Cast AI
自主 Kubernetes 优化,并具备新兴 AI 基础设施期权
Cast AI 兼具强劲的企业采用度、差异化的 Kubernetes 自动化能力和 $1B+ 估值,但 ARR、利润率及本轮融资经济条款均未披露,类似股票的风险收益仍只能评为 Track。
覆盖范围与披露说明
公开证据支持其强劲的企业采用、2026 年 1 月估值上升至 10 亿美元以上,以及向 GPU 编排扩张;但经审计的 ARR、利润率、留存率和融资轮次经济性仍未披露。
封面要素
公司概况
Cast AI 是一家成立于 2019 年的云基础设施自动化公司,帮助企业优化 Kubernetes 工作负载、调整计算资源规模,并在 AWS、Azure 和 Google Cloud 之间编排专用 GPU 容量。公司总部位于迈阿密,在维尔纽斯设有工程团队,服务超过 2,100 家组织;2026 年 1 月又通过 OMNI Compute 和 Oracle Cloud Infrastructure 合作支持扩充了产品线。 公开披露显示客户势头可观、估值超过 10 亿美元,但核心经营经济性仍属私有信息。
- 官网
- cast.ai
- 成立时间
- 2019-01-01
- 创始人
- Yuri Frayman, Laurent Gil, Leon Kuperman
- 创立地点
- Miami, Florida
- 总部
- Miami, FL (engineering: Vilnius, Lithuania)
- 产品
- 面向多云基础设施团队的自主 Kubernetes 优化软件,以及 OMNI Compute GPU 编排。
- 客户
- 在大规模运行 Kubernetes 的企业平台工程、DevOps 和 FinOps 团队。
- 商业模式
- 年度订阅为主,辅以与使用量和节省额挂钩的优化及 GPU 编排定价。
- 阶段
- Series C / Growth (unicorn)
- 融资情况
- 截至 2026 年,公司在十轮融资中累计融资约 $272M;另有一笔未披露的 2026 年 1 月战略投资,将估值推高至 $1B 以上。
执行摘要
主要优势
- 已覆盖超过 2,100 家企业,包括多家知名全球客户,企业采用度强
- Kubernetes 自主优化和 OMNI Compute 让平台超越被动式 FinOps 工具
- 2026 年 1 月的战略投资及 $1B+ 估值,验证了其赛道相关性和战略吸引力
主要风险
- 经审计 ARR、毛利率、NRR、现金消耗和现金跑道仍未披露
- 云超大规模厂商的原生工具及其他 FinOps 竞品,可能随着时间推移压缩定价权
- 与节省额和用量挂钩的变现模式,比纯订阅收入更难纳入投资判断
未决问题
- 当前 ARR 和收入增速尚无公开披露
- 2026 年 1 月战略投资的确切金额及经济条款仍未披露
- 毛利率、NRR、现金消耗和客户集中度数据不可得
目录
01公司概览
1.1 公司身份、创立与领导层
Cast AI Group Inc. 是一家成立于 2019 年、总部位于佛罗里达州迈阿密的云基础设施自动化公司,主要欧洲研发团队位于立陶宛维尔纽斯。公司由连续创业的企业软件创业者 Yuri Frayman、Laurent Gil 和 Leon Kuperman 创立;三人此前分别从 Viewdle 和 Zenedge 的领导岗位退出。Cast AI 是面向 Amazon Web Services、Google Cloud Platform 和 Microsoft Azure 云原生 Kubernetes 工作负载的自主性能优化引擎。它并非单纯的可观测性仪表盘,而是用机器学习代理持续调整 Pod 资源、对容器做装箱,并实时以动态配置的 Spot 实例替换昂贵的按需节点。CEO Yuri Frayman 负责公司方向和机构融资;总裁 Laurent Gil 主导全球合作伙伴生态与商业扩张;CTO Leon Kuperman 负责算法代理工程。公司高度依赖这三位创始人:他们一致的技术愿景和过往创业退出经历,既支撑客户信任,也支撑机构风投背书。[CO001, CO007, CO008, CO010]
| 人员 | 职务 | 背景 | 职能覆盖 | 关键人物依赖 |
|---|---|---|---|---|
| Yuri Frayman | 联合创始人兼 CEO | 连续创业的企业软件企业家;Viewdle(被 Google 收购)和 Zenedge(被 Oracle 收购)的联合创始人 | 高管管理、战略募资和全球市场拓展 | 高:公司愿景的首席设计者,也是机构融资轮次的主要发言人 |
| Laurent Gil | 联合创始人兼总裁 | Oracle 前高管,Viewdle 和 Zenedge 联合创始人;云基础设施与安全领域资深人士 | 商业执行、生态合作和 OMNI Compute 市场推广 | 高:超大规模云厂商联盟及行业遥测披露的主要对外联络人 |
| Leon Kuperman | 联合创始人兼 CTO | Oracle 前副总裁,Viewdle 和 Zenedge 联合创始人;在云架构和机器学习领域拥有深厚专业权威 | 技术架构、自主代理工程和核心知识产权 | 高:自动调配算法和预测性遥测背后的技术主脑 |
高管履历和过往公司退出已通过独立科技新闻报道及公司公告交叉核实;三位联合创始人此前都曾领导 Zenedge 和 Viewdle。
[CO001, CO007, CO008]1.2 资本结构、融资历程与独角兽估值
自 2019 年成立以来,Cast AI 已在六轮有记录的融资中累计募得超过 $180 million 股权资本。公司于 2020 年 11 月宣布获得 DNX Ventures、TA Ventures 和 Florida Funders 支持的 $8 million 种子轮;随后于 2021 年 12 月完成由 Cota Capital 领投、Samsung Next 参与的 $10 million A 轮。2023 年增长提速:3 月由 Creandum 领投 $20 million 早期 B 轮,11 月由 Vintage Investment Partners 领投 $35 million B 轮扩展。2025 年 4 月,Cast AI 完成由 G2 Venture Partners 和 SoftBank Vision Fund 2 共同领投的 $108 million C 轮,投后估值估计为 $900 million。2026 年 1 月,公司获得韩国新世界集团旗下企业创投机构 Pacific Alliance Ventures 的未披露战略投资,估值超过 $1 billion,跻身独角兽。虽然这一里程碑让 Cast AI 与 SoftBank 的 AI 基础设施组合并列于高关注位置,但收入运行率、毛利率和净现金消耗等关键财务基础仍是严密保密的私有指标,构成重大尽调缺口。[CO002, CO004, CO005, CO006, CO014]
| 指标 | 数值 / 状态 | 日期 / 时点 | 置信度 | 尽调缺口 |
|---|---|---|---|---|
| 投后估值 | $1,000M+(独角兽) | 2026-01 | 高 | 相对 Series C 的具体估值提升倍数为估算值 |
| 累计股权融资额 | $181M+ | 2026-01 | 高 | 战略投资的确切金额未披露 |
| 年度经常性收入(ARR) | 未披露的私募指标 | 2026-10 | 低 | 未公开披露经审计收入或收入运行率 |
| 活跃客户数 | 2,100+ 家组织 | 2024(Series C 时披露) | 中 | 近期文件未更新当前 2026 年数量 |
| 全球总部 | Miami, FL(研发:Vilnius, Lithuania) | 2026 | 高 | 已核实的公司实体和工程布局 |
| 核心工作负载方向 | Kubernetes 成本自动化与 OMNI GPU 市场 | 2026-01 | 高 | 已核实平台正式发布及与 Oracle OCI 的合作 |
截至 2026-10 的一级和二级私募市场数据;没有官方披露的财务 KPI 均标记为私募尽调缺口。
[CO001, CO002, CO003, CO004, CO006, CO009, CO014]| 利益相关方 | 职务 / 关系 | 控制权或经济重要性 | 尽调要求 |
|---|---|---|---|
| Pacific Alliance Ventures (Shinsegae Group) | 战略领投方(2026 年 1 月) | 将 Cast AI 推至 $1B+ 独角兽估值;亚洲扩张的战略伙伴 | 核实战略投资的确切金额及分配的治理董事席位 |
| SoftBank Vision Fund 2 | Series C 联合领投方(2025 年 4 月) | 主要成长期资本提供者;将 Cast AI 接入更广泛的 SoftBank AI 生态 | 确认二级持股及后续轮次中的任何共同投资权 |
| G2 Venture Partners | Series C 联合领投方(2025 年 4 月) | $108M 轮次领投方;专注于规模化的可持续运营效率 | 评估董事会代表权及其对公司盈利时间表的影响 |
| Vintage Investment Partners | Series B 领投方(2023 年 11 月)及 Series C 参与方 | 早期成长投资者,领投 $35M 轮;在多轮融资中持续投入并保持经济承诺 | 评估历史清算优先权及已行使的按比例认购权 |
| Creandum | Series B 早期领投方(2023 年 3 月)及 Series C 参与方 | 领投早期 $20M Series B;欧洲基础性创投伙伴,持有长期治理权益 | 审阅早期股东投票协议和转换条款 |
| Oracle Corporation | 战略超大规模云合作伙伴(2026 年 1 月) | 首家通过 OMNI Compute 市场释放闲置 GPU 容量的主要云服务商 | 检查涉及 Oracle Cloud Infrastructure 的收入分成条款和独家协议 |
投资者联合体和战略商业伙伴信息整理自公开融资公告及私募市场数据库;私募股权持仓仍未备案。
[CO002, CO003, CO004, CO005]1.3 平台架构、规模轨迹与里程碑演进
平台持续从基础容器资源调整扩展到完整的 AI 计算编排。2023 至 2024 年间,Cast AI 的市场势头快速增强,客户数翻倍以上,超过 2,100 家企业组织,其中包括 Akamai、BMW、Cisco、FICO 和 Hugging Face。平台创新最终在 2026 年 1 月推出 OMNI Compute 正式版:这是一个多云控制平面,可在不改代码的情况下,把外部计算容量和专用 GPU 集群转化为可互换的 Kubernetes 资源。OMNI Compute 首个超大规模云合作方是 Oracle Cloud Infrastructure,使运行在任意云上的组织都能按需调用 OCI 高性能 GPU 集群。该发布直接针对 Cast AI 对数万集群遥测后发现的系统性基础设施低效:企业 GPU 平均利用率仅约 5%,CPU 利用率为 8%。Cast AI 通过跨云 GPU 配置和自动回退机制,试图在企业 AI 推理工作负载激增之际盘活闲置计算容量。[CO003, CO009, CO010, CO012, CO013]
展示从 2019 年成立到 2026 年独角兽估值期间的关键融资、平台架构及运营规模里程碑。
[CO001, CO002, CO003, CO004, CO005, CO006, CO009]1.4 商业采用、运营摩擦与尽调判断
全球企业客户的采用验证了 Cast AI 自主降低云基础设施成本 50% 至 70% 的价值主张。Kubecost 等传统可视化工具只能生成需要开发者介入的静态分摊报告,而 Cast AI 会自主修改集群基础设施,直接消除浪费支出。不过,独立评测和企业部署也暴露出运营层面的注意事项。CyberNews 分析师指出,高级 Kubernetes 策略配置需要专业 DevOps 能力;与原生云厂商计费控制台相比,成本报告的颗粒度可能受限;定价模式也可能让小型开发团队感到阻力。此外,Cast AI 按使用量和实际节省额收费,企业采购团队必须结合波动的云消耗主动测算账单敞口。关键是,在客户留存、净收入留存率和运营现金消耗均缺乏经验证据的情况下,潜在利益相关方必须在 Cast AI 强大的平台能力和头部客户案例,与支持其 10 亿美元估值的经审计财务报表缺失之间权衡。[CO009, CO011, CO012, CO014]
1.5 展示项
02市场分析
2.1 市场边界、范围与现状替代方案
自主 Kubernetes 优化市场位于容器编排、云基础设施配置和持续云财务管理的交汇处。可服务边界涵盖运行在托管 Kubernetes 服务上的计算支出,主要包括 Amazon Elastic Kubernetes Service(EKS)、Google Kubernetes Engine(GKE)和 Microsoft Azure Kubernetes Service(AKS);动态工作负载在这些环境中需要持续调整资源、配置节点并管理 Spot 实例生命周期。边界外包括非容器化传统虚拟机、未托管的本地裸机,以及只做发票核对而非自主基础设施执行的通用 SaaS 财务管理层。当前企业部署中的现状替代方案,主要是 Kubernetes Cluster Autoscaler、Horizontal Pod Autoscaler(HPA)和 Karpenter 等原生开源自动扩缩组件;它们高度依赖手工 YAML 配置和反应式阈值规则,往往保留过于保守的资源余量。[CM008, CM009, CM012, CM014]
| 类别 | 纳入支出 | 排除支出 | 买方与付款方 | 战略相关性 |
|---|---|---|---|---|
| Kubernetes 自主自动扩缩容 | 计算节点配置、Spot 编排、Pod 调配 | 非容器化 VM 基础设施、裸机托管 | 平台 / 云基础设施副总裁预算 | 核心运营方向和自动化收入的主要驱动力 |
| AI / GPU 集群优化 | GPU 节点配置、时间切片、多云放置 | 专有模型训练知识产权、基础 LLM API 计费 | AI 基础设施负责人 / MLOps 预算 | AI 硬件成本飙升下单单位节省潜力最高 |
| 云财务管理(FinOps) | 集群级成本可见性、成本回溯、分配标签 | 企业 SaaS 支出、数据仓库查询、承诺合同 | FinOps 总监 / 工程财务预算 | 邻近工具层,通常通过合作伙伴生态集成 |
| 静态自动扩缩容工具 | 开源 Cluster Autoscaler、基础 Karpenter 和 HPA 规则 | 商业支持许可证、专有自动化套件 | DevOps 工程师 / 平台运营 | 现有基线和现状架构的主要替代方案 |
界定自主 Kubernetes 优化相对于邻近 FinOps 和现有云工具的可服务范围。
[CM008, CM009, CM012]2.2 市场规模视角与实际利用率基线
评估自主 Kubernetes 优化的可服务市场,必须用实际遥测数据校准规模估算,而不能依赖宽泛的自上而下云支出总量。Cast AI 分析的数万生产 Kubernetes 集群显示,计算浪费具有系统性且正在恶化:平均 CPU 利用率同比从 10% 降至 8%,内存利用率降至 20%。同时,CPU 过度配置从 40% 扩大到 69%,因为应用团队为避免限流和内存不足崩溃而人为抬高部署资源请求。生成式 AI 和机器学习工作负载的兴起进一步放大了低效:企业 GPU 平均利用率仅 5%,在高端硬件实例涨价 15% 的同时,95% 的稀缺加速器容量处于闲置状态。因此,这一机会对应云原生企业中规模达数十亿美元的效率回收池。[CM001, CM002, CM003, CM004, CM015]
| 发布方 | 年份 | 地理范围 | 数值指标 | CAGR | 方法论 | 置信度 | 局限 |
|---|---|---|---|---|---|---|---|
| Cast AI K8s 状态报告 | 2026 | 全球(AWS/GCP/Azure) | 8% CPU / 20% 内存利用率 | 未报告 | 来自 23,000+ 个未优化集群的直接遥测 | 高 | 公司发布的遥测;衡量技术浪费,而非美元支出总额 |
| Cast AI 加速器遥测 | 2026 | 全球(多云) | 5% 平均 GPU 利用率 | 未报告 | 对未优化 GPU 节点的直接测量 | 高 | 高端芯片的 Spot 可用性有限,导致企业预留习惯出现偏差 |
| Business Insider 分析 | 2026 | 全球 | 95% 闲置 GPU 容量 | 未报告 | 对企业囤积 AI 算力和集群数据的报道 | 中 | 反映 Cast AI 遥测而非独立普查的二手分析 |
| 竞争对手基准(nOps / Kubex) | 2026 | 全球 | 35% 至 50%+ 有效节省 | 未报告 | 竞争工具评估和客户基准 | 中 | 竞争对手发布的营销材料,缺乏标准化审计方法 |
综合一级遥测和第三方行业基准中的实证利用率指标及市场浪费代理指标。
[CM001, CM002, CM004, CM007, CM011]从公共云总支出逐层缩小至可服务的自主 Kubernetes 优化市场。
各层宽度仅表示层级,不代表数值比例。
规模估算层级将经验得出的 69% 过度配置率和利用率遥测,应用于企业容器基础设施预测。
[CM001, CM002, CM004, CM015]2.3 买方分层、用户画像与采购动态
自主 Kubernetes 优化的商业采用遵循分层的企业采购路径,由不同的技术和财务利益相关方共同决定。主要技术倡导者和预算负责人通常是平台工程副总裁、云基础设施负责人或站点可靠性工程(SRE)总监,他们对应用可用性和基础设施预算治理负责。在 AI 原生组织中,MLOps 和 AI 基础设施负责人越来越多地推动采购,以缓解严重的 GPU 可用性瓶颈,并在不牺牲推理延迟的情况下压低集群支出。技术团队优先考虑无扰动执行和与现有 GitOps 流水线的集成;企业采购与 FinOps 团队则审视定价透明度和多云覆盖。云市场、SoftwareOne 等软件分销商等商业渠道,是加速企业签约和账单整合的重要采购路径。[CM007, CM011, CM013, CM014]
| 细分市场 | 目标买方 | 主要用户 | 经济付款方 | 运营流程 | 预算负责人 | 采用触发因素 |
|---|---|---|---|---|---|---|
| 高增长 AI / 科技规模化企业 | 工程副总裁 / CTO | 平台工程师与 MLOps | 工程运营预算 | 自动化 GPU 集群放置与调配 | 工程副总裁 | 大规模 LLM 训练或推理上线后云账单飙升 |
| 企业 Kubernetes 用户 | 云基础设施负责人 | SRE 与 DevOps 团队 | 集中式云基础设施预算 | 用自主配置替代旧式节点自动扩缩容器 | 基础设施副总裁 | 集群无序扩张,应用频繁因内存不足崩溃 |
| 受监管金融与企业客户 | 首席信息官 / IT 副总裁 | 系统管理员与 SRE | IT 运营 / FinOps 预算 | 自动打补丁、节点生命周期升级和审计日志 | CIO / 企业 IT 副总裁 | 合规要求及时打补丁并保留基础设施审计轨迹 |
| 注重成本的数字原生中小企业 | DevOps 总监 / 技术负责人 | DevOps 工程师 | 通用技术预算 | Spot 实例自主再平衡和装箱 | 技术总监 | 董事会或投资者要求削减现金消耗并扩大毛利率 |
描绘主要目标细分市场中的企业角色、决策流程和商业触发因素。
[CM003, CM007, CM013, CM014]2.4 结构性增长驱动与关键采用约束
自主优化的市场动能来自强劲的结构性催化剂,但也受到明显的企业采用约束。首要加速因素是高成本 AI 加速器快速普及,以及 ARM Graviton 等替代架构结构性增长;在现代机群中,后者增速是 x86 核心的 3.5 times。同时,自动资源调整带来反直觉的可靠性红利:它会主动为受限 Pod 增加内存,同时缩减过度配置的节点,从而消除内存不足驱逐。不过,运营摩擦和架构锁定会压制采用:Cast AI 替换的是原生集群自动扩缩器,而不只是提供建议,因此企业平台团队要承担可观的迁移切换成本。此外,需要定制销售报价而非透明按单位定价的专有定价结构,也会在企业评估时增加预算摩擦。[CM003, CM005, CM006, CM007, CM008, CM010]
2.5 展示项
03竞争格局
3.1 竞争格局与市场分层
云成本优化和基础设施自动化市场呈现鲜明的架构分化,而非各供应商之间的功能同质化。管理容器化基础设施的组织面对一个碎片化生态,其中包括纯财务可视化平台、云账单承诺经纪商、专门的 Spot 编排器和集群内自主资源调整工具。在这一格局中,Cast AI 瞄准运行异构动态工作负载的生产 Kubernetes 集群;手动资源调整无法跟上需求变化。生产环境中,未优化的 Kubernetes 部署存在大量资源浪费,实测 CPU 平均利用率仅 8%,内存约 20%,意味着企业资产中存在显著的潜在低效。 竞争者从完全不同的切入点处理这些浪费。ProsperOps 完全位于云账单层,借助 IAM 权限运行无代理架构,动态交易 AWS Reserved Instances 和 Savings Plans,但完全不触及底层 Kubernetes Pod 的过度配置。相比之下,Kubecost 以 Helm chart 部署到集群内,接入 Prometheus 指标和云账单 API,提供按命名空间和 Pod 划分的细粒度成本分配及分摊报告。但 Kubecost 主要仍是可视化和治理工具,依赖工程团队手动查看仪表盘、提交工单和重新配置部署清单。Spot by NetApp 靠预测 Spot 实例中断来处理工作负载执行,但重点是 Spot 生命周期管理,而非整体 Pod 调整和 GPU 分时。此外,SoftwareOne Marketplace 等企业分销渠道已成为关键采购通路,把 Cast AI 的自动资源调整和安全扫描打包进统一的企业软件协议。[CP001, CP004, CP005, CP009, CP012, CP013]
| 竞争对手 | 类别 | 规模 / 融资 | 目标细分市场 | 核心差异化 | 主要局限 |
|---|---|---|---|---|---|
| Cast AI | K8s 自主优化器 | 已募资 $180M+(估值 $1B+) | 中大型多云 K8s 团队 | 实时自动调配和 GPU 时间共享 | 需要 K8s 基础设施;对非 K8s 场景支持有限 |
| Kubecost / OpenCost | 成本可见性与分摊 | 已募资 $20M+ / CNCF 项目 | DevOps 与平台工程 | 深度命名空间和 Pod 级成本分配 | 仅提供可见性;缺乏自主修复动作 |
| ProsperOps | 承诺管理 | 规模未披露 | 以 AWS 为主的企业用户 | RI 和 Savings Plans 自主优化 | 仅聚焦 AWS 计算承诺;不做 K8s 调配 |
| Spot by NetApp(Ocean)平台 | Spot 与工作负载自动化 | 被 NetApp 收购($450M) | 无状态和容器化工作负载 | 成熟的 Spot 预测及 Ocean 调度器集成 | 更聚焦 Spot 实例,而非原生 K8s Pod 调配 |
| ScaleOps / PerfectScale | Pod 级调配器 | 获风险投资支持的 Series A/B | 生产环境 K8s 平台团队 | 实时 Pod 资源建议与调优 | 生态规模较小,跨云 GPU 池化能力有限 |
| AWS Karpenter / Native HPA/VPA | 开源与云原生 | 超大规模云内置 | 所有云原生从业者 | 原生节点配置,且无许可费用 | 需要大量手工拼接策略和运营调优 |
数据综合截至 2026 年的从业者评价、供应商披露和市场列表;融资及估值数字反映最近报告的轮次。
[CP004, CP005, CP006, CP008, CP009, CP010, CP012]3.2 技术能力与架构差异化
Cast AI 的核心技术壁垒在于打通运行时 Kubernetes 工作负载遥测与自主基础设施配置之间的反馈回路。它不只是提醒工程团队存在过度配置的 Pod,而是安装主动代理,以程序方式调整 CPU 和内存上限、通过智能装箱整合工作负载,并在 AWS、Azure 和 Google Cloud Platform 上配置优化后的计算实例。独立基准和已披露的部署数据表明,自动资源调整可消除组织约 50% 的已配置 CPU 规模,扭转集群膨胀逐年加剧的趋势。 不过,平台最深的技术差异化在于专用 GPU 编排能力。在企业 Kubernetes 集群中,GPU 利用率平均仅 5%,约 95% 的稀缺昂贵加速器容量处于闲置。通用自动扩缩器和传统 FinOps 工具把 GPU 节点当作普通 CPU 实例处理;Cast AI 则整合 GPU 分时、Multi-Instance GPU(MIG)、Multi-Process Service(MPS)和 Kubernetes Dynamic Resource Allocation(DRA)等高级硬件共享机制。这种细粒度硬件虚拟化让多个 AI 推理工作负载安全共享物理卡,而不是为突发任务独占整块 $30,000 加速器。客户迁移的真实验证——例如 ALLEN Digital 将七个机器学习模型从 Amazon SageMaker 迁移到 Cast AI——确认成本下降 71%,推理延迟也更低。此外,跨区域自动放置带来显著套利空间:为 A100 GPU 工作负载选择有利云区域,可将计算成本最多降低 80%,相较当前 Spot 实例平均价格。[CP002, CP006, CP007, CP009, CP014, CP015]
| 采购标准 | Cast AI | Kubecost | ProsperOps | Spot by NetApp 平台 | AWS Karpenter |
|---|---|---|---|---|---|
| 自动 Pod 调配 | 完全自主 | 手工建议 | 不支持 | 自动化(Ocean) | 通过 VPA 支持(手工配置) |
| GPU 共享(MIG/MPS/装箱) | 原生自动化 | 仅可见性 | 不支持 | 部分支持 / 新兴 | 手工 DRA / 设备插件 |
| 多云覆盖 | AWS、Azure、GCP | AWS、Azure、GCP | 仅 AWS | AWS、Azure、GCP | 仅 AWS(Karpenter 核心) |
| 成本分配与分摊 | 集群与命名空间 | 深度 Pod/命名空间/标签 | 计费 / 账户级 | 工作负载级 | 仅 AWS 成本类别 |
| 部署形态 | 集群内代理 | Helm Chart 代理 | 无代理(IAM 权限) | 集群内代理 / SaaS | 集群内控制器 |
评估依据 2026 年发布的功能集和技术文档;不支持及部分支持的能力反映架构边界。
[CP004, CP005, CP006, CP009, CP012]按执行自主性和 Kubernetes 工作负载深度,定位云成本及基础设施优化方案。
| 条目 | X (2.8–8.8) | Y (3.5–8.5) | 说明 |
|---|---|---|---|
| 1. Cast AI | 8.8 | 8.5 | 在 AWS、Azure 和 GCP 上自主调整集群内资源并实现 GPU 共享 |
| 2. Kubecost | 8.2 | 3.5 | 深入展示 Kubernetes 成本并进行命名空间分摊,但工作流仅提供建议 |
| 3. ProsperOps | 2.8 | 8.2 | 完全自主管理 AWS 承诺和 Savings Plans,无需部署集群内代理 |
| 4. Spot by NetApp 平台 | 6.8 | 7.8 | 成熟的 Spot 中断预测及跨主要云厂商的容器工作负载扩展 |
| 5. ScaleOps / PerfectScale | 7.5 | 6.8 | 面向生产 Kubernetes 环境的动态 Pod 级资源调优和装箱优化 |
| 6. AWS Karpenter 与原生工具 | 5.2 | 4 | 开源及超大规模云厂商原生集群自动扩缩容,需要手动配置策略和阈值 |
坐标反映基于架构分析和 2026 年从业者评审的定性序数定位;刻度已归一化至 0 至 10。
[CP001, CP004, CP005, CP006, CP009, CP010, CP012]3.3 定价结构与 GTM 分销
云成本管理领域的商业打包,反映出供应商信心和客户 ROI 对齐程度的差异。传统企业 FinOps 平台按固定年度平台许可收费,并随未优化云支出总额扩大;这会形成错位激励,因为客户浪费越多,供应商收入越高。现代优化平台则越来越多采用与绩效挂钩或分层的订阅定价。ProsperOps 首创纯绩效模式,收取经核验 AWS 承诺节省额的 10% 至 15%,让客户免于前期财务风险,但范围严格限于财务折扣。 Cast AI 采用灵活商业策略:入门级集群订阅约 $200 /月起,高阶企业套餐则与经核验的基础设施成本下降挂钩。这种混合结构降低了中型市场工程团队的进入门槛,同时捕捉运行数百万美元计算预算的大型多集群企业部署所带来的上行空间。软件费用直接绑定已实现节省额,也使 Cast AI 与企业 FinOps 目标保持一致。融资与战略分销上,Cast AI 于 2026 年初从 Pacific Alliance Ventures 获得融资,估值超过 $1 billion。这笔资金加速了与 Oracle Cloud Infrastructure 合作推出 Omni Compute,打造统一的跨云计算市场,让企业可无缝把 AI 工作负载突发调度到 Oracle GPU 集群,从而绕开超大规模云厂商的容量约束。[CP004, CP005, CP008, CP011, CP013]
| 供应商 | 定价模式 | 入门价格 | 包含能力 | 商业权衡 |
|---|---|---|---|---|
| Cast AI | 混合 SaaS + 节省额百分比 | 约 $200/月,或按经核实节省额的百分比计费 | 自动扩缩容、GPU 优化、调配、安全 | ROI 高度一致,但月度支出随实际节省额波动 |
| Kubecost | 开放核心免费增值 | 免费层;企业版约 $1,000/月 | 实时成本监控、预算告警、分摊 | 软件成本可预测,但工程团队必须手动实现节省 |
| ProsperOps | 按绩效计费 | 经核实 AWS 节省额的 10-15% | 24/7 全天候自主进行 RI/Savings Plans 交易 | 财务风险为零,但解决的是承诺折扣,而不是浪费问题 |
| NetApp 的 Spot | 节省额分成 / SaaS | 定制企业合同 | Elastigroup Spot 故障转移、Ocean 容器扩缩容 | 企业合同体系成熟,但要求作出承诺 |
| AWS Karpenter / 开源 | 免费 / 开源 | 无软件费用($0) | 动态节点规格调整、EC2 上的装箱 | 无厂商许可费,但会把大量工程时间转移给内部运维 |
定价结构反映 2026 年公开的厂商报价和从业者比较;实际折扣会随支出规模和合同期限变化。
[CP004, CP005, CP009, CP011, CP012]3.4 壁垒持久性与超大规模云厂商商品化风险
评估 Cast AI 的竞争壁垒,必须审视结构性脆弱点以及公有云厂商激进的产品路线。首先,Cast AI 的价值主张根本上依附 Kubernetes 这一底层基座。对于运行单体传统应用、裸虚拟机、无服务器架构,或 Databricks、SageMaker 等托管 AI 平台的组织,Cast AI 用处很小。第三方从业者评测强调,工作负载可预测且静态的小型组织,从自主资源调整中获得的回报更低,平台的运营和财务开销因此难以证明合理。 其次,超大规模云厂商商品化构成持续的结构性逆风。云厂商不断增强原生计算管理工具,引入 AWS Karpenter 等开源控制器,以及 GKE Autopilot、EKS Auto Mode 等全托管环境。Karpenter 当前仍需要大量手工策略配置,也缺乏原生多云仲裁和跨云 GPU 池化;但超大规模云厂商拥有巨大的分销优势和零许可成本吸引力。如果云厂商开发出开箱即用、无摩擦的原生垂直 Pod 自动扩缩和 GPU 分时,客户部署第三方集群内代理的意愿可能下降。此外,目前仅 2% 的 Kubernetes GPU 工作负载使用 Spot 实例,企业普遍不愿采用可抢占基础设施,也为 Cast AI 的自动 Spot 故障转移算法设置采用障碍。Cast AI 的防线在于守住多云可互换性和供应商中立编排的技术领先,而超大规模云厂商在商业上缺乏合作的内在动力。[CP002, CP003, CP008, CP009, CP010]
3.5 展示项
04财务
4.1 收入模式与定价
公开证据显示其商业结构不止一种,而非统一价目表。一份独立供应商评测称,年度平台订阅的典型年度合同额为 $50,000 至超过 $1 million,部分绩效挂钩安排收取经核验节省额的 15% 至 25%。另一份竞争对手比较则称其采用使用量定价,由固定基础费用加每 CPU 费率构成。这些观察说明的是可能的计费机制,而非 Cast AI 的实际价格、收入确认政策、折扣、合同期限或组合。指定证据中没有当前价目表,因此订阅可预测性和收益分成波动仍需单独纳入投资判断。[CI003, CI004, CI005]
| 收入来源 | 机制 | 计费单位 | 公开状态 | 收入质量判断 | 尽调要求 |
|---|---|---|---|---|---|
| 年度平台订阅 | 合同约定的软件访问权 | 与集群支出挂钩的年度合同价值 | 据第三方报告,典型 ACV 为 $50K-$1M+ | 可能具备经常性;暂无续约证据 | 按队列拆分的 ARR、合同期限、续约情况和终止权 |
| 与绩效挂钩的定价 | 经核实客户节省额的分成 | 选定项目中经核实节省额的 15%-25% | 据独立厂商评测报告 | 与结果挂钩,但具有波动性且受基线影响 | 收入构成、节省基线政策、结算调整和争议 |
| 与算力挂钩的用量定价 | 基础费用加用量费 | 按 CPU 计费 | 竞争对手有此描述;未留存公开价目表 | 随用量增长,但可预测性可能较弱 | 发票样本、最低承诺、上限和折扣表 |
| GPU 优化能力 | 自动配置和共享 GPU 实例 | 变现单位未公开披露 | 产品能力已有记录;收入贡献未知 | 提供 SKU 和附加率证据前,无法验证 | SKU 定价、附加率、GPU 相关收入和交付成本 |
公开机制和区间来自报价单或市场观察,不代表实际定价或经审计的收入构成;未知字段需要管理层证据。
[CI003, CI004, CI005, CI014]| 证据项目 | 报告值 | 证据类别 | 支持的判断 | 局限 |
|---|---|---|---|---|
| 典型年度订阅 | $50K-$1M+ ACV | 独立厂商评测 | 企业软件的付费意愿 | 不是公司价目表,也不是实际平均价格 |
| 选定项目的节省额分成费 | 经核实节省额的 15%-25% | 独立厂商评测 | 与结果挂钩的变现 | 选择标准和核实条款属于私有信息 |
| Akamai 节省案例 | 云支出节省 40%-70% | 公司筛选的客户案例 | 潜在客户 ROI | 一个重点案例不能代表整个客户组合的平均水平 |
| GPU 基线利用率 | 平均 5% | 公司研究,有新闻报道佐证 | 可触达的浪费池规模很大 | 基线利用率不等于实际客户节省额 |
数值保留原始样本范围和证据类别;客户节省额与基线利用率都不是 Cast AI 的收入或利润率。
[CI003, CI004, CI009, CI010]展示从客户资源浪费到 Cast AI 变现的、有证据支持的路径,终点是尚未解决的毛利润环节。
| 节点 / 连接 | 节点 / 起点 | 终点 | 说明 |
|---|---|---|---|
| 节点 1 | 利用不足的 Kubernetes 资源池 | 报道的自动化前平均值包括 GPU 5%、CPU 8% 和内存 20% 的利用率。 | |
| 节点 2 | 工作负载与成本信号 | Cast AI 观测工作负载、基础设施、成本及服务等级信号。 | |
| 节点 3 | 自动化优化动作 | 资源调优、节点扩缩容、Spot 优化和 GPU 配置共同作用于这一机会。 | |
| 节点 4 | 客户节省结果 | Cast AI 强调 40%–70% 的云成本节省案例,但未披露整体组合节省。 | |
| 节点 5 | 合同化变现 | 据报道,其收费结构包括年度订阅、与节省额挂钩的费用及与计算量挂钩的收费。 | |
| 节点 6 | 已实现收入;毛利润未知 | 第三方披露了宽泛的 2025 年收入区间,但收入构成、COGS 和毛利率仍不可得。 |
该流程表示有证据支持的机制,不代表转化率、收入确认时点或量化的毛利润瀑布。
[CI001, CI008, CI013, CI014, CI016]4.2 公开采用进展与销售效率代理指标
Cast AI 首页声称已有超过 2,100 家公司采用;一份独立供应商评测则给出 2025 年年度收入 $15 million 至 $42 million 的宽泛区间。后者既不是当前 ARR,也不是经审计收入,不应按任一指标使用。同一评测称买方通常在两到四周内部署平台,这可以作为实施摩擦代理指标,但不能证明销售周期、CAC、转化率或回本周期。Cast AI 还突出 Akamai 案例,称云成本节省 40% 至 70%。这一客户结果可能支持付费意愿,但属于公司筛选的证据,不能证明整体客户留存或扩张经济性。[CI001, CI006, CI007, CI009]
4.3 客户经济性与利润率路径
最清晰的经济证据在于客户的浪费池,而非 Cast AI 自身的单位经济性。Cast AI 报告称自动化前 GPU 平均利用率为 5%、CPU 为 8%、内存为 20%;Business Insider 独立重复了 5% GPU 和 8% CPU 数据,并指出样本涵盖 23,000 个集群。报告方法说明,这些观察是在客户启用 Cast AI 自动化前收集的,因此是机会基线,而不是实测节省。Cast AI 将自动资源调整、扩缩和 GPU 配置描述为干预措施。这些事实解释了闲置计算转化为客户节省的潜在价值桥梁,但没有透露 Cast AI 的托管、支持、销售或服务交付成本,因此不能证明毛利率。[CI008, CI010, CI011, CI012, CI013, CI014]
| 指标 | 公开值 | 置信度 | 重要性 | 具体尽调要求 |
|---|---|---|---|---|
| 年收入 | 2025 年 $15M-$42M | 中;第三方区间 | 只能提供宽泛的历史规模指标 | 经审计的 2025 年收入和 2026 年月度收入桥接表 |
| 客户数量 | 2,100+ 家公司 | 中;公司口径 | 说明客户 Logo 规模,但不能说明客户经济性 | 按 ACV 档位统计的付费客户数量,以及前 20 大客户集中度 |
| 当前 ARR 与增长 | 无法获得 | 需要据此评估经常性规模和增长势头 | 当前 ARR、确认收入、订单额和同比增长 | |
| 毛利率 | 无法获得 | 决定扣除托管和服务交付成本后的可扩展性 | 按订阅、节省额分成、支持和 GPU 产品拆分的收入与 COGS | |
| NRR 与 CAC 回本周期 | 无法获得 | 检验留存、扩张和销售效率 | 客户队列 NRR、总留存率、CAC、销售周期和回本周期 | |
| 现金、烧钱速度、现金跑道与债务 | 无法获得 | 已融资金额无法判断资本充足性 | 现金余额、月度净烧钱、债务计划、现金跑道方案和下一轮融资触发条件 |
Null 表示指定的公开证据未披露该指标;每个 null 都有对应的管理层尽调要求,不得根据融资额或客户数量估算。
[CI001, CI002, CI006, CI016]4.4 资本充足性与财务结论
这一证据池中最新的量化融资证据,是第三方关于 2025 年 4 月 $108 million 成长轮和累计融资超过 $190 million 的报道。这段历史证明公司能获得外部资本,但已募资金不等于可用现金;没有现金、消耗、债务和营运资金数据,也无法证明 runway。如果年度订阅占主导,收入质量可能较有吸引力;但绩效挂钩和计算挂钩费用会随客户使用量及经核验节省额波动。一名竞争对手还认为 Cast AI 在 Kubernetes 之外没有覆盖完整成本,这可能限制钱包份额。因此财务结论是有希望但尚未完成投资判断:当前收入质量、利润率路径、销售效率、客户集中度和未来资本需求仍需私有尽调。[CI002, CI005, CI015, CI016]
4.5 展示项
05产品与技术
5.1 产品定义与客户工作流
Cast AI 提供自动化云基础设施优化平台,旨在消除企业 Kubernetes 环境中的资源浪费。在标准云原生工作流中,工程团队根据保守估计配置计算容量,为避免性能限流而严重过度配置。23,000 个企业集群的实际遥测凸显了低效规模:组织通常配置的 GPU 容量是实际使用量的 20 times,GPU 利用率约 5%,CPU 利用率 8%。Cast AI 通过在数分钟内部署集群内代理改变这一模式;代理先以不干扰的只读状态运行,用户随后才启用自动闭环优化。平台持续监测实际资源消耗,在毫核心级动态调整容器规模,并在无需工程团队手工介入或重构代码的情况下重新平衡节点机群。[CE001, CE002, CE003, CE005, CE011]
| 用户任务 | 当前人工工作流 | Cast AI 自动化 | 可量化收益 | 运营限制 |
|---|---|---|---|---|
| 容器资源规格调整 | 工程师根据峰值猜测设置静态 CPU / 内存请求和上限 | Autopilot 根据实时用量在毫核级别动态调整容器规格 | 在不降低响应延迟的情况下消除 20x 资源过度配置 | 启用生产环境自主运行前需要审核初始配置 |
| 节点集群扩缩容 | 静态节点池,或 Cluster Autoscaler 无协调的扩容 | 在最便宜的实例类型之间动态配置算力并激进装箱 | 减少算力浪费,目标是缩小与 50% 健康利用率之间的差距 | 如果无人管理,工作负载重新平衡偶尔会中断活动连接 |
| Spot 实例生命周期 | 手动回退脚本在实例被驱逐时可能造成集群算力突然不足 | 预测性机器学习提前最多 30 分钟预测 Spot 终止 | 支持优雅地排空工作负载并提前补充节点 | Spot 可用性取决于云厂商各区域 Spot 资源池容量 |
| GPU 算力采购 | 年度预留承诺为对冲稀缺性而囤积闲置加速器 | 通过 Omni Compute 动态放置工作负载,并配置可互换的跨云 GPU | 在无需长期锁定的情况下,把平均 GPU 基线利用率从 5% 拉高 | 出口成本和数据管道瓶颈可能限制跨云 GPU 的效用 |
工作流和效率结果综合自 2026 年遥测报告、G2 客户评价和企业案例研究数据。
[CE002, CE003, CE004, CE005, CE006, CE009, CE011]5.2 运行架构与预测型 Cast Engine
Cast AI 的核心技术基础是闭环控制架构和专有预测引擎。Horizontal Pod Autoscaler、Vertical Pod Autoscaler 等原生 Kubernetes 扩缩原语依赖反应式阈值规则,而 Cast Engine 使用基于数千个真实企业集群遥测训练的预测性机器学习模型。其关键能力之一是 Spot 实例终止预测:可提前最多 30 分钟预测云厂商驱逐 Spot 实例,并在节点被强制终止前协调 Pod 平滑撤离。架构不止支持单集群资源调整,还通过 Omni Compute 延展计算弹性,将外部计算和 GPU 容量(如 Oracle Cloud Infrastructure 的高性能集群)作为原生资源接入,无需修改配置。这一抽象让工作负载能在多种多云和本地环境中执行,并由统一的 FinOps 和运营策略管理。[CE001, CE003, CE004, CE005, CE006, CE012, CE013]
| 架构层 | 平台角色 | 核心依赖 | 技术 / 运营风险 |
|---|---|---|---|
| 集群内只读代理 | 收集集群拓扑、资源请求、实时用量和事件 | Kubernetes API 服务器、轻量 DaemonSet / Deployment 权限 | 开销低,但需要初始安全审计和凭证访问审批 |
| 优化控制器 | 执行扩缩容、节点替换和 Pod 在线迁移 | 云厂商 IAM 凭证(EKS、AKS、GKE),用于管理实例生命周期 | 过于激进的重新平衡操作可能造成 Pod 短暂抖动 |
| Cast 预测引擎 | 分析应用用量特征并预测 Spot 中断 | 使用数千个企业集群的遥测数据训练的专有 ML 模型 | 模型架构闭源;边缘工作负载可能不符合历史启发式规则 |
| Omni Compute 抽象层 | 将外部算力和 GPU 实例代理为本地集群节点 | 跨云网络连接、供应商 API 合作关系(如 Oracle OCI) | 跨云网络延迟和供应商 API 速率限制可能影响调度 |
架构角色、依赖和风险因素来自 Cast AI 文档及独立从业者评估。
[CE001, CE003, CE004, CE006, CE009, CE012, CE013]以层级堆栈展示 Cast AI 平台架构,从集群遥测一路到自主执行及多云计算联邦。
- 工作负载接入与传感层
- 轻量级集群内代理实时观测 CPU、内存、GPU 和 SLO 遥测,无需修改代码。
- 遥测与成本分析引擎
- 汇总集群指标,评估命名空间支出、分配效率及闲置容量基准。
- 预测智能核心
- Cast Engine 应用启发式和 ML 模型,提前最多 30 分钟预测 Spot 中断。
- 自主优化与资源调优
- 自动完成毫核容器资源调优、Pod 装箱以及 Horizontal/Vertical Pod Autoscaler 的动态协调。
- Omni Compute 与多云经纪层
- 借助 Cast AI Anywhere,在 AWS EKS、GCP GKE、Azure AKS、OCI 和本地环境之间编排实例。
- 安全与治理护栏
- 执行代理式运行手册、错误配置审计、容器漏洞扫描和 FinOps 策略控制。
基于技术文档和产品公告的概念性功能分层表示。
[CE001, CE003, CE004, CE005, CE007, CE012]5.3 产品模块、SKU 与工具集成
平台以一体化运营套件商业化,覆盖四大能力:自主成本优化、实时成本监控、容器安全态势管理,以及专用 AI/LLM 成本优化。成本优化引擎聚焦容器资源调整、自动 Spot 实例管理和装箱算法,把工作负载集中到更具成本效益的实例族。监控模块同时提供跨集群、命名空间和工程团队的 CPU、GPU、内存及存储支出细粒度可视化。安全能力持续审计集群配置漂移、漏洞和暴露的攻击面。为无缝接入现代 DevOps 工作流,Cast AI 原生集成 Terraform 基础设施即代码、Helm 部署打包、Prometheus 与 Grafana 指标可视化,以及 Jira 问题修复跟踪。[CE001, CE007, CE008, CE012, CE013]
| 模块 / SKU | 目标用户 | 成熟度 / 状态 | 核心差异化 | 尽调缺口 |
|---|---|---|---|---|
| 自主成本优化 | 平台 / DevOps 工程师 | 生产环境 / GA | 持续调整 Pod 规格、自动管理 Spot,以及节点装箱 | 未公开真实工作负载驱逐延迟的 SLA 基准 |
| Kubernetes 成本监控 | FinOps 与工程负责人 | 生产环境 / GA | 实时按集群、命名空间和工作负载分摊 CPU、GPU 与 RAM 支出 | 无法原生推导每笔业务交易的精细单位经济性 |
| Kubernetes 安全与态势 | 安全与 SRE 团队 | 生产环境 / GA | 持续配置审计、漏洞扫描和互联网暴露面跟踪 | 与专用云原生应用保护平台(CNAPP)存在重叠 |
| Omni Compute / GPU 市场 | AI / ML 基础设施团队 | 早期可用(2026) | 跨云 GPU 可互换性,以及与原生 OCI 和超大规模云厂商资源池的集成 | 对定价流动性、供应商折扣和多区域网络出口的可见性有限 |
模块状态、能力和缺口综合自官方文档、SoftwareOne 目录数据及 SiliconANGLE 截至 2026 年 10 月的公告。
[CE001, CE005, CE007, CE008, CE012, CE013]5.4 技术差异化、可靠性与运营风险
Cast AI 通过自主执行而非被动建议仪表盘实现差异化,直接用自动化 Kubernetes 操作改变云基础设施状态。不过,自主闭环优化带来需要谨慎尽调的运营和架构风险。客户评测显示,如果应用缺乏完善的 Pod Disruption Budget 或平滑终止处理,自动节点重新平衡偶尔会造成短暂工作负载中断。此外,初始接入需要充分验证策略边界和云 IAM 权限,组织才会放心在关键任务环境中启用自动执行。从经济角度看,Cast AI 按核心计价的模式会产生持续软件费用,在大型节点池中可能抵消部分基础设施节省。因此 CTO 必须在自动装箱已验证的效率提升,与组织对自动运营变更及软件许可开销的容忍度之间权衡。[CE004, CE006, CE009, CE010, CE011, CE014]
5.5 展示项
06客户
6.1 客户分层与买方画像
Cast AI 的商业客户主要分为多云企业平台团队、云原生规模化公司,以及在 Amazon Web Services(EKS)、Google Cloud Platform(GKE)和 Microsoft Azure(AKS)上管理大规模 Kubernetes 部署的 FinOps 从业者。主要经济买方通常是基础设施副总裁、平台工程负责人或云 FinOps 总监,职责是压低波动的云基础设施支出。技术评估和日常运营由站点可靠性工程师(SRE)及 DevOps 从业者推动,他们希望无需手写 YAML 就能实现集群稳定和工作负载调整。除面向科技与金融服务组织的直销外,Cast AI 还通过与 SoftwareOne 等企业 IT 服务商的渠道合作及市场分销协议扩大商业触达。平台服务全球超过 2,100 家组织,覆盖从对延迟敏感的金融交易系统到高吞吐机器学习模型推理流水线的各种工作负载。[CU001, CU002, CU004]
| 细分市场 | 买方与用户画像 | 目标用例 | 观察到的规模 | 战略价值与收入模式 | 尽调缺口 |
|---|---|---|---|---|---|
| 全球企业 | 基础设施副总裁、平台负责人 | 多云 Kubernetes 成本治理、Spot 自动化 | BMW、Cisco、Akamai | 通过直销获取高 ACV 多年期合同 | 未披露合同规模和事业部层面的渗透率 |
| 金融科技与金融服务 | 平台工程经理、FinOps 负责人 | 高可用、对延迟敏感的交易和银行工作负载 | Genial Investimentos | 生产集群规格调整和容器安全 | 与集成合作伙伴分成利润(例如 Delfia) |
| 云原生规模化企业 | DevOps 负责人、SRE 总监 | 持续调整集群规格、自动装箱 | Yotpo、2,100+ 家全球科技公司 | 按接入管理的算力分级收取订阅费 | 如果云预算收缩或趋稳,可能面临客户流失 |
| AI 与 ML 基础设施 | AI 平台工程师、MLOps 负责人 | GPU 集群规格调整、Omni Compute 外部算力 | 早期 GPU 集群客户(95% 基线闲置) | 在高端 GPU 稀缺之际,增长潜力高 | GPU 市场模式的持久性尚未得到验证 |
细分依据为公开案例研究、平台能力和合作伙伴列表;各细分市场的确切收入贡献属于私有信息。
[CU001, CU002, CU004]6.2 采用轨迹与遥测规模
Cast AI 的采用轨迹由专有遥测引擎进一步强化,该引擎持续观察数万个 Kubernetes 集群中的真实工作负载信号。公司 2026 年《Kubernetes 优化状况报告》汇集 AWS、Azure 和 Google Cloud 环境的基线遥测,显示自动化前普遍存在资源过度配置。关键发现是企业 GPU 容量平均利用率仅 5%,约 95% 的已配置 GPU 计算处于闲置,却持续产生大量成本。同时,CPU 平均利用率仅约 8%,说明传统容器扩缩机制跟不上动态企业需求。Cast AI 把自主代理闭环定位为即时补救机制,将广泛的集群观察转化为先落地、后扩张的销售对话。Akamai 等大型企业的部署正是这一轨迹的例证:启用自动成本分析和资源调整引擎后,数日内报告云支出下降 40% 至 70%。[CU006, CU007, CU011]
| 指标 | 数值 | 日期 | 来源 | 置信度 | 含义 | 缺失分母 |
|---|---|---|---|---|---|---|
| 全球客户基础 | 2,100+ 家公司 | 2026-10 | Cast AI 官方平台 | 中 | 在中型市场和企业市场实现广泛的国际渗透 | 付费层级与免费层级 / 评估集群之间的拆分 |
| 观测到的遥测规模 | 数万个集群 | 2026-04 | 2026 年 K8s 状况报告 | 高 | 大规模数据集驱动预测性机器学习规格调整引擎 | 纳入管理的算力总支出或 vCPU 数量 |
| 企业 GPU 利用率 | 平均利用率 5% | 2026-04 | Business Insider / Cast AI 遥测 | 高 | AI 集群的巨大低效为客户 ROI 叙事提供有力支撑 | 专用 GPU 节点相对于标准 CPU 集群的样本量 |
| 客户云支出节省 | 支出减少 40% 至 70% | 2026-10 | Akamai 与 Yotpo 案例披露 | 中 | 立竿见影的可量化 ROI 加快从评估到生产的转化 | 扣除 Cast AI 软件许可费后的净节省金额 |
指标来自官方披露、遥测报告和第三方媒体;客户队列留存率和流失率尚未披露。
[CU001, CU006, CU007, CU011]6.3 已命名企业部署与生产结果
经验证的生产案例和第三方新闻报道确认,Cast AI 已从最初的概念验证进入关键任务企业环境。BMW、网络集团 Cisco 等全球大型企业使用 Cast AI,自动化管理复杂容器足迹的云成本。在金融服务领域,巴西投资平台 Genial Investimentos 与集成伙伴 Delfia 部署 Cast AI,优化约 90% 工作负载运行于 Kubernetes 的环境。在 Genial 的主生产集群中,自动装箱和资源调整将平均 CPU 成本从 $5.88 降至 $5.40,在保持应用可用性的同时实现整体生产成本下降 40%。另一个非生产集群通过计划休眠和 Spot Instance 自动化,使 CPU 成本下降 80%,从每 CPU $13.29 降至 $5.29。这些部署验证了 Cast AI 关于算法优化能在持续替换实例期间保持工作负载韧性的主张。[CU003, CU005, CU008, CU009]
| 客户 | 细分市场 | 部署与用例 | 生产环境还是试点 | 报告结果 | 局限与风险 |
|---|---|---|---|---|---|
| Genial Investimentos | 金融科技 / 金融平台 | 每晚定时重新平衡和自动化 Spot 生命周期 | 生产环境(约 90% 的工作负载运行在 K8s 上) | 生产环境节省 40%($5.88 降至 $5.40/CPU);预发布环境节省 80% | 初始部署需要合作伙伴 Delfia 提供支持 |
| BMW | 汽车 / 全球企业 | 自动化 Kubernetes 成本优化和管理 | 生产环境 | 2026 年遥测分析引用的活跃企业客户 | 未披露具体事业部范围和年度合同价值 |
| Cisco | 网络与 IT 基础设施 | 多集群云成本优化和基础设施规格调整 | 生产环境 | Business Insider 引用的企业参考客户 | 未披露各业务部门的确切工作负载渗透率 |
| Akamai Technologies | 云与 CDN 基础设施 | 实时成本分析和自动化工作负载规格调整 | 生产环境 | 各集群实现云成本节省 40% 至 70% | 未披露长期合同价值和扩张条款 |
客户验证样本来自已核实的案例研究和新闻报道;合同价值和长期留存指标属于保密信息。
[CU003, CU005, CU008, CU009]6.4 扩张机制、渠道摩擦与持久性风险
Cast AI 的扩张循环通常从无摩擦的只读集群评估开始,几分钟内暴露潜在计算浪费,随后进入预发布自动化、生产环境实时资源调整,以及通过 Omni Compute 进行专用 GPU 优化。电商平台 Yotpo 等客户把部署从基础成本可视化扩展到自动 Spot 实例替换,在削减云支出 40% 的同时,让工程团队摆脱手工配置。尽管运营能力强,Cast AI 仍面临架构和商业逆风。CloudZero 的独立竞争分析指出,Cast AI 的成本分配模型严格局限于 Kubernetes 集群和 Pod,无法让组织在 Snowflake 等多服务数据平台之间分摊共享云支出,也无法按业务单元、团队或客户追踪成本。此外,由于 Cast AI 不披露队列留存率、毛收入留存率或客户集中度,潜在投资者无法验证快速的初始降本是否能形成持久的企业软件 ARR,亦无法排除云足迹完成资源调整后客户增长停滞的风险。[CU010, CU012, CU013, CU014]
从只读集群评估,依次推进到完整生产环境 Spot 自动化、Omni Compute GPU 扩张及企业许可。
[CU010, CU012, CU013]6.5 展示项
07风险
7.1 运营自主性、自动扩缩冲突与爆炸半径
自主 Kubernetes 优化把基础设施控制从人工工程团队转移给算法策略引擎,带来显著运营权衡。Cast AI 在集群内部署拥有管理权限的代理,可调整工作负载、驱逐 Pod,并动态配置或删除云计算节点。自动资源调整能快速消除闲置计算,但 Cast AI 与外部节点配置器并行运行,可能产生不可预测的竞态条件和冲突的扩缩决策。如果工程团队未配置完整的 Pod Disruption Budget,或有状态应用无法适应容器原地调整,激进装箱和 Pod 驱逐还可能打断敏感工作负载。组织必须执行严格护栏、分阶段上线模式和持续可观测性,控制运营爆炸半径。[CR001, CR002, CR006, CR010, CR012]
| 失效模式 | 可能性 | 严重程度 | 缓解措施成熟度 | 残余暴露 | 未解决缺口 |
|---|---|---|---|---|---|
| 自主节点扩缩容冲突 | 中等 | 高 | 必须确保节点自动扩缩容器的独占性 | 中等 | 跨非标准 Kubernetes 发行版的多集群策略编排 |
| 集群内代理凭证遭窃 | 低 | 严重 | 分阶段启用只读模式、ISO 27001、零 Secret 访问 | 低 | 验证租户侧针对自动修复操作的审计日志 |
| 整合期间工作负载驱逐中断 | 中等 | 中 | 原地调整、实时迁移并遵守 PDB | 低 | 传统有状态应用的工作负载韧性基准 |
| Spot 实例中断后的回退延迟 | 低 | 中 | 预测 Spot 中断并回退到按需实例 | 低 | 超大规模云厂商同步抽走 Spot 容量时的生产恢复时长 |
运营风险评估反映自动调整资源的权衡、代理权限及集群内中断场景。
[CR001, CR002, CR006, CR010, CR012]多维风险评估矩阵,按可能性、影响、缓解措施成熟度和残余严重程度,映射运营、合作伙伴、监管及财务风险敞口。
[CR001, CR002, CR003, CR007, CR013]7.2 安全权限、遥测治理与合规态势
将主动第三方代理直接部署进生产 Kubernetes 控制平面,会带来关键的数据安全和监管治理问题。为降低客户风险,Cast AI 持有 ISO 27001 认证和经审计的 SOC 2 Type II 鉴证,证明其安全控制经过外部验证。平台声称代理不会检查或收集敏感应用载荷、Kubernetes Secrets 或 ConfigMaps,只依赖资源利用率指标和集群调度遥测。不过,在严格监管司法辖区运营的企业仍须评估跨境遥测流,执行最小权限的基于角色访问控制,并在授予自动代理执行权限前审查租户数据保留政策。[CR001, CR007]
| 风险 / 要求 | 司法辖区 | 当前状态 | 可能性 | 严重程度 | 缓释成熟度 | 残余风险敞口 | 尽调路径 |
|---|---|---|---|---|---|---|---|
| 数据隐私与跨境遥测敞口 | 全球 / 欧盟 / 美国 | 持续合规 | 低 | 高 | SOC 2 Type II、ISO 27001,且无法访问 Secret / ConfigMap | 中等 | 检查遥测 DPA 和欧洲租户隔离控制 |
| 云厂商市场许可合规 | 多司法辖区 | 标准渠道协议 | 低 | 中 | 标准分销商和 ISV 市场合同条款 | 低 | 确认转售商协议中的契约义务及责任转移 |
| AI 基础设施出口管制与硬件限制 | 美国 / 国际 | 贸易审查进行中 | 低 | 中 | 平台提供软件治理,而非实体芯片经纪服务 | 低 | 核实客户针对受限 GPU 实例的合规条款 |
| Kubernetes 知识产权与开源许可 | 全球 | Apache 2.0 生态标准 | 低 | 低 | 构建于标准 Kubernetes 和开放 API 之上 | Low | 审查专有二进制文件与代理层开源组件 |
监管分析基于已发布的 SOC 2 Type II 鉴证、ISO 27001 范围及企业数据处理要求。
[CR001, CR007]7.3 超大规模云平台依赖与渠道合作伙伴动态
Cast AI 是依赖 Amazon Web Services、Google Cloud Platform 和 Microsoft Azure 超大规模云基础设施的专用自动化覆盖层。平台高频查询云厂商定价 API,评估 Spot 实例可用性、实例族价格和承诺利用率。云账单 API 的破坏性修改、价格数据源限流,或 Spot 驱逐政策大幅改变,都会直接影响 Cast AI 的优化速度。同时,AWS Karpenter 等超大规模云原生工具带来长期去中介化风险:如果云厂商原生打包高级工作负载资源调整,Cast AI 就必须维持更强的多云策略协调能力,并通过 SoftwareOne 等伙伴保持分销触达。[CR006, CR008, CR014]
| 依赖项 | 交易对手方 | 角色 | 集中度 | 失效场景 | 严重程度 | 缓解措施 | 残余暴露 |
|---|---|---|---|---|---|---|---|
| 超大规模云厂商 API 与定价遥测接入 | AWS、Microsoft Azure、Google Cloud | 核心基础设施与定价源 | 高 | API 破坏性变更或持续价格查询遭限流 | 高 | 冗余的多云定价接入与 SDK 抽象层 | 中等 |
| 云转售商与渠道分销 | SoftwareOne 与云市场 | 企业采购与销售渠道 | 中等 | 渠道合作伙伴降低优先级或分成不利 | 中 | 直销企业客户,同时推进市场联合销售 | 低 |
| 超大规模云厂商原生竞争工具 | AWS Karpenter / Azure / GCP 原生成本工具 | 直接的基准平台竞争者 | 高 | 原生工具开箱即用地加入自主工作负载调优 | 高 | 跨云统一策略引擎与专用 GPU 优化支持 | 高 |
| 核心容器编排原语 | CNCF / Kubernetes Project | 底层计算运行时标准 | 严重 | Kubernetes 自动扩缩容规范发生重大架构变化 | 中 | 持续与上游 Kubernetes 保持一致并参与社区贡献 | 低 |
生态与超大规模云厂商依赖登记表,评估 API 集成脆弱性、云渠道摩擦及市场动态。
[CR006, CR008, CR014]7.4 财务模型可持续性、定价不透明与终止标准
Cast AI 的商业价值主张建立在企业基础设施浪费之上;2026 年数据表明,未优化集群的 GPU 平均利用率仅 5%,CPU 为 8%。但客户反馈显示定价摩擦明显,尤其是定制报价结构、小集群最低消费,以及使财务预测复杂化的节省额百分比计费机制。持续存在的尽调难题,是年度经常性收入、毛利率、资本消耗率和客户队列留存数据完全没有公开披露。投资尽调必须验证高理论节省能否转化为持久软件利润率,并围绕企业流失、定价阻力或超大规模云平台锁定设定明确的终止标准。[CR003, CR004, CR005, CR011, CR013]
7.5 展示项
08估值
8.1 投资论点与建议框架
我们对 Cast AI 的投资尽调给出审慎的「跟踪」建议:风险评级为中等,估值判断偏高。Cast AI 已在云基础设施优化和应用性能自动化领域展现出清晰的技术与商业势头,企业客户已超过 2,100 家,其中包括 BMW、Cisco 和 Akamai 等知名客户。公司的 Kubernetes 自动调配和多云配置平台,直击企业云环境中的严重低效:真实集群利用率数据显示,已配置的 GPU 算力有 95% 处于闲置状态,CPU 平均使用率仅约 8%。此外,主要超大规模云厂商近期将 Nvidia H200 实例价格上调 15%,进一步强化了 CFO 层面自主降本的紧迫性。不过,Cast AI 在 Pacific Alliance Ventures 于 2026 年 1 月进行战略投资后,账面投后估值已超过 $1.0 billion,但交易金额及 ARR、净收入留存率、GAAP 毛利率等经营财务指标仍未披露。在 $1.0 billion 估值下,Cast AI 需要更高的收入倍数;若没有经审计的财务证明,公有云和 FinOps 软件可比公司的估值并不足以支撑这一水平。因此,我们建议继续跟踪该资产,争取结构化条款或二级市场折价,而不是按当前账面估值新建股权仓位。[CV001, CV002, CV003, CV004, CV014]
| 维度 | 立场 / 评级 | 关键证据与驱动因素 | 决策含义 |
|---|---|---|---|
| 建议 | 跟踪 | 技术产品市场匹配度强,但估值经济性和 ARR 未披露 | 未核实经审计 ARR 前,不要按 $1B 的表面估值投入新资金 |
| 信心 | 中 | 已验证客户进展(2,100+ 家企业)和 $1B 投后估值,但财务数据为私有信息 | 足以跟踪并开展竞争基准比较,但不足以无保留买入 |
| 风险评级 | 中 | 平均 GPU 利用率仅 5% 带来企业紧迫性,但云厂商定价波动增加摩擦 | 监测企业 FinOps 预算及原生超大规模云厂商的竞争反应 |
| 估值立场 | 偏高 | $1B 估值意味着较高的远期收入倍数(估计 ARR 的 >20x) | 要求定价纪律;争取二级市场折价或估值重置 |
评级将 Series C 及 2026 年 1 月战略投资中已验证的客户进展,与 ARR 私有披露缺口进行综合比较。
[CV001, CV002, CV003, CV004, CV014]将企业规模信号和 GPU 优化顺风,与估值悬置及最终跟踪立场连接起来的结构化决策树。
| 节点 / 连接 | 节点 / 起点 | 终点 | 说明 |
|---|---|---|---|
| 节点 1 | 企业规模信号 [scale-signal] | 2,100+ 个企业账户及一线客户验证(BMW、Cisco、Akamai) | |
| 节点 2 | AI 基础设施浪费顺风 [market-tailwind] | 95% GPU 闲置容量及不断上升的 AI 基础设施成本,推动 CFO 层面的 FinOps 紧迫需求 | |
| 节点 3 | 自主控制平面 [product-differentiation] | 实时自动调整资源规格,并为多云 GPU 提供 OMNI Compute 配置 | |
| 节点 4 | 财务披露不透明 [financial-opacity] | ARR、净留存率及战略轮条款未披露,带来多重不确定性 | |
| 节点 5 | 估值悬压($1.0B 标价) [valuation-overhang] | 定价达到独角兽级别($1.0B),在缺乏审计支持的情况下仍需更高收入倍数 | |
| 节点 6 | 建议:持续跟踪 [final-recommendation] | 跟踪 ARR 披露,寻求有纪律的二级市场或结构化进入机会 | |
| 连接 1 | 企业规模信号 [scale-signal] | AI 基础设施浪费顺风 [market-tailwind] | |
| 连接 2 | AI 基础设施浪费顺风 [market-tailwind] | 自主控制平面 [product-differentiation] | |
| 连接 3 | 自主控制平面 [product-differentiation] | 财务披露不透明 [financial-opacity] | |
| 连接 4 | 财务披露不透明 [financial-opacity] | 估值悬压($1.0B 标价) [valuation-overhang] | |
| 连接 5 | 估值悬压($1.0B 标价) [valuation-overhang] | 建议:持续跟踪 [final-recommendation] |
流程图呈现定性闸门顺序;不代表算法评分权重。
[CV001, CV002, CV004, CV006, CV010, CV014]8.2 核心论点与反证性反论点
Cast AI 的核心投资论点建立在三根支柱上:企业降本痛点迫切、技术具备自主性优势,以及跨云 AI 的选择权。企业因人工智能 FOMO 囤积稀缺 GPU 资源,云基础设施预算承受巨大压力,集群因此严重超配,服务器实际 GPU 利用率往往只有 5%。第一代 FinOps 工具只能提供静态监控面板和人工工单建议;Cast AI 则直接在基础设施层自主运行,实时完成 Pod 调配、节点装箱和动态 Spot 实例管理。OMNI Compute 的推出进一步扩大了公司的可服务市场:它建立统一控制平面,可在不同云之间配置 GPU 算力,无需改动应用代码。反论点则指出了可能摧毁股权上行空间的关键执行和市场风险。预算疲软的企业可能选择缩减已承诺的 GPU 总容量,而不是购买专业的第三方多云编排软件。此外,AWS Karpenter、Azure 原生工具等超大规模云厂商的原生自动扩缩容能力仍在快速提升。如果原生调度器缩小效率差距,Cast AI 的定价能力可能减弱,尤其是在多云客户流失加速或软件分销商拿走越来越多经济收益的情况下。[CV002, CV003, CV005, CV006, CV012]
| 角度 | 投资正题论点 | 反题论据 | 哪些变化会改变判断 |
|---|---|---|---|
| 市场机会 | GPU 浪费率高达 95%,AI 基础设施成本上升使自主优化成为 CFO 的关键优先事项 | 企业可能削减 GPU 总配置,而不是购买第三方多云控制平面 | OMNI Compute 加速采用,并验证 GPU 成本节省超过 50% |
| 产品与技术 | 自主控制平面不止于被动监控,还能实时调整集群资源并进行多云 GPU 配置 | 超大规模云厂商原生工具(Karpenter、Azure Advisor)及竞争初创公司可能压缩定价能力 | 独立基准证明其效率持续领先原生自动扩缩容器 3x |
| 财务表现 | 管理层称公司接近盈利,2,100+ 个企业账户具备健康的单位经济性 | 实际 ARR、现金消耗、毛利率和 NRR 仍完全不公开,真实现金消耗倍数因而无法判断 | 披露经审计 ARR 超过 $45M,且净收入留存率超过 125% |
| 估值倍数 | SoftBank、G2VP 和 Pacific Alliance Ventures 以 $1.0B 估值验证平台领导地位 | 估值从 $300M(2023 年 11 月)升至 $1B(2026 年 1 月),增速超过已披露的运营规模 | 获得下轮融资保护,或按低于 12x NTM ARR 倍数进行二级交易 |
论点评估市场驱动因素、竞争持久性、私有利润率假设及回报敏感性。
[CV002, CV003, CV005, CV006, CV012]8.3 融资历史与私募估值倍数
Cast AI 自 2019 年成立以来完成 10 轮融资,累计募资约 $272 million,持续获得顶级机构和企业投资者支持。2025 年 4 月,G2 Venture Partners 和 SoftBank Vision Fund 2 领投 $108 million 的 Series C 轮融资,Aglaé Ventures、Creandum、Hedosophia 和 Cota Capital 跟投,公司资本轨迹由此明显加速,估值约 $900 million。2026 年 1 月,韩国 $50 billion 零售集团 Shinsegae Group 的企业创投部门 Pacific Alliance Ventures 进行了一笔未披露金额的战略投资,Cast AI 的投后估值正式突破 $1.0 billion。根据 Multiples VC 的私募市场数据,Cast AI 的估值较 Series C 价格上升 1.1x,较 2023 年 11 月 $300 million 的 Series B 估值增长超过 3.3x。管理层称公司单位经济健康、具备盈利运营能力,但 ARR 完全没有公开披露,给分析带来很大不确定性。公有基础设施软件可比公司远期收入交易倍数约为 8x 至 12x;除非 Cast AI 的 ARR 收入运行率已超过 $80 million 至 $100 million,否则其 $1.0 billion 估值隐含的收入倍数偏激进。[CV001, CV007, CV008, CV009, CV010, CV011]
8.4 情景估值、下行触发因素与退出准备度
为评估不同市场状态下的风险调整回报,我们建立乐观、基准和悲观三种运营情景。乐观情景(概率 25%)下,OMNI Compute 成为企业 AI 推理的标准独立算力层,ARR 加速超过 $80 million,软件毛利率达到 80%+;按 18x 远期收入倍数计算,企业价值为 $1.44 billion 至 $1.80 billion,为 IPO 或被超大规模云厂商、网络巨头战略收购铺路。基准情景(概率 50%)下,Cast AI 在核心 Kubernetes 优化流程中保持稳健增长,ARR 达到 $40 million 至 $50 million,但私募市场倍数收缩将估值限制在 $450 million 至 $600 million,相比 $1.0 billion 的账面入场价,回报有限。悲观情景(概率 25%)下,云客户削减闲置 GPU 承诺,原生工具将基础调配能力商品化,增长停滞在 $30 million ARR 以下,估值重置至 $180 million 至 $240 million,并触发清算优先权压力。在投资委员会重新考虑「买入」建议前,必须完成三项尽调:第一,全面核验经审计的 GAAP 财务报表,确认当前 ARR 和毛利率;第二,取得客户群组留存数据,证明毛流失率低于 8%、NRR 高于 120%;第三,对 Series C 和战略轮的优先权结构进行法律审查。[CV001, CV008, CV010, CV013, CV014]
| 情景 | 运营与增长假设 | 目标倍数与估值逻辑 | 下行触发因素与风险 | 概率信号 |
|---|---|---|---|---|
| 乐观情景 | OMNI Compute 主导跨云 GPU 配置;ARR 在毛利率 80%+ 的情况下增长至 $80M 以上;账户数达到 2,500+ | 18x NTM ARR 倍数对应 $1.44B 至 $1.80B 企业价值;IPO 或被超大规模云厂商并购退出均具吸引力 | 超大规模云厂商限制 API、反锁定措施或原生工具激进跟价 | 25%(高上行选择权取决于 AI 推理规模) |
| 基准情景 | 核心 EKS/GKE 优化稳步扩张;ARR 达到 $40M–$50M;客户数 2,200+;现金消耗稳定 | 10x–12x NTM ARR 倍数支持 $450M 至 $600M 合理估值;普通股回报中等 | 企业采购周期放缓,专用 FinOps 厂商带来竞争压力 | 50%(反映当前增长轨迹及私募市场倍数压缩) |
| 悲观情景 | GPU 囤积泡沫破裂;企业转向原生 Karpenter/HPA;ARR 在 $30M 以下停滞 | 6x–8x NTM ARR 倍数带来 $180M 至 $240M 估值;Series C / 战略轮带来优先权 overhang | 下轮融资资本重组、客户流失及市场交易毛利率下滑 | 25%(若表面 $1B 估值遭遇倍数收缩,下行风险显著) |
估值区间将 ARR 估计值与公开基础设施软件及私募 FinOps 倍数进行比较。
[CV001, CV008, CV009, CV010, CV014]8.5 图表
免责声明
本报告是基于公开证据的尽调快照,不构成投资建议。重要的财务、法律、技术及合同事实仍未公开,任何投资决策前都应直接向管理层及通过一手文件核实。
证据索引
| 编号 | 陈述 | 可信度 | 来源 |
|---|---|---|---|
| CO001 | Cast AI was founded in 2019 in Miami, Florida, by Yuri Frayman, Leon Kuperman, and Laurent Gil to automate cloud-native Kubernetes infrastructure. | 高 | SO001, SO002, SO003 |
| CO002 | In January 2026, Cast AI secured strategic funding from Pacific Alliance Ventures, surpassing a $1 billion valuation to achieve unicorn status. | 高 | SO001, SO002, SO003 |
| CO003 | In January 2026, Cast AI launched OMNI Compute, a multi-cloud compute control plane and GPU marketplace partnered with Oracle Cloud Infrastructure. | 高 | SO001, SO002 |
| CO004 | Cast AI closed a $108 million Series C funding round in April 2025 led by G2 Venture Partners and SoftBank Vision Fund 2 at an estimated $900 million valuation. | 高 | SO001, SO002, SO003 |
| CO005 | Prior funding rounds included a $35 million Series B in November 2023 led by Vintage Investment Partners and a $20 million round in March 2023 led by Creandum. | 高 | SO003, SO006 |
| CO006 | Cast AI's cumulative lifetime equity capital raised totals over $180 million across six financing rounds through early 2026. | 高 | SO001, SO002, SO003 |
| CO007 | Yuri Frayman serves as Chief Executive Officer and co-founder, leading executive strategy and global expansion. | 高 | SO001, SO002 |
| CO008 | Laurent Gil serves as President and co-founder, overseeing commercial execution and product evangelism. | 高 | SO001, SO002, SO008 |
| CO009 | Cast AI reported doubling its enterprise customer base to over 2,100 organizations by 2024, with clients including BMW, Cisco, Akamai, and Hugging Face. | 高 | SO002, SO003, SO008 |
| CO010 | Cast AI's core automation engine combines pod rightsizing, node autoscaling, bin-packing, and Spot instance management across AWS, Azure, and Google Cloud. | 中 | SO004, SO009 |
| CO011 | Independent review by CyberNews identified drawbacks including advanced Kubernetes policy complexity, limited cost reporting granularity versus native tools, and higher pricing for smaller teams. | 中 | SO007 |
| CO012 | Cast AI telemetry published in its 2026 report indicates enterprise clusters average only 5% GPU utilization and 8% CPU utilization prior to automated optimization. | 高 | SO008, SO010 |
| CO013 | Cast AI differentiates from cost visibility tools such as Kubecost by executing autonomous infrastructure adjustments rather than passive recommendation dashboards. | 中 | SO005, SO007 |
| CO014 | Exact Annual Recurring Revenue, gross margin, net revenue retention, and GAAP operating profitability remain undisclosed private data points for Cast AI. | 中 | SO001, SO003 |
| CM001 | Production Kubernetes clusters sampled across major cloud providers exhibited an average CPU utilization of 8% and memory utilization of 20% in 2025. | 高 | SM001, SM002 |
| CM002 | Enterprise GPU utilization in non-optimized Kubernetes clusters averaged approximately 5%, leaving 95% of provisioned accelerator capacity idle. | 高 | SM001, SM009 |
| CM003 | Cloud provider pricing dynamics broke a historical cost reduction trend when AWS raised H200 Capacity Block prices by 15% in January 2026 citing supply scarcity. | 中 | SM001, SM002 |
| CM004 | Kubernetes CPU overprovisioning grew from 40% to 69% year over year as deployment teams systematically padded resource requests to prevent runtime throttling. | 中 | SM001, SM005 |
| CM005 | Less than 2% of cloud GPU capacity operated on interruptible Spot Instances in 2025 due to availability shortages and risk of long-running job disruption. | 中 | SM001 |
| CM006 | ARM-based CPU nodes expanded at 3.5 times the rate of x86 architecture between Q2 2024 and Q4 2025, reaching 9% of observed Kubernetes compute cores. | 中 | SM001 |
| CM007 | Automated workload rightsizing reduced provisioned CPU footprints by approximately 50% while decreasing out-of-memory container crashes from 40-50 per interval to near zero. | 中 | SM001 |
| CM008 | Autonomous node provisioning replaces native cluster autoscalers, introducing operational vendor lock-in that requires re-architecting infrastructure if decommissioned. | 中 | SM007, SM008 |
| CM009 | Alternative Kubernetes optimization architectures operate as advisory layers tuning pod resource requests without replacing platform node provisioners. | 中 | SM007 |
| CM010 | Cast AI commercializes its platform through private sales quoting rather than publishing fixed per-virtual-CPU price schedules. | 中 | SM007 |
| CM011 | Enterprise FinOps and cloud cost management buyers often look for commitment lifecycle management covering Reserved Instances and Savings Plans beyond cluster autoscaling. | 中 | SM008 |
| CM012 | Cast AI automates continuous bin-packing, rightsizing, and Spot instance lifecycle management across AWS, Azure, and Google Cloud Platform environments. | 中 | SM004, SM005 |
| CM013 | Independent market reporting ties Cast AI's valuation step-up and GPU marketplace launch to broader AI infrastructure optimization demand. | 中 | SM010 |
| CM014 | Cast AI was established by former Zenedge founders who built the autonomous performance platform in response to escalating enterprise cloud bills. | 中 | SM006 |
| CM015 | Independent multi-vendor industry market sizing reports isolating standalone autonomous Kubernetes optimization revenue remain unavailable publicly. | 中 | SM008 |
| CP001 | Production Kubernetes environments exhibit widespread overprovisioning, with average CPU utilization measured at 8% and memory utilization at 20% across unoptimized clusters. | 高 | SP001, SP007 |
| CP002 | Enterprise GPU utilization across managed Kubernetes services averages just 5%, indicating that approximately 95% of provisioned GPU compute capacity sits idle. | 中 | SP001, SP002, SP010 |
| CP003 | Only 2% of GPU workloads in Kubernetes clusters utilize discounted Spot Instances, leaving substantial infrastructure savings uncaptured. | 中 | SP001 |
| CP004 | ProsperOps specializes in agentless commitment automation for AWS Reserved Instances and Savings Plans, whereas CAST AI deploys an in-cluster agent for real-time Kubernetes scheduling and rightsizing. | 中 | SP003 |
| CP005 | Kubecost provides granular per-namespace cost visibility and chargeback reporting via open-core Helm deployment, but relies on manual execution rather than automated remediation. | 中 | SP003 |
| CP006 | Cast AI differentiates from generic cost-monitoring tools by combining GPU time-slicing, Multi-Instance GPU, Multi-Process Service, and bin packing to optimize hardware sharing. | 高 | SP002, SP005 |
| CP007 | ALLEN Digital consolidated seven AI models from Amazon SageMaker onto Cast AI's GPU stack, achieving 71% cost savings while maintaining comparable or lower inference latency. | 中 | SP002 |
| CP008 | Cast AI raised new capital from Pacific Alliance Ventures at a valuation exceeding $1 billion and partnered with Oracle Cloud Infrastructure to offer external GPU compute through Omni Compute. | 中 | SP009 |
| CP009 | Open-source built-in utilities such as Karpenter, Horizontal Pod Autoscaler, and Vertical Pod Autoscaler require substantial manual configuration and policy stitching across fragmented dashboards. | 高 | SP005, SP006 |
| CP010 | Cast AI requires an underlying Kubernetes environment and provides diminished return for small teams running static workloads or non-containerized virtual machines. | 中 | SP002, SP004 |
| CP011 | Cast AI offers paid commercial plans with entry pricing starting around $200 per month alongside value-based models taking a percentage of delivered cloud savings. | 中 | SP003, SP004 |
| CP012 | Spot by NetApp competes directly with Cast AI in cloud infrastructure optimization, using its Elastigroup and Ocean products to predict spot interruptions and manage container scheduling. | 中 | SP003 |
| CP013 | SoftwareOne distributes Cast AI via its marketplace to provide automated Kubernetes cost optimization and container security posture management for enterprise buyers. | 中 | SP008 |
| CP014 | Selecting favorable AWS regions for A100 GPU workloads can reduce GPU compute costs by up to 80% compared to average Spot Instance prices. | 中 | SP001 |
| CP015 | Autonomous rightsizing across Kubernetes clusters reduces provisioned CPU footprints by approximately 50%, countering year-over-year increases in overprovisioning. | 高 | SP001, SP007 |
| CI001 | An independent vendor review reports Cast AI's annual revenue in a $15 million to $42 million range as of 2025. | 中 | SI001 |
| CI002 | An independent vendor review reports a $108 million April 2025 growth round and more than $190 million of total funding for Cast AI. | 中 | SI001 |
| CI003 | An independent vendor review reports annual platform subscriptions typically ranging from $50,000 to more than $1 million in ACV depending on cluster spend. | 中 | SI001 |
| CI004 | An independent vendor review reports selected performance-linked contracts charging 15% to 25% of verified cost reductions against a baseline. | 中 | SI001 |
| CI005 | Competitor nOps describes Cast AI pricing as a fixed base fee plus a per-CPU usage charge. | 中 | SI006 |
| CI006 | Cast AI claims that more than 2,100 companies use its platform globally. | 中 | SI004 |
| CI007 | An independent vendor review reports that buyers typically deploy Cast AI in production within two to four weeks. | 中 | SI001 |
| CI008 | Cast AI describes automated workload optimization that rightsizes resources, scales nodes, and responds to changing demand. | 中 | SI004, SI005 |
| CI009 | Cast AI highlights an Akamai case study claiming 40% to 70% cloud savings. | 中 | SI004 |
| CI010 | Cast AI and Business Insider report average GPU utilization of 5% in the analyzed enterprise Kubernetes population. | 高 | SI002, SI008 |
| CI011 | Cast AI and Business Insider report average CPU utilization of 8% in the analyzed cluster population. | 中 | SI002, SI008 |
| CI012 | Cast AI reports average memory utilization of 20% in the analyzed clusters. | 中 | SI002 |
| CI013 | Cast AI says its utilization dataset was collected before the observed organizations enabled Cast AI automation. | 中 | SI003 |
| CI014 | Cast AI documentation says its autoscaler detects workloads requiring GPU resources and provisions GPU-optimized instances. | 中 | SI007 |
| CI015 | Competitor nOps says Cast AI does not provide full cost coverage for non-Kubernetes workloads. | 低 | SI006 |
| CI016 | Current ARR, revenue growth, gross margin, NRR, CAC payback, burn, runway, debt, and customer concentration remain unresolved in the assigned public evidence. | 低 | |
| CE001 | Cast AI operates an automated Kubernetes management platform that abstracts cloud infrastructure complexity across AWS, Azure, Google Cloud Platform, and hybrid environments via Cast AI Anywhere. | 高 | SE001, SE002 |
| CE002 | Cast AI's 2026 telemetry across 23,000 enterprise clusters found that organizations provision roughly 20 times more GPU capacity than consumed, averaging approximately 5% GPU utilization and 8% CPU utilization. | 高 | SE003, SE007, SE008 |
| CE003 | The platform deploys an in-cluster agent in minutes without requiring infrastructure re-architecting, beginning in a read-only monitoring posture before automated optimization is enabled. | 中 | SE001, SE002 |
| CE004 | Cast AI's predictive engine is designed to anticipate spot instance interruptions up to 30 minutes in advance, initiating graceful workload migrations before node terminations occur. | 中 | SE002 |
| CE005 | Autonomous rightsizing continuously adjusts CPU and memory at the millicore level to reduce overprovisioning while attempting to prevent application starvation and noisy-neighbor interference. | 中 | SE001, SE002 |
| CE006 | The platform integrates with native Kubernetes autoscaling primitives, extending Horizontal Pod Autoscaler and Vertical Pod Autoscaler mechanisms with automated node bin-packing and placement decisions. | 中 | SE002, SE006 |
| CE007 | In addition to compute optimization, the product suite includes Kubernetes security monitoring for misconfigurations and vulnerabilities, as well as LLM expense tracking and model selection optimization. | 中 | SE005 |
| CE008 | Software ecosystem integrations span infrastructure-as-code and observability tooling, including Terraform, Helm, Grafana, Prometheus, and Jira. | 中 | SE005 |
| CE009 | Enterprise practitioner feedback indicates that workload rebalancing activities can occasionally cause operational friction or unexpected behavior during live node reconfigurations. | 中 | SE006 |
| CE010 | Users report that while automated cost optimization reduces ongoing infrastructure maintenance, initial setup and policy configuration can require significant technical evaluation prior to production rollout. | 中 | SE006 |
| CE011 | Underutilization of cloud GPUs stems in part from structural orchestration bottlenecks, including primitive Kubernetes GPU scheduling, data pipeline latency, and rigid long-term reservation contracts. | 中 | SE007 |
| CE012 | Cast AI introduced Omni Compute to connect external GPU and compute capacity as native resources across clouds and on-premises environments without requiring code modifications. | 中 | SE009 |
| CE013 | Oracle Cloud Infrastructure partnered with Cast AI to make excess GPU capacity available directly to customers through the Omni Compute abstraction layer. | 中 | SE009 |
| CE014 | Customer reviews highlight that Cast AI's per-core pricing model can create substantial recurring software fees that offset a portion of realized cloud infrastructure savings. | 中 | SE006 |
| CU001 | Cast AI reports adoption by over 2,100 companies globally for Kubernetes workload automation. | 中 | SU005 |
| CU002 | Cast AI targets platform engineering, Site Reliability Engineering, and FinOps teams managing multi-cloud Kubernetes clusters. | 高 | SU005, SU008 |
| CU003 | Cast AI serves major global enterprises including automotive manufacturer BMW and networking technology leader Cisco. | 中 | SU003 |
| CU004 | Cast AI partners with IT solutions provider SoftwareOne to distribute its Kubernetes cost optimization and security platform through enterprise marketplace channels. | 中 | SU004 |
| CU005 | Brazilian investment platform Genial Investimentos deployed Cast AI across its Kubernetes clusters to achieve approximately 50% overall cloud computing savings. | 中 | SU002 |
| CU006 | Cast AI's 2026 telemetry report analyzed resource utilization across tens of thousands of Kubernetes clusters running on AWS, Microsoft Azure, and Google Cloud. | 高 | SU006, SU007 |
| CU007 | Telemetry collected by Cast AI revealed that average enterprise GPU utilization stood at only 5% across non-optimized Kubernetes clusters. | 高 | SU003, SU006 |
| CU008 | On its production cluster, Genial Investimentos achieved approximately 40% cost savings with average cost per CPU declining from $5.88 to $5.40. | 中 | SU002 |
| CU009 | Genial Investimentos achieved roughly 80% cost savings on its second cluster, lowering cost per CPU from $13.29 to $5.29 through automated rightsizing and rebalancing. | 中 | SU002 |
| CU010 | Digital experience platform Yotpo automated Spot Instance management using Cast AI to eliminate manual DevOps intervention and reduce cloud costs by 40%. | 中 | SU005 |
| CU011 | Content delivery and cloud provider Akamai reported 40% to 70% cloud cost savings following deployment of Cast AI's automated cost analytics. | 中 | SU005 |
| CU012 | Competitor analysis by CloudZero indicates that Cast AI restricts organizational cost allocation to cluster and pod boundaries rather than broader business dimensions like project, team, or product. | 中 | SU001 |
| CU013 | Cast AI does not consolidate non-Kubernetes SaaS infrastructure expenses or provide telemetry-driven shared cost splitting according to competitor benchmarks. | 中 | SU001 |
| CU014 | Cast AI does not publicly disclose net revenue retention, gross revenue retention, customer churn, or cohort renewal rates. | 低 | |
| CR001 | Cast AI requires in-cluster deployment and administrative permissions to alter workload sizing and dynamically provision or delete compute nodes. | 高 | SR001, SR007 |
| CR002 | Operating Cast AI concurrently with another node autoscaler creates conflicting scaling decisions and unpredictable cluster behavior unless Cast AI is granted exclusive autoscaler authority. | 中 | SR001 |
| CR003 | Enterprise criticism of Cast AI highlights small-cluster pricing friction, reporting depth limitations, documentation gaps, and advanced-policy complexity. | 中 | SR001 |
| CR004 | Cast AI's 2026 state of optimization research found enterprise GPU utilization averaging 5% and CPU utilization averaging 8% across unoptimized clusters. | 高 | SR003, SR004 |
| CR005 | Cast AI's utilization benchmark was compiled from 23,000 clusters from January 2025 through April 2026 across AWS, Google Cloud Platform, and Microsoft Azure. | 高 | SR004, SR006 |
| CR006 | Cast AI automates real-time instance selection, bin-packing, and pod rightsizing across major cloud platforms without requiring manual configuration tuning. | 中 | SR002, SR007 |
| CR007 | Cast AI holds ISO 27001 certification and SOC 2 Type II attestation, and asserts it does not access customer Kubernetes Secrets or ConfigMaps. | 中 | SR001 |
| CR008 | Cast AI is distributed across major enterprise software channels including SoftwareOne marketplace, AWS Marketplace, and cloud provider co-sell programs. | 中 | SR001, SR005 |
| CR009 | Cast AI claims its platform is trusted and deployed by more than 2,100 companies globally. | 中 | SR007 |
| CR010 | Cast AI supports in-place resource resizing on compatible Kubernetes versions to avoid pod restarts during workload rightsizing. | 中 | SR001, SR008 |
| CR011 | Idle GPU capacity generates substantial cloud waste because premium AI accelerators can cost up to 50 times more than comparable CPU instances. | 中 | SR004 |
| CR012 | Cast AI eviction and bin-packing automation can cause service disruption or pod recreation if applications lack robust Pod Disruption Budgets. | 中 | SR001 |
| CR013 | Cast AI relies on custom pricing quotes and savings-share commercial models that can create budget unpredictability for enterprise FinOps teams. | 中 | SR001 |
| CR014 | Cast AI depends on continuous hyperscaler telemetry ingestion and pricing APIs across AWS, GCP, and Azure for real-time node provisioning. | 高 | SR005, SR006, SR007 |
| CV001 | Cast AI reached a post-money valuation exceeding $1.0 billion in January 2026 following a strategic investment from Pacific Alliance Ventures. | 中 | SV002, SV003 |
| CV002 | Enterprise servers analyzed across thousands of clusters averaged 5% GPU utilization and 8% CPU utilization, indicating that roughly 95% of provisioned GPU capacity sits idle. | 高 | SV004, SV005 |
| CV003 | Enterprise GPU waste and rising AI infrastructure costs intensified the case for autonomous optimization in 2026. | 高 | SV004, SV005 |
| CV004 | Cast AI serves more than 2,100 enterprise organizations globally, including Fortune 500 accounts such as BMW, Cisco, and Akamai. | 高 | SV003, SV007 |
| CV005 | Long-term contract commitments and fear of GPU scarcity have driven enterprise organizations to overprovision compute capacity regardless of real-time workload requirements. | 中 | SV004 |
| CV006 | Cast AI introduced OMNI Compute as a unified compute control plane to discover and provision GPU resources across multiple clouds without requiring code changes. | 中 | SV002 |
| CV007 | The April 2025 Series C provided the last publicly priced valuation anchor before Cast AI's January 2026 strategic investment and unicorn step-up. | 中 | SV001, SV003 |
| CV008 | Multiples VC reports that Cast AI's valuation expanded 1.1x from $900 million in April 2025 to $1.0 billion in January 2026 across two priced funding rounds. | 中 | SV003 |
| CV009 | Cast AI's total lifetime equity funding reached approximately $272 million across ten financing rounds through 2026. | 中 | SV001 |
| CV010 | Cast AI does not publicly disclose its Annual Recurring Revenue (ARR), GAAP gross margins, net revenue retention, or cash burn rate. | 中 | SV003 |
| CV011 | Pacific Alliance Ventures is the corporate venture arm of South Korea's Shinsegae Group, a conglomerate with over $50 billion in commercial business interests. | 中 | SV002, SV003 |
| CV012 | Cast AI's platform provides real-time node lifecycle automation, pod rightsizing, and multi-cloud spot instance management across AWS, Azure, and Google Cloud. | 中 | SV007, SV008 |
| CV013 | SoftwareOne distributes Cast AI as an automated Kubernetes management platform targeting cost optimization, security monitoring, and resource bin-packing. | 中 | SV009 |
| CV014 | The diligence synthesis issues a 'track' recommendation with medium risk and stretched valuation stance due to the absence of audited ARR corroboration for the $1.0 billion headline valuation. | 中 | SV002, SV003, SV004 |
| 编号 | 出版方 | 标题 | 引文 |
|---|---|---|---|
| SO001 | SiliconANGLE | Cast AI raises funds from Pacific Alliance Ventures at $1B valuation to launch unified GPU marketplace | a startup that specializes in optimizing the efficiency of cloud-native environments, today announced it raised new funding from Pacific Alliance Ventures, surpassing a $1 billion valuation, to launch a unified cloud graphics processing unit marketplace. |
| SO002 | Business Wire | Cast AI Valued at Over $1 Billion With the Launch of Its GPU Marketplace | Cast AI, the leading Application Performance Automation platform, today introduced OMNI Compute, a unified compute control plane that automatically discovers available resources across cloud providers and regions and transparently extends existing Kubernetes clusters to consume them. The company also announced a strategic investment from Pacific Alliance Ventures (PAV), the U.S.-based corporate venture arm of Shinsegae Group, an over $50 billion Korean conglomerate with leading businesses across retail, consumer, and digital platforms. With this round of funding, Cast AI’s valuation exceeds $1 billion. |
| SO003 | Multiples.vc | CAST AI Valuation, Funding & Investors | Multiples | CAST AI Group Inc. is a Miami-based startup founded in 2019 that provides an application performance automation platform using machine learning to optimize Kubernetes environments in the cloud for cost savings, security, and efficiency through right-sizing resources, autoscaling, and spot instance management. |
| SO004 | Cast AI | Why Cast AI is the Best Tool for Kubernetes Cost Optimization in 2026 | Cast AI is a Kubernetes automation platform that reduces cloud costs by continuously optimizing workload resources and compute capacity across AWS, Google Cloud, and Azure. It combines pod rightsizing, node autoscaling, bin-packing, Spot instance management, and commitment optimization to reduce wasted capacity while maintaining workload performance and availability. |
| SO005 | Caplight | CAST AI Valuation & Funding History | Caplight | CAST AI Funding Rounds |
| SO006 | Cast AI | CAST AI Raises $35M Series B For Kubernetes Cost Optimization | Today, I’m excited to share that we’ve closed a $35M series B round from Vintage Investment Partners and existing investors Creandum and Uncorrelated Ventures. |
| SO007 | CyberNews | Cast AI Review 2026: Can This AI Really Cut Cloud Costs by 50%? | Advanced setup and policies can be difficult for teams new to Kubernetes ... Cost reporting granularity can be limited compared to native cloud tools ... Higher pricing, especially for smaller teams or clusters |
| SO008 | Business Insider | A New Cloud Report Finds That Companies Are Hoarding Unused AI Compute | Cast AI, an automated cloud cost optimization and management platform that works with companies like BMW and Cisco, published data in its 2026 State of Kubernetes Optimization Report showing that, on average, most organizations provide about 20 times more GPU capacity than they actively use at any given moment. |
| SO009 | SoftwareOne | Cast AI by CAST AI | Cast AI is designed to automate Kubernetes operations, focusing on cost optimization, security, and efficiency improvements. It helps companies reduce cloud expenses by optimizing resource allocation and automating tasks such as rightsizing and bin-packing. |
| SO010 | Cast AI | Cast AI’s 2026 State of Kubernetes Optimization Report | Cast AI, the leading automation platform, today released its 2026 State of Kubernetes Optimization Report, a comprehensive analysis of GPU, CPU, and memory utilization across non-optimized Kubernetes clusters. Drawing on data from tens of thousands of clusters, the report delivers a clear and urgent message: GPUs are poorly utilized at 5% despite their cost. |
| SM001 | Cast AI | 2026 State of Kubernetes Optimization Report | Average CPU utilization fell to 8% in 2025, down from 10% the prior year, while memory utilization declined from 23% to 20%. |
| SM002 | Cast AI | Cast AI's 2026 State of Kubernetes Optimization Report Reveals GPU Utilization at 5% | Average CPU utilization across clusters stood at just 8% in 2025, while memory utilization was 20%. |
| SM003 | Cast AI | Download the 2026 Kubernetes Optimization Report - Cast AI | This report is based on our analysis of tens of thousands of Kubernetes clusters across AWS, GCP, and Azure, covering January 1 to December 31, 2025, and until April 2026 for GPU utilization. |
| SM004 | Cast AI | Kubernetes Optimization Platform for Performance - Cast AI | Cast AI turns Kubernetes workload, infrastructure, cost, and SLO signals into safe automated actions: rightsizing pods, scaling nodes, optimizing GPUs and Spot, and fixing issues without manual tuning. |
| SM005 | Cast AI | Kubernetes Optimization for Reliability & Cost - Cast AI | Across thousands of clusters we analyzed, most Kubernetes spend is capacity that is requested, paid for, and never used. |
| SM006 | Cast AI | About Cast AI: The Application Performance Automation Platform | AWS, Azure, and Google Cloud integrations |
| SM007 | Kubex | Cast AI Alternative: Evaluation Across Six Dimensions | Cast AI replaces the Kubernetes Cluster Autoscaler entirely, so its own engine provisions nodes, selects instance types, and manages the fallback to spot instances (the discounted capacity a cloud provider can reclaim at short notice). |
| SM008 | nOps | Cast AI Alternatives: 11 Best Kubernetes Cost Optimization Tools | Cast AI focuses primarily on Kubernetes autoscaling and operates as a proprietary platform, which can introduce vendor lock-in and limit flexibility for teams seeking broader, multi-layer cloud cost optimization. |
| SM009 | Business Insider | Companies are hoarding AI compute because of FOMO — and they're sitting on most of it | According to data collected from 23,000 clusters across thousands of companies using Cast AI's AI agent, average GPU utilization across enterprise servers is at 5%, meaning roughly 95% of provisioned GPU capacity is not being used. |
| SM010 | SiliconANGLE | Cast AI raises funds from Pacific Alliance Ventures at $1B valuation to launch unified GPU marketplace | a startup that specializes in optimizing the efficiency of cloud-native environments, today announced it raised new funding from Pacific Alliance Ventures, surpassing a $1 billion valuation, to launch a unified cloud graphics processing unit marketplace. |
| SP001 | Conversational Geek Stats | What the Cast AI 2026 State of Kubernetes Optimization Report Says About Cloud Cost Optimization | Conversational Geek Stats | Average GPU utilization across Kubernetes clusters is only 5%, meaning organizations have roughly 20 times as much GPU capacity as their workloads consume at any given moment. |
| SP002 | Tapscape | Cast AI Review 2026: How It Tackles GPU Utilization for AI Workloads | Tapscape | Cast AI’s combination of time-slicing, MIG, MPS and bin packing is aimed directly at that 5% utilization figure and it’s the part of the platform that separates it from generic autoscaling or cost-monitoring tools that don’t touch GPU sharing at all. |
| SP003 | Bae Seokjae (Practitioner Blog) | AI Cloud Cost Optimization Tools 2026: ProsperOps vs CAST AI vs Kubecost Compared | ProsperOps is the top pick if you run significant AWS Reserved Instance or Savings Plans commitments, CAST AI wins for teams with complex Kubernetes workloads that need fully automated rightsizing, and Kubecost delivers the deepest cost visibility for engineering teams that want granular per-namespace or per-team chargeback without full automation lock-in. |
| SP004 | Cybernews | Cast AI Review 2026: Can This AI Really Cut Cloud Costs by 50%? | I found Cast AI best suited for larger companies looking to improve their efficiency, reduce waste, and maintain application performance across AWS, Azure, and GCP. |
| SP005 | Cast AI | Kubernetes Optimization Platform for Performance - Cast AI | Every workload has shifting CPU, memory, replica, latency, and cost needs. Karpenter, HPA, VPA, Spot, requests, limits, and node pools all help – but they still leave platform teams stitching together recommendations, tickets, dashboards, and manual YAML changes. |
| SP006 | Cast AI | Kubernetes Glossary | Avoiding overprovisioning in Kubernetes involves a multi-layered approach that addresses waste at the pod, node, and cluster levels. To build an efficient environment, you should combine built-in Kubernetes features with specialized third-party tools for visibility and automation. |
| SP007 | Cast AI | Cast AI's 2026 State of Kubernetes Optimization Report Reveals GPU Utilization at 5% | Average CPU utilization across clusters stood at just 8% in 2025, while memory utilization was 20%. |
| SP008 | SoftwareOne Marketplace | Cast AI by CAST AI | SoftwareOne Marketplace | Cast AI is designed to automate Kubernetes operations, focusing on cost optimization, security, and efficiency improvements. It helps companies reduce cloud expenses by optimizing resource allocation and automating tasks such as rightsizing and bin-packing. |
| SP009 | SiliconANGLE | Cast AI raises funds from Pacific Alliance Ventures at $1B valuation to launch unified GPU marketplace | Oracle Corp. joined Omni Compute as the first major cloud provider to make excess GPU capacity available through Cast AI customers. |
| SP010 | Business Insider | Companies are hoarding AI compute because of FOMO — and they're sitting on most of it | According to data collected from 23,000 clusters across thousands of companies using Cast AI's AI agent, average GPU utilization across enterprise servers is at 5%, meaning roughly 95% of provisioned GPU capacity is not being used. |
| SI001 | TechVendorIndex | Cast AI Review 2026 — Cloud FinOps Services | Cast AI offers a percentage-of-savings structure on selected engagements, typically charging 15–25% of verified cost reductions versus a baseline period. |
| SI002 | Cast AI | Cast AI's 2026 State of Kubernetes Optimization Report Reveals GPU Utilization at 5% | Average CPU utilization across clusters stood at just 8% in 2025, while memory utilization was 20%. |
| SI003 | Caplight | CAST AI Valuation & Funding History | Caplight | CAST AI Funding Rounds |
| SI004 | SiliconANGLE | Cast AI raises funds from Pacific Alliance Ventures at $1B valuation to launch unified GPU marketplace | a startup that specializes in optimizing the efficiency of cloud-native environments, today announced it raised new funding from Pacific Alliance Ventures, surpassing a $1 billion valuation, to launch a unified cloud graphics processing unit marketplace. |
| SI005 | Cast AI | Kubernetes Glossary | Cast AI: A platform that offers workload optimization features like adaptive memory tuning to prevent Out-of-Memory (OOM) crashes and automatic surge responses. |
| SI006 | nOps | Kubecost vs Cast.ai vs nOps | Pricing: Cast.ai employs a usage-based pricing structure, where costs are tied directly to compute consumption—with a fixed base fee plus a per-CPU rate. |
| SI007 | Cast AI | GPU Instances | Node Autoscaling | Cast AI Docs | Once activated, Cast AI's Autoscaler detects workloads requiring GPU resources and starts provisioning them. |
| SI008 | Business Insider | Companies are hoarding AI compute because of FOMO — and they're sitting on most of it | According to data collected from 23,000 clusters across thousands of companies using Cast AI's AI agent, average GPU utilization across enterprise servers is at 5%, meaning roughly 95% of provisioned GPU capacity is not being used. |
| SE001 | Cast AI | Cast AI Documentation | Getting Started | Cast AI Docs | Cast AI is an all-in-one Kubernetes automation, optimization, security, and cost management platform. It abstracts layers of provider-specific technical complexity so that you can easily manage Kubernetes operations on all three major cloud providers (GCP, AWS, Azure) and beyond with Cast AI Anywhere. |
| SE002 | Cast AI | Kubernetes Optimization Platform for Performance - Cast AI | Predicts spot interruptions up to 30 minutes before they happen, migrating workloads gracefully before your users feel a slowdown. |
| SE003 | Caplight | CAST AI Valuation & Funding History | Caplight | CAST AI Funding Rounds |
| SE004 | Cast AI | Download the 2026 Kubernetes Optimization Report - Cast AI | This report is based on our analysis of tens of thousands of Kubernetes clusters across AWS, GCP, and Azure, covering January 1 to December 31, 2025, and until April 2026 for GPU utilization. |
| SE005 | SoftwareOne Marketplace | Cast AI by CAST AI | SoftwareOne Marketplace | Supports integrations with Terraform, Grafana, Prometheus, Jira, Helm, and more. |
| SE006 | G2 | The G2 on Cast AI | While Cast AI's workload rebalancing capability is useful for optimizing cluster utilization, we have encountered a few issues during rebalancing activities. |
| SE007 | BizTech Weekly | 2026 Cast AI Report: Companies Waste GPU Capacity with Only 5% Utilization Amid Rising AI Infrastructure Costs - BizTech Weekly | Drawing on telemetry from 23,000 enterprise Kubernetes clusters—including environments run by BMW and Cisco—the report highlights a stark mismatch between what organizations *reserve* and what they *actually use*: enterprises are provisioning roughly 20× more GPU capacity than they consume, translating into ~5% GPU utilization (and ~8% CPU utilization). |
| SE008 | Business Insider | Companies are hoarding AI compute because of FOMO — and they're sitting on most of it | According to data collected from 23,000 clusters across thousands of companies using Cast AI's AI agent, average GPU utilization across enterprise servers is at 5%, meaning roughly 95% of provisioned GPU capacity is not being used. |
| SE009 | SiliconANGLE | Cast AI raises funds from Pacific Alliance Ventures at $1B valuation to launch unified GPU marketplace | With this funding, the company is expanding its product portfolio with Omni Compute, a capability that connects to external capacity, including GPUs, as native compute. |
| SU001 | CloudZero | CAST AI vs Kubecost: Kubernetes Cost Tools Compared (2026) | CAST AI limits allocating organizational costs by cluster and pod. With Kubecost, you can go further with cost allocation concepts such as project, team, product, environment, and more. |
| SU002 | Cast AI | Genial Investimentos | Using Cast’s automation, Genial Investimentos was able to save around 40% on its production cluster, with the average cost per CPU dropping from $5.88 to $5.40. |
| SU003 | Business Insider | Companies are hoarding AI compute because of FOMO — and they're sitting on most of it | Cast AI, an automated cloud cost optimization and management platform that works with companies like BMW and Cisco, published data in its 2026 State of Kubernetes Optimization Report showing that, on average, most organizations provide about 20 times more GPU capacity than they actively use at any given moment. |
| SU004 | SoftwareOne | Cast AI by CAST AI | SoftwareOne Marketplace | Cast AI is designed to automate Kubernetes operations, focusing on cost optimization, security, and efficiency improvements. It helps companies reduce cloud expenses by optimizing resource allocation and automating tasks such as rightsizing and bin-packing. |
| SU005 | Cast AI | Kubernetes Optimization Platform for Performance - Cast AI | Trusted by 2100+ companies globally |
| SU006 | Cast AI | Cast AI's 2026 State of Kubernetes Optimization Report Reveals GPU Utilization at 5% | Cast AI, the leading automation platform, today released its 2026 State of Kubernetes Optimization Report, a comprehensive analysis of GPU, CPU, and memory utilization across non-optimized Kubernetes clusters. Drawing on data from tens of thousands of clusters, the report delivers a clear and urgent message: GPUs are poorly utilized at 5% despite their cost. |
| SU007 | Cast AI | Download the 2026 Kubernetes Optimization Report - Cast AI | This report is based on our analysis of tens of thousands of Kubernetes clusters across AWS, GCP, and Azure, covering January 1 to December 31, 2025, and until April 2026 for GPU utilization. |
| SU008 | Cast AI | Kubernetes Glossary | Cast AI: is an automation platform that manages node lifecycles in real time, keeping applications reliable and responsive. It selects the right instance mix based on live workload signals, with cost savings following as a natural byproduct. |
| SR001 | Usage AI | CAST AI Reviews 2026: Capabilities, Value, and Fit | Criticism concentrates around small-cluster pricing, reporting, customization, documentation, and advanced-policy complexity. Some reviewers also want clearer explanations and better forecasting before changes are applied... CAST AI’s standard node automation must be the sole node autoscaler. Running it beside another node autoscaler can create conflicting and unpredictable scaling behavior. |
| SR002 | Inferensys | Kubecost vs Cast AI for Kubernetes GPU Optimization | Cast AI takes a fundamentally different approach by focusing on automated, policy-driven optimization and enforcement. Instead of just reporting on cost, Cast AI's engine continuously analyzes cluster state and automatically selects the most cost-effective GPU instances, bin-packs workloads, and can even evict low-priority jobs based on custom policies. |
| SR003 | Cast AI | Cast AI's 2026 State of Kubernetes Optimization Report Reveals GPU Utilization at 5% | Drawing on data from tens of thousands of clusters, the report delivers a clear and urgent message: GPUs are poorly utilized at 5% despite their cost. |
| SR004 | Business Insider | Companies are hoarding AI compute because of FOMO — and they're sitting on most of it | According to data collected from 23,000 clusters across thousands of companies using Cast AI's AI agent, average GPU utilization across enterprise servers is at 5%, meaning roughly 95% of provisioned GPU capacity is not being used. CPU utilization, the report says, is similarly low, at 8% of total capacity. |
| SR005 | SoftwareOne | Cast AI by CAST AI | SoftwareOne Marketplace | Cast AI is designed to automate Kubernetes operations, focusing on cost optimization, security, and efficiency improvements. It helps companies reduce cloud expenses by optimizing resource allocation and automating tasks such as rightsizing and bin-packing. |
| SR006 | Cast AI | Download the 2026 Kubernetes Optimization Report - Cast AI | This report is based on our analysis of tens of thousands of Kubernetes clusters across AWS, GCP, and Azure, covering January 1 to December 31, 2025, and until April 2026 for GPU utilization. All utilization data was collected before these organizations enabled Cast AI’s automation features, providing a baseline of how clusters perform under typical conditions without automated optimization. |
| SR007 | Cast AI | Kubernetes Optimization Platform for Performance - Cast AI | Cast AI turns Kubernetes workload, infrastructure, cost, and SLO signals into safe automated actions: rightsizing pods, scaling nodes, optimizing GPUs and Spot, and fixing issues without manual tuning. |
| SR008 | Cast AI | Kubernetes Glossary | Cast AI: is an automation platform that manages node lifecycles in real time, keeping applications reliable and responsive. It selects the right instance mix based on live workload signals, with cost savings following as a natural byproduct. |
| SV001 | Exa | Cast AI Funding Rounds & Investor Insights | Since its inception, Cast AI has built an impressive funding track record, successfully raising a total of $272 million across 10 funding rounds. The company’s funding journey has been characterized by large, strategically timed Series B and Series C rounds, involving prominent venture capital investors and a growth-focused approach. |
| SV002 | AI-TechPark | Cast AI Hits $1B Valuation & Launches OMNI GPU Marketplace | The company also announced a strategic investment from Pacific Alliance Ventures (PAV), the U.S.-based corporate venture arm of Shinsegae Group, an over $50 billion Korean conglomerate with leading businesses across retail, consumer, and digital platforms. With this round of funding, Cast AI’s valuation exceeds $1 billion. |
| SV003 | Multiples | CAST AI Valuation, Funding & Investors | Multiples | CAST AI is currently valued at $1B, based on its Series C in January 2026. The round was led by Pacific Alliance Ventures. CAST AI's valuation has grown 1.1x from $900M at its Series C in April 2025 to $1B in January 2026, across 2 priced rounds. |
| SV004 | Business Insider | Companies are hoarding AI compute because of FOMO — and they're sitting on most of it | According to data collected from 23,000 clusters across thousands of companies using Cast AI's AI agent, average GPU utilization across enterprise servers is at 5%, meaning roughly 95% of provisioned GPU capacity is not being used. CPU utilization, the report says, is similarly low, at 8% of total capacity. |
| SV005 | F4 Startup Intelligence | Cast AI | Developer Tools & Infrastructure | F4 | Cast AI | Developer Tools & Infrastructure | F4 |
| SV006 | Cast AI | Download the 2026 Kubernetes Optimization Report - Cast AI | This report is based on our analysis of tens of thousands of Kubernetes clusters across AWS, GCP, and Azure, covering January 1 to December 31, 2025, and until April 2026 for GPU utilization. |
| SV007 | Cast AI | Kubernetes Optimization Platform for Performance - Cast AI | Cast AI turns Kubernetes workload, infrastructure, cost, and SLO signals into safe automated actions: rightsizing pods, scaling nodes, optimizing GPUs and Spot, and fixing issues without manual tuning... Trusted by 2100+ companies globally |
| SV008 | Cast AI | Kubernetes Glossary | Cast AI: is an automation platform that manages node lifecycles in real time, keeping applications reliable and responsive. It selects the right instance mix based on live workload signals, with cost savings following as a natural byproduct. |
| SV009 | SoftwareOne | Cast AI by CAST AI | SoftwareOne Marketplace | Cast AI is designed to automate Kubernetes operations, focusing on cost optimization, security, and efficiency improvements. It helps companies reduce cloud expenses by optimizing resource allocation and automating tasks such as rightsizing and bin-packing. |