Emerald AI
面向用电可调数据中心与电网感知算力的 AI 控制层
Emerald AI 确实用 AI 电力灵活性切入了产品和市场痛点,但 2026 年 8 月 $1.05B Series A 已把大量未来收入转化计入估值,公开证据还没有证明。
封面要素
公司概况
Emerald AI 是一家总部位于华盛顿特区的软件公司,正在打造 Emerald Conductor 这个控制层,帮助 AI 数据中心在不牺牲关键工作负载的前提下响应电网和电力约束。公司销售对象位于超大规模云厂商、数据中心运营商、电力公用事业公司和电网机构的交集,把灵活性定位为提升拿电速度、释放容量的解决方案,而不是单纯的节能工具。公司 2024 年才成立,公开证据已显示异常强的早期技术验证和生态背书,但财务披露仍不足以让外部投资人充分论证当前 $1.05 billion 估值。
- 成立时间
- 2024-11-01
- 创始人
- Varun Sivaram
- 创立地点
- Washington, DC
- 总部
- Washington, DC
- 产品
- 编排 AI 工作负载和现场能源资源的软件,让数据中心能作为灵活电网资产运行。
- 客户
- 超大规模云厂商、AI 基础设施运营商、数据中心业主 / 运营商、电力公用事业公司和电网机构。
- 商业模式
- Emerald 销售软件控制层和相关部署工作流,围绕更快并网、负荷灵活性和感知电网的电力调度变现。
- 阶段
- Series A
- 融资情况
- 2026 年 8 月宣布 $150M Series A,估值 $1.05B;此前已披露的种子轮、延展融资和战略扩张融资合计约 $68M。
执行摘要
主要优势
- Emerald 切的是 AI 基础设施的真瓶颈:电力可用性和灵活负载管理。
- 以该阶段看,公开验证格外具体:Phoenix、英国、Santa Clara 和 Aurora 的证据都绑定了具名交易方和实测结果。
- 投资人和合作伙伴组合具备战略价值,涵盖公用事业公司、AI 生态参与者和工业现有企业,能帮公司打开市场入口。
主要风险
- $1.05B 估值默认收入规模和耐久性会远高于当前公开披露。
- 客户和生态仍集中在少数旗舰伙伴、公用事业公司和站点,未来收入质量可能被压缩。
- 如果公用事业公司和客户分不到足够价值,或电价结构缺乏吸引力,灵活负载经济性可能继续偏薄。
- 公开披露仍缺合同金额、利润率、现金跑道、留存和优先股堆叠细节,承销信心受限。
未决问题
- 收入、ARR、ACV、毛利率和客户续约数据未公开披露。
- 股权稀释、清算优先权和二级市场背景仍不够透明,无法充分承销入场回报。
- 按站点、合作伙伴和公用事业辖区划分的集中度仍不清楚,尽管旗舰验证很强。
- 公开证据还没有证明,早期试点和旗舰部署能复利成可重复的多站点商业项目。
目录
01公司概况
1.1 身份、总部与产品逻辑
Emerald AI 是一家总部位于华盛顿特区的气候与能源软件公司,成立于 2024 年末,瞄准 AI 基础设施里越来越具体的瓶颈:电力。公司不假设每个 AI 数据中心都必须像僵硬的 24/7 峰值负荷那样运行,而是主张现代 GPU 工作负载可以被编排,在不违反最关键服务级别要求的前提下,短暂降速、暂停或转移。Emerald 将这一逻辑封装进 Emerald Conductor:一个位于电力公用事业公司或电网运营商与数据中心运营团队之间的控制层。 公司材料始终把产品界定为软件,而不是发电设备。Conductor 摄取电网信号、工作负载优先级和本地能源状态,再调节设施用电需求或现场资源,让数据中心从被动负担变成可控资产。这一定位很关键,因为 Emerald 同时对超大规模云厂商、托管运营商和电力公用事业公司有意义。产品价值主张也因此直接绑定拿电时间:如果 Conductor 能让运营商更早接入、拿到更大的并网额度,或避开昂贵的电网升级,即便直接能源市场收入尚未验证,产品也能贴近关键任务预算。[CO001, CO002, CO007, CO008, CO031, CO034]
| 指标 | 数值 / 状态 | 日期 / 时点 | 置信度 | 缺口 / 注意事项 |
|---|---|---|---|---|
| 总部 | Washington, DC | 2026 | 高 | 官方联系页和投资者材料支持 |
| 其他办公室 | Boston, MA;San Francisco, CA 办公室 | 2026 | 高 | 官方联系页列出 |
| 成立时间 | November 2024 | 2024 | 中 | 具体日期未公开披露 |
| 阶段 | Series A 轮 / 早期独角兽 | Aug 2026 | 高 | 公司公告和 SEC 文件支持 |
| 最新一轮融资 | $150M Series A 轮 | 2026-08-25 | 高 | 公司宣布该轮超额认购 |
| 估值 | $1.05B | 2026-08-25 | 高 | 公司公布的估值 |
| 已披露累计融资 | 已宣布 ~$218M | 2026 | 中 | 根据已宣布轮次推算;私人股权结构细节未披露 |
| 创始人 | Varun Sivaram | 2026 | 高 | 公开材料只突出这一位创始人 |
| 已披露董事席位 | John Tough (Energize Capital) | 2026 | 中 | 观察员角色公开;控制权未披露 |
| 公开进展信号 | 5 次现场演示已完成 | 2026 | 中 | 公司称;并非所有地点都有独立文件记录 |
| 商业验证点 | SVP 试点;Aurora Virginia 设施 | 2026 | 高 | 电力公司和 S&P 来源支持 |
| 认可 | TIME100 + WEF Technology Pioneer | 2026 | 高 | 第三方认可 |
| 客户数 | 未披露 | 2026 | 高 | 公司披露客户类别,但不披露数量 |
| 员工数 | 未披露 | 2026 | 高 | 公开记录中没有经核实的员工数 |
累计融资根据公开宣布的 $24.5M 种子轮、$18M 种子延伸轮、$25M 战略扩张轮和 $150M Series A 轮加总推算; SEC 对已售出金额的披露与已宣布轮次规模不同。
[CO001, CO002, CO003, CO004, CO009, CO012]Emerald 如何围绕更快、可灵活调度的并网逻辑,把电力公司、AI 运营商和基础设施伙伴连起来。
[CO007, CO008, CO031, CO032, CO034, CO035]截至 2026 年 8 月,公开材料可见的资本、阶段、牵引力和证明点指标。
累计融资采用四舍五入,因为 Form D 已售金额和公司公布的轮次规模不是同一口径的融资额。
[CO009, CO012, CO018, CO027, CO029, CO030]1.2 创始人-市场匹配、管理梯队与治理
Emerald 的创始人-市场匹配,在基础设施控制创业公司里异常强。Varun Sivaram 有来自 Orsted 和 ReNew 的电力行业运营经验、美国外交工作的政策公信力,以及围绕能源系统约束的公共思想影响力。这些背景让 Emerald 面对监管机构和电力公用事业公司时具备可信度,而纯 AI 应用背景创始人很可能缺少这一点。早期管理梯队进一步加深了这种匹配:Ayse Coskun 是电网响应柔性计算领域最知名的学者之一,Shayan Sengupta 带来 AWS 和 Intel 的超大规模工程落地经验,Aroon Vijaykar 和 Mansi Shah 则补上根植于能源和企业基础设施的商业与产品领导力。 治理结构同样值得注意,因为它直接映射了 Emerald 的 GTM 策略。董事会和观察员名单包括 Energize、Radical、DCVC、NVentures、Lowercarbon 以及 Energy Impact Partners/Frontier Fund。顾问横跨公用事业、政策、AI 和能源市场圈层,从 Salt River Project 的 David Rousseau 到 Jason Bordoff 和 Gina Raimondo。上述网络是战略优势,因为 Emerald 必须跨行业销售。但它也是集中度风险,因为公开材料仍未披露投票控制权、保护性条款,或这一异常战略化股权结构的具体经济权利。[CO005, CO006, CO020, CO021, CO022, CO023]
| 人员 | 职务 | 过往背景 | 重要性 | 关键人物 / 治理风险 |
|---|---|---|---|---|
| Varun Sivaram | 创始人兼 CEO | 前 Orsted 首席战略与创新官;前 ReNew Power CTO;前美国国务院清洁能源官员 | 串联 AI 基础设施、电力公司政策和电力市场战略 | 高 — 创始人对融资、政策可信度和产品定位至关重要 |
| Ayse Coskun | 首席科学家 | Boston University 教授;弹性计算和 HPC 研究员 | 为电网感知计算提供技术可信度和研究领导力 | 中 — 深厚领域经验很难替代 |
| Shayan Sengupta | 工程负责人 | 前 AWS 和 Intel AI/HPC/云平台工程负责人 | 为企业级部署带来超大规模落地能力 | 中高 — 电力公司和数据中心规模下的可靠性关键 |
| Aroon Vijaykar | 首席商务官 | 前 Sunrun VPP、分销和制造业务负责人;前 AEE Solar CEO | 为 GTM 节奏补上电力公司和能源商业化经验 | 中 — 渠道和买方开发离不开商业负责人 |
| Mansi Shah | 产品负责人 | 前 VMware 企业数据和分布式系统首席技术专家 | 帮助把技术弹性落成可用的企业产品路线图 | 中 |
本表仅涵盖 Emerald 团队页面公开介绍的核心领导;公开材料未披露完整高管团队、薪酬或继任计划。
[CO005, CO006, CO020, CO022, CO023, CO024]| 利益相关方 | 类型 | 公开角色 | Emerald 战略价值 | 尽调问题 |
|---|---|---|---|---|
| Energize Capital | 领投 VC / 董事会 | Series A 联合领投;John Tough 被列为董事 | 能源转型可信度和电力公司网络 | 核实持股、董事会权利和后续跟投储备 |
| DCVC | 领投 VC / 观察员 | Series A 联合领投;Zachary Bogue 被列为董事会观察员 | 深科技投资判断和工业商业化支持 | 厘清经济条款和信息权 |
| NVentures / NVIDIA | 战略投资者 / 观察员 | 投资方和技术合作伙伴;Christina Buchanan 被列为观察员 | 把 Emerald 接入主导 GPU 生态和参考架构 | 评估对 NVIDIA 技术栈的依赖以及是否存在排他性 |
| Energy Impact Partners / Frontier Fund(投资方) | 战略财务投资者 / 观察员 | Shayle Kann 担任观察员;投资者网络有电力公司背书 | 带来电力公司入口和战略顾问覆盖 | 核实商业引荐与治理权利的边界 |
| Salesforce Ventures | 战略投资者 | 投资方,并公开背书创始人—市场匹配 | 企业软件可信度和 GTM 信号 | 评估是否存在产品或数据集成预期 |
| National Grid | 战略投资者 / 客户 | 战略投资者和英国演示项目对手方 | 在受监管市场验证电力公司买方逻辑 | 核查商业合同范围和经济性 |
| Silicon Valley Power | 客户 / 试点电力公司 | Santa Clara 官方试点伙伴 | 证明电力公司愿意因灵活性提供更大的电网接入额度 | 核实规模、期限以及从试点转成项目的路径 |
| Digital Realty / PJM / EPRI | 部署合作伙伴 | Virginia 的 Aurora AI Factory 生态 | 商业规模参考站点和标准影响力 | 厘清哪一方是付费客户,以及哪些成功指标决定扩张 |
公开来源披露了参与方,但未披露董事会投票权、清算优先权、按比例跟投条款或老股交易。
[CO010, CO011, CO012, CO019, CO020, CO021]1.3 融资历程、投资人基础与里程碑
Emerald 的融资节奏异常快。公司 2025 年 7 月凭 $24.5 million 种子轮公开亮相,随后通过 SEC Form D 文件和公司帖文披露了若干过渡融资,又在 2026 年 8 月 Series A 前完成 $25 million 战略扩张轮和 $18 million 种子轮延展。截至本报告日,公开披露指向累计宣布融资约 $218 million。融资路径体现一套有意设计的策略:先用早期试点证明技术可信度,再用战略投资人包围公司,等电力公用事业公司和数据中心买方开始把灵活性视为并网解决方案,而不是科研项目时,再募集更大一轮资金。 里程碑路径与这套资本策略相匹配。Emerald 的公开记录从 2024 年 11 月创立,推进到 2025 年 5 月 Phoenix 演示,再到 2025 年末英国和 Virginia 旗舰项目公告,最后到 2026 年 California 电力公用事业试点和独角兽 Series A。投资人组合本身也是商业信号。除财务 VC 外,本轮还包括芯片、公用事业、工业和能源公司,它们可能成为设计伙伴、客户或渠道关系。公司仍很年轻且尚未披露收入,生态强度因此成为公开可见的最清晰去风险证据之一。[CO003, CO009, CO010, CO011, CO012, CO013]
| 日期 | 事件 | 类型 | 金额 / 状态 | 参与方 | 含义 |
|---|---|---|---|---|---|
| 2024-11 | Emerald AI 成立 | 成立 | 公司设立 | Varun Sivaram | 开启「电力可调」的 AI 基础设施命题 |
| 2025-07 | 公开发布并完成种子轮 | 融资 | $24.5M 种子轮 | Radical Ventures、NVentures、Amplo、CRV、Neotribe 等投资方 | 支撑初始试点和公司发布 |
| 2025-08 | Form D 显示种子轮阶段发行规模更大 | 融资 | $35.3M 拟发行 / $34.17M 已售出 | Emerald AI, Inc. | 说明在广泛商业验证前已完成早期资本募集 |
| 2025-10 | 宣布 Virginia 的 Aurora AI Factory | 合作 | 计划中的 $96MW 参考设施 | Emerald AI、NVIDIA、Digital Realty、EPRI、PJM 等参与方 | 打造标杆商业规模参考设计 |
| 2026-02 | 提交额外 Form D | 融资 | $25.0M 拟发行 / $22.75M 已售出 | Emerald AI, Inc. | 扩张前过桥资本 |
| 2026-04 | 宣布 SVP 灵活负载试点 | 合作 | 商业级多 MW 试点 | Silicon Valley Power、NVIDIA、Emerald AI 等参与方 | 从演示推进到接入电力公司的部署 |
| 2026-08 | 战略生态获得认可 | 治理 | 据称有 12 家 Fortune Global 500 投资方 | 战略顾问委员会 | 显示 AI 和能源技术栈都有生态拉力 |
| 2026-08-25 | 宣布 Series A 轮 | 融资 | $150M,估值 $1.05B | Energize Capital、DCVC、大型战略财团 | 确立独角兽估值,并为全球商业推广提供资金 |
中间还有若干重要里程碑,包括种子延伸轮和英国试验;这里省略,是为了把本表保持为一条聚焦成立、融资、标杆合作和阶段变化的单一时间线。
[CO003, CO009, CO013, CO014, CO015, CO016]从 2024 年创立到 2026 年 8 月 Series A 的融资、试点和旗舰部署里程碑。
公开来源只披露月份、不披露具体日历日时,时间轴按月份标注。
[CO003, CO009, CO010, CO015, CO031, CO032]1.4 牵引信号与仍然关键的缺口
最强的外部证据在于,Emerald 已经越过 PPT 阶段,进入与一线对手方的现场演示。公开来源描述了 Phoenix 工作负载削减测试、London 电网响应试验、Santa Clara 电力公用事业试点,以及 Virginia 的 Aurora 参考设施。上述项目既显示技术认真度,也显示生态买账。它们还表明,产品正由真正影响市场形成的参与者塑形:NVIDIA、Digital Realty、EPRI、PJM、National Grid 和 Silicon Valley Power。 不过,公司概况尽调不能停在 logo。Heatmap 和 S&P Global 都点出核心商业化风险:电力公用事业公司和超大规模云厂商必须接受一种新的运营模型,让部分 AI 工作负载变成灵活性资源。这需要激励结构、运营手册和信任,而这些尚未在广泛生产规模上经受检验。公开材料也在收入、利润率、客户集中度、续约行为和控制权方面留下重大承销缺口。Emerald 已在叙事上迈入独角兽门槛,但公司仍未披露足以让外部投资人高置信度论证这一估值的运营证据。[CO026, CO029, CO030, CO031, CO032, CO034]
1.5 图表
02市场分析
2.1 市场边界与定义
Emerald AI 并不争夺所有 AI 基础设施相关支出。相关市场是一个狭窄的控制层:它帮助电力受限的数据中心更快拿到并网接入,参与公用事业或电网灵活性项目,并验证被削减或转移的工作负载仍满足服务约束。市场边界很重要,因为 AI 基础设施的大部分资金落在土地、建筑壳体、变电站、发电、网络和 GPU 上;只有当 Emerald 的软件让容量更早可用,或帮助避免额外电力系统成本时,公司才间接触达这些预算。 因此,纳入边界的支出包括编排软件、遥测、合规或验证工具、与公用事业公司或系统运营商的集成工作,以及与可用性、削减或调度挂钩的经常性软件费或绩效费。排除在外的支出包括通用托管租金、电力硬件、商用发电资产,以及一次性建设资本开支,除非 Emerald 能通过灵活性工作流捕获经济收益。现状替代路径仍是等待确定性并网、建设现场电力、使用没有专业软件的定制双边安排,或把部署转移到电力条件更好的地区。[CM028, CM029, CM030, CM031]
| 细分 / 类别 | 纳入支出 | 排除支出 | 买方 / 付款方 | 关联度 |
|---|---|---|---|---|
| 灵活并网编排 | 调度软件、遥测、限电逻辑、电力公司集成 | 输电建设、变电站资本开支、通用咨询 | 数据中心运营商 / 与电力公司分摊 | Emerald 核心切入点 |
| 电网项目参与层 | 验证、调度接口、报告、绩效结算支持 | 批发市场清算系统本身 | 电力公司、供电实体、运营商 | 把灵活性转成可变现的电网服务 |
| 表后电力协调 | 协调现场发电或储能与电网条件的软件 | 实体发电机、电池、燃料供应 | 数据中心运营商 | 混合供给驱动快速取电时适用 |
| 面向电力事件的 AI 工作负载编排 | 模型调度、策略控制、工作负载迁移接口 | GPU、基础 MLOps 技术栈、通用可观测性 | 超大规模云厂商 / 新云厂商 / 托管数据中心运营商 | 支撑 Emerald 承诺的技术用户工作流 |
| 排除的基础设施栈 | N/A | 土地、建筑壳体、冷却、芯片、变电站、商用发电、标准托管租金 | 基础设施开发商 | 相邻支出很大,但不是直接 TAM |
边界刻意收窄:Emerald 参与的是软件改变取电时间或运营灵活性的环节,而不是买方单纯花钱做通用数据中心建设的环节。
[CM028, CM029, CM030, CM031]2.2 规模测算视角与受边界约束的机会
公开来源有力证明宏观问题巨大且正在加速,但不能直接推出干净的软件总可用市场(TAM)。IEA 预计美国电力需求到 2030 年将年增近 2%,其中约一半增量来自数据中心;Berkeley Lab 则认为美国数据中心用电需求到 2028 年可能达到 325-580 TWh。JLL 又补上一条供给侧视角:2026 至 2030 年全球新增容量接近 97 GW,并由 14% 行业 CAGR 支撑。Bloom、CBRE 和 JLL 都把电力获取,而不是低价租金或连接性,描述为真正卡点。 因此,用 GW 测算比用软件金额测算更站得住。公开证据中置信度最高的空间逻辑来自 Duke/CFR 的表述:如果设施接受有限削减,约 100 GW 的美国数据中心需求可以更早接入。实际可服务市场(SAM)更小,因为商业化取决于电价机制、公用事业公司意愿,以及关键工作负载能扛住灵活性事件的证据。本章因此保留三条口径:宏观电力需求、数据中心供给增长,以及受约束的 25/50/100 GW 灵活并网区间,而不是假装公开证据已经披露 Emerald AI 实际可定价的软件收入池。[CM001, CM003, CM005, CM006, CM008, CM021]
| 发布方 | 年份 | 地域 | 数值 | CAGR | 方法 | 置信度 | 限制 |
|---|---|---|---|---|---|---|---|
| IEA | 2026 | 美国 | 截至 2030 年,用电需求年增速 ~2%;增量增长 ~50% 来自数据中心 | N/A | 宏观用电需求预测 | 高 | 不是软件市场估算 |
| Berkeley Lab | 2025 | 美国 | 到 2028 年美国数据中心用电需求 325-580 TWh | N/A | 自下而上用电需求情景 | 高 | 能耗,不是支出 |
| JLL | 2026 | 全球 | 到 2030 年新增 97 GW 数据中心容量 | 14% 供给 CAGR | 按地区和细分的行业供给预测 | 高 | 基础设施容量,不是 Emerald 收入 |
| Bloom Energy | 2026 | 美国 | 美国 IT 负载从 2025 年 ~80 GW 增至 2028 年 ~150 GW | N/A | 调研支持的行业综合 | 中 | 使用 IT 负载框架,而非已签约电力公司负载 |
| Duke/CFR 视角 | 2025 | 美国 | 有限限电下,近期 ~100 GW 余量 | N/A | 灵活并网思想实验 | 高 | 商业化假设未厘清 |
| Emerald 约束 SAM | 2026 | 北美 + 英国 | 25-100 GW 灵活并网机会视角 | N/A | 保留政策和验证不确定性的分析师区间 | 低 | 推导值,非发布方给出 |
本表保留彼此不兼容但对决策有用的视角,而不是把它们压成一个虚假的单一 TAM。最好的公开证据以 GW 或 TWh 呈现, 不是软件收入金额。
[CM001, CM003, CM005, CM011, CM021, CM046]边界约束的市场金字塔,从广义基础设施增长逐层收束到 Emerald AI 的近期商业切入口。
只有前三层来自发布方。底层是受约束的分析师视角,用来把公司实际可触达切口压窄到低于整体基础设施支出。
[CM005, CM003, CM021, CM046, CM048, CM049]单一指标的低 / 基准 / 高区间:以 GW 衡量的美国近期灵活并网机会。
前两行是分析师对 Duke/CFR 空间论点的转换,折扣来自 Heatmap、CBRE 和当前费率碎片化所暗示的商业化阻力;只有第三行是直接公开上限。
[CM021, CM048, CM049, CM050, CM041]2.3 买方、用户与付款方分层
Emerald AI 的早期市场是多边市场。直接运营用户通常是数据中心的能源、运营或基础设施团队,他们必须守住可用性,同时暴露一部分可调度灵活性。直接商业买方往往也是同一批团队,尤其在超大规模云厂商、新云厂商或大型托管开发商那里,拿电速度已变成生死问题。但当灵活性被嵌入电价、试点或并网协议时,经济赞助方也可能是电力公用事业公司、公共电力供应商或系统运营商。 换言之,Emerald 卖的不是普通业务线 SaaS 工具。它卖的是一个位于公用事业规划、并网、数据中心运营和 AI 工作负载调度交叉点的工作流。采用触发因素包括电网接入延迟、惩罚性成本分摊、非确定性服务机会,以及愿意配合的电力公用事业合作方。下面的分层图对估值很重要,因为买方预算负责人更接近能源战略和基础设施规划,而不是普通 IT 采购;采购周期会被拉长,但路径一旦跑通,战略价值也会提高。[CM032, CM033, CM034, CM035, CM046, CM047]
| 细分 | 买方 | 用户 | 付款方 | 工作流 | 预算负责人 | 采用触发因素 |
|---|---|---|---|---|---|---|
| 超大规模云厂商 AI 园区 | 基础设施 / 能源战略团队 | 站点运营 + 工作负载调度团队 | 超大规模云厂商 | 并网谈判 -> 试点 -> 运营政策 | 能源战略 / 基础设施资本开支负责人 | 供电延迟数月甚至数年 |
| Neocloud 或 AI 原生集群运营商 | 创始人 / 运营领导层 | 运营团队 | 运营商或融资 SPV | 公用事业协议 -> 软件部署 -> 验证事件 | COO / 基础设施负责人 | 需要快速锁定稀缺电网接入 |
| 托管数据中心开发商 / REIT | 开发 + 电力采购团队 | 设施运营 | 可向客户转嫁成本的开发商 | 园区设计 -> 公用事业沟通 -> 租户承诺 | 电力采购 / 开发负责人 | 大规模预租需要可信电力方案 |
| 公用事业或公共电力供应商 | 大负荷规划 / 创新团队 | 电网运营人员和客户经理 | 公用事业或电价机制 | 电价 / 试点设计 -> 客户加入 -> 调度 | 规划 / 监管 / 商务负责人 | 需要在不损害可靠性的前提下增加负荷 |
| RTO/ISO 或政策主导项目 | 间接发起方,而非典型软件买方 | 公用事业 + 客户参与方 | 项目专属成本分摊 | 市场规则 -> 电价 -> 本地落地 | 监管和市场设计团队 | 可靠性驱动的大负荷改革 |
公开合同不可得,预算归属只能定性判断。反复出现的模式是:经济买方更靠近电力规划,而不是中央 IT 采购。
[CM032, CM033, CM034, CM035, CM046, CM047]用序数刻画早期 Emerald AI 客群中谁买单、谁使用,以及采用阻力最高的位置。
单元格是有证据支撑的序数判断,不是问卷分数。它们总结了 JLL、CBRE、SEPA、DCK 和 Emerald 试点披露中的定性买方逻辑。
[CM032, CM033, CM034, CM035, CM046, CM047]从受约束负荷请求到经常性灵活性项目的商业路径。
由于没有公开漏斗数据,数值使用指数而非真实转化率。图形只是展示当下商业瓶颈所在。
[CM035, CM037, CM038, CM039, CM040, CM051]2.4 增长驱动、约束与矛盾
三股力量让这个市场在 2026 年踩中时点。第一,电力稀缺已经成为 AI 基础设施增长的一阶约束;JLL、CBRE、Bloom 和 IEA 在这一点上高度一致。第二,监管机构和电力公用事业公司正通过大负荷电价、灵活服务类别和明确考虑灵活需求的说明理由程序,主动搭建商业路径。第三,来自 Phoenix、英国以及更广泛 DCFlex 生态的现场证据表明,至少一部分 AI 工作负载可以大幅灵活调节,同时不关闭关键服务义务。 矛盾同样重要。Heatmap 捕捉到反向商业逻辑:只有电力公用事业公司把灵活性转化为更快并网或有意义的经济收益时,灵活性才有价值。运营商仍然保守,因为数据中心过去承诺近乎完美的可用性,而公开变现证据仍未揭示经常性合同结构或实际定价。结果是,一个市场战略重要性显而易见、技术可行性也可信,但软件收入捕获仍受当地监管、公用事业激励和少数旗舰证据调节,而不是由成熟品类预算直接支撑。[CM036, CM037, CM038, CM039, CM040, CM041]
| 驱动因素 / 约束 | 方向 | 时间 | 影响 | 尽调追问 |
|---|---|---|---|---|
| 核心枢纽电力稀缺 | 驱动因素 | 当前 | 让快速获得供电变得更紧迫 | 量化采用 Emerald 相比等待确定性供电的胜出案例 |
| 大负荷电价和弹性服务试验 | 驱动因素 | 2026 年以后 | 为弹性变现开出正式通道 | 梳理哪些公用事业今天已有真实经济让利 |
| 可控 AI 工作负载的现场验证 | 驱动因素 | 当前但仍早期 | 降低买方怀疑,支撑试点 | 复核事件级表现和 SLA 结果 |
| 转向现场供电或混合供电 | 混合 | 当前 | 既可能补强 Emerald 编排,也可能削弱纯电网弹性方案的需求 | 确认 Emerald 是否进入混合控制栈 |
| 运营商对可用性的保守态度 | 约束 | 持续 | 拖慢 AI 原生或公用事业支持试点之外的采用 | 测试客户对限电窗口和罚则的容忍度 |
| 州与公用事业落地碎片化 | 约束 | 持续 | 拉长销售周期,并迫使 GTM 本地化 | 按地区绘制活跃公用事业路径地图 |
| 经常性定价和价值捕获不清 | 约束 | 当前 | 软件收入 TAM 难以自圆其说 | 索取定价、合同基础和公用事业成本分摊数据 |
市场吸引力来自问题足够急,但商业化仍取决于本地项目设计,以及买方是否愿意用完全确定的供电保障换速度或经济性。
[CM036, CM037, CM038, CM039, CM040, CM041]2.5 图表
03竞争格局
3.1 竞争格局与品类地图
Emerald AI 竞争的市场仍在从相邻品类中拼出来,还不是同业边界清晰的稳定软件赛道。直接要完成的任务不是通用需求响应,而是让 AI 数据中心负荷足够灵活,从而释放更快电力接入、响应电网状态,并守住关键工作负载。因此,在本资料集中,Emerald 是最清晰的直接专业厂商。大多数其他供应商则从三类相邻位置切入:Voltus 和 CPower 等工商业需求响应聚合商;Virtual Peaker、EnergyHub、Uplight 和 Itron 等面向公用事业公司的灵活性平台;以及现场电力、迁往电力更有优势的地区、业主或公用事业嵌入式方案等替代路径。 原因在于,买方可能解决同一个电力受限问题,却从不发起 Emerald 与 Emerald 克隆厂商的直接采购对比。很多时候,真正的选择是在专业编排、既有能源平台能力、定制公用事业合同和资本密集型替代策略之间取舍。因此,竞争对手表把直接、相邻、既有厂商和替代类别分开,而不是假装每家供应商都是正面相撞的软件同业。[CP001, CP036, CP037, CP038, CP039, CP040]
| 竞争对手 | 类别 | 规模 / 融资 | 目标客群 | 差异化 | 局限 |
|---|---|---|---|---|---|
| Emerald AI | 直接专项厂商 | 私营;2026 年 8 月以 $1.05B 估值完成 $150M Series A 轮 | AI 数据中心、公用事业、电网运营商 | 面向电力受限数据中心的 AI 工作负载弹性 | 公开验证很早期,披露的商业规模有限 |
| Voltus | 工商业需求响应 / VPP 聚合商 | 覆盖美国 / 加拿大全部 9 个批发市场的大型多市场运营商 | 商业、工业、住宅弹性负荷 | 深厚的市场注册和变现基础设施 | 未明确围绕 AI 数据中心定位 |
| CPower | 工商业 VPP 平台 / NRG 旗下既有厂商 | 由 NRG 支持;美国站点覆盖广 | 工商业站点、分布式能源项目 | 变现平台宽,且有企业能源关系 | 弹性叙事偏通用,不专属数据中心 |
| Virtual Peaker | 公用事业需求响应 SaaS | 私营公用事业软件供应商 | 运营住宅 / 工商业弹性项目的公用事业 | 项目管理栈和设备集成 | 公用事业优先,不是 AI 集群优先 |
| EnergyHub | DERMS / 公用事业弹性平台 | 私营平台,有公开奖项和公用事业验证 | 公用事业与 DER 生态 | 公用事业级弹性和设备生态实力 | 抓取来源中直接面向数据中心的验证较弱 |
| Leap | DER 市场接入平台 | 私营平台,合作伙伴 logo 覆盖广 | 需要项目注册和收入的 DER 业主 | 偏结算和市场接入 | 对 AI 工作负载的控制表述较少 |
| Amperon | 预测 / 分析 | 服务 150+ 家能源领导企业的私营分析厂商 | 公用事业、电力交易商、可再生能源运营商 | AI 预测准确性和风险分析 | 互补性强于替代性 |
| Uplight / EnergyHub / Itron / Enel | 公用事业和清洁能源既有厂商 | 装机基础大或企业覆盖广 | 公用事业和大型能源买方 | 分销杠杆和更宽的解决方案包 | 未必能匹配 Emerald 的工作负载控制专长 |
本表区分直接同业、既有厂商和替代方案。「规模 / 融资」常常只能定性,因为公开的竞争对手材料更强调能力和客户类别,而不是经审计的细分财务。
[CP036, CP001, CP005, CP011, CP025]格局在专业数据中心相关性和现有厂商分发杠杆之间分化。
坐标轴是根据公开定位、客户类别和验证界面推导的序数判断,不是经审计的市场份额指标。
[CP001, CP018, CP015, CP044, CP049]3.2 能力、产品范围与分发对比
比较 Emerald 与同业,最清楚的口径是能力来源。Voltus、CPower 和 Leap 最强的场景,是客户已经拥有灵活资产,并希望完成市场注册、调度和结算。Virtual Peaker、EnergyHub、Uplight 和 Itron 更擅长的场景,是公用事业公司希望跨多种设备运营广泛客户项目。GridPoint 和 Enel 离 Emerald 的核心承诺更远,解决的是楼宇能源或广义清洁能源组合,而不是 GPU 集群编排。Amperon 大多是互补方,因为更好的预测本身并不会创造削减控制。 Emerald 主张的差异化不是宽泛品类规模,而是对最新买方痛点的狭窄相关性。JLL、CBRE 和 Bloom 描述的世界里,电力接入和交付时点主导数据中心决策。Emerald 正是围绕这个问题而建,而多数既有厂商更早诞生,服务于通用需求响应、DER 聚合或公用事业互动。定位强度取决于买方是否真的把 AI 工作负载灵活性视为一项值得付费的独立能力。[CP022, CP023, CP024, CP025, CP019, CP018]
| 采购标准 | Emerald AI | Voltus / CPower | Virtual Peaker / EnergyHub / Uplight | Leap / Amperon / GridPoint | 替代路径(Bloom / 运营商 / 公用事业) |
|---|---|---|---|---|---|
| 明确聚焦 AI 数据中心 | 高 | 低 | 低 | 低 | 中 |
| 公用事业 / 电网项目积累 | 中 | 高 | 高 | 中 | 高 |
| 市场注册 / 结算深度 | 中低 | 高 | 中 | 高 | 低 |
| 遥测 + 弹性负荷运营 | 高 | 高 | 高 | 中 | 中 |
| 在线数据中心弹性的公开验证 | 高 | 低 | 低 | 低 | 中高 |
| 设备 / 资产广度 | 低 | 中 | 高 | 中 | 高 |
| 与数据中心运营商的相关性 | 高 | 中 | 中低 | 低 | 高 |
各单元格是基于公开证据的顺序判断,不是基准测试结果。比较重点放在买方真正关心的能力来源,而非功能清单细节。
[CP022, CP023, CP024, CP025, CP027, CP019]Emerald 在 AI 数据中心专属能力上领先,现有厂商在通用公用事业或市场项目广度上领先。
数值是综合公开资料得出的定性品类强度评估。
[CP022, CP023, CP024, CP025, CP019, CP027]Emerald 在叙事契合度和验证新近度上得分较高,但公开定价和装机基础可见度薄弱。
[CP046, CP020, CP033, CP044, CP032]3.3 定价、包装与切换动态
这一竞争版图的公开定价透明度很差。Voltus 在需求响应侧异常透明,因为它发布了示意性的 MW-year 收益机会,但即便如此,那也不是软件标价。大多数其他供应商只描述结果、合作关系或解决方案族,并不披露合同基准、最低承诺、实施费或实际经济性。因此,本章的定价比较其实是包装方式比较:一些厂商像收入分成型聚合商,一些像公用事业 SaaS 或项目管理栈,另一些则把灵活性打包进更广泛的能源或基础设施解决方案。 切换动态同样不是二元。一旦数据中心、公用事业公司和遥测栈完成集成,流程、风险管理和利益相关方信任都会形成真实切换成本。但多栖部署也可行,因为 Emerald 可以与预测工具、公用事业 DR 软件或现场电力系统并行。因此,分发杠杆——尤其是既有公用事业和能源买方关系——与产品精巧度同样重要。Emerald 的风险在于,既有厂商可以先接触买方,再逐步缩小感知差距。[CP020, CP021, CP026, CP027, CP047, CP018]
| 供应商 / 类别 | 价格 / 单位 / 合同模式 | 标价 vs 实际价格 | 折扣 / 未知项 | 影响 |
|---|---|---|---|---|
| Emerald AI | 未披露;可能是企业合同或与绩效挂钩的合同 | Unknown | 未披露公开定价、实施费或结算分成 | 难以与同业对标 ACV 或利润率 |
| Voltus | 按市场公布总 MW-year 收益机会 | 结果示例,不是软件价格 | 净收入分成、客户分配和实施经济性不清 | 公开价值叙事最透明,但不可与 SaaS 价格直接比较 |
| CPower | 未披露 VPP / 变现合同 | Unknown | 收入分成、软件费和服务组合未公开 | 竞争更可能围绕变现结果,而非标价透明度 |
| Virtual Peaker / EnergyHub / Uplight | 未披露公用事业 SaaS 或平台合同 | Unknown | 未披露公开的公用事业合同基础、模块价格或实施费 | 公用事业采购和打包可能压过纯功能定价 |
| Leap | 平台 + 市场接入经济性未披露 | Unknown | 结算和抽佣条款未公开 | 在客户更看重市场接入而非专项控制时竞争 |
| 替代路径 | 资本开支、电力合同或公用事业电价经济性 | 因案例而异 | 需要电力硬件、公用事业让利或内部人员 | 可能直接拿到预算,甚至不给软件对比留空间 |
公开价格发现不足,本表比较的是合同逻辑,而不是假装各供应商之间存在干净的标价基准。
[CP020, CP021, CP047, CP015]3.4 护城河耐久性与替代风险
Emerald 最好的护城河论据,是一套整合证据:现场事件中的工作负载性能数据、公用事业特定运营手册,以及与 NVIDIA、National Grid、Silicon Valley Power 和更广泛 DCFlex 生态的伙伴可信度。上述都是真实资产,但仍处早期。公开记录显示,Emerald 的具名试点证据强于许多同业,却没有显示长期续约、大规模装机基础或定价权。换句话说,今天的护城河更像叙事加证据,而不是规模加锁定。 反向情景很直接。电力公用事业公司可能无法为灵活性创造足够经济价值,从而压缩整个品类。既有 DR 或公用事业平台厂商可能把软件适配到大负荷和数据中心用例。大型运营商或超大规模云厂商可能把这种能力内化。现场发电等替代路径也会缩小愿意接受基于削减权衡的买方池。在 Emerald 证明可重复的生产级采用之前,竞争风险与其说来自某一个对手,不如说来自更大平台和相邻替代方案对这个品类的吸收。[CP033, CP034, CP029, CP030, CP031, CP032]
| 护城河主张 | 威胁 | 严重性 | 缓释措施 / 尽调追问 |
|---|---|---|---|
| Emerald 占住直接 AI 数据中心弹性叙事 | 既有 DR 或公用事业厂商新增大负荷模块 | 高 | 索取对阵 Voltus、CPower 和公用事业平台既有厂商的竞争输赢数据 |
| 试点验证展示不伤工作负载的弹性 | 试点始终无法转成可重复的生产合同 | 高 | 要求查看已签续约、重复部署和生产 SLA 指标 |
| NVIDIA 与公用事业合作背书提高信任 | 公用事业决定通过更大的既有厂商标准化采购 | 高 | 按公用事业检查 pipeline,并确认 Emerald 是独家来源还是众多供应商之一 |
| 工作负载性能数据变成自有资产 | 超大规模云厂商或大型运营商将流程内化 | 中高 | 复核 IP 归属、模型数据权利和客户自研替代方案 |
| 专注提高产品契合度 | 品类过窄,可能太小,或太容易被替代路径吸收 | 中高 | 建模时只计入弹性能释放真实拿电速度价值的园区采用 |
| 多方集成形成粘性 | 与既有厂商共存会封顶定价权,因为买方把 Emerald 视作叠加层 | 中 | 追问 Emerald 是预算负责人、控制平面,还是可选优化层 |
本清单聚焦足以改变投资判断的耐久性问题,而不是琐碎功能缺口。
[CP033, CP034, CP029, CP030, CP031, CP049]3.5 图表
04财务情况
4.1 收入模式与变现逻辑
Emerald AI 的公开财务叙事,先从它不是什么说起。它不融资建设大型电厂,不拥有数据中心,也不销售大宗电力。公司把自己呈现为控制层,让 AI 数据中心变成用电可调的电网资产。核心变现逻辑因此几乎肯定由软件主导:Conductor 平台、针对具体部署的配置,以及让电力公用事业公司和运营商把电网状态转化为可接受算力响应的运营工作流。收入切入口来自经济价值,不是界面好看。如果 Emerald 帮客户更快接入、避开并网延迟,或从灵活性中捕获可靠性价值,即便还没有形成广泛机群规模,软件也能支撑有意义的合同价值。 挑战在于,公开来源尚未发布任何标价、平均合同价值、使用量定价或节省分成公式。因此,阅读 Emerald 当前收入模式的正确方式,是把它视为一个谈判型企业基础设施产品;其价值取决于当地电网瓶颈、客户工作负载关键性,以及哪一方捕获经济上行。方向上有吸引力,因为严重电力约束能支撑定价权;但也难以承销,因为投资人还无法把具名部署映射成披露过的收入密度。[CI010, CI011, CI013, CI014, CI017, CI018]
| 收入流 | 机制 | 单位 | 当前价值 / 状态 | 质量 | 尽调追问 |
|---|---|---|---|---|---|
| Conductor 软件平台 | 客户为工作负载编排和电网响应控制付费 | 订阅 / 许可(未披露) | 核心变现界面明确;经济性未披露 | 若可经常性收费,质量可能较高 | 索取合同结构、ACV 和续约基础 |
| 实施 / 集成服务 | 部署工程、站点配置和工作流集成 | 项目费或打包服务(未披露) | 早期部署中可能存在 | 中;可能非经常性 | 索取服务收入占比和附加率 |
| 公用事业项目参与支持 | 软件用于弹性或并网项目 | 项目费或服务费(未披露) | 在 SVP 式部署中可见 | 中;取决于项目设计 | 确认谁付费,以及收入是否经常性 |
| 商业试点 / 旗舰部署费用 | 付费价值验证或首站商业落地 | 试点合同或里程碑费用(未披露) | 近期最强的公开候选收入 | 低至中,直到重复性得到证明 | 索取合同期限和成功标准 |
| 潜在共享节省 / 价值定价 | 定价挂钩更快并网、避免电网升级或弹性价值 | 价值共享公式(未披露) | 概念上可行,但未公开 | Unknown | 索取定价逻辑和结算示例 |
| 战略设计伙伴项目 | 与战略投资人或生态伙伴的付费协作 | 商业 / 战略条款混合 | 可能存在,但未公开拆分 | Unknown | 区分战略融资和客户收入 |
除软件驱动核心模式的存在之外,每一行都依赖推断,因为 Emerald 披露的是用例和客户,不是合同模板或价目表。
[CI011, CI013, CI014, CI017, CI012]| 价格 / 单位 / 合同 | 标价 vs 实际价格 | 折扣 / 未知项 | 来源 |
|---|---|---|---|
| 企业软件合同 | 未公开标价 | 实际 ACV 和期限未知 | Emerald / Salesforce / NVIDIA 案例研究 |
| 站点部署 / 集成包 | 未公开套餐定价 | 可能被打包进首个部署项目的经济性 | SVP / National Grid / S&P 证据 |
| 公用事业方支持的灵活性项目费用 | 未披露与电价机制挂钩的 Emerald 费用 | 价值由谁拿走取决于具体项目 | SVP / Heatmap / S&P |
| 快速接入电力溢价 | 未披露明确计价公式 | 取决于避开的延误和当地电力稀缺程度 | CFR / DCD / Series A 公告 |
| 共享节省或按绩效计费部分 | 没有公开证据显示计价公式 | 私下可能存在,但无法纳入投资测算 | 无公开披露 |
| 战略方或渠道驱动的交易支持 | 商业折扣未知 | 投资人重叠可能影响实际价格 | Series A 投资方材料 |
本表有意把定价逻辑与实际报价费率拆开,因为公开记录支持前者,不支持后者。
[CI017, CI018, CI012, CI030]Emerald 的变现桥梁从电力痛点走向协商式软件收入。
[CI017, CI013, CI014, CI035]4.2 GTM 动作、成本结构与单位经济可见度
Emerald 的 GTM 动作更像战略型企业基础设施销售,而不是自助式 SaaS。为公司出资的同一组联盟,也解释了它可能怎样赢单:电力公用事业公司、数据中心运营商、NVIDIA 相关基础设施伙伴和战略投资人都贴近采购中心。这可以降低漏斗顶部的获客摩擦。但它也意味着销售周期长、部署范围需要定制、伙伴协同很重。换句话说,渠道杠杆真实存在,销售效率仍不透明。 可见成本结构也遵循同一模式。Emerald 的软件优先姿态,应让它远比拥有资产的能源基础设施模式轻资本,但公司很可能仍承担昂贵的工程、集成、基准测试和商业开发成本。公开证据没有披露获客成本(CAC)、回本周期、毛利率或贡献利润率,所以单位经济大多仍停留在定性层面。如果实施后软件毛利率占主导,业务最终可能极具吸引力;但今天的公开记录只支持较弱的说法:相比发电重资产替代方案,Emerald 可能更轻资本;相比普通横向 SaaS,它仍更重部署。[CI015, CI016, CI025, CI031, CI032, CI033]
| 指标 | 数值 / null | 置信度 | 为何重要 | 尽调索取项 |
|---|---|---|---|---|
| 获客成本(CAC) | null | 低 | 用于检验合作伙伴驱动的 GTM 是否实质降低获客成本 | 要求提供综合 CAC 和渠道来源 CAC |
| 销售周期长度 | null | 低 | 基础设施相邻交易可能周期长、耗现金 | 要求按公用事业方和运营商细分提供中位周期 |
| 毛利率 | null | 低 | 决定部署后是否由软件经济性主导 | 要求按合同类型提供毛利率 |
| 实施后贡献利润率 | null | 低 | 显示早期部署是否具备经济可扩展性 | 要求提供计入服务负担后的部署级 P&L |
| 单站点实施负担 | 定性偏高 | 中 | 定制化可能压住扩展性,推迟毛利率爬升 | 要求提供平均工程工时和集成步骤 |
| 相比自持资产方案的资本强度 | 低于重发电资产模式;高于纯 SaaS | 中 | 框定该模式需要多少融资 | 对标纯软件和重基础设施同行 |
公开证据支持该模式的相对定位,不支持绝对单位经济输出。
[CI031, CI032, CI033, CI034]单位经济模型缺失的部分,卡在企业需求和可规模化利润率之间。
[CI015, CI016, CI032, CI031]4.3 资本充足性与融资依赖
资本充足性是 Emerald 公开财务档案中最强的一部分,但仍留下重要盲点。融资序列异常快:以 $24.5 million 种子轮公开亮相,随后用 $18 million 延展把规模推至 $42.5 million,再用 $25 million 战略扩张轮把融资推至约 $68 million,最后以 $1.05 billion 估值完成 $150 million Series A。SEC Form D 文件佐证了发行规模越来越大、投资人参与越来越广的模式。截至 2026 年 8 月,公司已披露累计融资约 $217.5 million。 资本基础很重要,因为 Emerald 正从演示走向商业部署,需要工程支持、伙伴管理、产品硬化和全球商业扩张。但已披露融资额不等于手上现金,所审阅来源也没有发布烧钱速度或现金跑道。结果是一幅单边图景:投资人能看到 Emerald 相比多数气候软件创业公司资金充足,却看不到公司消耗这一优势的速度,也无法判断哪个里程碑会触发下一轮融资。实践中,下一轮触发因素很可能取决于当前旗舰部署能否在 Series A 现金优势被扩张成本吃掉之前,转化为可重复、收入密度高的商业项目。[CI001, CI002, CI003, CI004, CI005, CI006]
| 项目 | 公开数值 / 状态 | 日期 | 来源 | 含义 | 缺口 |
|---|---|---|---|---|---|
| 种子轮 | 披露 24.5M | 2025-07 | PR Newswire | 为演示和发布建立初始资本基础 | 剩余现金未知 |
| 种子轮扩展 | +18M 后披露累计 42.5M | 2026-02 | Emerald AI | 延长现金跑道,并扩大战略投资人基础 | 轮次之间烧钱速度未知 |
| 战略扩展轮 | +25M 后披露累计 68M | 2026-03 | Emerald AI | 在全面商业化规模前引入渠道属性重的战略资本 | 现金余额仍未披露 |
| Series A 轮 | 150M,估值 1.05B | 2026-08 | Emerald AI + SEC | 显著增强商业化扩张所需的资产负债表承载力 | 现金跑道仍未披露 |
| 2026 年 8 月 Form D 进展 | 已售 90.23M,剩余 59.77M,23 名投资人 | 2026-08-03 | SEC | 显示截至提交日该轮尚未完全交割 | 最终交割机制未知 |
| 债务 / 项目融资 | 无公开披露 | 2026 | 已审阅公开来源 | 未看到明显再融资负担 | 需要债务期限表和约束条款细节 |
本表聚焦前瞻性资本充足性,而不是逐字重复公司概况中的融资时间线。
[CI004, CI005, CI006, CI007, CI008, CI036]公开证据能给出较窄的融资区间,却只能给出较宽的实际财务表现区间。
这是融资可见度图,不是收入预测。公开资料支持资本区间,但不支持收入或烧钱区间。
[CI005, CI006, CI007, CI008]Emerald 的融资风险来自商业化转化,而不是厂站级资本开支。
[CI036, CI023, CI035, CI037]4.4 财务判断与承销缺口
公开记录足以支撑可信的财务叙事,但不足以完成干净承销。Emerald 显然有一个可以变现的问题要解决:电力约束正在恶化,数据中心开发商非常在意拿电速度,公司现在也有可信的商业化证据。上述条件让有意义的企业合同存在变得可信。但确认收入质量所需的财务证据仍然缺失。没有披露年经常性收入(ARR),没有合同价值分布,没有续约证据,没有利润率堆栈,也没有从烧钱速度到现金跑道的桥。 因此,正确判断不是 Emerald 缺少商业模式,而是公开证据只证明了商业模式的形状,还没有证明其经济性。投资人应把 Emerald 视为财务上有前景、但证据仍偏薄的公司。最有价值的尽调请求,是那些能最快压缩不确定性的问题:客户合同价值、实施负担、部署后的毛利率画像、销售效率、按 logo 和站点划分的集中度,以及与商业里程碑绑定的具体现金跑道计划。相邻的上市电力和数据中心平台也提醒了披露缺口:它们公开报告数十亿美元级收入基础,而 Emerald 仍未披露任何可比规模指标。在这些披露出现之前,公司的财务质量方向上有吸引力,但尚未被充分承销。[CI023, CI024, CI019, CI020, CI022, CI035]
| 未披露的私有指标 | 影响 | 精确尽调路径 |
|---|---|---|
| 按分部划分的 ARR / 收入 | 缺少这项,估值和资本效率无法锚定 | 要求按公用事业方 / 运营商 / 战略客户提供已签 ARR、确认收入和销售管线 |
| 平均合同价值和期限 | 缺少 ACV 和期限,收入质量和定价权无从判断 | 要求提供前 20 大合同的 ACV、期限、续约和定价依据 |
| 毛利率与服务组合 | 缺少毛利拆分,软件可扩展性只是推测 | 要求按合同类型提供毛利率和服务占比 |
| CAC、销售周期与回本周期 | 缺少效率指标,GTM 可扩展性尚未被证明 | 要求提供按渠道划分的 CAC、管线转化率和回本周期 |
| 烧钱速度与现金跑道桥接 | 看不清现金消耗,资本充足性判断就是单边的 | 要求提供当前现金、月度烧钱、招聘计划,以及不同情景下的现金跑道 |
| 按 logo 与站点划分的客户集中度 | 缺少集中度数据,收入耐久性容易被高估 | 要求提供最大客户占 ARR 和销售管线的比例 |
| 部署到收入的转化 | 缺少阶段转化数据,旗舰验证可能夸大货币化 | 要求提供试点到生产的转化率和实施时间表 |
这些路径最快能把 Emerald 从一个有吸引力的故事,变成可融资的投资判断。
[CI019, CI020, CI021, CI038, CI039, CI040]4.5 图表
05产品与技术
5.1 产品定义与公开模块地图
以当前阶段的公司而言,Emerald AI 的公开产品叙事异常具体。公司没有把自己描述成通用能源软件或通用数据中心管理工具,而是反复将产品定位为让 AI 数据中心成为用电可调电网资产的软件。换成客户语言,产品服务于希望获得更多电力接入或电网响应能力、同时不降低优先工作负载质量的运营商和电力公用事业公司。这条工作流比经典 DER 管理或公用事业需求响应工具窄得多。 公开模块地图仍稀疏,但真实存在。Emerald Conductor 是合作伙伴、公用事业和媒体来源中共同出现的核心平台名称。GridLink 则作为支持产品出现,尤其在 Aurora 架构叙事中,用于把电网要求连接到数据中心运营。除这些名称外,大多数能力按功能描述,而不是作为单独 SKU 呈现。这种呈现符合一家仍围绕单一旗舰运营产品商业化控制层的公司,而不是一家已经营销成熟多模块软件套件的公司。[CE001, CE002, CE003, CE005, CE006]
| 模块 / 资产 | 用户 | 状态 / 成熟度 | 差异化 | 尽调缺口 |
|---|---|---|---|---|
| Emerald Conductor | 数据中心运营商 / 公用事业对手方 | 相对成熟度高;演示和试点中的核心上线产品 | 直接做到 AI 基础设施的工作负载级电力灵活性 | 定价、部署数量和可靠性指标未公开 |
| GridLink | 电网 / 运营商集成层 | 中;有公开提及,但描述少于 Conductor | 把电网需求接到数据中心运营控制 | 与 Conductor 的功能边界未充分披露 |
| DSX Flex 集成 | AI 基础设施 / NVIDIA 栈用户 | 中高;有商业试点和路线图证据 | 把电力灵活性嵌入 AI 工厂运营栈 | 非 NVIDIA 环境可移植性的证据有限 |
| 公用事业调度接口 | 公用事业规划人员和运营人员 | 中;SVP 和 National Grid 场景提供了证据 | 让公用事业方请求或验证灵活响应 | 没有公开的标准化 API 或协议文档 |
| 优化策略库 | Emerald 运营 / 站点控制层 | 中;GitHub 伪代码和论文提供了证据 | 策略类型不止静态限频 | 没有公开的基准库或生产治理文档 |
| 遥测 / 验证层 | Emerald + 公用事业方 + 站点运营商 | 中;演示和试验中均有体现 | 闭合目标功率与工作负载约束之间的反馈环 | 没有公开的可观测性或审计报告规范 |
资产图谱来自公开产品名称和已展示功能。应把它看作逻辑模块图,而不是完整 SKU 目录。
[CE002, CE003, CE004, CE005, CE024]5.2 架构与运营工作流
最强的架构证据来自 GitHub 演示材料、Phoenix 论文、Latitude 访谈,以及 Emerald 与 NVIDIA 相关的发布帖。合在一起看,它们指向一个从外部电力约束开始、再下沉到工作负载级控制的工作流。电力公用事业或系统事件先定义目标。Emerald 为活跃任务画像,分类其灵活性,评估干预选项,并下发控制策略;策略可以包括功率上限、暂停、检查点或地理路由。随后,遥测检查结果中的电力轨迹和性能阈值是否仍可接受。 该架构很重要,因为它显示 Emerald 不只是预测或建议。产品坐进了运营闭环。公开材料也清楚表明,公司当前架构与 NVIDIA 系统深度交织,尤其是 DSX Flex、NIM 微服务和 Mission Control。好处是可信度提升更快、技术集成更紧;代价是明确的伙伴依赖,以及今天硬件无关部署的公开证据更少。[CE007, CE008, CE009, CE010, CE011, CE018]
| 用户任务 | 当前工作流 | 公司方案 | 可衡量收益 | 限制 |
|---|---|---|---|---|
| 响应公用事业电网事件 | 人工或粗粒度削减负载、备用发电,或不响应 | Conductor 为工作负载画像,并施加细粒度控制 | Phoenix 3 小时 25%;英国试验最高 40% | 证明基础仍限于少数公开部署 |
| 加快并网 | 等待可保障电力,或增加昂贵的现场发电 | 与公用事业框架绑定的灵活负载运营层 | 可能更快接入现有电网余量 | 取决于公用事业方是否提供真实灵活负载通道 |
| 在限电期间保护优先级 AI 任务 | 过度配置,或完全避开灵活调度 | 感知优先级的调度、暂停、DVFS 和恢复逻辑 | 公开来源称,测试中关键工作负载持续运行 | 没有公开 SLA 或长时段可靠性数据集 |
| 借助公用事业信号运营商业 AI 园区 | 运营商、公用事业方和供应商之间的定制化人工协调 | 与 DSX Flex 和调度接口集成的工作流 | 从演示推进到 SVP 的商业化多 MW 试点 | 当前叙事仍以 NVIDIA 为中心 |
| 支持地理感知灵活性 | 人工迁移负载,或完全不迁移 | ArXiv 和 WEF 材料描述了跨站点路由或迁移工作负载 | 可让算力匹配压力更低或更清洁的电网 | 多站点生产运营的公开证据仍有限 |
收益来自演示和合作伙伴表述;它们还不等同于广泛的生产基准集。
[CE007, CE010, CE012, CE013, CE022]| 层 / 组件 | 角色 | 依赖 | 风险 |
|---|---|---|---|
| 电网信号摄取 | 接收事件时间、功率目标和限电条件 | 公用事业方或电网运营商接口 | 对手方不提供可执行信号时没有价值 |
| 功率目标塑形 | 把事件定义转成分时段功率预算 | Emerald 控制逻辑 | 目标构造差会过度约束工作负载 |
| 工作负载画像 | 按灵活性、优先级和吞吐容忍度标记任务 | 访问工作负载遥测和历史画像 | 画像不准会拉低 QoS,或减少可实现灵活性 |
| 优化策略引擎 | 在任务和功率旋钮之间选择控制场景 | Conductor 逻辑、模型假设、站点策略 | 优化错误可能错过目标,或损害性能 |
| 执行控制 | 应用 DVFS、暂停、检查点、GPU 分配或路由 | 计算栈权限和 NVIDIA 绑定集成 | 硬件 / 软件依赖压缩可移植性 |
| 遥测与验证 | 对照阈值衡量实际功率和工作负载结果 | 电表、集群遥测、可观测性管线 | 审计能力不足可能削弱公用事业方信任 |
本架构把公开材料抽象成功能组件。Emerald 尚未发布完整内部技术设计文档。
[CE018, CE019, CE020, CE021, CE023]Emerald 的公开技术栈从电网信号下沉到工作负载控制,再通过遥测回到上层。
该技术栈综合自 GitHub 伪代码、研究论文和合作伙伴公告,不是 Emerald 官方架构图。
[CE018, CE019, CE020, CE021, CE011]Emerald 如何从一次电网事件走到已验证、对工作负载安全的降电。
该序列来自公开演示和伪代码,不是完整内部运行手册。
[CE007, CE008, CE009, CE010, CE011]5.3 部署、依赖与成熟度
Emerald 的成熟度主张更多依靠现场证据,而不是公开客户部署的广度。Phoenix 在 256-GPU 集群上演示了持续 3 小时 25% 功率降低。英国试验用超过 200 次模拟电网事件和更快响应动态扩展了证据基础,而 Silicon Valley Power 部署被描述为首个商业化、多兆瓦 DSX Flex 实施。路线图随后指向 96 MW Aurora 参考部署,以及 ERCOT 灵活负荷路径等更广泛的公用事业框架。 依赖关系很重。Emerald 依赖电力公用事业公司或电网运营商发送有意义的信号,依赖数据中心运营商允许运营控制,也依赖 NVIDIA 阵营基础设施来支撑公开描述的当前技术栈。上述依赖并不否定产品,但确实意味着 Emerald 应被视为多方部署业务,而不是简单的自助式软件工具。产品在技术上有差异化,但商业化仍高度依赖协同。[CE012, CE013, CE014, CE015, CE016, CE017]
| 控制项 / 指标 | 状态 | 范围 | 缺口 |
|---|---|---|---|
| 关于 AI 训练的隐私声明 | 公开称网站收集的个人数据不会用于训练 AI 模型 | 网站隐私处理 | 未直接说明客户运营数据如何使用 |
| 安全保障披露 | 公开称存在技术、行政和组织保障措施 | 网站和个人信息控制 | 没有公开认证、控制项映射或审计报告 |
| 前瞻性陈述免责声明 | 条款和条件中明确写明 | 所有公开网站声明和预计部署 | 体现管理层谨慎,不代表运营保障 |
| 禁止抓取 / 禁止模型训练条款 | 网站条款中明确写明 | 网站 IP 和数据挖掘限制 | 法律声明,不是产品安全态势证据 |
| 正式认证(SOC 2 / ISO 27001 等) | 已获取的公开来源中未见 | 会影响企业采购 | 需要安全尽调包或信任中心 |
| 可靠性 / 性能保障报告 | 仅有合作伙伴案例研究和演示 | 选定试点和测试 | 没有公开的广泛生产可用性或事故统计 |
本表区分网站法律 / 隐私控制与产品保障控制。前者存在;后者在公开记录中仍然薄弱。
[CE031, CE032, CE033, CE035, CE036, CE043]| 日期 / 阶段 | 功能 / 里程碑 | 状态 | 含义 | 来源 |
|---|---|---|---|---|
| 2025-05 演示 | Phoenix 256 块 GPU 现场演示 | 已完成 | 建立首个不伤工作负载的实时限电证明 | NVIDIA / Public Power / Phoenix 论文 |
| 2025-10 发布 | Aurora 电力灵活 AI 工厂参考设计 | 已宣布 | 范围从演示扩到参考架构和认证目标 | Emerald / Public Power |
| 2026-03 框架 | NVIDIA DSX Flex 商业试点框架和适配 ERCOT 的定位 | 已宣布 | 产品叙事转向商业部署和灵活并网项目 | Emerald |
| 2026-08 试验 | National Grid 英国试验 | 已完成 | 在欧洲补上快速响应和持续灵活性证据 | National Grid / NVIDIA |
| 2026-08 部署 | SVP 商业化多 MW 部署 | 进行中 / 已宣布 | 最接近实时商业推出的公开证明 | SVP / Emerald |
| 计划于 2026 年晚些时候 | Manassas 96 MW 商业规模部署 | 计划中 | 检验产品能否扩到试点规模之外 | Emerald / NVIDIA / SVP 框架 |
路线图按里程碑呈现,因为 Emerald 不发布常规产品发布日志。
[CE037, CE038, CE039, CE040, CE041, CE017]Emerald 位于多方部署链条中间。
该图聚焦公开资料可见的外部依赖,不覆盖每个内部软件服务。
[CE023, CE024, CE025, CE026]编排和现场验证证据最强,广泛信任披露和规模化运营证据较弱。
成熟度水平是分析师基于公开产品证据作出的判断,不是内部 QA 评分。
[CE012, CE013, CE016, CE027, CE036, CE043]5.4 差异化、信任与公开控制缺口
区分度最好的公开证据在于,Emerald 反复证明同一件狭窄的事:AI 工作负载可以响应电网需求,同时守住优先服务质量。这比广义公用事业灵活性厂商、预测供应商或通用数据中心软件提出的价值主张更具体。因此,最强的护城河候选不是品牌本身,而是现场性能数据、公用事业集成手册和 NVIDIA 相关运营知识的组合。 同时,公开控制面仍很薄。法律页面异常明确地说明,网站内容包含前瞻性陈述,不构成专业建议,也可能无法预测未来结果。隐私政策有用——它说明网站收集的个人信息不会用于训练 AI 模型,并且公司维护保障措施——但已抓取的公开记录并未暴露正式安全认证、模型治理审计或可靠性认证。尽调视角下,Emerald 的产品比纯概念更有技术根基,但在企业级信任和保障披露上仍处早期。[CE022, CE027, CE028, CE029, CE030, CE031]
5.5 图表
06客户情况
6.1 客户分层与买方地图
Emerald AI 的客户地图是多边的,因为只有电力系统参与者和算力参与者对齐时,产品才创造价值。电力公用事业公司和公共电力供应商可以是直接客户,因为它们可能部署软件、调度灵活负荷,并用它支持并网或可靠性目标。数据中心运营商、AI 工厂开发商、新云厂商和超大规模云厂商相邻运营商也是直接经济受益方,因为更快拿电或减少电网约束会实质性改善部署时点。PJM 和 EPRI 等电网机构更像生态促成方,而非直接经常性客户;战略投资人则可能充当渠道加速器和未来设计伙伴。 因此,不能用普通单一买方企业 SaaS 品类来建模 Emerald 的客户基础。同一账户中,用户、买方和付款方可能不同。市政电力公用事业公司可能赞助灵活负荷框架,运营商可能集成软件,云厂商或数据中心租户可能拿到运营收益。这种复杂性增加摩擦,但也意味着成功部署可以创造多个希望关系随时间加深的利益相关方。[CU001, CU002, CU003, CU004, CU031, CU032]
| 细分市场 | 买方 / 用户 / 付款方 | 用例 | 规模 | 收入 / 战略价值 | 缺口 |
|---|---|---|---|---|---|
| 公用事业方 / 公共电力供应商 | 买方:公用事业方;用户:电网规划人员 / 运营人员;付款方:公用事业方或电价机制 | 灵活负载调度、并网管理、可靠性 | 已点名:SVP、National Grid | 为 Emerald 打开监管和商业路径 | 无公开合同价值或续约数据 |
| AI 数据中心运营商 / 云运营商 | 买方:运营商或基础设施团队;用户:站点运营 / 工作负载调度人员;付款方:运营商 | 更快获得电力接入,并响应电力事件 | Phoenix、英国和 Aurora 生态均被点名 | 直接运营受益方,且可能成为未来 ACV 锚点 | 公开客户数未披露 |
| 数据中心业主 / 开发商 | 买方:园区或基础设施负责人;用户:运营 / 租赁支持 | 电力灵活性参考园区与租户支持 | 点名:Digital Realty Aurora | 潜在设施组合级扩张路径 | 商业状态仍多停留在路线图 |
| 电网机构 / 生态项目 | 买方:不清楚是否直接;用户:市场 / 项目人员;付款方:视项目而定 | 基准评测、测试验证和市场设计 | 点名:PJM、EPRI DCFlex、DOE Genesis | 渠道与信任放大器 | 不等同于经常性订阅客户 |
| 战略投资方 / 设计伙伴群体 | 买方:混合;用户:创新或战略负责人;付款方:混合 | 设计合作、渠道支持或未来客户路径 | 已披露 12 家 Fortune 500 共同投资方 | 可能加快进入企业客户 | 身份重叠会遮住独立市场宽度 |
这里有意把客户角色拆成买方、用户和付款方,因为 Emerald 的部署设计上就是多方参与。
[CU001, CU002, CU003, CU004, CU032]6.2 已点名客户证据与采用轨迹
公开记录最强的地方是具名证据项,而不是客户数量。Phoenix 围绕 Oracle、NVIDIA、Databricks 和 Salt River Project 建立了第一个持久运营叙事:一个 256-GPU 集群在 3 小时内降低 25% 功率。英国试验加入第二个地理区域和更强的快速响应证据,披露超过 200 次模拟电网事件,以及不到 1 分钟内超过三分之一的降幅。Silicon Valley Power 随后把叙事推近商业化,称其部署为首个商业化、多兆瓦 DSX Flex 实施。Aurora 通过 Digital Realty 和 PJM 补上旗舰数据中心业主与电网市场运营商语境,但今天它仍更像路线图,而不是广泛经常性收入的证据。 合在一起,采用路径看起来是从演示,到商业试点,再到旗舰参考部署。路径令人鼓舞,但仍不等于披露了大型客户基础或广泛生产机群。[CU005, CU006, CU007, CU008, CU009, CU011]
| 指标 | 值 | 日期 | 来源 | 置信度 | 含义 | 缺失分母 |
|---|---|---|---|---|---|---|
| 已披露现场演示 | 5 | 2026 | NVIDIA 案例研究 + WEF | 高 | 证据集不止单个展示事件 | 合格管线总量未知 |
| Phoenix 降载结果 | 25%,持续 3 小时 | 2025-05 | NVIDIA / Latitude / Public Power | 高 | 显示在 SLA 约束下可持续响应事件 | 重复频率未知 |
| 英国快速响应结果 | >33%,不到 1 分钟;最高 40% | 2026-08 | National Grid / NVIDIA | 高 | 显示在真实公用事业场景中的快速响应能力 | 能否转为经常性商业合同未知 |
| 英国模拟事件 | 5 天内 200+ | 2026-08 | National Grid | 中 | 说明能反复处理事件,而不只是一次脉冲测试 | 长期生产节奏未知 |
| SVP 商业状态 | 首个商业化多 MW DSX Flex 部署 | 2026-08 | SVP | 中 | 说明公司已不只停留在试点状态 | 收入金额或客户数未知 |
| Fortune 500 共同投资方 | 12 | 2026-08 | Series A 公告 | 中 | 显示战略需求面与渠道价值 | 多少是客户、多少只是投资方未知 |
轨迹表只保留公开记录真正支持的内容:里程碑和结果,而不是整齐的客户数量时间序列。
[CU005, CU007, CU009, CU010, CU012, CU004]| 客户 / 对手方 | 细分领域 | 部署 / 用例 | 生产环境与试点 | 结果 | 限制 |
|---|---|---|---|---|---|
| Salt River Project + Oracle / NVIDIA / Databricks 集群 | 公用事业 + 运营方生态 | 基于 256 块 GPU 的 Phoenix 电网压力响应 | 试点 / 演示 | 在 SLA 限制内降载 25%,持续 3 小时 | 单一站点;续约经济性未公开 |
| National Grid + Nebius | 公用事业 + AI 工厂运营方 | 英国电网响应型 AI 集群试验 | 试点 / 现场试验 | 不到 1 分钟削减 >33%;最高 40%;200+ 次事件 | 商业后续仍未披露 |
| Silicon Valley Power + NVIDIA 站点 | 公用事业主导商业试点 | 多 MW 规模灵活负载互联调度 | 商业试点 / 已公告部署 | 定位成首个商业化 DSX Flex 部署 | 合同收入或重复使用数据未公开 |
| Digital Realty + PJM + EPRI Aurora 项目 | 业主 / 电网生态旗舰 | 96 MW 参考 AI 工厂 | 参考部署 / 路线图 | 大规模设计伙伴证据和未来商业测试平台 | 尚不能证明广泛经常性收入 |
| Fortune 500 共同投资方群体 | 战略投资方 / 潜在客户渠道 | 潜在设计伙伴与客户引介入口 | 渠道信号,不是部署证据 | 显示与企业客户相关,不限于一家公用事业公司 | 群体身份及转化情况未公开 |
这里的点名证据不只是 logo:每行都把具体对手方与用例、状态,以及至少一个已披露结果或含义连起来。
[CU006, CU008, CU011, CU013, CU015, CU038]指数化漏斗,展示电力受限潜在客户如何走到可复制的项目铺开。
数值是指数化逻辑标记,不是公司披露的转化率;它们展示商业化瓶颈目前卡在哪里。
[CU017, CU039, CU037]Emerald 同时给出交易对手、部署类型和可测电力结果时,证明质量最强。
单元格是对证据质量和具体程度的序数判断;不是调查结果。
[CU015, CU029, CU030, CU038]6.3 持久性、重复使用与满意度
公开记录在客户持久性上明显变薄。Emerald 有可信的重复生态参与证据——NVIDIA、电力公用事业公司和电力市场参与者在多个证据点反复出现——但没有发布经典 SaaS 持久性指标,例如净留存率(NRR)、总留存率(GRR)、流失率、logo 留存或合同期限。已抓取证据也没有提供终端客户证言,明确讨论续约、长期实际 ROI 或试点后大规模生产推出。 最乐观的解读是,Emerald 正沿着从一次性证明到嵌入式商业角色的路径前进,尤其是在 SVP 灵活负荷项目等公用事业框架创造经常性运营需求的地方。谨慎解读则是,同一小圈伙伴可能撑起了全部可见需求叙事。承销时,真正未知的不是技术能不能跑,而是账户会不会变成持久、扩张的关系,而不是停留在展示性部署。[CU018, CU019, CU020, CU034, CU037, CU035]
| 指标 | 值 / null | 细分领域 | 置信度 | 尽调问题 |
|---|---|---|---|---|
| 净收入留存率(NRR) | null | 所有细分领域 | 低 | 索取按公用事业、运营方和旗舰站点划分的队列 NRR |
| 总收入留存率(GRR) | null | 所有细分领域 | 低 | 索取按部署类别划分的 GRR 和流失率 |
| 合同期限 | null | 公用事业与运营方账户 | 低 | 索取试点期限、续约选项和扩张权利 |
| 同一生态内重复部署 | 可见但未量化 | NVIDIA + 公用事业生态 | 中 | 索取重复账户数和生产环境转化数 |
| 独立客户满意度 / 评价语料 | null | 所有细分领域 | 低 | 索取客户推荐、NPS/CSAT 和试点后反馈报告 |
表中在公开记录无法支撑耐久性指标处有意保留 null。
[CU018, CU019, CU034, CU037]Emerald 的潜在客户旅程始于电力痛点,只有试点验证转成经常性运营关系,旅程才算走到终点。
[CU025, CU017, CU035, CU028]用代理指标按客户类型看持续性,显示公用事业嵌入式部署比展示型验证更有黏性。
百分比只是代理指标。Emerald 未发布实际队列,因此图表展示的是不同验证类型可能的相对持续性,而非公司披露的留存。
[CU018, CU019, CU020, CU037]6.4 扩张循环与集中度风险
扩张逻辑很直观:如果一个站点的用电可调部署跑通,就可以复制到新园区、公用事业辖区或多站点机群。可见路径是从 Phoenix 式可测量证明,推进到 SVP 等公用事业标准化项目,再到 Aurora 这样更大的 AI 工厂园区。战略投资人和伙伴生态可能放大这一动作,因为同一批参与者可以在资本、技术集成和客户介绍上帮忙。 但集中度风险同样清晰。大量公开证据依赖一个狭窄的具名伙伴圈:NVIDIA、National Grid 和 SVP 等电力公用事业公司、Digital Realty,以及少数演示站点。采购摩擦也很高,因为交易往往要求监管机构、电力公用事业公司、基础设施提供商和运营商对齐。Heatmap 的反向观点仍然重要:如果公用事业公司不能创造有意义的并网或经济价值,即便客户有兴趣,也可能无法规模化转化。因此,客户章节更多讨论协同成功下的扩张潜力,而不是已经证明的账户持久性。[CU026, CU027, CU022, CU023, CU024, CU028]
| 扩张驱动因素 | 集中度风险 | 影响 | 尽调路径 |
|---|---|---|---|
| 公用事业项目标准化 | 目前证据集中只看得到少数几家公用事业企业 | 高 | 按公用事业企业、阶段和已签项目类型梳理管线 |
| NVIDIA 生态杠杆 | 高度依赖单一计算栈生态 | 高 | 索取可移植性路线图和非 NVIDIA 商业证据 |
| 旗舰参考园区 | 少数标杆站点可能主导叙事和管线 | 高 | 索取按站点、客户标识和预期收入占比划分的集中度 |
| 战略投资方重叠 | 投资方不一定等于独立需求 | 中高 | 将战略内部人导入管线与自然需求管线分开 |
| 地理复制 | 公开证据覆盖多个地区,但总地域数不多 | 中 | 索取按地区划分的公用事业与园区扩张计划 |
| 多方采购 | 协调公用事业企业、运营方和监管机构会拉长销售周期 | 高 | 索取平均周期长度、阻碍因素和转化率 |
让 Emerald 具备战略重要性的那些特征,也同时带来集中度和采购风险。
[CU026, CU027, CU022, CU023, CU028, CU040]6.5 图表
07风险
7.1 监管与法律风险
Emerald 的产品正处在监管转型的正中央。需求因此受益,可预测性却受损。FERC 正迫使电网运营商说明或改革大负荷规则,PJM 也在明确考虑框架,要求新增大负荷带来容量或接受更早削减;全美电力公用事业公司则争相设计特殊电价,以保护现有缴费用户免受数据中心风险影响。上述进展验证了 Emerald 关于灵活性重要性的判断。监管变化也让商业化面对移动靶,因为奖励灵活性的同一套政策制度,也可能把抵押义务、削减权、最低期限或直接分摊成本转移给客户。 Emerald 自身法律披露进一步强化了谨慎必要性。网站条款称,试点和演示具有示意性且依赖具体条件,前瞻性陈述天然不确定,公司也不承诺更新公开声明。隐私政策体现了基础法律 / 隐私卫生,但不能替代公开信任中心、认证或企业保障材料。与此同时,2026 年中 FERC 到 NERC 围绕强制性计算负荷标准的转向显示,这一段监管边界正在快速硬化。因此,法律风险与其说来自可见诉讼,不如说来自对演进中电价制度的依赖,以及有限的公开合规证明。[CR001, CR002, CR006, CR008, CR009, CR010]
| 规则 / 牌照 / 案件 | 司法辖区 | 状态 | 可能性 | 严重性 | 缓释措施 | 剩余风险敞口 | 尽调路径 |
|---|---|---|---|---|---|---|---|
| FERC 大负载电价改革 / 说明理由令 | 美国 RTO/ISO 市场 | 2026 年改革周期正在推进 | 高 | 高 | Emerald 把产品对齐灵活负载路径,而不是与其对抗 | 规则仍可能实质改变价值、时点和客户义务 | 跟踪每项相关 RTO 申报,并询问管理层哪些电价路径对收入至关重要 |
| PJM IRAS / BYONC / 登记框架 | PJM / 州公用事业接口 | 2026 年 8 月提交;拟适用于 2027+ 负载 | 中高 | 高 | 面向能从灵活性和容量支持型互联中受益的客户销售 | 客户可能更快接入,但要承受优先被削减的风险敞口或更高合规负担 | 审查客户在 BYONC、削减权、遥测和补偿规则下的敞口 |
| 抵押品、最低期限、退出费、直接成本分摊等大负载电价保护 | 州公用事业电价 / 特殊合同 | 正在快速扩散 | 高 | 高 | 将 Emerald 定位为改善电价经济性和合规的工具 | 保护条款可能压缩或延后可触达需求 | 在给管线赋值前,按市场梳理目标公用事业电价条款 |
| Order 2222 / DER 协调不成熟 | 州 + 配电公用事业层 | 截至 2026 年初实施仍不完整 | 中 | 中高 | 先用更简单的双边公用事业项目 | 协调缺口会延迟或复杂化市场参与设计 | 询问哪些部署依赖尚未解决的配电 / 批发协调 |
| 隐私、安全和公开合规证据缺口 | 企业采购 / 隐私法 | 政策公开;认证未公开 | 中 | 高 | 已有法律 / 隐私政策,并陈述了基本防护措施 | 缺少公开保证材料可能拖慢企业交易,或提高尽调摩擦 | 索取信任中心材料、DPA 模板、认证和事件响应流程 |
该登记表认为,市场规则波动叠加合规证据单薄,比任何可见诉讼风险更关键。
[CR001, CR002, CR008, CR009, CR014]多数头部风险先传导到客户经济性,再传导到采用、收入质量和估值。
[CR002, CR029, CR022, CR028, CR037]7.2 运营、安全与依赖风险
运营层面,Emerald 的风险画像像一层控制层:它同时触及关键任务计算和面向电网的响应。以公司年龄看,公开验证 已经很亮眼,但仍集中在少数点名演示和标杆部署上。Emerald 自己的条款也强调,这些验证只是示例, 且绑定特定条件;因此,投资者不能把 Phoenix、英国试验或 SVP 的结果外推成普遍生产就绪。真实电网事件中, 一旦控制层表现不佳,Emerald 失去的不只是软件 KPI;客户工作负载可能受损,电力公司信任会受伤,灵活并网的 商业论证也会被削弱。 依赖风险同样明显。公开部署叙事与 NVIDIA 技术栈、电力公司项目设计,以及少数标杆对手方绑得很紧。今天, 这些关系是战略强项;反过来,也说明 Emerald 还没有证明自己能广泛迁移到不同生态、费率框架或客户类型。 替代风险也真实存在:部分客户可能认为,现场发电、容量采购或定制合同,比采用一层编排系统更简单;后者的 经济性取决于多方共同创造价值。[CR015, CR016, CR017, CR018, CR019, CR020]
| 故障模式 | 可能性 | 严重性 | 缓释成熟度 | 剩余风险敞口 | 未解决缺口 |
|---|---|---|---|---|---|
| 真实电网事件中,控制动作损害工作负载质量或未达 SLA | 中 | 极高 | 中 | 高 | 需要试点之外更广泛的生产环境 SLA 证据 |
| 遥测或通信故障打断调度协调 | 中 | 高 | 中低 | 高 | 需要故障安全和降级模式设计评审 |
| 编排或遥测层遭网络入侵 | 中低 | 极高 | 中低 | 高 | 无公开保证材料或事件历史 |
| 异构客户工作负载表现差于演示工作负载 | 中 | 高 | 中 | 中高 | 需要按工作负载类别划分的性能证据 |
| 支持组织跟不上商业化推广节奏 | 中 | 高 | 中低 | 中高 | 公开记录很少提及规模化现场运营 |
| 实测的试点结果无法在设施组合规模复现 | 中 | 高 | 中 | 中高 | 需要跨站点、跨时间的可重复性数据 |
运营风险被放大,因为 Emerald 的控制回路卡在客户正常运行时间和电网响应的交界处。
[CR015, CR016, CR017, CR018, CR027]| 依赖项 | 对手方 | 角色 | 集中度 | 失败场景 | 严重性 | 缓释措施 | 剩余风险敞口 |
|---|---|---|---|---|---|---|---|
| NVIDIA 软件和参考设计生态 | NVIDIA | 计算栈、可信度、部署路径 | 高 | 在非 NVIDIA 证据出现前,可移植性或合作关系走弱 | 高 | 拓宽集成并证明栈可移植 | 高 |
| 公用事业项目经济性 | SVP、National Grid、未来公用事业企业 | 调度权和经济价值 | 高 | 公用事业企业支付不足,或未能标准化项目 | 高 | 瞄准有明确灵活负载路径的市场 | 高 |
| 旗舰站点集中 | Digital Realty / Aurora / 少数试点站点 | 叙事和潜在管线锚点 | 高 | 一个展示站点延期或表现不佳,就会损害更广泛的需求故事 | 高 | 分散点名部署并发布重复证据 | 中高 |
| 合作伙伴主导的 GTM 动作 | 战略投资方和顾问委员会 | 引介、设计合作、渠道支持 | 中高 | 自然需求弱于合作伙伴辅助需求 | 中高 | 按独立渠道与合作伙伴渠道跟踪来源管线 | 中高 |
| 获取电力的替代路径 | 容量采购、现场发电、定制电价 | 客户痛点的替代解决方案 | 中 | 客户不用 Emerald 软件,也能解决快速接电问题 | 中高 | 证明经济性更优、复杂度更低 | 中高 |
目前几项优势——NVIDIA、电力公司、战略资本——同时也是最大的集中点。
[CR019, CR020, CR021, CR023, CR045]Emerald 残余风险最高的部分,集中在监管经济性、生态依赖和生产就绪度证明。
序数单元格概括有证据支撑的风险排序,而不是公司提供的评分模型。
[CR035, CR036, CR038, CR037, CR028]Emerald 的商业化路径依赖一条关系链:计算栈、公用事业公司、业主、监管方和客户。
[CR019, CR020, CR023, CR021, CR045]7.3 人员、执行与财务模型风险
Emerald 试图把大量执行压进很短时间:2024 年成立,到同行评审演示、电力公司试点、全球合作、商业规模标杆 公告,再到 August 2026 完成独角兽级 Series A 轮。技术团队和合作伙伴阵容能缓解风险,但不能消灭风险。 公司可以拥有顶尖研究员,却仍在现场支持、实施、安全运营或商业复制上失手。公开材料对演示背后的规模化 运营组织说得相对少。 财务模型风险又叠加在执行故事上。收入、利润率、客户集中度和现金跑道都未披露,投资者无法判断 Emerald 正走向 高毛利控制平台,还是部署周期很长、服务比重很高的集成商。私人公司有一定不透明度很正常;问题在于,高预期 现在也一起到来。独立 NERC 摘要和警报还暗示,Emerald 的可服务市场会继续在电网压力下演变,而不是落入稳定 规则集。管住风险的正确方式,是明确终止条件和尽调门槛,而不是模糊乐观。Emerald 已有足够外部验证, 说明风险组合可管理;前提是投资判断必须继续严守监管经济性、生产就绪度和生态集中度。[CR023, CR024, CR025, CR026, CR027, CR028]
| 角色 / 职能 | 依赖或缺口 | 可能性 | 严重性 | 缓释措施 | 尽调路径 |
|---|---|---|---|---|---|
| 创始人 / CEO | Varun Sivaram 撑起政策、融资和商业叙事 | 中 | 高 | 投资人联盟强,技术班底扎实 | 复核接班梯队深度和已下放的运营权责 |
| 技术班底 | 研究团队精英化,但面向现场规模的可靠性组织不够清晰 | 中 | 中高 | 团队半数拥有 PhD,论文积累较深 | 要求提供实施与可靠性组织架构图 |
| 商业运营 | 不到两年内从演示走向多区域部署 | 高 | 高 | 战略董事会和合作伙伴资源 | 要求提供管线阶段、人员计划和部署节奏 |
| 安全 / 合规职能 | 公开政策已经存在,但运营成熟度不透明 | 中 | 高 | 基础政策框架可见 | 要求提供安全负责人、控制措施和审计节奏 |
执行风险的核心不是团队够不够聪明,而是组织宽度能否支撑安全扩张。
[CR024, CR025, CR026, CR027]| 风险 | 可监控触发因素 | 阈值 / 事件 | 行动含义 |
|---|---|---|---|
| 监管经济性错配 | 灵活负载电价机制扩散,但补偿仍然偏弱 | 目标市场要求限电 / 抵押品,却没有清晰客户价值 | 暂停把快速商业化计入估值上行 |
| 可移植性风险 | 仍缺少非 NVIDIA 或非电力公司主导的证明 | 下一轮融资周期前没有可信可移植性证据 | 将生态依赖视为结构性问题,而非过渡问题 |
| 安全 / 可靠性风险 | 与控制层相关的公开事故、SLA 失败或重大中断 | 任何重要客户可见事件 | 升级为红色尽调,并要求事故复盘 |
| 客户集中风险 | 预期 ARR 过多绑定一两个旗舰站点 | 前两大站点或伙伴主导预期收入 | 下修商业规模假设 |
| 执行风险 | 实施积压增长快于已上线的经常性部署 | 服务负担或支持需求超过组织容量 | 下调利润率预期,并拉长规模化时间 |
| 财务不透明风险 | Series A 后烧钱速度、现金跑道和合同经济性仍未披露 | 核心经济性没有披露,也不给尽调访问 | 将其视为新资金投资逻辑的阻断点 |
终止标准必须可衡量,这样才能改变投资决策,而不只是改变备忘录语气。
[CR036, CR037, CR038, CR039, CR040, CR041]7.4 图表
08估值
8.1 投资逻辑与反向逻辑
Emerald AI 的投资逻辑成立。公司瞄准的是真瓶颈——AI 数据中心取电——并用控制层产品解决,成本和速度都可能 优于等待新电网基础设施。市场顺风不是猜想;多家独立来源都把电力描述为数据中心扩张的闸门。以公司年龄看, Emerald 的早期验证也异常具体,包括点名演示、电力公司背书的部署和标杆生态合作。这让它远比典型的收入前 气候软件概念更值得投资人认真看。 反向逻辑在于,公司质量好不等于价格好。灵活负载价值可能很难截留,客户经济性可能取决于费率和电力公司配合; Emerald 也还没有披露投资者需要的财务证据,无法把战略前景和持久商业质量分开。围绕小生态的集中度同样重要, 因为当前轮次价格已经假设 Emerald 能把早期验证转成可复制规模。因此,这个投资逻辑实质上有吸引力,但决策上 仍高度看价格。[CV004, CV005, CV006, CV007, CV008, CV009]
| 字段 | 评估 | 决策含义 |
|---|---|---|
| 建议 | 观察 | 保持主动尽调;仅凭公开证据,不按当前价格承诺投资 |
| 信心 | 中 | 市场需求和证明都真实存在,但经济性披露仍不足 |
| 风险评级 | 高 | 监管经济性、集中度和财务不透明仍是实质问题 |
| 估值立场 | 偏高 | 本轮已假设未来收入能有实质转化 |
| 最可能退出路径 | 战略收购或更晚 IPO | 需要比当前公开证据清晰得多的收入质量和持续性 |
| 促成买入的条件 | 合同价值披露 + 重复部署 + 利润率可见度 | 缺少这些,估值仍过度依赖假设 |
这张表刻意保持价格敏感,而不是给公司质量打泛化分数。
[CV010, CV011, CV012, CV013, CV036, CV034]| 论点 | 支撑 | 什么会改变判断 |
|---|---|---|
| 电力瓶颈真实存在,且在恶化 | Berkeley Lab、CFR 和 PJM 相关来源都指向同一件事:电力正在卡住 AI 扩张 | 如果电力稀缺比预期更快缓解,紧迫性和定价权都会下降 |
| Emerald 的早期证明强于常规水平 | 有具名电力公司和旗舰部署证据 | 需要证明试点会转成经常性付费项目 |
| 快速接电可以支撑软件溢价 | 避免延误可能比普通软件 ROI 更重要 | 需要披露合同来确认价值捕获 |
| 灵活负载经济性可能持续偏薄 | 电力公司和客户可能分不到足够价值 | 如果披露与电价挂钩的客户经济性,判断会改善 |
| 当前估值跑在公开经济性前面 | 没有公开收入、利润率或现金跑道支撑 | 如果披露真实财务数据,或入场价格降低,判断会改善 |
| 生态集中度尚未解决 | 与 NVIDIA 和电力公司相关的集中度仍然可见 | 如果可移植性更广、独立需求更强,判断会改善 |
这些论点的写法,是为了说明哪些证据真的会推动建议变化。
[CV004, CV006, CV005, CV007, CV008, CV009]8.2 估值背景与入场纪律
当前融资背景既亮眼,也让人不舒服。Emerald 在 August 2026 以 $1.05 billion 估值完成 $150 million Series A 轮,随附 Form D 显示截至 early August 该轮仍在配售。这是投资人需求的强信号。但公开市场投资者不应把需求 误当成估值证明。Emerald 尚未披露收入、ARR、毛利率或现金跑道,也没有公布股权结构表或优先股堆叠细节。因此, 当前估值还无法作为「现有销售额」故事来辩护。它是一份前瞻期权,押注市场领导地位、生态控制权和商业化快于预期。 有纪律的读法,是把价格倒过来算。即便按 8x 到 12x 销售收入倍数——高质量基础设施或电力转型平台在公开市场 能拿到这个区间,已经很慷慨——Emerald 也需要大约 $88 million 到 $131 million 年收入,才能支撑当前估值。 长期看并非不可能,但公开证据今天还没有证明。因此,以 $1.05 billion 入场,需要相信强劲的未来转化,而不只是 欣赏团队或赛道。[CV001, CV002, CV003, CV016, CV014, CV015]
只有 Emerald 做到远高于当前公开证据的收入规模,当前轮估值才更容易站住脚。
柱状图展示的是基于公开可比公司倍数的启发式估值结果,不是管理层指引。
[CV014, CV023, CV013]8.3 可比公司分析与情景框架
公开可比公司组合同时给出两点有用信号。第一,AI 电力和数字基础设施生态可以支撑健康估值倍数:Bloom、Equinix、 Digital Realty、Vertiv 和 Eaton 都能以可观的销售收入倍数交易,因为投资者奖励稀缺基础设施、电气化敞口和 AI 带动的增长。第二,这些公司披露的是数十亿美元收入,经营历史也比 Emerald 丰富得多。真正的启示不是 Emerald 今天就该按它们的平均倍数交易;而是如果 Emerald 执行到位,这些倍数提供了未来支撑的天花板和讨论语言。 由此自然进入情景法。乐观情景假设 Emerald 成为能跨多个站点和客户类型复制的控制层,拿到溢价倍数,并把隐含价值 推向数亿美元高位甚至更高。基准情景假设公司成功商业化,但节奏更慢,使当前轮次大致处于满估到偏高之间。悲观 情景假设战略重要性没有转化为广泛价值截留,最终证明当前价格过高。仅看公开证据,概率加权结果低于当前轮次, 这也是建议停在观察、而不是买入的原因。[CV018, CV019, CV020, CV021, CV022, CV023]
| 情景 | 假设 | 估值 / 回报逻辑 | 关键风险 | 概率信号 |
|---|---|---|---|---|
| 乐观 | 重复付费部署、可移植性扩大、电价经济性强,2028 年收入 ~160-240M | 10-12x 销售额 => ~1.6-2.9B;支撑相对本轮的上行 | 仍取决于集中度和执行 | 有可能,但需要多件事同时跑顺 |
| 基准 | 商业化继续推进,但收入扩张更慢,2028 年收入 ~70-110M | 6-8x 销售额 => ~420-880M;本轮估值显得充分到偏贵 | 价值捕获和披露仍不完整 | 公开证据下最合理 |
| 悲观 | 价值捕获弱、集中度高,2028 年收入 ~20-45M | 3-5x 销售额 => ~60-225M;相对本轮有较大下行 | 电价经济性或可移植性失败 | 如果早期证明不能持续叠加,这一情景就成立 |
| 入场纪律叠加层 | 本轮需要接近基准上沿或乐观情景的结果 | 没有更好披露,上行更像期权,难以真正承销 | 优先股堆叠可能进一步压低回报 | 当前价格要求更多证据 |
| 加权观点 | 如果没有重大去风险事件,公开证据指向的价值低于本轮 | 概率加权价值低于 1.05B | 披露和集中度是主要摇摆因素 | 支撑观察,而不是买入 |
情景区间是基于公开证据的启发式范围,不是管理层预测。
[CV025, CV026, CV027, CV028, CV029, CV030]| 可比对象 | 指标 | 倍数 / 估值 / 状态 | 相关性 | 局限 |
|---|---|---|---|---|
| Equinix | 2025 年收入 ~9.22B;市值 ~106.5B | ~10.8x P/S(市销率) | 高溢价数字基础设施 / 数据中心平台 | 成熟度和多元化高得多 |
| Digital Realty | 2025 年收入 ~6.11B;市值 ~73.1B | ~10.8x P/S(市销率) | 数据中心业主 / 互联与电力接入可比公司 | REIT 经济模型不同于软件控制层 |
| Vertiv | 2025 年收入 ~10.23B;市值 ~101.6B | ~8.8x P/S(市销率) | AI 电力 / 散热 / 基础设施受益可比公司 | 硬件和服务敞口不同于 Emerald |
| Bloom Energy | 2025 年收入 ~2.02B;市值 ~64.3B | ~20.6x P/S(市销率) | 电力瓶颈受益标的,带战略叙事溢价 | 硬件 / 项目属性和诉讼噪音不同 |
| Eaton | 2025 年收入 ~27.45B;市值 ~162.9B | ~5.4x P/S(市销率) | 相邻电气化和电力基础设施可比公司 | 大型多元化在位企业,不是风险投资阶段的专业公司 |
这些可比公司用于框定未来可支撑区间,并不是说今天可以直接类比。
[CV018, CV019, CV022, CV020, CV021, CV023]公开证据支撑一个很宽的区间,概率加权中心仍低于当前轮。
情景区间是基于证据的启发式判断,用于组织 IC 讨论;不是 DCF,也不是正式公平性意见。
[CV028, CV029, CV030, CV031, CV001]8.4 决策、触发因素与最终尽调问题
投委会信息应很直接:Emerald 应该进观察名单,而不是回避名单;但当前估值需要纪律。公司已有足够外部验证, 投资者应继续推进这个案子。它没有足够公开财务披露来支撑在本轮高置信度买入。这个区别很重要。许多优秀私人公司 在估值跑赢证据后,会变成糟糕投资;Emerald 今天已接近这条线。 什么会改变判断?正向重估触发因素包括披露合同金额、重复付费部署证据、超出当前生态的更广泛可移植性,以及更清晰 的利润率或现金跑道支撑。负向触发因素包括费率经济性弱、集中度高、无法取得财务资料,或发生安全 / 可靠性事件。 下一步最好不是围绕 AI 和电力做哲学辩论,而是围绕合同、股权结构表、利润率、集中度和现金跑道做定向尽调。如果 尽调通过,Emerald 可以支撑更积极的立场;如果不能,正确答案仍是耐心。[CV010, CV011, CV012, CV013, CV032, CV033]
| 触发因素 | 阈值 | 对投资逻辑的传导 | 行动含义 |
|---|---|---|---|
| 电价经济性失败 | 灵活负载项目给客户的价值弱,或限电 / 抵押品要求过重 | 削弱付费意愿,并拖慢转化 | 除非价格重置,否则转向回避 |
| 重复部署证明停滞 | 没有出现可信的多站点重复付费项目 | 削弱乐观和基准商业化假设 | 下调远期倍数支撑 |
| 集中度过高 | 一两个站点 / 伙伴主导预期 ARR | 压缩收入质量和退出吸引力 | 投资前要求披露集中度 |
| 财务访问仍受阻 | 尽调拿不到收入 / 利润率 / 现金跑道 | 即便投资逻辑仍吸引人,信心也应下降 | 不要承销本轮 |
| 安全或可靠性事故 | 与控制层相关的重大客户可见故障 | 损害信任和溢价倍数支撑 | 在复核完成前升级为回避 |
| 可移植性仍然狭窄 | 没有非核心生态证明 | 护城河看起来更弱,渠道依赖更强 | 降低估值容忍度 |
这张表用于支撑投 / 不投决策,而不是做叙事描述。
[CV035, CV032, CV033, CV034]| 主题 | 缺失证据 | 重要性 | 负责人 / 尽调路径 |
|---|---|---|---|
| 合同 | 按主要账户拆分的 ACV、期限、定价基础、续约权 | 检验收入质量最快 | 要求提供头部客户合同包 |
| 利润率 | 毛利率、服务组合、实施负担 | 区分软件经济性和服务拖累 | 要求提供按合同类型拆分的 P&L 视图 |
| 股权结构表 | 优先股堆叠、优先级、清算条款、老股交易 | 真实回报计算需要这些信息 | 要求提供完整资本结构表 |
| 集中度 | 按站点、电力公司和伙伴渠道拆分的 ARR 与管线 | 检验叙事是否比少数旗舰项目更宽 | 要求提供集中度明细 |
| 现金跑道 | 现金、烧钱速度、招聘计划、情景现金跑道 | 检验去风险前是否存在时间压力 | 要求提供董事会或财务计划 |
| 可移植性 | 非 NVIDIA、非核心电力公司的商业证明 | 检验生态依赖是否只是过渡 | 要求提供部署路线图和已签署证明 |
这些尽调要求按改变建议或估值立场的速度排序。
[CV038, CV039, CV040, CV041, CV042, CV034]Emerald 在公司质量上得分不错,但公开经济性还不足以支撑买入建议。
[CV004, CV006, CV008, CV013, CV010]市场和验证得分较好;经济性和估值支撑落后。
[CV043, CV044, CV045, CV046, CV012, CV047]8.5 图表
免责声明
本报告仅供参考,不构成投资建议。
证据索引
| 编号 | 陈述 | 可信度 | 来源 |
|---|---|---|---|
| CO001 | Emerald AI is a Washington, DC-based software company focused on making AI data centers power-flexible grid assets. | 高 | SO001, SO003, SO019 |
| CO002 | Official company materials list Washington DC as the primary location and Boston and San Francisco as additional office locations. | 高 | SO003, SO019 |
| CO003 | Emerald AI was founded in November 2024 and SEC filings identify 2024 as its year of incorporation. | 高 | SO012, SO015 |
| CO004 | Emerald AI, Inc. is a Delaware corporation with a business address at 4535 Westhall Drive NW, Washington, DC 20007. | 高 | SO015, SO016, SO017 |
| CO005 | Founder and CEO Dr. Varun Sivaram previously served as Chief Strategy and Innovation Officer at Orsted and Chief Technology Officer at ReNew Power. | 高 | SO004, SO019 |
| CO006 | Varun Sivaram also served as Managing Director for Clean Energy at the U.S. State Department and is a senior fellow for energy at the Council on Foreign Relations. | 中 | SO004 |
| CO007 | Emerald AI's flagship product is the Emerald Conductor platform, which orchestrates AI workloads and onsite energy resources to control facility power draw in real time. | 高 | SO001, SO009, SO019 |
| CO008 | The company positions Emerald Conductor as infrastructure that lets data centers respond to grid stress without compromising critical AI workloads. | 高 | SO001, SO009, SO021 |
| CO009 | On 25 August 2026 Emerald AI announced a $150 million oversubscribed Series A financing at a $1.05 billion valuation. | 高 | SO009, SO017 |
| CO010 | The Series A was co-led by Energize Capital and DCVC. | 高 | SO009, SO010 |
| CO011 | The Series A syndicate included NVIDIA, Samsung Ventures, Siemens, Aramco Ventures, Salesforce Ventures, GE Vernova, RWE, JERA Ventures, ADVentures, IQT, Radical Ventures, Energy Impact Partners, Lowercarbon Capital, John Doerr, Tom Steyer, Earthshot Ventures, Collective Global and General Catalyst's scout fund. | 高 | SO009, SO010 |
| CO012 | Emerald AI says twelve Fortune Global 500 companies are now investors and sit on its Strategic Advisory Board. | 中 | SO009, SO011 |
| CO013 | Emerald AI launched from stealth in July 2025 with a disclosed $24.5 million seed round led by Radical Ventures. | 中 | SO014, SO022 |
| CO014 | Emerald AI's August 2025 Form D disclosed a $35.3 million offering with $34.17 million sold at filing time. | 中 | SO015 |
| CO015 | A February 2026 Form D disclosed a $24.9996 million offering with $22.7496 million sold at filing time. | 高 | SO016, SO018 |
| CO016 | Emerald AI later announced an $18 million seed extension that brought total disclosed funding to $42.5 million. | 中 | SO012 |
| CO017 | Emerald AI subsequently announced a $25 million strategic expansion round that brought total funding to $68 million before the Series A. | 中 | SO011 |
| CO018 | Adding the announced $150 million Series A to the previously announced $68 million implies roughly $218 million of disclosed financing by the August 2026 run date. | 中 | SO009, SO011, SO014 |
| CO019 | The August 2026 Form D listed a $150 million offering with about $90.23 million sold at filing time, showing the round was still being closed when filed. | 高 | SO017, SO009 |
| CO020 | Emerald AI's board page names John Tough of Energize Capital as a director and David Katz of Radical Ventures, Zachary Bogue of DCVC, Christina Buchanan of NVentures, Clay Dumas of Lowercarbon Capital, and Shayle Kann of Frontier Fund/Energy Impact Partners as board observers. | 中 | SO002 |
| CO021 | The company also lists Gina Raimondo, David Rousseau, Jason Bordoff, Arushi Sharma Frank, Jesse Jenkins, Sean Kelly, Anna Patterson, Gaurav Desai, Richard Stuebi and Peter Hans Hirschboeck among its advisors. | 中 | SO002 |
| CO022 | Chief Scientist Ayse Coskun is a Boston University professor and an academic pioneer in flexible AI and high-performance computing for grid response. | 高 | SO005, SO019 |
| CO023 | Head of Engineering Shayan Sengupta joined from AWS, where he led specialized AI, HPC and Mac compute engineering teams supporting hundreds of millions of dollars in revenue. | 高 | SO006, SO019 |
| CO024 | Chief Commercial Officer Aroon Vijaykar previously led Sunrun's virtual power plant, distribution and manufacturing businesses and earlier served as CEO of AEE Solar. | 中 | SO007 |
| CO025 | Head of Product Mansi Shah previously served as a chief technologist at VMware focused on enterprise data products and distributed systems. | 中 | SO008 |
| CO026 | Emerald AI reported in 2026 that roughly half of its team are PhDs and that the team has produced more than 400 technical publications. | 中 | SO012 |
| CO027 | By August 2026 Emerald AI said it had completed five global demonstrations and moved into commercial deployments at multi-megawatt, full-data-center scale. | 中 | SO009, SO011 |
| CO028 | Those live demonstrations spanned Arizona, Illinois, Virginia, Oregon and London according to Emerald's 2026 funding and recognition posts. | 中 | SO009, SO027 |
| CO029 | Emerald AI's first Phoenix demonstration cut power use by 25% for three hours on a 256-GPU AI cluster while preserving acceptable workload performance. | 中 | SO022, SO021 |
| CO030 | National Grid and partners said a UK trial on a 96-NVIDIA-Blackwell-Ultra cluster cut electricity demand by more than a third in under a minute and by up to 40% while critical workloads continued. | 中 | SO026 |
| CO031 | Silicon Valley Power and Emerald AI launched a pilot in Santa Clara that offers expanded grid access to a participating data center in exchange for verifiable flexibility. | 中 | SO025 |
| CO032 | Emerald AI, Digital Realty, NVIDIA, EPRI and PJM are using the Aurora facility in Manassas, Virginia as the reference commercial-scale power-flexible AI factory, sized at roughly 96 MW. | 高 | SO013, SO024 |
| CO033 | TIME named Emerald AI one of its 2026 Most Influential Companies, and the World Economic Forum selected the company as a 2026 Technology Pioneer. | 高 | SO020, SO027 |
| CO034 | Official materials say Emerald AI serves customers across the AI power value chain, including leading AI firms, data center operators and electric utilities, but the company does not disclose a customer count. | 中 | SO009, SO001 |
| CO035 | Heatmap reported that Emerald's economic case still depends on utilities offering faster interconnection or other meaningful incentives and on hyperscalers accepting some degree of curtailment. | 中 | SO023 |
| CO036 | S&P Global noted that making data centers flexible can unlock power capacity, but operators have historically resisted curtailment because it is operationally tricky and risky. | 中 | SO024 |
| CO037 | Public evidence still does not disclose Emerald AI's revenue, audited margins, exact headcount, or detailed investor control rights, leaving material underwriting gaps at the new unicorn valuation. | 中 | SO009, SO002, SO019 |
| CM001 | IEA expects U.S. electricity demand to grow by nearly 2% annually through 2030, with roughly half of the increase driven by data centers. | 中 | SM001 |
| CM002 | IEA forecasts global electricity demand growth of 3.6% per year from 2026 through 2030. | 中 | SM001 |
| CM003 | Berkeley Lab said U.S. data center electricity consumption could rise from 176 TWh in 2023 to 325-580 TWh by 2028. | 中 | SM002 |
| CM004 | Berkeley Lab said data centers represented about 4.4% of U.S. electricity use in 2023 and could reach 6.7%-12% by 2028. | 中 | SM002 |
| CM005 | JLL projects about 97 GW of new global data center capacity between 2026 and 2030, effectively doubling the sector. | 中 | SM004 |
| CM006 | JLL frames the global data center sector at a 14% supply CAGR through 2030 in its base case. | 中 | SM004 |
| CM007 | JLL expects AI to represent about half of all data center workloads by 2030. | 中 | SM004 |
| CM008 | CBRE says power availability and grid infrastructure constraints are reshaping development timelines and site selection in major hubs. | 高 | SM005, SM006 |
| CM009 | CBRE says the ability to deliver 300 MW or more within 36 months is overtaking connectivity and power price as the key location criterion for many large deployments. | 中 | SM006 |
| CM010 | Northern Virginia vacancy fell to 0.3% in CBRE’s Q1 2026 global trends report despite 1.1 GW of year-over-year inventory growth. | 中 | SM005 |
| CM011 | Bloom cites estimates that U.S. IT load could roughly double from about 80 GW in 2025 to about 150 GW in 2028. | 中 | SM007 |
| CM012 | Bloom found that more than one-third of data centers are expected to use 100% onsite power by 2030. | 中 | SM007 |
| CM013 | Bloom reported that 73% of respondents were actively evaluating or selecting onsite power providers. | 中 | SM007 |
| CM014 | FERC’s June 2026 show-cause orders explicitly called for new transmission services for flexible large loads. | 高 | SM008, SM017 |
| CM015 | FERC grouped its large-load reforms into five categories, including cost transparency, co-location rules, and flexible-load services. | 中 | SM008 |
| CM016 | PNNL said no states had fully developed DER aggregator and distribution coordination frameworks as of early 2026 under Order 2222 implementation. | 中 | SM009 |
| CM017 | SEPA said its DELTa tracker covered 104 approved or pending large-load tariffs and service rules across more than 70 utilities in 37 states by July 2026. | 中 | SM010 |
| CM018 | SEPA said about one-quarter of tracked large-load tariffs included a concrete option for dispatchable large-load flexibility or another curtailment pathway. | 中 | SM010 |
| CM019 | Berkeley Lab’s August 2026 rate-design update analyzed a sample of 55 large-load tariffs, contracts, and related frameworks. | 中 | SM003 |
| CM020 | Berkeley Lab reported that the median minimum demand threshold across reviewed large-load tariffs was 25 MW and that 75% fell between 5 MW and 100 MW. | 中 | SM003 |
| CM021 | The Duke/CFR framing argues that roughly 100 GW of new U.S. data center demand could connect sooner if facilities accept limited curtailment. | 中 | SM011 |
| CM022 | The CFR summary says the flexibility requirement in the Duke study was about 25% curtailment for fewer than 200 hours per year while preserving 99.5% of annual energy use. | 中 | SM011 |
| CM023 | Utility Dive said flexible data center loads could address about 10% of the nation’s current aggregate peak demand if curtailed for 0.25% of maximum uptime. | 中 | SM012, SM011 |
| CM024 | National Grid said its UK flexibility trial showed data centers could cut power demand by more than one-third in under a minute and by up to 40% while critical workloads continued. | 中 | SM024 |
| CM025 | Utility Dive reported that Emerald AI’s Arizona demonstration achieved a 25% reduction in power consumption over three hours on a 256-GPU cluster. | 中 | SM012, SM026 |
| CM026 | IEEE Spectrum reported that the first DCFlex sites were designed to test distinct flexibility methods across Google, Nvidia/Oracle, and Data4 facilities. | 中 | SM016, SM015 |
| CM027 | IEEE Spectrum said DCFlex had 45 collaborators by mid-2025 and expected up to 10 sites that year, signaling ecosystem demand for flexibility pilots. | 中 | SM016 |
| CM028 | Emerald AI’s addressable market is not total data center capex; it is the control and market-integration layer that converts large AI loads into dispatchable or schedulable grid assets. | 高 | SM018, SM025, SM019 |
| CM029 | Included spend therefore covers workload-orchestration software, telemetry and verification, integration into utility or grid programs, and potentially recurring availability or performance fees. | 中 | SM018, SM019, SM010 |
| CM030 | Excluded spend includes GPUs, shells, substations, generation plants, and generic colocation rent unless Emerald AI directly monetizes them through the flexibility layer. | 中 | SM004, SM005, SM018 |
| CM031 | The status quo alternatives are waiting for firm interconnection, self-supplying with onsite power, relying on bespoke utility contracts without specialist software, or moving workloads to less-constrained regions. | 中 | SM007, SM005, SM013 |
| CM032 | Hyperscalers, neoclouds, and large colocation operators are early direct buyers because they control siting speed, workload placement, and uptime tradeoffs. | 中 | SM004, SM005, SM026 |
| CM033 | Utilities and grid operators are also economic sponsors because flexible-load programs, tariffs, and bespoke agreements determine whether faster interconnection creates monetary value. | 中 | SM010, SM008, SM013 |
| CM034 | The near-term budget owner is likely a combination of data center energy strategy teams and utility large-load planning or innovation teams, not a standard IT software buyer. | 中 | SM006, SM026, SM012 |
| CM035 | The primary adoption trigger is speed-to-power: buyers adopt flexibility when it yields faster interconnection, lower grid costs, or both. | 高 | SM008, SM010, SM006, SM007 |
| CM036 | The strongest macro driver is that power scarcity has become a first-order constraint on AI infrastructure growth. | 高 | SM004, SM005, SM007, SM001 |
| CM037 | A second driver is regulatory experimentation around large-load tariffs, flexible service classes, and faster non-firm connection structures. | 高 | SM008, SM010, SM003 |
| CM038 | A third driver is the emergence of field proof that some AI workloads can be scheduled or curtailed without shutting down critical services. | 高 | SM024, SM012, SM020 |
| CM039 | A major constraint is operator conservatism: many buyers still prefer no flex at all because uptime promises remain commercially sacred. | 中 | SM011, SM023, SM016 |
| CM040 | A second constraint is fragmented market design, because adoption depends on utility-by-utility tariffs, state policy, and local implementation rather than a single national program. | 中 | SM010, SM009, SM013 |
| CM041 | A third constraint is monetization uncertainty: public evidence supports a multi-GW opportunity but not a clean public software-dollar TAM for Emerald AI. | 中 | SM004, SM011, SM023 |
| CM042 | Heatmap preserved the key adverse thesis: flexibility only clears commercially if utilities provide meaningful interconnection advantage or compensation. | 中 | SM023 |
| CM043 | PJM’s proposed IRAS framework would treat 50 MW+ sites as new large loads and curtail uncovered demand before broader emergency measures. | 中 | SM013, SM014 |
| CM044 | POWER Magazine said PJM’s August 2026 filing tied about 30 GW of projected 2024-2030 peak-demand growth to data centers. | 中 | SM014 |
| CM045 | Bloom expects power constraints to reallocate U.S. growth toward power-advantaged regions such as Texas and the Southeast while legacy markets lose relative share. | 中 | SM007, SM005 |
| CM046 | Emerald AI’s pragmatic SAM is North America plus the UK markets where utilities, grid operators, and large-load customers are already testing flexibility pathways. | 中 | SM024, SM010, SM013, SM026 |
| CM047 | Emerald AI’s near-term SOM is better described as a handful of flagship campuses and utility-backed pilots than as a broad installed-base rollout. | 中 | SM026, SM016, SM015 |
| CM048 | A conservative U.S. flexible-interconnection lens is about 25 GW, representing only a quarter of the Duke/CFR 100 GW opportunity becoming commercially addressable in the near term. | 低 | SM011, SM023 |
| CM049 | A base-case U.S. flexible-interconnection lens is about 50 GW, assuming partial but material commercialization of the Duke/CFR headroom thesis in the most constrained markets. | 低 | SM011, SM004, SM005 |
| CM050 | A high-case U.S. flexible-interconnection lens is about 100 GW, matching the full Duke/CFR near-term headroom argument if policy and operational proof converge. | 中 | SM011 |
| CM051 | Converting GW opportunity into software revenue still requires private evidence on contract structure, pricing basis, utility cost-sharing, and realized performance payments. | 中 | SM023, SM026, SM018 |
| CP001 | Emerald AI is explicitly positioned around power-flexible AI data centers rather than generic DER or building loads. | 高 | SP001, SP002, SP023 |
| CP002 | Voltus serves commercial, industrial, and residential energy users across all nine wholesale power markets in the U.S. and Canada. | 中 | SP003 |
| CP003 | Voltus publishes gross earnings examples reaching as high as $350,000/MW-year in PJM and $470,000/MW-year in ISO-NE. | 中 | SP003 |
| CP004 | CPower positions itself as a broad C&I virtual power plant platform rather than a data-center-specific orchestration vendor. | 中 | SP004 |
| CP005 | CPower says NRG Energy has acquired CPower, giving it backing from a larger energy platform. | 中 | SP004 |
| CP006 | Virtual Peaker is utility-first software focused on launching and managing demand response and DER programs across residential, commercial, and industrial segments. | 中 | SP005 |
| CP007 | EnergyHub’s public proof is strongest in utility demand flexibility and DERMS programs rather than in hyperscale data center orchestration. | 中 | SP006 |
| CP008 | Leap competes as a market-access and revenue platform for distributed energy resources and virtual power plants. | 中 | SP007 |
| CP009 | Amperon is primarily a forecasting and analytics competitor rather than a direct dispatch-and-control replacement for Emerald AI. | 中 | SP008 |
| CP010 | GridPoint competes through commercial-building optimization and grid-interactive load management, not AI-cluster workload control. | 中 | SP009 |
| CP011 | Uplight combines customer engagement, rate engagement, and demand management across utilities and customers with 8.5 GW under management. | 中 | SP010 |
| CP012 | Enel North America sells integrated clean energy and flexibility solutions to corporate, industrial, utility, and city buyers. | 中 | SP012 |
| CP013 | Itron competes higher in the utility grid-management stack, making it more of an incumbent platform or partner than a direct application-layer peer. | 中 | SP011 |
| CP014 | Bloom’s data center power report highlights onsite power as a substitute path that can reduce the urgency of software-only flexibility in some campuses. | 中 | SP013 |
| CP015 | A buyer can address the same problem through onsite power, power-advantaged relocation, bespoke utility agreements, or internal workload scheduling without buying Emerald AI. | 中 | SP013, SP014, SP015, SP018 |
| CP016 | SEPA and FERC show that utilities and regulators are only beginning to create formal pathways for flexible large loads. | 高 | SP016, SP017 |
| CP017 | Emerald has fresher public proof in live data center pilots than most generic DER software incumbents because its public record includes SVP, National Grid, Phoenix, and NVIDIA-linked evidence. | 高 | SP024, SP025, SP020, SP026 |
| CP018 | Voltus, CPower, Uplight, EnergyHub, and Enel all have stronger pre-existing utility or energy-buyer distribution than Emerald AI. | 中 | SP003, SP004, SP010, SP006, SP012 |
| CP019 | Emerald’s differentiation is not broad VPP scale but a narrow specialization around AI workload flexibility under data center power constraints. | 高 | SP001, SP002, SP020 |
| CP020 | Public pricing is opaque across Emerald AI and most peers; Voltus is the clearest outlier because it publishes gross MW-year earning examples instead of software list prices. | 中 | SP003, SP001, SP004, SP005, SP007 |
| CP021 | The landscape mixes revenue-share aggregators, utility SaaS platforms, consulting-heavy solution sales, and broader energy-service bundles rather than one standard contract model. | 中 | SP003, SP004, SP005, SP010, SP012, SP007 |
| CP022 | Emerald is the clearest vendor in this source set making AI-workload choreography a headline capability rather than a side effect of generic DR software. | 高 | SP002, SP020, SP022 |
| CP023 | Virtual Peaker, EnergyHub, Uplight, and Itron show stronger utility-program and DERMS heritage than Emerald AI. | 中 | SP005, SP006, SP010, SP011 |
| CP024 | Voltus, CPower, and Leap show stronger market-participation and enrollment infrastructure than Emerald AI based on public surfaces. | 中 | SP003, SP004, SP007 |
| CP025 | Emerald, Bloom, CBRE, and JLL collectively suggest that the relevant buyer problem is power-constrained data center delivery, a job that most demand-response incumbents were not built around. | 高 | SP001, SP013, SP014, SP015 |
| CP026 | Switching costs become meaningful once a customer has utility relationships, telemetry, policy controls, and operating procedures integrated into a flexibility workflow. | 中 | SP016, SP024, SP025, SP005 |
| CP027 | Multi-homing is plausible because Emerald can coexist with utility DR software, forecasting vendors, or onsite-power providers instead of fully replacing them. | 中 | SP008, SP010, SP009, SP013, SP024 |
| CP028 | Large public data center operators such as Equinix and Digital Realty are not direct software peers, but their scale makes them likely partners, customer archetypes, or future entrants into flexibility orchestration. | 中 | SP027, SP028, SP014 |
| CP029 | Heatmap preserves the key adverse competitive risk: if utilities fail to attach real speed-to-power or compensation value to flexibility, Emerald’s narrow category may not sustain premium pricing. | 中 | SP019 |
| CP030 | If incumbent DR/VPP vendors adapt their platforms for large loads and pair that with existing utility relationships, Emerald could face pricing pressure before it establishes a moat. | 中 | SP003, SP004, SP005, SP006, SP010 |
| CP031 | Hyperscalers and top colocation developers could internalize parts of workload scheduling or utility coordination, especially if flexibility becomes strategically core. | 中 | SP015, SP014, SP020 |
| CP032 | Onsite generation and bring-your-own-power strategies are the most important non-software substitutes because they solve time-to-power without requiring as much curtailment tolerance. | 中 | SP013, SP014 |
| CP033 | Emerald’s moat claim strengthens materially if it can show repeatable production wins across multiple utilities and campuses rather than a few showcase pilots. | 中 | SP024, SP025, SP020, SP021 |
| CP034 | The most defensible moat candidate is a combination of workload-performance data, utility integration playbooks, and credibility with GPU and grid partners. | 中 | SP026, SP024, SP025, SP022 |
| CP035 | The public record is weak on competitor win rates, renewal, realized pricing, and share of wallet across nearly every vendor in this comparison. | 中 | SP003, SP004, SP005, SP006, SP007, SP008 |
| CP036 | The landscape is best understood as four overlapping categories: direct data-center-flexibility specialists, C&I DR/VPP aggregators, utility flexibility platforms, and substitute power/infrastructure strategies. | 中 | SP001, SP003, SP004, SP005, SP010, SP013 |
| CP037 | Voltus competes hardest when the buyer wants monetization of load flexibility in established wholesale programs rather than AI-specific workload control. | 中 | SP003, SP001, SP016 |
| CP038 | CPower competes hardest where energy-market monetization and enterprise energy management matter more than preserving GPU-workload QoS. | 中 | SP004, SP001 |
| CP039 | Virtual Peaker competes hardest where utilities own the buying decision and want a program-management stack rather than a data-center-specific operating layer. | 中 | SP005, SP001 |
| CP040 | EnergyHub competes hardest where device-network breadth and utility program scale matter more than large-load specialization. | 中 | SP006, SP001 |
| CP041 | Leap competes hardest when a customer already has controllable assets and primarily needs market access and settlement support. | 中 | SP007, SP001 |
| CP042 | Amperon is more complementary than substitutive because forecasting alone does not deliver dispatch or workload choreography. | 中 | SP008, SP001 |
| CP043 | GridPoint is a substitute mainly for commercial buildings and grid-interactive campuses, not for GPU-cluster orchestration. | 中 | SP009, SP001 |
| CP044 | Uplight, EnergyHub, Itron, and Enel have better utility-selling muscle than Emerald, which could matter if utilities standardize flexibility procurement. | 中 | SP010, SP006, SP011, SP012 |
| CP045 | Equinix and Digital Realty also matter competitively because large operators may prefer to embed flexibility in campus design, procurement, or landlord services rather than buy a standalone overlay. | 中 | SP027, SP028, SP015, SP014 |
| CP046 | Named utility and partner proofs from SVP, National Grid, NVIDIA, and Phoenix give Emerald better category storytelling than most peers, even though scale data remains thin. | 高 | SP024, SP025, SP026, SP020 |
| CP047 | The lack of public pricing means the pricing table in this chapter should be read as contract-model comparison, not as apples-to-apples list-price benchmarking. | 中 | SP001, SP003, SP004, SP005 |
| CP048 | Because the market is still forming, many vendors blur partner, substitute, and competitor roles at once. | 中 | SP016, SP017, SP013, SP022 |
| CP049 | Emerald wins the direct-comparison frame only if buyers decide AI-workload flexibility is a distinct problem worth specialized software rather than a feature of existing energy platforms. | 中 | SP001, SP019, SP020, SP005 |
| CI001 | Emerald launched publicly in July 2025 with a disclosed $24.5 million seed round. | 中 | SI003 |
| CI002 | Emerald said in February 2026 that it raised an additional $18 million, bringing total funding to $42.5 million. | 中 | SI004 |
| CI003 | Emerald said in March 2026 that it raised $25 million in a Strategic Expansion Round, bringing total funding to roughly $68 million. | 中 | SI005 |
| CI004 | Emerald announced a $150 million Series A at a $1.05 billion valuation on August 25, 2026. | 高 | SI001, SI002, SI008 |
| CI005 | Across the disclosed seed, extension, strategic expansion, and Series A rounds, Emerald has announced roughly $217.5 million of cumulative capital by August 2026. | 高 | SI003, SI004, SI005, SI001 |
| CI006 | The August 2025 Form D shows a $35.3 million offering amount, $34.17 million sold, and 37 investors. | 中 | SI006 |
| CI007 | The February 2026 Form D shows a $24.9996 million offering amount, $22.75 million sold, and 20 investors. | 中 | SI007 |
| CI008 | The August 2026 Form D shows a $150 million offering amount, $90.23 million sold, and 23 investors as of the filing date. | 中 | SI008 |
| CI009 | The Series A announcement says the new capital will be used to scale commercial deployments worldwide. | 中 | SI001 |
| CI010 | The company says its customers include leading AI firms, data center operators, and electric power utilities. | 高 | SI001, SI009 |
| CI011 | Emerald’s public monetization story is centered on Conductor software and orchestration rather than on owning large physical power assets. | 高 | SI001, SI032, SI018 |
| CI012 | No public list pricing, contract value, or pricing schedule is disclosed in the fetched sources. | 中 | SI001, SI032, SI015 |
| CI013 | The most plausible core revenue stream is enterprise software licensing or subscription tied to workload orchestration and grid-response control. | 中 | SI001, SI009, SI010 |
| CI014 | Early monetization likely also includes implementation and integration work because deployments require coordination with utilities, operators, and site systems. | 中 | SI011, SI012, SI013 |
| CI015 | Strategic investors and partners likely function as a distribution layer that can reduce top-of-funnel friction for early enterprise sales. | 中 | SI002, SI009, SI005 |
| CI016 | The GTM motion is likely slower than standard SaaS because deals require multi-party utility, operator, and infrastructure alignment. | 中 | SI013, SI016, SI017 |
| CI017 | Emerald’s willingness-to-pay wedge is speed-to-power and avoided interconnection delay rather than generic AI software productivity. | 高 | SI017, SI021, SI001 |
| CI018 | Because value depends on local power constraints and the split between utility and operator beneficiaries, contract pricing is likely negotiated rather than list-based. | 中 | SI009, SI013, SI016 |
| CI019 | No public revenue, ARR, GMV, or utilization metric is disclosed across company and third-party sources reviewed here. | 中 | SI001, SI015, SI014 |
| CI020 | Gross margin, contribution margin, and EBITDA are not publicly disclosed. | 中 | SI001, SI015, SI032 |
| CI021 | Monthly burn and cash runway are not publicly disclosed. | 中 | SI001, SI005, SI032 |
| CI022 | No debt facility, project finance structure, or other financing obligation is disclosed in the reviewed public materials. | 中 | SI001, SI005, SI008 |
| CI023 | Public sources do show the business moving from demonstrations toward named commercial deployments in 2026. | 高 | SI011, SI001, SI010 |
| CI024 | Visible commercialization still appears concentrated in a small number of flagship deployments and strategic relationships. | 中 | SI011, SI013, SI012 |
| CI025 | The Strategic Advisory Board and investor coalition likely improve enterprise access even though they do not prove organic standalone demand. | 中 | SI001, SI002, SI005 |
| CI026 | The cadence from seed to extension to strategic expansion to a unicorn Series A in roughly one year indicates unusually strong investor conviction. | 高 | SI003, SI004, SI005, SI001 |
| CI027 | The broader market has become more power-constrained, increasing the urgency of products that promise faster interconnection or flexible load economics. | 高 | SI019, SI020, SI022, SI023 |
| CI028 | Bloom’s 2026 report says 73% of operators are embedding onsite power into long-term strategies and over one-third expect 100% onsite power by 2030. | 中 | SI019, SI020 |
| CI029 | Berkeley Lab said data centers consumed 4.4% of U.S. electricity in 2023 and could reach 6.7% to 12% by 2028. | 高 | SI022, SI023 |
| CI030 | That external power bottleneck supports Emerald’s pricing power in principle because the alternative is often years of delay or more expensive onsite supply. | 高 | SI021, SI020, SI017 |
| CI031 | No public CAC, payback period, or sales-cycle metric exists, so sales efficiency cannot be underwritten directly. | 中 | SI032, SI009, SI015 |
| CI032 | The visible cost structure is likely dominated by engineering talent, site integration, partner support, and enterprise business development rather than commodity hardware. | 中 | SI032, SI010, SI011 |
| CI033 | Emerald appears materially less capital-intensive than developers that must finance generation or full data center buildouts, because its product is software and orchestration. | 高 | SI001, SI021, SI020 |
| CI034 | Even so, Emerald should not be modeled as frictionless horizontal SaaS because deployments are infrastructure-adjacent and site-specific. | 中 | SI011, SI012, SI013 |
| CI035 | Commercialization proof supports relevance, but revenue quality remains early-stage because contract size, recurrence, and churn are undisclosed. | 中 | SI011, SI001, SI012 |
| CI036 | A disclosed $150 million Series A gives Emerald a materially larger capital base than it had in March 2026, but not a disclosed self-funding profile. | 高 | SI005, SI001, SI008 |
| CI037 | If flagship commercial deployments fail to convert into repeatable revenue, the next financing will likely need to arrive before public economics are fully proven. | 中 | SI001, SI013, SI016 |
| CI038 | The core financial blocker is the absence of realized contract values, renewal terms, and deployment-to-revenue conversion data. | 中 | SI001, SI013, SI015 |
| CI039 | The second blocker is the absence of gross-margin and service-delivery-cost evidence. | 中 | SI032, SI001, SI015 |
| CI040 | The third blocker is the absence of burn and runway disclosure despite large recent fundraising. | 中 | SI001, SI005, SI008 |
| CI041 | Adjacent public infrastructure and data-center platforms report 2025 revenue bases in the billions, highlighting how early Emerald still is relative to financially transparent incumbents and enablers. | 中 | SI024, SI025, SI026, SI027, SI028, SI029 |
| CI042 | Public evidence supports a credible monetization path, but not a complete financial underwriting on revenue quality, margins, or capital efficiency. | 高 | SI001, SI011, SI016, SI020 |
| CE001 | Emerald AI’s core product is software that turns AI data centers into dispatchable or schedulable grid assets under power constraints. | 高 | SE003, SE008, SE024 |
| CE002 | Emerald Conductor is the flagship software platform publicly described across Emerald, NVIDIA, utility, and media sources. | 高 | SE003, SE006, SE008 |
| CE003 | Emerald also publicly references a GridLink product that links grid signals and data center controls. | 高 | SE004, SE010 |
| CE004 | Emerald’s current commercial narrative depends heavily on integration with NVIDIA DSX Flex and the broader DSX OS stack. | 高 | SE005, SE006, SE008 |
| CE005 | The public module map is still narrow: Conductor is explicit, GridLink is referenced, and other internal services are not productized publicly by name. | 中 | SE004, SE001, SE008 |
| CE006 | The product is designed for operators who need faster grid access or flexible dispatch without breaking AI workload performance. | 高 | SE003, SE006, SE007, SE017 |
| CE007 | The operating flow begins with a utility or grid signal that defines a target power reduction or flexibility event. | 高 | SE014, SE007, SE006 |
| CE008 | Emerald profiles jobs across flexibility, time sensitivity, and performance tolerance before or during an event. | 中 | SE013, SE014 |
| CE009 | Emerald then models power-reduction scenarios and chooses a control policy that balances grid targets against workload constraints. | 中 | SE013, SE014, SE015 |
| CE010 | Public materials indicate actuation can include DVFS power caps, job pausing, checkpointing, and workload migration or rerouting. | 中 | SE014, SE021, SE015 |
| CE011 | Emerald’s workflow ends with telemetry and verification against target power and workload-performance thresholds. | 中 | SE008, SE007, SE014 |
| CE012 | The Phoenix field demonstration reduced power demand by 25% for three hours on a 256-GPU cluster while staying within SLA constraints. | 高 | SE008, SE011, SE013 |
| CE013 | The UK trial showed up to 40% power reduction in under a minute while critical workloads continued. | 高 | SE007, SE008 |
| CE014 | National Grid said the UK test involved more than 200 simulated grid events over five days. | 高 | SE007, SE008 |
| CE015 | Emerald and partner sources say the platform has completed five live demonstrations at commercial data centers across two continents. | 高 | SE005, SE006, SE008 |
| CE016 | The SVP deployment is framed as the first commercial, multi-megawatt DSX Flex deployment. | 高 | SE005, SE006 |
| CE017 | A 96 MW power-flexible AI factory in Manassas is positioned as a large-scale reference deployment and certification standard. | 高 | SE004, SE010, SE005 |
| CE018 | Public materials say GridLink and Conductor leverage NVIDIA AI Enterprise components, including NIM microservices, with NVIDIA Mission Control. | 高 | SE004, SE010 |
| CE019 | The public architecture resolves into six layers: grid signal intake, power-target shaping, workload profiling, optimization policy, actuation/control, and telemetry/verification. | 中 | SE014, SE015, SE013 |
| CE020 | Key inputs include grid-event timing, target power, workload mix, flexibility scores, and performance thresholds. | 中 | SE014, SE013, SE015 |
| CE021 | The GitHub materials show policy families including DVFS-only, DVFS plus job pausing, and geographically distributed load shifting. | 中 | SE014, SE015 |
| CE022 | Emerald’s public proof repeatedly emphasizes protection of priority or critical workloads as a design constraint. | 高 | SE013, SE008, SE007, SE006 |
| CE023 | The product is highly partner-dependent on NVIDIA hardware/software, utility frameworks, and access to live commercial data center environments. | 高 | SE008, SE006, SE007, SE010 |
| CE024 | Utility or grid-operator participation is a functional dependency because the product’s value emerges when external actors send dispatch or interconnection signals. | 高 | SE006, SE007, SE017 |
| CE025 | Data center operators remain a critical dependency because Emerald must integrate into workload management and operating policies at the site level. | 中 | SE012, SE013, SE008 |
| CE026 | Commercial value also depends on local regulatory or tariff frameworks that reward flexible-load behavior. | 中 | SE005, SE006, SE017 |
| CE027 | The public GitHub repository is meaningful developer signal because it exposes pseudocode, datasets, and orchestration commands instead of pure marketing copy. | 高 | SE014, SE026 |
| CE028 | Nature, arXiv, CFR, and partner evidence collectively suggest Emerald’s product is grounded in an emerging technical field rather than generic energy rhetoric. | 高 | SE016, SE015, SE017, SE026 |
| CE029 | The strongest moat candidate is the combination of workload-flexibility profiling, operating data from live events, and utility/NVIDIA integration playbooks. | 中 | SE013, SE008, SE006, SE007 |
| CE030 | Emerald’s product is narrower than a general DERMS or utility program platform because it directly controls AI workload behavior inside data centers. | 高 | SE001, SE003, SE012 |
| CE031 | Emerald’s public privacy policy says personal information collected through its website is not used to train machine learning or AI models. | 中 | SE019 |
| CE032 | The privacy policy says Emerald maintains technical, administrative, and organizational safeguards to protect personal information, while explicitly warning that no system is perfectly secure. | 中 | SE019 |
| CE033 | The terms page says product and deployment descriptions may contain forward-looking statements and that actual results may differ materially. | 中 | SE018 |
| CE034 | Emerald’s terms explicitly say website content is not engineering, regulatory, legal, financial, or investment advice. | 中 | SE018 |
| CE035 | Emerald’s terms prohibit automated scraping and the use of website content to train or fine-tune AI models. | 中 | SE018 |
| CE036 | No public SOC 2, ISO 27001, model-governance audit, or formal reliability certification is visible in the fetched public record. | 中 | SE001, SE019, SE018, SE020 |
| CE037 | The roadmap began with the Phoenix field demonstration in 2025. | 中 | SE011, SE026 |
| CE038 | The 2025 Aurora announcement moved Emerald from proof-of-concept toward a reference-design and certification ambition. | 高 | SE004, SE010 |
| CE039 | The 2026 UK trial broadened proof to a European data center and a public utility partner. | 高 | SE007, SE008 |
| CE040 | The 2026 SVP announcement is the clearest transition from demonstration to commercial deployment. | 高 | SE006, SE005 |
| CE041 | The DSX framework post positions ERCOT-style flexible interconnection readiness as a next step for the product roadmap. | 中 | SE005 |
| CE042 | TIME100 and World Economic Forum recognition strengthen credibility but do not replace technical or compliance diligence. | 中 | SE023, SE021 |
| CE043 | Emerald still needs to prove repeatable production deployment, reliability over long periods, customer support at scale, and a stronger public trust/compliance posture. | 中 | SE006, SE007, SE018, SE019 |
| CU001 | Emerald’s current customer universe spans utilities/public power providers, hyperscalers or AI infrastructure operators, data center landlords/operators, and grid institutions. | 高 | SU010, SU013, SU019, SU012 |
| CU002 | Utilities are both customers and enabling partners because they can deploy Emerald software, send dispatch signals, and create the economic pathway for flexibility. | 高 | SU003, SU004, SU010 |
| CU003 | Data center operators and AI infrastructure owners are direct economic beneficiaries because the product can unlock faster power access and better capacity utilization. | 高 | SU018, SU010, SU011 |
| CU004 | The Series A announcement says 12 Fortune 500 companies participated as co-investors, suggesting the go-to-market motion may blur investor, design-partner, and customer roles. | 中 | SU001 |
| CU005 | Partner and company sources say Emerald has completed five live demonstrations across Arizona, Illinois, Virginia, Oregon, and London. | 高 | SU002, SU020 |
| CU006 | The Phoenix proof item involved Oracle, NVIDIA, Databricks, and Salt River Project around a 256-GPU cluster response event. | 高 | SU009, SU006, SU008 |
| CU007 | The Phoenix event achieved a 25% reduction in power consumption for three hours while workloads remained within SLA constraints. | 高 | SU002, SU009, SU006, SU031 |
| CU008 | The UK proof item involved National Grid, Nebius, NVIDIA, EPRI, and Emerald AI at a London-area data center. | 高 | SU004, SU005, SU002 |
| CU009 | The UK trial cut demand by more than one-third in under a minute and by up to 40% while critical workloads continued. | 高 | SU004, SU002 |
| CU010 | National Grid said the UK test sent more than 200 simulated grid events over five days. | 中 | SU004 |
| CU011 | The Santa Clara proof item centers on Silicon Valley Power and an NVIDIA workload site under a flexible load interconnection program. | 高 | SU003, SU002 |
| CU012 | SVP is framed as the first commercial, multi-megawatt DSX Flex deployment rather than just another demonstration. | 中 | SU003 |
| CU013 | Aurora in Manassas links Emerald AI with Digital Realty, PJM, EPRI, and NVIDIA around a 96 MW reference facility. | 高 | SU007, SU011, SU025 |
| CU014 | Aurora is better understood as a roadmap and flagship reference deployment than as proof of scaled recurring customer revenue today. | 中 | SU011, SU007 |
| CU015 | Emerald’s named proofs are stronger than simple logos because they include specific counterparties, geographies, workflows, and power outcomes. | 高 | SU009, SU004, SU003, SU011 |
| CU016 | Emerald does not publicly disclose customer count, retention cohorts, MW under management, or revenue concentration. | 中 | SU001, SU015, SU012, SU019 |
| CU017 | The visible adoption path runs from field demonstration to utility-backed commercial pilot to larger reference deployment. | 高 | SU006, SU003, SU011, SU002 |
| CU018 | The same NVIDIA-utility-data-center ecosystem reappears across multiple proofs, suggesting genuine land-and-expand potential but also ecosystem concentration. | 高 | SU002, SU003, SU004, SU007 |
| CU019 | The public record does not show formal renewals, multiyear contract durations, or cohort retention. | 中 | SU001, SU015, SU012 |
| CU020 | Most named proofs are still pilots, demonstrations, or pre-scale flagship deployments rather than a broad installed customer base. | 中 | SU011, SU003, SU004, SU006 |
| CU021 | Public proof covers Arizona, California, the UK, and Virginia, indicating geographic breadth but still a limited sample of utility frameworks. | 高 | SU006, SU003, SU004, SU011, SU020 |
| CU022 | Customer acquisition and proof are deeply entangled with NVIDIA’s platform and ecosystem. | 高 | SU002, SU003, SU007, SU026 |
| CU023 | Customer conversion is also deeply dependent on utility frameworks such as SVP’s flexible load interconnection program and National Grid’s trial model. | 高 | SU003, SU004, SU005, SU017 |
| CU024 | Much of the public customer evidence is partner-led or utility-led rather than end-user-led testimonials from buyers discussing realized ROI or renewals. | 中 | SU003, SU004, SU010, SU011 |
| CU025 | The customer journey is likely discovery through power-constraint pain, utility or partner engagement, scoped pilot, proof event, and then flagship commercial rollout. | 中 | SU010, SU003, SU004, SU011 |
| CU026 | The clearest expansion driver is reusing successful proof with new utilities, new campuses, or larger AI factory footprints. | 中 | SU002, SU011, SU001 |
| CU027 | The current evidence set implies concentration risk around a small number of flagship partners, utilities, and showcase sites. | 中 | SU002, SU011, SU010 |
| CU028 | Procurement friction is likely high because deals require multi-party coordination among utilities, data center operators, infrastructure vendors, and regulators. | 中 | SU011, SU010, SU014, SU017 |
| CU029 | The strongest public proof item today is the combination of Phoenix measured outcomes plus the UK trial’s rapid-response results because both disclose specific operating metrics. | 高 | SU002, SU004, SU009 |
| CU030 | The freshest 2026 proof items are the National Grid result, the SVP commercial deployment, and the Series A-backed strategic-customer narrative. | 高 | SU004, SU005, SU003, SU001, SU015 |
| CU031 | Strategic investors likely function as channel amplifiers and credibility anchors even when they are not named paying customers. | 中 | SU001, SU010, SU027 |
| CU032 | PJM, EPRI, and DOE-linked ecosystem roles widen channel access but are not equivalent to recurring customers. | 中 | SU023, SU021, SU022 |
| CU033 | TED and National Grid Partners materials suggest Emerald is already running a customer-education motion aimed at operators, utilities, and infrastructure stakeholders, not just investors. | 中 | SU028, SU029, SU030 |
| CU034 | No public NPS, CSAT, case-study renewal quote, or independent review corpus is visible for Emerald’s customer base. | 中 | SU015, SU019, SU012 |
| CU035 | If Phoenix-like proofs translate, Emerald could expand from site-level pilots into utility-standardized programs and multi-site AI factory fleets. | 中 | SU011, SU002, SU020 |
| CU036 | Heatmap’s adverse lens is that even interested customers may resist flexibility unless utilities attach meaningful economic or interconnection value. | 中 | SU017 |
| CU037 | Customer durability remains the main underwriting gap because public evidence shows freshness and technical feasibility more clearly than repeat usage. | 中 | SU001, SU003, SU004 |
| CU038 | The chapter should treat logos like NVIDIA, Digital Realty, Oracle, or PJM as proof amplifiers only when tied to a defined use case or measured outcome. | 中 | SU009, SU011, SU007 |
| CU039 | The shift from pilot to commercial is visible but incomplete: SVP is the clearest commercial proof point, while Aurora remains a reference deployment and broader installed-base evidence is still missing. | 高 | SU003, SU011, SU007 |
| CU040 | Investor-customer overlap may accelerate adoption but can also obscure whether demand is broad-based or concentrated among strategic insiders. | 中 | SU001, SU010, SU027 |
| CU041 | Axios reported Emerald’s first commercial deployment and described the company as already on its way to commercialization, adding a useful external marker between pilot proof and revenue-scale evidence. | 中 | SU032, SU003, SU011 |
| CR001 | FERC’s June 2026 show-cause orders force every major RTO/ISO under its jurisdiction to justify or reform rules for large-load integration. | 高 | SR001, SR006 |
| CR002 | PJM’s proposed IRAS framework would allow certain new large loads to connect before enough capacity exists, while exposing uncovered demand to earlier curtailment. | 中 | SR003, SR004 |
| CR003 | PJM defines a large load as 50 MW or more at a single site for the new framework. | 中 | SR003, SR004 |
| CR004 | Bring Your Own New Capacity is the main path for large loads to reduce or eliminate IRAS exposure. | 中 | SR003, SR004 |
| CR005 | PJM’s filings are driven by real capacity shortfalls and steep new large-load growth expectations. | 中 | SR004, SR003 |
| CR006 | SEPA says 104 approved or pending tariffs and service rules for large loads were being tracked across more than 70 utilities in July 2026. | 高 | SR006, SR007 |
| CR007 | About one quarter of tracked large-load tariffs include concrete dispatchable flexibility or curtailment pathways. | 中 | SR006 |
| CR008 | Large-load tariffs increasingly use design elements such as collateral requirements, minimum terms, exit fees, direct assignment of costs, and customer-specific procurement. | 高 | SR007, SR006 |
| CR009 | PNNL reported that no states had fully developed coordination frameworks for Order 2222-style DER aggregation communications as of early 2026. | 高 | SR002, SR008 |
| CR010 | NERC’s Large Loads Action Plan says existing reliability standards and processes are inadequate for reliable integration of emerging large computational loads. | 中 | SR008 |
| CR011 | Emerald’s terms explicitly say forward-looking statements about capabilities, deployments, and business plans are subject to risks and actual results may differ materially. | 中 | SR016 |
| CR012 | Emerald’s terms also say pilot and demonstration results are illustrative and specific to the conditions under which they were conducted. | 中 | SR016 |
| CR013 | Emerald’s privacy policy says the company maintains safeguards but cannot guarantee absolute security. | 中 | SR017 |
| CR014 | No public SOC 2, ISO 27001, public incident history, or trust-center package is visible in the fetched materials. | 中 | SR017, SR024, SR016 |
| CR015 | Most public proof comes from a small number of named pilots and flagship deployments under specific operating conditions rather than long fleet histories. | 高 | SR019, SR021, SR020, SR016 |
| CR016 | If Emerald underperforms during a grid event, it could simultaneously damage customer workload SLAs and the grid-flexibility value proposition. | 高 | SR021, SR020, SR019 |
| CR017 | Because Emerald’s value depends on telemetry and control over live compute and grid response, a cyber or communications failure would have outsized operational impact. | 高 | SR017, SR020, SR021, SR008 |
| CR018 | Workload diversity remains an execution risk because public proof does not yet show performance across a broad installed base with many customer profiles. | 中 | SR019, SR016, SR018 |
| CR019 | Emerald’s public commercialization story is deeply intertwined with NVIDIA’s stack, ecosystem, and reference designs. | 高 | SR019, SR018, SR030 |
| CR020 | Commercial value also depends on utilities and grid operators creating real economic pathways for flexible load participation. | 高 | SR023, SR020, SR021, SR006 |
| CR021 | The visible proof base is concentrated around a narrow set of counterparties such as NVIDIA, National Grid, SVP, Digital Realty, and a few flagship sites. | 高 | SR019, SR021, SR020, SR022 |
| CR022 | If connect-fast regimes make new large loads first in line for curtailment, Emerald’s customers may face a tougher product-sales conversation, not an easier one. | 中 | SR003, SR004, SR023 |
| CR023 | Strategic investors and advisory-board members help distribution, but they also raise dependence on a partner-led GTM motion. | 高 | SR026, SR018, SR031 |
| CR024 | Emerald is attempting an unusually fast transition from research and demos to multi-region commercial scaling. | 高 | SR025, SR026, SR018 |
| CR025 | Varun Sivaram is central to the company’s policy narrative, customer narrative, and technical narrative, creating obvious key-person risk. | 中 | SR025, SR018, SR027 |
| CR026 | The team-depth claim—half PhDs with 400+ technical publications—helps mitigate execution risk but does not replace a proven scaled field organization. | 高 | SR025, SR018 |
| CR027 | The public record still says little about a scaled implementation, customer success, security, or compliance organization. | 中 | SR024, SR018, SR017 |
| CR028 | Undisclosed revenue, margins, concentration, and runway create financial-model risk even if the technology works. | 中 | SR018, SR023, SR024 |
| CR029 | Many emerging tariff designs are built explicitly to protect ratepayers from large-load risk, which can shift more obligations and costs onto customers and developers. | 高 | SR007, SR006, SR001 |
| CR030 | Some risk is exogenous: data-center load growth itself is straining planning, permitting, and market design independent of any one startup’s execution. | 高 | SR028, SR029, SR004, SR034 |
| CR031 | The rise of onsite power as a default strategy is a substitute risk if customers choose generation-heavy self-help over software-led flexibility. | 高 | SR034, SR027, SR023 |
| CR032 | By mid-2026, FERC and NERC had moved from voluntary discussion to mandatory standards and alerts for computational-load reliability risk. | 高 | SR009, SR010, SR012, SR013, SR011 |
| CR033 | NERC’s May 2026 Level 3 alert required near-term action on modeling, planning, and commissioning for computational loads, showing regulators view the issue as urgent now rather than theoretical later. | 中 | SR013, SR009, SR015 |
| CR034 | Independent summaries of NERC’s 2026 reliability assessment reinforce that load growth and capacity shortfalls are becoming system-level risks across multiple regions. | 中 | SR014, SR004, SR028 |
| CR035 | The risk stack is serious but not automatically fatal because regulators, utilities, and infrastructure operators are actively building frameworks that can reward flexibility. | 高 | SR001, SR006, SR020, SR021 |
| CR036 | A thesis-break trigger would be a tariff regime that allows flexible interconnection rhetorically but denies enough economic value or curtailment certainty for customers to adopt. | 高 | SR023, SR006, SR003 |
| CR037 | A second thesis-break trigger would be a public security, reliability, or SLA event tied to Emerald’s control layer. | 中 | SR017, SR016, SR019 |
| CR038 | A third thesis-break trigger would be failure to show portability beyond the current NVIDIA- and utility-led ecosystem. | 高 | SR019, SR018, SR031 |
| CR039 | A fourth trigger would be learning that a single site or a small partner ring accounts for most expected revenue. | 中 | SR022, SR018, SR023 |
| CR040 | The first diligence priority is regulatory economics: tariff terms, curtailment rights, and who gets paid under flexible-load programs. | 高 | SR006, SR007, SR001 |
| CR041 | The second diligence priority is production readiness: telemetry, fail-safe behavior, support process, and measured SLA outcomes across heterogeneous workloads. | 高 | SR021, SR020, SR016, SR017 |
| CR042 | The third diligence priority is partner and ecosystem concentration across NVIDIA, utilities, landlords, and flagship sites. | 高 | SR019, SR022, SR030 |
| CR043 | The existence of named proofs with National Grid and SVP offsets some go-to-market and execution risk because they show external institutions are willing to pilot or deploy. | 高 | SR021, SR020, SR018 |
| CR044 | Emerald undertakes no obligation to publicly update website information, which raises diligence importance around stale or selectively refreshed claims. | 中 | SR016 |
| CR045 | Customers may choose onsite generation, capacity procurement, or bespoke tariff structures instead of buying Emerald’s orchestration layer. | 高 | SR034, SR003, SR006 |
| CV001 | Emerald announced a $150 million Series A at a $1.05 billion valuation on August 25, 2026. | 高 | SV001, SV002, SV003 |
| CV002 | The August 2026 Form D showed $90.23 million sold out of a $150 million offering as of the filing date. | 高 | SV003, SV001 |
| CV003 | Public evidence places Emerald at a unicorn valuation before it has disclosed public revenue, ARR, margin, or runway. | 高 | SV001, SV031, SV033 |
| CV004 | The positive thesis starts with a real bottleneck: power availability is now constraining AI data-center growth. | 高 | SV012, SV013, SV014, SV015 |
| CV005 | Emerald’s product thesis is that a software control layer can unlock speed-to-power and grid value faster than waiting for new infrastructure. | 高 | SV001, SV011, SV012 |
| CV006 | Customer proof is unusually concrete for the stage, with National Grid, SVP, Phoenix, and Aurora-style evidence rather than only logo slides. | 高 | SV005, SV006, SV007, SV032 |
| CV007 | The anti-thesis is that flexibility may be strategically valuable yet economically thin if utilities and customers do not share enough value. | 高 | SV009, SV006, SV007 |
| CV008 | The second anti-thesis is financial opacity: public investors cannot observe revenue quality, margins, concentration, or runway. | 高 | SV001, SV033, SV031 |
| CV009 | The third anti-thesis is concentration around a small number of counterparties and the NVIDIA-linked ecosystem. | 高 | SV005, SV010, SV008 |
| CV010 | The right headline recommendation is track rather than buy, because company quality appears promising but the price and evidence gap still matter. | 高 | SV001, SV009, SV005, SV033 |
| CV011 | Recommendation confidence should be medium: the market need and proof are real, but financial disclosure is thin. | 高 | SV001, SV031, SV010 |
| CV012 | Risk rating should be high because regulatory economics, concentration, and financial opacity are all material. | 中 | SV009, SV010, SV015 |
| CV013 | The current round looks stretched rather than attractive because public evidence does not yet prove enough revenue or margin support. | 高 | SV001, SV009, SV033 |
| CV014 | At public-style 8x to 12x sales multiples, Emerald would need roughly $88 million to $131 million of annual revenue to support a $1.05 billion value. | 中 | SV001, SV027, SV021, SV018 |
| CV015 | For an attractive venture return above the current round, Emerald likely needs either much higher revenue scale, a richer strategic premium, or both. | 中 | SV001, SV022, SV018, SV030 |
| CV016 | Cap-table detail and preference stack terms are not publicly disclosed, so return math cannot be fully underwritten. | 高 | SV003, SV004, SV001 |
| CV017 | Because current revenue is undisclosed, Emerald is better framed as an option-value or forward-milestone valuation than as a current-sales story. | 中 | SV001, SV033, SV009 |
| CV018 | Equinix is a relevant premium data-center-infrastructure comp, trading around a 10.8x P/S ratio with roughly $106.5B market cap and $9.2B 2025 revenue. | 中 | SV017, SV018, SV019 |
| CV019 | Digital Realty is a relevant landlord / interconnection comp, trading around a 10.8x P/S ratio with roughly $73.1B market cap and $6.1B 2025 revenue. | 中 | SV020, SV021 |
| CV020 | Bloom Energy is a relevant power-bottleneck beneficiary comp, trading around a 20.6x P/S ratio with roughly $64.3B market cap and $2.0B 2025 revenue. | 中 | SV022, SV023 |
| CV021 | Eaton is an adjacent electrification and power-infrastructure comp, trading around a 5.4x P/S ratio with roughly $162.9B market cap and $27.45B 2025 revenue. | 中 | SV024, SV025, SV026 |
| CV022 | Vertiv is a useful AI-infrastructure power-and-thermal comp, trading around an 8.8x P/S ratio with roughly $101.6B market cap and $10.23B 2025 revenue. | 中 | SV027, SV029 |
| CV023 | The relevant public comp band is roughly 5x to 21x sales, with richer multiples reserved for businesses that already disclose billions in revenue. | 中 | SV022, SV024, SV027, SV018, SV021 |
| CV024 | That comparison underscores the key problem: Emerald asks investors to price strategic option value without disclosing the revenue base that public comps disclose routinely. | 中 | SV001, SV018, SV024, SV027 |
| CV025 | A credible bull case needs repeat paid deployments, broader portability beyond the current ecosystem, and revenue scaling toward roughly $160 million to $240 million by 2028. | 中 | SV005, SV006, SV001, SV022 |
| CV026 | A reasonable base case assumes commercialization continues but revenue scales more slowly, into roughly a $70 million to $110 million range by 2028. | 中 | SV006, SV007, SV009, SV001 |
| CV027 | A bear case assumes value capture stays thin, concentration remains high, and revenue reaches only roughly $20 million to $45 million by 2028. | 中 | SV009, SV033, SV010 |
| CV028 | Applying 10x to 12x sales to the bull case suggests a rough $1.6 billion to $2.9 billion valuation range. | 中 | SV022, SV018, SV027 |
| CV029 | Applying 6x to 8x sales to the base case suggests a rough $420 million to $880 million valuation range. | 中 | SV024, SV027, SV021 |
| CV030 | Applying 3x to 5x sales to the bear case suggests a rough $60 million to $225 million valuation range. | 中 | SV024, SV021, SV009 |
| CV031 | A probability-weighted outcome across those scenarios lands below the current round unless Emerald quickly proves unusually strong revenue scale and durability. | 中 | SV001, SV009, SV027, SV021 |
| CV032 | Regulatory-economics risk transmits directly into valuation because flexible-load customers may face curtailment rights, collateral, or weak compensation structures. | 中 | SV009, SV015, SV012 |
| CV033 | Customer and ecosystem concentration reduce the quality of any future revenue base and therefore compress defendable multiple support. | 中 | SV010, SV005, SV008 |
| CV034 | The recommendation could improve if Emerald discloses real contract values, demonstrates non-NVIDIA portability, and converts flagship proofs into repeat multi-site programs. | 高 | SV001, SV006, SV007, SV008 |
| CV035 | The recommendation would worsen toward avoid if tariff economics stay weak, one or two sites dominate value, or a security / reliability incident occurs. | 高 | SV009, SV010, SV033 |
| CV036 | The most plausible exit path is a strategic-acquisition or later IPO once revenue and durability are clearer; neither is ready to underwrite today from public evidence alone. | 中 | SV001, SV030, SV026 |
| CV037 | A typical strategic acquisition may not clear the current round at venture-attractive returns unless Emerald becomes uniquely strategic or much larger. | 中 | SV030, SV026, SV018 |
| CV038 | The most important diligence ask is customer contract value and renewal structure. | 中 | SV001, SV006, SV008 |
| CV039 | The second key diligence ask is gross margin and services mix after implementation. | 中 | SV033, SV031, SV001 |
| CV040 | The third key diligence ask is full cap table, preference stack, and secondary liquidity context. | 高 | SV003, SV004 |
| CV041 | The fourth key diligence ask is concentration by site, utility, and partner channel. | 中 | SV010, SV007, SV006 |
| CV042 | The fifth key diligence ask is cash, burn, and runway under bull/base/bear commercialization paths. | 中 | SV001, SV033, SV009 |
| CV043 | On an IC scorecard, market attractiveness is high because the AI power bottleneck is real and worsening. | 高 | SV013, SV014, SV015 |
| CV044 | Proof quality is medium-high because Emerald has named deployments and measured outcomes, but the installed base remains small. | 高 | SV005, SV007, SV006, SV010 |
| CV045 | Moat is medium because coordination know-how and ecosystem access matter, but portability and standardization remain unresolved. | 中 | SV008, SV005, SV009 |
| CV046 | Economics confidence is low-medium because pricing power is plausible but financial evidence is thin. | 中 | SV009, SV001, SV011 |
| CV047 | Valuation support is low-medium at the current round because the price already assumes significant future scale. | 中 | SV001, SV018, SV021, SV027 |
| 编号 | 出版方 | 标题 | 引文 |
|---|---|---|---|
| SO001 | Emerald AI | Emerald AI homepage | |
| SO002 | Emerald AI | Our Team | |
| SO003 | Emerald AI | Contact Us | |
| SO004 | Emerald AI | Dr. Varun Sivaram | |
| SO005 | Emerald AI | Prof. Ayse Coskun | |
| SO006 | Emerald AI | Shayan Sengupta | |
| SO007 | Emerald AI | Aroon Vijaykar | |
| SO008 | Emerald AI | Mansi Shah | |
| SO009 | Emerald AI | Emerald AI Raises $150M Series A at $1.05B Valuation | |
| SO010 | Emerald AI | Our Series A Coalition, in Their Own Words | |
| SO011 | Emerald AI | Sharing our Strategic Expansion Round: Emerald AI Raises $25 Million to Transform AI Data Centers into Flexible Power Grid Assets | |
| SO012 | Emerald AI | Sharing Our Seed Extension: Emerald AI's Total Funding Reaches $42.5 Million to Scale Power-Flexible AI Infrastructure | |
| SO013 | Emerald AI | Emerald AI Teams with NVIDIA and Partners to Develop Power-Flexible AI Factory and Reference Design to Unlock 100 GW of Grid Capacity and Supercharge the AI Revolution | |
| SO014 | PR Newswire | Emerald AI Launches with $24.5M Seed Round to Transform AI Data Centers into Grid Allies | |
| SO015 | U.S. Securities and Exchange Commission | Emerald AI, Inc. Form D filing (August 2025) | |
| SO016 | U.S. Securities and Exchange Commission | Emerald AI, Inc. Form D filing (February 2026) | |
| SO017 | U.S. Securities and Exchange Commission | Emerald AI, Inc. Form D filing (August 2026) | |
| SO018 | Virginia Business | Emerald AI raises $22.7M, new SEC filing reports | |
| SO019 | Salesforce Ventures | Welcome, Emerald AI! | |
| SO020 | TIME | Emerald AI | |
| SO021 | CNBC | Emerald AI CEO Varun Sivaram: We transform AI data centers into 'flexible grid allies' | |
| SO022 | Newsweek | Emerald AI has a new approach to meeting AI's energy demand | |
| SO023 | Heatmap | The Software That Could Save the Grid | |
| SO024 | S&P Global Market Intelligence | 'Power-flexible' AI data center unveiled in Virginia, touted as template | |
| SO025 | Silicon Valley Power | News Release: Silicon Valley Power and Emerald AI Launch Pilot to Demonstrate Flexible Data Centers in Santa Clara and Unlock Power Capacity for AI | |
| SO026 | National Grid | UK-first trial of AI Grid Technology Successfully Demonstrates the Ability of Data Centres to Adjust Power Needs | |
| SO027 | Emerald AI | Emerald AI Is Selected as a 2026 Technology Pioneer by the World Economic Forum | |
| SM001 | International Energy Agency | Electricity 2026: Executive Summary | In the United States, electricity demand rose by 2.1% in 2025 and is projected to grow by nearly 2% annually through 2030, with around half of the total increase driven by the rapid expansion of data centres. |
| SM002 | Lawrence Berkeley National Laboratory | Berkeley Lab report evaluates increase in electricity demand from data centers | Lawrence Berkeley National Laboratory estimates data centers consumed about 4.4% of U.S. electricity in 2023 and could rise to 6.7% to 12% by 2028. |
| SM003 | Lawrence Berkeley National Laboratory | Electricity Rate Designs for Large Loads: Evolving Practices and Opportunities 2026 Update | The sample of 55 tariffs includes electric utility tariffs, contracts, and other frameworks. |
| SM004 | JLL Research | 2026 Global Data Center Outlook | Nearly 100 GW of new data centers will be added between 2026 and 2030, doubling global capacity. |
| SM005 | CBRE Research | Global Data Center Trends 2026 | Global power availability and grid infrastructure constraints are impacting development timelines and site selection, especially in established hubs in North America and Europe. |
| SM006 | CBRE Research | U.S. Real Estate Market Outlook 2026: Data Centers | Delivery of 300 MW or more within the next 36 months will become the most important location consideration, outranking power pricing and connectivity in most cases. |
| SM007 | Bloom Energy | 2026 Data Center Power Report | In parallel, 73% of respondents report actively evaluating or selecting onsite power providers. |
| SM008 | Federal Energy Regulatory Commission | FERC launches aggressive, targeted action to speed large load integration | Providing new transmission services for flexible large loads. |
| SM009 | Pacific Northwest National Laboratory | FERC Order 2222 DER Policy and Implementation Tracker Report | No states have fully developed coordination frameworks as of early 2026. |
| SM010 | Smart Electric Power Alliance | Where large-load tariffs fit in the future of data center flexibility | In the last year alone, our data have grown from 41 proposed and approved tariffs and rules in July 2025 to 104 in July 2026. |
| SM011 | Council on Foreign Relations | America May Not Need a Massive Energy Build-Out to Power the AI Revolution | Roughly 100 GW ... could be connected in the near term to power grids across the United States with no new power supply or delivery infrastructure upgrades. |
| SM012 | Utility Dive | It’s not a grid, it’s a system: tools and mindsets to optimize electric power delivery | Some utilities are already testing this flexibility. Salt River Project in Arizona saw a 25% reduction in power consumption over three hours from a data center cluster of 256 Nvidia GPUs using software from Emerald AI. |
| SM013 | Data Center Knowledge | PJM’s new deal for data centers: bring power or face cuts | PJM defines a “Large Load” as end-use customer demand with a cumulative peak of at least 50 MW at a single electrical site. |
| SM014 | POWER Magazine | PJM widens response to data center load as capacity shortfalls deepen | PJM’s Aug. 13 filing points to a forecast that shows peak demand surging by about 32 GW between 2024 and 2030—roughly 30 GW of it attributable to data centers. |
| SM015 | EPRI DCFlex | DCFlex initiative | Demonstration sites were launched to test different flexibility methods at grid-connected data centers. |
| SM016 | IEEE Spectrum | Big Tech Tests Data Center Flexibility for Local Power Grids | The selected hubs will serve as testbeds for solutions to the rising electricity demands of AI. |
| SM017 | Emerald AI | FERC is making flexibility its policy | The Commission explicitly called for new transmission services for flexible large loads. |
| SM018 | Emerald AI | NVIDIA DSX pilot framework | The DSX framework is designed to offer grid access and economic value to data centers that commit to flexibility. |
| SM019 | NVIDIA | AI factories and flexible power use | Flexible power use can let AI factories align compute with the needs of the power system. |
| SM020 | Nature Energy | Grid-interactive data centers can accelerate AI under power constraints | Grid-interactive control can reduce data center power consumption while preserving critical workloads. |
| SM021 | arXiv | Power-flexible AI clusters under inference-dominant workloads | Inference-heavy clusters still contain schedulable flexibility, though less than offline training regimes. |
| SM022 | Lawrence Berkeley National Laboratory | Data centers | Berkeley Lab studies data centers as a major and growing electricity end-use. |
| SM023 | Heatmap | Emerald AI, Nvidia, and the case for flexible data centers | If utilities do not create enough connection value or incentives, the economics for flexible data centers are harder to justify. |
| SM024 | National Grid | UK first trial of AI grid technology successfully demonstrates ability for data centres to adjust power needs | The trial demonstrated the ability of data centres to adjust power needs by more than a third in under a minute and by up to 40%. |
| SM025 | E&E News | Nvidia-backed startup wants data centers to be grid assets | Emerald AI is pitching data centers as controllable grid assets rather than fixed loads. |
| SM026 | Latitude Media | Nvidia and Oracle tapped this startup to flex a Phoenix data center | The Phoenix demonstration shows the commercial path requires utilities, data center operators, and AI infrastructure partners to align. |
| SP001 | Emerald AI | Emerald AI raises $150 million Series A | Emerald AI is building an AI platform for power-flexible data centers and grid management. |
| SP002 | Emerald AI | Launching the first power-flexible AI factory with NVIDIA | The offering is designed specifically around AI factories and power-flexible data centers. |
| SP003 | Voltus | Voltus | Voltus pays thousands of commercial, industrial, and residential energy users to support grid reliability. |
| SP004 | CPower | CPower | CPower’s Virtual Power Plant Platform monetizes your energy through demand response and energy flexibility programs. |
| SP005 | Virtual Peaker | Virtual Peaker | Modern utilities use Virtual Peaker to launch residential, commercial, and industrial demand flexibility programs. |
| SP006 | EnergyHub | EnergyHub | Flexible resources can operate at scale in a number of types of environments. |
| SP007 | Leap | Leap | Build and Scale Your Virtual Power Plants | Build and scale your virtual power plants. |
| SP008 | Amperon | Amperon | We provide the highest precision energy forecasting and analytics solutions to improve grid reliability, manage financial risk, and optimize renewable assets. |
| SP009 | GridPoint | GridPoint | By enabling dynamic load flexibility ... GridPoint uniquely serves both businesses and utilities with one platform. |
| SP010 | Uplight | Uplight | Our open, AI-powered platform combines personalized customer experiences with flexible load management. |
| SP011 | Itron | Itron Grid Management | Grid reliability: A reliable and resilient grid is essential. |
| SP012 | Enel North America | Enel North America | Create competitive advantage and maximize value with our suite of solutions that enable your organization to be more flexible in how you acquire and use energy. |
| SP013 | Bloom Energy | 2026 Data Center Power Report | Over one-third of data centers are expected to use 100% onsite power by 2030. |
| SP014 | CBRE Research | Global Data Center Trends 2026 | Global power availability and grid infrastructure constraints are impacting development timelines and site selection. |
| SP015 | JLL Research | 2026 Global Data Center Outlook | Power, not location or cost, will be the primary site selection criteria due to multiyear wait times for a grid connection. |
| SP016 | Smart Electric Power Alliance | Where large-load tariffs fit in the future of data center flexibility | One-quarter of the large-load tariffs and service rules we track include a concrete option for dispatchable large-load flexibility. |
| SP017 | Federal Energy Regulatory Commission | FERC launches aggressive, targeted action to speed large load integration | Providing new transmission services for flexible large loads. |
| SP018 | Data Center Knowledge | PJM’s new deal for data centers: bring power or face cuts | PJM defines a Large Load as end-use customer demand with a cumulative peak of at least 50 MW at a single electrical site. |
| SP019 | Heatmap | Emerald AI, Nvidia, and the case for flexible data centers | The business case gets much stronger if utilities offer meaningful connection advantages or compensation. |
| SP020 | IEEE Spectrum | Big Tech Tests Data Center Flexibility for Local Power Grids | Emerald AI will coordinate the choreography with local utilities at a Phoenix site. |
| SP021 | EPRI DCFlex | DCFlex initiative | Demonstration sites test different aspects of flexibility at live data centers. |
| SP022 | Public Power | NVIDIA, Emerald AI, EPRI, PJM and others develop power-flexible AI factory | The coalition is developing a power-flexible AI factory and reference design. |
| SP023 | Newsweek | Emerald AI Has a New Approach to Meeting AI’s Energy Demand | Emerald AI is attempting to turn data centers into grid allies. |
| SP024 | Silicon Valley Power | SVP and Emerald AI launch pilot to demonstrate flexible data centers in Santa Clara | The pilot offers expanded grid access in exchange for verifiable flexibility. |
| SP025 | National Grid | UK first trial of AI grid technology successfully demonstrates ability for data centres to adjust power needs | The trial validated that AI data centres can dynamically adjust power consumption without disrupting critical workloads. |
| SP026 | NVIDIA | Emerald AI case study | Emerald AI uses NVIDIA software and hardware stacks to enable power-flexible AI infrastructure. |
| SP027 | CompaniesMarketCap | Equinix market cap | Equinix is one of the world’s largest data center operators by public market capitalization. |
| SP028 | CompaniesMarketCap | Digital Realty market cap | Digital Realty is a large public data center operator. |
| SI001 | Emerald AI | Emerald AI raises $150 million Series A | Emerald AI announced it has raised $150 million in an oversubscribed Series A financing at a valuation of $1.05 billion. |
| SI002 | Emerald AI | Emerald AI $150 million Series A valuation investor quotes | The landmark $150 million Series A round included twelve Fortune Global 500 companies in the coalition. |
| SI003 | PR Newswire | Emerald AI launches with $24.5M seed round | Emerald AI launched with a $24.5 million seed round. |
| SI004 | Emerald AI | Sharing our seed extension | We’ve raised an additional $18 million, bringing our total funding raised to $42.5 million. |
| SI005 | Emerald AI | Sharing our strategic expansion round | Emerald AI has raised $25 million in a Strategic Expansion Round, bringing total funding to $68 million. |
| SI006 | SEC | Emerald AI Form D filing (August 2025) | The Form D filing shows a $35,299,903 offering amount with $34,169,168 sold to 37 investors. |
| SI007 | SEC | Emerald AI Form D primary document (February 2026) | The filing lists a $24,999,604 total offering amount, $22,749,615 sold, and 20 investors. |
| SI008 | SEC | Emerald AI Form D primary document (August 2026) | The filing lists a $150,000,000 offering amount, $90,229,639 sold, and 23 investors. |
| SI009 | Salesforce Ventures | Welcome Emerald AI | Hyperscalers, data center operators, and utilities need innovative solutions that support AI-scale workloads. |
| SI010 | NVIDIA | Emerald AI case study | The Emerald Conductor platform is proving that AI factories can be power-flexible grid assets. |
| SI011 | Silicon Valley Power | SVP and Emerald AI launch pilot to demonstrate flexible data centers in Santa Clara | SVP will deploy Emerald AI software to help manage and dispatch participating flexible data centers during limited periods of grid need. |
| SI012 | National Grid | UK first trial of AI grid technology successfully demonstrates ability for data centres to adjust power needs | The trial validated that data centres can dynamically adjust power consumption without disrupting critical workloads. |
| SI013 | S&P Global Market Intelligence | Power-flexible AI data center unveiled in Virginia, touted as template | The Aurora facility is designed with Emerald AI and NVIDIA technology to orchestrate AI workloads with grid needs. |
| SI014 | CNBC | Emerald AI CEO: We transform AI data centers into flexible grid allies | We transform AI data centers into flexible grid allies. |
| SI015 | Newsweek | Emerald AI Has a New Approach to Meeting AI’s Energy Demand | Emerald AI is taking a software-first approach to a massive energy bottleneck. |
| SI016 | Heatmap | Emerald AI, Nvidia, and the case for flexible data centers | Utilities still need to create enough value for flexible data center customers to care. |
| SI017 | Data Center Dynamics | Nvidia-backed Emerald AI raises $24.5m to turn data centers into grid assets | Emerald Conductor could enable data centers to obtain a grid connection significantly more quickly. |
| SI018 | E&E News | Nvidia-backed startup wants data centers to be grid assets | Emerald AI orchestrates AI workloads in real time to avoid straining the grid in times of peak demand. |
| SI019 | Bloom Energy | 2026 Data Center Power Report | Power availability has become the gating factor for data center expansion. |
| SI020 | Bloom Energy | 2026 Data Center Power Report PDF | Over one-third of data centers are expected to use 100% onsite power by 2030. |
| SI021 | Council on Foreign Relations | America may not need a massive energy build-out to power the AI revolution | Flexible data centers could connect swiftly to existing power grids without waiting up to a decade for new infrastructure. |
| SI022 | Berkeley Lab | Berkeley Lab report evaluates increase in electricity demand from data centers | Data centers consumed about 4.4% of total U.S. electricity in 2023 and are expected to consume between 6.7 and 12% by 2028. |
| SI023 | Berkeley Lab ETA | Data centers | Data center load growth has tripled over the past decade and is projected to double or triple by 2028. |
| SI024 | Stock Analysis | Vertiv Holdings Co revenue | Vertiv had annual revenue of $10.23B in 2025. |
| SI025 | Stock Analysis | Vertiv Holdings Co stock price and overview | Vertiv market cap was 101.56B with revenue (ttm) 11.48B. |
| SI026 | Stock Analysis | Equinix revenue | Equinix had annual revenue of $9.22B in 2025 and market cap of 106.53B. |
| SI027 | CompaniesMarketCap | Digital Realty market cap | Digital Realty had a market cap of $73.05B in August 2026. |
| SI028 | CompaniesMarketCap | Bloom Energy market cap | Bloom Energy had a market cap of $64.26B in August 2026. |
| SI029 | Stock Analysis | Eaton Corporation revenue | Eaton had annual revenue of $27.45B in 2025 and market cap of 162.91B. |
| SI030 | Public Power | NVIDIA, Emerald AI, EPRI, PJM and others develop power-flexible AI factory | The Aurora AI Factory in Virginia will implement a new reference design and certification standard. |
| SI031 | PR Newswire | Emerald AI teams with NVIDIA and partners to develop power-flexible AI factory | The coalition includes Digital Realty, PJM, and EPRI around the Aurora AI Factory. |
| SI032 | Emerald AI | Company | Emerald AI targets AI data centers and power grids. |
| SE001 | Emerald AI | Company | Emerald AI is the pioneer in AI-driven data center flexibility management. |
| SE002 | Emerald AI | Careers | Emerald AI is hiring against a growing product and deployment agenda. |
| SE003 | Emerald AI | Emerald AI raises $150 million Series A | Emerald AI is an AI platform for power-flexible data centers and grid management. |
| SE004 | Emerald AI | Launching the first power-flexible AI factory with NVIDIA | Emerald AI’s GridLink and Conductor products leverage NVIDIA AI Enterprise components, including NVIDIA NIM microservices, in coordination with NVIDIA Mission Control. |
| SE005 | Emerald AI | NVIDIA DSX pilot framework | Conductor dynamically modulates the facility’s power consumption in real time in response to utility signals, while preserving workload performance. |
| SE006 | Silicon Valley Power | SVP and Emerald AI launch pilot to demonstrate flexible data centers in Santa Clara | Emerald AI software will help manage and dispatch participating flexible data centers during limited periods of grid need. |
| SE007 | National Grid | UK first trial of AI grid technology successfully demonstrates ability for data centres to adjust power needs | Using Emerald AI’s software, Emerald Conductor, the trial validated that data centres can dynamically adjust power consumption in response to real-time signals, without disrupting critical workloads. |
| SE008 | NVIDIA | Emerald AI case study | Its Emerald Conductor platform transforms AI factories into power grid assets. |
| SE009 | NVIDIA | AI factories and flexible power use | Flexible power use can let AI factories align compute with the needs of the power system. |
| SE010 | Public Power | NVIDIA, Emerald AI, EPRI, PJM and others develop power-flexible AI factory | GridLink and Conductor products leverage NVIDIA AI Enterprise components, including NVIDIA NIM microservices. |
| SE011 | Public Power | SRP participates in artificial intelligence data center demonstration | The Emerald AI Conductor software platform orchestrates AI workloads in real-time. |
| SE012 | IEEE Spectrum | Big Tech Tests Data Center Flexibility for Local Power Grids | At the Phoenix site, Emerald AI will coordinate the choreography with local utilities. |
| SE013 | Latitude Media | Nvidia and Oracle tapped this startup to flex a Phoenix data center | The platform continuously profiles jobs across flexibility, time sensitivity, and performance tolerance and models thousands of optimization scenarios in seconds. |
| SE014 | GitHub | Emerald AI demo repository | The repository contains pseudocode for Emerald Conductor and key implementation code snippets from the AI orchestration layer. |
| SE015 | arXiv | Power-flexible AI clusters under inference-dominant workloads | The architecture integrates grid signals, workload scheduling, and power telemetry for fine-grained cluster power control. |
| SE016 | Nature Energy | Grid-interactive data centers can accelerate AI under power constraints | Grid-interactive control can reduce data center power consumption while preserving critical workloads. |
| SE017 | Council on Foreign Relations | America May Not Need a Massive Energy Build-Out to Power the AI Revolution | An even cheaper option is to orchestrate computational workloads across one or many data centers to precisely control power consumption while maintaining acceptable service quality. |
| SE018 | Emerald AI | Terms and Conditions | Information on this website may contain forward-looking statements ... actual results may differ materially. |
| SE019 | Emerald AI | Privacy Policy | We do not use personal information collected through our website to train machine learning or artificial intelligence models. |
| SE020 | Emerald AI | Contact Us | Resources: Terms & Conditions, Privacy Policy. |
| SE021 | World Economic Forum | Emerald AI organization profile | The Emerald Conductor platform dynamically interacts with local power grids, pausing AI workloads or routing them to a different area during times of grid stress. |
| SE022 | Emerald AI | Emerald AI joins DOE Genesis Mission Consortium | The consortium advances AI for scientific discovery and energy applications. |
| SE023 | TIME | TIME100 Most Influential Companies 2026: Emerald AI | Emerald AI is recognized for tackling AI’s energy bottleneck. |
| SE024 | Newsweek | Emerald AI Has a New Approach to Meeting AI’s Energy Demand | Emerald AI has a new approach to meeting AI’s energy demand. |
| SE025 | Data Center Dynamics | Nvidia-backed Emerald AI raises $24.5m to turn data centers into grid assets | Emerald Conductor could enable data centers to obtain a grid connection significantly more quickly by managing energy consumption through AI. |
| SE026 | arXiv | Turning AI Data Centers into Grid-Interactive Assets: Results from a Field Demonstration in Phoenix, Arizona | The field demonstration provides experimental evidence for grid-interactive AI data center control. |
| SU001 | Emerald AI | Emerald AI raises $150 million Series A | Twelve Fortune 500 companies participated as co-investors. |
| SU002 | NVIDIA | Emerald AI case study | Across five demonstrations at commercial facilities from Arizona to the United Kingdom, the Emerald Conductor platform is proving that AI factories can be power-flexible grid assets. |
| SU003 | Silicon Valley Power | SVP and Emerald AI launch pilot to demonstrate flexible data centers in Santa Clara | SVP will deploy Emerald AI software to help manage and dispatch participating flexible data centers during limited periods of grid need. |
| SU004 | National Grid | UK first trial of AI grid technology successfully demonstrates ability for data centres to adjust power needs | The trial validated that data centres can dynamically adjust power consumption without disrupting critical workloads. |
| SU005 | Emerald AI | National Grid and Emerald AI announce strategic partnership to demonstrate AI power flexibility in the UK | National Grid and Emerald AI announced a strategic partnership to demonstrate how AI data centres can work with the transmission network to adjust their energy use in real time. |
| SU006 | Public Power | SRP participates in artificial intelligence data center demonstration | The Emerald AI Conductor software platform orchestrates AI workloads in real-time. |
| SU007 | Public Power | NVIDIA, Emerald AI, EPRI, PJM and others develop power-flexible AI factory | The Aurora AI Factory in Virginia will serve as the implementation of a new reference design and certification standard. |
| SU008 | IEEE Spectrum | Big Tech Tests Data Center Flexibility for Local Power Grids | At the Phoenix site, Emerald AI will coordinate the choreography with local utilities, including Salt River Project. |
| SU009 | Latitude Media | Nvidia and Oracle tapped this startup to flex a Phoenix data center | In Phoenix, partners Oracle, Emerald AI, Nvidia, Databricks, and Salt River Project sought to reduce a data center’s power consumption by 25% for three hours. |
| SU010 | Salesforce Ventures | Welcome Emerald AI | Hyperscalers, data center operators, and utilities need innovative solutions that can support AI-scale workloads. |
| SU011 | S&P Global Market Intelligence | Power-flexible AI data center unveiled in Virginia, touted as template | The Aurora facility, being built by Digital Realty, is designed with Emerald AI and NVIDIA technology to orchestrate AI computing workloads with the needs of the grid. |
| SU012 | CNBC | Emerald AI CEO: We transform AI data centers into flexible grid allies | We transform AI data centers into flexible grid allies. |
| SU013 | E&E News | Nvidia-backed startup wants data centers to be grid assets | Emerald AI is pitching data centers as controllable grid assets rather than fixed loads. |
| SU014 | Axios | Utilities, Nvidia, and Emerald AI on power-flexible data centers | Utilities and AI infrastructure providers are experimenting with new models for data center flexibility. |
| SU015 | Newsweek | Emerald AI Has a New Approach to Meeting AI’s Energy Demand | Emerald AI has a new approach to meeting AI’s energy demand. |
| SU016 | TIME | TIME100 Most Influential Companies 2026: Emerald AI | Emerald AI is recognized for addressing AI’s energy bottleneck. |
| SU017 | Heatmap | Emerald AI, Nvidia, and the case for flexible data centers | Utilities still need to create enough value for flexible data center customers to care. |
| SU018 | Data Center Dynamics | Nvidia-backed Emerald AI raises $24.5m to turn data centers into grid assets | Emerald Conductor could enable data centers to obtain a grid connection significantly more quickly by managing energy consumption through AI. |
| SU019 | Emerald AI | Company | Emerald AI targets AI data centers and power grids. |
| SU020 | World Economic Forum | Emerald AI organization profile | The team has demonstrated this in five live commercial deployments across Arizona, Illinois, Virginia, Oregon, and London. |
| SU021 | Emerald AI | Emerald AI joins DOE Genesis Mission Consortium | Emerald AI is participating in a DOE-linked ecosystem effort. |
| SU022 | EPRI DCFlex | DCFlex initiative | Demonstration sites test real-world data center flexibility methods. |
| SU023 | Public Power | NVIDIA, Emerald AI, EPRI, PJM and others develop power-flexible AI factory | Digital Realty and PJM are part of the coalition around Aurora. |
| SU024 | PR Newswire | Emerald AI launches with $24.5M seed round | The launch positioned Emerald around data center customers and grid partners. |
| SU025 | PR Newswire | Emerald AI teams with NVIDIA and partners to develop power-flexible AI factory | The coalition includes Digital Realty, PJM, and EPRI around the Aurora AI Factory. |
| SU026 | NVIDIA | AI energy innovation climate research | NVIDIA describes Emerald as part of a broader energy innovation push around AI infrastructure. |
| SU027 | NGP Energy Technology Partners | Emerald AI | Emerald AI sits at the intersection of AI infrastructure and grid flexibility. |
| SU028 | National Grid Partners | Emerald AI whitepaper | The customer thesis centers on utilities and data center operators using flexibility to unlock power capacity. |
| SU029 | TED | How AI can solve its own energy crisis | Varun Sivaram presents flexible AI infrastructure to a practitioner audience. |
| SU030 | TED | The story you are not hearing about AI data centers | Ayse Coskun presents the technical and customer problem behind AI data center flexibility. |
| SU031 | Virginia Business | Emerald AI raises $22.7M, new SEC filing reports | The company released results from a May 2025 demonstration in Phoenix showing a 25% reduction for three hours while maintaining service quality. |
| SU032 | Axios | Cleantech veteran hopes to turn AI energy crisis on its head | Emerald AI had its first commercial deployment and was already well on its way to commercialization. |
| SR001 | FERC | FERC launches aggressive targeted action to speed large load integration | FERC directed six regional grid operators to justify or reform rules governing how data centers and other large loads connect to the grid. |
| SR002 | PNNL | FERC Order 2222 DER policy and implementation report | No states had fully developed coordination frameworks as of early 2026. |
| SR003 | Data Center Knowledge | PJM’s new deal for data centers: bring power or face cuts | Any portion of demand not backed by qualifying new capacity could be curtailed before other pre-emergency demand-response measures. |
| SR004 | POWER Magazine | PJM widens response to data center load as capacity shortfalls deepen | PJM points to capacity shortfalls and expects peak demand surging by about 32 GW between 2024 and 2030. |
| SR005 | Utility Dive | It’s not a grid, it’s a system | Data centers would prefer to never flex, but they do not need to draw 100% of maximum nameplate demand at all hours. |
| SR006 | SEPA | Where large-load tariffs fit in the future of data center flexibility | In July 2026 DELTa tracked 104 approved and pending tariffs and service rules across more than 70 utilities. |
| SR007 | Berkeley Lab / Brattle | Electricity rate designs for large loads: evolving practices and opportunities 2026 update | Utilities and regulators are managing operational and financial risks through tariffs, service agreements, collateral requirements, and direct assignment of costs. |
| SR008 | NERC | Large Loads Action Plan Q1 2026 update | Existing reliability standards, processes, and requirements are inadequate for the reliable integration of emerging large loads. |
| SR009 | POWER Magazine | FERC orders mandatory NERC reliability standards for data center and other computational loads | FERC ordered NERC to file new or modified reliability standards for computational loads by Dec. 31, 2026. |
| SR010 | FERC | RD26-7-000 order PDF | The order directs NERC to file new or modified reliability standards and associated registry criteria revisions by Dec. 31, 2026. |
| SR011 | Climate Solutions Law | NERC launches Project 2026-02 to address reliability risks from computational loads | Project 2026-02 signals that computational loads may face new registration obligations and reliability standards. |
| SR012 | Troutman Energy Report | FERC directs NERC to submit rules addressing risks associated with integration of computational loads into bulk power system | FERC made NERC’s computational-load schedule mandatory and enforceable rather than voluntary. |
| SR013 | EPG Solutions | Data center load risks to BPS reliability | NERC issued a Level 3 Essential Action Alert in May 2026 targeting grid stability risks posed by large computational loads. |
| SR014 | PureSky Energy | Key takeaways from NERC’s January 2026 long-term reliability assessment | NERC projected a 20-25% increase in nationwide peak demand over the next decade and warned of capacity shortfalls. |
| SR015 | KeenTel Engineering | NERC large loads: 2026 interconnection guide | Large computational loads are expected to play a more direct role in NERC’s reliability framework. |
| SR016 | Emerald AI | Terms and Conditions | Descriptions of pilots and demonstrations are illustrative and specific to the conditions under which they were conducted; past performance is not indicative of future results. |
| SR017 | Emerald AI | Privacy Policy | We maintain technical, administrative, and organizational safeguards, but no system is perfectly secure and we cannot guarantee absolute security. |
| SR018 | Emerald AI | Emerald AI raises $150 million Series A | The company says its technology now runs commercially at full data center scale. |
| SR019 | NVIDIA | Emerald AI case study | Across five demonstrations at commercial facilities, Emerald Conductor is proving AI factories can be power-flexible grid assets. |
| SR020 | Silicon Valley Power | SVP and Emerald AI launch pilot to demonstrate flexible data centers in Santa Clara | SVP will deploy Emerald AI software to manage and dispatch participating flexible data centers during limited periods of grid need. |
| SR021 | National Grid | UK first trial of AI grid technology successfully demonstrates ability for data centres to adjust power needs | The trial validated that data centres can dynamically adjust power consumption without disrupting critical workloads. |
| SR022 | S&P Global Market Intelligence | Power-flexible AI data center unveiled in Virginia, touted as template | Aurora is a flagship facility being built by Digital Realty with Emerald AI and NVIDIA technology. |
| SR023 | Heatmap | Emerald AI, Nvidia, and the case for flexible data centers | Utilities still need to create enough value for flexible data center customers to care. |
| SR024 | Emerald AI | Company | Emerald AI targets AI data centers and power grids. |
| SR025 | Emerald AI | Sharing our seed extension | Emerald says the team is now half PhDs with over 400 technical publications. |
| SR026 | Emerald AI | Sharing our strategic expansion round | Emerald launched a Strategic Advisory Board including seven Fortune 500 companies. |
| SR027 | Council on Foreign Relations | America may not need a massive energy build-out to power the AI revolution | Flexible data centers could connect swiftly to existing grids, but only if they accept limited flexibility during rare stress hours. |
| SR028 | Berkeley Lab | Berkeley Lab report evaluates increase in electricity demand from data centers | Data center electricity use could reach 6.7% to 12% of total U.S. electricity by 2028. |
| SR029 | Berkeley Lab ETA | Data centers | Data center load growth has tripled over the past decade and is projected to double or triple by 2028. |
| SR030 | Public Power | NVIDIA, Emerald AI, EPRI, PJM and others develop power-flexible AI factory | Aurora involves Digital Realty, EPRI, PJM, and NVIDIA alongside Emerald AI. |
| SR031 | Salesforce Ventures | Welcome Emerald AI | Hyperscalers, data center operators, and utilities need innovative solutions for AI-scale workloads. |
| SR032 | Data Center Dynamics | Nvidia-backed Emerald AI raises $24.5m to turn data centers into grid assets | The software could enable data centers to obtain a grid connection significantly more quickly. |
| SR033 | E&E News | Nvidia-backed startup wants data centers to be grid assets | Emerald pitches data centers as controllable grid assets rather than fixed loads. |
| SR034 | Bloom Energy | 2026 Data Center Power Report PDF | Over one-third of data centers are expected to use 100% onsite power by 2030. |
| SV001 | Emerald AI | Emerald AI raises $150 million Series A | Emerald AI announced it has raised $150 million in an oversubscribed Series A financing at a valuation of $1.05 billion. |
| SV002 | Emerald AI | Emerald AI $150 million Series A valuation investor quotes | Emerald says it now counts twelve Fortune Global 500 companies as investors. |
| SV003 | SEC | Emerald AI Form D primary document (August 2026) | The filing lists a $150,000,000 offering amount, $90,229,639 sold, and 23 investors. |
| SV004 | SEC | Emerald AI Form D filing (August 2025) | The August 2025 Form D shows 37 investors in a $35.3 million offering. |
| SV005 | NVIDIA | Emerald AI case study | Across five demonstrations, Emerald Conductor is proving AI factories can be power-flexible grid assets. |
| SV006 | Silicon Valley Power | SVP and Emerald AI launch pilot to demonstrate flexible data centers in Santa Clara | SVP framed the project as the first commercial, multi-megawatt DSX Flex deployment. |
| SV007 | National Grid | UK first trial of AI grid technology successfully demonstrates ability for data centres to adjust power needs | The UK trial showed data centres can adjust power needs without disrupting critical workloads. |
| SV008 | Salesforce Ventures | Welcome Emerald AI | Hyperscalers, data center operators, and utilities need innovative solutions for AI-scale workloads. |
| SV009 | Heatmap | Emerald AI, Nvidia, and the case for flexible data centers | Utilities still need to create enough value for flexible data center customers to care. |
| SV010 | S&P Global Market Intelligence | Power-flexible AI data center unveiled in Virginia, touted as template | Aurora is a flagship facility being built by Digital Realty with Emerald AI and NVIDIA technology. |
| SV011 | Data Center Dynamics | Nvidia-backed Emerald AI raises $24.5m to turn data centers into grid assets | Emerald Conductor could help data centers obtain grid connection significantly more quickly. |
| SV012 | Council on Foreign Relations | America may not need a massive energy build-out to power the AI revolution | Flexible data centers could connect swiftly to existing power grids without waiting up to a decade for new infrastructure. |
| SV013 | Berkeley Lab | Berkeley Lab report evaluates increase in electricity demand from data centers | Data centers could consume 6.7% to 12% of U.S. electricity by 2028. |
| SV014 | Berkeley Lab ETA | Data centers | Data center load growth has tripled over the past decade and is projected to double or triple by 2028. |
| SV015 | POWER Magazine | PJM widens response to data center load as capacity shortfalls deepen | PJM expects peak demand surging by about 32 GW between 2024 and 2030, roughly 30 GW from data centers. |
| SV016 | PureSky Energy | Key takeaways from NERC’s January 2026 long-term reliability assessment | NERC projected a 20-25% increase in nationwide peak demand over the next decade and warned of capacity shortfalls. |
| SV017 | Stock Analysis | Equinix stock price and overview | Equinix market cap was 106.53B and revenue (ttm) 9.90B. |
| SV018 | Stock Analysis | Equinix revenue | Equinix had annual revenue of $9.21B in 2025 and P/S ratio of 10.76. |
| SV019 | CompaniesMarketCap | Equinix market cap | Equinix had a market cap of $106.52B in August 2026. |
| SV020 | Stock Analysis | Digital Realty Trust stock price and overview | Digital Realty market cap was 73.05B and revenue (ttm) 6.76B. |
| SV021 | Stock Analysis | Digital Realty Trust revenue | Digital Realty had annual revenue of $6.11B in 2025 and P/S ratio of 10.81. |
| SV022 | Stock Analysis | Bloom Energy revenue | Bloom Energy had annual revenue of $2.02B in 2025 and P/S ratio of 20.64. |
| SV023 | CompaniesMarketCap | Bloom Energy market cap | Bloom Energy had a market cap of $64.26B in August 2026. |
| SV024 | Stock Analysis | Eaton Corporation stock price and overview | Eaton market cap was 162.91B and revenue (ttm) 30.03B. |
| SV025 | Stock Analysis | Eaton Corporation revenue | Eaton had annual revenue of $27.45B in 2025 and P/S ratio of 5.43. |
| SV026 | CompaniesMarketCap | Eaton market cap | Eaton had a market cap of $162.91B in August 2026. |
| SV027 | Stock Analysis | Vertiv Holdings Co revenue | Vertiv had annual revenue of $10.23B in 2025 with P/S ratio of 8.84. |
| SV028 | Stock Analysis | Vertiv Holdings Co stock price and overview | Vertiv market cap was 101.56B and revenue (ttm) 11.48B. |
| SV029 | Stock Analysis | Vertiv Holdings Co market cap | Vertiv had a market cap of $101.56B in August 2026. |
| SV030 | CompaniesMarketCap | Schneider Electric market cap | Schneider Electric had a market cap of $196.05B in August 2026. |
| SV031 | Newsweek | Emerald AI Has a New Approach to Meeting AI’s Energy Demand | Emerald AI has a new approach to meeting AI’s energy demand. |
| SV032 | Public Power | NVIDIA, Emerald AI, EPRI, PJM and others develop power-flexible AI factory | Aurora is a reference design and certification effort around a 96 MW AI factory. |
| SV033 | Emerald AI | Terms and Conditions | Forward-looking statements and pilot results are subject to risks and may differ materially from actual results. |