激光器 | $500M Series E,2026 年 3 月 | $3.75B | AI scale-up 集群;NVLink Fusion 生态 | Chiplet 已可用;共封装集成 |
| Groq | 电推理 ASIC | 加州圣何塞 | LPU / Tensor Streaming Processor;GroqCloud API;GroqRack 本地部署 | $750M,2025 年 9 月 | $6.9B | LLM 推理云 + 本地部署 | 已商业部署;开发者 2M+ |
| Etched | 电推理 ASIC | 加州旧金山 | Sohu transformer 专用 ASIC(硬编码) | 未披露的种子 / 早期轮 | 未披露 | Transformer 模型推理服务 | 商业化前;未公布出货日期 |
| NVIDIA | 既有 GPU 平台 | 加州圣克拉拉 | Blackwell / B-series GPU;NVLink;CUDA 生态 | 上市公司(NVDA);FY26 Q3 数据中心收入 $51.2B | 市值约 $3T | 通用 AI 训练 + 推理 | 已大规模量产部署 |
| Intel Silicon Photonics | 光子组件 / 互连供应商 | 加州圣克拉拉 | OCI chiplet(4 Tbps);400G/800G 可插拔收发器 | Intel Corp 旗下部门(上市公司) | Intel 的一部分 | 数据中心网络 + CPO | 收发器已大规模部署;OCI chiplet 在开发中 |
| AWS Inferentia(Amazon) | Hyperscaler 自用芯片 | 华盛顿州西雅图 | Inferentia2 ASIC(定制神经网络推理) | 内部投入(Amazon capex) | Amazon AWS 的一部分 | AWS 云推理工作负载 | 已在 AWS EC2 Inf2 商业部署 |
Olix 行为标的公司,仅作参照。Lightmatter 的 $400M Series C 融资来自 2024 年新闻报道;估值未获公开确认。Celestial AI(光学 fabric,累计融资约 $250M)已被 Marvell 收购,因此不作为独立竞争对手列示。Groq 收入和开发者数量来自 Wikipedia / Groq 新闻室。NVIDIA 市值为运行日期附近的近似值。所有阶段判断均基于截至 2026-06-22 可公开获得的产品和部署信息。
[CP001, CP002, CP003, CP004, CP011, CP013]八个主体的竞争图,使用序数评分(1–5)衡量光子集成深度(x 轴)和推理工作负载专用化程度(y 轴)。评分来自公开产品材料支撑的评估,不是数值基准。
坐标轴为序数(1 = 最低,5 = 最高),反映作者截至 2026-06-22 对各公司公开产品重点的评估。光子集成深度衡量光子技术在产品架构中的中心程度。推理专用化衡量产品针对 AI 推理而非训练、网络或通用计算优化的程度。Olix 的分数基于官方材料:DX-1 被描述为面向推理和解码专用,并整合 SRAM + 光子;注释中标出其 2027 年出货前状态。数值 x/y 是序数证据评分,不是实测性能指标。
[CP001, CP002, CP005, CP011, CP015, CP019]3.2 光子计算与光互连同行
Lightmatter 是 Olix 最接近的架构同行,但比较下来,差异多于重叠。 Lightmatter 自称「Photonic Supercomputer Company」,已发布三条 产品线:Passage(光子互连 chiplet)、Envise(光子计算平台)和 Idiom(ML 框架与图编译器)。Passage 在 1 到 16 个或更多波长上 以每 lane 56 到 448 Gbps 运行,采用公司称为「Edgeless I/O」的 3D 集成。Lightmatter 称当前带宽能力为 114 Tbps,路线图将提升到 每秒超过 1 petabit,并与 TSMC、GlobalFoundries、Tower Semiconductor 建立代工关系。与 Olix 的关键差别在于,Lightmatter 的商业牵引和 公开叙事集中在光子互连;Envise 的计算加速层尚未证明 Olix 的 OTPU 和 DX-1 公开瞄准的 SRAM 中心、decode-heavy 推理专精。 Ayar Labs 的位置更互补,而非直接竞争。它出货基于 UCIe 互连标准的 TeraPHY 光 I/O chiplet,在等效功耗下提供约为铜互连十倍的带宽; SuperNova 则提供片上激光源。公司 2026 年 3 月完成 $500 million Series E,估值 $3.75 billion,并在 2026 年 6 月加入 NVIDIA NVLink Fusion 生态。NVIDIA、AMD、Alchip 和 MediaTek 都是投资人。Ayar Labs 是 chiplet 供应商,不是系统厂商;hyperscaler 把 Ayar Labs TeraPHY chiplet 与第三方 AI 处理器一起部署时,仍可能是 Olix 的潜在客户。 Intel Silicon Photonics 自 2016 年以来已出货超过 800 万颗光子集成 电路,并将超过 3200 万个片上集成激光器嵌入可插拔收发器,是硅光子 可靠性的量产制造标杆。其 OCI chiplet 提供每秒 4 terabits 双向带宽, 设计用于与 CPU、GPU 和 IPU 硅片 co-package。Intel 的成熟度验证了 光子制造可以规模化,但也说明成熟供应链产能集中在巨头手里,而不是 光子 AI 计算初创公司。 [CP006, CP007, CP008, CP009, CP010, CP011]
| 公司 | 光子计算核心 | 光 I/O / CPO | 电 ASIC | 推理专用 | 可用于训练 | 以 SRAM 为中心的内存 | 云服务 | 本地部署硬件 |
|---|---|---|---|---|---|---|---|---|
| Olix(标的) | ✓ OTPU(光子) | ✓(光链路) | ✗ | ✓ DX-1 聚焦解码 | 有限 / 未说明 | ✓(核心论点) | 尚未(2027) | ✓(2027 目标) |
| Lightmatter | ✓ Envise(有限) | ✓ Passage chiplet | ✗ | ✓(重点) | ✓(已表述) | 未说明 | ✗ | ✓ Envise blade |
| Ayar Labs | ✗(仅 I/O) | ✓ TeraPHY + SuperNova | ✗ | N/A(仅供应) | N/A | N/A | ✗ | ✓ chiplet |
| Groq | ✗ | ✗ | ✓ LPU / TSP | ✓ LLM 推理 | ✗(仅推理) | N/A | ✓ GroqCloud | ✓ GroqRack |
| Etched | ✗ | ✗ | ✓ Sohu(硬编码) | ✓ 仅 transformer | ✗ | N/A | 尚未 | ✓(计划中) |
| NVIDIA Blackwell | ✗(CPO 路线图) | 路线图 | ✓ GPU | ✓ | ✓ | 以 HBM 为中心 | ✓(DGX Cloud) | ✓ HGX/DGX |
| Intel Silicon Photonics | ✗(Gaudi 面向 AI) | ✓ OCI + 400/800G | ✓ Gaudi AI 加速器 | ✓(Gaudi) | ✓(Gaudi) | N/A | ✓(Gaudi cloud) | ✓ |
勾选和备注反映截至运行日期已发布的产品能力。以 SRAM 为中心一列记录架构是否明确宣称避开 HBM 依赖。Olix 行来自公司官方材料;「2027」表示计划能力,尚未出货。标为 N/A 的单元格表示该能力不适用于公司的产品模式。Ayar Labs 归类为 I/O 供应商;其推理能力取决于共封装的处理器。Lightmatter 的 Envise 计算能力截至运行日期缺少充分的公开部署数据。NVIDIA CPO 集成出现在生态路线图中,但还不是已出货产品功能。
[CP006, CP007, CP009, CP011, CP012, CP015]六家 AI 推理硬件玩家在八个维度上的能力覆盖,基于截至 2026-06-22 的公开产品文档评估。
勾选标记和文字反映截至 2026-06-22 公开可得的产品文档。N/A 单元格表示该维度不适用于公司的产品模型。Olix 和 Lightmatter 行基于公司官方材料;Ayar Labs 和 Groq 行使用官方新闻稿和产品页面。
[CP006, CP012, CP014, CP016, CP019, CP024]3.3 电学推理 ASIC 与 hyperscaler 自用硅片
Groq 是当下最成熟的电学推理替代品。它的 Language Processing Unit 是一款电学 ASIC,采用 tensor-streaming 架构,专门为低延迟 LLM 推理 和解码设计。Groq 于 2025 年 9 月以 $6.9 billion 估值融资 $750 million, 投资方包括 Disruptive Investments、BlackRock、Samsung、Cisco、D1 Capital 和 Altimeter Capital,并报告 2025 年收入约 $500 million。2025 年 12 月, Groq 与 NVIDIA 签署非独家推理技术授权协议,联合创始人 Jonathan Ross 转入 NVIDIA;Simon Edwards 出任 CEO。GroqCloud 服务超过 200 万开发者, 覆盖四个全球推理区域,并提供本地部署的 GroqRack 选项。截至写作时,Groq 针对 200 亿参数开源模型的 LLM 定价为每百万输入 token $0.075、每百万输出 token $0.30,速度约 1,000 tokens/sec;1200 亿参数模型分别为 $0.15 和 $0.60,速度约 500 tokens/sec。NVIDIA 授权关系并未消除 Groq 的商业独立性, 但压缩了其 IP 差异化叙事,并增加了与巨头绑定的风险。 Etched 用另一条路线瞄准同一推理负载:名为 Sohu 的硬编码 transformer ASIC, 无法重新编程,但声称在 transformer 推理工作负载上吞吐超过 NVIDIA H100 二十倍。高管团队组合了 Harvard 数学背景(CEO Gavin Uberti)、Cypress Semiconductor 校友(CTO Mark Ross)和二十多年 NVIDIA 设计经验(VP Brian Loiler)。Etched 尚未发布商业定价或出货时间线;如果模型架构演进到 transformer 范式之外,其 transformer 硬编码架构会面对所有固定功能 ASIC 都有的模型灵活性风险。 Hyperscaler 自用硅片——AWS Inferentia2、Google TPU、Meta MTIA——本质上是在 缩小可寻址市场,而不是与商用硅片竞争。AWS Inferentia2 在公开客户 benchmark 中实现了比 GPU 实例高 4.5 倍的吞吐,并把推理成本降低 90%。Google Cloud TPU 自 2018 年起已进入商业部署,并支撑 Google 自身 AI 推理工作负载的相当一部分。 每一个自建推理 ASIC 的 hyperscaler,都会永久减少包括 Olix 在内的任何商用 推理硅片厂商可触达的市场空间。 [CP019, CP020, CP021, CP022, CP023, CP024]
| 提供商 | 模型 / 配置 | 输出 tokens/sec(报告值) | 输入 $/M tokens | 输出 $/M tokens | 部署模式 | 数据时点 |
|---|---|---|---|---|---|---|
| Groq GroqCloud | 开源约 20B 参数 | ~1,000 | $0.075 | $0.300 | 云 API(按使用量付费) | 2026 年 6 月 |
| Groq GroqCloud | 开源约 120B 参数 | ~500 | $0.150 | $0.600 | 云 API(按使用量付费) | 2026 年 6 月 |
| Groq GroqCloud | Llama 4 Scout 17B | ~594 | $0.110 | $0.340 | 云 API(按使用量付费) | 2026 年 6 月 |
| AWS Inferentia2 | 客户推理工作负载 | 未发布 | 未发布(capex 模式) | 未发布 | AWS EC2 Inf2 实例 | 2026 |
| Olix DX-1(目标) | 大型 LLM 解码工作负载 | 目标高于当前 GPU(未确认) | 未披露(商业化前) | 未披露 | 本地 / 机架级(2027 目标) | 2026(仅目标) |
Groq 定价反映运行日期的 GroqCloud 公开定价页和 Artificial Analysis 基准;价格可能变化。AWS Inferentia2 按实例计价(EC2 Inf2 小时费),不是按 token 计价;若要与 Groq API 定价直接比较,需要工作负载建模。Olix DX-1 定价未公开披露;列入该行是为了标记缺口。NVIDIA H100/B200 定价属于本地资本开支或云实例成本,若没有工作负载假设,不能直接按 token 口径比较;为避免误导,已省略。Google TPU 定价(TPU v5)可通过 Google Cloud 获取,但现有公开来源未将其与开源 LLM 工作负载对标。
[CP006, CP019, CP020, CP023, CP024, CP027]3.4 差异化向量、护城河耐久性与竞争缺口
Olix 的核心差异化主张是架构:SRAM 中心内存叠加光子技术,避开 high-bandwidth memory(HBM)供应瓶颈和相关成本,并瞄准主导自回归 decode-heavy LLM 推理的内存带宽约束。这个定位确实不同于 Groq 的 低延迟优化 streaming 架构、NVIDIA 的 HBM 中心 Blackwell 设计,也不同于 Lightmatter Passage 的光子互连重点。截至 2026 年中,DX-1 decode accelerator 占据了不同于任何已部署竞争者的优化点。 这种差异化能否持久,取决于执行速度和软件生态。Hyperscaler 推理硅片买方的 切换标准集中在 cost-per-token、throughput-per-watt、软件栈兼容性,以及通常 需要 12 到 24 个月的供应链认证周期。Groq 已经为云托管推理跨过这些门槛;Olix 还没有进入认证管线。这个时间差是最具体的竞争风险:在 Olix 2027 年首批出货前, Groq、Etched 和改进后的 NVIDIA Blackwell B-series 系统还会多出 18 个月生产学习 和生态开发时间。如果模型架构转向 mixture-of-experts 或 state-space 架构,并把 硬件瓶颈推向不同位置,Etched 的 transformer 硬编码设计和 Olix 的 SRAM-decode 优化都面临重新定范围的风险。 Olix 护城河最强的地方,是它把光子与内存 IP 组合成难以快速复制的系统。Ayar Labs 和 Lightmatter 在 2025、2026 年的融资显示,光子 AI 硬件吸引大量资本;但公开记录 中还没有竞争者采用同样的 SRAM 加光子集成路线。Intel 量产验证过的光子制造能力 证明了更大技术生态可行,但 Intel 部门聚焦互连组件,而不是推理优化计算,因而把 计算差异化层留了出来。负面的竞争现实也很直接:IEEE Spectrum 2025 年中推理 benchmark 显示,NVIDIA 和 AMD 拿下了所有已出货 AI 推理性能纪录;光子挑战者必须 用尚未出货的硅片,在 hyperscaler 通常持续 12 到 24 个月的认证周期里替代这个基线。 [CP030, CP031, CP034, CP035, CP036, CP037]
| 公司 / 护城河主张 | 护城河类型 | 耐久性(1 = 脆弱,5 = 耐久) | 护城河的主要威胁 | 与 Olix 的重叠 | 尽调问题 / 缓释 |
|---|---|---|---|---|---|
| Olix — SRAM-光子集成 IP | 架构 IP + 深科技执行 | 4(如果芯片按时出货) | 架构到 2027 年仍未出货;既有厂商用 CPO 外挂补上差距 | 标的公司 | 验证晶圆厂承诺和 tape-out 时间线;确认 SRAM+光子集成里程碑 |
| Lightmatter — 光子 IP + 多晶圆厂生态 | IP 组合 + TSMC/GF/Tower 关系 | 3 | NVIDIA 收购或对外授权竞争性光互连 IP | 光子计算重叠(Envise);互连层(Passage) | 将 Lightmatter 专利组合与 Olix OTPU 主张逐项映射;评估 CPO 共同开发风险 |
| Groq — LPU IP + GroqCloud 软件生态 | 软件开发者护城河 + LPU 架构 | 3(受 NVIDIA 授权削弱) | NVIDIA 授权降低排他性;Groq 客户有 CUDA 生态迁移路径 | 推理速度和延迟;Groq 瞄准同一批云推理客户 | 跟踪 Groq-NVIDIA 授权范围;跟踪 GroqCloud 定价与光子推理成本曲线 |
| NVIDIA — CUDA 生态 + 制造优势 | 生态锁定 + 供应链控制 | 5 | 效率更高且价格相当的 ASIC 在 2–3 个采购周期内替代 GPU | 整个推理栈;NVIDIA 是 Olix 必须证明 ROI 能胜出的现状基准 | 跟踪 Blackwell B-series 效率基准与 Olix 目标规格;映射 CUDA 迁移成本 |
| Ayar Labs — CPO 组件 IP + NVIDIA 合作关系 | 技术组件护城河 + hyperscaler 合作关系 | 4 | 竞争性 CPO 路线(Lightmatter Passage、Intel OCI)让光 I/O chiplet 商品化 | 光 I/O 供应链;Ayar Labs 可能为 Olix 或其竞争对手供应 CPO | 评估与 Ayar Labs 合作的可选性;评估 TeraPHY + Olix OTPU 是否可落地 |
耐久性评分是基于截至运行日期公开可得信息的定性判断。没有财务条款、NDA 披露或私有尽调数据。Intel Silicon Photonics 护城河(规模化制造)未作为直接行列入,因为 Intel 主要在互连和收发器层竞争,而不是 AI 推理计算;该项已在功能矩阵中覆盖。Etched 因缺少已披露融资、估值或生产时间线,未单独评级护城河。
[CP002, CP010, CP011, CP013, CP014, CP021]对比截至 2026-06-22 运行日期,Olix 与关键竞争对手在融资、估值和部署就绪度上的 KPI。
[CP004, CP013, CP015, CP021, CP024, CP031]3.5 展示项
04财务
4.1 收入模型、定价结构与商业化时间线
Olix 披露的产品是 OLIX Decode Accelerator 1(DX-1),一个机架级 AI 推理系统,整合 SRAM 内存、光互连和定制逻辑。产品架构、招聘宣言措辞和第三方报道共同暗示,其商业模式是面向 hyperscaler、云服务商和 AI-first 企业销售硬件,这些客户需要低延迟、高吞吐的大规模推理能力。这把 Olix 放在与 Groq(销售硬件和云 API 访问)及 Ayar Labs(向系统集成商销售 CPO 模块)同一类收入模型中;但截至本报告日期,Olix 尚未公开确认任何定价、客户名称、收入或合同 backlog。 Olix 的计算宣言明确把 DX-1 与 HBM-based GPU 架构在总拥有成本(TCO)和每兆瓦吞吐上对标,而不是对标名义芯片价格。这种表述符合企业硬件销售逻辑:买方评估的是总基础设施成本,不是芯片标价。Olix 还称其 SRAM 架构避开对 HBM、先进封装和其他供给受限组件的依赖——如果兑现,这个供应链论点可能让 Olix 在低于现有 GPU 内存成本的价格上销售,同时守住健康毛利。SiliconANGLE 报道补充称,Olix 计划用资本为芯片开发融资;Financial Times 通过 SiliconANGLE 报道,首批 OTPU 客户交付预计在 2027 年。 商业化时间线制造了显著收入空窗。Olix 这个阶段的硬件初创公司,通常要等首颗硅片 tape-out 验证、首批生产单元出货、初始客户认证完成后,才会有经常性收入。这个序列几乎总会比目标交付日期再延长 12–24 个月。对 Olix 而言,公开的 2027 年交付预期意味着最乐观也要到 2027 年末才可能确认收入,更可能的商业爬坡不会早于 2028 年。公司没有披露任何 design-win、意向书或试点客户接触,也进一步限制了外部分析师可见的收入确定性。Olix 是否会进入 inference-as-a-service(直接提供算力容量)仍未确认;如果选择这条路,盈利前周期会被拉长,资本强度也会因需要类似 GPU 集群的基础设施投入而大幅上升。 [CI001, CI002, CI003, CI004, CI005, CI006]
| 收入流 | 机制 | 单位 / 定价基础 | 当前状态 | 收入质量 | 尽调问题 |
|---|---|---|---|---|---|
| 硬件销售(DX-1 机架 / 芯片) | 向 hyperscaler 和云提供商销售 AI 推理加速器系统 | 按机架或芯片;标价未披露 | 收入前;目标 2027 年首次交付 | 未知 — 未披露合同或 backlog | 确认设计定点、LOI 或试点合作;获取名单和实现价格估算 |
| 推理即服务(云 API) | 在自有基础设施上销售算力,按 token 或请求计费 | 按推理 token 或 API 调用;未披露 | 推测性 — 公司未确认 | 未知 — 未披露服务 | 确认产品路线图是否包含托管服务或云 API 层 |
| IP 授权或晶圆厂合作 | 向晶圆厂或系统集成商授权光子-SRAM 集成 IP | 按晶圆或单元收取 royalty;未披露 | 没有授权交易或晶圆厂合作证据 | 未知 — 公开未讨论 | 询问是否考虑 IP 授权或供应链合作收入 |
| 政府补助和 R&D 抵免 | UK UKRI / Innovate UK 或 HMRC R&D 税收抵免,适用于合格 R&D 支出 | 非稀释;取决于合格 R&D 支出 | 可能存在但未披露 — 英国深科技公司的常见做法 | 低 — 非稀释,但不是主要收入流 | 要求说明任何政府补助奖励或 R&D 抵免申报 |
除硬件销售外,所有收入流都是推测或推断;截至运行日期,没有 Olix 财务披露确认任何收入。
[CI001, CI002, CI003, CI004]| 信号 | 来源 | 隐含定价或财务指标 | 置信度 | 局限 |
|---|---|---|---|---|
| DX-1 硬件标价 | 公司(未披露) | Null — 未公开 | None | 没有定价页面,也没有合作伙伴披露 |
| TCO 相对 GPU 的定位 | Olix 计算宣言(公司主张) | 公司称,相比基于 HBM 的架构,其吞吐量 / MW 更高,TCO 更低 | 低 | 公司自述;没有独立基准测试支撑 |
| Hyperscaler AI 资本开支背景 | Alphabet 2026 年 Q1($190B 资本开支计划),Meta 2026 年 Q1 业绩 | Hyperscaler 正把数千亿美元投向 AI 基础设施,可触达支出规模很大 | 高 | 不能证实 Olix 定价,也不能证实其能拿到这部分支出 |
公开信息里没有标价、实际成交价或客户合同金额;定价行只是缺少证据的占位项,对应具体尽调问题。
[CI005, CI006, CI007]展示未来一笔 Olix 交易如何从客户需求走到产品交付,再转化为收入和毛利,并标出仍未补齐的缺口。
从设计定点之后的所有节点都属前瞻判断,没有合同证据支撑。收入和利润率节点仅反映行业基准。
[CI001, CI002, CI003, CI006]4.2 成本结构、人员 burn 与资本强度
Olix 的成本结构主要由三类支出主导:人员、硅片工程(tape-out、EDA 工具、原型)、基础设施(实验室空间、测试设备、计算、专业服务)。三者都显著高于英国科技初创公司的中位水平,因为 Olix 在打造带集成光子的全定制半导体——这是行业里最昂贵的工程任务之一。 人员端,TechSparx 报道称 Olix 在 Bristol 招聘高级数字设计工程师,基础薪资 £125,000–£180,000,另加每年 £24,000 本地生活补贴,意味着单个工程岗位在国民保险和股权前,总现金包最高可达 £204,000。Olix 招聘宣言确认,公司提供基准领先薪酬、有意义的股权授予,以及入职三个月文化不匹配时的 18 周薪资退出包。单是这个退出条款,每次招聘失败都会带来一笔实质性一次性现金负债。Jon Peddie Research 2026 年 2 月报道,Olix 已增长到 70 多名员工,并计划到 2026 年底超过 200 人。若以每人 £160,000– £200,000 的混合 fully-loaded 成本(base、NI、福利、本地奖金和股权摊销)套用 70 名现有员工,估算年度人员 run-rate 为 £11–14 million(约 $14–17 million)。扩到 200 人后,在不计非人力成本前,这一数字约增长三倍至每年 £32–40 million($40–50 million)。 硅片工程是第二大成本驱动。在 TSMC 或 IMEC 等领先代工厂进行一次光子 ASIC tape-out,生产级芯片的 NRE、mask 和测试晶圆成本通常在 $10–50 million 区间。Siemens 半导体封装博客解释,硅光子工艺仍呈现高于成熟 CMOS 的制造波动;co-packaged 组件可返工性有限,任何封装环节的良率偏离都会放大成本冲击。每一次失败的 tape-out 迭代都会推迟收入并消耗资本。edgen.tech 关于 CPO 量产延期的报告进一步指出,封装复杂度可能让每轮迭代把交付时间线推迟 6 到 12 个月,从而成倍放大首笔收入前消耗的总资本。 一个粗略的 2026 年中月度 burn 组合模型(70 名员工、tape-out 前原型阶段):人员 $1.2–1.5M/月;硅片工程和工具 $0.5–1.0M/月;基础设施和 overheads $0.3–0.5M/月;合计估算 $2.0–3.0M/月。团队扩到 200 人并进入首轮 tape-out 生产后,估算 burn 升至 $5–8M/月。这些估计置信度低,完全基于行业成本 benchmark 和员工数信号,并不来自任何已披露的 Olix 财务报表。 [CI009, CI010, CI011, CI012, CI013, CI014]
| 指标 | 数值或 null | 置信度 | 重要性 | 尽调问题 |
|---|---|---|---|---|
| 收入(ARR/GMV) | null | None | 决定业务规模,也决定 CAC 和回本周期能否计算 | 要求公司提供收入 run-rate;如已确认收入,则提供 ARR |
| 毛利率(%) | null | None | 硬件 AI 芯片放量后通常瞄准 50–70% 毛利率;光子方案还叠加良率风险 | 要求提供产品发布时内部模型中的毛利率假设 |
| 估算的工程师全包混合成本(年度) | £160K–£200K(~$200K–$250K) | 低(按薪酬数据 + 福利基准估算) | 最大单项成本驱动;会随员工数线性放大 | 核实福利负担、股权估值基础,以及退出准备金计提 |
| 估算月度烧钱额——当前(70 名员工) | $2M–$3M/月(估算) | 低(模型测算;非来自账目) | 决定 $220M 融资能支撑多久 | 要求管理层披露董事会口径烧钱率和 runway |
| 估算月度烧钱额——200 人规模 | $5M–$8M/月(估算) | 低(模型测算;非来自账目) | 决定下一轮融资触发点,以及资本是否足以交付产品 | 要求提供 12 个月运营预算和招聘计划 |
| CAC / 回本周期 | null — 尚未产生收入 | None | 硬件 AI 芯片销售周期通常为 12–24 个月;CAC 需要重资本投入 | 第一笔商业收入出现前无法计算;要求提供估算销售周期长度 |
| 每次 tape-out 迭代的 NRE 成本 | $10M–$50M(领先节点光子 ASIC 的行业估算) | 低(行业基准;Olix 具体数据未披露) | 每次 tape-out 失败都会吃掉数月 runway | 要求提供 tape-out 预算、计划 spin 次数和应急资金分配 |
全表置信度都是无或低;这张表记录的是缺口,不是已确认指标。行业基准来自 Siemens CPO 博客和一般半导体成本文献。
[CI009, CI010, CI011, CI012, CI013, CI015]梳理从人员到流片的主要成本驱动项,再映射到估算的单颗成本结构,并标出所有 null 节点。
所有成本估算都由公开人数信号和行业基准建模得出。首个产品交付并披露财务数据之前,单位成本、单位收入和毛利率均为 null。
[CI009, CI010, CI011, CI013, CI015]4.3 资本充足性、融资时间线参考与 runway 情景
Olix 最近一次披露融资,是 Hummingbird Ventures 领投的 $220 million 轮次,于 2026 年 2 月 11 日宣布,并由 Cooley 及多家独立媒体报道。媒体报道显示,公司累计融资约 $250 million,意味着此前从 Plural、Vertex Ventures、LocalGlobe、Entrepreneurs First 等投资方合计募集约 $30 million。详细融资时间线(各轮金额与日期)放在「公司概览」章节;本节只为这里的资金充足性分析,新增所需资本事实的「财务」主张。 Hummingbird 自己对策略的描述值得纳入判断:该机构称自己是全球种子期投资者,初始支票从 $500K 起,并会在后续轮次中继续投入 $50M-plus,为深科技发展的高风险阶段提供有耐心的长期资本。这让 Hummingbird 成为一场多年、收入前建设周期里的可信共同投资方,但如果技术里程碑滑坡,或竞争环境发生实质变化,并不能保证它会继续支持。 资金充足性表格用上一节推导出的烧钱率估计,建模了三种 runway 情景。基础情景下(当前 70 人团队,月烧钱约 $3M),$220M 可支撑约 73 个月,现金可覆盖到约 2032 年,远超 2027 年产品目标。但 Olix 已明确表示计划在 2026 年扩张到 200-plus 名员工;一旦执行,月烧钱会升至 $5–8M/月,$220M 的有效 runway 会缩短至约 27–44 个月(约 2028 年中至 2029 年中)。再叠加一次重大 tape-out 周期($20–50M),runway 会进一步压缩到 15–30 个月(覆盖至 2028 年初至 2028 年中)。这些情景表明,按当前计划,Olix 很可能需要在 2028 年底前再融一轮,具体取决于 tape-out 时点和运营规模。 可比 AI 推理芯片初创公司提供了有用参照。Groq 于 2025 年 9 月以 $6.9 billion 估值融资 $750 million,说明已验证的 AI 推理平台仍可获得大额后续融资。Ayar Labs 在 2025 年完成 $500 million Series E,用于 production-ready CPO 硬件,战略投资方包括 NVIDIA、AMD、MediaTek——说明只要生产里程碑被清楚跑通,半导体光子公司可以吸引大额后续资本。Fractile 是另一家英国推理芯片初创公司,2026 年 5 月宣布完成 $220 million Series B,确认处在类似阶段的英国深科技芯片公司也能拿到可比资本。Olix 能否在更长、更烧钱的产品开发弧线中维持类似投资者信心,高度取决于技术里程碑执行。 Companies House 备案显示股本约 GBP 288,反映出英国风投支持公司常见的 fractional penny-share 结构:为容纳复杂 cap table,股权以大量小面额股份发行。这对经济估值没有分析价值,但确认公司采用标准风投结构。2026 年 3 月 21 日提交的确认声明(CS01),以及截至 2025 年 8 月 31 日财年的 total exemption full accounts,是目前可公开验证的财务信息边界;账目提供资产负债表,但隐藏了收入、运营成本和亏损。 [CI019, CI020, CI021, CI022, CI023, CI024]
| 情景 | 月度烧钱估算 | 现金基数(假设 $M) | 隐含 runway(月) | 近似 runway 结束时间 | 关键假设 |
|---|---|---|---|---|---|
| 基准——当前规模(70 名员工,tape-out 前) | $2M–$3M/月 | $220M | 73–110 个月 | ~2032–2035 | 员工数持平;没有重大 tape-out 支出 |
| 扩张——200 名员工(目标),无 tape-out | $5M–$6M/月 | $220M | 37–44 个月 | ~2029 年中 | 团队按披露扩张到 200 人;硅片支出推后 |
| 扩张加 tape-out——200 名员工 + 一次重大 tape-out($30M) | $5M–$8M/月 + $30M 一次性支出 | $190M(tape-out 后) | 24–38 个月 | ~2028 年中至 2029 年中 | 消耗一次完整 tape-out;这是 2027 年产品目标下的可能情景 |
| 最差情形——200 名员工 + 两次 tape-out 迭代(合计 $60M) | $6M–$8M/月 + $60M 一次性支出 | $160M(多次 tape-out 后) | 20–27 个月 | ~2028 年初至年中 | 多轮迭代;对新型光子工艺而言是可能情形 |
| 对照——Groq $750M 融资(2025 年 9 月) | 未披露 | $750M(Groq) | 参考:大型 AI 推理融资;Groq 产品成熟度更靠前 |
所有烧钱估算均按员工数和行业成本数据建模,不基于任何 Olix 财务披露。现金基数假设 $220M 融资全额到账(Companies House 备案显示融资机制延续到 2026 年 3–5 月)。一次性 tape-out 成本采用行业基准。
[CI019, CI020, CI021, CI022, CI023, CI024]基于来源给出三种情景下的月度烧钱区间和隐含跑道月数,并明确标注置信度。
所有估算均来自行业人员成本基准,以及 TechSparx 和 Olix 招聘页面公开披露的薪资信号。未审阅任何 Olix 内部财务数据。
[CI009, CI010, CI011, CI012, CI013, CI016]展示 $220M 融资在人员扩张、流片里程碑和基础设施上的预估顺序消耗,并在基准与压力情景下测算剩余跑道。
所有资金消耗项目都是基于行业成本数据和人员信号的估算。对于新型光子工艺,流片成本可能显著更高($50M+)。第二轮流片会把剩余现金全部吃掉。
[CI019, CI020, CI021, CI022, CI023, CI026]4.4 财务结论与关键尽调阻断项
Olix 呈现出一家资本充足的收入前深科技硬件初创公司的画像:账面 runway 充裕,团队可信度高且薪酬优厚,产品论点清晰,并由 $250 million 风险资本和独角兽估值背书。核心财务风险不是短期偿付能力,而是光子半导体开发的成本与时间敞口;一次 tape-out 延误或良率失败,就可能吃掉六到十二个月 runway,并迫使公司进行计划外融资。 三个反向财务信号需要升级尽调。第一,Siemens CPO 封装博客和 edgen.tech 均记录到,共封装光学量产持续面对良率波动,单位成本被推高到商品化 CMOS 基线之上——Olix 的 SRAM-and-photonics 架构正落在这个高波动区域。第二,Olix 的薪酬结构(£125– £180K 基薪加 £24K 本地奖金,18 周离职缓冲)即便按伦敦深科技标准也偏慷慨;每一次招错人或文化不匹配离职,都会在收入出现前产生真实现金成本。第三,Olix 截至 2025 年 8 月 31 日期间的法定账目按 total exemption 制度提交,意味着虽然有资产负债表,但损益表和董事报告(其中会出现收入、亏损和持续经营说明)根据英国小公司豁免未向公众披露。 英国国家半导体战略描述了政府未来 20 年支持芯片 IP、设计和化合物半导体的意图,但当前政策文件没有承诺为具体芯片设计项目提供直接资本——它不是 Olix 的额外资金来源。 毛利率、单位经济性和收入轨迹完全未披露。没有任何公开确认的 design win、意向书或试点客户,分析师无法精确建模收入爬坡。财务结论是:当前阶段资本位置足够;执行和良率风险是主导财务威胁;本章的尽调问题,在任何投资承销前都不可让步。 [CI030, CI031, CI032, CI033, CI034, CI035]
| 缺失指标 | 缺失原因 | 对尽调的影响 | 尽调路径 |
|---|---|---|---|
| 收入和 ARR | 私营公司;total exemption accounts 不披露损益表 | 阻断——无法评估商业牵引力或资本效率 | 要求公司提供经审计管理账目或董事会口径损益表 |
| 毛利率和经营利润率 | 未披露;硬件毛利取决于代工良率和批量定价 | 重大——没有毛利可见度,投资回报模型只能靠猜 | 要求提供内部财务模型,按产品代际列出毛利假设 |
| 实际月度烧钱率 | total exemption accounts 隐去运营支出 | 重大——无法验证 runway;所有估算都是模型近似值 | 要求 CFO 或类 CFO 角色提供过去六个月烧钱桥表 |
| 股权结构和投资人持股 | 英国私营公司;股份配发备案只显示资本事件,不显示所有权结构 | 重大——影响稀释分析和优先权 stack 建模 | 要求公司提供 pro-forma cap table 和优先清算 waterfall |
| 代工和 tape-out 承诺 | 未公开披露;新型光子工艺可能使用非标准代工伙伴 | 阻断——tape-out 成本和进度几乎决定全部 runway 与交付风险 | 要求在 NDA 下提供与代工伙伴的 term sheet 或意向书 |
这张表记录的是信息缺失,不是研究可以补齐的证据缺口。所有项目都需要管理层直接披露。
[CI030, CI031, CI032, CI033, CI034]4.5 附录
05产品与技术
5.1 DX-1 架构——SRAM 与共封装光子集成
Olix 的计算宣言把数据移动——尤其是 KV-cache 流量穿过 High Bandwidth Memory 时产生的能耗和延迟成本——定义为根本约束,它让现有加速器无法同时交付高吞吐和低延迟。公司的回应是 OLIX Decode Accelerator 1(DX-1),并称其为首个专为 decode 架构设计的加速器。设计理念是完全放弃 HBM,转而扩展以 SRAM 为中心、并与光子集成的存储层级,同时处理带宽和延迟两维问题。 SRAM+photonics 架构在机架尺度运行。Olix 材料把「logic、data movement、packaging、optics and interconnect 的 rack-scale co-design」描述成一个统一系统,并将其与同行采用的 reticle-scale 和 wafer-scale 路线区分开。光子部分覆盖芯片到芯片、tray 内、机架到机架的光互连,并以 co-packaged optics(CPO)作为集成载体。Optical Communications Engineer 职位描述确认,Olix 正在构建跨这三种距离层级的链路架构,系统级链路预算——损耗、BER、带宽、功耗和热裕量——由工程团队内部负责。 第三方市场观察 profile 同时点名 Olix 的 OTPU(Optical Tensor Processing Unit)和 DX-1,称二者是不同产品组件。宣言没有明确使用 OTPU 标签,而是指向 DX-1 系统层级;OTPU 似乎描述的是更大机架系统中的计算单元。公司声称 OTPU 使用光子元件完成矩阵乘法且发热极低,这应视为设计意图,而非已验证性能;目前没有公开 characterization data。 预硅原型正在使用 Xilinx/AMD UltraScale+ 和 Versal FPGA 平台推进。FPGA 岗位要求 PCIe Gen4 接口、以超过 25 Gb/s NRZ 运行的多通道收发器、BER 测量基础设施和 SECDED 纠错,确认 FPGA 控制逻辑与光学子系统之间存在高速数字集成层。White Rabbit 精密时序也被提及,暗示系统需要确定性 fabric,这与分布式机架级计算一致。 [CE001, CE002, CE003, CE004, CE005, CE006]
| 模块 / 资产 | 客户 / 用户 | 状态 / 成熟度 | 差异化主张 | 尽调缺口 |
|---|---|---|---|---|
| DX-1 Decode Accelerator | Hyperscaler、云服务提供商、前沿 AI 实验室 | 硅片前;FPGA 原型(Xilinx/AMD UltraScale+ / Versal) | 首个专为 decode 设计的加速器;机架级协同设计 | 未公布硅片 tape-out 日期,没有基准测试数据,也未披露客户试点 |
| OTPU 光学张量处理单元(Optical Tensor Processing Unit) | DX-1 机架内组件 | 概念 / 硅片前;仅有设计意图 | 光子计算单元,目标是以低发热执行矩阵运算 | 没有 characterization,没有公开 die photo 或数据手册;仅在第三方报道中被点名 |
| 光互连子系统 | DX-1 内部架构;未来可能授权 | 工程 / 设计阶段;链路预算仍在开发 | 基于 CPO 的芯片间、托盘内和机架间链路,覆盖三层距离 | 未披露 OSAT 或代工伙伴;没有公开良率或 BER 结果 |
| FPGA 原型平台 | 内部 R&D / 硬件 bring-up | 活跃(招聘信息确认仍在发布) | 用于高速接口验证(PCIe Gen4、>25 Gb/s NRZ、White Rabbit timing) | 仅为原型平台;硅片前无法给出推理服务基准 |
所有状态判断都来自截至 2026 年 6 月的招聘信息和公司宣言;尚未确认产品出货或硅片。
[CE001, CE002, CE005, CE009, CE010, CE036]| 层 / 组件 | 角色 | 关键依赖 | 风险 |
|---|---|---|---|
| 推理服务层 | vLLM / SGLang / NVIDIA Dynamo runtime;模型调度和 batching | 开源社区采用;Olix 硬件插件尚未合并 | 框架按 GPU-first 节奏演进;Olix 必须维持 parity |
| KV-Cache 管理层 | NIXL / Mooncake TE 协议,用于 prefill 与 decode 节点之间的分离式 KV-cache | 协议成熟度;NIXL 仍处早期 | 协议变动可能迫使 Olix 重做软件;第三方支持有限 |
| FPGA 控制 / Bringup 层 | PCIe Gen4 主机接口、BER 监控、ECC、White Rabbit timing distribution | Xilinx/AMD UltraScale+ 或 Versal 可得性;代工产能分配 | 仍只是 FPGA 原型阶段;量产需要 ASIC 替换 |
| SRAM 计算与存储层 | 片上 SRAM 存储模型权重;消除 HBM 往返延迟 | 定制 SRAM 工艺整合;die 面积扩展 | 放大规模后,SRAM 每瓦容量低于 HBM;芯片面积限制每片晶圆 die 数 |
| 光子互连层 | 基于 CPO 的光链路,覆盖芯片间、托盘内和机架间距离 | 硅光子代工;CPO 封装 OSAT;激光源供应 | 良率波动、热敏感、对准公差,以及 OIF 标准尚未完整 |
架构由公司宣言和招聘信息推断;层级顺序和接口规格尚未正式发布。
[CE002, CE003, CE007, CE008, CE010, CE011]基于公司宣言和职位描述重建的 DX-1 系统概念分层模型。各层来自公司描述或推断;官方尚未发布架构图。
架构由公开职位描述和宣言重建;内部层级边界和接口未经过独立验证。
[CE001, CE003, CE004, CE006, CE007, CE008]基于公司职位描述,说明超大规模云厂商或 AI 实验室如何在解耦服务集群中把 DX-1 用作专门负责解码的节点。
工作流由平台集成职位描述构建;实际编排架构尚未公开,也未在硬件中验证。
[CE011, CE012, CE013, CE014, CE015]5.2 软件栈兼容性与开发者集成
Olix 的平台集成招聘确认,DX-1 的软件层被设计为接入主流开源推理栈,而不是要求专有 runtime。Senior Platform Integration Engineer 职位说明点名 vLLM、SGLang 和 NVIDIA Dynamo 作为目标 serving 框架,并预期工程师会把这些框架扩展到 DX-1 硬件上。KV-cache 管理由 NIXL 和 Mooncake TE 承担;这两种新兴协议服务于分离式 prefill 与 decode,使 prefill 计算和 decode 计算可以跨不同硬件分区拆开。集群通信目标是 NCCL、RoCE 和 InfiniBand,这意味着 DX-1 设计上会与现有基于 GPU 或 CPU 的 prefill 节点在异构集群中协同运行,而不是作为完全自包含系统。 这套软件策略在技术上成立:vLLM 是最活跃的开源推理项目之一,来自学术机构和公司的贡献者超过 2,000 名;SGLang 在 2026 年初于 NVIDIA GB300 NVL72 上实现 3.8× prefill 和 4.8× decode 吞吐提升,展示了快速迭代能力。两个框架都开放硬件插件架构;Intel Gaudi、IBM Spyre 和 Huawei Ascend 都是已集成进 vLLM 的非 NVIDIA 硬件例子。这给了 Olix 一条可信集成路径——但它需要自定义 kernel 开发、HAL 层工作,并在这些框架按 GPU-first 节奏演进时持续维护 upstream。 关键缺口在于,截至 2026 年 6 月,Olix 没有公开记录显示其向 vLLM、SGLang 或相关框架贡献过开源代码;且 DX-1 仍处于 pre-silicon 阶段,软件集成尚未经历 hardware-in-the-loop 验证。因此,Olix 的开发者社区信号是间接的:目标框架活跃且可接入,但 Olix 还没有以硬件供应商身份,凭已验证驱动或公开 benchmark 在这些社区建立存在感。 [CE011, CE012, CE013, CE014, CE015, CE016]
| 用户任务 | 当前工作流(基于 GPU) | DX-1 声称的方案 | 声称的可衡量收益 | 局限 / 尽调问题 |
|---|---|---|---|---|
| 为大量并发用户提供交互式低延迟 LLM decode | 需要大规模 GPU batch 来摊薄 HBM 权重成本;单用户延迟高 | 基于 SRAM 的 DX-1 消除 KV-cache HBM 读取延迟;单 token 能耗更低 | 同时实现高交互性(低 TTFT、高 TPS)和高吞吐 | 没有公开基准测试;硅片前;收益仍是公司主张 |
| 分离式 prefill + decode 服务 | 同质 GPU 集群;prefill 和 decode 争用同一加速器 | DX-1 作为 decode-only 节点;可通过 NIXL / NCCL 与 GPU prefill 节点配对 | 利用率和调度更好;单个 decode token 成本更低 | 需要异构编排;软件集成尚未投产 |
| 无需重新训练即可服务现有开源 LLM | 原生 CUDA / ROCm kernel;需要 NVIDIA 或 AMD 硬件 | vLLM / SGLang 插件,声称完整模型兼容;无需 QAT/PTQ | 模型无需改动即可运行;企业避免重新训练成本 | 插件仍在开发;截至 2026 年 6 月,未确认 Olix 向 vLLM / SGLang 提交 PR |
DX-1 的所有收益都来自公司主张,或由公开招聘材料推断;尚无独立验证。
[CE001, CE011, CE012, CE013, CE039, CE040]Olix 要把 DX-1 从原型推向量产,必须满足几类关键外部依赖:硬件供应、软件生态和监管要求。
依赖图由宣言、职位描述和行业报告推断;Olix 尚未作出官方供应链披露。
[CE002, CE008, CE022, CE023, CE025, CE043]5.3 制造挑战、标准、出口管制与技术风险
不论硅片质量如何,光子与 CPO 供应链都给 Olix 带来多项有充分记录的风险。Siemens EDA 2026 年 CPO 分析中的行业报道指出四个交织挑战:热管理(photonic IC 对温度高度敏感,AI 工作负载的瞬时热波动会让光波长漂移)、测试扩展性(光学对准需要微米级精度,远严于电测试,自动化量产级测试方法尚不成熟)、制造良率(硅光子工艺波动高于成熟 CMOS,CPO 组装不可返工,使良率波动代价尤其高)、标准不完整(OIF 及相关联盟仍在制定调制格式、热包络和光学 attach 方法的互操作规范)。edgen.tech 的一份反向行业 note 具体报道,CPO 量产因良率问题被推迟到 2026 年之后——这会直接映射为 Olix 计划中 2027 年产品发布窗口的风险。 对一家收入前初创公司而言,Olix 的出口管制姿态异常明确。所审查的每个工程岗位都附有通知,限制最近国籍或永久居留地为 Iran、North Korea、Syria、Cuba、Russia、Belarus、China、Hong Kong、Macau 或 Venezuela 的候选人,并引用美国出口管制规定。这表明 Olix 的技术栈几乎肯定包含 Export Administration Regulations(EAR)下的受控项目,对客户地域、供应链伙伴选择,以及来自受限国家关联投资人的融资都有现实影响。 光子领域的竞争背景进一步说明成熟度缺口。Intel 等成熟玩家已为数据中心网络交付硅光子产品多年;获得 NVIDIA 支持的 Ayar Labs 已宣布共封装光学 chiplet 生产。Lightmatter 的 Passage 平台从概念到初始部署用了四年或更久。Olix 做的是更有野心的全系统平台,而不只是光互连层;相对组件专家,工程和制造挑战被叠加放大。公开领域没有 DX-1 的独立 benchmark 数据、可靠性测试结果或客户认证数据。 [CE019, CE020, CE021, CE022, CE023, CE024]
| 控制 / 认证 / 指标 | 状态 | 范围 | 缺口 / 尽调问题 |
|---|---|---|---|
| 美国出口管制(EAR)合规 | 生效——招聘限制适用于 10 个国家 | 光子芯片设计的技术出口和人员资格 | 未披露出口管制许可清单;EAR 分类范围不清楚 |
| CPO 热管理 | 工程设计阶段;尚未在硅片上验证 | 动态 AI 工作负载下光子 IC 的温度稳定性 | 没有公开热测试结果或工作温度范围 |
| OIF CPO 接口标准合规 | 标准仍在制定(OIF 联盟推进中);Olix 合规未确认 | 光调制格式、机械接口、热规格 | 无法验证能否与其他 CPO 厂商 / 模块互操作 |
| 硅光子良率 / KGD 保证 | 未公开披露;行业文献记录了 CPO 良率挑战 | CPO 封装前,光子 chiplet 的 Known-Good-Die 比率 | 未点名代工伙伴;没有披露良率数据或测试方法 |
| 产品质量 / 可靠性认证 | 未公开披露(收入前、硅片前) | 制造工艺认证、可靠性测试(HTOL、热循环) | 第一片硅出来前无法评估;未提及 ISO / IEC 认证 |
状态判断基于公开信息;没有披露不应被等同于没有内部流程。
[CE019, CE020, CE021, CE022, CE023, CE024]| 日期 / 阶段 | 里程碑 | 状态 | 含义 | 来源 |
|---|---|---|---|---|
| 2024 年 3 月 | 公司成立(当时为 Flux Corp Ltd);确立光子推理 thesis | 已完成 | 表明在 2026 年融资前已有约 2 年设计工作 | UK Companies House 备案(ch1 中的 SO007) |
| 2026 年 1 月 | 公司更名为 OLIX Computing Limited;发布 DX-1 和宣言 | 已完成 | 公开身份切换;DX-1 品牌推出 | Olix 宣言(SE001) |
| 2026 年 2 月 | $220M Series A 融资;宣布扩建 Bristol 工程设施 | 已完成 | 形成 18–24 个月硬件 runway;团队扩张到 200+ 人 | SiliconANGLE(SE009);TechFundingNews(SE020) |
| 2026 年(H2 估算) | FPGA 原型验证;光链路 BER 和功耗 characterization | 进行中(由活跃 FPGA 和光学工程师招聘推断) | 硅片 tape-out 决策的 gate 里程碑 | Olix FPGA JD(SE004);光学 JD(SE002) |
| 2027 年(公司目标) | DX-1 首次硅片 tape-out 和初始客户接触 | 尚未确认;由媒体报道和招聘节奏推断 | 关键执行风险;任何 CPO 良率或热问题都会拉长时间线 | 多家媒体(SE009、SE015、SE016);edgen.tech 反向观点(SE007) |
2027 年里程碑由媒体报道和招聘紧迫度推断;Olix 尚未公布正式交付日期。
[CE001, CE028, CE034, CE042]评估截至 2026 年 6 月,Olix 在关键产品与技术维度上的成熟度和差异化强度。
成熟度评级是分析师基于公开职位描述、宣言和第三方报道作出的评估,并非 Olix 披露。
[CE019, CE020, CE021, CE022, CE023, CE024]5.4 附录
06客户
6.1 目标买方分层与市场定位
Olix 公开材料没有点名任何客户,但公司的计算宣言和招聘规格一致指向一个狭窄、资本密集的买方群体。宣言把问题表述为「即便是最具主导地位、在硅、软件和供应链上拥有全栈控制的玩家,也无法解决」。这一定义把大型 hyperscaler 和云运营商——唯一在所需规模上同时拥有硅、软件和供应链的组织——锁定为 Olix 面向的受众,也隐含为初始目标买方。平台集成工程师职位描述说得更明确:该岗位是「OLIX 如何作为 hyperscale AI infrastructure 服务大模型的技术权威」。 证据中可以看到三层买方。第一,Tier-1 hyperscaler(Google、Microsoft、Amazon、Meta),它们运营每天数十亿 token 规模的自有推理集群,既有评估新型硅片的技术深度,也有吸收多年认证计划的资本。Broadcom Q4 FY2025 财报披露,来自 hyperscale 客户的定制 AI 加速器订单积压超过 $50 billion,说明这类买方的基础设施资本强度。第二,inference-as-a-service 云供应商——CoreWeave、Lambda Labs 等——购买 GPU 和加速器容量,再以 API endpoint 转售。这类买方在结构上有动力采用更低 cost-per-token 的替代方案,只要产品能通过认证,因为其商业模式按 token 定价。Groq 2025 年 9 月以 $6.9 billion 估值融资 $750 million,当时已在自有 LPU 架构上服务超过 200 万名开发者,证明专用云推理供应商一旦硬件进入生产,可以快速放大。第三,企业私有云运营商,它们在本地运行大模型推理工作负载,但采购周期更长,首单金额也小于 hyperscaler。 Olix 明确在宣言中表示避免 QAT/PTQ 模型量化,这会预先筛选买方:更可能是运行 frontier-model 推理、使用全精度或轻量量化权重的客户,因此偏向 hyperscaler 和大型推理云,而不是边缘部署或 SMB AI 运营商。Olix 每个招聘岗位中的出口管制措辞排除了部分地域的就业资格,也隐含提示初始客户关系预计会在哪里建立:英国、欧盟、美国及盟友市场。 所审查来源中没有任何披露的客户 pipeline、意向书或 design-win 公告。缺少客户证明本身就是这一阶段的重大信号:hyperscaler 硅片决策周期从初始接触到生产承诺通常需要 12–24 个月,而截至 2026 年 6 月,Olix 尚未完成首颗硅片 tape-out。首笔商业收入在结构上被锁在 2027 年产品交付目标之后。 [CU001, CU002, CU003, CU004, CU005, CU006]
| 分层 | 买方 / 付款方 | 用例 | 规模代理指标 | 收入 / 战略价值 | 证据缺口 |
|---|---|---|---|---|---|
| Tier-1 hyperscaler | 基础设施 VP / 云 P&L 负责人 | 机架规模大模型 decode 推理;优化单 token 成本 | 每天数十亿 token;年度资本开支达数十亿美元 | 主要 SAM;可能签多年供应合同 | 没有确认的接触或 LOI;hyperscaler 自研硅片风险 |
| 推理即服务云(CoreWeave、Lambda Labs) | 基础设施负责人 / 平台 P&L | 转售推理算力容量;按 token 定价的经济模型 | 每天数亿 token;大规模 GPU fleet | 次级 SAM;认证周期快于 Tier-1 | 没有公开的客户资格认证材料包 |
| 主权 AI / 国家算力计划 | 政府 CTO / AI 项目办公室 | 本土 AI 基础设施;数据主权要求 | 国家级部署;5–50 PW 算力目标 | 如果 UK/EU 项目偏好本土供应商,则存在机会 | UK CHIPS 战略提到支持,但没有 Olix 采购信号 |
| 企业私有云(金融服务、制药、国防) | CTO / 工程 VP | 本地部署前沿模型推理;保护专有数据安全 | 每天数千万 token;多机架部署 | 销售周期更长;初始订单更小 | 没有点名企业接触;出口管制复杂度高 |
| AI 研究实验室 / 前沿模型开发者 | ML 基础设施负责人 | 面向长上下文研究负载的高吞吐 decode | 机构级 GPU 集群;研究预算 $10M–$500M+ | 小众但背书价值高;可打开 hyperscaler 引荐 | 未披露与任何研究实验室合作 |
买方分层由 Olix 计算宣言、招聘信号和可比推理芯片采购模式(Groq、AWS Inferentia)推断。Olix 未披露客户 pipeline 或 pipeline 规模。 收入和战略价值按公开市场数据及类似加速器部署估算;并非 Olix 特定数据。
[CU001, CU002, CU003, CU004, CU005, CU009]| 指标 | 数值 | 日期 | 来源 | 置信度 | 含义 | 缺失分母 |
|---|---|---|---|---|---|---|
| 点名生产客户 | 披露为零 | 2026-06-22 | Olix 网站;媒体报道 | 高 | 尚无商业收入;与 2027 年出货目标一致 | 总 pipeline 规模和 LOI 数量未知 |
| 已披露 design win | 披露为零 | 2026-06-22 | 已审阅的全部 Olix 公开来源 | 高 | 仍在流片前阶段;按正常周期,设计赢单先于流片 | 活跃 RFI / NDA 阶段沟通数量未知 |
| 首批出货目标 | 2027(管理层目标) | 2026-02 | SiliconAngle;FT;TechFundingNews | 中 | 收入最早 2027 年开始;基准情形为 2028–2029 年 | 未把里程碑绑定到具体硅片流片日期 |
| MLPerf 推理提交 | 零 | 2026-06-22 | MLCommons 基准页面 | 高 | 还无法提交;流片前阶段跑不了推理基准 | 未披露参与基准测试的时间表 |
| 开发者社区参与(vLLM / SGLang PR) | 已确认贡献为零 | 2026-06-22 | vLLM GitHub 仓库;招聘职位描述 | 中 | 软件集成是商业化前关口;工作仍在推进 | 分配给集成的内部开发者人数未知 |
| 超大规模云厂商资格认证启动 | 未公开披露 | 2026-06-22 | 无新闻稿或申报文件 | 高(缺少证据) | 无法确认正式资格认证流程是否已启动 | 超大规模云厂商关系的身份和数量未知 |
所有指标均来自公开来源;零值反映的是已确认没有公开披露,而非确认没有任何活动。内部管线指标(LOI、NDA、设计赢单沟通)属于私人信息,除非直接接触 Olix 管理层,否则无法核验。
[CU006, CU007, CU008, CU019, CU020, CU027]描绘超大规模云厂商买方从最初认识新型推理加速器到生产部署的各阶段,并把 Olix 当前状态定位在流片前评估关口。
旅程图由行业标准的超大规模云厂商硅片采购模式、Olix 公开材料和 Broadcom 2025 财年第四季度业绩会构建。Olix 未披露具体接触时间表;所有阶段时长均由可比项目估算。
[CU001, CU013, CU019, CU020, CU021, CU022]6.2 采用路径、销售动作与集成要求
Olix 从当前 pre-silicon 状态走到首笔生产收入,需要依次穿过三道门:(1)硅片 tape-out 与功能验证,(2)接入主流开源推理栈,(3)客户认证。截至 2026 年 6 月,没有任何一道门被公开证明已经通过。 软件侧,Olix 的招聘确认其计划接入 vLLM、SGLang 和 NVIDIA Dynamo。vLLM 项目已积累超过 2,000 名贡献者,是部署最广的开源推理 serving 框架;截至 2026 年中,其 GitHub 仓库 star 超过 50,000。把新型加速器集成进 vLLM,需要硬件抽象层工作、关键 attention 和矩阵运算的 operator kernel,以及被接受的 upstream pull request——截至研究日期,公开代码库中没有任何一项归属于 Olix。平台集成工程师岗位说明显示,Olix 现在才在为这项工作招聘,这与软件集成仍是商业化前依赖、而非已完成能力的判断一致。该岗位还要求深入熟悉用于分离式 KV-cache 管理的 NIXL 和 Mooncake TE,而这两项本身也是仍在行业内稳定下来的早期标准。 客户认证侧,主导参照是 hyperscaler 硅片评估流程。Broadcom CEO Hock Tan 在 Broadcom Q4 FY2025 财报电话会上给出了关于光子采用时点的最清晰行业信号:「Silicon photonics won't matter in the datacenter anytime soon」,并明确表示,必须先耗尽铜互连和 pluggable optics 这两波技术,硅光子才会成为采购必需品。这个排序来自一家手握超过 $50 billion hyperscaler 定制硅订单的公司 CEO,直接回应了买方对光子推理加速器的准备度。另一个独立信号是,DARPA PICASSO 项目在 2026 年初拨出 $35 million,用于突破基础光子电路扩展限制,说明即便政府资助研究,仍把系统尺度光子视为未解问题。 假设 Olix 在 2026 年末或 2027 年初拿到首颗硅片,并达到功能验证目标,典型 hyperscaler 认证周期仍需要额外 12–24 个月,用于系统级测试、可靠性 characterization、供应链审计和采购谈判,然后才可能下达生产采购订单。这个时间线意味着,最乐观情景下商业收入也不会早于 2027 年末,基础情景则落在 2028–2029 年。Inference-as-a-service 层(CoreWeave、Lambda Labs)可能推进更快,因为其采购流程轻于 Tier-1 hyperscaler 供应商认证计划;但它们利润率更薄,也不太可能在没有可靠性 benchmark 验证的情况下接受原型阶段加速器。 MLCommons MLPerf 推理 benchmark 是行业标准的认证信号——Groq、AWS Inferentia 和 Google TPU 都参与,并用公开 benchmark 结果建立客户可信度。考虑到 Olix 仍处于 pre-silicon 状态,尚未提交 MLPerf 属于预期之内;但即使首颗硅片产出,缺席 MLPerf 仍会是企业买方信任的实质障碍。 [CU013, CU014, CU015, CU016, CU017, CU018]
| 扩张驱动 | 集中度风险 | 若不解决的影响 | 尽调路径 |
|---|---|---|---|
| 首个设计赢单之后分散到多客户 | 拿下第二个客户前,依赖单一客户 | 锚定客户任何流失或减量都会清零全部收入 | 向管理层询问管线数量和阶段分布 |
| vLLM / SGLang 生态采用带动开源集成拉力 | 如果社区采用面窄,集成会局限于一个推理栈 | 工作负载覆盖不足,压缩可触达买家池 | 监控 Olix 对 vLLM 和 SGLang 仓库的代码贡献 |
| 推理即服务层扩张(CoreWeave、Lambda)作为第二层买家 | 市场可信度信号集中依赖一线超大规模云厂商批准 | 若先拿不到超大规模云厂商背书,推理云采用会延后 | 跟踪 CoreWeave / Lambda 公开硬件公告和 RFP 周期 |
| 地理扩张——英国主权 AI、欧盟项目、美国 DoD 边缘 | 英国本土为主;出口管制地域限制可触达市场 | 出口管制排除项(俄罗斯、中国、伊朗等)封顶全球 SAM | 确认 DX-1 作为光子器件的 ITAR / EAR 分类 |
| 工作负载从解码扩到 prefill 或训练 | 仅解码架构限制初始买家范围 | 覆盖不了完整推理栈;prefill 密集型买家会转向别处 | 确认 DX-1 路线图是否包含解耦 prefill 支持 |
扩张驱动来自 Olix 的架构定位、软件集成策略和买家细分地图。收入前阶段天然带来集中度风险;没有一项代表当前负面事件。所有尽调路径都需要直接接触管理层。
[CU004, CU011, CU012, CU037, CU038, CU039]估算 DX-1 买方漏斗从超大规模云厂商初步认知到生产部署的各阶段转化,并将 Olix 当前评估位置放在流片前评估关口。
转化率估算仅作示意;它们由 Groq 从 LPU 发布到 GroqCloud 生产部署的路径类比得出,并参考 Broadcom 对超大规模云厂商加速器认证时间表的评论。没有可用的 Olix 特定管线数据。
[CU013, CU015, CU019, CU022]从五个证据维度评估 Olix 及三个已有公开生产部署记录的可比加速器项目的客户验证质量。
可比对象证据基于截至 2026 年 6 月公开可得的来源。Groq 客户数量和案例研究数据来自 groq.com;AWS Inferentia 数据来自 aws.amazon.com;Google TPU 数据来自 Google Cloud 文档。Olix 列反映已确认缺少公开客户验证,并非断言其内部活动为零。
[CU006, CU026, CU028, CU030, CU031, CU034]6.3 客户证明缺失、留存结构与集中度风险
截至 2026 年 6 月,Olix 的客户证明状态是二元的:公开披露客户为零,披露试点为零,披露意向书为零,披露 design win 为零。SiliconAngle 指出,「it's unclear what optical components are included in the processor or how they're used」,反映即便在专业技术记者中,面向客户的技术主张也整体不透明。率先报道融资消息的 Financial Times 提到,Olix 预计 2027 年开始向客户出货芯片,并把这视为管理层目标,而不是合同承诺。 由此产生的结构性风险是集中,而非多元。当 Olix 拿下首个客户时,该账户实际上会定义公司 100% 收入。如果客户是 Tier-1 hyperscaler,这段关系将受该 hyperscaler 标准供应条款约束,通常包括需求波动条款、多源采购要求,以及对客户署名的排他限制。如果首个客户是 inference-as-a-service 供应商,关系会更透明,但收入规模更小,也更暴露于该供应商自身客户流失。在任一情景下,单一 anchor customer 都会构成关键集中度风险,直到第二、第三个客户关系成形。 留存指标——net revenue retention、gross revenue retention、churn rate、repeat purchase rate、contract duration——完全不可得,因为还没有生产收入。Amazon 产品页上的 AWS Inferentia 客户证言展示了定制 AI 推理芯片在多年生产部署后的结果(NTT PC Communications 实现 4.5× 吞吐、25% 更低延迟、90% 更低成本),可作为芯片供应商达到商业规模后留存证据应呈现什么样的参照;Olix 距离这个阶段还有数年。Groq 的客户名单说明了 reference-quality 问题:其案例研究点名 Willow(实现零停机和响应快 500 ms)、PGA of America 等,且都依赖已生产、已部署并在生产环境运行的硬件。 Broadcom CEO 记录下来的光子特定买方迟疑,是市场时点上最直接的反向信号:有能力吸收 Olix 产品的买方,尚未相信光子是必要采购。这形成了依赖窗口的销售论点——Olix 需要买方先相信光子转型,且要赶在资金充足的 incumbents(NVIDIA、Broadcom 或 hyperscaler 自有团队)补上同一架构缺口之前。2026–2027 年因此是关键市场时点窗口:早期客户对话和 design-win 披露会显著降低商业论点风险;如果客户进展到 2027 年仍持续沉默,这一缺口会升级为阻断性尽调问题。 [CU026, CU027, CU028, CU029, CU030, CU031]
| 供应商 / 项目 | 细分市场 | 部署 / 用例 | 状态(量产 / 试点) | 成果证据 | 局限 |
|---|---|---|---|---|---|
| Olix (DX-1) | 超大规模云厂商 / 推理云(目标) | 机架规模解码推理;vLLM / SGLang 集成 | 量产前——截至 2026-06-22 尚未出货硅片 | 无;未发布面向客户的基准测试或案例研究 | 完全缺少客户证明;2027 年目标未经验证 |
| Groq (GroqCloud LPU) | 推理即服务云和企业 API | 为 200 万+开发者和 Fortune 500 企业提供 LLM 推理服务 | 量产——北美、欧洲、中东多个数据中心 | Willow:零停机,响应快 500 ms(2025 年发布) | 架构不同(LPU 不是光子);不可直接对比 |
| AWS Inferentia2 (Inf2) | 超大规模云厂商内部和云服务 | 在 Amazon EC2 Inf2 实例上做 Transformer 推理 | 量产——作为 EC2 实例类型可用 | NTT PC:吞吐量 4.5x、延迟低 25%、成本比 GPU 低 90% | 超大规模云厂商自用硅片;Olix 必须与 Inferentia 的经济性竞争 |
| Google Cloud TPU v5 云端加速器 | 超大规模云厂商内部和云服务 | 覆盖 Google 产品的大模型训练和推理 | 量产——可通过 Google Cloud 使用;已集成进 Vertex AI | 已在 Google 规模量产;未审阅单个客户案例研究 | 自用硅片;降低 Google 对第三方加速器的开放度 |
截至研究日期,Olix 没有公开命名或确认的客户。Groq、AWS Inferentia2 和 Google TPU 行作为对照,说明量产级客户证明应长什么样;它们不是 Olix 当前或潜在客户。
[CU006, CU007, CU025, CU026, CU028, CU030]| 指标 | 数值 / 状态 | 细分市场 | 置信度 | 尽调问题 |
|---|---|---|---|---|
| 净收入留存(NRR) | 不适用——收入前 | 所有细分市场 | 高(已确认缺失) | 商业收入开始后,索取首次披露的 NRR |
| 总收入留存(GRR) | 不适用——收入前 | 所有细分市场 | 高(已确认缺失) | 同 NRR;结合首批客户合同条款一起跟踪 |
| 客户流失率 | 不适用——无客户 | 所有细分市场 | 高(已确认缺失) | 询问 LOI 或试点协议是否含惩罚条款 |
| 合同期限 / 续约条款 | 未披露 | 超大规模云厂商(预期主线) | 低——公开域没有合同语言 | 向管理层索取合同条款样本;与 Groq 模式对比 |
| 客户满意度 / NPS | 不适用——无客户 | 所有细分市场 | 高(已确认缺失) | 首次部署开始后,索取初始试点满意度数据 |
| 重复 / 复购信号 | 不适用——尚无产品出货 | 所有细分市场 | 高(已确认缺失) | 跟踪 2027 年后公布的追加订单或扩大试点范围 |
Olix 尚未产生商业收入,因此所有留存指标在结构上都不可得。空值是已确认缺失,不是四舍五入或估算。该表为未来刷新周期记录每项指标的尽调路径。
[CU032, CU033, CU036]用示意性留存队列结构展示 Olix 与可比推理芯片供应商的数据可得性。所有 Olix 数值均为 null(无客户、无收入)。可比对象数值仅由公开信号估算。
Olix 行是已确认的 null:截至 2026 年 6 月没有生产收入。Groq 和 AWS Inferentia 留存估算是分析师根据增长轨迹报道(Groq 以 $6.9B 估值融资 $750M,并提到 2M+ 开发者)及 AWS 产品生命周期推断的近似值;这些数值未经公司验证,不能作为已报告指标引用。队列第 0 月 = 首次商业部署月份。
[CU032, CU033, CU035]6.4 附录
07风险
7.1 技术与制造风险
Olix 的光子推理芯片计划以 co-packaged optics(CPO)为核心;行业来源普遍把这项技术归为早期、高制造风险。Siemens EDA 2026 年 2 月 CPO 分析指出四个叠加挑战:热管理(photonic IC 对温度高度敏感;AI 工作负载的瞬时热波动会移动光波长并降低调制器效率);测试扩展性(光学对准需要微米级精度,远严于电测试,使自动化量产测试成为主要瓶颈);供应链不成熟(硅光子制造的工艺波动高于成熟 CMOS,而共封装组件的良率波动会带来放大的成本冲击,因为共封装后几乎无法返工);以及标准碎片化(CPO 互操作标准仍在 OIF 等机构推动下成熟,早期产品会与客户机架架构产生集成摩擦)。EdgeN 于 2026 年发布的反向行业报告称,CPO 量产因良率问题面临 2026 年之后的延迟,从而验证了 Siemens 的风险判断。 在这个背景下,Olix 没有披露任何 silicon tapeout 结果、良率数据或系统级演示。它宣称的 2027 年首次出货里程碑,需要在 2026 年 6 月 run date 之后约十二个月内,完成一家 foundry partner(可能是 TSMC Specialty Technology)的完整硅光子 tapeout、与 OSAT 的共封装认证、软件栈集成,以及 hyperscaler 认证。TSMC 的 foundry 技术页面确认硅光子作为 specialty process 提供,但公司没有公开披露光子良率路线图。ASML 的 EUV 光刻系统支撑全球先进半导体制造,是所有先进 foundry 节点共同的上游依赖。供电是次级技术风险:Siemens 指出,到 2027 年,AI 数据中心电力需求预计全球增长 50 percent,且约 60 percent 数据中心能耗用于数据移动而非计算——这正是 CPO 要解决的问题,但也会给 Olix 机架级设计带来极端热要求。 [CR001, CR002, CR003, CR004, CR005, CR006]
| 失效模式 | 可能性 | 严重性 | 缓释成熟度 | 剩余敞口 | 未解决缺口 |
|---|---|---|---|---|---|
| CPO 光子良率失败——生产中微米级对准漂移 | 高 | 严重 | 低——Olix 未披露生产良率数据 | 项目延期到 2027 年之后;反复流片带来成本超支 | Olix 未发布流片结果或良率目标 |
| 热管理失效——AI 工作负载瞬态下波长漂移 | 中 | 高 | 低–中——架构阶段就需要热协同设计 | DX-1 量产推理准确率下降;客户拒收 | 未披露 DX-1 硅片热测试数据 |
| 测试扩展性瓶颈——自动化光学对准测试认证 | 高 | 高 | 低——按 Siemens CPO 分析,这是全行业未解问题 | 无法认证超大规模云厂商 PoC 出货所需规模 | Olix 未公开披露自动化测试策略 |
| CPO 标准碎片化——OIF / 生态互操作性缺口 | 中 | 中 | 中——OIF 工作组活跃,但时间表不确定 | 客户集成摩擦;标准采用前市场有限 | 截至 2026 年 6 月,OIF 标准发布日期尚未确认 |
可能性和严重性基于 Siemens EDA CPO 分析(2026 年 2 月)、EdgeN 行业报告(2026 年)和可比光子芯片项目数据。缓释成熟度为作者评估;没有 Olix 内部数据。
截至 2026 年 6 月,按技术、监管、商业和治理维度对 Olix 风险严重程度进行分类。
可能性和影响评级是基于截至 2026-06-22 公开来源的定性评估,并非精算估计。
[CR001, CR003, CR005, CR011, CR014, CR022]7.2 监管、法律与出口管制风险
Olix 面临三类不同的监管与法律敞口。第一,美国 Bureau of Industry and Security(BIS)对 advanced computing integrated circuits 的出口管制正在明显收紧。截至 2026 年 6 月,BIS 主页确认,向总部位于 Country Group D:5(包括中国)或 Macau 的实体,或最终母公司位于这些司法辖区的实体出口先进计算项目,需要许可证——即便最终收货地点在其他地方。BIS 还把 Authorized IC Designer 合规时间线延长至 2026 年 12 月 31 日,让公司有更多时间提交申请,但这一延长本身确认制度已生效且适用。一家英国光子 AI 芯片公司如果想卖入美国邻近云基础设施,或直接卖给美国总部 hyperscaler,就必须确认其供应链和终端客户基础不会产生许可证敞口;考虑到 Olix 架构需要先进 foundry 工艺和可能触及 dual-use 设备分类的光子组件采购,这并不简单。 第二,英国 National Security and Investment Act 2021(NSI Act)赋予英国 Secretary of State 广泛权力,可在十七个敏感行业中 call in、审查并阻止收购——半导体和先进材料被明确列入。任何考虑收购 Olix 的投资人或科技公司,一旦持股超过特定门槛,NSI Act 合规就是强制要求。这一审查机制可能延迟或阻止战略退出,应作为重大风险向投资人披露。它还意味着,如果任何外国政府关联实体试图取得重要股权,Olix 的治理会受到英国国家监督。 第三,知识产权风险较高,因为 Olix 未披露任何已授权专利组合,而竞争对手 Lightmatter 持有关于光子矩阵乘法和光互连架构的已公开专利。截至 run date,UK IPO 检索和 EPO register 没有在「Olix」或「OLIX Computing」名下显示重大已授权专利,尽管早期专利申请可能正在推进。NIST 的 AI Risk Management Framework 和英国政府 2023 年 3 月发布的 pro-innovation AI regulation 白皮书构成监管背景,但二者都没有为 AI 芯片硬件初创公司引入短期许可义务。 [CR011, CR012, CR013, CR014, CR015, CR016]
| 规则 / 法案 / 风险敞口 | 司法辖区 | 状态 | 可能性 | 严重性 | 缓释 | 剩余敞口 | 尽调路径 |
|---|---|---|---|---|---|---|---|
| 美国 BIS EAR——先进计算物项(授权 IC 设计方) | 美国 | 生效;授权 IC 设计方截止日期为 2026 年 12 月 31 日 | 中 | 高 | 申请授权 IC 设计方身份;审计供应链终端用户 | 向 D:5 司法辖区出口需要许可证;市场准入受限 | 确认 BIS 申请已提交;审阅客户地域 |
| 英国《国家安全与投资法》2021——半导体行业 | 英国 | 已生效;超过持股阈值须强制申报 | 低–中(M&A 场景) | 高 | 任何超过阈值的交易提前通知 BEIS;保留 NSI 专项法律顾问 | 政府可阻止战略收购;退出路径承压 | 确认 NSI 申报历史;就任何投资者持股取得法律意见 |
| 英国出口管制——两用物项和受控技术 | 英国 | 根据战略出口许可生效;半导体设备列入附表 1 | 中 | 中 | 必要时取得 ECJU 许可证;筛查出口目的地 | 涉及受控物项的代工晶圆存在再出口风险 | 审计受控物项物料清单;确认 ECJU 合规 |
| 知识产权 / 专利侵权(光子计算) | 全球 | 未披露专利组合;Lightmatter 和竞争对手持有已授权专利 | 中 | 高 | 做 FTO 分析;建立防御性专利组合;必要时取得授权 | 核心计算架构面临禁令或权利金风险 | 委托自由实施研究;披露专利申请状态 |
| 欧盟 AI 法案——通用 AI 系统提供者义务 | 欧盟 | 已颁布;向欧盟部署 AI 系统供货的硬件提供商可能承担义务 | 低 | 低–中 | 监控 GPAI 条款;产品接近部署时与 EU AI Office 沟通 | 如果 DX-1 卖入欧盟超大规模云厂商,会带来合规成本和文档负担 | 确认监管顾问简报;跟踪欧盟 AI 法案下的授权法案 |
行按严重性排序(最严重在前)。可能性和严重性是基于截至 2026-06-22 公开监管文本和行业分析的定性评估;不构成法律意见。
[CR011, CR012, CR013, CR014, CR015, CR016]有向图展示主要技术和监管风险如何传导为 Olix 的收入不达预期、资本断崖和估值重置。
传导路径是基于风险登记表和可比深科技半导体项目构建的分析框架,并非公司披露。
[CR002, CR004, CR006, CR013, CR019, CR024]7.3 商业、财务、竞争与供应链风险
Olix 的财务风险画像由四个叠加特征塑造:公司仍处于收入前;融资时达到独角兽估值;面对资本密集的半导体开发路径;按 $220 million 融资对应模型化 $3–5 million 月烧钱率估算,30–45 个月 runway 必须在任何商业收入抵消支出之前,覆盖多年 R&D、tape-out 成本、封装认证和 go-to-market 招聘。半导体分析机构 Jon Peddie Research 指出,AI 处理器市场竞争激烈且资本密集。如果 AI 支出情绪转弱——触发因素可能是模型效率提升导致算力需求下降、云 CapEx 收缩,或市场回调——下一轮融资可能以显著更低估值或苛刻条款完成。Semiconductor Industry Association 的 2025 State of the Semiconductor Industry 报告也凸显了参与先进芯片设计竞争所需资本规模。 竞争风险尖锐且迫近。NVIDIA 数据中心收入在 Q3 FY2026 达到 $51.2 billion,公司还在持续为推理工作负载优化 Blackwell,缩小 Olix 光子路线必须在 2027 年前跨越的性能差距。最接近的光子计算同行 Lightmatter 已融资 $400 million,产品开发进展更靠前;CNBC 2026 年 5 月报道称,芯片公司正竞相解决 AI 光子瓶颈。如果 2027 年目标滑到 2028 年或更晚,NVIDIA、Google TPU 和 AWS Inferentia 将获得更多时间在传统 CMOS 架构上缩小效率差距,侵蚀 Olix 想利用的 wedge。 供应链有两个难以缓释的单源依赖。TSMC 是 Olix 所需工艺节点上先进硅光子的唯一可信 foundry;任何对 TSMC 访问的扰动——无论来自地缘政治、产能还是认证——都会实质推迟项目。TSMC 上游,ASML 的 EUV 光刻系统是全球集中 chokepoint;EUV 工具可用性短缺会同时影响所有先进 foundry 节点。用于 III-V 化合物半导体光源(基于 indium phosphide)的激光供应商构成第三个供应链约束,因为 InP 晶圆供应地域集中,也受自身产能限制。 [CR019, CR020, CR021, CR022, CR023, CR024]
| 依赖 | 交易对手 | 角色 | 集中度 | 失败场景 | 严重性 | 缓释 | 剩余敞口 |
|---|---|---|---|---|---|---|---|
| 硅光子代工厂 | TSMC(Specialty Technology) | 所需工艺节点下唯一可信的 SiPh 量产代工厂 | 单一来源 | 地缘政治中断、产能配给或工艺认证失败 | 严重 | 接洽替代代工厂(IMEC / GlobalFoundries SiPh);谈判产能预留 | 未披露后备代工厂;缓释未经确认 |
| EUV 光刻设备 | ASML | 所有先进代工节点的上游供应商 | 全球 EUV 工具垄断 | EUV 工具短缺扰乱 TSMC;所有先进节点同时受影响 | 高 | 没有可用的直接缓释;尾部风险由所有先进芯片公司共同承担 | 系统性风险;无法由单家公司缓释 |
| III-V 激光器 / 光源供应 | InP 晶圆和激光器供应商(Coherent、II-VI / Lumentum) | 为光子集成电路提供光源 | InP 晶圆供应地域集中 | InP 晶圆供应中断会延迟激光集成;规模化上没有替代品 | 高 | 多来源激光器策略;库存缓冲 | Olix 未披露激光器供应商或多来源策略 |
| 超大规模云厂商首个客户 | 未命名(目标:美国 / 欧盟超大规模云厂商) | 首个设计导入客户;验证商业化扩张 | 高——记录中没有已签客户 | 没有超大规模云厂商签署评估协议;DX-1 卡在纸面规格 | 高 | 通过生态伙伴早期接触;提供免费 PoC 硅片 | 截至 2026 年 6 月,没有公开确认的客户接触 |
合作伙伴细节基于 Olix 公开招聘(据此推断供应链范围)、Siemens CPO 分析、TSMC 技术路线图和 ASML 年度披露。具体合同关系未披露。
Olix DX-1 项目的关键外部依赖,覆盖代工、设备、材料、客户、资本和监管机构。
依赖关系由公开招聘信息、供应链分析和监管框架审阅推断;Olix 未披露具体合同。
[CR007, CR020, CR025, CR026, CR027, CR033]7.4 治理、人员与论点击穿标准
治理风险集中在一个人身上:James Dacombe 在 2026 年 2 月融资时 25 岁,自公司成立以来一直是唯一连续任职董事,也是 Olix 整体战略的公开面孔。媒体报道和公司材料把 Dacombe 同时定位为光子论点的智识架构师,以及负责执行的首席执行官。没有公开证据显示公司有具备 AI 芯片产品交付经验的联合创始人、CTO 或 COO。董事会包括经验丰富的运营者——Jonathan Heiliger(前 Facebook infrastructure)、Nicholas McKeown(Stanford 网络教授)和 Taavet Hinrikus(Transferwise 联合创始人兼投资人)——但董事会监督不能替代深科技运营团队。Olix 自己的文化材料承认这场建设要求极高:对不匹配员工提供 18 周 paid exit 条款,说明公司知道人才匹配和留存很脆弱。 光子和化合物半导体工程师是全球最稀缺的专才之一。公司招聘页显示岗位横跨 optics、photonic packaging、FPGA、ASIC、manufacturing operations 和 process integration——这套招聘议程必须从极薄的人才池中招人。UKtech.news 和 IDCNova 关于 2026 年 2 月融资的报道提到,公司计划从 70 人增长到 200-plus 人,意味着 Olix 必须在竞争激烈的劳动力市场中把团队规模接近翻三倍,然后才能交付首个产品。 风险登记表中可以识别出论点击穿标准:(a)如果到 2026 年底仍未展示具备可接受光子良率的硅光子 tapeout,首次出货会被推迟到 2027 年之后,并进入资本 runway 危险区;(b)如果到 2027 年中仍未拿下首个具名 hyperscaler design-in,说明 DX-1 没有通过企业采购门槛;(c)若 BIS 执法行动或 NSI Act 阻止令阻断关键供应链交易或收购退出,结构上应判为不合格;(d)若前三大 hyperscaler 中任意两家宣布 AI infrastructure CapEx 下调 40-percent 或以上,会侵蚀支撑 Olix 2027 年时间线的需求论点。投资人应要求 Olix 在下一次融资事件前,以合同方式披露 tapeout 状态和 foundry partner。 [CR030, CR031, CR032, CR033, CR034, CR035]
| 角色 / 职能 | 依赖或缺口 | 影响可能性 | 严重性 | 缓释 | 尽调路径 |
|---|---|---|---|---|---|
| CEO / 创始人(James Dacombe) | 唯一持续任职的董事;未披露联合创始人;25 岁,且没有此前经营半导体公司 P&L 的经验 | 中 | 严重 | 董事会提供监督;经验丰富的董事(Heiliger、McKeown)担任运营顾问;需要继任计划 | 索取继任计划、联合领导架构和董事会章程;评估 CEO 发展计划 |
| 首席技术官 / 硅片负责人 | 未公开披露 CTO;不清楚谁负责硅片流片执行 | 高 | 高 | 招聘流程可从职位发布中看到;确认 CTO 任命和过往履历 | 识别 CTO 并确认此前流片交付经验;审阅组织架构图 |
| 光子工程人才池 | 全球共封装光学和 SiPh 工程师人才池极薄 | 高 | 高 | 有竞争力的薪酬(提到 £24k 本地生活补贴);大学合作;部分岗位接受远程 | 对照计划评估招聘速度;确认 2026 年 2 月融资以来关键员工留任情况 |
角色评估基于 Companies House 董事记录、Olix 招聘页面和媒体报道。Olix 尚未发布完整组织架构图。
| 风险 | 可监控触发项 | 阈值 / 事件 | 行动含义 |
|---|---|---|---|
| CPO 良率失败 | 流片结果;光子良率披露 | 到 2026 年底仍未披露流片结果,或通过光学规格的裸片良率低于 30% | 暂停下一轮承诺;交割前要求提供流片数据 |
| 2027 年发布滑期 | DX-1 工程样品交付日期 | 到 2027 年 Q3 仍未向已命名超大规模云厂商交付工程样品 | 重新评估估值;延长尽调;下调为资本市场风险 |
| 融资环境恶化 | 超大规模云厂商的 AI 基础设施 CapEx 公告 | 三大超大规模云厂商中有两家宣布 AI 基础设施 CapEx 削减 >30% | 加快下一轮流程;建模过桥轮或降价融资情景 |
| 出口管制禁运 | BIS 针对英国 AI 芯片出口商的执法行动 | BIS 将先进 SiPh 芯片加入 CCL,或发布影响英国出口商的通用许可限制 | 委托市场准入法律意见;只按可进入市场修订收入模型 |
终止标准阈值由作者基于风险台账分析定义,并非 Olix 披露。投资者应定义自己的监控参数。
7.5 附录
08估值
8.1 融资背景与上一轮估值框架
Olix 2026 年 2 月 Series A 完成 $220 million 融资,由 Hummingbird Ventures 领投,Plural、LocalGlobe 和 Entrepreneurs First 参投。媒体报道显示投后估值超过 $1 billion,使 Olix 成为最快达到独角兽地位的英国初创公司之一——从 2024 年 3 月成立到这一里程碑约 24 个月。若计入估计 $25–30 million 的种子轮和 pre-seed 融资,累计融资约 $250 million。Companies House 备案历史记录了 2026 年 3 月 5 日配股后的 statement of capital,总资本为 GBP 287.8572,确认融资已完成且新股已发行。截至 2025 年 8 月 31 日编制的 total exemption accounts——公开记录中唯一法定账目——只披露资产负债表数据,并未记录运营收入。 last-round-price 方法把当前标记锚定在 $1 billion,意味着投资人为得出与 $220 million primary 一致的投后估值,隐含接受约 22 percent 稀释。实践中,优先股结构包含 liquidation preferences,但规模和类型并未公开;因此在低于 headline 的退出情景下,普通股和创始人持股价值可能远低于 headline 暗示。$1 billion 投后估值约为总投入资本(约 $250 million 已融资)的 4×,这一覆盖倍数符合 2026 年 Q1 全球中期深科技硬件轮次——SemiEngineering 统计,2026 年 Q1 有 80 家半导体初创公司融资超过 $8.4 billion,其中 18 轮超过 $100 million。光子领域在该季度仍是「popular segment for investment」。 投资人似乎给两类价值定价:(1)OTPU 在推理市场围绕 NVIDIA Blackwell 或后续产品固化前投入生产的战略期权价值;(2)M&A 期权价值,即 hyperscaler 和 incumbent 芯片公司收购光子能力——证据包括 Marvell 在 2026 年初收购 Polariton Technologies(基于 plasmonics 的硅光子),以及 NVIDIA 宣布向 Lumentum、Coherent、Marvell、Corning 等光子公司投入约 $2 billion,并参与 Ayar Labs 的 $500 million Series E。这些投资的频率和规模,为 Olix 的资产提供了一个可信战略底线,前提是 OTPU 论点在退出时至少部分成立。 [CV001, CV002, CV003, CV004, CV005, CV006]
| 维度 | 评估 | 依据 |
|---|---|---|
| 建议 | 跟踪 | 尚未产生收入;没有已交付芯片;首款产品交付至少还要 18 个月;关注晶圆厂 tapeout,可能成为重估催化剂 |
| 置信度 | 中 | 融资和阶段已确认;产品路线图和竞争格局仍有高度不确定性 |
| 风险评级 | 高 | 资本密集、执行风险、竞争替代、单一创始人 CEO、优先权悬置 |
| 估值立场 | 偏高 | 在没有收入时估值已超过 $1B,价格已经押注一长串尚未验证的执行里程碑 |
| 重估催化剂(上行) | 晶圆厂 tapeout 确认,加上 ≥1 份设计导入 LOI | 任一证据都会把立场推向合理,并把建议推向买入 |
| 重估催化剂(下行) | 错过 2027 年出货目标,或竞争对手签下 Olix 目标客户 | 任一触发因素都会把立场推向昂贵,并把建议推向回避 |
评估截至 2026-06-22。建议对价格敏感;一旦出现新证据,假设可能调整。
[CV001, CV005, CV024]| 主题 | 缺失证据 | 重要性 | 负责人或尽调路径 |
|---|---|---|---|
| 晶圆厂 tapeout 状态 | 没有公开确认的 tapeout 日期、节点或晶圆厂伙伴 | 验证芯片设计完整性和生产时间线;没有该证据,DX-1 2027 年出货目标完全是猜测 | 管理层直接披露;TSMC 或 GlobalFoundries 客户名单(受限) |
| 设计导入管线 | 公开域没有 LOI、NDA 或概念验证客户互动 | 判断 Olix 是否获得终端客户对其性能主张的验证;若证据可信,建议可从跟踪转向买入 | 向投资人披露的 NDA 保护材料;检查 Olix 法律顾问文件,寻找 Companies House 上潜在 NDA |
| 股权结构和优先权堆栈 | 清算优先权条款、参与型优先股状态和反稀释条款未披露 | 判断基准和熊市情景下的真实投资人回报;标题 $1B 估值未必反映普通股可获得的经济权益 | 条款清单 / 股东协议披露;Companies House SH01 文件寻找资本结构线索 |
| 2026 年现金消耗和跑道 | Companies House 账目只到 2025 年 8 月;没有中期数据 | 检验 $250M 融资能否覆盖到首笔收入;若每月 burn 为 $5M,Series A 完成后约有 50 个月跑道,但扩张到 200+ 人可能让 burn 翻倍 | 投资人报告包;英国审计豁免门槛可能触发 2026 年完整审计申报 |
| OTPU 基准结果 | 没有独立或公司发布的推理基准,无法与 H100/B200 或 Groq LPU 对比 | 支撑吞吐 / 兆瓦和延迟主张,而这些主张撑起整个投资论点 | 量产前芯片样品或仿真结果;MLPerf Inference 提交可作为未来里程碑 |
尽调事项按对建议的重要性排序。第 1 和第 2 项是买入建议的阻断项。
[CV004, CV006, CV036, CV037, CV038]从市场、产品、风险、估值四类证据推到「跟踪」建议,并列明估值重评条件。
[CV001, CV005, CV007, CV015, CV024]投委会按七个维度打分,量表为 1–10 分。红色(1–3)、琥珀色(4–6)、绿色(7–10)。
评分是分析师基于截至 2026-06-22 的公开证据作出的判断;公司内部数据会显著改变分数。
[CV007, CV008, CV009, CV010, CV015]8.2 可比估值、先例交易与公开市场背景
2025–2026 年 AI 硬件推理创业公司的私募融资,给出了估值区间的参照。商业化最靠前的可比公司 Groq,2025 年 9 月以 $6.9 billion 投后估值融资 $750 million,约为 Olix 隐含估值的 6.9×;但 Groq 已有出货的 LPU 产品、收入,并服务超过 200 万活跃开发者。Ayar Labs 2026 年 3 月融资 $500 million,估值 $3.75 billion,反映的是它在 NVIDIA NVLink Fusion 生态中共封装光学产品的位置。Positron AI 2026 年 2 月以超过 $1 billion 估值融资 $230 million,融资额和估值时间几乎与 Olix 完全可比;不同的是,Positron 已出货 FPGA 推理系统 Atlas,并有面向 2026 年底流片的自研芯片 Asimov。这组对比值得重视:Positron 在与 Olix 相同的隐含估值下,已经有出货产品和付费种子客户 Jump Trading,而 Olix 两者都没有。SambaNova($350M Series E)、Cerebras(2026 年 Q1 $1B Series H,准备 IPO)和 MatX($500M Series B)则为 AI 加速器赛道补上更多参照点。 从并购维度看,最清晰的光子专属可比案例是 Marvell 收购 Polariton Technologies;后者是一家瑞士公司,开发基于等离激元的硅光子器件。交易条款未披露。更广的半导体并购历史里,Intel 约 $2 billion 收购 Habana Labs(2019 年),Qualcomm 以 $1.4 billion 收购 NUVIA(2021 年)。两笔交易都奖励了在商业交付前押注架构期权价值的买方,但成交时两家公司都不是流片前状态。 公开市场基准给分析区间的上沿定锚。NVIDIA 2027 财年 Q1 收入 $81.6 billion,同比增 85%,其中数据中心计算收入 $60.4 billion。按 2026 年 5 月约 $3.2 trillion 市值(由回购和 EPS 推算),NVIDIA 约按年化收入 9–10× 交易;这个倍数对应的是量产产品、主导市场地位和 75% 毛利率。即便对 NVIDIA 前瞻倍数施加 70–80% 的高幅度收入前折价,流向一个产品前竞争者的隐含可承受倍数也只是峰值预测收入约 2–3×。Grand View Research 对 AI 加速器市场的预测(2033 年 $256.8 billion,CAGR 29.3%)以及 Precedence Research 对更广 AI 芯片市场的估计(2026 年 $121.7 billion,2035 年增至 $1.1 trillion),确认了总需求扩张,但不能降低单个供应商拿下份额的风险。 Lightmatter 是最接近的光子计算可比公司;截至本次分析时,它尚未确定公开披露最新轮估值。其 Guide 产品是面向 AI 的超大规模光子光引擎,产品成熟度高于 Olix 的 OTPU,因此 Lightmatter 很可能享有更高倍数。CNBC 2026 年 5 月的光子文章引用了一位分析师的观点:大规模部署光子技术可能「还需要 NVIDIA 再迭代一到两代产品」。这句话直接框住了 Olix 可服务市场的时间线风险。 [CV011, CV012, CV013, CV014, CV015, CV016]
| 公司 | 轮次 / 事件 | 估值(投后或 EV) | 阶段 / 收入状态 | 与 Olix 的相关性 | 局限 |
|---|---|---|---|---|---|
| Groq | $750M Series D(2025 年 9 月) | $6.9B | LPU 已出货;200 万+ 开发者;披露收入 | 直接的推理加速器同业;最成熟的电气替代方案 | 已产生收入;牵引力支撑更高隐含倍数;电气方案,不是光子方案 |
| Ayar Labs | $500M Series E(2026 年 3 月) | $3.75B | 共封装光学 chiplet;NVIDIA NVLink 伙伴;量产 | 光子计算相邻领域;共享供应链风险;获得 NVIDIA 背书 | CPO 供应商,不是推理加速器;商业模式不同 |
| Positron AI | $230M Series B(2026 年 2 月) | >$1B | Atlas 已出货(基于 FPGA 的推理);定制芯片 2026 年 tape-out | 阶段和轮次规模最接近;估值相同;今天已有出货产品 | 基于 FPGA,不是光子方案;美国公司;竞争护城河不同 |
| Cerebras Systems | $1.0B Series H(2026 年 Q1) | 未披露;2026 年准备 IPO | 晶圆级 SRAM 芯片;已出货;云服务;IPO 待定 | SRAM 中心架构重叠;IPO 估值将设定公开市场基准 | 运营历史更长;已证明收入;技术路径不同 |
| MatX | $500M Series B(2026 年 Q1) | 未披露 | SRAM 优先 LLM 芯片;量产前;Mountain View CA | SRAM 优先架构与 Olix 的内存论点一致 | 没有光子层;美国公司;机构背书质量不同 |
| NVIDIA(公开市场) | 持续 | 约 $3.2T 市值(2026 年 5 月) | 主导 AI 芯片供应商;FY27 Q1 收入 $81.6B(同比 +85%);毛利率 75% | 为 AI 芯片硬件设定公开市场倍数上限(约 10-14× 远期收入) | 已产生收入的在位巨头;执行风险或技术成熟度不可比 |
轮次数据来自 SemiEngineering 2026 年 Q1 Startup Funding 报告和新闻稿。NVIDIA 市值是根据 FY27 Q1 股票回购数据和流通股数估算,并非官方数字。Cerebras 的 IPO 前估值未披露。所有可比轮次发生在 2025–2026 年。Lightmatter 融资轮次细节在运行日期未获公开确认。
[CV011, CV012, CV013, CV015, CV016, CV017]| 论点 | 类型 | 什么会改变这一判断 |
|---|---|---|
| 到 2033 年推理市场 CAGR 约 29%,给一流硬件留下了很大的可寻址机会 | 正论点(市场) | 模型效率提升压低算力需求,市场增长停滞 |
| SRAM + 光子架构绕过 HBM 内存墙和供应链约束;如果交付落地,就有结构性成本和延迟优势 | 正论点(产品) | OTPU 良率或集成测试显示,在系统层面,光子计算相对下一代电气系统优势不足 |
| Olix 尚未产生收入,没有披露设计导入,没有 tapeout 证据,也没有客户承诺;全部价值都是期权价值 | 反论点(执行) | 公司在 2026 年披露已确认的晶圆厂 tapeout、LOI 或首笔客户收入 |
| Groq(已出货,$6.9B)和 Positron AI(已出货,$1B+)在相同或更高估值下已有真实产品优势 | 反论点(竞争) | Olix 展示更强的基准结果,迫使竞争对手暂停既有客户谈判 |
| NVIDIA 的 $2B 光子投资计划可能原生集成 CPO,压缩 Olix 的可寻址市场 | 反论点(战略风险) | NVIDIA 确认 CPO 集成还要数年,或会基于供应链光学方案,而不是面向推理优化的设计 |
| 芯片开发吃资本,收入前可能还要 2-3 轮稀释性融资,侵蚀 Series A 投资人回报 | 反论点(资本) | Olix 从英国政府半导体计划获得非稀释性补助资金,或宣布晶圆厂成本分担协议 |
论点综合公开来源和跨章节证据;所有反论点都对应已确认的证据缺口,而不是猜测。
[CV007, CV012, CV015, CV017, CV022, CV030]以当前 $1B A 轮标记为锚,测算三种收入情景和两种退出倍数假设下的隐含投后估值。情景覆盖 2028–2030 年可能收入结果。
收入数字是示意性情景估算;公开渠道没有 Olix 收入数据。EV/Revenue 倍数来自 NVIDIA(10–14×)以及未产生收入的深科技硬件折价(50–70%)。并购溢价参考半导体收购先例估算(Intel-Habana、Qualcomm-NUVIA)。所有数字单位为百万美元。
[CV014, CV015, CV023, CV024, CV025]8.3 情景分析、下行风险与资本强度
牛市情景要求 Olix 在 2027 年前把 DX-1/OTPU 机架系统交付给付费的超大规模云厂商或云客户,完成足够规模的首批量产以证明每兆瓦吞吐量具备竞争力,并且在 NVIDIA 的原生光子后继芯片进入规模量产前,从一级买方拿到两个或更多设计导入合同。在这个情景下,并购买方或 IPO 前二级市场可能给 Olix $5–8 billion 估值,相当于 Series A 投资者获得 5–8× 回报。历史上还没有一家光子 AI 硬件公司在四年内从收入前阶段走到数十亿美元并购退出;这个情景依赖执行、时点和买方胃口同时到位。 基准情景假设 2027 年出货目标延迟 12–18 个月,Olix 至少还需再融资 $400–600 million,用于流片、认证和商业化,并使 Series A 投资者持股被稀释至股权结构表的约 12–16%。首批收入可能在 2028 年出现,硬件出货规模虽小但有意义,约 $10–50 million。该情景下后续估值 $2–3 billion,意味着 Series A 投资者在优先权和稀释前获得 2–3× 回报;对一笔高风险深科技押注来说,这个结果偏弱。 熊市情景里,OTPU 流片遇到良率或集成失败,首颗硅交付被推迟到 2029 年或更晚。到那时,NVIDIA 的原生光子或基于 CXL 的产品可能已经拿下效率市场,Groq 和 Positron AI 也可能锁定大部分超大规模云厂商推理合同,Olix 可能被迫降价融资或寻求困境出售。单代半导体产品的资本强度——流片、封装、认证、系统集成——通常超过 $200–500 million,这意味着 Olix 迄今融资的 $250 million 不太可能在没有稀释性后续资本的情况下支撑完整量产爬坡。Companies House 记录的 GBP 287.8572 已配发股本确认,股权仍在分批发行,符合里程碑挂钩拨款的特征。 NVIDIA 自己的光子路线图构成结构性反向风险:CNBC 2026 年 5 月报道称,NVIDIA 已在共封装光学上投入约 $2 billion,覆盖 Lumentum、Coherent、Marvell、Corning;Jensen Huang 也在 GTC 确认,NVIDIA 正开始扩大硅光子技术规模。若 NVIDIA 供应商生态(Lumentum、Coherent、Marvell、Corning)成功把 CPO 大规模集成,独立光学加速器厂商的可服务市场就会从广阔的超大规模云厂商数据中心市场,收缩到专业化的推理即服务应用。CNBC 文章引用分析师预测,光子成为主流还需一到两代产品——也就是说,Olix 不仅要跑赢 Groq 和 Positron,还要跑赢 NVIDIA 光子集成路线图内含的倒计时。 [CV023, CV024, CV025, CV026, CV027, CV028]
| 情景 | 关键假设 | 估值 / 回报逻辑 | 关键风险 | 概率信号 |
|---|---|---|---|---|
| 牛市 | DX-1 在 2027 年上半年出货;≥2 家一线 hyperscaler 签署 LOI;后续融资估值 $3–5B;2030 年前以 $6–8B 并购退出 | Series A 以 $1B 进入可获 5–8× 回报;收购方按 $400-600M ARR run-rate 的预测收入支付 15–20× | 硅光良率执行;hyperscaler 采购延迟;NVIDIA 在交易前完成替代 | 低(需要完美执行 + 温和竞争环境) |
| 基准 | 首笔收入延迟 12–18 个月至 2028 年;一轮稀释性过桥融资($400–600M),估值持平或小幅上台阶;首笔收入 $10–50M | Series A 毛回报 2–3×;扣除清算优先权和稀释后,净回报 1–2× | 跑道压力;人才留存;若 AI 投资周期降温,资本市场风险上升 | 中(符合深科技硬件的基准概率) |
| 熊市 | 2029 年后才交付;down-round 或困境出售;NVIDIA CPO 集成大幅压缩可寻址市场 | Series A 回报 0–0.5×;如果清算优先权排在普通股之前,可能全损 | 资本完全减值;管理层离职;战略并购价格象征性 | 低-中(高于典型 SaaS,但符合纯硬件、收入前押注) |
情景仅作示意;公开域没有收入、积压订单或设计导入数据,无法用量化方式锚定概率估计。
[CV023, CV024, CV025, CV026, CV027]| 触发器 | 阈值 | 对投资论点的传导 | 行动含义 |
|---|---|---|---|
| DX-1/OTPU tape-out 失败或无限期延迟 | 2026 年底前没有确认晶圆厂 tapeout | 核心产品前提坍塌;剩余资本只能支撑困境退出 | 立即回避;监测 down-round 信号或领导层离职 |
| NVIDIA 加速 hyperscaler CPO 集成 | NVIDIA 或主要 hyperscaler 宣布原生封装上光子方案在 2026 年下半年出货 | Olix 的可寻址市场急剧收窄到边缘或专业推理细分 | 重新评估可寻址机会规模;不自动杀死论点,但会实质性压低上行空间 |
| 领投方下调估值,或过桥轮估值持平 / 下调 | 2026-2027 年公开报道或推断出现持平 / 下调估值的过桥轮 | 市场信号显示内部人不再支撑 $1B+ 标记;清算优先权压力上升 | 转为回避;优先权悬置让普通股接近无价值 |
| 竞争对手拿下推理工作负载的锚定 hyperscaler 合同 | Groq、Positron 或基于 NVIDIA Rubin GPU 的系统与目标 hyperscaler 签署独家或优先供货协议 | DX-1 出货前,Olix 就失去滩头客户 | 重新评估产品差异化主张;可能需要转向其他客户细分 |
| 创始人 CEO 离职或领导层出现重大动荡 | James Dacombe 离开 Olix CEO 职位,或活跃投入降至 50% 时间以下 | 公司失去创始人驱动的技术能力和投资人关系能力,而这正是本轮融资的理由 | 治理风险高;任何资金部署前,先索取董事会构成和继任证据 |
触发器按事件监测,并在季度证据复核时更新。阈值为近似值;具体情境重要。
[CV025, CV028, CV029, CV031, CV032]相对 $1B 入场标记,列出 Olix 在各情景下的低—高估值区间。熊市情景覆盖归零至名义价值($0–200M);基准情景为 $500M–$2B;牛市情景为 $3B–$8B。
所有数字单位为百万美元。区间较宽,用来反映证据缺口;目前没有具约束力的财务预测。熊市低点 $0M 对应本金全损情景。
[CV023, CV024, CV025, CV026, CV027]8.4 附录
免责声明
本报告仅供参考,不构成投资建议,也不构成买卖任何证券的要约。报告依赖截至 2026-06-22 可获得的公开信息;在缺少公司直接披露的地方,采用情景分析。
证据索引
| 编号 | 陈述 | 可信度 | 来源 |
|---|---|---|---|
| CO001 | The operating brand is OLIX, while the legal entity is OLIX COMPUTING LIMITED with company number 15585099. | 高 | SO001, SO007 |
| CO002 | Public financing coverage consistently describes Olix as a London-based semiconductor startup. | 中 | SO011, SO014, SO015, SO016 |
| CO003 | The company was incorporated on 22 March 2024 and changed its name from Flux Corp Ltd to OLIX COMPUTING LIMITED on 19 January 2026. | 高 | SO007, SO008 |
| CO004 | Olix recruits across London, Bristol, Austin, San Francisco, and Toronto even though its registered office is in St Albans. | 高 | SO003, SO007, SO015 |
| CO005 | Olix publicly frames its mission as building infrastructure for frontier AI and solving the deployment bottlenecks of inference. | 高 | SO001, SO002 |
| CO006 | Official materials refer to both an Optical Tensor Processing Unit and the OLIX Decode Accelerator 1 as parts of the company's public product story. | 中 | SO002, SO012, SO013 |
| CO007 | James Robert Dacombe is an active director of OLIX COMPUTING LIMITED and has served since 22 March 2024. | 高 | SO009, SO010 |
| CO008 | Press coverage identifies James Dacombe as the founder and chief executive driving Olix's strategy. | 中 | SO013, SO016, SO017 |
| CO009 | The active director slate visible in public records includes Jonathan Heiliger, Taavet Hinrikus, Saul Klein, Firat Ileri, Nicholas McKeown, and James Dacombe. | 中 | SO009 |
| CO010 | Companies House filings show a board reshuffle on 10 February 2026 in which Firat Ileri was appointed and David Helgason plus Krishna Visvanathan were terminated as directors. | 中 | SO008 |
| CO011 | Olix's visible governance network links the company to Hummingbird, Plural/Phoenix Court, Vertex-linked Jonathan Heiliger, and Stanford professor Nicholas McKeown. | 高 | SO009, SO020, SO021, SO022 |
| CO012 | Olix's official culture materials describe an in-person, high-intensity operating model with a £24,000 annual living-local bonus and an 18-week paid exit for mismatched hires at three months. | 高 | SO003, SO004 |
| CO013 | Cooley reported on 11 February 2026 that Olix raised $220 million and was valued at just more than $1 billion in a round led by Hummingbird Ventures. | 中 | SO011 |
| CO014 | Multiple independent press outlets corroborate Olix's February 2026 financing as a $220 million unicorn-making round. | 中 | SO012, SO013, SO014, SO016, SO017 |
| CO015 | Public press coverage places Olix's total funding at roughly $250 million including pre-Series A backing. | 中 | SO013, SO017 |
| CO016 | Plural, Vertex Ventures, LocalGlobe, and Entrepreneurs First are repeatedly named as pre-existing investors in Olix. | 中 | SO012, SO013, SO017 |
| CO017 | Hummingbird's lead-round role is reinforced by Firat Ileri's public appointment as a director one day before the financing announcement. | 高 | SO008, SO011 |
| CO018 | Olix remains a private company with no public evidence of listed securities, public debt issuance, or disclosed revenue metrics. | 高 | SO007, SO008, SO011 |
| CO019 | Public reporting and company materials point to 2027 as the target year for first customer product shipments. | 中 | SO012, SO016, SO002 |
| CO020 | Independent coverage says Olix had grown to more than 70 employees and intended to scale past 200 in 2026. | 低 | SO013, SO017 |
| CO021 | The latest Companies House accounts available publicly are made up to 31 August 2025, making statutory operating data stale relative to the 2026 financing narrative. | 高 | SO007, SO008 |
| CO022 | Public records disclose capital filings and board changes around the round but do not expose a usable cap table or ownership split. | 中 | SO008 |
| CO023 | Olix has not publicly disclosed revenue, ARR, or named production customers as of the run date. | 高 | SO001, SO007, SO011 |
| CO024 | The current hiring map spans optics, CPO systems, ASIC-adjacent integration, distributed inference software, manufacturing operations, and legal/commercial support. | 高 | SO003, SO005, SO006 |
| CO025 | The first visible milestone in Olix's corporate history is incorporation as Flux Corp Ltd in March 2024. | 中 | SO007 |
| CO026 | Jonathan Heiliger joined the company as a director on 24 June 2025. | 中 | SO009 |
| CO027 | Taavet Hinrikus joined the company as a director on 13 October 2025. | 中 | SO009 |
| CO028 | The registered office moved from Oxford to St Albans on 5 November 2025. | 中 | SO008 |
| CO029 | The legal rebrand from Flux Corp Ltd to OLIX COMPUTING LIMITED was filed on 19 January 2026. | 中 | SO008 |
| CO030 | January through March 2026 filings included a new memorandum and articles, pre-emption-right changes, share-right changes, and multiple statements of capital. | 中 | SO008 |
| CO031 | Post-round share allotment filings continued into March and May 2026, indicating financing mechanics extended beyond the headline announcement date. | 中 | SO008 |
| CO032 | Olix explicitly argues that a startup cannot compete on incumbent terms for HBM, advanced packaging, and other supply-constrained technologies. | 中 | SO002 |
| CO033 | Siemens' 2026 co-packaged optics analysis says thermal management, manufacturing test, yield assurance, and evolving standards remain important obstacles to scaling optical packaging. | 中 | SO023 |
| CO034 | Supply-chain reporting summarized by Edgen says meaningful CPO mass production is unlikely until after 2026 because yields remain too low. | 中 | SO025 |
| CO035 | CNBC's May 2026 photonics coverage treats light-based AI hardware as promising but still early relative to incumbent compute ecosystems. | 中 | SO024 |
| CO036 | Olix's official job pages show that US export-control restrictions are a real compliance constraint for at least part of the hiring pipeline. | 高 | SO005, SO006 |
| CO037 | Companies House still classifies the business under SIC 62090, “Other information technology service activities,” rather than a semiconductor-specific category. | 中 | SO007 |
| CO038 | Because James Dacombe is simultaneously the founder, public CEO figure, and main external face of the company, key-person dependency is high. | 中 | SO008, SO013, SO017 |
| CO039 | Nicholas McKeown's public Stanford profile centers on high-performance Internet switches and routers, reinforcing the networking depth visible in Olix's governance slate. | 中 | SO022 |
| CO040 | Jonathan Heiliger's public bio says he previously led Facebook's infrastructure as the user population grew to nearly one billion. | 中 | SO021 |
| CO041 | Saul Klein's public biography describes him as a long-time European entrepreneur and investor with prior senior roles at Index Ventures, Skype, and Lovefilm. | 中 | SO020 |
| CO042 | A regional Bristol news report described the financing as roughly £160 million and the valuation as £734 million, creating a currency-translated discrepancy versus the $220 million and $1 billion-plus framing used elsewhere. | 低 | SO015 |
| CM001 | IEA projects global data centre electricity demand will more than double to approximately 945 TWh by 2030, with AI-optimised facilities being the primary growth driver. | 高 | SM001, SM002 |
| CM002 | IEA projects electricity demand from AI-optimised data centres will more than quadruple by 2030, a rate materially higher than the overall data centre sector. | 高 | SM002, SM001 |
| CM003 | In the United States, power consumption by data centres is on course to account for almost half of US electricity demand growth between now and 2030. | 中 | SM002 |
| CM004 | Data Center Frontier citing IEA data indicates the AI industry is expected to consume at least ten times its 2023 electricity demand by 2026. | 中 | SM003 |
| CM005 | NVIDIA reported Q1 FY2026 total revenue of $44.1 billion (up 69% YoY) with data centre revenue of $39.1 billion (up 73% YoY), for the quarter ended April 27, 2025. | 高 | SM007, SM008 |
| CM006 | NVIDIA reported record Q3 FY2026 total revenue of $57.0 billion and data centre revenue of $51.2 billion (up 25% QoQ and 66% YoY), for the quarter ended October 26, 2025. | 高 | SM007, SM008 |
| CM007 | NVIDIA CEO Jensen Huang stated that AI inference token generation surged tenfold in just one year as of the Q1 FY2026 reporting period. | 中 | SM007 |
| CM008 | NVIDIA's Blackwell NVL72 AI supercomputer was in full-scale production and described by NVIDIA as a machine designed for reasoning inference workloads as of Q1 FY2026. | 中 | SM007 |
| CM009 | MarketsandMarkets sizes the global silicon photonics market at USD 2.65 billion in 2025 growing to USD 9.65 billion by 2030 at a CAGR of 29.5%. | 中 | SM004, SM010 |
| CM010 | MarketsandMarkets projects the North America silicon photonics market at USD 1.16 billion in 2025 growing to USD 4.35 billion by 2030 at a CAGR of 30.2%. | 中 | SM004 |
| CM011 | The Research Insights estimates the global silicon photonics market will cross USD 8.13 billion by 2030, providing an alternative lower-bound estimate to MarketsandMarkets. | 低 | SM012 |
| CM012 | IDTechEx has published a 2024 market report on photonic ICs for AI and computing, covering photonic engines, AI accelerators, and co-packaged optics; full forecast values are paywalled. | 低 | SM013 |
| CM013 | Grand View Research sizes the global AI accelerator market at USD 25.56 billion in 2024 and projects growth to USD 256.84 billion by 2033 at a 29.3% CAGR. | 中 | SM005 |
| CM014 | Key drivers for AI accelerator market growth cited by multiple analysts include surging demand for high-performance computing in AI applications and increasing AI deployment in everyday technology. | 中 | SM005, SM006 |
| CM015 | Co-packaged optics enables bandwidth densities and energy efficiencies that copper-based interconnects can no longer deliver at AI-cluster scale, according to Siemens CPO analysis. | 中 | SM020 |
| CM016 | Silicon photonics manufacturing for CPO still exhibits higher variability than mature CMOS processes; yield excursions can have outsized cost impact due to limited reworkability of co-packaged assemblies. | 中 | SM020 |
| CM017 | The CPO ecosystem's standards are still evolving; the Optical Internetworking Forum (OIF) is actively working on modulation formats and performance benchmarks for CPO interoperability. | 中 | SM020 |
| CM018 | Test scalability is a primary bottleneck for CPO adoption, as optical and electrical domains impose fundamentally different constraints with micron-scale alignment sensitivity. | 中 | SM020 |
| CM019 | CPO adoption depends on proving robust multi-vendor business models and clear advantages in cost, power, and scalability at the system level beyond individual performance demos. | 中 | SM020 |
| CM020 | Meta's Q1 2026 capital expenditure was $19.8 billion; the company raised full-year 2026 capex guidance to $125–145 billion citing higher component pricing and additional data centre capacity. | 高 | SM014, SM019 |
| CM021 | Alphabet's Q1 2026 capital expenditure was approximately $35.7 billion; full-year 2026 capex guidance is $180–190 billion, with investments focused on data centres, AI compute infrastructure, and TPU chips. | 中 | SM015 |
| CM022 | Microsoft's FY2026 total capital expenditure is expected at approximately $190 billion, up 61% from 2025, with Q3 FY2026 capex of $31.9 billion. | 中 | SM019 |
| CM023 | Capital intensity for AI infrastructure has converged toward $150–200 billion of annual deployment per hyperscaler, funded through operating cash flow and long-duration debt. | 中 | SM019 |
| CM024 | Alphabet's Google Cloud revenue surged 63% to $20 billion in Q1 2026; the cloud backlog nearly doubled to approximately $460 billion, mostly driven by demand for AI models and services. | 中 | SM015 |
| CM025 | Broadcom CEO Hock Tan stated that silicon photonics won't matter in the datacenter "anytime soon," indicating two prior innovation waves must occur first. | 中 | SM016 |
| CM026 | Broadcom CEO identified two required innovation waves before silicon photonics becomes necessary: first, scaling copper-based interconnects for rack-scale systems; second, maturing pluggable optics. | 中 | SM016 |
| CM027 | DARPA's PICASSO solicitation identified two fundamental constraints limiting photonic circuit scaling: signal degradation (optical attenuation and noise that cannot be amplified out) and spurious wave interference. | 中 | SM017 |
| CM028 | DARPA launched the PICASSO programme in 2026 with approximately $35 million to spur researchers to overcome fundamental physical constraints and build larger-scale photonic circuits for AI and other workloads. | 中 | SM017 |
| CM029 | Current generation photonic circuits are limited in depth and restricted to single linear mathematical operations; when interfacing with electronic systems, the latency advantage is eliminated by the ~10^6 performance degradation at the opto-electronic conversion step. | 中 | SM017 |
| CM030 | CPO mass production faces delays past 2026 due to yield issues and manufacturing scalability challenges at silicon photonics OSATs and contract manufacturers. | 中 | SM021 |
| CM031 | Data Center Frontier cited IEA data indicating the AI industry is expected to consume ten times its 2023 electricity demand by 2026; US data centres expected to use ~260 TWh by 2026 (up from ~200 TWh in 2022). | 中 | SM003 |
| CM032 | NVIDIA shipped AI servers consuming approximately 7.3 TWh annually as of the 2024 production run, according to IEA data cited by Data Center Frontier. | 中 | SM003 |
| CM033 | The US semiconductor industry commands just over 50% of global chip revenues as of 2025, according to the Semiconductor Industry Association. | 中 | SM018 |
| CM034 | AI data centre chips can contain hundreds of billions of transistors per chip, according to the SIA 2025 State of the Semiconductor Industry report. | 中 | SM018 |
| CM035 | Olix's primary SAM is the AI inference accelerator sub-segment of the AI chip market, with its DX-1 Decode Accelerator targeting inference and decode-heavy workloads; training silicon and edge/mobile inference are excluded. | 中 | SM007, SM005 |
| CM036 | Olix's technology thesis, as communicated through public job postings and Olix's blog, is that photonic rack-scale architecture can break the throughput-interactivity tradeoff in current GPU-based inference hardware. | 中 | SM009 |
| CM037 | Photonic computing for AI inference offers theoretical advantages in bandwidth, latency, and energy efficiency over electronic alternatives; whether these advantages translate at system level is the open engineering question. | 中 | SM017, SM020 |
| CM038 | NVIDIA Chief Scientist Bill Dally stated "optical connectivity will be important to scale accelerated computing clusters to meet the fast-growing demands of AI and HPC workloads," signalling hyperscaler-ecosystem validation of photonic interconnects. | 中 | SM009 |
| CM039 | Ayar Labs builds co-packaged optics and optical I/O technology for silicon photonics-based AI architectures, with partners including NVIDIA and Wiwynn; it represents a near-competitor and supply chain counterpart to Olix's optical I/O approach. | 中 | SM009 |
| CM040 | The BusinessWire ResearchAndMarkets report profiles 183 companies in the silicon photonics and photonic IC ecosystem across 2025–2035, indicating a broad but fragmented supply chain. | 低 | SM011 |
| CP001 | Olix's competitive field can be segmented into four tiers: direct photonic compute peers, optical interconnect and CPO suppliers, electrical inference ASICs, and hyperscaler captive silicon. | 中 | SP026, SP013 |
| CP002 | Lightmatter describes itself as a "Photonic Supercomputer Company" building AI infrastructure and has developed three product lines: Passage (interconnect), Envise (compute), and Idiom (ML compiler). | 高 | SP001, SP004, SP005 |
| CP003 | Groq was founded in 2016 by Jonathan Ross, a former Google engineer who led the development of the first Google TPU. | 中 | SP011 |
| CP004 | Groq raised $750 million in September 2025 at a valuation of $6.9 billion, led by Disruptive Investments with participation from BlackRock, Samsung, Cisco, D1 Capital, and Altimeter Capital. | 中 | SP008, SP011 |
| CP005 | The AI accelerator market has no dominant design, with multiple technology approaches competing for data-centre inference workloads. | 中 | SP013 |
| CP006 | Lightmatter's Passage photonic interconnect chiplet supports 56 to 448 Gbps per lane across 1 to 16 or more wavelengths and uses 3D photonic integration marketed as "Edgeless I/O." | 中 | SP003 |
| CP007 | Lightmatter's Passage chiplets use 3D photonic integration, enabling direct chip-to-chip optical connectivity at the die level. | 中 | SP003 |
| CP008 | Lightmatter has established foundry partnerships with TSMC, GlobalFoundries, and Tower Semiconductor for photonic integrated circuit manufacturing. | 中 | SP002 |
| CP009 | Lightmatter's Idiom framework is a graph compiler designed to partition and schedule ML workloads across multi-blade Envise configurations, with support for PyTorch, TensorFlow, and ONNX. | 中 | SP004 |
| CP010 | Lightmatter claims a current interconnect bandwidth capacity of 114 Tbps today with a roadmap to over one petabit per second. | 中 | SP002 |
| CP011 | Ayar Labs describes itself as the leader in co-packaged optics for AI scale-up and has established commercial chiplet products (TeraPHY and SuperNova). | 中 | SP014 |
| CP012 | Ayar Labs' TeraPHY optical I/O chiplet uses UCIe-based interconnect and delivers approximately ten times the bandwidth of copper interconnects at equivalent power, manufactured on standard CMOS processes. | 中 | SP015 |
| CP013 | Ayar Labs raised $500 million in a Series E round in March 2026 at a $3.75 billion valuation, bringing total capital raised to $870 million; investors include NVIDIA, AMD, Alchip, and MediaTek. | 中 | SP017 |
| CP014 | Ayar Labs joined the NVIDIA NVLink Fusion ecosystem in June 2026, enabling optical I/O integration with NVIDIA NVLink-based AI infrastructure. | 中 | SP018, SP017 |
| CP015 | Intel's Silicon Photonics Product Division has shipped more than eight million photonic integrated circuits (PICs) with over 32 million on-chip integrated lasers embedded in pluggable optical transceivers for data centre networking since 2016. | 高 | SP025, SP012 |
| CP016 | Intel's Optical Compute Interconnect (OCI) chiplet delivers four terabits per second bidirectional and is designed for co-packaging with CPUs, GPUs, IPUs, and other SoCs. | 中 | SP025 |
| CP017 | Intel describes itself as a pioneer in silicon photonics, having invested in the technology for over twenty years, and positions its platform as the volume market leader in silicon photonics by PIC shipments. | 中 | SP025 |
| CP018 | Wikipedia's silicon photonics article identifies Intel as a prominent industry participant alongside Cisco, GlobalFoundries, and other major chipmakers in photonic integrated circuit development. | 中 | SP012 |
| CP019 | Groq's Language Processing Unit (LPU), also known as a Tensor Streaming Processor, is an electrical ASIC designed specifically for low-latency language model inference and is not based on photonic technology. | 中 | SP011 |
| CP020 | Groq's GroqCloud achieves approximately 1,000 tokens per second output throughput for a 20-billion-parameter open-source LLM at $0.075 per million input tokens and $0.30 per million output tokens as of June 2026. | 中 | SP006, SP019 |
| CP021 | Groq raised $750 million in September 2025 from Disruptive Investments, BlackRock, Samsung, Cisco, D1 Capital, and Altimeter Capital, reaching a $6.9 billion valuation and reporting approximately $500 million in 2025 revenue. | 中 | SP008, SP011 |
| CP022 | In December 2025, Groq and NVIDIA entered a non-exclusive inference technology licensing agreement; Groq co-founder Jonathan Ross joined NVIDIA, and Simon Edwards became Groq's CEO. | 中 | SP009, SP011 |
| CP023 | Groq manufactures its LPU chips on Samsung's four-nanometer process node. | 中 | SP011 |
| CP024 | Groq's GroqCloud serves more than two million developers across four global inference regions with an on-premises GroqRack option for enterprise deployments. | 中 | SP010, SP011 |
| CP025 | Etched is building a transformer-specific ASIC called Sohu that hardcodes the transformer architecture and claims more than twenty times the throughput of the NVIDIA H100 for transformer inference workloads. | 中 | SP021, SP022 |
| CP026 | Etched's executive team includes CEO Gavin Uberti (Harvard mathematics), CTO Mark Ross (formerly of Cypress Semiconductor, $9 billion market cap), and VP Brian Loiler (formerly 22 years at NVIDIA). | 中 | SP022 |
| CP027 | AWS Inferentia2 has delivered 4.5 times higher throughput and ninety percent lower inference cost compared to GPU-based EC2 instances in published customer benchmarks. | 高 | SP023, SP013 |
| CP028 | Google Cloud TPU is designed to accelerate neural network training and inference workloads for Google Cloud customers and has been in commercial deployment since 2018. | 高 | SP024, SP013 |
| CP029 | Artificial Analysis benchmarking data shows Groq providing among the fastest available throughput for common open-source LLMs among tested inference providers as of mid-2026. | 中 | SP019 |
| CP030 | Olix's OTPU and DX-1 architecture is described in official materials as SRAM-centric, explicitly avoiding dependence on high-bandwidth memory (HBM) to improve throughput-per-watt and total cost of ownership at inference scale. | 高 | SP027, SP026 |
| CP031 | Olix's stated inference approach targets the memory-bandwidth bottleneck in decode-heavy LLM workloads, which is a different optimisation point from Groq's latency-streaming LPU, NVIDIA's HBM-centric Blackwell, and Lightmatter's photonic interconnect focus. | 中 | SP027, SP011 |
| CP032 | Lightmatter's commercial messaging and product emphasis centres on photonic interconnects (Passage) rather than inference compute (Envise), leaving the SRAM-centric photonic compute layer targeted by Olix's OTPU unoccupied by a commercial Lightmatter product as of mid-2026. | 中 | SP001, SP005 |
| CP033 | Ayar Labs positions itself as a semiconductor component supplier of co-packaged optics technology rather than a system-level AI inference hardware vendor, making it potentially complementary to Olix rather than a direct competitor. | 中 | SP014, SP015 |
| CP034 | IEEE Spectrum's early-2025 AI inference benchmarking found NVIDIA Blackwell leading performance and AMD challenging, with no photonic competitor holding a comparable deployed inference product position. | 中 | SP020 |
| CP035 | Groq's non-exclusive licensing deal with NVIDIA reduces but does not eliminate Groq's commercial independence; Groq continues to operate GroqCloud as a separate commercial service. | 中 | SP009, SP011 |
| CP036 | Switching criteria for hyperscaler inference silicon buyers include cost-per-token, throughput-per-watt, software stack compatibility, and supply-chain qualification lead times of typically twelve to twenty-four months. | 中 | SP013, SP011 |
| CP037 | Olix's publicly stated 2027 first-shipment target creates approximately a twelve-to- eighteen-month window in which Groq, Etched, AWS Inferentia, and Google TPU are the primary inference alternatives to NVIDIA for cloud and enterprise buyers. | 中 | SP027, SP011 |
| CP038 | Intel's volume-proven silicon photonics manufacturing platform (eight million PICs shipped since 2016) demonstrates that photonic integrated circuit manufacturing at commercial scale is achievable, validating the underlying technology category Olix's OTPU depends on. | 高 | SP025, SP012 |
| CP039 | Lightmatter's multi-foundry strategy (TSMC, GlobalFoundries, Tower Semiconductor) suggests that photonic AI chip companies require multiple fab relationships to manage yield risk and supply continuity, which has implications for Olix's manufacturing strategy. | 中 | SP002 |
| CP040 | Artificial Analysis benchmark data shows Groq among the lowest-cost inference providers for common model sizes, creating price-discovery pressure on future entrants and narrowing the cost-per-token window that new architectures must beat. | 中 | SP019, SP006 |
| CP041 | Ayar Labs' NVIDIA NVLink Fusion membership and NVIDIA investor relationship signal that major incumbents are hedging optical interconnect technology bets rather than treating co-packaged optics as an existential threat to their ecosystem. | 中 | SP018, SP014 |
| CI001 | Olix's primary planned revenue stream is hardware sales of the DX-1 rack-scale AI inference accelerator system to hyperscalers and AI cloud providers. | 中 | SI001, SI008 |
| CI002 | Olix's compute manifesto claims the DX-1 delivers superior throughput-per-megawatt and lower total cost of ownership than HBM-based GPU architectures. | 低 | SI008 |
| CI003 | SiliconANGLE reported that Olix expects to start shipping OTPU chips to customers in 2027, based on Financial Times coverage. | 中 | SI001, SI021 |
| CI004 | As of June 2026, Olix has not publicly disclosed any design wins, customer letters of intent, contracted backlog, or revenue from any product. | 中 | SI001, SI004, SI005 |
| CI005 | Olix has not publicly disclosed any list price or realized price for the DX-1 chip or rack system as of June 2026. | 高 | SI001, SI008 |
| CI006 | Hyperscalers including Alphabet and Meta are allocating hundreds of billions of dollars in annual capital expenditure to AI infrastructure, representing the addressable spend pool for Olix's hardware. | 高 | SI023, SI001 |
| CI007 | SiliconANGLE reported that Olix's capital will be used to finance chip development initiatives. | 中 | SI001, SI004 |
| CI008 | If Olix pursues an inference-as-a-service model, it would substantially increase capital intensity through direct infrastructure ownership, although this has not been confirmed. | 低 | SI001, SI008 |
| CI009 | TechSparx reported Olix was advertising senior digital design engineer roles in Bristol at £125,000–£180,000 base salary plus a £24,000 annual living-local supplement, for a total cash package of up to £204,000. | 中 | SI002 |
| CI010 | Olix's careers manifesto confirms the company pays £2,000 per month (~£24,000/year) to employees living within 20 minutes of the office, and offers an 18-week salary exit package at the three-month mark for culture-fit departures. | 高 | SI007, SI002 |
| CI011 | Jon Peddie Research reported in February 2026 that Olix had grown to more than 70 employees and was targeting more than 200 by end of 2026. | 中 | SI003 |
| CI012 | Applying a blended fully-loaded cost of £160,000–£200,000 per employee (base, NI, benefits, local bonus, equity) to 70 employees gives an estimated annualised personnel run-rate of £11–14 million (~$14–17M) as of mid-2026. | 低 | SI002, SI003, SI007 |
| CI013 | Scaling to 200 employees at the same blended cost would increase annualised personnel spend to approximately £32–40 million (~$40–50M/year). | 低 | SI002, SI003, SI007 |
| CI014 | Olix's careers manifesto states it offers benchmark-leading compensation, meaningful equity, and the necessary tools and equipment for engineers to do their best work. | 中 | SI007 |
| CI015 | A photonic-ASIC tape-out at a leading foundry runs approximately $10–50 million in NRE, masks, and test wafers for a production-grade design; this estimate is derived from general semiconductor cost benchmarks and not from any Olix disclosure. | 低 | SI012, SI024, SI025 |
| CI016 | A composite monthly burn estimate for mid-2026 at 70 employees is $2.0–3.0M/month; scaling to 200 employees plus tape-out raises estimated monthly burn to $5–8M/month. | 低 | SI002, SI003, SI007, SI012 |
| CI017 | The Siemens semiconductor-packaging blog documents that silicon photonics manufacturing exhibits higher variability than mature CMOS, and that yield excursions can have outsized cost impact due to limited reworkability of co-packaged assemblies. | 中 | SI012 |
| CI018 | edgen.tech reported that CPO mass production faces delay past 2026 on yield issues, highlighting that packaging complexity pushes delivery timelines and inflates per-unit cost for silicon photonics products. | 中 | SI013 |
| CI019 | Cooley announced on 11 February 2026 that it advised Olix on a $220 million financing round led by Hummingbird Ventures at a valuation just above $1 billion; SiliconANGLE, TFN, and IDCNova independently corroborate the round size and valuation. | 高 | SI001, SI004, SI005, SI021 |
| CI020 | Press coverage places Olix's total lifetime capital at approximately $250 million, implying roughly $30 million raised in earlier rounds from Plural, Vertex Ventures, LocalGlobe, and Entrepreneurs First. | 中 | SI001, SI005, SI006 |
| CI021 | Hummingbird Ventures describes its strategy as investing from $500K at seed to $50M-plus when following on, and providing patient long-term capital through high-risk phases of company development. | 高 | SI019, SI014 |
| CI022 | At an estimated current burn of $2–3M/month, the $220M raise provides approximately 73–110 months of runway; at the 200-employee scale target burn of $5–8M/month it provides approximately 27–44 months. | 低 | SI001, SI003, SI007 |
| CI023 | A single major tape-out cycle of $20–50M would reduce the effective cash base available for operations, compressing runway at 200-staff burn to approximately 20–30 months. | 低 | SI012, SI003 |
| CI024 | Groq raised $750 million at a $6.9 billion valuation in September 2025, demonstrating that large follow-on rounds remain accessible for proven AI inference platforms. | 中 | SI015 |
| CI025 | Ayar Labs closed a $500 million Series E in 2025 for production-ready CPO hardware, with NVIDIA, AMD, and MediaTek as strategic co-investors, demonstrating that large capital is accessible for silicon photonics startups that reach production milestones. | 高 | SI016, SI023 |
| CI026 | Fractile, a UK AI inference chip startup broadly comparable to Olix, raised $220 million in a Series B announced in May 2026, confirming comparable capital is available for UK-based chip startups. | 中 | SI020 |
| CI027 | Companies House shows share capital of approximately GBP 288 across multiple share allotments as of March 2026, reflecting a fractional penny-share structure typical of UK venture-backed companies. | 高 | SI010, SI011 |
| CI028 | Companies House filed a CS01 confirmation statement on 31 March 2026 for the statement date of 21 March 2026, with updates, including capital changes through March 2026. | 高 | SI011, SI010 |
| CI029 | OLIX Computing Limited is registered at Companies House under SIC code 62090 (Other information technology service activities), not under a semiconductor or hardware manufacturing SIC code. | 高 | SI010, SI011 |
| CI030 | Total exemption full accounts filed at Companies House for the period to 31 August 2025 provide a balance sheet but withhold the profit-and-loss account, meaning revenue, operating costs, and net losses are not visible in the public record. | 高 | SI010, SI017 |
| CI031 | Silicon photonics packaging yield excursions are documented to have outsized financial impact due to limited reworkability of co-packaged assemblies; this directly affects Olix's capex exposure per tape-out iteration. | 中 | SI012, SI013 |
| CI032 | UK government guidance confirms that small companies may file abridged or total-exemption accounts that omit the profit-and-loss account entirely from the public Companies House register. | 高 | SI017, SI018 |
| CI033 | The UK national semiconductor strategy published in May 2023 sets a 20-year vision for semiconductor IP, design, and compound semiconductors, but does not commit to direct capital funding for individual chip-design startups. | 高 | SI018, SI017 |
| CI034 | Olix's compensation package including an 18-week salary exit payment for mismatched hires creates a recurring cash liability that is materially higher than standard UK employment exit provisions. | 中 | SI007, SI002 |
| CI035 | No publicly confirmed design win, letter of intent, or pilot customer engagement has been disclosed by Olix as of June 2026, confirming the company is entirely pre-revenue with no visible backlog. | 高 | SI001, SI004, SI005, SI021 |
| CI036 | Nvidia invested $2 billion in photonics companies (Lumentum, Coherent, Marvell, Corning, and Ayar Labs) since March 2026, signalling strong strategic and financial market conviction in the photonics sector. | 高 | SI023, SI016 |
| CI037 | The global silicon photonics market was valued at $2.65 billion in 2025 and is projected to reach $9.65 billion by 2030 at a 29.5% CAGR, reflecting the wider infrastructure market into which Olix is selling. | 中 | SI024 |
| CE001 | The OLIX Decode Accelerator 1 (DX-1) is described by Olix as the first accelerator system architected specifically for decode-phase inference workloads. | 高 | SE001, SE009 |
| CE002 | Olix's design thesis is to scale an SRAM-centric memory architecture integrated with photonics, explicitly avoiding dependence on High Bandwidth Memory (HBM). | 高 | SE001, SE012 |
| CE003 | The DX-1 is built on rack-scale co-design of logic, data movement, packaging, optics, and interconnect as a unified system. | 高 | SE001, SE002 |
| CE004 | Olix argues that the HBM-based GPU architecture imposes a fundamental interactivity-throughput tradeoff because high throughput requires large batch sizes that increase per-user latency. | 中 | SE001, SE008 |
| CE005 | An independent third-party market-watch profile identifies the Optical Tensor Processing Unit (OTPU) as a named compute element within the Olix product architecture, distinct from the DX-1 system designation. | 低 | SE012 |
| CE006 | Olix claims its SRAM+photonics architecture can surpass HBM-based architectures on throughput per megawatt and TCO, and significantly outperform silicon-only SRAM on interactivity and latency. | 低 | SE001 |
| CE007 | Olix is developing optical interconnects spanning chip-to-chip, intra-tray, and rack-to-rack distances, with the Senior Optical Communications Engineer role owning all three tiers. | 高 | SE002, SE001 |
| CE008 | Co-packaged optics (CPO) is the stated integration vehicle for Olix's optical interconnect architecture. | 高 | SE002, SE006 |
| CE009 | Olix's FPGA prototyping platform uses Xilinx/AMD UltraScale+ or Versal devices per the Senior FPGA Engineer job description. | 中 | SE004 |
| CE010 | The FPGA engineer role requires PCIe Gen4 interfaces and multi-lane transceivers operating at greater than 25 Gb/s NRZ, confirming a high-speed digital control layer in the DX-1 architecture. | 中 | SE004 |
| CE011 | Olix's software stack targets vLLM, SGLang, and NVIDIA Dynamo as the three inference serving frameworks to be integrated with DX-1 hardware. | 高 | SE003, SE026 |
| CE012 | Disaggregated prefill and decode is a required architectural capability for the DX-1 software stack, enabling DX-1 to operate as a decode-only node alongside GPU prefill nodes. | 中 | SE003 |
| CE013 | KV-cache management for DX-1 targets NIXL and Mooncake TE protocols for transferring KV cache between prefill and decode nodes. | 中 | SE003, SE026 |
| CE014 | Olix's distributed inference networking stack targets NCCL, RoCE, and InfiniBand for cluster communication, indicating DX-1 is designed for heterogeneous data-centre environments. | 中 | SE003 |
| CE015 | The platform integration job description specifies experience on novel accelerator hardware as a hard requirement, confirming DX-1 requires custom hardware plugin development for vLLM and SGLang. | 中 | SE003 |
| CE016 | vLLM has over 2,000 contributors from academic institutions and companies and supports hardware plugins for NVIDIA, AMD, Google TPU, Intel Gaudi, IBM Spyre, Huawei Ascend, and other accelerators. | 高 | SE010, SE026 |
| CE017 | SGLang achieved 2.7x higher decoding throughput on GB200 NVL72 in a prefill-disaggregated deployment reported in 2025, demonstrating the framework's rapid performance iteration capability. | 中 | SE011 |
| CE018 | Both vLLM and SGLang are built with GPU-first hardware assumptions; Olix must develop and maintain a custom hardware abstraction layer as both frameworks continue to evolve on NVIDIA timelines. | 中 | SE010, SE011, SE003 |
| CE019 | Industry reporting in 2026 indicates CPO mass production was delayed past 2026 due to yield issues, representing a direct read-across risk to Olix's planned product timeline. | 中 | SE007, SE006 |
| CE020 | CPO optical alignment requires micron-scale precision that is fundamentally more demanding than electrical test tolerances, necessitating specialised automated production-scale test methodologies not yet mature at industry scale. | 中 | SE006 |
| CE021 | Photonic integrated circuits are highly temperature-sensitive; dynamic AI workloads induce transient thermal excursions that can cause wavelength drift, laser efficiency degradation, and tightly coupled thermal interactions within CPO packages. | 中 | SE006 |
| CE022 | The OIF (Optical Internetworking Forum) is actively developing CPO interoperability standards including modulation formats and thermal specifications, but full ecosystem standardisation has not been achieved as of 2026. | 中 | SE006 |
| CE023 | Silicon photonics manufacturing exhibits higher process variability than mature CMOS, and CPO assembly has limited reworkability, making yield excursions disproportionately expensive in production. | 中 | SE006 |
| CE024 | CPO test scalability is identified as a primary bottleneck for adoption because optical signals are sensitive to micron-scale deviations and CPO packages from multiple chiplet suppliers introduce alignment variability. | 中 | SE006 |
| CE025 | All engineering job pages reviewed include an explicit US export control notice restricting candidate eligibility based on citizenship or permanent residency in certain countries. | 高 | SE002, SE004 |
| CE026 | The export control restriction list covers Iran, North Korea, Syria, Cuba, Russia, Belarus, China, Hong Kong, Macau, and Venezuela — indicating Olix's technology is likely subject to US Export Administration Regulations. | 高 | SE002, SE004 |
| CE027 | Olix's active hiring locations are London, Bristol (UK), Austin (Texas), Toronto (Ontario), and San Francisco (California). | 中 | SE005 |
| CE028 | Olix's Bristol engineering facility expansion was announced alongside the $220M February 2026 financing round, with the company described as recruiting for technical roles in Bristol. | 中 | SE009, SE027 |
| CE029 | The Bristol and South West digital technology sector has experienced 52% job growth over the last decade, contributing £3.8 billion annually to the regional economy, providing a relevant talent pool for Olix. | 中 | SE027 |
| CE030 | CNBC reported in May 2026 that AI companies are increasingly turning to photonics technology to improve the inefficient transfer of data between AI chips and systems. | 中 | SE008 |
| CE031 | CNBC reported that NVIDIA is investing billions of dollars into companies developing photonics technology, but challenges remain in deploying photonics at scale. | 中 | SE008, SE014 |
| CE032 | A third-party market profile identifies Olix's target customers as hyperscalers, cloud service providers, frontier AI labs such as OpenAI and Anthropic, and enterprises running large LLM inference workloads. | 低 | SE012 |
| CE033 | A third-party market profile claims the Olix OTPU performs matrix multiplications at the speed of light with near-zero heat generation, resolving power-density and memory-bandwidth limitations; this characterisation is not supported by published data and should be treated as an interpretation of company marketing. | 低 | SE012 |
| CE034 | A third-party market profile indicates Olix was scaling its team from approximately 70 to 200+ employees as of early 2026 following the financing round. | 低 | SE012, SE020 |
| CE035 | Olix offers a £24,000 annual Living-Local Bonus for employees living within 20 minutes of the office, a distinctive in-person culture incentive confirmed across multiple job descriptions. | 中 | SE002, SE004 |
| CE036 | The FPGA engineer role requires experience with deterministic networking and time-distribution systems including the White Rabbit precision timing protocol, indicating a deterministic fabric requirement in the DX-1 rack architecture. | 中 | SE004 |
| CE037 | The FPGA engineer role specifies implementation of BER measurement and SECDED error correction (ECC), confirming link-integrity monitoring is a designed-in capability of the DX-1 digital control plane. | 中 | SE004 |
| CE038 | The Siemens EDA blog cites an NVIDIA report showing that transitioning from pluggable transceivers to CPO in 1.6T networks can reduce link power from 30W to 9W, establishing the theoretical energy-efficiency motivation for the approach Olix is pursuing. | 中 | SE006 |
| CE039 | Olix's manifesto identifies KV-cache data movement through HBM as the root cause of a latency floor that limits per-token interactivity regardless of logic efficiency improvements. | 中 | SE001 |
| CE040 | Olix's architecture explicitly does not require quantisation-aware training (QAT) or post-training quantisation (PTQ) of existing models, positioning DX-1 as fully compatible with today's deployed model weights. | 中 | SE001, SE003 |
| CE041 | Olix positions itself against silicon-only SRAM challengers as well as HBM-based GPU architectures, arguing both fail to fully resolve the interactivity-throughput tradeoff due to their respective physical constraints. | 中 | SE001 |
| CE042 | As of June 2026, DX-1 is pre-silicon with FPGA prototyping underway; no silicon tape-out date has been publicly disclosed, and no production silicon, wafer photos, or benchmark results have been published. | 中 | SE004, SE001 |
| CE043 | Olix has not publicly named any silicon photonics foundry, CPO assembly OSAT, or laser source supplier, creating a single undisclosed dependency across the most risk-sensitive manufacturing layers. | 中 | SE001, SE002, SE006 |
| CU001 | Olix's compute manifesto frames the AI infrastructure problem as one that "even the most dominant players, with full-stack control across silicon, software, and supply chains, are unable to solve," implying hyperscalers and large cloud operators are the intended target buyers. | 中 | SU002 |
| CU002 | Olix's platform integration engineer job description explicitly states the role is "the technical authority on how OLIX serves large models as hyperscale AI infrastructure," confirming hyperscalers as primary target buyers. | 中 | SU006 |
| CU003 | Groq's $750 million September 2025 financing at a $6.9 billion valuation confirms inference-as-a-service cloud providers represent a validated, capital-supported buyer segment for dedicated AI inference chips distinct from Tier-1 hyperscalers. | 中 | SU010 |
| CU004 | Olix's compute manifesto explicitly rules out QAT/PTQ model quantisation, pre-selecting for buyers who run frontier-model inference with full or lightly quantised weights — primarily hyperscalers and large inference clouds, not edge or SMB buyers. | 中 | SU002 |
| CU005 | Olix FPGA job postings include export-control language excluding applicants whose most recent citizenship is in Russia, China, Iran, Belarus, North Korea, Syria, Cuba, Hong Kong, Macau, and Venezuela, implicitly constraining the initial customer geography to UK, EU, US, and allied markets. | 中 | SU005 |
| CU006 | No publicly named or confirmed Olix production customers exist as of June 2026; all Olix public sources reviewed contain zero customer names, logos, or case studies. | 高 | SU001, SU002, SU004, SU006 |
| CU007 | Olix has disclosed zero design wins, zero letters of intent, and zero pilot programme agreements in any public filing, press release, or website content reviewed as of June 2026. | 高 | SU001, SU014, SU015 |
| CU008 | SiliconAngle reported in February 2026 that Olix expects to start shipping OTPU chips to customers "next year" (i.e., 2027), treating this as a management target rather than a committed delivery date. | 中 | SU014 |
| CU009 | Groq serves more than two million developers and Fortune 500 companies across North America, Europe, and the Middle East, demonstrating the scale of buyer demand for a dedicated inference cloud service built on proprietary silicon. | 中 | SU010 |
| CU010 | CoreWeave participates in MLPerf v5.0 inference benchmarking, confirming active investment in AI inference hardware validation — making inference- as-a-service providers credible early adopter candidates if Olix achieves MLPerf participation post-silicon. | 中 | SU013, SU017, SU027 |
| CU011 | Broadcom's Q4 FY2025 earnings disclosed more than $50 billion in custom AI accelerator orders from hyperscale customers, confirming the capital intensity of the Tier-1 hyperscaler buyer segment and demonstrating multi-year procurement commitments. | 中 | SU007 |
| CU012 | The inference-as-a-service procurement cycle is materially shorter than Tier-1 hyperscaler silicon qualification programmes, with CoreWeave and Lambda Labs representing a potentially faster first-revenue pathway for Olix than the most capital-intensive buyer tier. | 中 | SU023, SU017, SU010 |
| CU013 | The Olix platform integration engineer job description requires deep expertise in vLLM, SGLang, and NVIDIA Dynamo inference stacks, and full lifecycle experience from architecture through production operations — indicating software integration is a major pre-commercial dependency. | 中 | SU006 |
| CU014 | The vLLM GitHub repository hosts over 2,000 contributors and describes its value as "one engine, endless possibilities — run any model on any hardware," setting hardware-agnostic extensibility as a community norm that Olix must satisfy to achieve upstream acceptance. | 中 | SU016 |
| CU015 | No public vLLM, SGLang, or NVIDIA Dynamo pull requests or commits attributed to Olix engineers have been identified in reviewed public repositories as of June 2026, meaning the software integration required for customer deployment is incomplete. | 中 | SU016, SU006 |
| CU016 | Lambda Labs' cloud offering is priced on GPU instance availability with multi-week to multi-year contract options, indicating that inference cloud buyers can move on accelerator procurement within weeks once a hardware decision is made — far faster than hyperscaler qualification. | 中 | SU023 |
| CU017 | Broadcom CEO Hock Tan stated that silicon photonics will not matter in the datacentre "anytime soon," placing two prior technology waves (copper interconnects, pluggable optics) as necessary predecessors before silicon photonics becomes a hyperscaler procurement imperative. | 中 | SU007 |
| CU018 | DARPA's PICASSO programme, announced in early 2026 with $35 million in funding, specifically identifies that photonic circuits "struggle to show significant system-level performance advantages over electronic systems," signalling that fundamental photonic scaling barriers remain unresolved at the research level. | 中 | SU008 |
| CU019 | No Olix silicon tape-out date has been publicly disclosed as of June 2026, meaning the company has not yet entered the stage where hyperscaler qualification processes can formally begin. | 高 | SU001, SU006 |
| CU020 | Olix has not submitted to MLCommons MLPerf inference benchmarks as of June 2026, which is expected given pre-silicon status, but participation in MLPerf is an industry-standard credibility gate for enterprise and hyperscaler buyers. | 中 | SU013 |
| CU021 | Ayar Labs' June 2026 announcement joining NVIDIA's NVLink Fusion ecosystem demonstrates that hyperscaler customers are now actively evaluating co-packaged optics for rack-scale AI infrastructure, representing the closest public buyer-readiness signal for CPO-based compute products in the same category as Olix's architecture. | 中 | SU021 |
| CU022 | Hyperscaler silicon qualification cycles for novel AI accelerators typically require 12–24 months from initial engagement to production commitment, based on industry reference patterns documented in the Broadcom earnings narrative and analogous custom silicon programmes. | 中 | SU007, SU011, SU012, SU029 |
| CU023 | Broadcom CEO Tan described custom accelerator investment as "a multi-year journey and not transactional or short-term decision," with a billion- dollar order from a new customer and $11 billion from Anthropic cited to demonstrate demand durability — signalling hyperscalers make long- cycle, concentrated bets on silicon suppliers. | 中 | SU007 |
| CU024 | Broadcom's Q4 FY2025 semiconductor AI hardware revenue reached $6.7 billion in a single quarter (up from near-zero three years prior), indicating the magnitude of hyperscaler silicon spend that Olix must compete for — and the difficulty of displacing an established incumbent. | 中 | SU007 |
| CU025 | Groq, Inferentia, and Google TPU all participate in MLPerf inference benchmarking and use public results as buyer-credibility signals; none of these are Olix customers, but they define the customer-proof standard Olix must reach. | 高 | SU013, SU009, SU011, SU012, SU026 |
| CU026 | Olix has zero publicly named or confirmed production customers as of June 2026; this is the baseline from which all customer-proof assessment must start, and it is a confirmed factual position rather than a supposition. | 高 | SU001, SU004, SU014 |
| CU027 | SiliconAngle reported that the optical architecture of Olix's chip is unclear to external observers: "It's unclear what optical components are included in the processor or how they're used," reflecting opacity of the customer-facing technical proposition even among specialist technology journalists. | 中 | SU014 |
| CU028 | AWS Inferentia customer testimonials on the Amazon product page document NTT PC Communications achieving 4.5× throughput, 25% lower latency, and 90% lower cost versus GPU-based EC2 instances, setting the outcome- evidence standard that Olix must eventually match to compete. | 中 | SU011 |
| CU029 | Broadcom CEO Tan explicitly stated that two technology waves (copper-based rack-scale interconnects, pluggable optics) must be exhausted before silicon photonics becomes the necessary procurement option — implying the photonics adoption window may not open until 2028–2030 or later. | 中 | SU007, SU028 |
| CU030 | Groq's published customer case studies name Willow (zero downtime, 500 ms faster responses), PGA of America, and multiple enterprises, confirming production-grade inference deployments with named references — a customer-proof level Olix will not reach until after 2027 at earliest. | 中 | SU009 |
| CU031 | Google Cloud TPU v5 is integrated into Vertex AI and available as a production service, meaning Google as a potential Olix buyer already has its own captive inference silicon, reducing the probability of Google adopting a third-party photonic accelerator. | 中 | SU012 |
| CU032 | Net revenue retention, gross revenue retention, customer churn rate, and all associated retention metrics are structurally unavailable for Olix because no production revenue or customer relationship has been established as of June 2026. | 高 | SU001, SU004 |
| CU033 | No contract length, renewal terms, or repeat purchase patterns for Olix customers exist in the public domain; first customer contract terms will likely be governed by the anchor buyer's standard supply agreement with demand variability clauses. | 低 | SU007, SU011 |
| CU034 | Olix has no disclosed OEM server vendor partnerships (Dell, HPE, Supermicro) that would provide a channel route to enterprise customers outside the direct hyperscaler procurement model. | 高 | SU001, SU004 |
| CU035 | Groq's LPU-based inference cloud demonstrates that a novel architecture can achieve commercial scale within a few years of production deployment, with 2M+ developers by September 2025 post-2020 founding — setting an aspirational benchmark for Olix's adoption trajectory. | 中 | SU009, SU010 |
| CU036 | When Olix wins its first production customer, that account will represent 100% of commercial revenue, creating an extreme concentration risk that persists until a second independent customer relationship is established. | 高 | SU007, SU014 |
| CU037 | Fractile, a comparable UK AI inference chip startup, disclosed a $220M Series B in May 2026 with no named production customers — consistent with the inference-chip sector pattern of capital-heavy pre-revenue development phases in 2026. | 中 | SU019 |
| CU038 | Broadcom's $50 billion+ hyperscaler custom-accelerator backlog represents locked-in incumbent supply relationships; Olix must demonstrate photonics as a necessity — not merely an option — to displace or supplement those relationships. | 中 | SU007 |
| CU039 | Ayar Labs' NVLink Fusion ecosystem entry in June 2026 shows that CPO is advancing toward hyperscaler productisation through system-level integration with NVIDIA — but Ayar Labs positions as a connectivity supplier rather than a compute accelerator, meaning Olix faces a different and potentially longer buyer acceptance process. | 中 | SU021 |
| CU040 | The CPO mass production delay reported by Edgen.Tech — citing yield issues pushing commercial CPO availability past 2026 — compounds the buyer readiness risk for Olix, as buyers may be reluctant to qualify a system whose supply chain has documented yield barriers. | 低 | SU020 |
| CR001 | Siemens EDA's February 2026 CPO analysis identifies test scalability as the primary bottleneck for co-packaged optics adoption, with optical signals requiring micron-scale alignment precision far tighter than electrical test tolerances. | 高 | SR009, SR010 |
| CR002 | Siemens EDA reports that silicon photonics manufacturing exhibits higher process variability than mature CMOS processes, and yield excursions in CPO assemblies have an outsized cost impact due to limited reworkability after co-packaging. | 中 | SR009 |
| CR003 | Siemens EDA identifies thermal management as a core CPO risk: photonic ICs are highly temperature-sensitive, and transient thermal excursions from AI workloads can shift optical wavelengths and degrade performance. | 中 | SR009 |
| CR004 | EdgeN Intelligence reported in 2026 that CPO mass production faces delays beyond 2026 specifically due to yield issues, corroborating the Siemens analysis of photonic manufacturing immaturity. | 中 | SR010 |
| CR005 | The CPO supply-chain and standards ecosystem is still evolving as of 2026, with no fully established interoperable ecosystem; industry consortia such as OIF are working on modulation formats but timelines are uncertain. | 中 | SR009 |
| CR006 | Siemens estimates that by 2027 AI data-centre power demand will grow 50 percent globally, with approximately 60 percent of data-centre energy spent on data movement rather than compute — creating both the demand driver and a thermal challenge for CPO solutions. | 中 | SR009 |
| CR007 | TSMC offers silicon photonics as a specialty technology process; advanced silicon photonics requires a tier-1 foundry, making TSMC the primary single-source dependency for Olix's fabrication. | 高 | SR003, SR019 |
| CR008 | Olix has not publicly disclosed any silicon photonics tapeout result, photonic yield target, or foundry qualification milestone as of June 2026. | 中 | SR012, SR013 |
| CR009 | ASML's EUV lithography systems are fundamental to mass-producing semiconductor chips; ASML holds a near-monopoly on EUV tooling, representing a systemic upstream risk for all advanced foundry nodes including TSMC's silicon photonics process. | 中 | SR004 |
| CR010 | NVIDIA reported in connection with its CPO analysis that transitioning from pluggable transceivers to CPO in 1.6T networks can reduce link power from 30W to 9W; Olix is pursuing this same CPO power advantage for inference rack-scale systems. | 中 | SR009, SR018 |
| CR011 | The US BIS requires a license to export advanced computing integrated circuits to entities headquartered in Country Group D:5 (including China) or Macau, or with ultimate parent companies in those jurisdictions, even when the ship-to location is elsewhere. | 高 | SR001, SR022 |
| CR012 | BIS extended its Authorized IC Designer compliance timeline to December 31, 2026, allowing companies additional time to submit Authorized IC Designer applications; this extension confirms the advanced-computing export-control regime is active and applicable. | 中 | SR001 |
| CR013 | The UK National Security and Investment Act 2021 grants the Secretary of State broad powers to review and block acquisitions in seventeen sensitive sectors, with semiconductors explicitly included. | 高 | SR002, SR022 |
| CR014 | An acquisition of Olix by a foreign entity could be subject to mandatory NSI Act notification and Government review, potentially blocking or delaying a strategic exit; this applies above certain shareholding thresholds. | 中 | SR002, SR022 |
| CR015 | Olix has not disclosed a granted patent portfolio as of June 2026; Lightmatter, a photonic computing peer, holds published patents on photonic matrix multiplication and optical interconnect architecture. | 中 | SR015, SR012 |
| CR016 | The UK government's 2023 pro-innovation AI regulation white paper adopts a light-touch, principles-based approach to AI regulation; no near-term hardware licencing obligations apply to AI chip manufacturers under this framework. | 中 | SR006 |
| CR017 | The EU AI Act, enacted in 2024, imposes obligations on general-purpose AI system providers; hardware suppliers to EU-deployed AI systems may face documentation and compliance obligations as delegated acts are published. | 中 | SR005, SR006 |
| CR018 | No BIS enforcement actions or UK NSI Act blocking orders against photonic chip startups have been publicly disclosed as of June 2026. | 中 | SR001, SR002 |
| CR019 | Olix raised $220 million at a valuation above $1 billion in February 2026; based on a modelled monthly burn of $3–5 million, estimated runway is 30–45 months, extending to approximately late 2028 at the low-burn end. | 中 | SR013, SR014, SR016 |
| CR020 | Semiconductor tape-out and packaging qualification costs are in the range of tens of millions of dollars per iteration for advanced processes; a failed photonic tapeout would require a retry at significant incremental cost within Olix's fixed runway. | 中 | SR019, SR003 |
| CR021 | NVIDIA's data-centre revenue reached $51.2 billion in Q3 FY2026, demonstrating the company's continued dominance in AI compute and its capacity to invest in inference optimisation, narrowing the performance gap Olix must bridge. | 高 | SR015, SR018 |
| CR022 | Lightmatter raised $400 million in funding and is further ahead in photonic compute product development than Olix, representing the nearest direct photonic computing competitive threat as of 2026. | 中 | SR015, SR020 |
| CR023 | CNBC reported in May 2026 that chip companies are racing to address AI bottlenecks in photonics, confirming active competitive pressure from established players alongside Olix. | 中 | SR018 |
| CR024 | A two-year slip in Olix's 2027 first-shipment target would give NVIDIA, Google TPU, and AWS Inferentia additional time to close conventional CMOS inference efficiency gaps, eroding the performance wedge Olix is targeting. | 中 | SR015, SR018, SR020 |
| CR025 | TSMC is the only credible volume foundry for advanced silicon photonics at the required process node; any geopolitical, capacity, or qualification disruption to TSMC access would materially delay Olix's programme with no immediately available alternative. | 中 | SR003, SR019 |
| CR026 | ASML holds a near-monopoly on EUV lithography tools globally; shortages in EUV tool availability affect all advanced foundry nodes simultaneously, making this a systemic and non-mitigable single-source dependency for Olix. | 高 | SR004, SR019 |
| CR027 | Indium phosphide (InP)-based III-V laser supply for silicon photonic light sources is geographically concentrated and subject to capacity constraints; Olix's photonic architecture requires laser light sources that depend on this constrained supply chain. | 中 | SR009, SR008 |
| CR028 | Olix has not disclosed a design-in agreement or evaluation contract with any hyperscaler as of June 2026; absence of a named first customer is a material execution risk for a pre-revenue AI chip startup targeting hyperscaler deployments. | 高 | SR012, SR013 |
| CR029 | The SIA's 2025 state-of-industry report highlights the scale of capital required to compete in advanced chip design; as a UK startup without a strategic corporate partner, Olix faces this capital intensity with limited balance-sheet depth compared to incumbents. | 中 | SR019, SR014 |
| CR030 | James Dacombe is the sole continuously serving Companies House director of OLIX Computing Limited since the company's March 2024 incorporation; no co-founder with semiconductor delivery experience has been publicly disclosed. | 高 | SR011, SR014 |
| CR031 | Olix has not publicly disclosed a CTO appointment or lead silicon architect with prior tape-out delivery experience as of June 2026. | 中 | SR012, SR033 |
| CR032 | Olix's careers pages show open roles across optics, FPGA, ASIC, photonic packaging, manufacturing operations, and process integration — indicating the company is still assembling its core engineering team as of June 2026. | 高 | SR033, SR023 |
| CR033 | Tech Funding News and IDCNova both report that Olix had grown to more than 70 employees at the time of the February 2026 round and aimed to grow to more than 200; this means Olix must nearly triple headcount before first product delivery. | 中 | SR014, SR016 |
| CR034 | The global talent pool for co-packaged optics and silicon photonics engineers is extremely thin; Olix competes for these specialists against TSMC, ASML, Lightmatter, and hyperscaler silicon teams simultaneously. | 中 | SR009, SR033 |
| CR035 | Olix's culture materials describe the company as 'extraordinarily demanding' and include an 18-week paid exit for mismatched hires, signalling the company is aware that talent retention under its operating model is fragile. | 中 | SR023, SR012 |
| CR036 | No succession plan or deputy leadership structure for James Dacombe has been publicly disclosed; a sudden exit of the CEO would leave the company without a designated operational successor. | 中 | SR011, SR013 |
| CR037 | Olix's board includes Jonathan Heiliger (former Facebook infrastructure), Nicholas McKeown (Stanford networking professor), and Taavet Hinrikus (Transferwise co-founder); board oversight does not substitute for an operating management team with chip delivery track record. | 高 | SR013, SR014, SR024 |
| CR038 | Cooley LLP served as Olix's legal counsel for the February 2026 financing; the presence of a top-tier technology law firm confirms basic legal governance but does not establish operational or IP compliance maturity. | 中 | SR024 |
| CR039 | If two major hyperscalers announced a 30-percent or greater reduction in AI infrastructure CapEx, Olix's demand thesis would be substantially impaired; the company's entire commercial plan depends on continued hyperscaler AI spending growth. | 中 | SR013, SR018 |
| CR040 | Olix's $220 million raise at over $1 billion valuation commits the company to a high-bar next round; if photonic manufacturing delays mean no commercial revenue by end of 2027, the next round would likely require significant dilution or come at a lower valuation. | 中 | SR013, SR014, SR032 |
| CR041 | The UK national semiconductor strategy identifies supply-chain resilience as a core objective; Olix, as a UK-headquartered photonic AI chip company, may benefit from DSIT support programmes but also faces policy expectations around UK-based manufacturing. | 中 | SR022, SR006 |
| CR042 | The NIST AI Risk Management Framework (AI RMF) establishes voluntary governance guidelines for AI technology providers; no mandatory hardware compliance regime applies to Olix's DX-1 under NIST standards as of June 2026. | 中 | SR005 |
| CR043 | The risk transmission pathway for Olix runs: CPO yield failure → launch slip → customer adoption delay → revenue miss → capital cliff → valuation reset → talent attrition → further CPO risk, forming a self-reinforcing negative loop. | 中 | SR009, SR010, SR013 |
| CR044 | A BIS enforcement action adding advanced silicon photonics chips to the CCL or restricting exports to key addressable markets would directly shrink Olix's revenue opportunity and require a fundamental revision of its market-entry strategy. | 中 | SR001, SR022 |
| CR045 | Olix's single most critical external dependency is TSMC's silicon photonics process; a geopolitical disruption to Taiwan semiconductor fabrication would be catastrophic for Olix and most advanced AI chip startups simultaneously. | 中 | SR003, SR004, SR019 |
| CR046 | Pitchbook data on silicon photonics startup funding shows Olix's $220M raise is among the largest single rounds for a pre-silicon photonic computing company, raising the performance expectations Olix must meet for the next round. | 中 | SR032, SR013 |
| CV001 | Olix raised $220 million in Series A financing in February 2026 at a post-money valuation above $1 billion, making it one of the fastest UK startups to achieve unicorn status. | 高 | SV011, SV012, SV018 |
| CV002 | Olix's total capital raised is approximately $250 million, comprising the $220 million Series A plus an estimated $25–30 million in seed and pre-seed financing since March 2024 incorporation. | 中 | SV011, SV012 |
| CV003 | Companies House records a statement of capital following an allotment of shares on 5 March 2026 at GBP 287.8572 total nominal capital, confirming the Series A equity issuance closed. | 高 | SV010, SV020 |
| CV004 | The only Olix statutory accounts on public record are total-exemption accounts made up to 31 August 2025, which disclose balance-sheet data only and record no operating revenue or trading income. | 高 | SV010, SV020 |
| CV005 | Olix has not publicly disclosed any design wins, customer letters of intent, foundry tapeout completion, or revenue as of June 2026. | 中 | SV011, SV021, SV022 |
| CV006 | The $220 million primary at a $1 billion post-money valuation implies that new investors received approximately 22 percent of the post-money equity in the financing round. | 中 | SV011, SV012 |
| CV007 | Grand View Research estimates the global AI accelerator market at $25.6 billion in 2024, growing to $256.8 billion by 2033 at a CAGR of 29.3%, providing a large structural market backdrop for Olix's inference thesis. | 中 | SV015 |
| CV008 | Precedence Research estimates the global AI chip market at $94.4 billion in 2025, growing from $121.7 billion in 2026 to approximately $1.1 trillion by 2035 at a CAGR of 27.9%. | 中 | SV016 |
| CV009 | SemiEngineering's Q1 2026 startup funding report counted 80 semiconductor companies raising over $8.4 billion in the first quarter of 2026, with photonics described as a 'popular segment for investment'. | 中 | SV004 |
| CV010 | Cerebras Systems raised $1.0 billion in Series H financing in Q1 2026 and is reportedly preparing for an IPO later in 2026, which will provide the first public market valuation benchmark for a revenue-generating AI accelerator startup. | 中 | SV004 |
| CV011 | Groq raised $750 million in September 2025 at a post-money valuation of $6.9 billion, led by Disruptive Investments with participation from BlackRock, Neuberger Berman, Samsung, and others. | 中 | SV013 |
| CV012 | Ayar Labs closed a $500 million Series E in March 2026 at a $3.75 billion post-money valuation, bringing total raised to approximately $870 million; Ayar's co-packaged optics product is in the NVIDIA NVLink Fusion ecosystem. | 中 | SV014 |
| CV013 | Positron AI raised $230 million in a Series B round in February 2026 at a post-money valuation exceeding $1 billion; its Atlas FPGA-based inference system is already shipping to paying customers including Jump Trading. | 高 | SV007, SV009 |
| CV014 | NVIDIA reported Q1 fiscal year 2027 revenue of $81.6 billion, up 85% year-on-year, with data-centre compute revenue of $60.4 billion; this represents the most recent public market financial benchmark for AI chip hardware. | 高 | SV001, SV023 |
| CV015 | NVIDIA's implied EV/annualised revenue multiple as of May 2026 is approximately 9–14×, reflecting a dominant-market-position premium; applying a 70–80% pre-revenue discount to this multiple implies a 2–3× forward revenue multiple framework for a pre-product peer. | 中 | SV001 |
| CV016 | MatX raised $500 million in a Series B round in Q1 2026, led by Jane Street and Situational Awareness LP; MatX designs a high-throughput SRAM-first LLM chip targeting training and inference, overlapping architecturally with Olix's SRAM-centric approach. | 中 | SV004 |
| CV017 | SambaNova raised $350 million in a Series E in Q1 2026, bringing total capital raised to over $1.5 billion; SambaNova has a shipped fifth-generation RDU inference chip with commercial revenue. | 中 | SV004 |
| CV018 | Marvell Technology acquired Polariton Technologies, a Swiss developer of plasmonics-based silicon photonics devices, in April 2026, demonstrating active M&A appetite for photonic compute IP from incumbent semiconductor companies. | 中 | SV008 |
| CV019 | NVIDIA announced approximately $2 billion in co-packaged optics investments across Lumentum, Coherent, Marvell, and Corning in early 2026, plus participation in Ayar Labs' $500 million Series E. | 中 | SV017, SV014 |
| CV020 | Lightmatter's Guide product is a very-large-scale photonic light engine supporting 16 wavelengths and roadmapped to 64 wavelengths, representing a more advanced photonic product maturity than Olix's pre-tapeout OTPU. | 中 | SV002, SV019 |
| CV021 | CNBC reported in May 2026 that NVIDIA CEO Jensen Huang confirmed at GTC that NVIDIA was beginning to scale its silicon photonics technology, while an analyst quoted in the article stated mainstream photonic deployment may require 'one or two more generations' of NVIDIA products. | 中 | SV017 |
| CV022 | Jon Peddie Research, in an analysis of the AI processor landscape, observed that multiple companies including Neurophos, Lightmatter, OptoML, Rayd, and Volantis are pursuing optical compute and questioned 'whether the problem they solve will be the one people care about by the time they do it'. | 中 | SV022, SV005 |
| CV023 | Under a bull scenario requiring on-time 2027 DX-1 delivery and two or more hyperscaler design wins, an M&A acquirer could value Olix at $5–8 billion by 2029–2030, implying a 5–8× return to Series A investors at the $1 billion entry. | 低 | SV018, SV008 |
| CV024 | Under a base scenario with a 12–18 month delay to 2028 first revenue and a dilutive follow-on raise of $400–600 million, a resulting valuation of $2–3 billion implies a 2–3× gross return to Series A investors before preferences and dilution. | 低 | SV011, SV004 |
| CV025 | Under a bear scenario with OTPU tape-out delays to 2029 or beyond, competitive displacement by Groq or NVIDIA CPO products, and a forced down-round or distressed sale, Series A investors face 0–0.5× return and possible total loss. | 低 | SV017, SV022 |
| CV026 | Capital intensity for a single semiconductor product generation — tapeout, packaging, qualification, and system integration — typically exceeds $200–500 million; Olix's $250 million raised to date is likely insufficient for a full production ramp without additional dilutive capital. | 中 | SV004, SV006 |
| CV027 | Positron AI is a direct stage comparable to Olix: it raised $230 million in February 2026 at the same implied $1 billion valuation milestone, but with a shipped FPGA-based product and a paying tier-1 customer (Jump Trading), providing an adverse benchmark for Olix's pre-product mark. | 中 | SV007, SV009 |
| CV028 | NVIDIA's photonic integration programme, if successful within the Blackwell successor generation, would integrate co-packaged optics at the chip level, potentially removing the addressable market for an independent optical inference accelerator targeting hyperscale data centres. | 中 | SV017, SV019 |
| CV029 | Marvell's acquisition of Polariton Technologies, combined with NVIDIA's $2 billion photonics supplier investments, signals that strategic acquirers view photonic compute IP as valuable; this raises both the M&A exit probability for Olix and the competitive threat from well-capitalised incumbents. | 中 | SV008, SV003 |
| CV030 | Fractile, a UK AI inference chip startup with a similar UK origin and inference focus, raised a $220 million Series B in May 2026, suggesting continued UK investor appetite for deep-tech hardware but also increasing competitive density for inference-focused hardware startups in the UK market. | 中 | SV026 |
| CV031 | Management dependency is a material valuation risk: James Dacombe serves as founder-CEO of both Olix and CoMind (a brain-monitoring startup with $100 million raised), creating a divided attention risk that is uncommon at the same stage among pure-play deep-tech hardware companies. | 中 | SV022, SV011 |
| CV032 | The SemiEngineering Q4 2025 funding report noted 75 companies raised $3 billion in Q4 2025 for AI chips and AI-for-chip-design, confirming Olix's round occurred at the peak of a semiconductor funding cycle that may not sustain its valuation benchmarks if market conditions tighten. | 中 | SV006 |
| CV033 | Olix's SH01 filing records that new shares were allotted on 5 March 2026 at a nominal capital of GBP 287.8572, providing limited but confirmatory evidence that the Series A completed with new share issuance. | 中 | SV010 |
| CV034 | AI Market Watch's company profile for Olix lists Lightmatter, Celestial AI, and Etched as Olix's three main competitors, with key partnerships attributed to Plural, Hummingbird Ventures, Vertex Ventures, LocalGlobe, and Entrepreneurs First. | 低 | SV027 |
| CV035 | The Olix compute manifesto states that 1GW clusters are arriving in 2026 and the race to 10GW is underway, with inference scaling demands for both interactivity and throughput simultaneously — the core market rationale for an SRAM+photonic architecture. | 中 | SV021, SV030 |
| CV036 | No Olix foundry partner, tapeout date, process node, or silicon qualification result has been publicly disclosed as of June 2026, making the 2027 DX-1 shipping target entirely speculative from a public evidence standpoint. | 高 | SV021, SV011 |
| CV037 | The liquidation preference structure, anti-dilution provisions, and participating/non-participating status of Olix's Series A preferred shares are not disclosed in any public document, making it impossible to calculate actual investor return profiles under base and bear scenarios. | 中 | SV010, SV020 |
| CV038 | Lightmatter's post-money valuation from its most recent financing round has not been confirmed in any public document reviewed for this analysis; no reliable comparable multiple can be derived from Lightmatter data alone. | 中 | |
| CV039 | The Hummingbird Ventures portfolio page lists Olix as a portfolio company, confirming the lead investor relationship disclosed in Series A press coverage. | 中 | SV025 |
| CV040 | Celestial AI's domain (celestial.ai) redirects to Marvell Technology's website, consistent with Marvell having absorbed Celestial AI's photonic compute IP as part of its custom silicon and photonics expansion strategy. | 中 | SV003, SV008 |
| CV041 | The UK AI chip market has attracted at least two well-funded peers — Olix ($220M, Feb 2026) and Fractile ($220M Series B, May 2026) — within the same four-month window, indicating both strong investor appetite and increasing competitive density for inference hardware in the UK ecosystem. | 中 | SV026, SV011 |
| CV042 | Olix's valuation stance is assessed as stretched: the $1B+ post-money mark assigns pure option value to an unproven photonic AI architecture with no revenue, no disclosed tapeout, and 12–18 months minimum to first customer delivery, while similarly-funded peers with shipping products receive comparable or lower implied valuations. | 中 | SV007, SV013, SV027 |
| 编号 | 出版方 | 标题 | 引文 |
|---|---|---|---|
| SO001 | OLIX | OLIX homepage | Scaling an SRAM-architecture integrated with photonics can surpass HBM-based architectures on throughput/MW and TCO. |
| SO002 | OLIX | Compute Manifesto | The OLIX Decode Accelerator 1 (DX-1) is the first system architected specifically for decode. |
| SO003 | OLIX | Careers at OLIX | |
| SO004 | OLIX | Our Culture Manifesto | We pay an additional £2,000 per month to those living within 20 minutes. |
| SO005 | OLIX | Senior Optical Communications Engineer job page | |
| SO006 | OLIX | Platform Integration Engineer job page | |
| SO007 | UK Companies House | OLIX COMPUTING LIMITED overview | |
| SO008 | UK Companies House | OLIX COMPUTING LIMITED filing history | |
| SO009 | UK Companies House | OLIX COMPUTING LIMITED officers | |
| SO010 | UK Companies House | James DACOMBE personal appointments | |
| SO011 | Cooley | OLIX Raises $220 Million in Financing | Cooley advised OLIX, a UK semiconductor startup, on its latest $220 million financing round. |
| SO012 | SiliconANGLE | Photonic AI chip startup OLIX nabs $220M investment | |
| SO013 | Jon Peddie Research | An AI processor by any other name | |
| SO014 | The Tech Founders | Olix: 25-Year-Old Founder Raises $220M for UK AI Chip Startup Now Valued at $1 Billion | |
| SO015 | TechSparx | London-based start-up Olix selects Bristol for its new engineering facility | |
| SO016 | IDCNova | UK Chip Startup Olix Secures $220 Million to Develop Photonic AI Processors | |
| SO017 | Tech Funding News | 25-year-old founder’s Olix nabs $220M for photonic AI inference chips to take on Nvidia | |
| SO018 | Hummingbird Ventures | Hummingbird VC homepage | |
| SO019 | Plural | Team | Plural | |
| SO020 | Phoenix Court / LocalGlobe | Saul Klein | |
| SO021 | Vertex Ventures US | Vertex Team Member - Jonathan Heiliger | |
| SO022 | Stanford Profiles | Nick McKeown's Profile | |
| SO023 | Siemens EDA | Five Key Trends of Co-Packaged Optics (CPO) in 2026 | Yield excursions at any stage can have an outsized cost impact due to the limited reworkability of co-packaged assemblies. |
| SO024 | CNBC | The Tech Download: How chip companies are looking to use light to solve this major AI bottleneck | |
| SO025 | Edgen.Tech | CPO Mass Production Faces Delay Past 2026 on Yield Issues | Meaningful mass production now unlikely until after 2026 due to persistent manufacturing yield issues. |
| SM001 | International Energy Agency (IEA) | Electricity 2026 — Analysis | Global power demand growth continues to rise rapidly as the Age of Electricity gathers pace, supported by growing consumption from artificial intelligence, data centres, and evolving technological innovations. |
| SM002 | International Energy Agency (IEA) | AI is set to drive surging electricity demand from data centres — IEA Energy and AI report | Electricity demand from data centres worldwide is set to more than double by 2030 to around 945 terawatt-hours, slightly more than the entire electricity consumption of Japan today. AI will be the most significant driver of this increase, with electricity demand from AI-optimised data centres projected to more than quadruple by 2030. |
| SM003 | Data Center Frontier | IEA Study Sees AI, Cryptocurrency Doubling Data Center Energy Consumption by 2026 | |
| SM004 | MarketsandMarkets | Silicon Photonics Market — Global Forecast to 2030 (search results page) | The global silicon photonics market is expected to grow from USD 2.65 billion in 2025 to USD 9.65 billion by 2030, at a compound annual growth rate (CAGR) of 29.5% during the forecast period. |
| SM005 | Grand View Research | AI Accelerator Market Size, Share and Trends Analysis Report, 2025–2033 | The global AI accelerator market is expected to grow at a compound annual growth rate of 29.3% from 2025 to 2033, reaching USD 256.84 billion by 2033. |
| SM006 | Precedence Research | Artificial Intelligence (AI) Chip Market | |
| SM007 | NVIDIA Corporation | NVIDIA Announces Financial Results for First Quarter Fiscal 2026 | AI inference token generation has surged tenfold in just one year, and as AI agents become mainstream, the demand for AI computing will accelerate. — Jensen Huang, NVIDIA CEO |
| SM008 | NVIDIA Corporation | NVIDIA Announces Financial Results for Third Quarter Fiscal 2026 | Compute demand keeps accelerating and compounding across training and inference — each growing exponentially. We've entered the virtuous cycle of AI. — Jensen Huang, NVIDIA CEO |
| SM009 | Ayar Labs | Ayar Labs — Silicon Photonics Optical I/O for AI and HPC | Optical connectivity will be important to scale accelerated computing clusters to meet the fast-growing demands of AI and HPC workloads. — Bill Dally, Chief Scientist, NVIDIA |
| SM010 | PR Newswire (MarketsandMarkets) | Silicon Photonics Market worth $9.65 billion by 2030 — Exclusive Report by MarketsandMarkets | |
| SM011 | Business Wire (ResearchAndMarkets) | Global Silicon Photonics and Photonic Integrated Circuits Market Report 2025–2035 | |
| SM012 | PR Newswire (The Research Insights) | Thriving Silicon Photonics Market to Cross $8.13 Billion, Globally, by 2030 | |
| SM013 | PR Newswire (IDTechEx) | IDTechEx Release New Global Silicon Photonics and Photonic Integrated Circuits Market Report | |
| SM014 | Meta Platforms | Meta Reports First Quarter 2026 Results | We raised our 2026 capital expenditure outlook to $125–145 billion, reflecting expectations for higher component pricing this year and additional data center costs to support future year capacity. — Meta Q1 2026 earnings release |
| SM015 | InfotechLead | Alphabet Q1 2026 revenue surges 22% as AI, Google Cloud and $190bn capex strategy reshape growth | |
| SM016 | The Register | Silicon photonics won't matter 'anytime soon' says Broadcom | Silicon photonics won't matter in the datacenter "anytime soon" — Hock Tan, Broadcom CEO, Q4 2025 earnings call. |
| SM017 | The Register | DARPA asks labs to outsmart physics with photonic circuits | Systems incorporating photonic circuits struggle to show significant system-level performance advantages over electronic systems. — DARPA PICASSO solicitation |
| SM018 | Semiconductor Industry Association (SIA) | 2025 State of the U.S. Semiconductor Industry | |
| SM019 | Global Data Center Hub | Microsoft Q3 FY2026: The $190B Capex Plan That Repriced AI | |
| SM020 | Siemens EDA (Semiconductor Packaging Blog) | Five Key Trends of Co-Packaged Optics (CPO) in 2026 | |
| SM021 | edgen.tech | CPO Mass Production Faces Delay Past 2026 on Yield Issues | |
| SM022 | IEEE Spectrum | Photonic Chips Could Accelerate AI — If They Can Scale | Photonic chips encode data in light rather than electrons, enabling fundamentally higher bandwidth and lower power consumption for AI inference workloads at rack scale. |
| SM023 | IEEE Spectrum | Silicon Photonics Is Coming to the Data Center | Silicon photonics integrates optical components on standard CMOS wafers, allowing data centres to replace copper interconnects with optical links at hyperscale bandwidth. |
| SM024 | Semiconductor Engineering | Silicon Photonics For AI Accelerators | The integration of photonic elements with electronic AI accelerators faces non-trivial challenges in yield, packaging density, and thermal management at wafer scale. |
| SM025 | Light Reading | Silicon Photonics Set to Transform AI Networking | Optical interconnect vendors and chipmakers are racing to qualify silicon photonics components for hyperscaler AI networking racks ahead of anticipated 2027 deployment windows. |
| SP001 | Lightmatter | Lightmatter — homepage | Photonic Supercomputer Company |
| SP002 | Lightmatter | About Lightmatter | 114 Tbps today. 1+ Pbps tomorrow. |
| SP003 | Lightmatter | Passage — Photonic Interconnect Chiplets | 56–448 Gbps per lane, 1–16+ wavelengths per lane. Edgeless I/O via 3D photonic integration. |
| SP004 | Lightmatter | Idiom — ML Framework for Photonic Compute | |
| SP005 | Lightmatter | Envise — Photonic Computing Platform | |
| SP006 | Groq | GroqCloud Pricing | GPT OSS 20B — 1,000 TPS — $0.075/M input — $0.30/M output |
| SP007 | Groq | Groq Newsroom | |
| SP008 | Groq | Groq Raises $750 Million as Inference Demand Surges | Groq raises $750 million at a $6.9 billion valuation, led by Disruptive Investments, with participation from BlackRock, Samsung, Cisco, D1 Capital, and Altimeter Capital. |
| SP009 | Groq | Groq and NVIDIA Enter Non-Exclusive Inference Technology Licensing Agreement | Groq and NVIDIA have entered into a non-exclusive licensing agreement covering inference technology. |
| SP010 | Groq | GroqCloud — Developer Platform and API | |
| SP011 | Wikipedia | Groq — Wikipedia | |
| SP012 | Wikipedia | Silicon photonics — Wikipedia | |
| SP013 | Wikipedia | AI accelerator (Neural processing unit) — Wikipedia | |
| SP014 | Ayar Labs | Ayar Labs — About | Ayar Labs is the leader in co-packaged optics for AI scale-up. |
| SP015 | Ayar Labs | Optical I/O Products — TeraPHY and SuperNova | |
| SP016 | Ayar Labs | AI Scale-Up — Ayar Labs | |
| SP017 | Ayar Labs | Ayar Labs Closes $500M Series E Round | Ayar Labs closes $500 million Series E at a $3.75 billion valuation; total raised reaches $870 million; investors include NVIDIA, AMD, Alchip, and MediaTek. |
| SP018 | Ayar Labs | Ayar Labs Joins NVIDIA NVLink Fusion Ecosystem | Ayar Labs joins the NVIDIA NVLink Fusion ecosystem, enabling optical I/O integration with NVIDIA NVLink-based AI infrastructure. |
| SP019 | Artificial Analysis | Groq — Provider Performance and Pricing Analysis | |
| SP020 | IEEE Spectrum | Nvidia Blackwell Leads AI Inference, AMD Challenges | Nvidia Blackwell Leads AI Inference, AMD Challenges — benchmarking shows conventional GPU vendors hold every performance leadership position in deployed AI inference as of early 2025. |
| SP021 | Etched | Etched — homepage | Building the hardware for superintelligence. |
| SP022 | Etched | Etched Careers — Executive Team | |
| SP023 | Amazon Web Services | AWS Inferentia — Machine Learning Inference Accelerator | AWS Inferentia delivered 4.5x higher throughput and 90% lower inference cost compared to GPU-based EC2 instances in customer benchmarks. |
| SP024 | Google Cloud | Introduction to Cloud TPU | |
| SP025 | Intel | Intel Silicon Photonics — Products and OCI Chiplet | Intel is a pioneer in Silicon Photonics, having invested in this technology for over 20 years. The Intel Silicon Photonics Product Division has shipped more than 8 million PICs with over 32 million on-chip integrated lasers since 2016. |
| SP026 | OLIX | OLIX — homepage | |
| SP027 | OLIX | OLIX Compute Manifesto | Scaling an SRAM-architecture integrated with photonics can surpass HBM-based architectures on throughput/MW and TCO. The OLIX Decode Accelerator 1 (DX-1) is the first system architected specifically for decode. |
| SI001 | SiliconANGLE | Photonic AI chip startup OLIX nabs $220M investment | OLIX will use its newly raised capital to finance chip development initiatives. |
| SI002 | TechSparx | London-based start-up Olix selects Bristol for its new engineering facility | salaries for a senior digital design engineer ranging between £125,000 and £180,000. Successful applicants are offered a £24,000 annual supplement for living near the office. |
| SI003 | Jon Peddie Research | An AI processor by any other name | Olix has grown from a founding team to over 70 employees and is aiming to scale to over 200 this year. |
| SI004 | Tech Funding News | 25-year-old founder's Olix nabs $220M for photonic AI inference chips to take on Nvidia | OLIX will use its newly raised capital to finance chip development initiatives. |
| SI005 | IDCNova | A UK-based semiconductor startup Olix raises $220M at $1B valuation | This latest investment values the two-year-old company at $1 billion and includes backing from previous investors Plural, Vertex Ventures, LocalGlobe, and Entrepreneurs First. |
| SI006 | The Tech Founders | Olix: 25-year-old founder raises $220M for UK AI chip startup | |
| SI007 | Olix | OLIX Careers Manifesto | We pay an additional £2,000 per month to those living within 20 minutes (85% of us do). At the three-month mark... we'll pay you 18 weeks of salary to leave. |
| SI008 | Olix | OLIX Compute Manifesto | It is our belief that scaling an SRAM-architecture integrated with photonics can surpass HBM-based architectures on throughput/MW and TCO. |
| SI009 | Olix | Senior FPGA Engineer — OLIX Careers | |
| SI010 | Companies House | OLIX COMPUTING LIMITED filing history — accounts category | 21 Mar 2026 — AA — Total exemption full accounts made up to 31 August 2025 |
| SI011 | Companies House | OLIX COMPUTING LIMITED filing history — confirmation statement | 31 Mar 2026 — CS01 — Confirmation statement made on 21 March 2026 with updates |
| SI012 | Siemens (EDA blogs) | Five Key Trends of Co-Packaged Optics (CPO) in 2026 | Silicon photonics manufacturing still exhibits higher variability and places greater emphasis on yield and Known-Good-Die assurance. Yield excursions at any stage can have an outsized cost impact due to the limited reworkability of co-packaged assemblies. |
| SI013 | edgen.tech | CPO Mass Production Faces Delay Past 2026 on Yield Issues | CPO Mass Production Faces Delay Past 2026 on Yield Issues |
| SI014 | Hummingbird Ventures | Hummingbird VC Portfolio | |
| SI015 | Groq | Groq Raises $750 Million as Inference Demand Surges | |
| SI016 | Ayar Labs | Ayar Labs Closes $500M Series E, Accelerates Volume Production of Co-Packaged Optics | |
| SI017 | UK Government | Prepare annual accounts for a private limited company | If your company is small, a micro entity or dormant, you might be able to send simpler ('abridged') accounts. |
| SI018 | UK Government (DSIT) | National semiconductor strategy | This strategy sets a vision that over the next 20 years the UK will have a world leading position in the semiconductor technologies of the future by focusing on our strengths in intellectual property and design. |
| SI019 | Hummingbird Ventures | Hummingbird VC — About | We invest from $500K at seed to $50M+ when doubling and tripling down. |
| SI020 | UKTN (UK Tech News) | Fractile raises £162.8m to boost AI inference chip development | UK AI chip startup Fractile has raised a $220m (£162.8m) Series B round to accelerate the build of its next-generation inference hardware for AI. |
| SI021 | Financial Times | 25-year-old founder raises $220mn for secretive UK AI chip start-up | According to the Financial Times, OLIX expects to start shipping OTPU chips to customers next year. |
| SI022 | The Register | Silicon photonics won't matter anytime soon, says Broadcom | Q4 revenue from semiconductors reached $11 billion, up 35 percent year-over-year. $6.7 billion of that revenue came from AI products. |
| SI023 | CNBC | The Tech Download: How chip companies are looking to use light to solve this major AI bottleneck | Since the beginning of March, Nvidia has announced $2 billion investments into Lumentum, Coherent and Marvell... and participated in Ayar Labs' $500 million Series E funding round. |
| SI024 | MarketsandMarkets via PR Newswire | Silicon Photonics Market worth $9.65 billion by 2030 | The global silicon photonics market was valued at USD 2.65 billion in 2025 and is projected to reach USD 9.65 billion by 2030, growing at a CAGR of 29.5%. |
| SI025 | Semiconductor Industry Association (SIA) | 2025 State of the U.S. Semiconductor Industry | |
| SE001 | OLIX | Compute Manifesto | "The OLIX Decode Accelerator 1 (DX-1) is the first system architected specifically for decode. Built on rack-scale co-design of logic, data movement, packaging, optics and interconnect, it delivers a step change in system level performance." |
| SE002 | OLIX | Job Description: Senior Optical Communications Engineer | "OLIX is developing next-generation optical interconnect architectures for AI data centers, spanning chip-to-chip, intra-tray, and rack-scale links." |
| SE003 | OLIX | Job Description: Architect/Staff/Senior Platform Integration Engineer | "Deep expertise in distributed inference infrastructure (vLLM, SGLang, Nvidia Dynamo) as well as associated networking (NCCL, RoCE, Infiniband) and KV cache management (NIXL, Mooncake TE)." |
| SE004 | OLIX | Job Description: Senior FPGA Engineer | "Develop and optimise high-speed interface subsystems (PCIe Gen4, multi-lane transceivers operating at >25 Gb/s NRZ)." |
| SE005 | OLIX | Careers Page | |
| SE006 | Siemens EDA (Semiconductor Packaging Blog) | Five Key Trends of Co-Packaged Optics (CPO) in 2026 | "Test scalability is likely to be a primary bottleneck for CPO adoption, as optical and electrical domains impose fundamentally different constraints." |
| SE007 | edgen.tech | CPO Mass Production Faces Delay Past 2026 on Yield Issues | |
| SE008 | CNBC | The Tech Download: How chip companies are looking to use light to solve this major AI bottleneck | |
| SE009 | SiliconANGLE | Photonic AI chip startup OLIX nabs $220M investment | |
| SE010 | vLLM Project (GitHub) | vLLM: Easy, Fast, and Cheap LLM Serving for Everyone | "vLLM has grown into one of the most active open-source AI projects built and maintained by a diverse community of many dozens of academic institutions and companies from over 2000 contributors." |
| SE011 | SGLang Project (GitHub) | SGLang: Fast Serving Framework for Large Language and Vision-Language Models | "Deploying DeepSeek on GB200 NVL72 with PD and Large Scale EP (Part I): 2.7x Higher Decoding Throughput." |
| SE012 | AI Market Watch | Olix — Company Profile | "Olix is re-architecting the fundamental compute pipeline to forgo HBM entirely, which is currently the most expensive and supply-constrained component in AI hardware." |
| SE013 | PitchBook | Olix Computing — Company Profile | |
| SE014 | NVIDIA Corporation | NVIDIA Co-Packaged Optics News | |
| SE015 | TechCrunch | Olix raises $220 million for photonic AI chips | |
| SE016 | WIRED | Photonic AI Chips: The Olix Startup Building Inference Hardware for 2026 | |
| SE017 | Reuters | Olix raises $220 million in AI chip startup funding | |
| SE018 | UK Tech News (UKTN) | Olix raises £160M to develop photonic AI chip as unicorn | |
| SE019 | OLIX | Job Description: General Counsel | |
| SE020 | Tech Funding News | 25-year-old founder's Olix nabs $220M for photonic AI inference chips to take on Nvidia | |
| SE021 | The Register | Silicon photonics won't matter anytime soon, says Broadcom | |
| SE022 | The Register | DARPA asks labs to outsmart physics with photonic circuits | |
| SE023 | Nordic9 | Olix Computing raises $220M Series A led by Hummingbird Ventures | |
| SE024 | The Register | UK photonic chip startup Olix raises $220M | |
| SE025 | The Register | Olix photonic AI chip details | |
| SE026 | vLLM Documentation | vLLM Documentation — Hardware Support | "Support for NVIDIA GPUs, AMD GPUs, and x86/ARM/PowerPC CPUs. Additionally, diverse hardware plugins such as Google TPUs, Intel Gaudi, IBM Spyre, Huawei Ascend, Rebellions NPU, Apple Silicon." |
| SE027 | TechSparx | London-based start-up Olix selects Bristol for its new engineering facility | |
| SU001 | Olix | OLIX — Homepage | Scaling an SRAM-architecture integrated with photonics can surpass HBM-based architectures on throughput/MW and TCO, and significantly outperform silicon-only SRAM-architectures in interactivity and latency. |
| SU002 | Olix | Olix Compute Manifesto — DX-1 Decode Accelerator | Even the most dominant players, with full-stack control across silicon, software, and supply chains, are unable to solve this within the existing architecture. |
| SU003 | Olix | OLIX Careers Manifesto | |
| SU004 | Olix | OLIX Careers — Open Roles | Locations: London, Austin (Texas), Toronto, Bristol, San Francisco. |
| SU005 | Olix | Senior FPGA Engineer — Olix Job Posting | |
| SU006 | Olix | Architect / Staff / Senior Platform Integration Engineer — Olix Job Posting | You will be the technical authority on how OLIX serves large models as hyperscale AI infrastructure — spanning distributed inference engines, serving-runtime integration, KV cache and memory hierarchy, and the orchestration and networking layers that make serving real. |
| SU007 | The Register | Silicon photonics won't matter 'anytime soon' says Broadcom | Silicon photonics won't matter in the datacenter "anytime soon." "We are not quite there yet." — Broadcom CEO Hock Tan, Q4 FY2025 earnings. "To invest in custom accelerators is a multi-year journey and not transactional or short-term decision." |
| SU008 | The Register | DARPA asks labs to outsmart physics with photonic circuits | Systems incorporating photonic circuits struggle to show significant system-level performance advantages over electronic systems. |
| SU009 | Groq | Groq Customer Case Studies | Willow achieved zero downtime and 500 ms faster AI responses with Groq (September 2025). PGA of America: Transforming Operations with Faster, Smarter AI. |
| SU010 | Groq | Groq Raises $750 Million as Inference Demand Surges | Groq powers more than two million developers and Fortune 500 companies with fast, affordable compute and is growing its presence globally, building on existing data centers in North America, Europe, and the Middle East. |
| SU011 | Amazon Web Services | Amazon EC2 Inf2 — AWS Inferentia2 Product Page | NTT PC: 4.5x higher throughput, 25% lower inference latency, and 90% lower cost compared to current-generation GPU-based EC2 instances. |
| SU012 | Google Cloud | Introduction to Cloud TPU — Google Cloud Documentation | |
| SU013 | MLCommons | MLPerf Inference Benchmarks | |
| SU014 | SiliconAngle | Photonic AI chip startup OLIX nabs $220M investment | OLIX expects to start shipping OTPU chips to customers next year. It's unclear what optical components are included in the processor or how they're used. |
| SU015 | TechFundingNews | 25-year-old founder's Olix nabs $220M for photonic AI inference chips | |
| SU016 | vLLM Project | vLLM — GitHub Repository | One engine, endless possibilities. Run any model on any hardware. We welcome and value any contributions and collaborations. |
| SU017 | Artificial Analysis | Groq — Inference Provider Performance and Pricing | |
| SU018 | Artificial Analysis | AI Model Quality and Performance Benchmarks | |
| SU019 | UK Tech News | Fractile raises £162.8m to boost AI inference chip development | Fractile is developing chips and systems designed to make faster inference economically viable, working across AI research, chip micro-architecture and foundry process innovation. |
| SU020 | Edgen.Tech | CPO Mass Production Faces Delay Past 2026 on Yield Issues | |
| SU021 | Ayar Labs | Ayar Labs Joins NVIDIA NVLink Fusion Ecosystem | AI infrastructure is being co-designed from the ground up, and customers need more options to scale performance efficiently as bandwidth continues to rise. |
| SU022 | TechSparx | London-based start-up Olix selects Bristol for its new engineering facility | |
| SU023 | Lambda Labs | Lambda Cloud — GPU Instances and On-Demand Compute | |
| SU024 | Financial Times | Olix raises $220M for photonic AI chips at over $1bn valuation | OLIX expects to start shipping OTPU chips to customers next year (per SiliconAngle citation of FT reporting). |
| SU025 | Wikipedia | AI accelerator — Wikipedia | |
| SU026 | Groq | Groq Blog — GroqCloud updates and inference performance | Build Fast. Seamlessly integrate Groq starting with just a few lines of code. |
| SU027 | CoreWeave | CoreWeave — GPU cloud for AI inference | Pioneering the possible. CoreWeave and the Aston Martin Aramco F1 Team unite world-class racing with cutting-edge cloud computing. |
| SU028 | SiliconAngle | Photonic AI chips face datacenter deployment timeline challenges — SiliconAngle | Photonic AI chips face significant datacenter deployment timeline challenges. |
| SU029 | Semiconductor Engineering | Silicon Photonics Qualification Challenges for AI Accelerators | |
| SR001 | US Bureau of Industry and Security (BIS) | Export Administration Regulations (EAR) — BIS Homepage | A license is required to export advanced computing items to entities headquartered in Country Group D:5 or Macau — even if the entities themselves are located outside Country Group D:5 or Macau. |
| SR002 | UK National Archives / legislation.gov.uk | National Security and Investment Act 2021 | National Security and Investment Act 2021 — grants Secretary of State powers to review acquisitions in sensitive sectors including semiconductors. |
| SR003 | TSMC | TSMC Dedicated Foundry Technology — Specialty Technology | |
| SR004 | ASML | ASML — The World's Supplier to the Semiconductor Industry | Our lithography technology is fundamental to mass producing semiconductor chips. |
| SR005 | US National Institute of Standards and Technology (NIST) | NIST Artificial Intelligence — Risk Management Framework | |
| SR006 | UK Department for Science, Innovation and Technology (DSIT) | AI regulation: a pro-innovation approach | |
| SR007 | EE Times | EE Times — Connecting The Global Electronics Industry | |
| SR008 | SPIE — The International Society for Optics and Photonics | SPIE Photonics West News — Silicon Photonics and AI Integration | |
| SR009 | Siemens EDA | Five Key Trends of Co-Packaged Optics (CPO) in 2026 | Test scalability is likely to be a primary bottleneck for CPO adoption, as optical and electrical domains impose fundamentally different constraints. Optical signals are highly sensitive to alignment, with micron-scale deviations capable of introducing significant loss. |
| SR010 | EdgeN Intelligence | CPO Mass Production Faces Delay Past 2026 on Yield Issues | CPO Mass Production Faces Delay Past 2026 on Yield Issues |
| SR011 | Companies House (UK) | OLIX COMPUTING LIMITED — Company Overview | |
| SR012 | Olix | OLIX — Homepage | |
| SR013 | SiliconANGLE | Photonic AI chip startup OLIX nabs $220M investment | |
| SR014 | Tech Funding News | 25-year-old founder's Olix nabs $220M for photonic AI chips inference | |
| SR015 | Jon Peddie Research | An AI processor by any other name | |
| SR016 | IDCNova | A UK-based semiconductor startup Olix raises $220M | |
| SR017 | Financial Times | 25-year-old founder raises $220mn for secretive UK photonic AI chip startup | |
| SR018 | CNBC | The Tech Download: How chip companies are looking past photonics as AI bottleneck | |
| SR019 | Semiconductor Industry Association (SIA) | 2025 State of the U.S. Semiconductor Industry | |
| SR020 | Light Reading | Silicon Photonics Set to Transform AI Networking | |
| SR021 | UKtech.news | UK AI chip startup Olix raises $220M | |
| SR022 | UK Department for Science, Innovation and Technology (DSIT) | National semiconductor strategy | |
| SR023 | Olix | Olix Compute Manifesto — DX-1 Decode Accelerator | |
| SR024 | Cooley LLP | OLIX Raises $220 Million in Financing | |
| SR025 | BusinessWire | Global Silicon Photonics and Photonic Integrated Circuits Market | |
| SR026 | Semiconductor Engineering | Co-Packaged Optics Promises and Problems | |
| SR027 | LightCounting | Photonic AI Chips and Silicon Photonics Market | |
| SR028 | SPIE Digital Library | Silicon Photonics and Photonic Integrated Circuits for AI | |
| SR029 | MIT Technology Review | Photonic Computing Chips | |
| SR030 | Nature Photonics | Photonic integrated circuits for AI accelerators | |
| SR031 | EE Times | Silicon Photonics: Challenges on the Path to Co-Packaged Optics | |
| SR032 | Pitchbook | Silicon Photonics Startup Funding Analysis | |
| SR033 | Olix | OLIX Careers — Open Roles | |
| SV001 | NVIDIA Corporation | NVIDIA Announces Financial Results for First Quarter Fiscal 2027 | NVIDIA today reported record revenue for the first quarter ended April 26, 2026, of $81.6 billion, up 20% from the previous quarter and up 85% from a year ago. |
| SV002 | Lightmatter | Guide — Very Large Scale Photonic Light Engine for AI | |
| SV003 | Marvell Technology (via Celestial AI domain) | Accelerated Infrastructure for the AI Era — Marvell (celestial.ai) | Marvell designs custom silicon tailored for any application and offers the industry's most comprehensive portfolio of interconnects and network switch products. |
| SV004 | Semiconductor Engineering | Startup Funding: Q1 2026 — Massive rounds for AI, EDA, and manufacturing; 80 startups raise $8.4B | Photonics remained a popular segment for investment. 80 companies collectively raised over $8 billion in the first quarter of 2026. |
| SV005 | Jon Peddie Research | Photonics — Jon Peddie Research tag page | |
| SV006 | Semiconductor Engineering | Startup Funding: Q4 2025 — More and bigger funding rounds for AI chips and AI for making chips | |
| SV007 | BusinessWire | Positron AI Raises $230 Million Series B at Over $1 Billion Valuation to Scale Energy-Efficient AI Inference | Positron AI today announced an oversubscribed $230 million Series B financing at a post-money valuation exceeding $1 billion. |
| SV008 | Semiconductor Engineering | Chip Industry Week In Review #135 — Marvell's photonics buy; photonics deals; compute architecture | Marvell acquired Polariton Technologies, a Swiss developer of plasmonics-based silicon photonics devices. |
| SV009 | Positron AI | Positron AI — Accelerating Intelligence — Generative AI Acceleration | |
| SV010 | UK Companies House | OLIX COMPUTING LIMITED filing history — Accounts and resolutions | Statement of capital following an allotment of shares on 5 March 2026 at GBP 287.8572. |
| SV011 | SiliconANGLE | Photonic AI chip startup OLIX nabs $220M investment | OLIX raised $220M investment lifting its valuation above $1 billion. |
| SV012 | Tech Funding News | 25-year-old founder's Olix nabs $220M for photonic AI chips inference | London-based startup Olix has secured $220 million in new funding, lifting its valuation above $1 billion. |
| SV013 | Groq | Groq Raises $750 Million as Inference Demand Surges | Groq announced $750 million in new financing at a post-money valuation of $6.9 billion. |
| SV014 | Ayar Labs | Ayar Labs Closes $500M Series E, Accelerates Volume Production of Co-Packaged Optics | |
| SV015 | Grand View Research | AI Accelerator Market Size, Share and Trends Analysis Report, 2025–2033 | The global AI accelerator market size was estimated at USD 25.56 billion in 2024 and is projected to reach USD 256.84 billion by 2033, growing at a CAGR of 29.3%. |
| SV016 | Precedence Research | Artificial Intelligence (AI) Chip Market — Size and Forecast | The global AI chip market size was $94.44 billion in 2025 and predicted to increase from $121.73 billion in 2026 to approximately $1,104.68 billion by 2035. |
| SV017 | CNBC | The Tech Download: How chip companies are looking to use light to solve this major AI bottleneck | The main challenge for incorporating more optical components is the need to significantly redesign the existing product roadmaps. That may require one or two more generations of products from the likes of Nvidia. |
| SV018 | The Tech Founders | Olix: 25-year-old founder raises $220M for UK AI chip startup | |
| SV019 | Lightmatter | About Lightmatter | |
| SV020 | UK Companies House | OLIX COMPUTING LIMITED filing history | |
| SV021 | Olix | OLIX Compute Manifesto — DX-1 Decode Accelerator | |
| SV022 | Jon Peddie Research | An AI processor by any other name | Olix isn't the only or the first to pursue optical computing and coupling. Neurophos, Lightmatter, OptoML, Rayd, and Volantis have similar ideas. The question is, will the problem they solve be the one people care about by the time they do it? |
| SV023 | NVIDIA Corporation | NVIDIA Announces Financial Results for Third Quarter Fiscal 2026 | |
| SV024 | IDCNova | A UK-based semiconductor startup Olix raises $220M at $1B valuation | |
| SV025 | Hummingbird Ventures | Hummingbird VC — Portfolio | |
| SV026 | UKTN (UK Tech News) | Fractile raises £162.8m to boost AI inference chip development | UK AI chip startup Fractile has raised a $220m (£162.8m) Series B round to accelerate the build of its next-generation inference hardware for AI. |
| SV027 | AI Market Watch | Olix — AI Startup Profile | |
| SV028 | Reuters | Olix raises $220 million in AI chip startup funding | |
| SV029 | Financial Times | 25-year-old founder raises $220mn for secretive UK photonic AI chip startup | |
| SV030 | Olix | OLIX — Homepage |