Poolside
技术野心卓越的主权 AI 软件工程公司,但商业基本面仍不透明。
Poolside 拥有出色的主权 AI 软件工程能力,也获得了 Nvidia 支持的变革性流动性事件;但公开证据仍不足以支撑激进的估值判断。
覆盖范围与披露说明
公开证据主要集中在产品架构、融资历史和 2026 年 8 月与 Nvidia 的交易;经审计收入、客户数、留存率和利润率质量仍未披露。
封面要素
公司概况
Poolside 是一家专注软件工程工作流的主权企业 AI 公司。公司销售 Poolside Platform 和 Laguna 模型家族,支持本地部署、VPC 和气隙环境,让专有代码留在客户的安全边界内。2026 年与 Nvidia 的交易显著改善了公司的流动性和市场验证,但公开证据仍无法确定其商业规模、留存率和单位经济性。
- 创始人
- Jason Warner, Eiso Kant
- 创立地点
- United States
- 总部
- San Francisco, CA and Paris, France
- 产品
- Poolside 销售主权软件工程 AI 系统,包括 Laguna 模型家族和 Poolside Platform,用于安全代码生成、重构和自主开发者工作流。
- 客户
- 大型企业、国防与情报机构,以及受监管的金融机构
- 商业模式
- 企业软件订阅、前置部署服务和战略许可
- 阶段
- Series C / growth-stage private
- 融资情况
- 已募集 $1.626B 股权资本,并在 2026 年 8 月获得 Nvidia 对 Model Factory 的 $6B 非排他性许可承诺。
执行摘要
主要优势
- 面向受监管企业和公共部门买家的差异化主权部署模式。
- 代码专用基础模型和智能体软件工作流的技术可信度较强。
- Nvidia 及此前风险投资支持方提供了重要的财务与战略验证。
主要风险
- 在缺少经审计 ARR、收入或留存披露的情况下,商业规模仍不透明。
- CoreWeave 崩溃和 Nvidia 交易之后,基础设施与人才集中度风险上升。
- 知识产权来源、许可体系的异质性及治理复杂度可能拖慢采用。
未决问题
- 当前经审计的 ARR、收入质量、毛利率、现金消耗和资金续航。
- 直接客户数量、生产环境部署范围,以及留存 / 扩张指标。
- 董事会构成、投资者权利,以及任何附函或治理义务。
- 训练数据来源、知识产权赔偿责任,以及开放权重许可风险敞口。
目录
01公司概览
1.1 公司身份、使命与运营模式
Poolside(对应法律实体 Poolside, Inc.,于 2023 年 4 月在美国注册)是一家为软件工程打造人工智能模型和自主系统的公司。公司由前 GitHub 首席技术官 Jason Warner 和连续创业的工程智能企业家 Eiso Kant 于 2023 年初创立,运营总部横跨美国加州旧金山和法国巴黎。Poolside 的根本判断是:软件开发将是人工智能首个达到并超越人类水平能力的主要技术领域,因此需要全栈模型架构,而不是浅层的 IDE 外壳扩展。 为在高保障环境中落实这一使命,Poolside 成立了 Poolside Federal LLC。这是一家注册于美国的运营实体,专门服务联邦机构、国家安全组织和国防工业基地客户。不同于按 token 持续收费、并将数据远程路由的面向消费者云端代码助手,Poolside 提供开放权重基础模型(包括 Laguna 模型家族),专门支持本地部署、VPC 和气隙部署。客户无需向外传输数据即可在本地无限推理,专有企业代码库也因此始终留在机构控制范围内。 [CO001, CO002, CO003, CO004, CO005]
| 指标 / 维度 | 报告值 | 截至日期 | 信心度 | 分析备注与证据缺口 |
|---|---|---|---|---|
| 公司法定名称 | Poolside, Inc. / Poolside Federal LLC | 2026-08 | 高 | 主要公司实体于美国注册(2023 年 4 月);专门的联邦子公司负责美国公共部门业务。 |
| 总部与运营地点 | San Francisco, CA (US) & Paris (France) | 2026-08 | 高 | 覆盖北美和欧洲的分布式运营基地,由通过审查的美国人员提供支持。 |
| 运营阶段 | Series C / 成长阶段 | 2026-08 | 高 | 2024 年 10 月完成 Series B($500M);2026 年 8 月完成 Nvidia 提供的 $1B Series C 股权融资。 |
| 投前股权估值 | $12,000M ($12.0B) | 2026-08 | 高 | 由 2026 年 8 月 Nvidia 股权融资轮确定;隐含约 $13.0B 投后资本化规模。 |
| 累计股权融资总额 | $1,626M ($1.626B) | 2026-08 | 高 | 包括 $26M Seed、$100M Series A、$500M Series B,以及 Nvidia $1,000M 股权投资。 |
| 许可资本承诺 | $6,000M ($6.0B) | 2026-08 | 高 | Nvidia 对 Model Factory 的非独家许可;预计在 2027 年底前分配给历史投资人。 |
| 年化收入 / ARR | 未报告($50M-$66.4M 外部估算) | 2026-08 | 低 | 公司未披露 GAAP 收入;私下估算来自第三方模型外推。 |
| 员工人数 / 团队规模 | 347 人(Nvidia 重组前记录) | 2026-08 | 中 | Tracxn 记录 347 名员工;109 名基础模型研究工程师收到 Nvidia 的聘用邀约。 |
| 核心基础架构 | Laguna 开放权重模型家族 | 2026-07 | 高 | 模型包括 Laguna M.1、XS.2 和 Laguna S 2.1 开放权重版本,采用 RLCEF 方法训练。 |
| 部署形态 | 本地、VPC 与物理隔离主权部署 | 2026-08 | 高 | 无强制外部 API 依赖或出站要求;向客户交付完整权重。 |
公司事实基线汇总自主要实体披露、交易条款清单和行业跟踪数据库。未经审计的收入估算来自外部跟踪模型,而非 经验证的公司申报。
[CO001, CO003, CO004, CO010, CO011, CO013, CO014, CO015]1.2 执行团队、治理与组织规模
Poolside 由联合创始人 Jason Warner(首席执行官)和 Eiso Kant(首席技术官,运营文件中也曾任联合 CEO)领导。共同创办 Poolside 之前,Warner 曾任 GitHub 首席技术官,负责工程孵化和发布 GitHub Copilot、GitHub Actions、GitHub Codespaces 等定义品类的开发者工具,此前还在 Heroku、Canonical 和风险投资机构 Redpoint Ventures 担任领导职务。Kant 则带来了互补的领域经验:他是 AI 辅助代码分析的先行者,2015 年创办早期代码库分析初创公司 source{d},之后又打造了工程智能平台 Athenian。 公司还聘请前 GitHub 首席营收官 Paul St. John 负责企业市场拓展,并于 2025 年收购 Fern Labs,补充了多智能体编排基础设施和前置部署工程师。商业和机构支持来自领投机构 Bain Capital Ventures、Felicis Ventures,以及 Nvidia、SentinelOne 的 S Ventures、Citi Ventures 和 Adams Street Partners 等企业投资方。2026 年 8 月发生了结构性领导变动:109 名专注基础模型开发的核心工程师在一项企业许可交易后转至 Nvidia,而 Warner 和 Kant 留在 Poolside,继续负责公司的独立产品路线。 [CO006, CO007, CO008, CO009]
| 高管姓名 | 高管职务 | 职业背景与领域匹配 | 职能领域覆盖 | 关键人物依赖与状态 |
|---|---|---|---|---|
| Jason Warner | 联合创始人兼首席执行官 | 曾任 GitHub 首席技术官(负责 Copilot、Actions、Codespaces);Redpoint Ventures 董事总经理;Heroku 工程副总裁; Canonical。 | 负责管理层战略、产品架构愿景、投资者关系、公司发展和联邦事务。 | 关键人物依赖度高;Poolside 投资逻辑和联邦定位的主要公开架构者;Nvidia 交易后留任。 |
| Eiso Kant | 联合创始人兼首席技术官 | source{d} 创始人兼 CEO(2015 年率先将机器学习用于代码库);Athenian 创始人兼 CEO(工程指标平台)。 | 负责应用 AI 核心研究、代码执行反馈强化学习(RLCEF)和 Model Factory 架构设计。 | 关键人物依赖度高;代码机器学习系统领域权威;交易后仍留在 Poolside。 |
| Paul St. John | 首席营收官 | 曾任 GitHub 首席营收官;负责企业商业化扩张和开发者工具上市推广。 | 负责全球企业客户获取、战略渠道联盟和前置部署工程合同。 | 职能依赖度显著;把主权防务和金融企业试点转化为多年期承诺的关键人物。 |
高管名单已与投资人公告和行业登记信息核对。两位联合创始人在 Nvidia 许可交易后仍继续领导 Poolside。
[CO006, CO007, CO008]1.3 资本化轨迹与 2026 年 8 月 Nvidia 交易
Poolside 的资本化历史跻身企业软件史上融资速度最快的案例之列。公司于 2023 年 5 月完成 $26M 种子轮,2023 年 8 月完成 $100M A 轮,随后于 2024 年 10 月完成 $500M B 轮。该轮由 Bain Capital Ventures 领投,Nvidia、DST Global、StepStone Group、Citi Ventures、HSBC Ventures 和 SentinelOne(S Ventures)参与,投后估值为 $3.0B。2026 年春季,Poolside 与 CoreWeave 谈判接入规划中的 2 吉瓦数据中心集群失败,导致 Project Horizon 终止;公司筹备中的 $2B C 轮融资也因此停滞,并失去原计划中的 40,000 台 GB300 GPU 集群。 2026 年 8 月 20 日,Poolside 与 Nvidia 达成总额 $7B 的多部分交易,彻底解决了这一基础设施瓶颈。交易包括两项内容:第一,Nvidia 为 Poolside 的专有 Model Factory 训练与评估平台支付 $6B 非排他性商业许可费,这笔资本计划在 2027 年底前分配给 Poolside 现有投资人,而不是作为公司运营现金;第二,Nvidia 以 $12B 投前估值向 Poolside 直接投资 $1B(投后约 $13B)。关键在于,该协议是非排他性许可和少数股权投资,而非收购,因此 Poolside 保留了公司独立性和商业行动自由。 [CO010, CO011, CO012, CO013, CO014, CO015]
| 利益相关方 / 投资人 | 角色与进入轮次 | 资本与交易概况 | 战略重要性与治理权利 | 尽调要求与剩余考量 |
|---|---|---|---|---|
| Nvidia | 战略投资人与商业伙伴(Series B、Series C、被许可方) | 参与 2024 年 10 月 Series B;2026 年 8 月投入 $1,000M Series C 股权;签署 $6,000M 非独家许可协议。 | 主导算力和分销伙伴;获得 Model Factory 平台许可;招募 109 名 Laguna 工程人员。 | 评估许可分配限制或员工转移是否削弱 Poolside 在 2026 年后的模型创新独立性。 |
| Bain Capital Ventures | 领投机构投资人(Series B) | 领投 $500M Series B(2024 年 10 月),投后估值 $3B。 | 主要机构董事会席位及初期多智能体商业化规模背后的首要财务赞助方。 | 核查董事会席位分配、表决控制权区块,以及 $6B 许可清算的分配瀑布条款。 |
| Felicis Ventures | 核心早期机构支持者(Series A、Series B) | 共同领投 $100M Series A(2023 年 8 月);参与 Series B。 | 支持公司拓展欧洲和北美软件工程中心的早期财务赞助方。 | 评估优先认购权的维持情况,以及其在 2026 年 8 月资本结构安排中的参与度。 |
| Redpoint Ventures | 早期机构支持者(Seed、Series B) | 早期风险投资赞助方(Jason Warner 担任董事总经理前的机构);参与 Series B。 | 与创始人网络治理立场一致,并参与早期资本联合体组建。 | 确认历史股权持有和二级市场参与记录。 |
| SentinelOne (S Ventures) | 战略企业安全投资人(Series B) | $500M Series B(2024 年 10 月)的战略少数股东。 | 为主权 SOC 环境中的安全本地 AI 部署和遥测集成提供战略验证。 | 评估联邦情报和商业企业领域联合上市推广管线的转化。 |
| Adams Street Partners | 风险成长投资人(Series B) | $500M Series B 的成长阶段投资人。 | 为企业开发者生产力经济性和全栈软件生成系统提供财务验证。 | 审查后续投资参数和长期流动性预期。 |
利益相关方地图汇总自已披露的公司公告和 Tracxn 联合投资记录。不包括保密天使投资人和未披露的二级联合体实体。
[CO008, CO010, CO012, CO013, CO014]按时间梳理 Poolside 自 2023 年创立、Series B 扩张,到 2026 年春季 CoreWeave 集群取消,以及 2026 年 8 月 Nvidia 许可与股权交易这一里程碑事件。
日期反映公开报道的公告月份。历史 Series B 估值采用已披露的交易数据。
公司注册成立 [CO001]
种子轮融资($26M) [CO010]
Series A 轮融资($100M) [CO010]
Series B 轮融资($500M) [CO010]
AWS Bedrock 合作 [CO004]
CoreWeave Project Horizon 宣布 [CO012]
CoreWeave Horizon 交易告吹 [CO012]
Laguna 模型家族发布 [CO003]
1.4 展示项
02市场分析
2.1 市场边界、核心支出类别与传统替代方案
软件工程人工智能的商业市场跨度很大,从面向消费者的自动补全插件一直延伸到主权基础模型基础设施。Poolside 明确处于基础设施与企业市场的交界处,服务范围不同于依赖公有多租户云 API 的普通消费者代码助手。Poolside 的核心可服务支出包括自管推理引擎、开放权重代码模型、企业智能体编排平台,以及直接部署在客户安全边界内的专用 Model Factory 工具链。目标客户包括 Global 2000 企业、受监管金融机构和公共部门国防机构;这些组织因知识产权政策或法定数据驻留要求,禁止将专有代码传出至第三方多租户云。该边界不包括低 ACV 的个人开发者订阅、未受管理的公有 SaaS copilot 和通用消费者聊天机器人外壳。现有替代方案包括静态分析工具、基于规则的 IDE 扩展、人工软件测试工作流,以及托管在商业超大规模云上的通用 LLM API 端点。[CM001, CM002, CM004, CM005, CM009]
| 市场细分 | 纳入支出类别 | 排除支出类别 | 主要买方 / 付款方 | 对主权 AI 的相关性 |
|---|---|---|---|---|
| 主权基础模型与工具链 | 模型权重、本地 Model Factory 许可、定制 RL 微调 | 通用消费级 LLM API token、无人管理的公共 SaaS | CIO、CTO、首席 AI 官 | 自主管理企业 AI 的直接核心可寻址市场 |
| 企业智能体工作流与 IDE 扩展 | 智能体 CLI 工具、多文件重构智能体、开发环境集成 | 开箱即用的单行自动补全订阅、消费级 IDE | 工程副总裁、开发者生产力负责人 | 企业边界内的主要高速部署场景 |
| 受监管与防务 AI 工程 | 物理隔离算力设备、符合 STIG 的部署、通过审查的支持服务 | 仅支持商业云的编程副驾、公共共享租户 SaaS | 防务机构、联邦总包商、CISO | 对数据出站零容忍的高毛利主权细分市场 |
| 算力与 Model Factory 基础设施 | 交钥匙硬件设备(Dell)、私有 VPC 集群、GPU 推理编排 | 公共多租户 AI 试验场托管、消费级 GPU 云 | 基础设施负责人、云运营负责人 | 决定模型规模和执行可行性的支撑基础设施 |
纳入和排除支出分类反映企业边界要求,将主权自主管理软件工程工作流与普通消费级 AI 助手区分开来。
[CM001, CM002, CM004, CM005]2.2 可服务市场规模、增长动力与估值视角
评估 Poolside 的可服务市场,必须把广义软件工程支出拆成受证据约束的具体规模层级。宏观基线是:全球软件工程师劳动报酬和企业开发预算每年超过 $1.2T,背后是合格开发者长期短缺和工程交付积压。以此为基础,AI 辅助软件开发工具的总可寻址市场(TAM)预计到 2026 年为 $25B–$35B;随着企业从简单代码补全转向自主多文件智能体工作流,复合年增长率为 30%–45%。主权、本地部署和受 VPC 限制的企业 AI 可服务市场(SAM)预计为 $6B–$10B,主要由航空航天、国防、银行、医疗和公共部门治理等受监管行业驱动。在这一 SAM 中,Poolside 的可获取市场(SOM)是高安全子市场,规模为 $1.5B–$3B,覆盖气隙、主权和国防部署。2026 年 8 月 20 日,Nvidia 承诺为 Poolside 的 Model Factory 软件提供 $6B 非排他性许可,并以 $12B 投前估值进行 $1B 股权投资,市场估值基准也因此被彻底重塑。[CM001, CM002, CM005, CM006, CM007, CM009, CM010, CM012, CM016]
| 测算视角 / 发布方 | 发布年份 | 地理范围 | 估计价值(USD) | 预计增长 / CAGR | 方法 / 依据 | 信心度与局限 |
|---|---|---|---|---|---|---|
| 全球软件工程支出(宏观基线) | 2024–2026 | 全球 | $1.2T+ 年度开发者劳动力与软件预算 | N/A(宏观基线) | 专业开发者总薪酬和企业软件工程预算 | 基线劳动力规模很大,但 AI 自动化工具只能触达其中一部分 |
| AI 辅助软件开发工具(TAM) | 2025–2026 | 全球 | $25B–$35B 预计市场支出 | 30%–45% CAGR | 企业开发者工具采用模型、代码生成订阅和智能体工具支出 | 信心度中等;从基础自动补全转为企业智能体的转化率不同,结果差异很大 |
| 主权与受监管企业 AI(SAM) | 2026 | 北美与欧洲 | $6B–$10B 可寻址细分市场 | 35%+ CAGR | 限定为需要私有基础设施的国防工业基地、金融服务、医疗和公共部门机构 | 信心度中等;受到本地硬件容量和企业合规周期限制 |
| 自主管理智能体基础设施(SOM) | 2026 | 目标企业账户 | 保密 / 未披露 | N/A | Poolside 面向运行物理隔离部署的《财富》500 强企业和防务机构的直接管线 | 公开可见度低;商业 ARR 和管线指标仍为私有且未经验证 |
测算视角综合全球开发者支出的分析师估计和已报告交易,并按合规和数据驻留限制筛选,以识别主权企业机会。
[CM006, CM009, CM010, CM012, CM016]将可触达市场从全球软件工程支出逐层拆解至气隙隔离的主权编码部署。
各层宽度仅表示层级,不代表数值比例。
市场规模分层综合第三方开发者工具市场预测与公开主权部署基准;除战略交易承诺外,金额均代表年度支出。
全球软件工程劳动力与预算 [CM010]
总可触达市场(TAM):AI 编码平台 [CM009, CM010]
可服务市场(SAM):受监管与主权 AI [CM001, CM002, CM005]
2.3 买方画像、采购机制与部署授权
主权 AI 软件工程的采用路径和采购周期,与传统开发者工具采购有明显差异。个人席位制代码助手通常通过工程部门信用卡列支,或由开发者自下而上采用;Poolside 平台则需要由上而下的企业采购,涉及 CIO、CISO 和专门的风险委员会。商业企业的经济买方是企业 CIO 和基础设施负责人,他们希望现代化改造遗留应用,同时严格执行防止数据泄露的规则,不暴露专有源代码库。联邦和国防部门则通过正式国防合同工具、联邦项目执行办公室,以及 Vibrint、Sterling Computers 和 Hunted Labs 等战略系统集成商采购。在这些敏感领域获得部署授权,需要经过审查的人员、符合 STIG 的加固,以及在专用本地硬件上完成气隙一体化部署。由于 Poolside 模型不产生数据外传,也不收取按 token 计费的 API 费用,经济买方可以把软件工程自动化列为可预测的资本支出,而不是不受约束的可变运营成本。[CM001, CM002, CM003, CM005, CM011, CM012, CM013, CM014, CM015]
| 客户细分 | 经济买方 | 主要用户 | 授权机构 / 付款方 | 目标工作流 | 预算归属 | 关键采用触发因素 |
|---|---|---|---|---|---|---|
| 防务与情报机构 | 项目执行官 / 机构负责人 | 防务软件工程师、情报分析师 | 联邦采购、机密项目预算 | 任务遗留代码现代化、物理隔离安全软件开发 | 国防拨款 / 特殊任务资金 | 强制数据驻留、零出站安全政策、物理隔离任务运营 |
| 国防工业基地总包商(例如 RTX) | 工程副总裁 / 首席技术官 | 系统软件工程师、嵌入式固件开发者 | 企业研发和国防合同工具计费 | 自动代码重构、合规文档、加速 ATO | 工程间接预算与直接项目成本池 | 供应链安全要求、CMMC 合规、专有 IP 保护 |
| 受监管金融机构 | 首席信息官 / 技术基础设施负责人 | 核心银行软件工程师、量化开发者 | 企业 IT 采购、技术指导委员会 | 遗留代码库迁移、主机现代化、安全代码审查 | 集中的企业技术资本预算 | 监管审计压力、严格的数据丢失防护(DLP)要求、私有 VPC 要求 |
| 一线科技与主权云提供商 | AI 基础设施负责人 / 平台工程副总裁 | 内部平台工程师、基础模型研究人员 | 战略研发投资、云合作伙伴联合开发资金 | Model Factory 许可、定制模型训练、交钥匙推理交付 | 战略企业研发与基础设施资本开支 | 垂直基础设施集成、算力变现、主权云产品 |
买方细分区分了防务项目办公室与商业企业 IT 的不同决策路径;在后者,采购高度受安全认证和数据驻留规则把关。
[CM001, CM002, CM003, CM005, CM011, CM013, CM014, CM015]2.4 结构性增长动力、基础设施约束与尽调风险
多项长期顺风因素正在推高企业对主权 AI 软件工程的需求,包括遗留任务软件不断累积的技术债、对代码隐私日益严格的监管审查,以及软件需求与工程人才供给之间不断扩大的生产率缺口。Poolside 从代码执行反馈中进行强化学习,为多步推理和自动重构提供了有防守力的架构基础。但商业扩张仍受明显的采用摩擦和结构性瓶颈制约。企业部署销售和实施周期很长,往往需要前置部署研究工程师进入客户团队完成定制集成。此外,主权基础设施模式需要大量算力来完成客户侧模型微调和推理。2026 年春季,Poolside 与 CoreWeave 就得州 Pecos County 2 吉瓦数据中心园区的谈判破裂,直接暴露了这一算力脆弱性,打乱扩张计划并最终促成与 Nvidia 的交易。再加上公司没有公开收入、客户留存基准或可供核验的生产案例,独立验证经济性成为尽调的关键障碍。[CM003, CM004, CM006, CM007, CM008, CM014, CM015, CM016]
2.5 展示项
03竞争格局
3.1 竞争格局:直接同业、既有巨头与替代方案
企业生成式 AI 软件工程市场分为三个竞争层级:云原生既有助手、风险投资支持的代码专用前沿实验室,以及通用前沿基础模型。在既有巨头阵营中,Microsoft 的 GitHub Copilot 通过深度整合 Visual Studio Code 和企业 Microsoft 许可协议保持强大分发,成为 AI 辅助开发的现状基准。直接的风险投资支持挑战者包括 Magic(2024 年融资 $320M,用于构建超长上下文软件合成模型)和 Codeium(凭借广泛的多语言支持达到 $1.25B 估值)。此外,OpenAI、Anthropic 和 Mistral AI 等前沿基础模型实验室也通过云 API 提供前沿推理模型。云端助手在个人开发者体验上占优,但国防、情报和一级银行的受监管买方会评估气隙主权方案,因为专有源代码不会穿越公有云边界。[CP001, CP002, CP003, CP004]
| 竞争对手 | 类别 | 规模 / 融资 | 目标细分 | 差异化 | 局限 |
|---|---|---|---|---|---|
| GitHub Copilot (Microsoft) | 现有云助手 | Microsoft 子公司;$100M+ ARR 规模 | 全球企业和个人开发者 | 深度 IDE 集成、原生 GitHub pull request 工作流、企业分销 | 依赖云端,代码 / 上下文必须出站;主权物理隔离支持有限 |
| Magic | 直接 AI 编程实验室 | 已融资 $465M+(2024 年 8 月 $320M 轮) | 企业软件开发团队 | 拥有超长代码库上下文窗口的专有前沿模型 | 封闭云 API 部署;本地主权物理隔离交付尚未验证 |
| Codeium (Exafunction) | 直接编程助手 | 已融资 $243M+($150M Series C,估值 $1.25B) | 个人开发者和企业团队 | 支持多种语言和混合部署的自研生成模型 | 历来更关注开发者体验,而非主权防务工作负载 |
| Anthropic (Claude) | 前沿基础模型实验室 | 从 Amazon、Google 和风险投资机构融资数十亿美元 | 普通企业、金融和软件团队 | 在经验证的编程基准上具备业界领先推理能力(Claude 3.5/4.6 系列) | 主要通过云 API 交付;推理成本高,且严格采用云多租户 |
竞争对手规模、融资和定位汇总自 Crunchbase、Tracxn 及截至 2024–2026 年的公开市场披露。
[CP001, CP002, CP003, CP004]3.2 能力差异化:主权权重与多租户云 API
Poolside 直接把完整模型权重交付到客户环境,因此能在没有遥测和外发网络调用的情况下完全气隙运行,从而形成区别于云优先竞争者的结构性优势。GitHub Copilot 和 Anthropic 需要把代码片段或 AST 表示传输到多租户云数据中心,而 Poolside 的 Laguna 基础模型运行在客户自有基础设施或安全的一体化集群上。平台提供带 Model Context Protocol 连接器的容器化智能体运行时,可接入企业内部系统,完整记录推理轨迹,并在运行时清除凭据。在独立评测中,Poolside 专有 Laguna M.1 在 SWE-bench Pro 上达到 46.9%,在 SWE-bench Verified 上达到 72.5%,超过 Devstral 2 并接近云端前沿模型;开放权重 Laguna XS.2 则以轻量级 330 亿参数规模在 SWE-bench Pro 上达到 44.5%。[CP005, CP006, CP007, CP008]
| 购买标准 | Poolside | GitHub Copilot | Magic | Anthropic Claude |
|---|---|---|---|---|
| 物理隔离 / 主权部署 | 原生本地权重交付,零出站网络流量 | 不支持;需要连接 Microsoft 多租户云 | 不支持;专有云平台 | 仅限第三方云 VPC;不交付原始权重 |
| 推理成本与定价模式 | 通过硬件订阅提供不限量推理,不按 token 收费 | 按用户按月 SaaS 订阅($19-$39/user/mo) | 按用量计费的云定价 / 企业合同 | 按 token 收取 API 费用,或按席位签订企业协议 |
| 智能体执行与工具调用 | 在容器中编排 MCP 服务器,连接 Jira、Slack 和 GitHub | Copilot Workspace / Extensions 通过云端访问工具 | 在专有云环境中运行自主智能体工作流 | 计算机使用和工具使用 API 由外部 API 调用中介 |
| 模型权重主权 | 提供完整模型权重(Laguna 系列;XS.2 采用 Apache 2.0 开放许可) | 仅由 Microsoft/OpenAI 运营的专有封闭权重 | 仅由 Magic 运营的专有封闭权重 | 由 Anthropic 或云合作伙伴运营的专有封闭权重 |
能力信息反映截至 2026 年公开的技术架构、官方产品文档和厂商披露。
[CP005, CP006, CP007, CP008]按部署隔离程度与代理自主性,对软件工程 AI 解决方案进行序数定位。
| 条目 | X (3–9.5) | Y (6.8–9.2) | 说明 |
|---|---|---|---|
| 1. Poolside(Laguna M.1 / 平台) | 9.5 | 8.5 | |
| 2. GitHub Copilot(企业版) | 3 | 8 | |
| 3. Anthropic(Claude 3.5/4.6) | 3.5 | 9.2 | |
| 4. Magic | 3 | 8.7 | |
| 5. Codeium | 5.5 | 7.2 | |
| 6. 本地开源(Laguna XS.2 / Ollama) | 9 | 6.8 |
部署隔离度(X 轴,1–10)衡量本地、气隙隔离与零数据外传能力;代理自主性(Y 轴,1–10)反映多步推理能力,以及公开评估中的工具执行自主性基准。
[CP005, CP007, CP008, CP011]3.3 定价与产品打包:可预测的算力经济学与 token 波动
AI 软件开发的商业定价,在按用量计费的云 API 与基于基础设施的许可之间存在显著差异。传统云平台按席位收取订阅费,或按输入和输出 token 收取消费费;自主多步智能体持续执行推理和测试循环时,预算会不可预测地上升。相比之下,Poolside 把主权企业方案打包为与硬件绑定的可预测订阅,在客户专用 GPU 上提供无限推理。为推动开发者采用并扩大研究生态影响力,Poolside 还以宽松的 Apache 2.0 开源许可发布 Laguna XS.2,让开发者无需持续许可费即可在标准工作站 GPU 上运行本地智能体工作流。但本地部署模型需要大量硬件资本投入,必须使用高内存工作站芯片或企业服务器集群。[CP009, CP010, CP011]
| 厂商 / 方案 | 定价与许可模式 | 包含能力 | 折扣或未知项 | 对企业尽调的影响 |
|---|---|---|---|---|
| Poolside 平台与模型 | 与硬件绑定的年度订阅;不限量本地推理 | 完整模型权重、Atlas 推理引擎、容器化智能体运行时、MCP 服务器 | 私有合同定价未披露;前置部署工程费用未知 | 消除 token 价格波动,但需要大量前期算力 / 服务器投资 |
| Laguna XS.2(开放权重) | 按宽松的 Apache 2.0 开源许可免费提供 | 可在单块 GPU 本地运行的 33B 参数(3B 激活)MoE 权重 | 商业支持和微调工具不包含在开放权重内 | 支持社区零成本评估和离线边缘开发者部署 |
| GitHub Copilot 企业版 | 列表价为每用户每月 $39 | 聊天、代码补全、pull request 摘要、自定义文档索引 | Microsoft Enterprise Agreements 提供企业批量折扣 | 进入门槛低、按席位成本可预测,但代码元数据会离开边界 |
| Anthropic Claude 企业版 | 按席位 SaaS,加按用量计费的 API token 档位 | Claude 模型、500k 上下文窗口、SSO、数据保留控制 | 按用量和承诺档位协商定制企业合同价格 | token 支出会随智能体循环增加;需要持续向云端发起出站调用 |
定价结构反映厂商公开列表价;企业折扣和定制专业服务费用仍未公开。
[CP009, CP010, CP011]3.4 护城河持久性:战略联盟、人才流失与替代威胁
Poolside 的竞争防守力主要来自主权监管定位,而不是原始算法智能上的不可逾越领先。Anthropic 等前沿基础模型实验室仍在不断刷新基准,只要买方能获得数据保护例外,封闭云 API 就是强有力的替代方案。此外,Poolside 在 2026 年 8 月与 Nvidia 交易后经历重大组织转型:Nvidia 为 Poolside 核心 Model Factory 训练软件支付 $6B 非排他性许可费,并吸收了 Poolside 130 名员工中的 109 人。Jason Warner 和 Eiso Kant 两位联合创始人保留了独立性、公司股权和客户关系,但约 20 人的精简团队要对抗超大规模既有企业,交付风险明显上升。为维持差异化,Poolside 依赖与 Amazon Web Services 和 Atos 的企业分销合作,以及嵌入国家安全机构、具备安全许可支持的解决方案架构师。[CP012, CP013, CP014]
3.5 展示项
04财务
4.1 收入来源与定价架构
Poolside 通过专用企业和主权部署变现软件智能,把 Laguna 基础模型和多智能体系统定位在高安全环境。与按 token 收取 API 推理费的公有超大规模云模型形成鲜明对比,Poolside 向政府和受监管企业客户提供完整模型权重,可在本地或气隙边界内部署,不产生外部数据外传。在这一主权许可架构下,客户组织支付固定年度软件许可费,并在自己的硬件集群上无限推理,为企业 IT 和国防项目建立高度可预测的采购模式。平台还集成专有 Atlas 推理引擎;Poolside 报称它最多可将硬件算力利用率提升 30%,从而优化客户基础设施的总拥有成本。除经常性主权软件订阅外,Poolside 还证明了通过战略技术变现获得非稀释性收入的能力,标志性案例就是 Model Factory 平台的非排他性软件许可交易。[CI001, CI008, CI012]
| 收入来源 | 机制 | 单位 | 当前状态 | 质量 | 尽调问题 |
|---|---|---|---|---|---|
| 主权企业许可证 | 按年或多年订阅本地部署 / 气隙环境中的模型权重 | 按客户组织或站点部署 | 商业部署活跃 | 经常性收入潜力高,且客户迁移障碍较大 | 核实许可证续约率及合同 SLA 承诺 |
| 政府与国防合同 | 通过 Poolside Federal LLC 签订固定价或成本加成的主权联邦合同 | 按项目或机构部署 | 活跃(与 Vibrint、Hunted Labs 合作) | 稳定性高,预算科目黏性强 | 查阅 CAGE / NAICS 任务订单及涉密人员配置义务 |
| 战略 IP 授权 | 非排他性技术授权(例如 Model Factory) | 平台交易里程碑费 | 已实现(2026 年 8 月 Nvidia 交易 $6B) | 一次性、非股权性质的战略现金流入 | 审查授权范围、IP 权利负担及排他性限制 |
| 专业服务与前置部署 | 定制集成、微调及安全加固 | 工时费、材料费或里程碑费用 | 企业客户业务活跃 | 毛利率较低的服务收入,用于推动软件采用 | 审计专业服务毛利率及可计费利用率 |
反映 Poolside 在公开文件、合作伙伴公告及公司文档中披露的商业产品和授权结构。实际实现的美元金额仍属私有信息。
[CI001, CI008, CI012]从企业主权客户接洽到长期实现毛利率的运营转化流程。
| 节点 / 连接 | 节点 / 起点 | 终点 | 说明 |
|---|---|---|---|
| 节点 1 | 客户需求识别 [N1] | Regulated enterprise or public agency identifies need for air-gapped sovereign coding AI | |
| 节点 2 | 本地架构与部署 [N2] | Full model weights delivered inside customer security perimeter with zero external egress | |
| 节点 3 | Atlas 引擎优化 [N3] | Proprietary inference engine increases hardware compute utilization by up to 30% | |
| 节点 4 | 前置部署定制 [N4] | Solutions architects integrate internal code repositories and CI/CD pipelines | |
| 节点 5 | 主权许可转化 [N5] | Transition to multi-year recurring license with predictable flat-rate pricing | |
| 节点 6 | 扩张与战略变现 [N6] | Contract value expansion across agency divisions and strategic technology licensing |
展示定性的商业与运营转化阶段;底层合同层面的利润率属于非公开信息。
[CI001, CI003, CI012, CI013]4.2 商业进展与销售效率代理指标
Poolside 的市场进入依靠高接触度的直接企业合作和联邦系统伙伴关系,而不是低摩擦的开发者自助获客。借助专门的公共部门载体 Poolside Federal LLC,公司已与 Vibrint、Hunted Labs 等国家安全承包商建立集成合作,把气隙代码模型部署到敏感的联邦任务中。企业客户通常需要前置部署解决方案架构师配置安全边界、在内部代码库上微调模型,并接入现有持续集成流水线。但这种高接触度的主权交付模式意味着企业销售周期长、前期部署成本高。公开报道显示,截至 2026 年第一季度,Poolside 年化企业收入运行率仍低于 $50M。尽管 SentinelOne 的 S Ventures 等企业投资方支持商业采用,客户获客成本、净收入留存率和实际合同回本周期等完整指标仍是未披露的私有数据。[CI001, CI002, CI006, CI012, CI014]
| 产品线 | 价格 / 合同模式 | 标价与实际价格 | 折扣 / 未知项 | 来源 |
|---|---|---|---|---|
| 公共部门 / 政府 | 无限本地推理的固定基础设施许可证 | 标价:按 token 推理费为零 | 硬件规格、维护和支持费用按客户私下报价 | Poolside 官方政府页面 |
| Laguna 开放权重开发者 API | 通过 OpenRouter / API 按用量计收 token 费用 | 标价:Laguna S 2.1 输入 $0.09/M,输出 $0.18/M | 2026 年 8 月记录到 10% 降价;批量折扣未知 | Tech Insider 市场调查与定价指数 |
| 企业主权平台 | 年度软件许可证,加上计算基础设施转收费 | 实际 ACV 估计超过 $1M;未公布标价 | 折扣按前置部署工程承诺定制 | 行业分析与市场画像 |
| 战略基础模型授权 | 定制的双边技术转让与软件许可证 | Model Factory 软件实际交易金额为 $6B | 条款为非排他性;持续版税或支持义务未披露 | 金融分析师文件与公司画像报告 |
开发者端点和 API 层级的标价已公开;企业及主权条款由双方按客户逐一私下协商。
[CI001, CI012, CI014]4.3 成本结构与基础设施经济性
前沿代码模型开发的成本结构资本密集度极高,主要由预训练算力和集群运营驱动。2026 年以前,Poolside 的年化算力支出估计约为 $400M,在实现正毛利前造成了大量现金消耗。为获得主权算力独立性,Poolside 于 2025 年 10 月宣布 Project Horizon,计划在得州 Pecos County 建设雄心勃勃的 2 吉瓦数据中心园区,并通过表后空气衍生涡轮机利用廉价的放空天然气。根据初始协议,CoreWeave 将作为首个 250MW 阶段的 15 年锚定租户,并同意提供一个配备超过 40,000 台 Nvidia GB300 NVL72 GPU 的 AI 集群。但 Poolside 同期筹备的 $2B C 轮融资于 2026 年春季未能完成后,CoreWeave 终止租约,并将 40,000 台 GPU 调配至其他设施。Poolside 随后采取防守性现金保全措施,暂停交付周期长的燃气涡轮机承诺,并重组场地开发以保护运营流动性。[CI003, CI004, CI005, CI006, CI007, CI011, CI013]
| 指标 | 数值 | 置信度 | 意义 | 尽调问题 |
|---|---|---|---|---|
| 年度经常性收入(ARR) | 低于 $50M(据报道为 2026 年第一季度) | 中 | 对照已融资 $1.6B 以上资本,衡量商业进展基线 | 查阅经审计 ARR 明细及流失客户群 |
| 年度计算支出 | 约 $400M(Horizon 之前,估算值) | 中 | 决定毛利率和烧钱速度的主要成本驱动因素 | 查阅云托管协议及 GPU 供应商承诺 |
| 软件毛利率 | 未公开披露 | 低 | 决定长期现金流状况及经营杠杆 | 拆解 GPU 计算、托管和支持之间的 COGS |
| 现金余额与资金续航 | 由 $1B 股权融资 + $6B 授权收入支撑(2026 年 8 月) | 中 | 表明 Nvidia 交易后具备多年资本自给能力 | 核实资金配置、托管扣留款及税务负债 |
据报道的估算值和分析师区间构成这些数值;精确 GAAP 数据属于保密的私营公司信息。
[CI006, CI008, CI009, CI010, CI014]4.4 资本充足性与尽调结论
Poolside 的融资轨迹既暴露出前沿人工智能资本化的脆弱性,也显示出其韧性。公司在 2024 年 10 月 B 轮融资 $500M、估值 $3B 后,$2B C 轮于 2026 年 4 月在投资人质疑每年 $400M 算力支出、而收入不足 $50M 的情况下失败,形成生死攸关的融资瓶颈。然而,Poolside 于 2026 年 8 月通过与 Nvidia 的双重交易大幅重塑资产负债表:Nvidia 以 $12B 投前估值进行 $1B 主要股权投资,同时为 Model Factory 技术签署 $6B 非排他性软件许可协议。这笔交易把 Poolside 的累计主要股权融资扩大至 $1.63B,并带来非股权现金储备,在无需立即依赖融资的情况下确保数年现金跑道。不过,从投资判断角度看,财务结论仍需谨慎:许可收入隔离了近期偿付能力风险,但底层软件业务仍需就经常性毛利、客户留存和可持续单位经济性开展实证尽调。[CI002, CI005, CI006, CI008, CI009, CI010, CI014]
4.5 展示项
05产品与技术
5.1 产品界面与客户工作流
Poolside 的产品不只是模型端点:公开产品描述涵盖开发者界面、集中管理的集成、容器化执行和管理控制。Poolside 称,智能体可运行于 VS Code、Visual Studio、Zed、IntelliJ 和终端;AI Engineer 还介绍了一款 macOS Desktop Assistant,能够跨项目运行多个代码智能体。设计中的工作流从开发者现有工作环境开始,通过受管理连接向智能体提供代码库和工具权限,在受控容器中执行每次会话,并为管理员记录最终操作。对于不能接受源代码外传的买方,这种设计差异化最明显;但指定证据没有量化开发者采用率或生产率提升。[CE001, CE002, CE003, CE012]
| 模块或界面 | 主要用户 | 有证据支持的状态 | 差异化 | 尽调缺口 |
|---|---|---|---|---|
| Poolside 平台与控制台 | 企业平台与安全团队 | 2026 年 5 月宣布为生产级平台 | 在客户边界内执行 agent,并集中管理控制 | 独立的生产规模可靠性与采用证据 |
| Malibu Point 与 Laguna 产品线 | 软件工程师与受监管组织 | 独立目录报道的当前 2026 年产品线 | 覆盖复杂工作完成与高安全部署的编码任务专业化 | 核心规格及独立复现的模型基准 |
| Atlas 推理引擎 | 基础设施与模型服务团队 | 作为政府技术栈的一部分提供 | 公司声称在现有算力上最多提升 30% 性能 | 基准测试方法、硬件基线及工作负载分布 |
| IDE 终端与桌面界面 | 开发者与工程团队 | 官方及开发者社区来源均描述了多个界面 | 接入开发者已有工具,同时保持控制集中 | 活跃用户留存、使用频率及扩展采用率 |
这些状态区分公司公告与第三方描述;任何一行都不应被解读为独立认证或采用证明。
[CE001, CE002, CE006, CE008, CE012]5.2 架构、模型与训练机制
证据支持的是分层运营模型,而非单一代码助手。开发者界面位于智能体执行和 Model Context Protocol 集成之上;治理控制约束凭据、网络和审计记录;Poolside 模型和 Atlas 推理引擎提供智能与服务层;部署则运行在客户控制的基础设施或云端虚拟私有网络上。第三方资料称,2026 年产品线包括 Malibu、Point 和 Laguna;Poolside 政府页面则报称 Atlas 可将模型性能提高最多 30%。AI Engineer 和 SentinelOne 都描述了来自代码执行的强化学习,即由编译和测试结果提供反馈。这些架构和性能说法大多来自厂商或投资人,而非独立基准测试。[CE004, CE005, CE006, CE008, CE013, CE015]
| 层 | 角色 | 有证据支持的依赖 | 主要风险 |
|---|---|---|---|
| 开发者界面 | 在 IDE、桌面和终端界面捕捉编码任务 | 客户端集成与代码库访问 | 公开使用量和兼容性遥测有限 |
| Agent 运行时与集成 | 执行会话并连接获批的企业工具 | 容器、托管 MCP 服务器及凭证 | 工具循环错误、连接器缺口及特权操作风险 |
| 治理控制 | 应用网络策略、脱敏、访问控制和可审计性 | 正确的客户策略配置与完整轨迹捕捉 | 配置错误、日志不完整及控制效果未经验证 |
| 模型与训练 | 提供代码推理和任务执行能力 | Laguna 或其他选定模型,加上代码执行反馈 | 仅文本能力,以及不确定的未来前沿模型路线图 |
| 推理与部署 | 在客户控制的基础设施中提供模型服务 | Atlas 客户算力或云 VPC 容量 | 硬件规格、成本性能及更新负担 |
该架构综合自官方产品描述和第三方技术画像;未披露的实施细节仍是尽调缺口。
[CE002, CE003, CE004, CE005, CE006, CE013, CE014]公开证据描述了从开发者界面到客户控制部署的六层架构,各层之间存在治理与模型服务依赖。
- 开发者界面
- IDE、桌面端与终端界面收集工程任务。
- 模块
- IDE 集成
- 桌面端与终端代理
- 代理执行
- 容器化会话执行文件与工具操作。
- 模块
- 会话容器
- 代理轨迹
- 企业集成
- 集中管理的 MCP 服务器连接获批准的业务与工程工具。
- 模块
- MCP 服务器
- 工具连接器
- 治理
- 网络策略、机密处理、脱敏与审计记录约束执行。
- 模块
- 策略控制
- 可搜索审计记录
- 模型与推理
- 面向编码的模型与 Atlas 提供推理和本地服务。
- 模块
- Poolside 模型
- Atlas 推理
- 客户基础设施
- 裸机、气隙隔离的合作伙伴硬件或客户 VPC 部署,使执行始终处于选定边界内。
- 模块
- 本地或气隙隔离
- AWS、Azure 或 Google Cloud VPC
这是对已描述组件的综合,并非已披露的底层网络图;组件排序不能证明实际实现或控制有效性。
[CE004, CE005, CE015]5.3 部署、实施与运营成熟度
Poolside 称,Platform 可运行在客户裸金属、合作伙伴交付的气隙硬件,或现有 AWS、Azure 和 Google Cloud 虚拟私有云中。其政府方案更进一步,承诺交付完整模型权重,并在没有外部调用、数据外传或共享系统的情况下运行。实施需要服务支持:Poolside 称,前置部署研究工程师会与客户团队一起识别工作流,在数周内交付初始自动化,然后加固并衡量效果。这种支持模式可能降低实施摩擦,但也让交付能力成为依赖项。指定来源没有提供公开服务级别协议、状态历史、事故记录、部署时长分布或大规模生产可靠性的独立证据。[CE004, CE005, CE007]
| 用户任务 | 当前约束 | Poolside 方案 | 公开资料可支持的收益 | 限制或尽调问题 |
|---|---|---|---|---|
| 在敏感代码库上运行编码 agent | 源代码不能离开安全边界 | 在客户控制的基础设施中部署 agent 和模型 | 避免 Poolside 所描述的云端出站模式 | 通过架构审查和渗透测试核实边界控制 |
| 将 agent 接入企业工具 | 凭证和工具访问需要中央策略 | 使用托管 MCP 服务器,在运行时注入并脱敏凭证 | 集中工具访问,同时保留可搜索的轨迹 | 测试连接器覆盖范围、权限边界及故障处理 |
| 在断网政府环境中运行 | 公有云 API 可能不可用或被禁止 | 提供完整模型权重和气隙推理 | 按 Poolside 的说法,去除对外部推理端点的依赖 | 验证硬件规格、打补丁、更新及模型治理流程 |
| 从试点转入生产工作流 | 客户团队可能缺乏 agent 实施经验 | 用前置部署的研究工程师搭建并加固初始工作流 | Poolside 表示初始工作流可在数周内上线 | 获取部署周期分布、人员配置比例及实测客户结果 |
这些收益是供应商描述的设计结果;本分配部分没有受控生产力研究或部署客户群数据。
[CE002, CE003, CE004, CE005, CE007]5.4 差异化、信任控制与技术风险
Poolside 最清晰的差异化是主权能力:完整权重、本地推理、网络策略执行、密钥处理和可搜索轨迹,都是为受监管或断网环境设计的。现有证据无法证明这些控制已通过指定第三方认证或独立渗透测试,因此应将其视为设计主张,而不是已验证的保障。外部模型目录提供了有用但有限的信号:LM Market Cap 列出 Poolside 的四个变体,四者都支持函数调用和推理,但都不支持视觉能力;Artificialus 也将这些模型描述为纯文本。AI Engineer 还报道,Poolside 在 2026 年 8 月与 Nvidia 达成安排后停止前沿模型开发,却没有披露更新后的愿景,造成路线图不确定性,需要直接审查合同和产品规划。[CE003, CE005, CE009, CE010, CE011, CE014, CE016]
06客户
6.1 企业与公共部门客户分层
Poolside 将商业和公共部门市场进入战略集中在高度受监管、对安全极为敏感的大型组织。这些组织使用标准多租户云 AI 代码工具时会遇到严格合规障碍。公司的主要买方是企业 CISO、CIO 和联邦项目经理,其机构政策严禁向第三方端点传出源代码。Poolside 不销售通用 SaaS 订阅,而是围绕三个核心目标细分市场提供 Poolside Platform:需要气隙部署的国防和国家安全组织;在监管监督下管理复杂遗留代码库的一级金融服务机构;以及拥有大量内部工程团队的大型科技企业。在这些细分市场,Poolside 提供可预测的订阅价格和对模型权重的完整控制,而不是按 token 消费计费,直接解决大企业 AI 采用中经常导致项目停滞的预算波动和数据主权问题。[CU001, CU002, CU003]
| 客户群体 | 目标买方与付费方 | 核心软件工程用例 | 目标部署规模 | 商业与战略价值 | 关键尽调缺口 |
|---|---|---|---|---|---|
| 国防与情报 | CISO、项目执行官、国防集成商负责人 | 传统任务现代化、气隙代码生成、安全 DevSecOps 流水线 | 气隙集群、边缘工作站、战术服务器 | 高 ACV 的多年主权许可证;战略性国防锁定 | 未公布的涉密合同金额和确切席位数 |
| 受监管金融服务 | 首席信息安全官、企业工程负责人 | VPC 隔离的自动化测试、代码库搜索、无需云端出站的代码重构 | 拥有 5,000 名以上开发者的大型企业 | 可预测的年度订阅,取代波动的云 token 账单 | 企业安全例外试点在全面推出前的持续时间 |
| 企业基础设施与 AI 科技 | 工程副总裁、开发者平台负责人 | 专业化 agent 编码编排、在专有代码库上执行定制模型推理 | 多代码库工程组织(100–1,000 名以上工程师) | 战略共创、基准可引用性、模型使用反馈 | 早期 AI 基础设施合作伙伴的客户集中度 |
| 联邦民用机构 | 机构 CIO、联邦技术采购负责人 | 传统政府代码库重构、加速 ATO 合规文档编制 | 通过 Poolside Federal LLC 服务联邦机构 IT 分支(CAGE 11R53) | 任务级多年机构授权,零出站风险 | 采购周期以及 FedRAMP / STIG 合规时间线 |
客户分群综合自 Poolside 官方公共部门及平台架构披露,以及截至 2026 年 10 月的独立行业研究报告;合同金额未公开。
[CU001, CU002, CU003, CU005, CU006]6.2 采用轨迹与增长指标
Poolside 的目标客户是拥有 5,000 名或更多开发者的企业,商业扩张重点是高年度合同价值关系,而非面向大众的开发者自助服务。公司尚未公开发布经审计的 GAAP 收入数据或正式付费账户数,但 GetLatka 的私有市场情报和调查估算显示,Poolside 年化收入运行率在 2024 年 12 月达到约 $30M,并在 $500M B 轮融资后于 2025 年初接近 $50M。这一估计轨迹显示,公司已从早期实验性试点大幅扩展到每年数百万美元的软件和专业服务合同。但要评估 Poolside 的真实商业采用,必须看到公开报道缺失的分母:付费生产账户总数、平均合同金额,以及经常性软件许可与前置部署工程专业服务之间的确切划分,仍是未经确认的私有数据。此外,Tracxn 识别出更广义 AI 软件工程领域有超过 1,100 家活跃竞争者,说明 Poolside 必须持续证明相对于既有工具的差异化投资回报。[CU001, CU009, CU011, CU012]
| 指标 | 报告 / 估算值 | 观察日期 | 来源 | 置信度 | 战略含义 | 缺失分母 |
|---|---|---|---|---|---|---|
| 估算年化收入 | $50M | 2025-03 | GetLatka Private Equity Survey | 中 | 显示收入正从早期试点合同加速迈向数百万规模的商业承诺 | 经审计的 GAAP 确认收入、经常性 ARR 拆分及合同起始日期 |
| 上年度估算收入 | $30M | 2024-12 | GetLatka Private Equity Survey | 中 | 继 2024 年末 $500M B 轮扩张后,显示快速上行轨迹 | 合同数量、平均售价(ASP)以及已确认与已签约积压订单的区分 |
| 每个账户的目标开发者数量 | 5,000+ 名开发者 | 2025-10 | Sacra Company Diligence Report | 中 | GTM 聚焦能够承接大量定制部署的超大型企业账户 | 企业账户总清单及已部署开发者席位的实际渗透率 |
| 活跃企业竞争者 | 1,183 个市场实体(159 个已融资) | 2023-08 | Tracxn 企业情报 | 中 | 突显云托管与自托管开发者助手赛道的激烈拥挤 | 提供真正气隙本地化主权部署的竞争者子集 |
所有金融采用数据均为第三方分析师估算和调查数据;Poolside 未公开发布经审计财务指标、ARR 或官方账户数量。
[CU001, CU009, CU011, CU012]6.3 已披露客户部署与国防验证
Poolside 可公开引用的客户部署主要集中在国家安全、国防科技和专业基础设施工程。软件供应链安全公司 Hunted Labs 已部署 Poolside 模型,既用于加强内部安全代码开发,也用于支持美国作战人员的软件安全项目。在联邦公共部门,Poolside 通过 Poolside Federal LLC(CAGE Code 11R53)运营,与 Vibrint、Sterling Computers 等成熟国防系统集成商合作,向机密政府设施交付气隙代码模型和全栈 AI 工作站。此外,国防行业情报报告称,包括 RTX Corporation 在内的美国主要航空航天和国防承包商,已在安全计算设施中评估或部署 Poolside 的可定制基础模型。这些国防级部署验证了 Poolside 独特的气隙价值主张,但合同细节和生产席位数仍受保密约束。[CU004, CU005, CU006]
| 客户 / 组织 | 客户群体 | 部署与工作流用例 | 生产与试点状态 | 已记录结果与影响 | 证据限制 |
|---|---|---|---|---|---|
| Hunted Labs | 软件供应链安全 | 面向国家安全任务的内部安全代码开发和软件安全工具 | 生产部署 | 从根本上改变内部安全编码工作流;已部署以支持作战人员国防任务 | 未披露具体席位数、合同金额和量化吞吐指标 |
| Vibrint | 联邦国防系统集成商 | 为敏感联邦任务和作战人员战备提供气隙模型交付 | 生产集成 / 战略合作伙伴 | 在不牺牲安全性的情况下,将先进 AI 能力部署到涉密政府计算环境中 | 合同规模、活跃国防项目名称和部署时间线已列为机密或被隐去 |
| Sterling Computers | 公共部门技术提供商 | 为公共部门机构提供交钥匙气隙 AI 辅助开发环境 | 生产渠道合作 | 为要求零云端出站的公共部门客户提供主权全栈 AI 开发能力 | 终端机构身份、席位数量和渠道收入分成未公开 |
| RTX Corporation | 航空航天与国防工业基础 | 本地代码生成、代码库重构及安全工程辅助 | 试点 / 企业评估 | 证明国防级客户对可定制模型的兴趣,这些模型适配公司特定编码实践 | 公司官方新闻确认及转入生产的状态仍未经验证 |
枚举反映供应商公告和国防行业报道中披露的公开参考客户与集成伙伴;银行和零售领域的商业企业账户仍匿名。
[CU004, CU005, CU006]6.4 客户旅程、扩张循环与留存机制
Poolside 的客户采用旅程与依靠开发者自下而上自助采用的传统开发者工具明显不同。合作通常从 CISO 和基础设施架构层面开始,安全团队会验证 Poolside Platform 能否在客户的安全边界内运行,包括裸金属、气隙集群或私有云 VPC,并且无需外部连接。为跨过从初始部署到可衡量生产率的鸿沟,Poolside 派出前置部署研究工程师,与客户工程团队并肩工作,识别高杠杆工作流,并通过 Model Context Protocol 服务器将模型接入内部工具链。基础设施伙伴 Baseten 已实现 Laguna 模型家族的快速部署和亚秒级响应,这说明运营性能对开发者满意度至关重要。但 Poolside 没有公开净收入留存率、毛续约率或合同期限分组数据,因此扩张循环的持久性,以及初始试点能否转化为全企业开发者推广,仍是重要尽调问题。[CU002, CU007, CU008, CU010, CU014]
从 CISO 架构评估到前置部署工程扩展与主权多代理自动化的六阶段企业客户采用生命周期。
旅程阶段基于截至 2026 年 10 月的 Poolside Platform 架构披露、企业案例研究与合作伙伴部署工作流建模。
[CU002, CU003, CU008, CU010]6.5 采购摩擦、客户集中度与竞争压力
尽管主权定位强劲,Poolside 仍面临由企业销售周期长和竞争压力大带来的实质性客户采用风险。Sacra 的独立研究强调,向《财富》500 强银行和国防机构销售复杂的本地部署、气隙 AI 系统,需要 9–18 个月采购周期、严格安全审计和高客户获客成本。Quasa 的从业者评测还指出,Poolside 明确不是开箱即用的消费者工具,要取得最佳效果必须配置专用基础设施、整合数据并建设内部平台工程。在更广泛的市场中,Poolside 必须直接对抗 Microsoft 和 GitHub Copilot;后者拥有超过 2,000 万用户,在《财富》100 强中的渗透率达到 90%,同时还要面对开放权重代码模型的快速进步。这些竞争动态意味着,任何企业预算收紧或 AI 平台落地延迟,都可能增加现金消耗,并提高 Poolside 早期核心客户的集中度风险。[CU011, CU012, CU013, CU014]
6.6 展示项
07风险
7.1 监管、知识产权来源与许可风险
Poolside 发布开放权重模型的基础战略,为企业采用者带来了重大监管和知识产权风险。公司没有详细公开 Laguna 模型家族预训练数据集和筛选来源,给企业法务审查带来潜在版权暴露和合理使用审查风险;相关数据来源的合规性也尚未得到证实。此外,Poolside 在不同模型版本中采用异质许可模式:Laguna S 2.1 使用 OpenMDW-1.1,同时 Laguna M.1 继续使用 Apache 2.0,使企业无法建立统一的开源治理框架。为缓解主权和国防部门的采购限制,公司成立了专门的美国实体 Poolside Federal LLC,配备 CAGE code 11R53 和通过安全审查的解决方案架构师;但生成软件成果的长期责任仍未获得赔偿保障。[CR001, CR002, CR004]
| 规则 / 许可证 / 案件 | 司法辖区 | 状态 | 可能性 | 严重性 | 缓释措施 | 剩余暴露 | 尽调路径 |
|---|---|---|---|---|---|---|---|
| 训练数据版权与合理使用 | US / EU | 尚未解决;未披露数据集审计 | 中 | 高 | 通过 RLCEF 执行 GPL/AGPL 过滤流程及合成数据生成 | 高 | 查阅数据集排除项及客户 IP 赔偿条款 |
| 模型权重授权一致性(OpenMDW 与 Apache 2.0) | 全球 | Laguna 各代产品采用活跃的多许可证制度 | 高 | 中 | 在 Hugging Face 模型卡片上发布宽松许可证声明 | 中 | 核实 OpenMDW-1.1 相对于标准 Apache 2.0 的采购审批 |
| 联邦安全许可与 CAGE 注册 | 美国 | 合规;Poolside Federal LLC CAGE 11R53 | 低 | 中 | 由美国注册实体配备持证人员 | 低 | 确认现行国防合同发起方及设施安全许可 |
| 跨境双地运营(旧金山 / 巴黎) | US / France | 公司双地存在仍在运营 | 中 | 低 | 特拉华州母公司控股结构,在巴黎设有研究团队 | 低 | 审计跨境 IP 转让定价与出口管制分类 |
基于公开实体申报文件、Hugging Face 模型卡和供应商文档的监管与合规评估,截至 2026 年 10 月。
[CR001, CR002, CR004]7.2 运营可靠性、安全架构与模型生命周期
将生成式代码智能体直接部署到企业环境,会带来尖锐的运营和运行时安全挑战。Poolside 通过气隙本地架构和 Atlas 运行引擎,阻止数据越过客户防火墙外传,从而降低外部数据泄露风险。其 Reinforcement Learning from Code Execution Feedback 架构每分钟可在 3.5 亿个已索引代码库上执行最多 10,000 次容器化代码运行,试图用自动执行反馈消除幻觉语法。但由于没有公布模型弃用路线图,运营持久性仍受约束,客户可能突然面对架构过时。此外,验证气隙安装、执行定制安全审查和配置开发环境集成所需的漫长评估周期,也会造成严重的企业销售摩擦。[CR003, CR005, CR006, CR014]
7.3 伙伴依赖、算力波动与战略重整
Poolside 的基础设施路线曾遭遇严重外部干扰,技术和商业依赖因此被重新定义。2025 年 10 月,公司宣布与 CoreWeave 合作在得州建设专用 2 吉瓦 AI 算力园区;但协议于 2026 年 4 月双方共同终止,导致原计划的 40,000 台 GPU 集群落空,并迫使公司依赖次级云资源。2026 年 8 月,Nvidia 为 Poolside 的 Model Factory 执行了一项标志性的 $6B 非排他性软件许可,同时直接投资 $1B,投前估值为 $12B。虽然这项合作带来巨额非稀释性流动性,但约 109 名核心研究和工程员工转至 Nvidia,威胁人才连续性和技术独立性。与此同时,商业分销仍绑定于与 Amazon Web Services 就 Bedrock 和 Trainium 硬件达成的多年协议。[CR007, CR008, CR009, CR010]
| 依赖项 | 交易对手 | 角色 | 集中度 | 失效场景 | 严重程度 | 缓解措施 | 残余暴露 |
|---|---|---|---|---|---|---|---|
| 战略计算资源与 Model Factory 许可 | Nvidia | 战略股权投资者与软件被许可方 | 关键 | Nvidia 开发专有内部工具,使已许可软件变得多余 | 高 | $6B 非独家许可与保留独立性的公司治理 | 高 |
| 高性能计算集群 | CoreWeave | 前 Project Horizon 核心计算资源供应商 | 高 | 合同终止后只能依赖二级云供应商 | 高 | AWS Bedrock 与多云备用部署选项 | 中 |
| 企业云分发与模型托管 | Amazon Web Services (AWS) | 托管 Bedrock API 与 EC2 Trainium 基础设施合作伙伴 | 中 | AWS 在 Bedrock 目录中偏好自有或替代编码模型 | 中 | 气隙隔离的本地交付降低纯云锁定 | 中 |
| 网络安全战略分发与联合销售 | SentinelOne (S Ventures) | 战略风险投资者与网络安全分发合作伙伴 | 低 | 渠道销售优先级下降或竞争性产品捆绑 | 低 | 多渠道企业销售策略与直接联邦销售团队 | 低 |
按交易对手风险暴露排序的第三方商业、计算资源与平台依赖分析。
[CR007, CR008, CR009, CR010]7.4 战略风险综合、财务不透明与论点失效触发器
2026 年 8 月 Nvidia 交易的结构性后果,似乎是 Poolside 从前沿基础模型研究转向应用层企业智能体工作流,同时据报道暂停了专有预训练工作。Poolside 在四轮融资中累计获得约 $1.63B 股权资本,另有 $6B 软件许可收入,为公司提供数年现金跑道,隔离近期市场下行风险。尽管如此,根本性的财务不透明仍在:截至 2025 年,估计年收入 $50M 尚未验证,GAAP 软件利润率未披露,底层托管费用也不透明。关键论点失效触发器包括人才流失超过交易后的留存目标、云主机积极蚕食一方 Copilot,或针对开放权重模型输出的知识产权侵权挑战未获缓解。[CR011, CR012, CR013, CR014]
风险评估矩阵,从影响、可能性、缓解措施与残余暴露等维度评估主要运营、依赖、治理和法律风险。
定性风险评级综合已报道交易条款、许可文档与公有云合作披露。
[CR001, CR008, CR009, CR010, CR011, CR012]7.5 展示项
08估值
8.1 投资论点与主权 AI 定位
Poolside 放弃标准公有云多租户交付,转而采用完全主权化、物理隔离的本地基础模型部署,由此在企业市场 站稳了差异化位置。Poolside 的 Laguna 模型家族采用代码执行反馈强化学习(RLCEF)训练,专为严格客户 安全边界内的复杂软件开发任务设计。这套架构面向高安全等级的公共部门、国防工业基地和受监管金融机构; 这些机构在制度上无法把专有源代码发送给多租户商业云 API。 但这套投资逻辑仍受到显著运营和商业不确定性的制约。CoreWeave Project Horizon 合作于 2026 年 4 月 双方终止,凸显 Poolside 对波动性第三方基础设施承诺的暴露。同时,公司没有公开客户案例或经审计的 ARR 指标,外部无法独立验证商业产品市场匹配度。技术价值主张对主权防务应用仍很有吸引力,但商业规模尚未 尚未得到证明。[CV002, CV010, CV006, CV007]
| 正论 / 反论 | 核心论点 | 证据基础 | 哪些变化会改变判断 |
|---|---|---|---|
| 正论:主权企业需求 | 国防与金融监管企业需要零数据外传的气隙隔离本地代码 LLM | 与 Vibrint、Sterling Computers 和 Hunted Labs 的合作 | 更广泛的企业流失,或企业普遍采用公有云托管的副驾驶工具 |
| 正论:专用 Model Factory 优势 | 基于代码执行反馈的强化学习(RLCEF)打造专有的自动化模型开发资产 | Nvidia 对 Model Factory 软件平台的 $6B 非独家许可 | 前沿实验室使用商品化合成流水线取得相当的编码基准成绩 |
| 反论:资本与计算资源负担 | CoreWeave 2GW 德州园区失败,令 $2B C 轮融资尝试告吹,也加剧资本脆弱性 | CoreWeave 项目于 2026 年 4 月双方终止;失去 40k GB300 集群 | 获得永久的主权计算基础设施,并分散电力与硬件接入 |
| 反论:商业收入不透明 | 未披露经审计收入、ARR 或净留存指标,$12B 估值缺乏基本面支撑 | 公开估算为 $30M-$50M,而官方披露仍为零 | 披露超过 $100M 且企业净收入留存率处于行业前四分位的经审计 ARR |
核心论点综合公司定位与独立技术报告;null 表示不适用的触发条件。
[CV002, CV006, CV007, CV010]8.2 融资历史与 Nvidia 交易结构
Poolside 的资本化历史显示,战略硬件合作推动估值快速攀升。早期轮次累计融资 $126 million 后,公司于 2024 年 10 月完成 $500 million 的 Series B,投后估值 $3 billion,由 Bain Capital Ventures 领投, Nvidia 和多家大型机构跟投。2026 年 8 月,Nvidia 做出总额 $7 billion 的标志性承诺:其中包括为 Poolside Model Factory 软件支付 $6 billion 的非独家许可费,以及按 $12 billion 投前估值进行的 $1 billion 直接股权投资。 这笔交易的独特结构对公司治理和财务有深远影响。$6 billion 许可费计划直接分配给早期投资人,而不是 充入 Poolside 的日常运营资产负债表,实际上成为老股东的流动性退出事件。此外,Nvidia 向 Poolside 的 109 名研究和系统工程师发出聘用邀约。联合创始人 Jason Warner 和 Eiso Kant 仍保持公司独立性, 但运营经验转移至 Nvidia、资金流向被拆分,形成了非同寻常的治理结构,潜在股权投资人必须严格评估。[CV001, CV003, CV004, CV005, CV009]
8.3 情景分析、可比公司与回报逻辑
评估 Poolside 的合理企业估值,必须正视私募市场表面估值与可验证软件基本面的巨大背离。在 缺乏经审计财务披露的情况下,二手估算显示年收入 $50 million;按 $12 billion 估值计算,企业价值倍数 超过收入的 240x。即使拿高增长 AI 基础设施同业和高估值公开 DevOps 公司(如 IPO 时的 GitLab 与 Snowflake)作基准,Poolside 仍享受极高增长溢价,仿佛未来多年都能无瑕疵地执行商业化。 我们的情景模型把这些结构性矛盾纳入三个可能结果。基准情景(概率 55%)假设企业 ARR 稳步接近 $100M-$150M,倍数压缩后趋于 70x-100x,对应标准化估值 $10B-$14B。乐观情景(概率 20%)预计估值 达到 $20B-$30B,前提是主权防务订单带来垄断式定价能力,且 Model Factory 产生持续许可收入。相反, 悲观情景(概率 25%)计入 Nvidia 交易后的人员流失和企业销售阻力,预计倍数压缩至 Series B 重组水平 $3B-$5B。[CV007, CV008, CV011, CV012]
| 情景 | 关键假设 | 估值 / 回报逻辑 | 关键下行触发因素 | 概率信号 |
|---|---|---|---|---|
| 乐观情景 | 主权本地编码成为国防 / 金融领域的必需品;Model Factory 带来经常性许可收入 | 通过 IPO 或超大市值战略收购实现 $20B-$30B 终值(相对 $12B 基准的 2.0x-2.5x MOIC) | 超大规模云厂商推出经认证的气隙隔离编码设备,以更低价格竞争 | 20% |
| 基准情景 | Poolside 拿下专门的主权国防细分市场,Nvidia 集成则推动生态稳步采用 | $10B-$14B 估值区间(约 1.0x MOIC),2027 年平价或小幅上调融资 | 企业销售周期拉长,推理与研发持续高额烧钱 | 55% |
| 悲观情景 | 编码模型商品化;Nvidia 向 109 名员工发出邀约后人才流失;折价轮或清算 | $3B-$5B 估值(Series B 重组水平,资本损失 60%-75%) | 无法将联邦试点转化为企业 ARR,或投资者分配资金耗尽 | 25% |
情景回报假设入场估值为 $12B;概率信号反映作者基于计算资源接入与商业销售漏斗的权重判断。
[CV008, CV011, CV012, CV004]从主权企业需求出发,经由资本冲击与估值张力,追踪至最终建议的决策路径。
| 节点 / 连接 | 节点 / 起点 | 终点 | 说明 |
|---|---|---|---|
| 节点 1 | 企业主权需求 [tech-eval] | 国防与金融机构需要不发生外部数据泄漏的气隙隔离本地代码 LLM。 | |
| 节点 2 | Nvidia Model Factory 交易 [licensing-deal] | 非独家 $6B 许可验证了自动化模型创建平台与软件工程专长。 | |
| 节点 3 | 投资者分配条款 [governance-overhang] | $6B 收益流向原股东,而不是运营研发资产负债表账户。 | |
| 节点 4 | CoreWeave 园区崩溃 [capital-shock] | 德州 2GW 园区终止且 $2B Series C 融资失败,造成基础设施集中。 | |
| 节点 5 | 估值立场:偏高 [valuation-gap] | 在缺少公开 ARR 指标的情况下,账面 $12B 估值意味着超过 240 倍的预估收入倍数。 | |
| 节点 6 | 建议:继续研究 [final-recommendation] | 要求直接审计联邦销售漏斗转化、经常性毛利率与交易后的治理。 | |
| 连接 1 | 企业主权需求 [tech-eval] | Nvidia Model Factory 交易 [licensing-deal] | 验证技术 |
| 连接 2 | Nvidia Model Factory 交易 [licensing-deal] | 投资者分配条款 [governance-overhang] | 分配现金 |
| 连接 3 | 投资者分配条款 [governance-overhang] | CoreWeave 园区崩溃 [capital-shock] | 叠加影响 |
| 连接 4 | CoreWeave 园区崩溃 [capital-shock] | 估值立场:偏高 [valuation-gap] | 支撑立场 |
| 连接 5 | 估值立场:偏高 [valuation-gap] | 建议:继续研究 [final-recommendation] | 得出评级 |
基于定性交易分析与量化估值倍数的顺序决策层级。
[CV002, CV003, CV004, CV006, CV008, CV013]8.4 最终建议、信心度与尽调要求
我们给出“继续研究”建议,信心度中等、风险评级高,并认为估值已明显透支。Poolside 拥有出色的技术资产, 也契合主权防务需求;但在 $6B 新进资金被分配到公司账外的情况下,按 $12B-$13B 估值入股,财务上并不 平衡。资本配置者不应把表面估值当作经营现金流或短期流动性的验证。 严谨的投资委员会必须把未来投资决策建立在严格的尽调门槛上。关键要求包括核查经审计的 GAAP ARR、客户 客户群体扩张指标,以及扣除内部推理和集群成本后的毛利率情况。此外,法律顾问必须审计 Model Factory 与保留企业工具之间明确的 IP 边界,以及与 Nvidia 的附函治理条款。明确的论点失效触发因素包括 联邦试点不续约、管理层意外离职,或妨碍模型持续开发的算力供应限制。[CV013, CV014, CV004, CV006]
| 维度 | 评估 | 证据锚点 | 决策含义 |
|---|---|---|---|
| 建议 | research-more | 前沿模型能力已获证明;商业收入可见度不透明 | 在审计确认收入与企业留存前,暂缓直接股权承诺 |
| 信心 | medium | 交易结构经二手报道确认;客户指标未披露 | 将商业采用情况与企业参考客户及销售漏斗转化率交叉核对 |
| 风险评级 | high | 对计算资源高度依赖,且关键客户 / 投资者集中于 Nvidia | 在条款清单中要求下行保护与隔离的 IP 审计 |
| 估值立场 | stretched | 投资前估值随 $6B 投资者分配许可扩大至 $12B | 在没有已发布 ARR 支撑的情况下,以 $12B-$13B 入场要求极端增长假设 |
综合截至 2026 年 10 月已核实的公开报道与融资记录;null 值表示公司尚未披露经审计数据的指标。
[CV001, CV003, CV004, CV013]8.5 展示材料
免责声明
本报告基于截至运行日期的公开及第三方证据作出尽职调查评估,不构成投资、法律或会计建议。
证据索引
| 编号 | 陈述 | 可信度 | 来源 |
|---|---|---|---|
| CO001 | Poolside was founded in early 2023 with corporate registration in the United States and dual operational headquarters established across San Francisco and Paris. | 高 | SO006, SO007 |
| CO002 | Poolside Federal LLC operates as a United States domiciled corporate entity dedicated exclusively to US public sector and defense missions. | 中 | SO006 |
| CO003 | Poolside develops open-weight coding models including the Laguna model family (Laguna M.1, XS.2, and Laguna S 2.1) trained via reinforcement learning from code execution. | 高 | SO001, SO011, SO012 |
| CO004 | Poolside delivers foundation models for on-premises, VPC, and air-gapped deployments with unlimited inference on customer hardware and no outbound data egress. | 高 | SO006, SO008 |
| CO005 | Poolside's stated mission is that software engineering will be the first broad capability where artificial intelligence reaches and surpasses human-level intelligence. | 高 | SO004, SO008 |
| CO006 | Chief Executive Officer Jason Warner previously served as Chief Technology Officer at GitHub where he oversaw the development and release of GitHub Copilot and Actions. | 高 | SO008, SO010 |
| CO007 | Co-founder and Chief Technology Officer Eiso Kant previously pioneered codebase machine learning by founding source{d} in 2015 and engineering platform Athenian. | 高 | SO008, SO010 |
| CO008 | Former GitHub Chief Revenue Officer Paul St. John joined Poolside to direct enterprise commercialization and go-to-market execution. | 中 | SO010 |
| CO009 | In August 2026, 109 Poolside foundation model research engineers received job offers from Nvidia, while co-founders Warner and Kant remained at Poolside. | 中 | SO005, SO012 |
| CO010 | Poolside raised $626M in total venture capital through its October 2024 Series B round, which was led by Bain Capital Ventures at a $3.0B post-money valuation. | 高 | SO003, SO004, SO007 |
| CO011 | Third-party industry trackers estimate Poolside's annualized revenues between $50M and $66.4M, though Poolside does not publicly publish GAAP revenue figures. | 中 | SO002, SO009 |
| CO012 | In spring 2026, negotiations with CoreWeave for a 2GW Texas AI campus collapsed, terminating Project Horizon and causing a planned $2B Series C to fail. | 中 | SO005, SO011 |
| CO013 | On August 20, 2026, Nvidia agreed to a $6B non-exclusive commercial license for Poolside's Model Factory software platform, with proceeds slated for distribution to early investors. | 中 | SO005, SO011, SO012 |
| CO014 | Nvidia committed a separate $1B direct equity investment in Poolside in August 2026 that valued the startup at $12B pre-money, bringing total equity raised to $1.626B. | 中 | SO005, SO007, SO011 |
| CO015 | The August 2026 transaction with Nvidia was explicitly structured as a non-exclusive license and minority equity investment rather than an acquisition, leaving Poolside independent. | 中 | SO005, SO012 |
| CO016 | Poolside established a landmark sequence of funding, product, and infrastructure milestones spanning its 2023 inception through its 2026 Nvidia recapitalization. | 中 | SO003, SO005, SO011 |
| CM001 | Poolside positions its foundation models and software agent workflows primarily for enterprise and infrastructure deployment within customer-controlled environments rather than consumer developer workflows. | 中 | SM001, SM004 |
| CM002 | Enterprise deployment options for Poolside's platform span on-premises data centers, air-gapped turnkey Dell hardware, bare metal, AWS VPC, Microsoft Azure, and Google Cloud. | 中 | SM001 |
| CM003 | Poolside's platform pricing is enterprise-only with custom annual contracts and no public self-serve subscription tier or free trial evaluation path. | 中 | SM001 |
| CM004 | Poolside trains coding foundation models from scratch using reinforcement learning from code execution feedback across hundreds of millions of real-world tasks. | 中 | SM002 |
| CM005 | Poolside's self-managed deployment enables organizations to run model inference inside their VPC, on-premises data center, or air-gapped environment without code or prompt data leaving the boundary. | 高 | SM002, SM005 |
| CM006 | On August 20, 2026, Nvidia agreed to a non-exclusive $6 billion license for Poolside's Model Factory software platform alongside a $1 billion equity investment valuing the startup at $12 billion pre-money. | 中 | SM003 |
| CM007 | The $6 billion Nvidia Model Factory license payment is scheduled for distribution to Poolside's existing investors by the end of 2027 rather than directly funding company operating accounts. | 中 | SM003 |
| CM008 | CoreWeave data center negotiations for a two-gigawatt campus in Texas collapsed in spring 2026, creating severe infrastructure constraints that preceded Poolside's transaction with Nvidia. | 中 | SM003 |
| CM009 | Poolside operates in a crowded market landscape with over 1,180 active competitors in artificial intelligence systems, including major foundation model providers such as OpenAI, Mistral AI, and Anthropic. | 中 | SM006 |
| CM010 | Global demand for software development continues to outpace the supply of skilled engineers, driving enterprise interest in generative AI as a force multiplier for engineering velocity. | 中 | SM007 |
| CM011 | Poolside has executed contracts across the defense industrial base including the United States Department of Defense and RTX Corporation, and has pursued software engine sales to Israeli defense organizations. | 中 | SM008 |
| CM012 | Poolside delivers open-weight foundation models purpose-built for air-gapped and on-premises deployment without per-token charges, allowing public sector agencies to run unlimited inference on internal hardware. | 中 | SM005 |
| CM013 | Public sector partners including Vibrint, Sterling Computers, and Hunted Labs deploy Poolside models into classified and national security mission compute environments. | 中 | SM005 |
| CM014 | Poolside employs cleared personnel and solutions architects experienced in hardened Security Technical Implementation Guide environments to facilitate air-gapped enterprise and government deployments. | 中 | SM005 |
| CM015 | Poolside's Forward Deployed Research Engineers embed directly with customer engineering teams to reduce implementation risk and deploy production workflows within weeks. | 中 | SM001 |
| CM016 | Disclosed annualized recurring revenue, recognized contract values, and referenceable production deployment metrics for Poolside remain undisclosed in public domain reporting. | 低 | |
| CP001 | Market intelligence monitors over 1,100 active developer AI tools, categorizing OpenAI, Mistral AI, and Anthropic as leading foundation model alternatives. | 中 | SP004 |
| CP002 | Direct venture-backed AI coding competitors include Magic, which raised $320 million in August 2024, and Codeium, which secured $150 million at a $1.25 billion valuation. | 中 | SP003 |
| CP003 | Poolside raised a $500 million Series B in October 2024 led by Bain Capital Ventures with participation from Nvidia, Citi Ventures, and SentinelOne's S Ventures to fund frontier coding model development. | 高 | SP003, SP006 |
| CP004 | Market analysts identify Microsoft's GitHub Copilot as the dominant incumbent coding assistant across enterprise software engineering workflows. | 中 | SP007 |
| CP005 | Poolside designs its architecture for air-gapped and on-premises deployment, delivering full model weights rather than multi-tenant cloud API access. | 高 | SP005, SP006 |
| CP006 | The Poolside Platform incorporates centrally managed Model Context Protocol servers to integrate agents with tools including Slack, Jira, GitHub, Workday, and Salesforce within customer network boundaries. | 中 | SP002 |
| CP007 | Poolside trains its models using Reinforcement Learning from Code Execution Feedback to enable autonomous multi-step software engineering rather than static code completions. | 中 | SP006 |
| CP008 | On the SWE-bench Pro benchmark, Poolside's proprietary Laguna M.1 achieved a 46.9% score, while its open-weight 33-billion parameter Laguna XS.2 achieved 44.5%. | 中 | SP008 |
| CP009 | Laguna M.1 scored 72.5% on SWE-bench Verified, outperforming Devstral 2 at 72.2% while trailing Claude Sonnet 4.6 at 79.6%. | 中 | SP008 |
| CP010 | Poolside provides government and defense customers with unlimited inference on their own hardware under predictable subscriptions rather than charging per-token cloud API fees. | 中 | SP005 |
| CP011 | Running Laguna XS.2 locally requires at least 36 GB of unified memory on Apple Silicon or a 24 GB to 32 GB VRAM GPU using 4-bit quantization. | 中 | SP008 |
| CP012 | In August 2026, Nvidia agreed to pay Poolside a $6 billion non-exclusive license fee for its Model Factory software while investing $1 billion at a $12 billion valuation. | 中 | SP001, SP007 |
| CP013 | The Nvidia transaction absorbed 109 of Poolside's approximately 130 employees as new Nvidia hires, leaving three co-founders and approximately 20 employees at the independent company. | 中 | SP001 |
| CP014 | Poolside maintains strategic cloud and distribution agreements with Amazon Web Services for Bedrock and EC2 as well as sovereign solutions partnerships with Atos in Europe. | 中 | SP007 |
| CI001 | Poolside provides government and enterprise customers with full model weights for on-premises and air-gapped deployment with unlimited inference and zero per-token fees. | 中 | SI003 |
| CI002 | In October 2024, Poolside raised a $500 million Series B round led by Bain Capital Ventures at a $3 billion valuation with strategic participation from SentinelOne's S Ventures. | 中 | SI006, SI007 |
| CI003 | In October 2025, Poolside announced Project Horizon, a planned 2-gigawatt AI data center campus in Pecos County, Texas, with CoreWeave signing on as anchor tenant for an initial 250MW phase. | 高 | SI002, SI004, SI005 |
| CI004 | CoreWeave contracted to deliver an AI compute cluster comprising more than 40,000 GPUs based on Nvidia GB300 NVL72 architectures to support Poolside's model development. | 高 | SI002, SI005 |
| CI005 | CoreWeave terminated its 15-year anchor tenant lease for Project Horizon in April 2026 after Poolside's parallel $2 billion Series C financing round failed to close. | 中 | SI001 |
| CI006 | Poolside's planned $2 billion Series C at a $14 billion pre-money valuation collapsed in spring 2026 after investors cited concerns over sub-$50 million ARR and roughly $400 million in annual compute spend. | 中 | SI001 |
| CI007 | Following the cancellation of the Project Horizon anchor lease, CoreWeave reallocated the 40,000 Nvidia GB300 GPUs originally planned for West Texas to other operational data centers. | 中 | SI001 |
| CI008 | In August 2026, Nvidia paid $6 billion to Poolside in a non-exclusive software licensing agreement for its Model Factory technology and extended offers to 109 staff while Poolside remained independent. | 中 | SI008 |
| CI009 | Alongside the software licensing agreement in August 2026, Nvidia invested $1 billion in primary equity in Poolside at a $12 billion pre-money valuation. | 中 | SI006, SI008 |
| CI010 | Poolside's cumulative primary equity funding reached approximately $1.63 billion across four venture and strategic financing rounds through August 2026. | 中 | SI006, SI008 |
| CI011 | In April 2026, Poolside implemented cash preservation measures to extend its operating runway, including pausing capital commitments on long-lead gas turbines for Project Horizon. | 中 | SI001 |
| CI012 | Poolside delivers software engineering solutions through custom on-premises enterprise models, forward-deployed engineering teams, and its Atlas inference engine. | 高 | SI003, SI007 |
| CI013 | Project Horizon was designed around behind-the-meter aero-derivative gas turbines intended to utilize discounted flared natural gas in the Permian Basin to reduce power costs. | 中 | SI001, SI004 |
| CI014 | Poolside has not published audited financial statements, recognized ARR, software gross margins, or customer retention metrics, leaving key underwriting quantities as private data gaps. | 中 | SI001, SI008 |
| CE001 | Poolside markets a software-engineering AI platform that combines coding models and agent systems for enterprise use. | 中 | SE003, SE005 |
| CE002 | Poolside says its agents work through VS Code, Visual Studio, Zed, IntelliJ, and terminal interfaces and connect to enterprise tools through centrally managed MCP servers. | 中 | SE003 |
| CE003 | Poolside says each agent session runs in a container with secret management and network policy controls while actions are recorded as searchable trajectories. | 中 | SE003 |
| CE004 | Poolside describes deployment on customer bare metal, partner-delivered air-gapped hardware, and existing AWS, Azure, or Google Cloud virtual private clouds. | 中 | SE003, SE008 |
| CE005 | Poolside says its government deployment provides full model weights and can operate without external calls, data egress, or shared systems. | 中 | SE006 |
| CE006 | Poolside claims that its Atlas inference engine improves model performance by up to 30% on customer-controlled infrastructure. | 中 | SE006 |
| CE007 | Poolside says its forward-deployed research engineers work with customer teams to ship an initial automated workflow in weeks and then harden and measure it. | 中 | SE003 |
| CE008 | Artificialus reports a 2026 Poolside model lineup comprising Malibu for complex engineering, Point for rapid completion, and Laguna for open-weight high-security deployment. | 中 | SE001 |
| CE009 | Artificialus reports that Laguna XS.2 and M.1 are text-only models without vision or multimodal inputs. | 中 | SE001 |
| CE010 | On the access date, LM Market Cap cataloged four Poolside models with an average composite score of 40 out of 100. | 中 | SE002 |
| CE011 | On the access date, LM Market Cap showed function calling and reasoning for all four cataloged Poolside models but vision for none of them. | 中 | SE002 |
| CE012 | AI Engineer reports that Poolside offers a macOS Desktop Assistant for running multiple coding agents across projects and repositories. | 中 | SE004 |
| CE013 | AI Engineer and SentinelOne describe Poolside's training approach as reinforcement learning from code execution, using compilation and test results as automated feedback. | 中 | SE004, SE007 |
| CE014 | AI Engineer reports that Poolside stopped building frontier models after an August 2026 Nvidia arrangement and had not disclosed its updated vision. | 中 | SE004 |
| CE015 | SentinelOne reports that Poolside designed its own data-curation pipelines and trains models with execution-based feedback from programming environments. | 中 | SE007 |
| CE016 | AI Market Watch identifies on-premises operation and customer-specific adaptation as Poolside's principal privacy and customization differentiators. | 中 | SE008 |
| CU001 | Poolside targets large enterprise customers with more than 5,000 developers, concentrating on highly regulated verticals including defense, financial services, and retail. | 中 | SU009, SU010 |
| CU002 | Enterprise deployments of the Poolside Platform are delivered exclusively within customer security perimeters, supporting on-premises bare metal, air-gapped hardware, and private VPCs. | 高 | SU003, SU006 |
| CU003 | Poolside provides customer organizations with direct control of model weights and predictable subscription billing rather than variable consumption-based per-token fees. | 高 | SU003, SU006 |
| CU004 | Hunted Labs deploys Poolside models to secure software supply chains internally and deliver secure development tools for national security missions. | 高 | SU003, SU006 |
| CU005 | Federal systems integrators Vibrint and Sterling Computers partner with Poolside Federal LLC to deliver air-gapped coding models for classified defense and intelligence customers. | 中 | SU006 |
| CU006 | RTX Corporation and other defense industrial base suppliers have evaluated or deployed Poolside's customizable coding models within secure defense computing facilities. | 中 | SU009 |
| CU007 | Former GitHub Chief Revenue Officer Paul St. John joined Poolside to lead go-to-market initiatives targeting enterprise accounts with large developer populations. | 高 | SU005, SU010 |
| CU008 | Baseten deployed a dedicated OpenAI-compatible model API infrastructure for Poolside's Laguna model family, achieving sub-second time-to-first-token benchmarks on Triton MoE backends. | 中 | SU004 |
| CU009 | Third-party research estimates Poolside's annualized revenue run-rate reached approximately $50 million as of early 2025, although the company has not publicly disclosed audited financials. | 中 | SU007, SU010 |
| CU010 | Poolside employs Forward Deployed Research Engineers who co-locate with enterprise development teams to identify high-value workflows and operationalize production agents. | 中 | SU003 |
| CU011 | Independent analyst assessments identify extended enterprise procurement cycles and high customer acquisition costs as material risks to Poolside's commercial scaling. | 中 | SU010 |
| CU012 | Poolside faces competition from incumbent cloud developer assistants including GitHub Copilot, which possesses over 20 million users and 90 percent Fortune 100 penetration. | 中 | SU007, SU010 |
| CU013 | Practitioner evaluations highlight that Poolside requires custom enterprise data integration and infrastructure setup rather than functioning as an off-the-shelf consumer developer utility. | 中 | SU001 |
| CU014 | Poolside does not publicly disclose net retention rates, gross renewal figures, or cohort contract durations, limiting visibility into long-term enterprise expansion durability. | 中 | SU009, SU010 |
| CR001 | Poolside does not publish a detailed disclosure of the datasets used to train its Laguna model family, creating copyright and IP provenance compliance risks for enterprise buyers. | 中 | SR001 |
| CR002 | Laguna model releases utilize differing software licenses across generations, ranging from OpenMDW-1.1 for Laguna S 2.1 to Apache 2.0 for Laguna M.1, preventing uniform enterprise legal approvals. | 中 | SR001 |
| CR003 | Poolside has published no formal support lifecycle or commitment detailing how long specific versions of its open-weight Laguna models will remain maintained. | 中 | SR001 |
| CR004 | Poolside maintains a dedicated subsidiary, Poolside Federal LLC, with CAGE code 11R53 and cleared personnel to mitigate security and compliance risks for public sector and defense customers. | 高 | SR003, SR005 |
| CR005 | Poolside deploys models directly inside customer on-premises and air-gapped environments to eliminate external data egress and third-party cloud telemetry risks. | 高 | SR003, SR004 |
| CR006 | Poolside's Reinforcement Learning from Code Execution Feedback architecture executes up to 10,000 containerized code executions per minute across 350 million indexed repositories to train models on verified execution outcomes rather than unvalidated syntax. | 高 | SR002, SR004 |
| CR007 | Poolside entered a multi-year partnership with Amazon Web Services to distribute its models via Amazon Bedrock and execute inference on AWS Trainium chips and EC2 instances. | 中 | SR002 |
| CR008 | CoreWeave and Poolside announced Project Horizon in October 2025 to deploy an AI campus in Texas, but the agreement was mutually terminated in April 2026 resulting in the loss of a planned 40,000 GPU cluster. | 中 | SR005, SR007 |
| CR009 | In August 2026, Nvidia agreed to a $6 billion non-exclusive software licensing deal for Poolside's Model Factory alongside a separate $1 billion equity investment at a $12 billion pre-money valuation. | 中 | SR006, SR007, SR009 |
| CR010 | As part of the August 2026 Nvidia transaction, approximately 109 Poolside research and engineering employees received job offers to transition to Nvidia, creating operational and talent continuity risks. | 中 | SR007, SR009 |
| CR011 | Following the Nvidia Model Factory licensing agreement, Poolside reportedly paused internal development of new frontier foundation models while maintaining independence under its original co-founders. | 中 | SR007, SR009 |
| CR012 | Industry estimates indicate Poolside annualized revenue of approximately $50 million as of 2025, but actual recognized revenue, GAAP gross margins, and operating cash burn remain undisclosed. | 中 | SR008 |
| CR013 | Poolside has raised a total of approximately $1.63 billion in equity capital across four funding rounds, supplemented by $6 billion in non-equity software licensing proceeds. | 中 | SR006, SR007 |
| CR014 | Poolside's enterprise sales motion faces protracted procurement evaluation cycles due to buyer demands for custom security clearances, on-premises testing, and formal compliance certifications. | 中 | SR001, SR003 |
| CV001 | Poolside raised a $500 million Series B funding round in October 2024 at a $3 billion post-money valuation led by Bain Capital Ventures. | 高 | SV005, SV001, SV008 |
| CV002 | Poolside delivers foundation coding models engineered for on-premises and air-gapped deployment within customer security perimeters without external data egress. | 高 | SV003, SV005 |
| CV003 | In August 2026, Nvidia agreed to a non-exclusive $6 billion licensing agreement for Poolside's Model Factory software platform alongside a separate $1 billion equity investment valuing the company at $12 billion pre-money. | 中 | SV008, SV009, SV006 |
| CV004 | The $6 billion proceeds from Nvidia's Model Factory license are scheduled to be distributed directly to Poolside's existing investors by the end of 2027 rather than retained as operational balance sheet capital. | 中 | SV009, SV006 |
| CV005 | Nvidia extended employment offers to approximately 109 Poolside engineers who developed the Laguna model series while co-founders Jason Warner and Eiso Kant remain with the independent company. | 中 | SV008, SV009, SV006 |
| CV006 | Poolside's planned 2-gigawatt Project Horizon Texas data center partnership with CoreWeave was mutually terminated in April 2026, causing a planned $2 billion Series C round to fail and forfeiting access to a 40,000 GB300 GPU cluster. | 中 | SV006, SV009 |
| CV007 | Poolside has not publicly published audited revenue, ARR, gross margins, or customer retention metrics, with secondary estimates placing annual revenue at approximately $50 million as of 2025. | 中 | SV007, SV002 |
| CV008 | At an estimated $50 million run-rate, Poolside's $12 billion pre-money valuation reflects an implied forward revenue multiple exceeding 240x, representing a significant premium over mature public devops and data peers. | 中 | SV002, SV007 |
| CV009 | Poolside's historical equity funding across Seed, Series A, and Series B rounds totaled approximately $626 million prior to the August 2026 Nvidia equity commitment. | 中 | SV001, SV002, SV008 |
| CV010 | Poolside uses Reinforcement Learning from Code Execution Feedback (RLCEF) across indexed repositories to optimize foundation models for autonomous multi-step software development tasks. | 中 | SV005, SV008 |
| CV011 | In a base-case scenario with enterprise revenue scaling toward $100M-$150M, Poolside's fair enterprise valuation normalizes to $10B-$14B under prevailing AI infrastructure multiple compression. | 中 | SV008, SV009 |
| CV012 | A bear-case downside trigger would involve enterprise sales cycle elongation, depletion of investor payouts, and customer churn, compressing valuation to Series B recapitalization marks of $3B-$5B. | 中 | SV006, SV008, SV009 |
| CV013 | The diligence stance recommends a 'research-more' rating with medium confidence, high risk, and a stretched valuation profile pending verification of commercial ARR quality and federal contract pipeline. | 中 | SV006, SV008, SV009 |
| CV014 | Exit readiness requires transitioning federal pilot programs into binding production renewals and demonstrating standalone compute stability independent of sole-source hardware suppliers. | 中 | SV003, SV006, SV009 |
| 编号 | 出版方 | 标题 | 引文 |
|---|---|---|---|
| SO001 | Poolside | The Model Factory series | The Model Factory series A six-part blog series on how we train, evaluate, and iterate on foundation models. From GPU-to-GPU weight transfers and automated architecture ablations to reinforcement learning from code execution at scale. |
| SO002 | Compworth | poolside – Financials, Industry Trends & Funding Info – 2026 | poolside’s business performance suggests annual revenues of around $66.4M. |
| SO003 | Sacra | Poolside valuation, funding & news - Sacra | Poolside closed a $500 million Series B in October 2024 at approximately $3 billion valuation. The round was led by Bain Capital Ventures with participation from Nvidia, eBay Ventures, DST Global, StepStone Group, Citi Ventures, Felicis Ventures, and Redpoint Ventures. |
| SO004 | Crunchbase News | AI-Coding Startup Poolside Raises Massive $500M Series B | Poolside closed a $500 million Series B led by Bain Capital Ventures. The new round values the startup at $3 billion, Bloomberg reported. |
| SO005 | Enera Labs | Nvidia's $7B Poolside Deal: Enterprise AI Impact Decoded | On August 20, 2026, Nvidia disclosed a deal that reframes its long-term AI strategy: a non-exclusive $6 billion license for Poolside AI’s Model Factory software platform, packaged with a $1 billion equity investment that values the Paris-based coding-AI startup at $12 billion before the new capital. |
| SO006 | Poolside Government | Government - Poolside Sovereign AI | Poolside is a U.S.-based AI company delivering frontier open-weight models purpose-built for on-premises and air-gapped deployment—giving government agencies full sovereignty over their AI infrastructure, data, and costs. |
| SO007 | Tracxn | Poolside - 2026 Company Profile, Team, Funding & Competitors | Poolside has raised a total funding of$1.63B over 4 round s.Its first funding round was on May 24, 2023.Its latest funding round was a Series C round on Aug 21, 2026 for $1B.1 investor participated in its latest round.Poolside has 30[institutional investor s](https://platform.tracxn.com/a/d/company/6001b59628a00f1e2cf20fa5#a:funding-and-investors) and 6[angel investor s](https://platform.tracxn.com/a/d/company/6001b59628a00f1e2cf20fa5#a:funding-and-investors). |
| SO008 | SentinelOne S Ventures | S Ventures Invests in poolside, Next Generation AI for Software Engineering | Co-founders Jason Warner (CEO at poolside) and Eiso Kant (CTO at poolside) share decades of industry and company-building expertise between them, having seeded innovative solutions previously for GitHub, Ubuntu and created Athenian, a data-enabled engineering platform and source{d}, the world’s first company applying AI code and software, respectively. |
| SO009 | GetLatka | Poolside AI Revenue 2025: $50M Est., $12B Est. Valuation | Poolside AI has not published its revenue. GetLatka estimates it at about $50M a year (March 2025). |
| SO010 | Adams Street Partners | Why We Invested in poolside – Revolutionizing Software Development | Chief Executive Officer Jason Warner incubated pivotal products such as Copilot and Code Search during his four-year tenure as Chief Technology Officer (CTO) at GitHub. Jason co-founded poolside with CTO Eiso Kant, a front runner in AI-powered code generation tools since launching source{d} in 2015. The recent addition of former GitHub Chief Revenue Officer (CRO) Paul St. John, former GitHub CRO, brings extensive experience in leading go-to-market initiatives across the enterprise landscape. |
| SO011 | AI Market Watch | Poolside AI - AI Startup Profile | AI Market Watch | Total funding$1.626B equity raised (Seed $26M + Series A $100M + Series B $500M + Nvidia Strategic Investment $1B Aug 2026 at $12B pre-money / ~$13B post); prior $2B Series C attempt failed spring 2026. Separately Nvidia paid $6B non-exclusive Model Factory license (Aug 2026), not equity. |
| SO012 | AI Engineer | Poolside: AI Coding Models and Enterprise Agents | In August 2026, Nvidia reportedly agreed to a $6 billion non-exclusive licence for the Model Factory, job offers to 109 employees, and a separate $1 billion investment at a $12 billion pre-money valuation. Poolside remains independent, with its founders staying; the arrangement is not an acquisition. Poolside reportedly stopped building frontier models and had not disclosed its updated vision. |
| SM001 | The AI Agent Index | Poolside Review (2026): Enterprise Foundation Model Lab | Poolside is the foundation model lab building large language models specifically for long-horizon work, starting with software engineering agents, and the enterprise systems to deploy them at scale. Founded in 2023 by Jason Warner (former CTO of GitHub) and Eiso Kant, Poolside has raised $626M+ at a $3B valuation from DST Global, Bain Capital Ventures, and NVIDIA. |
| SM002 | Poolside | Poolside overview - Poolside | Poolside is an AI lab that trains foundation models from scratch for software engineering. Poolside provides agentic coding models, self-managed inference, APIs, and agent workflows for organizations that need to run AI inside environments they control. |
| SM003 | Enera Labs | Nvidia's $7B Poolside Deal: Enterprise AI Impact Decoded | Enera | On August 20, 2026, Nvidia disclosed a deal that reframes its long-term AI strategy: a non-exclusive $6 billion license for Poolside AI’s Model Factory software platform, packaged with a $1 billion equity investment that values the Paris-based coding-AI startup at $12 billion before the new capital. The combined $7 billion commitment, reported by Newcomer from a letter to investors and confirmed by The Information, is among the largest single transactions in AI history, and it carries specific implications for every enterprise team building on AI coding infrastructure. |
| SM004 | NeuronFeed | Poolside Review 2026: AI Coding Foundation Model | NeuronFeed | Unlike consumer-facing coding assistants, Poolside positions itself primarily as an enterprise and infrastructure player, offering models that can be deployed within an organization's own environment. This deployment model targets large companies and regulated industries that have strict requirements around code privacy, data residency, and security. |
| SM005 | Poolside | Mission-grade AI for public sector organizations | Poolside is a U.S.-based AI company delivering frontier open-weight models purpose-built for on-premises and air-gapped deployment—giving government agencies full sovereignty over their AI infrastructure, data, and costs. |
| SM006 | Tracxn | Poolside | Developer of artificial intelligence models and systems for software engineering. It builds models purpose-built to excel at software engineering. It allows models to learn through trial and error, navigating ambiguity to discover solutions. The intelligence understands architecture, development processes, and how software reflects business. It delivers intelligence that integrates across the entire development ecosystem. |
| SM007 | SentinelOne S Ventures | S Ventures Invests in poolside, Next Generation AI for Software Engineering | SentinelOne | The demand for software continues to surpass the supply of skilled developers, and worse, many coding tasks today are repetitive and time-consuming, taking valuable time away from already stretched developers – time that could and should be focused on strategic problem solving and innovation. |
| SM008 | Wikipedia | Poolside AI | The company competes in two sectors with other AI companies, building both large language models and autonomous development products. |
| SP001 | Value Add VC | Nvidia's $7B Poolside Deal: The Reverse Acquihire Playbook | Nvidia announced on August 21, 2026 that it will pay AI coding startup Poolside $6 billion to non-exclusively license "Model Factory" — the internal system Poolside built to train its own frontier coding models — and will absorb 109 of Poolside's roughly 130 employees as new Nvidia hires. |
| SP002 | Poolside | AI, your way: introducing the Poolside Platform | Today we’re announcing the Poolside Platform: a production-grade system for running AI agents inside your boundary, anywhere your business needs it to run. The Poolside Platform is built to help enterprises adopt AI while not compromising on security or governance. |
| SP003 | Crunchbase News | AI-Coding Startup Poolside Raises Massive $500M Series B | In August, San Francisco-based Magic, which also develops AI models to write software, raised a $320 million round from new investors including Eric Schmidt and Sequoia. |
| SP004 | Tracxn | Poolside - 2026 Company Profile, Team, Funding & Competitors | The company has 1183 active competitors, including 159 funded and 49 that have exited |
| SP005 | Poolside | Mission-grade AI for public sector organizations | Air-gapped and on-premises deployment are the foundation for all of our architectural decisions, not capabilities retrofitted later. Every query, every agentic workflow, every response runs entirely within your infrastructure. No external calls. No data egress. No shared systems. |
| SP006 | SentinelOne S Ventures | S Ventures Invests in poolside, Next Generation AI for Software Engineering | SentinelOne | We believe that poolside’s relentless focus on building frontier models in-house and offering features such as on-premise delivery, agentic behaviors, and larger context windows sets them apart from others in the industry. |
| SP007 | AI Market Watch | Poolside AI - AI Startup Profile | AI Market Watch | Main competitors GitHub Copilot Key partnerships AWS (strategic partnership for Amazon Bedrock and EC2, announced Dec 2024); Atos (sovereign AI solutions for Europe, 2026); CoreWeave (2GW AI campus in Texas, Oct 2025) |
| SP008 | VentureBeat | American AI startup Poolside launches free, high-performing open model Laguna XS.2 for local agentic coding | When I asked on X why government agencies would seek to use Poolside instead of leading proprietary U.S. labs like Anthropic, OpenAI and Google, Poolside post-training engineer George Grigorev told me in a reply that: "we think that we can be faster to deploy our models to enterprise customers, and we can literally ship weights in fully isolated environments on-prem, so it can work offline. which might be critical for gov/public sectors :) but ofc anthropic enterprise is hard to beat" |
| SI001 | Tech Insider | Poolside AI's $14B Texas Collapse: CoreWeave Out [2026] | Poolside AI is still trying to rescue its $14 billion bet on a 2-gigawatt Texas data center more than three months after CoreWeave walked away from a 15-year anchor lease and a planned $2 billion funding round, including a rumored $1 billion Nvidia anchor check, fell apart in early April 2026. |
| SI002 | CoreWeave | CoreWeave Announces Partnership with Foundation Model Company Poolside to Deliver AI Cloud Services | Under the agreement, CoreWeave will provide a state-of-the-art cluster of NVIDIA GB300 NVL72 systems, including more than 40,000 GPUs. Separately, CoreWeave plans to provide Poolside with its market-leading cloud solutions for Project Horizon, Poolside’s 2GW AI campus in West Texas. |
| SI003 | Poolside | Mission-grade AI for public sector organizations | Unlike cloud-based AI services, there are no per-token fees: agencies run unlimited inference on their own hardware, making Poolside a predictable, scalable investment. |
| SI004 | Air Street Capital | Compute, power, intelligence: Poolside goes full-stack with Project Horizon | Through a partnership with CoreWeave, which will anchor and operate the first 250MW phase, poolside is combining compute, power, and intelligence under one roof to create a vertically integrated foundation for the agentic enterprise. |
| SI005 | CoreWeave Investor Relations | CoreWeave Announces Partnership with Foundation Model Company Poolside to Deliver AI Cloud Services | CoreWeave plans to serve as the anchor tenant and operational partner for the first phase of the project, which comprises 250MW of gross power and includes an option to further expand capacity by an additional 500MW. |
| SI006 | Tracxn | Poolside - 2026 Company Profile, Team, Funding & Competitors | Poolside has raised a total funding of$1.63B over 4 round s.Its first funding round was on May 24, 2023.Its latest funding round was a Series C round on Aug 21, 2026 for $1B. |
| SI007 | SentinelOne | S Ventures Invests in poolside, Next Generation AI for Software Engineering | In October 2024, SentinelOne’s venture arm, S Ventures, made a strategic investment in poolside as part of the AI assistive coding startup’s $500 million Series B round. |
| SI008 | AI Market Watch | Poolside AI - AI Startup Profile | AI Market Watch | Total funding$1.626B equity raised (Seed $26M + Series A $100M + Series B $500M + Nvidia Strategic Investment $1B Aug 2026 at $12B pre-money / ~$13B post); prior $2B Series C attempt failed spring 2026. Separately Nvidia paid $6B non-exclusive Model Factory license (Aug 2026), not equity. |
| SE001 | Artificialus | Poolside | Poolside is an enterprise and government-grade AI software development platform with a proprietary model family trained on production code. |
| SE002 | LM Market Cap | poolside Models - Rankings, Pricing & Benchmarks (2026) | poolside currently offers 4 AI models on our platform, with an average quality score of 40/100. |
| SE003 | Poolside | AI, your way: introducing the Poolside Platform | Every session runs in a containerized environment with built-in secret management and network policy control. Every agent action, including tool calls, file edits, reasoning steps, and decisions, is recorded as a searchable trajectory. |
| SE004 | AI Engineer | Poolside - AI Coding Models and Enterprise Agents | Poolside offers Laguna, an open-weight family of agentic coding models, and tools for developers and enterprise teams. |
| SE005 | Tracxn | Poolside | Developer of artificial intelligence models and systems for software engineering. It builds models purpose-built to excel at software engineering. |
| SE006 | Poolside | Mission-grade AI for public sector organizations | Every query, every agentic workflow, every response runs entirely within your infrastructure. No external calls. No data egress. No shared systems. |
| SE007 | SentinelOne | S Ventures Invests in poolside, Next Generation AI for Software Engineering | RLCEF works by exposing models to real-world coding challenges found within extensive codebases. |
| SE008 | AI Market Watch | Poolside AI - AI Startup Profile | Ability to run AI models on-premises without requiring data to leave the customer’s environment, enhancing privacy and regulatory compliance |
| SU001 | Quasa | Poolside Review: The Secure AI Platform Built for Real Enterprise Development | It’s perfect for engineering organizations, large enterprises, DevOps teams, and companies building internal AI agents who need powerful coding intelligence without sending sensitive code to third-party clouds... Downsides: As a frontier lab focused on enterprise, it’s not a consumer tool (pricing is custom/enterprise); best results require proper integration and data setup |
| SU002 | Amazon Web Services | How poolside pioneers AI assisted software development on AWS | point offers us the ability to place code completion capabilities in the hands of developers. malibu is our flagship—and largest ever—model, built for the most challenging software engineering problems like code generation, testing, refactoring, and more. |
| SU003 | Poolside | AI, your way: introducing the Poolside Platform | We built the Poolside Platform for a specific kind of team: the kind whose approved base images are a matter of policy; whose network architectures make sure no single node has both GPU access and outbound connectivity; whose CISO has been forced, over the past two years, to relax controls for an AI vendor because there wasn’t a real alternative. |
| SU004 | Baseten | How Baseten powered Poolside's model launch in record time | Poolside has rapidly become one of the most well-funded and closely watched names in enterprise AI by building frontier coding models exclusively for self-hosted deployments... Baseten deployed a comprehensive inference platform for Poolside that combines dedicated model hosting with the Baseten Inference Stack |
| SU005 | Adams Street Partners | Why We Invested in poolside – Revolutionizing Software Development with Generative AI | Unlocking high-fidelity, reliable AI for code generation and engineering workflows is perhaps the largest opportunity within the generative AI landscape today. This is because demand for code far exceeds the supply of coders... The recent addition of former GitHub Chief Revenue Officer (CRO) Paul St. John, former GitHub CRO, brings extensive experience in leading go-to-market initiatives across the enterprise landscape. |
| SU006 | Poolside | Mission-grade AI for public sector organizations | Unlike cloud-based AI services, there are no per-token fees: agencies run unlimited inference on their own hardware, making Poolside a predictable, scalable investment... Partnering with Poolside allows Vibrint to deliver cutting-edge AI capabilities into the most sensitive government environments without compromise. |
| SU007 | Tracxn | Poolside | The company has 1183 active competitors, including 159 funded and 49 that have exited... Developer of artificial intelligence models and systems for software engineering. It builds models purpose-built to excel at software engineering. It allows models to learn through trial and error, navigating ambiguity to discover solutions |
| SU008 | SentinelOne | S Ventures Invests in poolside, Next Generation AI for Software Engineering | SentinelOne | In October 2024, SentinelOne’s venture arm, S Ventures, made a strategic investment in poolside as part of the AI assistive coding startup’s $500 million Series B round... poolside Enterprise where users can dictate what data sources (repositories and knowledge bases, for example) the model should learn from. |
| SU009 | AI Market Watch | Poolside AI | Target customer Large enterprises, particularly in highly regulated industries such as finance, defense, and retail... Notable customers RTX Corporation and other companies in the American defense industrial base; Israeli defense companies (in discussions) |
| SU010 | Sacra | Poolside | The go-to-market approach focuses on enterprises with 5,000+ developers, including major banks and defense contractors... Enterprise sales cycles: Poolside’s focus on large enterprise customers with complex security and compliance requirements creates long sales cycles and high customer acquisition costs. Delays in enterprise AI adoption or budget cuts could significantly impact revenue growth and cash burn rates. |
| SR001 | Layer3Labs | Poolside AI and the Laguna Models Explained (2026) | Poolside does not publish a detailed disclosure of what the models were trained on, which some legal reviews will treat as a blocker for code generation. |
| SR002 | Amazon Web Services | poolside and AWS announce strategic agreement to enable secure, customized generative AI for software engineering on Amazon Bedrock and Amazon Elastic Cloud Compute (EC2) | poolside will also leverage AWS Trainium chips to power inference for its malibu and point foundation models, enabling optimal price performance for customers. |
| SR003 | Poolside | Mission-grade AI for public sector organizations | Poolside Federal LLC is a US-domiciled entity dedicated exclusively to US public sector missions. |
| SR004 | SentinelOne | S Ventures Invests in poolside, Next Generation AI for Software Engineering | SentinelOne | SentinelOne congratulates poolside on their recent funding round, allowing them to scale their training clusters to 10,000 GPUs. |
| SR005 | Silicon Valley Investclub | Poolside AI Company Profile | CoreWeave announced a partnership with Poolside to deliver advanced AI cloud services and provide a cluster of NVIDIA GB300 NVL72 systems and cloud solutions for Project Horizon. |
| SR006 | Tracxn | Poolside Company Profile | Poolside has raised a total funding of$1.63B over 4 round s.Its first funding round was on May 24, 2023.Its latest funding round was a Series C round on Aug 21, 2026 for $1B. |
| SR007 | AI Market Watch | Poolside AI - AI Startup Profile | AI Market Watch | CoreWeave Project Horizon announced Oct 2025 then mutually terminated Apr 2026, $2B Series C attempt failed spring 2026 (lost 40k GB300 cluster) |
| SR008 | GetLatka | Poolside AI Revenue 2025: $50M Est., $12B Est. Valuation | Poolside AI has not published its revenue. GetLatka estimates it at about $50M a year (March 2025). |
| SR009 | AI Engineer | Poolside: AI Coding Models and Enterprise Agents | AI Engineer | In August 2026, Nvidia reportedly agreed to a $6 billion non-exclusive licence for the Model Factory, job offers to 109 employees, and a separate $1 billion investment at a $12 billion pre-money valuation. Poolside remains independent, with its founders staying; the arrangement is not an acquisition. Poolside reportedly stopped building frontier models and had not disclosed its updated vision. |
| SV001 | Wikipedia | Poolside AI | announced that it would invest up to $1 billion in Poolside which would quadruple its valuation to $12 billion. |
| SV002 | Silicon Valley Investclub | Poolside AI — Funding, Valuation & Leadership | Estimated dilution per round, computed from disclosed round sizes and post-money valuations. The full shareholder register is not public. |
| SV003 | Poolside | Bring frontier AI to your mission. | Poolside is a U.S.-based AI company delivering frontier open-weight models purpose-built for on-premises and air-gapped deployment—giving government agencies full sovereignty over their AI infrastructure, data, and costs. |
| SV004 | Tracxn | Poolside - 2026 Company Profile, Team, Funding & Competitors | Developer of artificial intelligence models and systems for software engineering. It builds models purpose-built to excel at software engineering. |
| SV005 | SentinelOne | S Ventures Invests in poolside, Next Generation AI for Software Engineering | In October 2024, SentinelOne’s venture arm, S Ventures, made a strategic investment in poolside as part of the AI assistive coding startup’s $500 million Series B round. |
| SV006 | AI Market Watch | Poolside AI - AI Startup Profile | CoreWeave Project Horizon announced Oct 2025 then mutually terminated Apr 2026, $2B Series C attempt failed spring 2026 (lost 40k GB300 cluster) |
| SV007 | GetLatka | Poolside AI Revenue 2025: $50M Est., $12B Est. Valuation | Poolside AI has not published its revenue. GetLatka estimates it at about $50M a year (March 2025). |
| SV008 | Sacra | Poolside valuation, funding & news | Poolside closed a $500 million Series B in October 2024 at approximately $3 billion valuation. The round was led by Bain Capital Ventures with participation from Nvidia, eBay Ventures, DST Global, StepStone Group, Citi Ventures, Felicis Ventures, and Redpoint Ventures. |
| SV009 | Enera Labs | Nvidia's $7B Poolside Deal: Enterprise AI Impact Decoded | The $6 billion license is non-exclusive and covers Model Factory. Poolside retains the right to license it to others. The $6 billion is expected to be distributed to Poolside’s existing investors by the end of 2027, not to the company’s operating accounts. |