Ineffable Intelligence
获得创纪录融资的英国前沿 AI 实验室,押注经验式强化学习通向超级智能
Ineffable Intelligence 兼具欧洲最强创始人驱动前沿 AI 逻辑之一,也有最难承销的公开估值结构之一:种子轮 $5.1B、零收入、零客户,技术和治理风险仍未解开。
封面要素
公司概况
Ineffable Intelligence 是一家总部位于伦敦的前沿 AI 初创公司,由 David Silver 于 2025 年底创立,目标是打造一种强化学习「超级学习者」,能从经验而非人类生成数据中发现知识。公司于 2026 年 4 月公开亮相,同时宣布完成 11 亿美元种子轮,投后估值 51 亿美元,投资方包括 Sequoia、Lightspeed、NVIDIA、Google 及英国公共部门投资者。它仍是尚未产生收入的研究实验室,没有披露客户、财务报表、安全框架或商业化时间表,但已迅速拿下 NVIDIA 和 Google Cloud 的战略基础设施合作。
- 成立时间
- 2025-11-19
- 创始人
- David Silver
- 创立地点
- London, United Kingdom
- 总部
- London, United Kingdom
- 产品
- 一套仍处于商业化前阶段的强化学习研究平台和基础设施栈,用于训练能生成经验、评估结果并持续改进、且不依赖人类数据的「超级学习者」。
- 客户
- 潜在客户包括主权 AI 项目、超大规模计算合作伙伴,以及未来的科学、工程和企业研发用户;截至本次报告日期,尚无公开披露的付费客户。
- 商业模式
- 商业模式尚未公开;当前活动是研究、招聘、资本部署和基础设施建设,而非产品销售。
- 阶段
- seed-stage private
- 融资情况
- 2026 年 4 月 27 日完成 11 亿美元种子轮,投后估值 51 亿美元,是迄今报道的欧洲最大种子轮融资。
执行摘要
主要优势
- David Silver 是全球少数多次证明强化学习能打出前沿突破的研究者,这给公司带来异常强的创始人与市场匹配。
- 种子轮财团和基础设施组合堪称顶级:Sequoia、Lightspeed、NVIDIA、Google Cloud 和英国主权资本同时提供背书和算力入口。
- 作为一家英国前沿实验室,Ineffable 在欧洲占据差异化战略位置:它追求体验式强化学习,而不是又一条文本优先的 LLM 路线。
主要风险
- 公司没有披露收入、客户、烧钱速度、员工数、财务报表或商业路线图,常规承销几乎无从下手。
- 技术逻辑依赖强化学习从受限环境扩展到开放式发现;公开证据还没有验证这个前提。
- 公司核心依赖 David Silver、NVIDIA 硬件路线图、Google Cloud 基础设施和后续融资轮,集中度风险彼此叠加。
- 与美国领投方相比,英国公共部门投资人似乎治理杠杆有限,由此留下主权、知识产权控制和公共利益问题。
未决问题
- 当前烧钱速度、预算内算力支出、现金跑道假设,以及撑到首个技术里程碑的融资计划均未披露。
- 没有公开证据说明公司的里程碑框架、基准设计、安全治理流程或外部评估机制。
- 公司没有披露付费客户、试点、合同或定价界面,因此需求证明和商业化时间点仍属推测。
- 完整股权结构表、清算优先权、按比例跟投权、董事会控制条款和公共投资人治理保护均未公开。
目录
01公司概览
1.1 身份、产品与运营版图
Ineffable Intelligence Ltd(公司编号 16865241)于 2025 年 11 月 19 日在英格兰和威尔士注册为私人有限公司。其注册地址为 3rd Floor, 1 Ashley Road, Altrincham, Cheshire, WA14 2DT——这是英国新注册科技公司常见的专业服务地址;运营总部在伦敦,所有投资方、政府和新闻材料都一致引用这一地点。SIC 代码为 74909(其他未另分类的专业、科学及技术活动),符合公司尚未产生收入的研究状态。官方网站为 https://www.ineffable.ai/。 公司宣称的使命是通过创造所谓「超级学习者」来「与超级智能进行第一次接触」:一种能够从自身经验中发现全部知识的 AI 系统,从基础运动技能一直到深刻的智力突破。核心技术押注在于,超级智能将靠经验式学习实现,而不是靠学习人类数据,并由全球最强大的强化学习(RL)算法驱动。不同于在互联网规模语料上训练的大语言模型,Ineffable 的系统将实时生成并评估自身经验,对计算基础设施的互联、内存带宽以及训练—推理紧耦合循环提出极高要求。公司自注册后一直隐身运营,并在 2026 年 4 月 27 日宣布种子轮融资时同步公开亮相。 截至本次报告日期,Ineffable 没有披露产品、收入或客户基础;公司处在纯研究和基础设施阶段,这与其融资结构和网站使命表述一致。公司自称计划创造一段时间窗口,让「雄心勃勃的研究能够蓬勃发展,而不必屈从于渐进式产品和短期利润的要求」——这是对标准 VC 支持产品路线图的有意偏离。所有公开证据都把它指向一家前沿 AI 研究实验室;短期产出预期是科学突破,而非商业产品。 [CO001, CO002, CO003, CO004, CO005, CO006]
| 指标 | 数值 / 状态 | 日期 | 置信度 | 缺口 / 尽调要求 |
|---|---|---|---|---|
| 投后估值 | $5.1 billion | 2026-04-27 | 高 | 单轮估计;没有二级交易可交叉核验 |
| 累计股权融资 | $1.1 billion(仅种子轮) | 2026-04-27 | 高 | 无此前轮次;未披露债务或信贷 |
| 运营总部 | 英国伦敦 | 2026-06-22 | 高 | 注册地址为 Altrincham;运营总部依据所有新闻和公司来源 |
| 阶段 | 收入前研究实验室 | 2026-06-22 | 高 | 未披露产品、客户或收入 |
| 收入 / ARR | 未披露 | 低 | 私有信息;在数据室索取 | |
| 员工人数 | 未公开披露 | 低 | 留存来源中没有任何员工人数;在数据室索取 |
所有财务数据来自 2026-04-27 种子轮公告。估值仅为种子轮投后估值。收入和员工人数未公开;截至运行日期,公司是收入前研究实体。
[CO001, CO003, CO019, CO020, CO043]身份、研究使命、资本、基础设施合作和治理依赖如何在公司层面连起来。
[CO006, CO008, CO009, CO019, CO033, CO036]1.2 创始人、领导层与治理
David Silver 是 Ineffable Intelligence 的 CEO 和创始人。他曾任 Google DeepMind 强化学习副总裁,在创立 Ineffable 之前在 DeepMind 工作约二十年。他是 University College London(UCL)计算机科学教授,被广泛视为过去十年最有影响力的 AI 研究者之一。他在 DeepMind 参与的项目包括作为主要架构师或核心贡献者推动 AlphaGo、AlphaZero、AlphaStar、AlphaFold 和 AlphaProof——这些项目代表了强化学习和博弈求解 AI 史上最突出的展示。AlphaGo 在 2016 年 3 月以 4:1 击败 18 次世界围棋冠军李世石,比赛观看人数超过 2 亿。Silver 于 2026 年 1 月 16 日被任命为该法律实体董事。对于一家以 RL 为中心追求超级智能的实验室,他的创始人—市场匹配度极强:他在 Ineffable 所追逐的同一领域已有二十年经同行评议的发表成果。 Sesamers(引用 tech.eu)报道称,另有三名创始团队成员同样来自 DeepMind:Wojciech Czarnecki、Lasse Espeholt 和 Junhyuk Oh——均被描述为过去十年一直处在强化学习研究前沿。这些人没有出现在 Companies House 的董事记录中,但 Sesamers 的说法与 Silver 本人关于 Ineffable 团队将由「只为这一使命而来的杰出个人」组成的表述一致。在官方披露前,这一信息应视为二级来源报道,置信度中等。 种子轮完成时,治理结构迅速成型。Sequoia Capital 合伙人 Alfred Lin——地址列为 2800 Sand Hill Road, Menlo Park——于 2026 年 4 月 17 日获任董事。Ravi Mhatre 的地址列为 2200 Sand Hill Road, Menlo Park(Lightspeed 地址),同样于 2026 年 4 月 17 日获任董事。George Samuel Rose 是居住在加拿大的加拿大国民,也在 Companies House 记录中列为董事。Oakwood Corporate Secretary Limited 在公司注册时获任公司秘书。Sequoia 和 Lightspeed 合伙人在种子轮首日即进入董事会,符合共同领投方获得董事会席位的安排。关键人物依赖高度集中于 Silver;他主导科学愿景、外部叙事、资本形成故事和战略合作公告。若二级来源所称的创始团队得到确认,将降低但无法消除这种集中度。 [CO009, CO010, CO011, CO012, CO013, CO014]
| 人物 | 角色 / 头衔 | 背景 | 创始人—市场匹配 / 覆盖范围 | 关键人依赖 |
|---|---|---|---|---|
| David Silver | CEO 兼创始人 | UCL 教授;前 Google DeepMind 强化学习副总裁;AlphaGo、AlphaZero、AlphaProof 架构师 | 异常强——在 Ineffable 追求的同一领域深耕并发表强化学习研究已有二十年 | 关键;牵引科学愿景、对外叙事、融资和合作伙伴叙事 |
| Alfred Lin | 董事(2026-04-17 任命) | Sequoia Capital 合伙人;地址 2800 Sand Hill Road, Menlo Park | 领投方董事会代表;未披露运营角色 | 仅限董事会治理;职能依赖低 |
| Ravi Mhatre | 董事(2026-04-17 任命) | Lightspeed Venture Partners 合伙人;地址 2200 Sand Hill Road, Menlo Park | 领投方董事会代表;未披露运营角色 | 仅限董事会治理;职能依赖低 |
| George Samuel Rose | 董事 | 加拿大国籍,居住在加拿大;注册地址匹配,可能承担公司 / 法务职能 | 公司秘书 / 法务基础设施支持 | 运营依赖低 |
| Wojciech Czarnecki | 创始团队成员(二手来源报道) | 前 DeepMind 强化学习研究员;里程碑强化学习论文共同作者 | 强化学习研究纵深强——印证创始团队的技术班底 | 中等;仅二手来源——角色和正式头衔未确认 |
| Lasse Espeholt | 创始团队成员(二手来源报道) | 前 DeepMind 强化学习研究员 | 强化学习研究和工程纵深 | 中等;仅二手来源 |
Sesamers/tech.eu(二手来源)称,Czarnecki、Espeholt 和 Oh 是加入创始团队的前 DeepMind 校友;他们未出现在 Companies House 高管记录中。Alfred Lin 和 Ravi Mhatre 由 Companies House 以及 Sand Hill 地址对应的投资人身份推断确认。George Rose 的角色 依据地址和国籍模式推断;尚无官方头衔发布。
[CO009, CO010, CO011, CO012, CO013, CO014]1.3 融资、投资方与资本结构
2026 年 4 月 27 日,Ineffable Intelligence 宣布完成 11 亿美元种子轮,投后估值 51 亿美元——这是欧洲史上最大种子轮,也是全球 AI 史上规模最大的首轮外部融资之一。本轮由 Sequoia Capital 和 Lightspeed Venture Partners 共同领投。Cooley LLP(伦敦合伙人 Eric Davison)为公司融资提供法律顾问。按 UKTN 报道所隐含汇率,英镑等值约为 8.14 亿英镑。 参投方包括 NVIDIA、Google、Index Ventures、DST Global、EQT Ventures、Flying Fish Ventures、Evantic Capital、BOND Capital 和英国 Wellcome Trust,另有两家英国公共部门载体:British Business Bank(其自身新闻稿确认投资 2000 万美元 / 1480 万英镑)和 UK Sovereign AI Fund(金额「商业敏感」,政府披露显示其典型投资区间为 100 万至 1000 万英镑)。Sovereign AI Fund 和 British Business Bank 让英国以共同投资者身份进入由美国风投主导的轮次。英国科学技术大臣 Liz Kendall 和 AI 大臣 Kanishka Narayan 均发表背书,将这笔投资定义为英国能够成为「AI maker, not taker」的证明。 作为一家没有披露产品或收入的种子阶段公司,Ineffable 在这一轮之前没有融资历史;2026 年 4 月前的股权结构未知。51 亿美元投后估值对一家仅成立数月、没有商业产品的实体而言极高,使其在种子轮即进入「五角兽」状态(估值超过 50 亿美元的公司)。未见债务、信贷额度、老股转让或此前融资的报道或确认。公司承诺在一个「雄心勃勃的研究能够蓬勃发展,而不必屈从于渐进式产品和短期利润的要求」的窗口内运营,这将影响投资者预期商业回报的时间线。 David Silver 已公开承诺,将其 Ineffable 股权产生的任何个人收益 100% 通过 Founders Pledge 网络捐给高影响力慈善机构。Sesamers 报道称,按本轮隐含股权价值计算,这是 Founders Pledge 史上最大的一笔承诺。TechCrunch 和 Hotminute 也印证了这一承诺。Founders Pledge 网站本身不公布个人成员承诺。 [CO019, CO020, CO021, CO022, CO023, CO024]
| 利益相关方 | 角色 / 类别 | 控制权 / 经济重要性 | 尽调问题 |
|---|---|---|---|
| Sequoia Capital | 领投方;Alfred Lin 进入董事会 | 共同领投;通过 Alfred Lin 直接占有董事会席位;在 VC 中治理影响力最强 | 确认董事会席位条款、按比例认购权、信息权 |
| Lightspeed Venture Partners | 领投方;Ravi Mhatre 进入董事会 | 共同领投;通过 Ravi Mhatre 直接占有董事会席位 | 确认董事会席位条款;核验 Mhatre 身份与 Companies House 记录是否一致 |
| NVIDIA | 投资方 + 深度工程合作伙伴 | 财务持股叠加强化学习基础设施共设计;Jensen Huang 亲自背书合作 | 确认投资金额;审查共设计 IP 归属条款 |
| Google / Google Cloud | 投资方 + 首选云合作伙伴 | 财务持股;围绕 AI Hypercomputer 签订排他性首选云协议;部署 A5X / Vera Rubin | 确认投资金额;审查云合同排他性和切换成本 |
| British Business Bank | 公共共同投资方 | $20 M(£14.8 M)已确认;公开信息未显示董事会席位 | 确认 BBB 持股是否附带治理权 |
| UK Sovereign AI Fund | 公共共同投资方 | 金额未披露(“商业敏感”);下一轮优先拒绝权;典型区间约 £1–10 M | 索取准确投资金额;厘清是否存在可执行的英国锚定或 IP 条件 |
| Index Ventures、DST Global、EQT Ventures、Flying Fish、Evantic Capital、BOND Capital 等投资方 | 财团参与方 | 少数财务持股;未公开披露董事会代表 | 确认按比例认购权、反稀释权和信息权;核查无侧函 |
| UK Wellcome Trust | 战略 / 使命一致投资方 | 少数财务持股;与有益 AI 和健康突破论点一致 | 确认投资金额以及任何附带治理条件 |
除 British Business Bank(已确认 $20 M)外,所有参与方的投资金额均未公开披露。投资者名单 汇总自 British Business Bank 新闻稿、EU-Startups、Hotminute 和 Cooley LLP 新闻稿;这些来源 对具名参与方存在一些差异,此处呈现合并名单。
[CO019, CO020, CO021, CO022, CO023, CO024]种子阶段可公开获得的关键指标,凸显超常资本规模与收入前研究姿态之间的脱节。
[CO019, CO020, CO024, CO026, CO043, CO050]1.4 合作伙伴、基础设施与里程碑
种子轮公告后的数周内,Ineffable 快速推进前沿强化学习所需的计算基础设施。公司已公开宣布两项重大基础设施合作。 2026 年 5 月 13 日,Ineffable 与 NVIDIA 宣布开展工程层面的合作,共同构建大规模运行超级学习者系统所需的强化学习流水线。双方工程师正在共同设计训练基础设施,初期基于 NVIDIA Grace Blackwell 硬件,并计划成为最早探索 Vera Rubin 平台的团队之一。Jensen Huang 表示:「We are thrilled to partner with Ineffable Intelligence to codesign the infrastructure for large-scale reinforcement learning as they push the frontier of AI and pioneer a new generation of intelligent systems.」NVIDIA 已参投种子轮,因此既是投资方,也是深度工程合作伙伴。 2026 年 6 月 16 日,Ineffable 与 Google Cloud 在 Google Cloud Summit London '26 上宣布战略合作。根据协议,Google Cloud 是 Ineffable 的首选云服务商,将部署规模最大的 A5X 集群之一,底层由 NVIDIA Vera Rubin NVL72 驱动。合作中特别提到 Google Cloud 的 AI Hypercomputer 架构——一种 GPU、网络和存储的系统级整合——是其区别于标准 GPU 租赁的关键。David Silver 表示:「We chose Google Cloud as the best fit for our reinforcement learning infrastructure. We aren't just looking for processors; we are building a resilient and scalable environment to make 'first contact' with superintelligence.」Google Cloud CEO Thomas Kurian 也就获选发表了相应致辞。Google 同样参投了种子轮。 这些里程碑合在一起说明,Ineffable 在走出隐身状态后 60 天内,已为其研究计划锁定两项关键的非资本投入——计算基础设施和工程人才。UCL 关联和其作为伦敦 AI 锚点的定位,也被 UCL 在伦敦 AI 生态成长语境中引用。 [CO033, CO034, CO035, CO036, CO037, CO038]
| 日期 | 事件 | 类型 | 金额 / 估值 / 状态 | 参与方 | 含义 |
|---|---|---|---|---|---|
| 2025-11-19 | INEFFABLE INTELLIGENCE LTD 在英格兰和威尔士注册成立 | 创立 | 公司编号 16865241;SIC 74909 | Oakwood Corporate Secretary(同日任命秘书) | 法人实体成立;注册于 Altrincham 地址 |
| 2026-01-15 | David Silver 在 Ineffable 博客发布个人创始说明 | 创立 | 个人博客文章日期为 2026 年 1 月 15 日 | David Silver | Ineffable 存在的首个公开信号;界定使命和研究哲学 |
| 2026-01-16 | David Silver 被任命为董事 | 创立 | Companies House 备案 | David Silver | 正式治理计时开始;确认 Silver 自 2026 年 1 月初起就是主要负责人 |
| 2026-04-17 | Alfred Lin 和 Ravi Mhatre 被任命为董事 | 融资 | 交割前董事会成型 | Sequoia Capital(Lin)、Lightspeed(Mhatre)董事席位 | 公开宣布前已锁定领投方董事会席位;治理结构成型 |
| 2026-04-27 | 种子轮公布;公司走出隐身 | 融资 | $1.1 B,投后估值 $5.1 B | Sequoia、Lightspeed、NVIDIA、Google、Index、DST、EQT、Flying Fish、Evantic、BOND、Wellcome、BBB、Sovereign AI 等投资方 | 欧洲史上最大种子轮;创立首日即达 pentacorn 估值 |
| 2026-04-27 | 英国政府背书投资;Sovereign AI Fund 和 British Business Bank 共同投资 | 治理 | BBB $20 M / £14.8 M 已确认;Sovereign AI 金额未披露 | British Business Bank、Sovereign AI Fund、UK DSIT(Liz Kendall 声明) | 英国公共资本与主导美国 VC 并投;引发主权和治理疑虑 |
| 2026-04-30 | Newspage.news 发布治理疑虑评论 | 负面 | 仅为咨询评论;无法律行动 | Rohit Parmar-Mistry(Pattrn Data)、Katrina Young(KYC Digital)等治理评论人士 | 重大不利信号:评论者质疑公共资金是否买到主权或治理影响力 |
| 2026-05-13 | NVIDIA 与 Ineffable 宣布强化学习基础设施工程合作 | 合作伙伴关系 | 工程共设计;未披露 $ 金额 | NVIDIA、Ineffable Intelligence;引用 Jensen Huang 和 David Silver | 锁定 Grace Blackwell 和 Vera Rubin 上的深度软硬件合作;对大规模强化学习关键 |
| 2026-06-16 | Google Cloud 宣布成为首选云合作伙伴 | 合作伙伴关系 | 战略云协议;部署 A5X / Vera Rubin NVL72 集群 | Google Cloud(Thomas Kurian)、Ineffable Intelligence(David Silver);在 Cloud Summit London '26 宣布 | 确认算力基础设施策略;通过 Google Cloud AI Hypercomputer 与 Google Cloud 形成首选云锁定 |
日期来自 Companies House、官方新闻稿和已发布博客。创始博客说明日期(2026 年 1 月 15 日) 取自网站文本;该说明在 2026 年 4 月下旬网站上线时公开发布,但署期为 2026 年 1 月。
[CO001, CO013, CO014, CO015, CO019, CO020]从 2025 年 11 月注册成立到 2026 年 6 月 Google Cloud 合作的关键里程碑,展示公司如何在数月内从创立走到 pentacorn 状态,并拿下首批重大基础设施合作。
[CO006, CO007, CO020, CO026, CO028, CO035]1.5 负面信号与治理担忧
公开记录中最主要的负面信号不是运营层面,而是结构层面:英国政府以少数股权身份与占主导地位的美国风投共同投资后,AI 行业评论者围绕治理和主权提出了一组担忧。 Newspage.news(2026 年 4 月 30 日)刊登了两位 AI 顾问的评论。Rohit Parmar-Mistry(Pattrn Data)认为,「种子阶段的公共资金不应只换来一篇新闻稿和少数股权」,政府在没有治理权、下游用途透明度或可信保障的情况下参与投资,「有助于为上行风险去风险化,却让私人投资者捕获战略价值」。Katrina Young(KYC Digital)将英国政府的参与描述为对可能产生战略重要知识的系统拥有约 1% 的影响力,并警告称,「主权不是通过出现在股权表上实现的——它要靠控制、杠杆和问责来守住。」 Electronics Weekly(2026 年 4 月 28 日)也观察到,作为投资的一部分,英国政府得到的是一个伦敦地址、一小笔股权和下一轮优先购买权——但没有结构性保证,能确保 Ineffable 的发现、IP 或商业价值留在英国。 截至本次报告日期,保留证据中未发现诉讼、监管调查、领导层争议、制裁或产品召回。公司成立不到八个月,限制了负面事件的暴露面。治理担忧对长期公共利益和英国 AI 主权叙事具有实质意义,但目前并不构成公司的运营或法律风险。尽调应核实公共共同投资者实际持有哪些治理权;Sovereign AI Fund 将其投资条款描述为商业敏感。 [CO040, CO041, CO042, CO043, CO025]
1.6 展示项
02市场分析
2.1 市场边界、纳入支出与替代方案
界定 Ineffable Intelligence 的可寻址市场需要先划清边界,因为公司追逐的是一种能力——通向超级智能的自主强化学习——而不是成熟产品类别中的某个产品。三个重叠视角最有分析价值:(1)前沿 AI 训练基础设施,即训练全球最先进 AI 系统所需的硬件、软件和计算编排;(2)AI-for-science 和知识发现应用,在这些场景中,RL 及相关深度学习方法正在替代传统科学模拟和人类专家工作流;(3)主权和战略 AI 能力,英国、欧盟及盟友政府正把前沿 AI 能力作为国家安全和产业政策议题积极投入。 在没有商业产品的前提下,以下不应计入可寻址市场:(a)企业聊天机器人、推荐引擎、NLP-as-a-service 等通用 AI 应用,它们是下游消费者,不是买方;(b)作为大宗商品的 AI 硬件(NVIDIA 芯片是 Ineffable 的投入成本,不是它销售的市场);(c)面向 SME 或消费者买家的 AI 软件订阅。分析师引用的表面庞大的「全球 AI 市场」数字——从 2000 亿美元到 1 万亿美元以上不等,取决于方法——混合了这些被排除的类别,不能在没有边界逻辑的情况下作为 Ineffable 的 TAM。 Ineffable 所提出方案的现状替代品包括:基于人类整理数据训练的大语言模型(占主导的监督学习范式);AI-for-science 场景中的人类科学专家和领域专用仿真软件;以及 Google DeepMind、OpenAI、Anthropic 等超大规模厂商和前沿实验室的内部研究项目,这些机构同时最可能成为早期合作伙伴 / 买方,也是最有能力的 incumbent。纯 RL 流水线 incumbent 很少;David Silver 的团队曾是 DeepMind 中全球最突出的团队之一,因此替代威胁主要来自内部能力,而非外部供应商。 [CM001, CM002, CM003, CM004, CM005, CM006]
| 类别 | 纳入支出 | 排除支出 | 买方 / 付款方 | 与 Ineffable 的相关性 |
|---|---|---|---|---|
| 前沿 AI 训练基础设施 | GPU / 加速器集群;AI 优化网络和存储;共设计强化学习训练流水线 | 仅推理负载;通用云 VM;游戏 GPU | 超大规模云厂商、前沿实验室、主权算力计划 | 核心 TAM——超级学习器训练的直接算力买方 |
| AI for science / 知识发现 | 用于药物发现、蛋白质折叠、材料科学、基因组学的 AI 算力 | 面向终端消费者或 SME 销售的 AI 应用;通用 NLP 服务 | 药企研发预算、国家实验室、基因组公司、政府拨款 | SAM 邻近市场——强化学习能力可支撑科学发现项目 |
| 主权 / 战略 AI 能力 | 政府出资的前沿 AI 集群、国家 AI 实验室、国家支持算力 | 通用政府 IT 采购;非 AI 公有云支出 | DSIT(英国)、BEIS、EC AI Office、NAIRR(美国)、EuroHPC、盟友防务项目 | 核心 SAM——主权买方看重战略回报,而非商业 ROI |
| 通用 AI 应用市场 | 未纳入——这些是下游用例 | 企业聊天机器人、NLP API、推荐引擎、面向 SME 的 AI-as-a-service | 不适用 | 排除:这些买方不采购前沿强化学习训练系统 |
| AI 硬件和芯片 | 未纳入——NVIDIA、AMD 芯片是 Ineffable 的成本投入 | GPU 制造、芯片设计收入(Nvidia、AMD、Intel) | 不适用 | 排除:组件市场,不是 Ineffable 可触达的收入市场 |
| 现状替代方案 | 监督学习 / LLM 训练(当前主导范式) | 人类研究员、传统仿真、规则系统 | 所有前沿 AI 买方目前都使用监督学习 | 切换成本高:既有流水线、工具和人才都围绕 LLM |
支出类别和买方分类基于代理证据构建,包括 NVIDIA 财报、英国政府披露、EU AI Act 监管框架和 分析师市场报告。没有公开来源直接测算“前沿强化学习超级学习器”市场规模;这些类别体现尽调团队 的边界逻辑。“纳入”和“排除”是评估判断,不是来源给出的数字。
[CM001, CM002, CM003, CM004]2.2 市场规模:多重视角与相互矛盾的估算
没有一个已发布数字能干净衡量 Ineffable 正进入的市场。下文使用三个视角,并明确方法,同时保留矛盾。 视角一——AI 计算基础设施(最直接相关的代理指标)。NVIDIA 数据中心业务——全球领先 AI GPU 供应商的主导收入线——在 FY2026 第一季度(截至 2025 年 4 月 27 日的季度)实现 391 亿美元收入,同比增长 73%,环比增长 10%。年化后,仅 NVIDIA 数据中心收入就约为每年 1570 亿美元;这本身仍低估总 AI 计算支出,因为它排除了 AMD、Google TPU、AWS Trainium 和定制芯片。这是覆盖所有买方层级、同时包含训练和推理计算支出的最宽指标,并不只包括 Ineffable 所瞄准的前沿研究。尽管美国对中国 H20 产品出口管制带来约 80 亿美元逆风,NVIDIA 仍指引 FY2026 第二季度总收入约为 450 亿美元。 视角二——投资 / 资本开支流。Goldman Sachs Research 预测,到 2025 年,全球 AI 相关投资将接近 2000 亿美元,美国将在大型企业采用中领先,尤其是信息服务和专业 / 科学服务领域。Goldman Sachs 同时指出一个重要约束:AI 对生产率的影响可能要到本十年后半段才显现——这会实质性制约企业 ROI,也会影响超大规模厂商之外的可自由支配 AI 支出节奏。 视角三——前沿 RL 训练计算(最贴近 Ineffable 的目标)。Epoch AI 的研究显示,自 2010 年以来,前沿 AI 模型训练计算量每年增长 4–5 倍;近年发布时算力排名前十的前沿模型约每年增长 5 倍。训练单个前沿模型如今大约需要 10²⁴–10²⁵ FLOPs 计算量,每次训练成本估计在 5000 万美元到数十亿美元之间。关键是,Epoch AI 还指出 AlphaGo 和 AlphaGo Zero 等 RL 重型模型是计算离群值——其训练消耗的计算量远高于同时代典型深度学习模型——这意味着纯 RL 超级学习者的训练成本可能高于传统前沿模型的趋势线。 矛盾估算:新闻稿中引用的广义 AI 市场 TAM,从 Goldman Sachs 约 2000 亿美元投资流,到 Statista 对 2030 年数千亿美元市场的展望不等,但方法不可比(支出 vs. 投资 vs. 收入 vs. 经济影响)。没有任何一个数字单独隔离经验式 RL 训练这一类别。超级学习者系统的具体 SAM 无法从独立来源获得;本文的规模分析由代理指标构建,应视为高不确定性估算。 [CM007, CM008, CM009, CM010, CM011, CM012]
| 发布方 / 来源 | 年份 | 地理范围 | 数值 / 指标 | CAGR / 增长 | 方法论 | 置信度 | 对 Ineffable 的局限 |
|---|---|---|---|---|---|---|---|
| NVIDIA(新闻稿) | 2025(FY2026 Q1) | 全球 | 单季数据中心收入 $39.1B | 同比 +73%,环比 +10% | 公开财报;训练 + 推理合并 | 高 | 包含推理;高估纯训练强化学习市场 |
| Goldman Sachs Research | 2024 | 全球 | 到 2025 年 AI 投资接近 ~$200B | 未说明 | 分析师投资流模型;包括资本开支、并购、研发 | 高 | 投资 ≠ 收入;生产率影响推迟到本十年后期 |
| Statista Market Outlook 市场展望 | 2026 | 全球 | AI 市场预计到 2030 年将显著增长 | 预计两位数 CAGR | 自上而下,并以自下而上验证;B2B、B2G、B2C 合并 | 中 | 高度聚合;混合应用、基础设施和服务 |
| Epoch AI(研究) | 2024 | 全球 | 前沿训练算力每年增长 4–5× | 每年 4–5× | 基于前沿模型的自下而上算力估算;经验趋势 | 高 | 衡量算力而非支出;趋势中未纳入强化学习异常值 |
| British Business Bank / UK Sovereign AI Fund 公共资本 | 2026 | 英国 | $20M BBB + 未披露 SAIF(估计 <£10M) | N/A(拨款 / 股权) | 直接公共投资;单笔交易证据 | 高 | 相对全球算力市场,英国公共投资规模小;存在主权溢价 |
| Ineffable Intelligence / NVIDIA 合作 | 2026 | 全球 | Google Cloud 上的 A5X(Vera Rubin NVL72)集群——规模居前 | N/A | 新闻稿;定性规模参考 | 中 | 集群规模和资本开支未披露;仅作代理 |
所有数字都是代理指标,并非“超级学习器强化学习”类别的直接测量市场规模。NVIDIA 数据中心收入 (训练 + 推理、所有 AI 负载)是可触达前沿算力支出最接近的代理,但包含许多不在 Ineffable 范围内的 负载。Goldman Sachs 的 $200B 是投资流(累计或年度未说明);同一来源明确将生产率和 ROI 证据 推迟到后期。Epoch AI 算力趋势按模型衡量,不是整体市场规模。英国主权投资数字是已确认的下限值。
[CM007, CM008, CM009, CM010, CM011, CM013]三层规模测算框架,从宽口径 AI 算力基础设施市场下探到 Ineffable 可信的近期机会,所有数值都受证据约束。
TAM 数字由 NVIDIA Q1 FY2026 数据中心收入($39.1B)年化得出;不含非 NVIDIA 算力,且推理口径低估了训练专项支出。SAM 是尽调团队估算,没有独立来源;区间($20–50B)反映主权项目预算和 hyperscaler RL 配额的不确定性。收入前阶段无法直接测量 SOM。
[CM007, CM010, CM016, CM015]来自多个独立公开来源的 AI 基础设施市场规模或投资低 / 基准 / 高估计,全部以十亿美元计。
所有数值均为十亿美元。区间代表已发布的低 / 中 / 高值或分析师置信区间;没有区间时,则采用尽调团队解读。各来源方法不兼容(年收入 vs. 累计投资 vs. 单次训练成本),不得相加。Statista 数字被标记为付费墙且高度聚合,只能作方向性参考。
[CM007, CM008, CM009, CM011, CM023]2.3 买方、用户与付款方分层
Ineffable 的买方图景与标准 B2B SaaS 市场截然不同,因为截至本次报告日期,公司没有产品、没有客户、没有收入。因此相关分层只能是前瞻性的:如果 Ineffable 取得技术进展,谁会资助或采用超级学习者系统? 第一层——超大规模厂商和云服务商。Google、Microsoft、Amazon 和 Meta 共同控制着全球最大的 AI 训练集群,也是前沿 AI 模型的主要建设者。Google Cloud 已与 Ineffable 建立首选供应商关系,使其既是基础设施合作伙伴,也是 RL 能力的潜在买方。NVIDIA 作为股权投资者和工程共同设计方,也拥有独特的双重角色。这些机构有预算(每年数百亿美元级资本开支)、有整合超级学习者系统的技术成熟度,也有在竞争对手之前获取能力的战略动机。采用触发点将是自主知识发现能力在规模上的可验证进展。 第二层——主权和国家支持实验室。英国的 British Business Bank(2000 万美元)、Sovereign AI Fund 和美国 National AI Research Resource(NAIRR)代表了一个独立买方层级;其中采购由国家战略目标驱动,而非商业 ROI。欧盟的 AI factories 和 EuroHPC 计划是欧洲对应项。NVIDIA 在 ISC 2026 宣布欧洲将新增 35 台 NVIDIA AI 超级计算机,包括 Forschungszentrum Jülich 的 JUPITER,说明主权基础设施投资正在规模化落地。对这一买方层级而言,付款方是政府项目或公共研究机构,采用触发点是地缘政治紧迫性和科学雄心,而非短期商业回报。 第三层——前沿 AI 实验室。其他前沿实验室(Anthropic、OpenAI、xAI、Mistral 等)是 Ineffable 的主要竞争对手;不过,Ineffable 开发的 RL 特定能力也可能通过授权或协作方式部署,而非直接竞争。这些机构的计算预算相对超大规模厂商受限,近期不太可能成为大规模付款方。 第四层——科学研发机构。AlphaFold 的成功——该项目由 Google DeepMind 在 David Silver 更广泛的 RL 研究体系下开发——证明制药公司、国家实验室和基因组学机构对 AI-for-science 能力存在真实需求信号。这些买方按研究经费周期运作,当科学价值具有变革性时,对计算成本不那么敏感。但 Ineffable 的超级学习者论点与可立即部署的科学工具之间仍有很大距离;这个买方层级至多属于中期机会。 企业研发是投机性的长期层级,需要先有产品和业绩记录,采购才可能开始。 [CM017, CM018, CM019, CM020, CM021, CM022]
| 分层 | 买方 / 用户 | 付款方 | 工作流契合 | 预算负责人 | 采用触发点 |
|---|---|---|---|---|---|
| 超大规模云厂商 / 云服务商 | Google、Microsoft、Amazon、Meta AI 等超大规模云厂商 | 云资本开支预算(各自每年数百亿美元) | 强化学习研究基础设施合作 / 授权 | 首席技术官 / AI 研究副总裁 | 已展示大规模自主知识发现 |
| 主权 / 国家支持实验室 | UK DSIT、BBB、EU AI Office、NAIRR、EuroHPC 成员 | 政府 AI 基金 / 拨款项目 | 战略合作或研究拨款 | 国家 AI 项目主任 | 国家安全 / 战略 AI 自主叙事 |
| 前沿 AI 实验室(潜在合作方) | Anthropic、OpenAI、xAI、Mistral 等前沿 AI 实验室 | 投资人支持的资本开支 | 强化学习 IP 授权或联合研究 | 研究负责人 / 董事会 | 相对监督学习存在强化学习能力缺口 |
| 科学研发组织 | 药企、国家实验室、基因组研究所 | 研发预算 + 政府研究拨款 | 面向特定领域的 AI for science 应用 | 研究主管 / 首席科学官 | 经过同行评审的科学发现强化学习演示(AlphaFold 类比) |
| 企业研发(推测) | 拥有研发实验室的大型工业公司 | 研发预算(资本开支中的小部分) | 最终将超级学习器输出部署为决策工具 | CTO / 首席数据官 | 商业产品路线图 + 证据基础——目前缺失 |
| NVIDIA / Google Cloud(基础设施共同投资方) | NVIDIA(投资方 + 工程合作伙伴);Google Cloud(首选提供商) | 合作伙伴基础设施承诺;实物算力 | 共设计强化学习训练流水线;供应集群 | 合作伙伴 / OEM 销售负责人 | 已经激活——NVIDIA 和 Google Cloud 合作已确认 |
分层是前瞻性的:截至运行日期,Ineffable 没有收入或客户。预算规模是数量级估算,来自公开的 超大规模云厂商资本开支披露和英国政府投资公告。“采用触发点”描述每个分层从观察者转为付款方 所需的里程碑。NVIDIA 和 Google Cloud 单列,是因为它们兼具基础设施提供商与生态合作伙伴 / 投资者的独特双重角色。
[CM017, CM018, CM019, CM020, CM021, CM022]在本报告运行日 Ineffable 仍处收入前研究姿态的前提下,对各买方细分市场在五个采用维度上的评分评估(1=很低,5=很高)。
评分(1–5)是尽调团队基于公开证据做出的评估;未获独立来源验证。风险 / 障碍评分为反向指标(5=最高障碍)。预算规模反映可用 AI capex 的相对数量级,不是绝对数字。
[CM017, CM018, CM019, CM022, CM024, CM025]2.4 增长驱动因素与采用约束
AI 计算基础设施市场正经历前所未有的需求增长。NVIDIA CEO Jensen Huang 表示,AI 推理 token 生成量一年内增长了十倍;NVIDIA FY2026 第一季度业绩也确认,数据中心需求「incredibly strong」,短期看不到需求饱和。NVIDIA 在沙特、阿联酋和欧洲扩展 AI factories,证明主权买方正在把政治承诺转化为真实计算采购。英国 AI Opportunities Action Plan(2025 年 1 月)和在 Seoul Summit 签署的英国 Frontier AI Safety Commitments 表明,英国政府政策与支持前沿 AI 发展一致,为 Ineffable 的融资提供了需求信号和合法性框架。 RL scaling 正显示实际进展:Kimi k1.5 论文(2025 年 1 月)证明,将 RL 应用于 LLM 时,即使没有 Monte Carlo tree search 或 process reward models,也能在多个推理基准上匹配 OpenAI 的 o1。AlphaFold 的原子级蛋白质结构预测让 RL 和深度学习成为科学研究中不可或缺的工具。这些技术概念验证在监督学习等价物层面验证了 RL 投资论点,但尚未验证 Ineffable 所追求的自主经验式学习。 约束端有五个结构性障碍。第一,EU AI Act 的 GPAI 义务已于 2025 年 8 月 2 日生效,对高能力前沿模型引入安全评估、版权披露要求和系统性风险缓释;这会给 Ineffable 在欧盟的任何商业部署增加合规成本和法律不确定性。第二,美国对 NVIDIA H20 芯片的出口管制造成 FY2026 第一季度约 45 亿美元费用和约 70 亿美元被阻断或损失的收入,证明地缘政治供应冲击可实质打乱 AI 基础设施计划,与需求无关。第三,RL 训练天然比同等监督学习训练更消耗计算(Epoch AI 分析显示 AlphaGo 系列系统是计算离群值)。第四,通用 AI 投资的 ROI 不确定:Goldman Sachs Research 明确指出生产率影响可能推迟到本十年后半段,引发对可自由支配支出节奏的疑问。第五,Ineffable 没有任何商业产品、收入,甚至没有可展示的研究里程碑,这意味着截至本次报告日期,整个采用路径都是推测性的;英国政府稀薄股权(且没有可执行 IP 或主权条件)带来的治理担忧,也会增加合法性风险。 [CM026, CM027, CM028, CM029, CM030, CM031]
| 因素 | 方向 | 时间 | 对 Ineffable 的含义 | 尽调问题 |
|---|---|---|---|---|
| AI 算力需求指数级增长(推理同比增长 10×) | 驱动因素 | 当前——NVIDIA FY2026 Q1 已确认 | 前沿训练基础设施正在加速建设;Ineffable 受益于顺风 | 确认 Ineffable 的集群访问权已承诺,而不是取决于未来融资 |
| 主权 AI 投资(英国、欧盟、海湾国家) | 驱动因素 | 近期(2026–2028) | 政府正在成为买方;英国 BBB / SAIF 投资确认政治背书 | 对照 Ineffable 算力需求,绘制主权项目预算规模 |
| 强化学习扩展产出商业结果(Kimi k1.5、o1 系列) | 驱动因素 | 2024–2025 年已有证据 | 验证强化学习投资论点;降低投资人 / 买方对该路线的怀疑 | 核验 Ineffable 的强化学习架构是否能对标已发布的 RL-LLM 基准 |
| AI for science 需求(AlphaFold 类比) | 驱动因素 | 当前——AlphaFold 已被全球药企 / 研究机构采用 | 科学研发买方层级已活跃;若完成演示,超级学习器可承接类似需求 | 识别 Ineffable 计划早期部署或合作的具体科学领域 |
| EU AI Act GPAI 义务(2025 年 8 月生效) | 约束 | 当前生效——适用于欧盟前沿模型 | 任何欧盟商业部署都会增加安全、版权和系统性风险合规负担 | 评估 Ineffable 研究阶段是否触发 GPAI 义务;获取范围法律意见 |
| 美国对 AI 芯片的出口管制(H20、A100 变体) | 约束 | 已生效——仅 FY2026 Q1 即影响 NVIDIA 收入约 $7B | 即便已有合作承诺,供应冲击也可能扰乱计划中的算力采购 | 确认 Google Cloud 集群承诺在出口管制情景下是否稳固;检查合同保护 |
| 强化学习算力强度高于监督学习 | 约束 | 结构性 / 持续 | AlphaGo / AlphaZero 是算力异常值;超级学习器成本可能达到同类 LLM 运行的 10–100× | 索取 Ineffable 第一阶段训练运行的内部算力预算预测 |
| ROI 不确定且生产率影响滞后 | 约束 | 中期(2026–2030) | Goldman Sachs 指出 AI 对生产率的影响推迟到本十年后期;企业买方采用缓慢 | 假设 Ineffable 在 2030 年前没有商业产品,建模收入 / 授权路径 |
方向和时点是基于运行日已发布证据作出的判断。「约束」项代表重大风险,不代表确定结果。出口管制影响来自 NVIDIA 2026 财年一季度公开财报。RL 计算强度参考 Epoch AI 对 AlphaGo 计算异常值的分析。ROI 判断参考 Goldman Sachs Research。治理和安全约束参考《EU AI Act》官方监管文本。
[CM026, CM027, CM028, CM029, CM030, CM031]分阶段采用路径,展示 Ineffable 在每一阶段必须达成什么,才能打开买方漏斗的下一层级,从当前研究阶段走向商业收入。
数值代表各融资层级中前沿 RL 实验室推进到下一阶段的大致比例,基于既往前沿 AI 队列(DeepMind、OpenAI、Anthropic)的观察成功率。它不是对 Ineffable 的具体预测。
[CM026, CM027, CM029, CM036, CM037]2.5 展示项
03竞争者
3.1 竞争格局概览
Ineffable Intelligence 位于三个重叠竞争类别的交汇处。第一,它是与 OpenAI、Anthropic、Google DeepMind、xAI 等共享 AGI 使命的前沿 AI 实验室的直接范式竞争者——这些机构都在朝越来越自主、越来越强大的 AI 系统推进,只是技术架构不同。第二,它是超大规模厂商内部研究团队的替代项:尤其是 Google DeepMind,它是 Ineffable 创始人的制度来源,并继续在 Alphabet 的计算和分发基础设施内运营全球最深厚的强化学习项目。第三,它的英国 / 欧盟主权 AI 定位,使其成为 Mistral AI 的相邻竞争者;后者按估值(2024 年约 60 亿美元)和已部署产品组合计算,均是欧洲领先的前沿实验室。 一个潜在直接竞争者的独立类别——据行业评论者报道,包括 AMI Labs 和 Recursive Superintelligence——被称为正在追逐相关 AGI 或自我改进系统范式。截至本次报告日期,这两个实体都缺少足够公开披露(官方网站、融资公告或发表研究),无法做有意义画像;它们的存在既代表该范式可信度的先例,也代表证据缺口。 现状和内部自建替代方案与具名竞争者同样重要:未来可能部署超级学习者级系统的大型企业或政府,可以选择内部构建(通过超大规模厂商 API 访问前沿模型)、向 OpenAI 或 Anthropic 采购,或继续使用当下基于 LLM 的工具——这些都构成 Ineffable 必须从需求端替代的选项。 竞争定位图(FP001)把主要竞争者放在两个轴上:商业部署成熟度,以及强化学习 / 经验式学习聚焦度。Ineffable 在 RL 聚焦度上最高,在部署成熟度上最低——与 OpenAI 构成战略镜像。作为研究实验室,这一象限位置有防御性;但若 incumbent 加大 RL 投入,同时继续握有分发优势,它在结构上会变得脆弱。 [CP025, CP026, CP027, CP028, CP029]
每个竞争者按两个序数轴定位:RL / 体验式学习聚焦度(y 轴,1=仅监督 / LLM,10=不依赖人类数据的纯体验式 RL)和商业部署准备度(x 轴,1=仅研究,10=广泛商业部署)。评分是有证据支撑的序数估计,不是公开基准;轴标签包含在标题和注释中。
序数轴评分来自作者对各实体已发布产品成熟度、RL 研究姿态和分发证据的评估。评分不来自任何公开指数或排名;只能作为框架用的方向性定性估计。
[CP025, CP026, CP030]3.2 前沿实验室竞争者画像
OpenAI 是全球商业化最成熟的前沿 AI 实验室。公司成立时的使命是确保 AGI 造福全人类,如今以 OpenAI Group(公共利益公司)形式运营,由非营利组织 OpenAI Foundation 治理。2024 年 10 月,OpenAI 完成 66 亿美元融资,投后估值 1570 亿美元——这是截至当时史上最大单笔私营科技融资。OpenAI 的产品组合包括 ChatGPT(据报每周活跃用户超过 2 亿)、GPT-4o、o1/o3 推理链家族、DALL-E、Sora 和 OpenAI API,并公布了基于 token 的定价。OpenAI 的 RL 投入集中在基于人类反馈的强化学习(RLHF)和思维链过程奖励模型(o1/o3 家族展示了这一点)——该范式仍依赖人类生成训练数据,与 Ineffable 的无数据方法在结构上不同。Microsoft 合作(130 亿美元投资、Azure 集成)给了 OpenAI 企业分发和计算能力,Ineffable 无法复制。 Anthropic 由 Dario Amodei、Daniela Amodei 及离开 OpenAI 的同事于 2021 年创立。其企业使命明确绑定 AI 安全:公司公开承认自己可能正在「building one of the most transformative and potentially dangerous technologies in history」,并坚持安全导向实验室必须站在前沿的论点。截至 2025 年 1 月,Anthropic 据报正在以约 600 亿美元估值融资。Amazon 已承诺最高投资 40 亿美元。其产品组合包括 Claude 3 家族(Haiku、Sonnet、Opus)、Claude.ai 和公布价格的 API。Constitutional AI 框架和 Responsible Scaling Policy 构成了行业内最系统化的公开安全机制。Anthropic 的技术范式以 LLM 为先,并用 RLHF 增强对齐——并非没有人类数据的经验式 RL。 Google DeepMind 是 Alphabet Inc. 旗下整合后的 AI 组织,由 Google Brain 和 DeepMind 于 2023 年合并而来——David Silver 在这里工作了二十年。按其自身描述,它拥有「unparalleled computing infrastructure」,可访问 Google TPU pods,规模不是独立初创公司所能匹敌;研究组合覆盖 Gemini、AlphaFold、AlphaProof、机器人和 AGI 对齐。Google Search、Android 和 Google Cloud Platform 带来的全球分发,使其消费者和企业触达远超任何独立实验室。作为 Silver 领导下产出 AlphaGo、AlphaZero 和 AlphaProof 的机构,它拥有该领域最深厚的公开 RL 记录。 xAI 由 Elon Musk 于 2023 年创立,并于 2025 年 3 月以约 800 亿美元估值融资 60 亿美元。其 Grok 聊天机器人嵌入 X(原 Twitter)平台(约 5 亿注册用户),形成独特的消费者分发渠道。xAI 的 Colossus GPU 集群(据报约 100,000 块 H100 GPU)和「truth-seeking」AI 定位,使其区别于安全对齐型实验室。xAI 尚未发布可与 OpenAI system cards 或 Anthropic RSP 相比的正式安全框架。 Mistral AI 于 2023 年在法国创立,创始人来自 Google DeepMind 和 Meta Fundamental AI Research;公司在 2024 年 6 月以约 60 亿美元估值融资约 6.4 亿美元。其产品组合包括 Mistral Large/Small API 模型、开源权重发布(Mistral 7B、Mixtral 8x7B)和 Le Chat(面向企业和消费者的 AI 助手,支持本地部署和主权云部署选项)。按估值和产品广度计算,Mistral 是欧洲领先的前沿实验室;其对欧盟监管的适配,使其成为 Ineffable 面向欧盟和英国主权 AI 买方的主要相邻竞争者。 [CP001, CP002, CP003, CP004, CP005, CP007]
| 竞争者 | 类别 | 融资 / 估值(最新披露) | 主要目标客群 | 核心差异点 | 相对 Ineffable 的主要短板 |
|---|---|---|---|---|---|
| Ineffable Intelligence | 直接竞争 — 范式开创者 | 已融资 $1.1 B / 投后估值 $5.1 B(2026 年 4 月) | 前沿研究机构、主权 AI 买家、未来 AGI 被授权方 | 不用人类数据的体验式 RL;David Silver 创始人与市场匹配;英国主权定位 | 没有产品、没有收入,除创始人履历外没有业绩记录;范式落地前执行风险 |
| OpenAI | 直接竞争 — 范式 + 在位者 | 融资 $6.6 B / 估值 $157 B(2024 年 10 月;后续很可能继续融资) | 企业开发者、消费者、政府,通过 API 和 ChatGPT 触达 | ChatGPT 周活用户 200 M+;GPT-4o / o1 / o3 系列;Microsoft Azure 分发;已发布定价 | RL 范式是 RLHF / chain-of-thought,而不是无数据体验式 RL;商业化压力压缩纯研究机动性 |
| Anthropic | 直接竞争 — 范式 + 安全在位者 | 2025 年 1 月报道称估值约 $60 B;Amazon 已承诺投入 $4 B | 企业、安全敏感型开发者、受监管行业买家 | Constitutional AI;Responsible Scaling Policy;Claude 3 系列;Amazon AWS 分发 | LLM 优先 / RLHF 架构,不是体验式 RL;没有开放权重模型;分发重心在美国 |
| Google DeepMind | 在位者 + 内部自建替代 | Alphabet 子公司 — 未独立融资;Alphabet 市值约 $2 T | Google 消费产品、GCP 企业、全球 AI-for-science 买家 | TPU 算力无可匹敌;AlphaGo/AlphaZero/AlphaFold 的 RL 履历;全球产品分发;David Silver 的来源机构 | 快速跟进风险:可能在内部重建 Ineffable 的项目;Alphabet 内部优先级竞争;创业公司机动性较弱 |
| xAI | 相邻竞争 — 范式 + 消费端 | 已融资 $6 B / 估值约 $80 B(2025 年 3 月) | X 平台用户、企业 Grok API 消费者、美国科技买家 | Grok 嵌入 X(500 M 用户);Colossus 约 100 K H100 集群;Musk 品牌分发 | RL 重心不同(LLM 优先,不是体验式);安全姿态较弱;美国监管集中风险;没有欧盟主权定位 |
| Mistral AI | 相邻竞争 — 欧洲主权 AI | 已融资约 $640 M / 估值约 $6 B(2024 年 6 月) | 欧洲企业、开发者、要求欧盟驻留的主权 AI 买家 | 开放权重模型(Mistral 7B、Mixtral);Le Chat 企业助手;对齐 EU AI Act;法国 / 欧洲锚点 | LLM 优先架构,没有体验式 RL;算力足迹更窄;模型质量上限低于 OpenAI/Anthropic |
融资和估值数据来自各实体最新引用来源,到 2026 年 6 月运行日可能已过时——OpenAI(2024 年 10 月之后)和 Anthropic(2025 年 1 月之后)尤其如此。Google DeepMind 是子公司,估值未独立披露。所有数字均为近似值;标注 「已报道」或「可能」的单元格反映二级来源证据。
[CP001, CP003, CP007, CP009, CP012, CP014]| 实体 | 产品 / 层级 | 定价模型 | 已发布价格点(截至运行日) | 关键限制 | 对 Ineffable 的含义 |
|---|---|---|---|---|---|
| OpenAI | API — 全模型范围(GPT-4o、o1、o3-mini 等) | 按百万输入 / 输出 token 收费;按模型分层 | 已在 openai.com/api/pricing 发布;随模型而变(例如此前定价中,GPT-4o 每 MTok 输入 / 输出为 $2.50/$10;需查看当前页面) | 标价;实际企业定价可能不同;费率频繁变化 | 为 AI API 定价设定市场参照;Ineffable 没有可比产品或定价 |
| Anthropic | API — Claude 3 Haiku、Sonnet、Opus、Claude 4 系列 | 按百万输入 / 输出 token 收费;按模型分层 | 已在 anthropic.com/pricing 发布(例如截至 2026 年 6 月的 Claude 3.5 Sonnet 层级、Opus 4 层级) | 可用提示缓存、延长 TTL 定价、批量折扣;企业定制定价 | 证明聚焦前沿安全的实验室也能发布并维持商业定价——Ineffable 若要部署,也必须搭出这种模型 |
| Mistral AI | API — Mistral Large、Mistral Small、Codestral、Pixtral;Le Chat 企业版 | 按 token 的 API 定价,叠加 Le Chat 企业订阅 / 本地部署 | 企业定价需联系销售;API 定价已在 mistral.ai 发布;因只能通过 JS 访问,未提取具体费率 | 企业和本地部署选项让公开比较中的定价不透明 | Mistral 的欧盟主权定位意味着竞争不只靠价格,也靠价值观和数据驻留——这是 Ineffable 未来商业策略最直接的参照 |
| Google DeepMind (Gemini) | Gemini API 通过 Vertex AI 和 AI Studio 提供 | 通过 Google Cloud 按 token 定价;AI Studio 提供免费层 | 通过 Google Cloud 定价页发布(本轮未直接抓取) | 深度嵌入 Google Cloud 定价;企业定价通过 GCP 合同 | Google 的规模使其能够补贴定价,独立实验室无法匹配 |
| Ineffable Intelligence | 截至运行日没有商业产品或 API | 未披露 / 研究阶段 — 未宣布定价模型 | N/A — 不适用 | 公司已承诺近期不推出商业产品的窗口期 | Ineffable 没有可比较的定价策略;这构成任何收入或货币化分析的阻断性证据缺口 |
定价数据来自本轮抓取的公司官方定价页;现货价格仅作指示,且频繁变化。写有「未提取」或「未直接抓取」 的单元格反映抓取时遇到的访问限制。Ineffable 行用于明确记录定价缺失,而不是作估算。
[CP005, CP010, CP024, CP038]3.3 能力与功能对比
在五家主要竞争者中,最重要的单一能力维度是技术范式边界:五家均以人类生成互联网数据上的监督预训练作为模型架构基础,并把 RL 作为对齐和推理增强层。没有一家追逐 Ineffable 所提出的、从零开始且不使用人类数据的经验式 RL 范式。这意味着 Ineffable 占据了能力矩阵中目前没有 incumbent 填补的空白格——这既可能是蓝海位置,也同时是验证风险位置,因为该范式尚未在开放世界领域完成规模化验证。 在商业 API 可用性上,Ineffable 为零:OpenAI 公布了六个模型的按 token 定价,Anthropic 公布了三个模型,Mistral 公布了多个层级。Google DeepMind 通过 Google Cloud 的 Vertex AI 和 AI Studio 提供 Gemini。xAI 向开发者提供 Grok API。这些都代表 Ineffable 当前不具备的商业成熟度;要达到这一水平,需要多年研究成功。 在安全框架上,OpenAI(system cards、使用政策)、Anthropic(Constitutional AI、RSP、Frontier Red Team)和 Google DeepMind(Responsible AI team)均有公开、可审计的安全架构。截至本次报告日期,Ineffable 没有任何公开框架。作为一家在 AI Seoul Summit 承诺下运营的英国实验室,这一缺口在任何部署前都需要补上。 在开放权重模型可用性上——这对希望控制模型权重的主权 AI 买方越来越重要——只有 Mistral 将其作为核心产品类别。OpenAI、Anthropic、Google DeepMind 和 xAI 都不公开发布前沿权重。Ineffable 在这一维度上没有任何公告。 功能 / 能力矩阵(FP002)提供按竞争者和能力划分的视图;不支持的单元格明确标记。矩阵确认,Ineffable 唯一的近期竞争差异化在于经验式 RL 范式本身——所有其他能力维度都明显有利于 incumbent。 [CP026, CP027, CP004, CP035, CP036, CP037]
| 能力 | Ineffable Intelligence | OpenAI | Anthropic | Google DeepMind | xAI | Mistral AI |
|---|---|---|---|---|---|---|
| 体验式 RL / 无数据学习 | 规划中(核心使命) | 否 — RLHF / chain-of-thought RL | 否 — Constitutional AI / RLHF 对齐 | 部分 — 历史上的游戏领域 RL(AlphaGo/AlphaZero);不是无数据 AGI | 否 — LLM 优先,配合 RLHF | 否 — LLM 优先,开放权重 |
| 前沿 LLM 系列 | 未披露 | GPT-4o、o1、o3、o4-mini | Claude 3 Haiku / Sonnet / Opus | Gemini 2.5 Pro / Flash 系列 | Grok 3 系列 | Mistral Large / Small / Codestral / Pixtral 模型 |
| 已发布定价的商业 API | None | 是 — 已发布按 token 定价 | 是 — 已发布按 token 定价 | 是 — Gemini API 通过 Vertex AI / AI Studio 提供 | 是 — 面向开发者的 Grok API | 是 — 已发布按 token 和企业定价 |
| 拥有大用户基数的消费产品 | None | ChatGPT(周活约 200 M+) | Claude.ai(消费端和企业端) | Gemini app(Google Search + Android)消费产品 | Grok(X / Twitter 平台,约 500 M 用户) | Le Chat(消费端 + 企业 AI 助手) |
| 开放权重模型发布 | None | 否 — 闭源权重 | 否 — 闭源权重 | 否 — 闭源权重 | 否 — 闭源权重 | 是 — Mistral 7B、Mixtral 8x7B、Devstral(开放) |
| 已发布安全框架 | 未披露 | System cards、使用政策、安全评估 | Constitutional AI、RSP、Frontier Red Team 安全框架 | Responsible AI 团队、Gemini 安全报告 | 有限 — 没有正式发布的 RSP 等价框架 | 已发布安全框架很少 |
| 主权 / 欧盟监管定位 | 强 — 英国主权 AI 基金锚定,毗邻欧盟 | 弱 — 美国 PBC 架构,重心在美国 | 弱 — 重心在美国,Amazon 支持 | 中等 — 全球运营;欧盟主权重心有限 | 弱 — 美国中心,Musk 控制 | 强 — 法国总部,符合 EU AI Act,提供开放权重选项 |
能力项基于运行日公司官网和二级新闻来源整理。「规划中」表示公司宣称的未来能力,但没有商业证据;「未披露」 表示没有公开证据显示该能力。带限定词的单元格反映部分或有条件能力;没有支撑或未知的单元格已明确写出。
[CP004, CP026, CP027, CP035, CP036]Ineffable 及其五个主要竞争者在七个维度上的能力覆盖。Ineffable 唯一差异化位置在体验式 RL 范式;其他所有能力维度都偏向既有玩家。未获支持或未知的单元格已明确标出。
能力条目来自本次运行中抓取的官网和二级来源。"None" 和 "No" 单元格表示未公开披露该能力,并非确认架构上不可能。"Planned" 仅表示公司声称的未来状态。
[CP026, CP027, CP030]3.4 分发、计算与结构性力量
前沿 AI 中最持久的竞争优势不是模型质量或研究新意,而是分发能力、计算获取和人才密度。在这三个维度上,相比成熟玩家,Ineffable 都面临严重结构性劣势。 分发能力:ChatGPT 每周活跃用户超过 2 亿(OpenAI)、Claude.ai(Anthropic)、集成进 Google Search 和 Android 的 Gemini(Google DeepMind)、嵌入 X 的 5 亿注册用户基础的 Grok(xAI),这些消费者和企业分发护城河耗费多年和数十亿美元才建成。Ineffable 分发为零——没有产品、没有 API、没有用户。即便 Ineffable 产出突破性超级学习者系统,也需要从零建设分发,而 incumbent 已经握有全球企业和消费者 AI 买方的注意力和 API 关系。 计算获取:Google DeepMind 在 Alphabet 的 TPU Pod 基础设施内运行,规模不是外部实体能够复制;Microsoft Azure 为 OpenAI 的训练运行提供合同承诺计算;Amazon 对 Anthropic 的投资包含 AWS 计算额度。Ineffable 的 Google Cloud 合作意义重大,但并不提供与超大规模厂商支持实验室相同水平的承诺分配、多十年组织关系或冗余计算基础设施。xAI 的 Colossus 集群(约 100,000 块 H100)在原始 GPU 数量上超过 Ineffable 当前披露的集群(通过 Google Cloud 使用 NVIDIA Vera Rubin NVL72 驱动的 A5X)。 超大规模厂商合作:五家具名竞争者均至少拥有一个主要超大规模厂商合作(OpenAI/Microsoft、Anthropic/Amazon+Google、Google DeepMind/Google 内生、Mistral/Azure+Google)。Ineffable 拥有 Google Cloud 首选供应商和 NVIDIA 工程共同设计伙伴关系——有意义,但范围窄于部分竞争者享有的多超大规模厂商接入。 人才:仅 Google DeepMind 就在全球雇用数百名 RL 研究者。OpenAI 和 Anthropic 各自拥有规模相当的研究组织。Ineffable 公开确认的研究团队包括 David Silver 和另外三名 ex-DeepMind 创始人(Wojciech Czarnecki、Lasse Espeholt、Junhyuk Oh)——个人都很出色,但四人团队面对拥有数百乃至数千名研究者的组织,是资本无法迅速补齐的人才密度缺口。 [CP030, CP031, CP034, CP036, CP015, CP019]
3.5 护城河持久性、商品化与替代风险
Ineffable 的主要护城河主张是范式领导力:现存唯一一个以可信资本规模阐明并融资推进纯经验式 RL 超级学习者愿景的团队,由全球被引用最多的活跃 RL 研究者领导。今天这是真实差异化。护城河能否持久取决于两个因素:(i)资金充足的 incumbent 多久会追上这一范式;(ii)在复制发生之前,该范式能否产出可商业化部署的结果。 快速跟随风险很高。Google DeepMind 可以在内部重组一个不依赖人类数据的经验式 RL 研究计划:它保留同一套公开 RL 研究语料,仍有 David Silver 过去的团队成员在职,并拥有远超 Ineffable 的计算获取。如果超级学习者范式被证明可行,对 DeepMind 而言,将资源重新导向这一方向的增量努力远低于任何其他实体。AI 历史支撑这一担忧:一旦商业机会得到验证,OpenAI 就从通用语言建模转向基于 RL 的推理(o1 家族);DeepMind 本身也以极快速度和巨大影响从游戏 RL 转向蛋白质折叠(AlphaFold)。 OpenAI 在 2025 年转为公共利益公司,削弱了此前将非营利锚定实验室与商业机构区分开来的治理差异。随着 OpenAI 如今商业化运营、Anthropic 获得 Amazon 深度资助,Ineffable 所讲的「纯研究」定位已成为更拥挤的叙事——尽管 Ineffable 对反对短期产品的承诺比任何竞争者都更绝对。 Newspage.news 及其引用评论者提出的负面证据引发担忧:Ineffable 的英国政府少数股权并不能阻止 IP、人才或突破性发现流向控制董事会的美国总部投资者——这是一个治理缺口,不仅影响主权主张,也影响公司在主要投资者推动商业化时维持独立战略方向的能力。 商品化风险真实存在,但属于中期:驱动 DeepMind 历史突破的核心 RL 算法已经发表且公开可得。如果 Ineffable 的超级学习者路径纯粹依赖算法(而非依赖专有数据护城河或硬件整合),复制只需要足够的计算和人才——这两样 incumbent 都非常充足。 护城河持久性和竞争风险登记表(TP004)以及竞争就绪 KPI 仪表盘(FP003)以结构化形式总结了六项护城河主张及其主要威胁。 [CP031, CP032, CP033, CP034, CP039, CP040]
| 护城河主张 | 主要威胁 | 严重性 | 支撑证据 | 缓释措施 / 尽调问题 |
|---|---|---|---|---|
| 范式领导力:唯一一家以规模化方式追求无数据体验式 RL 的实验室 | Google DeepMind 或 OpenAI 将资源转向同一范式;先发优势在 12–36 个月内被侵蚀 | 高 | DeepMind 在 Silver 领导下做出了 AlphaGo/AlphaZero;仍保有 RL 人才;算力更强。OpenAI 在观察到商业可行性后约 24 个月内转向基于 RL 的推理(o1)。 | 定义并发布里程碑,在在位者缩小差距前证明进展;为关键算法创新锁定 IP 保护;跟踪 DeepMind RL 论文发布节奏 |
| 创始人履历:David Silver 是全球被引用次数最高的在职 RL 研究者 | 关键人物离职、丧失履职能力,或次级创始团队离开;风险集中在单一个人 | 严重 | 对外叙事、战略合作和科学可信度都通过 Silver 传导。次级创始人(Czarnecki、Espeholt、Oh)仅由二级来源报道,尚未在官方备案中确认。 | 确认并披露完整创始团队构成;建立科学领导层继任和冗余机制;谈判长期对齐激励的 IP 归属条款 |
| 通过 Google Cloud 优选供应商伙伴关系获得算力 | 出口管制、供应中断,或 Google Cloud 战略降级;不可抗力情形下合同可执行性不足 | 重大 | NVIDIA H20 出口管制导致 NVIDIA 单季度计提 $4.5 B 费用;Ineffable 的集群依赖通过 Google Cloud 采购的 NVIDIA Vera Rubin 硬件,受同类地缘政治供应链风险影响。 | 获取并审阅完整 Google Cloud 合同条款,包括算力配额承诺、硬件交付保证和不可抗力条款;探索多云选择 |
| 英国主权 AI 定位:British Business Bank + Sovereign AI Fund 少数共同投资人 | 英国政府少数股权不提供可执行的 IP 锚定或主权条件;突破性 IP 可能流向控制董事会的美国投资人 | 重大 | Newspage.news 评论及其引用专家明确质疑,在该治理结构下英国公共资本能否实现既定主权目标。美国 VC 董事会控制权(Sequoia、Lightspeed)可能压过英国战略利益。 | 要求并审阅 Ineffable 的投资人权利协议和董事会治理文件,评估英国政府股权是否带有 IP 锁定、否决权或主权条件;若不存在,应升级为尽调阻断项 |
| 体验式 RL 基础设施协同设计的先发优势(NVIDIA 伙伴关系) | NVIDIA 向其他前沿实验室(OpenAI、Google DeepMind)提供同等协同设计支持;工程伙伴关系并不排他 | 中等 | NVIDIA 博文称为 Ineffable 协同设计 RL 流水线,但 NVIDIA 同时也是 OpenAI 的伙伴和投资人,并且是 DeepMind、xAI 的主要供应商。尚未披露排他性。 | 确认 NVIDIA-Ineffable 协同设计伙伴关系是否包含任何排他期或优先访问权;厘清 Vera Rubin NVL72 集群访问权是排他还是与其他 NVIDIA 客户共享 |
| 人才集中:前 DeepMind RL 专家团队,拥有十年前沿 RL 研究积累 | 在位者用留任包抢夺人才;DeepMind、xAI 和 OpenAI 都在从同一 RL 人才池积极招聘 | 中等 | 前沿 AI 研究招聘竞争极其激烈;DeepMind、OpenAI 和 Anthropic 各自拥有数百名 ML 研究员。Ineffable 的次级创始团队公开确认有限。 | 核验雇佣协议和股权归属时间表;评估此前 DeepMind 雇佣关系中的任何竞业限制或禁止招揽条款是否仍有效;确认次级团队已全职投入 |
严重性评级(高 / 严重 / 重大 / 中等)是分析师基于引用证据作出的判断;不是定量概率估计。证据来源按 source ID 引用;每行的交叉印证主张见章节 claimRefs。
[CP031, CP032, CP033, CP034, CP039, CP040]六项竞争耐久度 KPI,概括截至 2026 年 6 月 Ineffable 相对现有前沿实验室的护城河位置和关键风险维度。
[CP034, CP039, CP030, CP033]3.6 展示项
04财务
4.1 收入模式与变现路径
截至 2026 年 6 月 22 日,Ineffable Intelligence 没有披露收入、客户或产品。公司的公开使命明确排斥近期商业产出:它寻求「一个让宏大研究能够繁荣的窗口,不必屈从于渐进式产品和短期利润的要求」。这是一种有意选择的战略姿态,不是暂时缺口。公司目前是纯研究实体,没有披露任何商业合同、价格表或合作伙伴收入安排。 在没有确认收入来源的情况下,分析只能从公司的技术定位和可比先例推演未来可能的变现路径。四条路径站得住脚:(a)向企业、政府和研究机构许可已训练模型权重或 RL 能力——前沿 AI 实验室能力成熟后的主流模式;(b)面向政府(B2G)的主权 AI 合同,英国政府通过 British Business Bank 和 Sovereign AI Fund 共同投资,构成结构性入口;(c)当能力达到商业可用后,向已部署的 superlearner 提供 API 访问;以及(d)由系统自身产生的科学 IP 和专利带来的版税或许可费,尤其是在科学、数学和工程领域。上述路径均未被公司确认、披露或给出时间戳。Sequoia 的「Act Two」分析(2023)警告,若 AI 实验室短期内没有产品市场匹配,算力成本膨胀会带来融资风险,投资者对收入前估值的耐心也有历史上限。 定价侧没有可分析的定价模型。任何关于 Ineffable 未来收入的模型,都必须假设商业部署时间、目标客户细分(主权、企业、研究机构)以及定价机制(API 计量访问、权重许可、版税)。这些问题仍未解决,也未披露。收入流表(TI001)列出可能路径、机制及其证据质量。定价表(TI002)记录当前定价认知状态,以及数据室需要补充的材料。收入模式桥图(FI001)展示在最可能的许可场景下,客户互动如何转化为收入和毛利。 [CI001, CI002, CI023, CI024, CI027, CI037]
| 来源 | 机制 | 单位 / 定价代理 | 当前状态 | 收入质量 | 尽调问题 |
|---|---|---|---|---|---|
| 已部署 superlearner 的 API 访问 | 按 token 或订阅计量 superlearner 推理 | 按 API 调用或分层订阅;未披露标价 | 不可用 — 尚未产生收入,没有产品 | 推测;一旦可用,毛利率上限高 | 确认是否规划 API 商业化;索取产品路线图 |
| 权重 / 能力授权 | 将训练好的 RL 模型权重一次性或经常性授权给企业和政府 | 预付费或版税;未披露定价 | 不可用 — 没有可授权能力 | 推测;IP 授权典型毛利率 80–90%+ | 索取商业化时间表,以及潜在客户的任何 LOI 或 MoU |
| 主权 / B2G 合同 | 政府采购 Ineffable 能力,用于国家 AI 项目 | 合同制;BBB 和 Sovereign AI Fund 共同投资提供结构性入口 | 未披露合同;共同投资传递意图,不等于收入 | 推测;政府合同可有高毛利,但销售周期长 | 厘清 UK Sovereign AI Fund 共同投资是否附带任何采购偏好或优先查看权 |
| 科学 IP 版税 | superlearner 产出的科学发现(药物靶点、材料、数学)带来许可费或版税 | 可变;取决于 IP 制度和专利组合建设 | 未披露 IP 组合;公司尚未进入发现阶段 | 推测;若发现可商业授权,毛利率高 | 索取 IP 所有权政策和任何规划中的 IP 战略文件 |
| 研究合作 / 资助 | 来自研究委员会、ARIA、EU Horizon 或美国同类机构的非稀释资金 | 基于资助;不是经常性收入,但可降低烧钱 | 未公开宣布资助;UK AI Growth Zones 和 DSIT 创造路径 | 不是收入;只降低有效烧钱率 | 确认是否已申请或获得任何 UKRI、ARIA 或 EPSRC 资助 |
所有收入来源均为推测;截至 2026 年 6 月 22 日,Ineffable Intelligence 未披露收入、客户或商业合同。本表基于公司技术定位和可比前沿 AI 实验室先例(OpenAI、Anthropic、DeepMind)列举可行未来路径。收入质量评级是结构性判断,不是已确认指标。
[CI001, CI023, CI024, CI028, CI041]| 维度 | 已知 / 已披露 | 来源 / 依据 | 尽调问题 |
|---|---|---|---|
| 标价 | 未披露;不存在商业产品 | ineffable.ai 或伙伴披露中没有任何定价页 | 向数据室索取任何定价框架草案或商业条款清单 |
| 合同结构 | 未知;未披露客户协议 | 没有监管备案或投资人声明提及合同 | 索取与潜在客户或伙伴签署的任何 LOI、MoU 或条款概要 |
| 实际价格相对标价折扣 | 不适用;不存在定价 | N/A | 将取决于商业化时的竞争动态;没有可用代理 |
| 收入确认方法 | 未知;公司尚未进入收入阶段 | 未提交 IFRS/UK GAAP 财务报表(Companies House,2026 年 6 月 22 日) | 审阅数据室中的会计政策草案;确认是否已准备 IFRS 15 框架 |
| 渠道 / 经销经济性 | 未披露;未宣布渠道伙伴 | 除 Google Cloud(基础设施,不是分发)外,没有伙伴公告 | 确认是否规划通过 Google Cloud Marketplace 或 NVIDIA 生态做企业分发 |
| 政府资助 / 补贴会计 | 未披露;UK AI Growth Zones 代表潜在非稀释路径 | Gov.uk AI Growth Zones 发布文件(2025);未确认向 Ineffable 提供资助 | 确认任何 UKRI 或 DSIT 资助申请;索取其在管理账中的处理方式 |
定价表反映截至运行日商业货币化层的完全缺失。Ineffable 未披露产品、定价页或客户协议。所有单元格记录的是缺乏证据,而不是反向证据。
[CI001, CI002, CI012, CI023, CI028]展示未来超级学习者能力在最可行的授权或 API 访问场景中如何转化为收入和毛利,并标出当前仍为空的节点。
“算力基础设施”和“RL 研究”之后的所有节点都代表未来状态,目前没有确认的商业化计划。该桥接图只展示结构,用于说明,不是预测。
[CI013, CI014, CI024, CI036, CI040]4.2 成本结构与资本强度
现阶段,Ineffable 的成本结构几乎全是研究和基础设施成本。公开资料中没有销售成本(COGS)、毛利率或经营杠杆数据;公司也没有收入可用于衡量成本比例。 最大成本驱动项是算力。与标准 transformer 预训练不同,强化学习负载需要紧密集成的训练—推理循环;NVIDIA 与 Ineffable 联合署名博客称,这种循环会「以预训练不会出现的方式,对互联、内存带宽和服务能力施压」。系统必须持续行动、观察、打分并更新——与静态数据集预训练相比,对集群级编排提出数量级更高要求。算力强度是公司最重大的财务约束:Epoch AI 的分析(2025)显示,前沿模型训练成本自 2016 年以来每年增长 2.4 倍,硬件占前沿模型开发成本的 47–67%,研发人员占 29–49%,能源占 2–6%。Epoch AI 预计,若趋势延续,到 2027 年最大规模训练运行单次成本将超过 $1 billion。 Google Cloud 和 NVIDIA 两个合作关系都显著缓解了成本压力。Google Cloud 作为 Ineffable 的首选云提供商,正在部署由 NVIDIA Vera Rubin NVL72 驱动的最大 A5X 集群之一——在标准云采购模式下,这类基础设施不需要 Ineffable 直接资本开支融资。NVIDIA 的联合设计合作提供工程资源和硬件访问,否则公司需要直接采购。两个合作关系合在一起,把原本巨额资本开支(或前沿规模集群每年数亿美元级云租赁支出)转化为与合作商业条款绑定的更灵活成本结构——但相关细节并未公开披露。Arxiv 研究(2022)发现,大规模 ML 模型所需算力比标准深度学习高 10-100×,深度学习时代训练算力约每 6 个月翻一倍,因此成本轨迹是长期结构性风险,不是静态成本项。 研发人员是第二大成本类别。Ineffable 于 2026 年 2 月 5 日通过员工股票期权计划(含美国子计划),显示公司打算以有竞争力的股权薪酬水平招聘并留住人才。股份支付薪酬会按 IFRS 2 或同等准则确认为非现金费用,即使现金人力成本尚未扩张,也会新增一条未来成本线。公司没有披露员工人数,也尚未提交任何财务报表。资本强度与现金流图(FI004)在这些约束下勾勒结构性成本流向。 [CI013, CI014, CI015, CI016, CI017, CI018]
梳理 $1.1B 种子轮资金流向主要支出类别和成本节点,展示 Google Cloud 与 NVIDIA 合作在哪些环节替代直接资本开支,以及哪些位置仍存在私有信息缺口。
现金流为结构性、方向性描述;各节点之间的实际分配未公开披露。箭头权重和节点大小仅用于说明。
[CI003, CI006, CI011, CI015, CI016, CI021]4.3 资本充足性与现金续航
Ineffable 于 2026 年 4 月 27 日完成 $1.1 billion 种子轮融资,投后估值 $5.1 billion。这是欧洲史上最大种子轮,由 Sequoia Capital 和 Lightspeed Venture Partners 共同领投。已确认参与方包括 British Business Bank(确认 $20 million)、UK Sovereign AI Fund(金额未披露)、NVIDIA、Google、Index Ventures、DST Global、EQT Ventures、Flying Fish、Evantic Capital、BOND Capital 和 Wellcome Trust。 公司未公开披露任何债务工具、信贷额度、风险债或项目融资义务。Companies House 尚无年度账目;公司于 2025 年 11 月 19 日成立,首份账目尚未达到法定提交期限。Companies House 备案历史显示三次股本声明:2025 年 11 月成立时 GBP 10;2026 年 1 月 21 日 GBP 320(与 Silver 被任命为董事及重要控制人相关);2026 年 3 月 12 日 GBP 350。这些是名义法定数值,不是实际募集的股本溢价;$1.1 billion 投资会以股份溢价流入,并在首份财务报表提交时出现。 前沿 AI 实验室基准显示,$1.1 billion 融资支撑多年资金续航具备可能性。Anthropic 和 OpenAI 在达到类似资本水平前,已用数亿美元级年度算力加人员预算运营。但这些比较并不精确:给定同等能力水平,Ineffable 以 RL 为中心的负载可能比 transformer 预训练结构性地更耗算力,精确月烧钱额也未知。即便按每月 $60–100 million 的烧钱率计算,$1.1 billion 融资也只能提供约 11–18 个月资金续航——明显低于使命所暗示的「多年研究周期」,除非 Google Cloud 和 NVIDIA 的合作补贴能显著抵消直接支出。因此,资本充足性可以成立,但尚未被证明;公司的下一轮股权融资很可能由科学里程碑、种子资金耗尽或商业部署启动触发。UK AI Growth Zones 和更广泛的 DSIT 政策环境可能带来非稀释资金入口,但目前没有公开宣布给 Ineffable 的此类补助。资本充足性表(TI004)和财务估算区间图(FI003)给出关键资本指标的结构化区间估计。 [CI003, CI004, CI005, CI006, CI007, CI008]
| 指标 | 值 / 状态 | 日期 / 依据 | 置信度 | 尽调问题 |
|---|---|---|---|---|
| 总股权融资 | $1.1 billion(仅种子轮) | 2026-04-27 | 高 | 让法律顾问确认股权结构和股份类别条款 |
| 投后估值 | $5.1 billion | 2026-04-27 | 高 | 估值来自种子轮条款;没有二级交易可交叉校验 |
| 手头现金 | 未披露;扣除初始搭建成本后,推定接近 $1.1B | 2026-06-22 | 低 | 索取最近月末经审计或管理账口径现金余额 |
| 月度烧钱速度 | 未披露;按可比基准估算为 $30–100M/月 | 2026-06-22 | 低 | 索取管理账和现金流预测;确认是否计入合作伙伴补贴 |
| 现金跑道估算 | 未披露;按种子轮资金估算为 11–36 个月(取决于烧钱速度和补贴) | 2026-06-22 | 低 | 索取列明跑道情景的财务模型;厘清 Google Cloud / NVIDIA 经济条款 |
| 债务 / 信贷额度 | 未披露 | 2026-06-22 | 中 | 向法律顾问确认;索取任何 venture debt 或项目融资问询记录 |
| 计划资金用途 | 未正式披露;推定用途:算力基础设施、研发人员、工程合作 | 2026-06-22 | 低 | 从投资人材料或董事会 deck 中索取正式资金用途拆分 |
| 下一轮触发条件 | 未披露;推定为:科学里程碑、资金消耗,或商业部署阶段 | 2026-06-22 | 低 | 向管理层确认 Series A / 成长期融资时间表和里程碑触发条件 |
资本充足性指标主要来自 2026 年 4 月种子轮公告。烧钱速度、现金跑道和现金头寸均未公开披露;估算区间以 Anthropic、OpenAI 等前沿 AI 实验室为基准,并按 Ineffable 算力密集型负载特征调整。Google Cloud 和 NVIDIA 的合作补贴可能显著拉长跑道,但商业条款未披露。
[CI003, CI005, CI006, CI008, CI009, CI010]对无法从公开证据精确测量的关键财务参数,给出有来源支撑的估计区间。所有区间置信度较低,并标注依据。
区间来自前沿 AI 实验室基准(Epoch AI、NVIDIA、Anthropic/OpenAI 公开披露),不是 Ineffable 管理账。实际值可能有重大差异。
[CI018, CI019, CI020, CI025, CI034, CI036]4.4 公开财务缺口与单位经济模型
Ineffable Intelligence 的财务画像几乎完全处于私有状态。作为收入前研究实体,公司没有提交账目、没有披露产品定价,也没有客户指标,标准单位经济模型框架无法按常规方式适用。没有客户获取成本(CAC)、没有客户生命周期价值(LTV)、没有回本周期、没有 ARR、没有 GMV,也没有单位量。这些缺口不是暂时披露延迟,而是公司刻意保持商业化前状态的结果。 真正可分析的是商业化启动后,未来单位经济模型的结构性驱动因素。前沿规模 RL 算力的资本强度意味着,一旦开始产生收入,COGS 将主要由算力成本而非人力劳动主导;结构上更接近超大规模云厂商或云原生软件毛利,而不是服务业务。如果 Ineffable 通过 API 访问或权重许可变现,服务更多用户的边际成本可能较低(毛利率天花板较高),但摊销后的研究和基础设施成本基数会很重。即使缺少真实数字,单位经济模型桥图(FI002)也从结构上绘制了这一逻辑。市场进入动作完全未定:没有销售团队、没有渠道伙伴、没有披露的上市计划、没有客户引用,也没有销售管线数据。公开财务缺口表(TI005)逐项列出缺失指标及其具体尽调路径。单位经济模型表(TI003)记录每个关键指标已知、估算或不可得的状态,并给出明确信心等级。 [CI029, CI030, CI031, CI033, CI034, CI036]
| 指标 | 值 / 状态 | 置信度 | 重要性 | 尽调问题 |
|---|---|---|---|---|
| ARR / 收入运行率 | $0 / 未披露 | 高(已确认尚未产生收入) | 基线收入指标;确立毛利率和 LTV 背景 | 索取首笔收入时间表和任何已承诺客户管线 |
| 客户获取成本(CAC) | 不适用 — 没有客户 | 高(已确认没有客户) | B2B 或 B2G 商业模型的关键效率指标 | 索取财务模型中的预计 CAC 假设 |
| 客户生命周期价值(LTV) | 不适用 — 没有客户或合同 | 高 | LTV/CAC 比率是 SaaS 或授权模型的核心承销输入 | 向数据室索取预计 LTV 假设和定价模型 |
| 毛利率 | 未披露;IP 授权模型估算上限为 70–90% | 低(参考可比 IP / 软件授权商基准估算) | 决定规模化后的盈利能力和增长资本效率 | 定价模型明确后,向管理层索取销售成本估算 |
| 月烧钱率 | 未披露;基于前沿 AI 实验室基准估算为 $30–100M/月 | 低(估算;未确认) | 跑道和下一轮融资时点的主要驱动因素 | 索取显示现金头寸和月烧钱的管理账 |
| 跑道(来自 $1.1B 种子轮) | 估算为 11–36 个月,取决于烧钱率和伙伴补贴 | 低(估算;取决于未披露的伙伴商业条款) | 对下一轮融资时点和财务风险评估至关重要 | 索取含有和不含 Google Cloud / NVIDIA 补贴假设的详细现金流模型 |
| 回收期 | 不适用 — 没有收入 | 高 | 收入开始后可显示销售周期和资本效率 | N/A,直到商业模型明确 |
| COGS / 交付成本 | 未披露;以算力为主(预计占开发成本的 47–67%) | 低(参考 Epoch AI 前沿模型成本基准估算) | 决定未来定价的毛利率上限 | 索取算力成本结构,以及与 Google Cloud 已约定的任何云合同条款 |
指标值反映 Ineffable Intelligence 公开单位经济数据完全缺失。烧钱率和跑道是基于前沿 AI 实验室成本基准(Epoch AI、NVIDIA 财务披露)得出的粗略估算;公司并未确认。毛利率上限是基于 IP 授权模型可比项作出的结构性估算,不是公司披露。
[CI001, CI017, CI018, CI025, CI029, CI030]| 缺失指标 | 对承销判断的影响 | 具体尽调路径 |
|---|---|---|
| 经审计或管理口径财务报表 | 阻断项——没有损益表、资产负债表或现金流表,无法判断烧钱速度、COGS 或收入质量 | 通过 data room 索取;公司尚无法定申报义务,但应有管理账 |
| 月度现金烧钱速度 | 阻断项——没有烧钱速度,就算不出跑道;下一轮融资风险无法量化 | 索取 2026 年 4–6 月月度管理账;要求提供过去 3 个月平均值 |
| $1.1B 种子轮资金用途拆分 | 重大项——没有分配数据,无法判断多少资金留给算力、人员和运营 | 索取投资人 deck 中的资金用途页,以及任何经董事会批准的预算 |
| Google Cloud 商业合同条款 | 重大项——算力补贴可能让有效烧钱速度减半或翻倍;未披露条款是关键变量 | 索取经济条款摘要:折扣、最低承诺、信用期,以及任何股权或收入分成安排 |
| NVIDIA 合作商业条款 | 重大项——工程共同设计合作可能包含硬件折扣、免费算力时长,或 IP 共享义务 | 索取 NVIDIA 商业协议条款摘要和 IP 归属安排 |
| 员工人数与薪酬结构 | 重大项——研发人员占前沿 AI 开发成本的 29–49%;公司未披露人数 | 按职能索取员工人数、平均薪酬、股票期权池规模和条款 |
| 种子轮后股权结构表(完全摊薄) | 重大项——投资人占比、期权池摊薄和任何清算优先权都会影响回报画像 | 索取完整 cap table,包含种子轮前后摊薄表、期权池明细和 warrant 安排 |
本表列出承销判断所需、但截至 2026 年 6 月 22 日任何公开来源均未提供的具体私有财务指标。「阻断项」缺口会使可信承销判断无法成立; 「重大项」缺口会影响财务模型精度,但不妨碍定性投资判断。
[CI001, CI005, CI012, CI013, CI029, CI030]梳理未来 API 或授权业务的单位经济结构逻辑,并用空节点明确标出当前数值未知的位置。当前没有数据,输入为定性判断,并附上近似说明。
目前没有实际单位经济数据;除“输入:RL 算力成本”外,所有节点均为假设。毛利率上限参考 IP 和软件授权基准推断。
[CI015, CI017, CI025, CI029, CI031, CI035]4.5 财务结论与尽调阻碍项
传统意义上,Ineffable Intelligence 无法被财务承销。收入质量为零——没有收入。毛利率路径是推测——没有可建模的 COGS 结构。资本强度极高——前沿规模 RL 叠加 Vera Rubin NVL72 集群,只能由商业条款未披露的合作补贴部分抵消。烧钱率和资金续航未确认——公司没有发布财务报表,也没有公开烧钱估计。下一轮融资风险显著——Ineffable 需要额外资本(很可能仍是十亿美元级)才能在初始研究阶段后维持前沿规模运营,融资时点也不确定。 对种子阶段前沿 AI 研究投资而言,这并不构成一票否决。$1.1 billion 种子轮规模明确用于资助研究周期,而非商业里程碑;Sequoia、Lightspeed、NVIDIA、Google 和 Wellcome Trust 的参与,也验证了资本形成假设。但若没有证据显示商业化时间表已经定义且可合理实现,投资者不应在标准 3-5 年风险投资周期内建模正向现金流。核心尽调阻碍项包括:(1)没有任何已提交或审计财务报表;(2)没有披露烧钱率、现金头寸或资金续航;(3)没有 $1.1 billion 种子轮资金用途拆分;(4)没有商业路线图或收入模式;(5)算力合作商业条款抵消资本需求,但结构未知;(6)没有公布员工人数或薪酬结构。每一项都是治理良好的 Series A 或更后期流程中标准的数据室问题。 [CI023, CI024, CI025, CI026, CI033, CI037]
4.6 图表证据
05产品与技术
5.1 产品定义——Superlearner 研究平台
Ineffable Intelligence 不提供商业产品、服务或 API。它当前交付的是一个研究项目和基础设施栈,目标是打造公司所谓的「superlearner」:一个通过强化学习,从自身经验中发现全部知识、而不依赖人类生成数据的 AI 系统。公司明确表示,自己处在「一个让宏大研究能够繁荣的时间窗口,不必屈从于渐进式产品和短期利润的要求」。这是有意选择的商业化前姿态,不是暂时缺口。 根据公开材料,superlearner 概念包含三个功能组件。第一,经验生成层:系统在模拟环境中行动,生成自己的训练数据,而不是消费固定数据集。第二,打分与奖励层:系统观察自身行动结果,依据奖励信号评估结果,并更新内部状态。第三,训练—推理循环:系统基于打分后的经验持续细化参数;NVIDIA 称这种「紧密循环」对互联、内存带宽和服务基础设施提出异常高的要求。David Silver 曾把系统的预期范围描述为学习「从基础运动技能到深刻智力突破」的一切,并称其为「量级可与达尔文相比的科学突破」。 从任何商业意义上看,这个平台都还不是产品。没有定价、没有客户、没有部署,也没有除创始人在 DeepMind 既有工作公开描述之外的能力文档。最接近的类比是早期研究算力项目:一组基础设施投资、招聘计划和算法开发冲刺,目标是最终产出一个具备科学意义的训练系统。产品模块表(TE001)把研究平台的功能组件,与当前成熟度、目标用户、差异化因素和尽调缺口逐一对应。 [CE001, CE002, CE003, CE004, CE005, CE021]
| 模块 / 资产 | 用户 / 受益方 | 状态 / 成熟度 | 差异化因素 | 尽调缺口 |
|---|---|---|---|---|
| 经验生成层(仿真环境) | 研究团队(内部) | 概念 / 基础设施规划 | 必须生成丰富的非人类经验;相对标准游戏引擎的新颖性未知 | 未披露仿真环境描述、领域覆盖或奖励模型设计 |
| RL 训练循环(行动-观察-评分-更新管线) | 研究团队(内部) | 基础设施建设中;从 Grace Blackwell 起步 | 与 NVIDIA 共同设计;瞄准算力与互联的紧耦合 | 未发布架构规格、模型设计或超参数体系 |
| 奖励 / 评分层 | 研究团队(内部) | 概念 / 早期设计 | 开放式发现的核心未解难题就是奖励设计 | 未披露奖励模型类型、评估标准或反奖励黑客措施 |
| Superlearner 模型(训练后的 RL agent) | 研究受益方(科学、数学、技术) | 训练前;尚无已训练系统 | 若成功:不依赖人类数据发现新的科学知识 | 未公开承诺能力里程碑、benchmark 或时间表 |
| Google Cloud A5X 部署集群(Vera Rubin NVL72) | 研究团队(内部) | 合同已承诺;硬件处于生产中 | Google Cloud 上规模最大的 A5X 集群之一;集成 AI Hypercomputer | 未披露集群规模、合同条款、定价和排他性 |
所有模块都处于研究或基础设施建设阶段。状态判断基于公开合作公告和 NVIDIA 硬件生产确认。截至 2026 年 6 月 22 日,公司未披露产品规格、 设计文档或能力输出。
[CE001, CE002, CE004, CE005, CE007, CE011]5.2 架构与运营模式——RL 训练管线
核心运营模式是一条基于经验的强化学习管线:智能体在模拟环境中行动,生成观察和奖励,模型权重在连续循环中更新。这与静态数据集预训练有根本差异。NVIDIA 与 Ineffable 联合署名博客解释称,「系统必须在紧密循环中持续行动、观察、打分并更新,这会以预训练不会出现的方式,对互联、内存带宽和服务能力施压。」系统还会在「与人类语言截然不同的丰富经验形态」上训练,可能需要「新型模型架构和训练算法」。 创始团队为这种架构带来深厚技术先例。联合创始团队成员 Lasse Espeholt 是 IMPALA 架构(2018)的第一作者;该架构确立了可扩展分布式深度 RL,并采用重要性加权的 actor-learner 设计。A3C 框架(Mnih 等,2016)是异步并行 RL 训练的基础,也是 Ineffable 正在扩展方法的直接祖先。DeepMind 团队的 AdA 论文(2023)——《Human-Timescale Adaptation in an Open-Ended Task Space》——证明大规模多任务 RL 可以产生通用智能体,并在数百项任务中以人类时间尺度学习并适应。这些先例提供背景,但并不定义 Ineffable 的系统;公司没有发布任何关于实际 superlearner 的架构描述、模型规格或设计文档。 运营流程仍处研究阶段:模拟环境生成经验,奖励模型给经验打分,训练更新传递进模型,然后循环在规模上重复。工作流表(TE002)把这一研究工作流与未来预期用例对应。架构表(TE003)把技术栈层级与依赖和风险对应。运营流程图(FE002)展示经验生成、打分和训练如何在研究管线中连接。 [CE005, CE006, CE015, CE023, CE024, CE025]
| 用户任务 / 目标 | 当前工作流(Superlearner 之前) | Ineffable 目标方案 | 声称收益 | 限制 / 证据缺口 |
|---|---|---|---|---|
| 科学知识发现(如新材料、药物靶点) | 人类研究者借助既有文献和实验迭代假设 | Superlearner 在大规模仿真中通过 RL 生成并测试假设 | 加速发现,超出人类研究者能穷尽探索的范围 | 未部署能力;没有任何 Ineffable 系统产出科学成果的证据 |
| 数学证明发现 | 数学家靠直觉和启发式搜索,证明助手负责验证 | RL agent 通过经验发现新证明(可类比 AlphaProof) | 首个自主解决奥赛级问题的 AI 系统 | AlphaProof 属于 DeepMind,不属于 Ineffable;Ineffable 尚无可比系统 |
| 工程优化(排序、矩阵乘法) | 领域专家手工设计算法;搜索依赖启发式方法 | RL agent 在仿真中试错,发现更快算法 | DeepMind 已演示(排序、矩阵乘法)——Ineffable 尚未做到 | 前代项目证明概念可行;Ineffable 尚未复现或扩展 |
| 通用知识获取(任意领域) | LLM 基于互联网规模的人类数据训练;知识边界受人类记录限制 | Superlearner 通过经验生成自有知识;原则上不受边界限制 | 摆脱人类数据受限 AI 系统的「化石燃料」约束 | 通用知识规模尚未验证;奖励函数设计是关键瓶颈 |
工作流描述基于 David Silver 的公开表述、NVIDIA-Ineffable 博客、Wired 访谈,以及 Silver 在 DeepMind 任职期间的先例。截至运行日期, Ineffable 尚未部署或演示任何自有用例。所有收益主张要么属推测,要么来自 DeepMind 前代项目。
[CE002, CE003, CE004, CE021, CE022, CE026]| 层 / 流程 / 组件 | 系统角色 | 关键依赖 | 主要风险 |
|---|---|---|---|
| NVIDIA Vera Rubin NVL72 GPU 集群 | 核心训练算力;每机柜通过 NVLink 6 连接 72 个 Rubin GPU + 36 个 Vera CPU | NVIDIA 硬件产能爬坡;供应分配 | 供应延迟,或被转向竞争性 hyperscaler 负载 |
| NVIDIA Vera CPU 机柜(每机柜 256 个 CPU) | RL 仿真环境;训练循环中基于 CPU 的 agent 执行 | NVIDIA Vera CPU 生产;Spectrum-X Ethernet 网络 | CPU 环境吞吐量限制经验生成规模 |
| Google Cloud AI Hypercomputer(Jupiter 网络 + 存储) | 云部署层;网络结构;经验数据存储 | Google Cloud 可用性;多年合同条款 | 合作条款(定价、credits、SLA)未披露;集中在单一云上 |
| 经验生成 / 仿真层 | 即时生成训练数据;定义奖励信号和环境 | 内部研究团队设计;仿真引擎未披露 | 奖励黑客、分布错配、仿真到现实落差 |
| RL 训练算法与模型架构 | 将经验转成权重更新;策略改进循环 | Silver 团队专长;新型架构设计(尚未披露) | 非游戏领域需要新架构;这一规模没有先例 |
| 推理 / 评估层 | 测试已训练 agent 的能力;为发现评分;验证输出 | 尚未设计或披露 | 未公开定义评估框架、benchmark 或成功指标 |
架构推导自 NVIDIA-Ineffable 合作博客、Google Cloud 新闻稿和 NVIDIA Vera Rubin 平台公告。Ineffable 未发布任何架构规格。所有内部层 (仿真、训练算法、推理)均由公开合作伙伴披露和 David Silver 的技术表述推断而来。
[CE005, CE006, CE007, CE008, CE009, CE011]展示 Ineffable 计划中的强化学习研究管线里,经验生成、评分、训练和评估如何衔接。
[CE004, CE005, CE006, CE015, CE021, CE022]5.3 基础设施——NVIDIA 与 Google Cloud 依赖
Ineffable 的算力基础设施建立在双合作伙伴模式上。NVIDIA 提供硬件平台;Google Cloud 提供云部署环境。2026 年 5 月宣布的 NVIDIA 合作是一项工程级协作,目标是「联合设计大规模强化学习基础设施」。工作从 NVIDIA Grace Blackwell 开始,并且「首批探索即将推出的 NVIDIA Vera Rubin 平台」。Vera Rubin NVL72 机架整合 72 个 Rubin GPU 和 36 个 Vera CPU,通过 NVLink 6 连接,并包含一个专用 CPU 机架,配备 256 个为 RL 环境模拟专门设计的 Vera CPU——这一硬件能力与 Ineffable 的经验生成层直接相关。Vera Rubin 平台每瓦推理吞吐量最高可达 Blackwell 的 10 倍,并能用四分之一 GPU 数量训练大型模型。 2026 年 6 月,Google Cloud 在「对基础设施市场进行严格评估」后被选为首选云合作伙伴。根据合作安排,Ineffable 将在 Google Cloud 上部署「最大 A5X 集群之一」,由 Vera Rubin NVL72 驱动。Google Cloud 的「AI Hypercomputer」架构——结合 Jupiter networking、性能工程化 GPU 和优化存储——被 David Silver 称为胜过标准「一箱芯片」方案的决定性因素:「训练前沿模型需要的不只是原始算力;它需要复杂的硬件和软件编排。」合作还覆盖 Google Cloud 的网络结构(Jupiter networking)和存储系统,使训练与推理负载能够紧密耦合。 这些基础设施依赖带来实质性集中风险。Ineffable 的整个算力项目依赖 NVIDIA Vera Rubin 的生产爬坡(已确认在 GTC 2026「全面投产」)、NVIDIA 的软件栈(CUDA、NVLink),以及 Google Cloud AI Hypercomputer 的规模化可用性和定价。两项合作的条款——定价、算力额度、排他性和服务水平协议——均未公开披露。关键依赖图(FE003)映射完整依赖链。产品架构图(FE001)展示分层技术栈。 [CE007, CE008, CE009, CE010, CE011, CE012]
分层展示 Ineffable 研究平台:从物理算力到预期科学输出,并标出每一层当前成熟度。
层级结构根据 NVIDIA 博客、Google Cloud 新闻稿和 NVIDIA Vera Rubin 平台公告推断。Ineffable 未公开披露内部架构(仿真设计、奖励模型、训练算法)。
[CE005, CE006, CE007, CE008, CE009, CE011]用有向图展示 Ineffable 的关键外部依赖:硬件供应商、云服务商、监管机构和合作伙伴路线图。
[CE007, CE011, CE013, CE018, CE027, CE028]5.4 部署状态与路线图
截至 2026 年 6 月 22 日,Ineffable 没有已部署产品、API 或公开可访问系统。公司仍处纯研究和基础设施建设阶段。没有公开披露发布时间表、产品路线图、能力目标或训练计划。 唯一能证明前瞻规划的证据来自基础设施合作:NVIDIA 协作被描述为探索「当 AI 世界从人类数据转向通过模拟和经验学习的模型时所需的下一代硬件和软件」;Google Cloud 合作则描述部署「最大 A5X 集群之一」,这是一项面向未来的基础设施承诺,而非当前能力。 Silver 的公开沟通描述的是多年研究窗口:「一段让宏大研究能够繁荣的时间和机会窗口」。这与 $1.1 billion 种子资本用于资助研究阶段、而非上市冲刺的定位一致。Sovereign AI 共同投资也明确把该投资框定为长期支持一家「定义品类的公司」,而不是近期产品周期。没有能力里程碑(例如首个已训练 RL 智能体、首个科学发现、首个基准测试结果)被宣布、承诺或泄露。 路线图表(TE005)映射 Ineffable 发展项目中公开可观察的阶段。这些条目基于基础设施合作时间线、NVIDIA 硬件可用性和公司自述研究周期推断;没有任何条目是公司披露的目标日期。基础设施建设之后的每个阶段都存在重大不确定性。 [CE017, CE020, CE021, CE028, CE031]
| 阶段 / 时间窗口 | 里程碑 / 功能 | 当前状态 | 含义 | 来源 |
|---|---|---|---|---|
| Nov 2025 – Apr 2026 | 公司注册、团队组建、隐身研发、种子轮融资 | 已完成——$1.1B 种子轮于 2026 年 4 月 27 日关闭 | 证明融资能力;没有产品或技术里程碑 | Companies House;CNBC;Cooley 公告 |
| May 2026 | 宣布 NVIDIA 工程合作;Grace Blackwell 工作启动 | 已宣布——基础设施共同设计推进中 | 硬件合作降低 NVIDIA 获取风险;不证明已训练能力 | NVIDIA 博客;Ineffable 博客 |
| June 2026 | 宣布 Google Cloud A5X 合作;承诺 Vera Rubin NVL72 集群 | 已宣布——计划部署集群;Vera Rubin 已投产 | 云基础设施已锁定;集群交付后即可开始训练 | Google Cloud 新闻稿;NVIDIA Vera Rubin 公告 |
| 2026 下半年 – 2027(推断) | 基础设施建设;首次大规模 RL 训练运行 | 未宣布——由合作时间线和 NVIDIA Vera Rubin 可用性推断 | 预计这一窗口会出现系统能力(或缺乏能力)的首批证据 | 分析师基于合作伙伴路线图推断;Ineffable 未作承诺 |
| 2027 – 2029(推测) | 首批科学能力里程碑(若研究成功) | 未宣布——推测性研究周期 | 若没有突破,预计这一窗口内不会产生收入 / 商业化 | Silver 使命表述;Sovereign AI 对长期支持的定位 |
第 1–3 阶段日期为已确认事件。第 4–5 阶段来自合作伙伴路线图和 David Silver 使命表述的推断;Ineffable 未公开承诺任何日期、里程碑或时间线。 本表是在公司未发布路线图情况下,基于最佳证据重建的版本。
[CE007, CE008, CE011, CE021, CE028, CE036]5.5 差异化与知识产权
Ineffable 的差异化主要来自人力资本,而不是 IP、产品或已部署能力。David Silver 的履历是最清晰的差异点:AlphaGo、AlphaZero、AlphaStar、AlphaFold 和 AlphaProof 代表了该领域历史上最重要的一组已发表 RL 成果,Silver 在每项成果中都担任领导或核心贡献角色。Wojciech Czarnecki、Lasse Espeholt 和 Junhyuk Oh 组成的联合创始团队,被描述为过去十年一直处在 RL 研究前沿;Espeholt 是 IMPALA 分布式 RL 架构的第一作者,该架构直接支撑 Ineffable 正在建设的这类规模化训练基础设施。 技术赌注是把基于经验的学习作为 LLM 范式的替代路径。Silver 在 Wired 采访中提出,LLM「像某种化石燃料」——是一条捷径,但最终会耗尽,因为它受人类知识边界约束;而从自身经验中学习的 RL 智能体则是「一种可再生燃料——可以不停地学、一直学、无限地学」。Silver 与 Sutton 合著的「Era of Experience」立场论文把这一论点形式化。没有任何具名现有巨头提出过与这一框架竞争的主张;最接近的类似物是 RL 增强型 LLM(OpenAI o-series、DeepSeek-R1),而 Silver 的立场实际上把它们批评为渐进式改良,而非范式跃迁。 Ineffable 没有披露专利、商业秘密、专有数据集或授权技术。公司名下没有发表论文、没有模型权重,也没有基准测试结果,无法在创始人履历之外支撑技术差异化主张。产品成熟度图(FE004)把 Ineffable 当前能力状态映射到 superlearner 栈的相关维度。 [CE023, CE024, CE025, CE034, CE035, CE038]
从四个维度评估 Ineffable 研究平台的能力与成熟度,并映射当前证据和缺口状态。
[CE017, CE019, CE020, CE021, CE023, CE026]5.6 信任、安全、合规与质量控制
截至 2026 年 6 月 22 日,Ineffable 没有发布任何安全、对齐或合规披露。没有模型卡、安全评估框架、对齐研究议程、红队协议,也没有等同于 Responsible Scaling Policy 的文件。没有提交或披露合格评定、AI Act 合规声明或英国监管备案。公司没有公开签署首尔 AI 峰会(2024 年 5 月)达成的 Frontier AI Safety Commitments;截至 2025 年 2 月,该承诺已吸引包括主要前沿实验室在内的更多签署方。 缺少安全披露本身并不证明实践不安全——Ineffable 还处在非常早期的研究阶段,尚无可评估的已训练系统。但前沿 AI 实验室的惯例是在大规模训练前发布安全框架,而不是训练之后。Anthropic 的 Responsible Scaling Policy(现为 3.3 版,2026 年 5 月更新)在每一层级训练推进前,为能力进阶设定阈值和治理关口。OpenAI 已发布 Preparedness Framework 文件。UK AI Safety Institute 已发布前沿模型评估标准。Ineffable 没有对应文件。这个缺口本身就是重大尽调证据:投资者和政策制定者无法评估训练、评估或部署决策将由什么安全治理结构约束。 英国监管环境相关,但目前较宽松。UK AI Regulation White Paper(2023)采取支持创新、按行业监管的路径,没有专门的前沿 AI 规则。EU AI Act 对具有系统性风险的通用 AI 模型提供方提出要求;一旦 Ineffable 模型训练达到阈值算力(10^25 FLOPs),很可能适用。Ineffable 注册地在英国(目前不在 EU AI Act 辖区内),但若在欧洲部署,将面临 EU Act 义务。公司没有披露合规路径、数据治理政策或隐私框架。信任与合规表(TE004)将所有相关控制项与其当前公开状态对应。 [CE017, CE018, CE019, CE020, CE030, CE032]
| 控制 / 认证 / 框架 | 当前状态(截至 2026 年 6 月 22 日) | 范围 / 适用性 | 缺口 / 尽调要求 |
|---|---|---|---|
| Model card 或能力披露 | 缺失——未发布 model card | 适用于任何训练后的 Ineffable 系统 | 索取 model card 政策,以及首个训练里程碑发布的时间表 |
| 安全评估框架(red-teaming、evals) | 缺失——未披露评估方法 | 按 UK AI Safety Institute 指引,前沿规模训练前需要具备 | 确认是否已接触 AISI;索取安全评估设计 |
| Responsible Scaling Policy 或同等文件 | 缺失——未发布 RSP 或类似文件 | 可比对象(Anthropic RSP v3.3、OpenAI Preparedness)设定能力阈值 | 索取说明训练升级控制的治理框架 |
| Frontier AI Safety Commitments(Seoul Summit 2024)安全承诺 | 截至 2025 年 2 月更新,不在签署方名单 | 自愿承诺;Anthropic、OpenAI、Google DeepMind 等已签署 | 确认 Ineffable 是否计划签署;若不签署,解释替代治理安排 |
| EU AI Act GPAI 系统性风险合规 | 缺失——无合格评定,EU 内无指定代表 | 一旦在 EU 部署,适用于训练算力超过 10^25 FLOPs 的 GPAI 模型 | 确认是否计划在 EU 部署;确认合规路径 |
| UK AI Regulation(pro-innovation white paper,2023)监管白皮书 | UK 框架非规定式;Ineffable 目前尚无具体义务 | 按行业分拆监管;截至运行日期,UK 尚无有约束力的前沿 AI 规则 | 跟踪 UK AI Act 进展;确认与 DSIT 和 AISI 的沟通 |
| 数据治理与隐私政策 | 缺失——未发布隐私政策、数据留存政策或 GDPR 声明 | 若 Ineffable 处理个人数据或用抓取数据训练,则相关 | 确认仿真输入的数据来源方法;确认 GDPR 合规计划 |
状态判断基于截至 2026 年 6 月 22 日未见任何公开文档,并与 Frontier AI Safety Commitments 签署方名单和 Anthropic RSP 版本历史交叉核对。 安全披露缺失反映的是 Ineffable 仍处训练前阶段,而非已确认不合规;但这对投资人和政策制定者构成重大治理缺口。
[CE017, CE018, CE019, CE020, CE030, CE032]5.7 技术风险
四类结构性技术风险已经存在且重要。第一,算力集中:整个研究项目依赖 NVIDIA Vera Rubin NVL72 的按时规模化交付,以及 Google Cloud AI Hypercomputer 的可用性。NVIDIA 已确认 Vera Rubin 截至 GTC 2026「全面投产」,但供应约束、超大规模云厂商(Amazon、Microsoft、Google 自身负载)之间的优先分配,以及软件成熟度仍未知。如果 NVIDIA Vera Rubin 爬坡慢于预期或供应被转移,Ineffable 的基础设施时间线会直接受威胁。Ineffable 获得优先或保证供应的条款没有披露。 第二,模拟到现实的落差:完全在模拟环境中训练的 RL 系统一旦应用到分布外环境,就有有据可查的失败模式。Ineffable 用于生成经验的模拟环境质量和多样性未知。目标能力越抽象、越多样(例如科学发现、数学证明),就越难指定奖励函数和环境,使系统产生有用的泛化,而不是奖励黑客或狭窄解法。Ineffable 没有披露任何模拟环境设计、奖励模型架构或领域覆盖策略。 第三,泛化未被证明:Ineffable 在 DeepMind 的前身成果(AlphaGo、AlphaZero、AlphaProof)在边界清晰、完全可观测、奖励信号明确的组合领域取得了超人表现。把这种方法泛化到开放式现实世界知识发现——奖励信号模糊、环境部分可观测、领域无边界——是难度质变的研究挑战。任何公开披露中都没有证据证明这一鸿沟已经跨越。 第四,关键人物与团队扩张风险:使命与 David Silver 的研究愿景直接绑定。科学项目依赖留住 Silver 和创始团队,也依赖在竞争激烈的人才市场中招聘更多前沿 RL 研究员。Wired 采访描述团队正在从「只献身于这一使命的杰出个体」中组建,暗示人才画像窄而深。创始团队任何波动,都可能实质影响科学项目和投资者信心。 [CE006, CE007, CE026, CE027, CE028, CE029]
5.8 图表证据
06客户
6.1 客户基础分层——收入前基线
截至 2026 年 6 月 22 日,Ineffable Intelligence 没有商业客户。公司按设计完全处于收入前阶段。它公开的使命是开展前沿 AI 研究,尤其是构建一个能够从自身经验而非人类数据中发现新知识的「superlearner」系统。公司明确表示,它运营时「不屈从于渐进式产品和短期利润的要求」。目前没有产品、API、定价页、企业级产品或面向客户的服务。 因此,Ineffable 有意义的客户分层只能是前瞻性的——也就是如果 superlearner 平台实现科学目标,谁可能购买或许可其产出。公开记录和公司叙事中浮现出五类潜在买方。第一,主权 AI 实验室和国家 AI 项目(例如 UK Sovereign AI Fund 本身)是近期制度契合度最高的细分:承担建设本国 AI 能力使命的政府,可能寻求 superlearner 发现成果的访问权。第二,超大规模云厂商——Google Cloud、AWS、Microsoft Azure——可能通过许可 superlearner 能力并嵌入自身 AI 服务而成为客户,这与 Google Cloud 现有基础设施合作关系一致。第三,制药、材料科学、气候研究和药物开发领域的科研组织,是能够独立发现新科学知识的 AI 的高价值终端用户。第四,企业研发组织(在技术、工程和金融领域有内部研究任务的大公司)在能力阈值被证明后,构成更广泛商业市场。第五,寻求自主科学发现能力的国防和国家安全机构补足潜在买方图谱。 除当前投资者 / 合作伙伴关系外,上述细分均没有披露与 Ineffable 的关系。客户分层表(TU001)把五类细分与买方 / 付款方角色、用例、规模潜力、战略价值和证据缺口逐一对应。客户旅程图(FU001)展示不同细分的概念性采用路径。 [CU001, CU002, CU003, CU013, CU014, CU020]
| 细分 | 买方 / 用户 / 付款方 | 用例 | 规模潜力 | 收入 / 战略价值 | 证据缺口 |
|---|---|---|---|---|---|
| Sovereign AI 实验室 / 国家 AI 计划 | 政府机构(付款方和名义用户) | Superlearner 发现的授权权利;本土 AI 能力获取;政策杠杆 | 中——受政府预算周期限制;可能形成大型锚定合同 | 对 UK/EU 政府具备高战略价值;近期收入潜力中等 | 无 LOI、MOU 或采购记录;UK 持股仅为投资人股权 |
| Hyperscalers(Google、AWS、Azure) | 科技公司(付款方 = 被授权方或云基础设施用户) | 将 superlearner 能力嵌入云 AI 服务;先发授权 | 极大——云 AI 市场;hyperscalers 拥有深厚研发预算 | 若 superlearner 能力可大规模授权,收入可能被彻底改写 | 现有 Google Cloud 关系中,Ineffable 是买方而非反向供应方;没有授权交易 |
| 科学研究组织(制药、材料、气候) | 研究机构和生物制药研发实验室(付款方和用户) | AI 自主药物靶点发现;材料属性预测;气候建模 | 大——全球企业制药和材料科学研发预算超过 $200B | 高;经验证的药物发现科学突破可带来可观授权费 | 未披露试点、合作或商业化前协议;IP 条款未解决 |
| 企业研发组织 | 有内部研究职责的大型企业(付款方) | 面向产品创新、工程和竞争情报的自主发现 | 大——全球企业研发支出超过 $1T;但 AI 预算竞争激烈 | 中;企业承诺前需要证明特定领域突破 | 未披露企业沟通;没有产品或 demo 可转化兴趣 |
| 国防 / 国家安全机构 | 政府国防机构和国家实验室(付款方) | 面向战略优势的自主科学发现;材料、密码学、物流 | 中——国防 AI 预算增长,但采购保密且缓慢 | 战略价值高,但难以透明变现;这里采购摩擦最大 | 未披露国防关系;治理和出口管制复杂性未解决 |
所有细分都只是潜在方向;截至 2026 年 6 月 22 日,公司没有商业客户或付费合同。收入 / 战略价值估算参考公开市场报告和可比前沿 AI 授权, 但仍属推测。所有缺口均由公开来源没有任何客户披露所确认。
[CU013, CU014, CU020, CU021]Ineffable Intelligence 超级学习者平台的潜在客户分群,以及从认知到扩张的概念性采用路径。
机构验证之后的所有旅程阶段都只是前瞻性、概念性描述。公司未披露任何商业客户接触。该旅程基于类似前沿 AI 从研究走向商业化的模式。
[CU013, CU021, CU031]6.2 采用轨迹——合作伙伴信号与商业证明缺失
传统采用轨迹指标——活跃用户数、部署实例、年度合同价值、复购率——并不存在于 Ineffable Intelligence,因为公司没有发布商业产品。现阶段无法用 SaaS、API 或企业软件客户指标评估公司。 最接近采用准备度的公开代理指标,是 Ineffable 2026 年 4 月出隐身后数月内宣布的战略合作。NVIDIA 宣布一项工程级协作,共同设计强化学习基础设施栈,从 NVIDIA Grace Blackwell 起步,并延伸至即将推出的 Vera Rubin 平台。该协作涉及两家公司工程师共同建设训练管线,是 Ineffable 已披露的最具运营实质的外部关系。2026 年 6 月,Google Cloud 宣布首选合作伙伴协议,Ineffable 将在 Google Cloud AI Hypercomputer 上部署由 NVIDIA Vera Rubin NVL72 驱动的最大 A5X GPU 集群之一。两项关系都传递出基础设施层面的技术可信度,但它们是「合作伙伴即基础设施客户」关系(Ineffable 是算力买方),不是产品客户关系。 公共部门侧,UK Sovereign AI Fund 和 British Business Bank 合计向种子轮承诺约 $20 million;Sovereign AI Fund 还提供英国最大 AI 超级计算机访问、签证支持,以及「英国国家的独特杠杆」。这代表对研究项目的制度性认可,但不构成产品采用。公司或任何第三方来源均未披露试点项目、意向书、商业化前协议或客户管线。采用轨迹表(TU002)在这一基线下记录所有可衡量信号,采用漏斗图(FU002)映射从研究走向商业化的理论路径各阶段,以及漏斗目前停住的位置。 [CU004, CU005, CU006, CU007, CU008, CU015]
| 指标 | 数值 | 日期 | 来源 | 置信度 | 含义 | 缺失分母 / 缺口 |
|---|---|---|---|---|---|---|
| 付费客户(数量) | 0 | 2026-06-22 | 公开披露复核(未宣布客户) | 高 | 公司完全处于收入前阶段;没有商业产品 | 未披露 pipeline、LOI 或商业化前协议 |
| 年度经常性收入(ARR) | 2026-06-22 | 未找到收入来源 | 高 | 现阶段不适用 | 产品前阶段不适用收入概念 | |
| 战略技术伙伴(基础设施) | 2 (NVIDIA, Google Cloud) | 2026-06-22 | NVIDIA 2026 年 5 月博客;Google Cloud 2026 年 6 月新闻稿 | 高 | 技术可信度信号;不代表商业客户 | 没有客户部署;Ineffable 是基础设施买方,不是供应方 |
| 公共部门投资人 / 验证方 | 2 家(UK Sovereign AI Fund、British Business Bank) | 2026-06-22 | British Business Bank 2026 年 4 月公告 | 高 | 机构以投资人身份背书;不是产品购买方或用户 | $20M BBB 持股占本轮 <2%;按专家评论,治理权有限 |
| 招聘岗位(人才需求信号) | 多个开放岗位(研究工程师、ML 科学家) | 2026-06-22 | jobs.ashbyhq.com/ineffable/,2026 年 6 月访问 | 中 | 活跃招聘显示研究项目仍在推进;未见销售 / GTM 招聘 | 未列出客户成功、销售或客户管理岗位 |
所有指标都反映 Ineffable Intelligence 截至 2026 年 6 月 22 日处于收入前、产品前状态。合作伙伴和投资人关系与商业客户分开记录。空值 表示数据不存在,而不是未披露的正向数字。
[CU001, CU002, CU004, CU005, CU006, CU015]从公众认知到商业生产的分阶段采用漏斗,展示截至 2026 年 6 月 Ineffable Intelligence 的采用管线停在哪一环。
伙伴 / 支持方阶段的漏斗数值代表实体数量,不是标准化百分比。认知阶段为定性描述。漏斗说明公司顶部可信度强,但商业转化为零。
[CU001, CU004, CU005, CU015, CU025]6.3 具名客户证明——验证者和合作伙伴不是客户
客户证明的尽调标准,需要具名部署、明确用例、可衡量结果,以及生产阶段使用或至少活跃试点的证据。截至 2026 年 6 月,Ineffable Intelligence 一项标准也不满足。没有具名付费客户、没有披露试用、没有案例研究、没有 G2 或 Gartner Peer Insights 评论、没有具名用户的会议发言、没有政府采购记录,也没有引用客户产品使用的新闻稿。 最接近客户类比关系的实体,必须清晰区分。NVIDIA 是技术伙伴和基础设施联合开发方——它与 Ineffable 工程师共同设计 RL 训练管线,但这是 Ineffable 受益于 NVIDIA 硬件的供应商—客户关系,不是任何组织部署或购买 superlearner 产品的关系。Google Cloud 是 Ineffable 的首选基础设施提供方——Ineffable 是 Google Cloud 的客户,而非相反。UK Sovereign AI Fund 和 British Business Bank 是持有股权的财务投资者;政府宣布的「下一轮优先拒绝权」是投资选择权,不是产品购买或部署证据。按标准尽调口径,这些实体都不构成产品客户。 具名客户证明表(TU003)列出最接近客户状态的三个实体,并明确记录其真实关系类型;客户证明矩阵(FU003)映射这些维度上的证据质量。TU003 中没有任何合格行,本身就是重大尽调发现。David Silver 和公司没有披露任何目标客户接触或商业对话,这与其声明的优先研究、而非近期商业化一致。Google Cloud 的 Thomas Kurian 表示 Ineffable 正在「利用我们的全栈 AI Hypercomputer」——这确认 Ineffable 的角色是基础设施消费者,而不是拥有自身客户的产品供应商。 [CU017, CU018, CU019, CU026, CU033, CU034]
| 实体 | 实际关系类型 | 细分 | 部署 / 用例 | 生产 vs 试点 | 结果证据 | 限制 / 为什么不是客户 |
|---|---|---|---|---|---|---|
| NVIDIA Corporation | 技术共同开发伙伴 / 硬件供应商 | 基础设施伙伴 | 在 Grace Blackwell 和 Vera Rubin 平台上共同设计 RL 训练管线 | 都不是——产品前研究合作 | 联合博客描述管线共同设计;Jensen Huang 背书 | NVIDIA 是供应商;Ineffable 获得硬件访问,而不是产品购买方 |
| Google Cloud (Alphabet) | 首选基础设施云提供商 | 云基础设施供应商 | 为超级学习器研究提供 A5X GPU 集群和 AI Hypercomputer | 均不是——Ineffable 是 Google Cloud 的客户,不是反向关系 | Google Cloud 2026 年 6 月新闻稿;Thomas Kurian 声明 | Ineffable 向 Google Cloud 购买算力;没有产品售予 Google Cloud |
| UK Sovereign AI Fund / British Business Bank 公共投资方 | 股权投资方和机构背书方 | 公共部门投资方 / 验证方 | 基金使命要求建设本土 AI 能力;Sovereign AI 提供超级计算机访问 | 均不是——这是股权投资,附带「下一轮优先拒绝权」,不是产品使用 | British Business Bank 公告;Sovereign AI Fund 2026 年 4 月帖文 | 仅为投资者关系;治理专家指出缺少可执行的产品使用权 |
本表列出与 Ineffable Intelligence 最接近类客户关系的三类实体。按标准商业尽调定义,三者都不构成产品客户。表名沿用既定 schema 要求;实质结论是证据缺口——没有任何具名产品客户。
[CU017, CU018, CU019, CU034, CU035]对与 Ineffable Intelligence 存在关系的实体,按关系类型(行)和证明维度(列)映射证据质量。
矩阵单元格反映证据缺失,而不是存在但未确认。所有空白 / None 单元格均经截至 2026 年 6 月 22 日的公开来源穷尽审查确认。
[CU017, CU018, CU019, CU026, CU027]6.4 留存、持久性与满意度——商业化前真空
净收入留存、毛收入留存、客户流失、合同续签率、分同期群留存、客户满意度分数和 NPS 数据,对 Ineffable Intelligence 都不存在,因为没有商业产品产生收入,也没有获取客户。所有标准留存和持久性指标在现阶段都不适用。 这一阶段公司的留存问题,应改用结构性持久性代理指标回答:研究合作模式的粘性、Sovereign AI Fund 承诺的持久性,以及 Ineffable 对 Google Cloud 多年基础设施承诺所暗含的锁定。Google Cloud 基础设施合作和 NVIDIA 联合设计关系,都暗示研究项目具备中期运营连续性,但这些不是客户留存指标——它们是供给侧承诺。 人才市场提供了一个间接持久性信号:招聘信息确认 Ineffable 仍在招聘高级研究员和工程师。公司能从 Google DeepMind 和其他前沿实验室吸引研究员,是研究项目在技术社区内仍具可信度的代理指标。但在商业产品出现前,没有客户满意度调查、Net Promoter Score、评论平台条目或任何用户层反馈可用,也不应期待其出现。 留存表(TU004)用明确空值捕捉所有可得持久性指标,并为每个缺失数据点提出尽调要求。留存同期群图(FU004)在乐观、基准和悲观商业化假设下,按买方细分类型展示未来留存的情景分析,并明确标注为预测分析。 [CU016, CU023, CU028, CU029, CU030, CU033]
| 指标 | 数值 / 状态 | 适用细分 | 置信度 | 尽调要求 |
|---|---|---|---|---|
| 净收入留存率(NRR) | null — 不存在客户 | N/A | 高(已确认不存在) | 商业化里程碑时再询问;未来任何条款清单都必须要求 NRR 目标和续约结构 |
| 总收入留存率(GRR) | null — 不存在客户 | N/A | 高(已确认不存在) | 同上;投入前确保合同条款包含最低年度续约率 |
| 客户满意度 / NPS | null — 未部署产品 | N/A | 高(已确认不存在) | Series A 前要求 beta / 试点满意度调查;任何增长轮都需要参考客户 |
| 研究项目连续性(代理指标) | 活跃——截至 2026 年 6 月,NVIDIA 和 Google Cloud 合作已确认 | 基础设施合作伙伴 | 高(合作公告已核验) | 跟踪算力合同续约条款;弄清 Google Cloud 承诺期限和退出条款 |
公司仍处于收入前阶段,所有商业留存指标均不适用。研究项目连续性是唯一可衡量的耐久性代理指标,但它对应供应商关系,不是客户留存。null 值表示数据不存在。
[CU016, CU023, CU028, CU029]在乐观、基准、悲观商业化假设下,按买方分群类型预测留存率——所有数字都是分析预测,不是历史数据。
所有数值都是分析师预测,基于可比前沿 AI 研究访问协议和企业研发授权续约率。没有实际客户分组数据。悲观情景(30/20/15%)模拟没有承诺锚定合同的低留存交易式访问。乐观情景参考主权实验室多年研究服务的留存模式。三种情景都假设已获得至少一个付费客户;截至运行日期,这尚未发生。
[CU029, CU030, CU039]6.5 扩张路径与集中风险分析
Ineffable Intelligence 走向商业客户,需要穿过几个不同阶段,每个阶段都依赖前一阶段成功。第一阶段——研究项目执行——必须证明可信的技术突破,商业接触才可能发生。第二阶段需要把突破转化为可部署能力或研究访问产品。第三阶段需要识别并签下首批客户合同;对前沿 AI 研究产出而言,这通常是长周期企业或政府采购。只有在初始客户基础形成后,先落地再扩张的动态才会启动。 在客户前阶段,集中风险极端。公司的整个基础设施依赖集中在单一云提供商(Google Cloud)和单一硬件合作伙伴(NVIDIA)。任何商业、技术或政治层面的关系扰动,都会实质威胁研究项目。公共部门侧,英国政府 $20 million 股权相对 $1.1 billion 总融资太小,基本无法对 Ineffable 方向形成结构性杠杆;newspage.news 引述的 AI 治理专家称,这只带来「约百分之一影响力」。 可能买方细分的采购摩擦很高。主权 AI 实验室和国家 AI 项目按年度预算周期运行,采购时间线跨多年。制药和材料科学科研组织有复杂的 IP 所有权要求,任何 AI 生成发现用于商业之前都必须先谈妥。企业 R&D 买方在承诺前需要概念验证里程碑和供应商验证。扩张与集中风险表(TU005)按严重程度评估和尽调路径映射这些风险。 [CU010, CU011, CU012, CU024, CU027, CU031]
| 扩张驱动 / 风险因素 | 集中风险 | 实质化后的影响 | 尽调路径 |
|---|---|---|---|
| Google Cloud 是唯一披露的算力提供商 | 关键——整个研究基础设施依赖单一云提供商 | 研究项目中断;失去 A5X 集群访问;多年延误 | 获取承诺期限;要求多云应急方案;审查 SLA 和退出条款 |
| NVIDIA 硬件依赖 | 高——整个架构押注 Vera Rubin NVL72 路线图 | 若 Vera Rubin 出货延迟或联合设计关系变化,技术进度会被拖慢 | 审查硬件供应条款;确认 Blackwell 备选方案;压力测试 NVL72 时间线 |
| 英国政府持股过小,难以取得治理权 | 重大——<2% 的轮次份额换不来多少控制权;专家称缺少治理条件 | 公共部门客户转化风险;政治撤退;负面媒体叙事 | 要求完整治理条款清单;就政府条件取得独立法律审查 |
| 目标买方群体采购周期长达多年 | 高——主权实验室和企业 R&D 通常按 18-36 个月采购周期运作 | 从研究突破到首个商业合同,收入可能推迟 4-7 年以上 | 绘制买方细分采购日历;及早接触主权 AI 实验室,争取 LOI 信号 |
| 拿下首个客户前不存在 land-and-expand 动态 | 阻断——没有可扩张的客户基础;采购也没有参考客户 | 无法向未来投资者证明采用曲线,也难以拿下下一层级买方 | 未来 18 个月优先锁定一个锚定客户(主权实验室或科学研究机构) |
风险等级由分析师基于公开披露和可比前沿 AI 研究公司动态评估。没有可用于验证的运营数据。所有尽调路径都是成长阶段投资承诺前的最低要求。
[CU010, CU024, CU031, CU039, CU040]6.6 图表证据
07风险
7.1 监管、法律与主权风险
Ineffable Intelligence 身处快速变化的监管环境,却没有稳定法律框架可依。公司注册并主要运营于英国;截至 2026 年 6 月,英国对 AI 监管采取「支持创新」路径——有意避免行业专项立法,转而依赖现有监管机构。这个姿态减轻了短期合规负担,却带来前瞻监管风险:一旦英国 AI 法案落地并加入有约束力的安全或算力义务,Ineffable 的训练项目可能面临追溯合规成本或运营限制。 EU AI Act 自 2024 年 8 月开始分阶段执行,带来实质域外风险。Ineffable 的 superlearner 一旦在任何欧盟辖区部署或营销,几乎必然会按该法规第 3(63) 条被认定为通用 AI(GPAI)模型。GPAI 提供商必须准备技术文档、维护最新模型登记;若按第 51 条被归为「系统性风险」,还需接受对抗测试并报告事故。公司尚未披露任何 EU AI Act 合规方案。 2021 年《National Security and Investment》(NSI)法案对投资人构成结构性法律风险。该法赋予英国国务大臣权力,可审查并阻止、或附条件批准对 AI 技术领域实体取得重大控制权的交易。未来任何二级融资若把重大控制权转移给非英国投资人,都可能触发强制通知义务。Sequoia Capital(美国)和 Lightspeed Venture Partners(美国)已持有董事会席位;任何进一步提高外资持股的动作都需要法律意见。 Information Commissioner's Office(ICO)关于 AI 与数据保护的指引明确,部署 AI 的组织必须适用 UK GDPR 的问责与治理原则,包括默认数据保护、目的限制,以及自动化决策公平性。尽管 Ineffable 称 superlearner 不依赖人类数据学习,RL 训练闭环仍会生成、处理并评估经验。系统在评估任务中是否接触个人数据,仍未披露。公司没有公开确认的数据保护官、数据处理协议或 ICO 沟通记录。 知识产权方面:David Silver 在 Google DeepMind 工作约二十年,多名同事也来自 DeepMind 和 UCL。Silver、其 UCL 教授身份与 Ineffable 公司实体之间的 IP 转让协议尚未公开披露。如果任何核心算法部分形成于 DeepMind 或 UCL 任职期间,商业化临近时 IP 归属可能发生争议。公司位于 Altrincham 的注册地址(专业服务地址)并不显示其拥有独立于运营公司的正式 IP 持有实体。 英国政府宣布共同投资后,独立专家直接提出了主权治理担忧。Newspage 和 Electronics Weekly 引述的评论人士指出,政府通过 British Business Bank 和 UK Sovereign AI Fund 持有少数股权,并不赋予政府对所创造 IP、发现成果或技术下游用途的任何监督权。Sovereign AI Fund 自身公布的投资策略确认,其使命是财务回报加公共利益,但 Ineffable 投资所附带的具体治理保护,没有出现在任何公开文件中。 [CR001, CR002, CR003, CR004, CR005, CR006]
| 风险 / 规则 / 案例 | 司法辖区 | 状态 | 可能性 | 严重性 | 主要缓释措施 | 剩余敞口 | 尽调路径 |
|---|---|---|---|---|---|---|---|
| EU AI Act,GPAI 条款(Art. 3(63)、Art. 51);超级学习器一旦在欧盟部署,很可能被认定为具系统性风险的通用 AI 模型 | 欧盟 | 生效(2024 年 8 月义务;2026 年 8 月全面部署期限) | 高——缺少 opt-out 证据时应按默认情形处理 | 高——技术文档、透明度、对抗性测试义务;若系统性风险评估未通过,可能被禁 | 任何欧盟部署前先对 GPAI 门槛作法律分析;接触 DPA;预登记 GPAI 登记册 | 重大——合规计划尚未启动;未披露欧盟法律顾问 | 获取 GPAI 适用性法律意见;确认合规计划启动前不会面向欧盟用户部署 |
| 截至 2026 年 6 月,英国没有有约束力的 AI 专项立法;AI Regulation White Paper(2023)采用由既有监管机构分工执行的原则性路径;UK AI Bill 尚未提交 | 英国 | 结构性——预计至少持续 2–3 年;若 AI Bill 引入行业专项义务,风险会上升 | 中——2026–2027 年可能立法 | 中——有约束力的算力门槛、强制安全审计或模型登记可能带来合规成本和运营限制 | 主动接触 DSIT / AISI;参与自愿安全承诺框架;签署 Bletchley 承诺 | 重大——未披露政府联络职能;独立评论者指出治理缺口 | 确认 Ineffable 是否已接触 AISI;要求提供任何自愿安全承诺文件副本 |
| National Security and Investment Act 2021 要求非英国收购方取得英国 AI 实体重大影响力时履行通知义务;Sequoia 和 Lightspeed(美国机构)在种子轮即持有董事席位 | 英国 | 生效——NSI Unit 已运行;AI 属强制通知行业 | 种子轮低(既有结构);若未来轮次外国投资者增持,则为中 | 高——政府可能阻止二级收购、对既有外国投资者权利附加条件,或要求结构性补救 | 维持英国多数董事会控制;对未来任何外国投资者交易作法律审查;可通知收购须强制预通知 | 重大——未公开披露 NSI 法律意见或合规框架 | 要求 NSI 法律意见;确认拟议 Series A 投资者国籍结构和董事会结构 |
| ICO 关于 AI 与数据保护的指引(2023 年 3 月)要求 AI 系统落实默认数据保护、问责、目的限制和公平性义务;UK GDPR 适用于 RL 评估任务中处理的个人数据 | 英国 | 执行中——ICO 正在主动开展 AI 审计并发布指引 | 中——RL 管线可能不处理个人数据;未知 | 中——ICO 调查可能导致执法通知、罚款,或要求修改训练管线 | 任命 Data Protection Officer;委托数据映射;确认 RL 评估环境是否涉及个人数据 | 中等——未披露 DPO、隐私政策或 ICO 接触 | 确认 DPO 任命;审查 RL 数据管线的个人数据暴露;要求 DPA 审计计划 |
| 部分算法在 UCL / DeepMind 任职期间开发,可能存在 IP 归属争议;未公开披露 IP 转让协议;UCL 教授身份仍在;学术发表压力可能抢在专利提交前公开成果 | 英国 / 全球 | 潜伏——商业化或专利提交时会急剧升温 | 中——前雇主 IP 条款常见;UCL IP 政策适用于员工发明 | 高——核心 RL 算法 IP 争议可能损害许可、估值和商业部署 | 完整 IP 转让契约;UCL IP 许可或转让;覆盖 Silver 既往工作的 freedom-to-operate 意见 | 重大——未披露 IP 登记册、专利组合或转让情况 | 要求 IP 转让协议;确认 UCL IP 政策;获取 freedom-to-operate 法律意见 |
本表覆盖截至 2026 年 6 月 22 日从公开来源识别的风险。未披露的数据室材料(法律意见、IP 登记册、DPA 文件)会显著改变剩余敞口评级。可能性和严重性依据公开证据评估;并非法律意见。
[CR001, CR002, CR003, CR004, CR005, CR006]7.2 运营、技术与安全风险
Ineffable Intelligence 的运营风险主要由三类相互咬合的问题主导:算力脆弱性、RL 扩展不确定性,以及安全与对齐治理完全缺位。 算力脆弱性是眼下最可衡量的运营风险。Ineffable 的整条训练管线围绕 NVIDIA Grace Blackwell 和下一代 Vera Rubin NVL72 平台共同设计,并计划部署在 Google Cloud AI Hypercomputer 的 A5X 集群上。NVIDIA 于 2026 年 5 月发布 Vera Rubin NVL72;面向 Ineffable 的生产配额时间表和容量承诺并未公开披露。NVIDIA 服务的超大规模客户队列庞大且仍在增长——Microsoft、Google、Meta、Amazon,以及全球主权 AI 项目都可能争夺产能。没有披露的第二算力安排。一旦出现重大延误或产能重新分配,研究项目就会停摆。 Ineffable 所需尺度和通用性的强化学习,在开放式环境中尚未被技术验证。此前里程碑式 RL 成果——AlphaGo、AlphaZero、AlphaStar——都运行在边界清晰、奖励信号明确的游戏环境里。把这些方法扩展到开放世界知识发现,会引入尚未解决的挑战:奖励黑客(系统优化代理奖励,而非真实目标)、规格博弈(满足字面规格,而非其精神),以及仿真到现实的鸿沟(仿真中学到的技能无法迁移)。学术文献已充分记录这些失败模式,Ineffable 也未披露任何缓释方法或评估框架。Technology Review 曾指出,多个研究项目中的 RL 扩展已撞墙。 安全与对齐风险对系统性风险最关键。Ineffable 的既定目标是「与超级智能进行首次接触」——也就是打造能力超越人类水平的系统。截至 2026 年 6 月 22 日,公司没有公开披露安全框架、负责任扩展政策(RSP)、对齐研究议程、红队项目,也没有披露安全负责人(没有安全主管、安全委员会或外部安全顾问机构)。Anthropic、OpenAI 和 Google DeepMind 都已发布 RSP 或等同安全承诺,Ineffable 却没有类似披露。这不只是治理缺口;它是商业化前的缺口,能力越往前推进,越难补上。Seoul Summit 的 Frontier AI Safety Commitments 已由主要前沿实验室签署,但 Ineffable 未公开背书。 superlearner 尺度下,能源和基础设施风险也很实质。前沿 RL 需要紧密的训练—推理闭环,单位 FLOP 能耗高于标准预训练负载。总电力和冷却需求尚未披露。 网络安全风险偏高。Ineffable 的模型权重和训练工件一旦积累起来,就是极其宝贵的知识产权。公司未披露任何网络安全态势、数据驻留政策或事故响应框架。 [CR009, CR010, CR011, CR012, CR013, CR014]
| 失败模式 | 可能性 | 严重性 | 缓释成熟度 | 剩余敞口 | 未解决缺口 |
|---|---|---|---|---|---|
| NVIDIA 将 NVL72 产能重新分配给竞争性 hyperscaler 或主权 AI 客户;或产能爬坡晚于承诺时间线 | 中——NVIDIA 面临大量竞争需求;截至 2026 年 6 月,NVL72 尚未 GA | 关键——训练项目停摆;没有替代硬件能达到同等 RL 闭环性能 | 低——未披露合同分配保证;联合设计合作略有缓释 | 关键 | NVIDIA 合作商业条款未披露;未确认分配保证;没有第二套算力安排 |
| 超级学习器优化代理奖励函数,而非预期目标;RL agent 学会钻评估空子,而不是获得真实能力 | 高——这是所有大规模 RL 都有记录的失败模式;开放式奖励环境尤其突出 | 高——可能让数月训练失效并浪费巨额资金;能力声明变得无法核验 | 低——未披露评估方法、奖励设计框架或 red-teaming 项目 | 高 | 未发布奖励规格、评估框架或基准套件 |
| 超级学习器取得显著能力,但未披露对齐框架;系统追求的目标偏离预期;没有内部或外部 red-teaming 提前发现 | 中——公司明确瞄准超级智能;没有安全项目 | 关键——若高度能力但未对齐的系统被部署,可能造成不可逆伤害;也会触发 EU AI Act 下系统性风险认定 | 关键-低——没有 RSP、没有安全团队、没有 red-teaming;截至 2026 年 6 月,缓释完全缺位 | 关键 | 没有 RSP、没有安全负责人、没有对齐研究议程、没有外部安全顾问委员会;这是论文级别的担忧 |
| 对抗性入侵(国家行为体或竞争对手)窃取已训练模型权重、算法规格或专有训练基础设施 | 中——前沿 AI 实验室是高价值目标;未披露安全态势 | 关键——竞争地位、IP 价值和投资者信心都会严重受损 | 未知——未披露网络安全、SOC、渗透测试或保险确认 | 关键 | 未公开披露网络安全态势、数据驻留政策或事件响应框架 |
运营风险按严重性排序。缓释成熟度仅依据公开披露评估;未披露的内部控制不能计入。
[CR009, CR010, CR011, CR012, CR013, CR016]7.3 伙伴与基础设施依赖风险
Ineffable Intelligence 的基础设施建在一组很窄的关键依赖之上;其中任何一个环节都可能独立损害研究项目。 NVIDIA 是集中度最高的依赖。Vera Rubin NVL72 集群为 Ineffable 的 RL 工作负载定制共同设计;没有其他硬件能在 superlearner 所需训练—推理紧密闭环中匹配其互连和内存带宽规格。NVIDIA 还对 Ineffable 进行了股权投资,因此同时扮演投资人和唯一硬件供应商。硬件协议的商业条款——定价、配额保证、排他性,以及升级到未来 NVIDIA 世代的路径——完全未披露。 Google Cloud 是第二个关键依赖,作为 Ineffable 部署 AI Hypercomputer A5X / Vera Rubin NVL72 集群的「首选云提供商」。这项合作于 2026 年 6 月宣布;商业条款同样未披露。Google 也单独投资了 Ineffable,形成平行的投资人—云提供商双重角色。Google Cloud 同时服务 Google DeepMind、OpenAI(通过竞争对手 Microsoft Azure),以及自身内部 AI 项目;Ineffable 没有披露相对于这些竞争优先事项的优先级或容量保证。 资本依赖集中在领投人层面。Sequoia Capital 和 Lightspeed Venture Partners 共同领投种子轮并持有董事会席位。按 VC 支持公司的惯例,种子轮领投人通常被期待为下一轮背书并继续领投。如果任一机构对技术进展或竞争格局产生保留,它们退出 Series A 背书会被其他投资人解读为强烈负面信号——也就是「抽梯子」情景。 British Business Bank 和 UK Sovereign AI Fund 合计在种子轮投入约 $30–40 M。其治理权和信息权未公开披露。如果政府调整 AI 政策,或认定投资中的公共利益条款没有得到满足,即便没有合同追索权,也可能制造声誉和政治压力。 人才管线依赖 David Silver 的个人学术网络。已披露且具名来源的多数雇员来自 DeepMind 和 UCL。Silver 既是 CEO,也是主要招聘者。如果他离开,不仅科学方向承压,人才管线也很可能同时枯竭。 [CR017, CR018, CR019, CR020, CR021, CR022]
| 依赖 | 交易对手 | 角色 | 集中度 | 失败情景 | 严重性 | 缓释措施 | 剩余敞口 |
|---|---|---|---|---|---|---|---|
| 主要算力硬件 | NVIDIA | 唯一硬件合作伙伴;Vera Rubin NVL72 联合设计;投资者 | 关键——同等 RL 闭环性能下没有可行替代;NVIDIA 同时也是投资者,形成双重角色 | NVL72 分配被转向或推迟;产能爬坡滑坡;NVIDIA 重新优先服务 hyperscaler 客户 | 关键——训练项目停摆 | 联合设计合作和投资者关系提供一定筹码;商业条款未披露 | 关键——商业条款未披露;单一来源依赖;未发布备选方案 |
| 云基础设施 | Google Cloud | 首选云提供商;AI Hypercomputer A5X / Vera Rubin 部署;投资者 | 部署高度关键;Azure / AWS 可替代,但联合设计基础设施带来高切换成本 | GCP 故障、商业争议,或 Google 将优先级转向自身 AI 工作负载(DeepMind、Gemini) | 高——训练显著中断;切换成本高 | 非独家首选安排;SLA 条款未披露;Google 同时也是投资者(双重角色) | 高——SLA 未披露;未披露多云应急;Google 自身 DeepMind 工作负载的战略优先级更高 |
| 下一轮资本 | Sequoia Capital 和 Lightspeed Venture Partners(共同领投) | 董事会控制(董事席位);后续融资信号;创始人支持 | 高——按惯例,种子轮领投方预期会领投或为 Series A 释放信号;董事席位带来治理影响 | 投资者信心崩塌;任一机构拒绝参与 Series A;共同领投方冲突 | 关键——向市场释放负面信号;Series A 难度大幅上升;可能出现契约或董事会冲突 | Alfred Lin(Sequoia)和 Ravi Mhatre(Lightspeed)是经验丰富的前沿 AI 投资人;当前关系看起来正向 | 高——治理文件未披露;投资者权利协议未公开;Series A 时间线取决于尚未定义的技术里程碑 |
| 政府合法性和监管通道 | UK Government(DSIT、British Business Bank、Sovereign AI Fund)英国政府主体 | 共同投资者;监管对话渠道;公共合法性提供方 | 中——政府持股带来可信度和监管入口,但不带来对 British Business Bank 支持的合同权利 | 政策变化(AI Bill 通过);2029 年选举后政府更替;Sovereign AI Fund 使命变化;外界批评交易条款 | 中——声誉和监管通道风险;除非政策性约束义务生效,否则没有直接运营影响 | 少数投资者关系;BBB 董事会代表(未披露);持续对齐公共利益 | 中——治理权未披露;专家已公开指出,当前持股结构并未确保公共利益义务 |
| 研究人才管线 | UCL 和前 DeepMind 网络 | 主要高级研究员招聘渠道;学术可信度锚点 | 高——Silver 的个人网络是已识别的前沿 RL 人才主要来源;未披露 HR 职能或独立招聘 | Silver 离开会带走网络;UCL 关系变化;DeepMind 人才被反向报价或竞业限制锁住 | 高——人才流失风险直接拖慢研究项目速度 | 维持 UCL 任职关系;NVIDIA 和 Google Cloud 背书带来额外招聘信号 | 高——未披露独立人才管线;招聘和科学路线都存在单人依赖 |
合作伙伴风险登记表仅反映公开披露。三项基础设施合作(NVIDIA、Google Cloud、British Business Bank)的商业条款均未披露。严重性评级假设未来 12 个月内的合理最坏情形。
[CR017, CR018, CR019, CR020, CR021, CR022]7.4 财务、资本与模式风险
Ineffable 的财务风险主要来自两个因素叠加:完全没有商业活动,以及前沿算力层极高的资本强度。 $1.1 B 种子轮是欧洲史上最大种子轮,但它是否足以支撑使命仍未知,因为公司未披露烧钱速度、员工人数、资金用途拆分或算力成本估算。基于 Epoch AI 对前沿训练成本的建模(显示成本每 9–18 个月翻倍)以及 Sequoia 对领先前沿实验室每年在算力上花费数十亿美元的观察,$1.1 B 可能只够支撑所需节奏下 18–36 个月的 runway。这意味着公司需要在 2027 年末或 2028 年完成 Series A——届时任何产品可能都尚未部署或获得商业验证。 公司明确拒绝短期商业产出,称希望拥有「一段让雄心勃勃的研究蓬勃发展的窗口,不必屈从于增量产品和短期利润的要求」。战略上这自洽,但也带来投资人耐心风险:如果前沿 AI 融资环境改变(正如 Sequoia 2023 年「Act Two」文章所警告),收入前前沿实验室的募资条件会困难得多。Wired 和 FT 的报道已指出,前沿 AI 资本配置正转向能够证明产品牵引力的实验室。 收入模式风险高。公司没有定价模型、商业伙伴、API,也没有披露许可框架。四条可能的变现路径(模型许可、B2G 主权合同、API 访问、科学 IP 版税)都要求 superlearner 先达到商业可用能力——这是种子轮无法解决的二元不确定性。 $5.1 B 投后估值意味着种子轮价格 / 资本比为 4.6×。下一轮若以持平或下行估值融资,将释放投资论点恶化信号,并触发早期投资人按种子轮价格标记持仓后的减记连锁反应。British Business Bank 参与种子轮,也意味着公共资金暴露在同一下行情景中。 [CR025, CR026, CR027, CR028, CR029, CR030]
7.5 人员、执行与治理风险
Ineffable 的人员风险结构性尖锐。David Silver 是唯一具名创始人、CEO 和主要科学权威。公司没有公开确认的联合创始人、COO、CTO、CFO 或首席安全官。Ashby 招聘页面显示,公司仍在招聘高级 RL 研究员、ML 基础设施工程师和研究科学家——说明成立十二个月后、完成 $1.1 B 融资两个月后,团队仍在搭建。这对一家隐身实验室并不罕见,但也让投资人看不到通常应支撑如此宏大论点的板凳深度证据。 关键人集中度处在前沿 AI 实验室同业中的极端水平。Anthropic 由 Dario Amodei、Daniela Amodei 和另外六名 OpenAI 校友共同创立;OpenAI 拥有深厚管理梯队;即便 Google DeepMind 也有分布式科学领导层。Ineffable 的全部智力资产——Silver 的 RL 研究议程、前沿研究员网络,以及与 NVIDIA 和 Google 的关系——都集中在一个人身上,且没有披露继任计划。 Silver 的研究科学家画像进一步放大执行风险,而非运营者画像。他的论文和公开评论显示出非凡科学深度,但没有证据表明他此前有打造、扩张或商业化一家科技公司的经验。前沿 AI 实验室要同时面对多重执行挑战:招揽人才、采购算力、设计安全项目、维护投资人关系、开展监管沟通,最终还要开发产品。这些挑战与前沿研究在管理上不同,需要另一套能力。 治理风险实质存在。公司是标准私人有限公司结构,没有披露公共利益契约、独立安全委员会、外部顾问委员会,也没有发布负责任 AI 原则。英国政府少数股权制造了治理问责预期,但公司当前披露并未满足这一预期。Newspage 和 Electronics Weekly 引述的批评者称,现有安排是公共资金为公司提供合法性,却没有获得相称治理权。 [CR032, CR033, CR034, CR035, CR036, CR037]
| 角色 / 职能 | 依赖或缺口 | 可能性 | 严重性 | 缓释措施 | 尽调路径 |
|---|---|---|---|---|---|
| CEO 和创始人——David Silver | 唯一具名高管;整个研究论点、人才网络、投资者信心和外部可信度都锚定在单个人身上;未披露继任者或联合创始人 | 中——Silver 看起来投入度高且科学动机强;健康、冲突或竞争性机会仍是不可知风险 | 关键——一旦离开,论点立即失效;投资者信心崩塌;人才管线卡死 | 留任安排(未披露);股权持有形成财务一致性;NVIDIA 和 Google 合作让关系部分机构化 | 要求披露 CEO 留任协议、vesting cliff 和继任计划 |
| CTO / 研究负责人 | 未公开确认 CTO、研究负责人或科学共同负责人;技术项目的科学治理存在单点 | 高——可能已由内部人员担任但未披露 | 高——科学领导没有冗余;Silver 离开后没有技术负责人接棒 | 未知——无披露 | 要求确认 CTO / 研究负责人任命和汇报结构 |
| CFO / 财务 | 未公开确认 CFO 或财务负责人;公司仍在种子阶段运作,未披露财务控制框架 | 中——该阶段常见缺口,但考虑到已融资 $1.1 B,显得不寻常 | 中——财务管理不当风险;burn rate 无法由外部监控 | Unknown | 要求确认 CFO 任命;要求审计师任命和管理账框架 |
| 安全 / 对齐负责人 | 未公开识别安全负责人、对齐研究员或 responsible AI officer;未披露 RSP 或安全框架;公司瞄准超级智能,却没有任何公开安全项目 | 高——有意推迟;公司处于纯研究姿态 | 高——系统性风险;监管敞口;声誉风险;若安全失败迫使研究停摆,则构成论点风险 | 未确认任何措施 | 将任命安全负责人设为投资条件;要求 RSP 或同等文件草案;要求 Seoul Summit 承诺签署时间线 |
人员风险登记表完全基于公开披露。内部组织架构、留任协议和股权 cap table 未公开。实际团队配置可能比公开证据显示的更成熟。
[CR032, CR033, CR034, CR036, CR037]7.6 缓释措施、监测指标与终止标准
每个优先风险簇都有不同的缓释路径、监测信号和明确阈值;一旦越过阈值,就构成投资论点破裂事件,需要投资人重新评估。 监管风险方面,主要缓释措施是主动与 DSIT、AISI 和 ICO 沟通,并在下一轮融资前委托法律分析 NSI Act 持股限制和 EU AI Act GPAI 义务。监测信号是英国 AI 法案提交议会,以及欧盟 GPAI 登记册发布;论点破裂阈值是具有约束力的算力或安全义务生效,而 Ineffable 若不对训练项目做实质重组就无法遵守。 算力依赖方面,缓释措施是与 NVIDIA 和 Google Cloud 谈判合同化配额保证,并至少正式评估第二算力提供商(Lambda Labs、CoreWeave 或 AWS Trainium)的可行性。监测信号是 NVIDIA 公开业绩电话会中关于 Vera Rubin NVL72 爬坡时间表的表述;论点破裂事件是确认延迟超过十二个月,或明确把 NVL72 产能重新优先分配给其他客户。 RL 扩展不确定性方面,缓释措施是定义并发布技术里程碑路线图,包含可验证里程碑(具体基准成绩、能力展示),且可由独立方评估。监测信号是 Ineffable 及其前沿竞争者经同行评审的 RL 泛化结果;论点破裂阈值是 Ineffable 开始训练后三十个月内,任何前沿来源都没有展示开放世界 RL 泛化改进,同时 Ineffable 也拿不出反证。 关键人风险方面,缓释措施是任命至少一名等同联合创始人的科学联合负责人,形成书面继任计划,并与高级研究员签订留任安排。论点破裂事件是 David Silver 因任何原因离开 CEO 职位。 下一轮融资方面,监测信号是前沿 AI 风险投资活动轨迹,以及 Sequoia / Lightspeed 投资组合公司的募资可比样本;论点破裂事件是在种子轮后三十六个月内无法以不低于 $5.1 B 估值完成 Series A,或确认 runway 低于十二个月且没有已承诺跟投。 汇总后的终止标准与缓释表(TR005)概括五个风险簇、触发事件和建议投资人行动。风险热力图(FR001)按可能性和严重性映射所有重大风险。传导图(FR002)展示主要风险如何传导至收入、估值和投资人回报结果。依赖图(FR003)可视化必须主动管理的关键外部依赖。 [CR039, CR040, CR041, CR042]
| 风险 | 监测触发器 / 信号 | 论点破裂阈值 / 否决事件 | 触发后的建议动作 |
|---|---|---|---|
| 算力集中(NVIDIA / Google Cloud) | NVIDIA 公开财报电话会;Vera Rubin NVL72 产能爬坡表述;NVIDIA 客户优先级公告;季度检查 | NVIDIA 确认 Vera Rubin NVL72 延迟超过 12 个月,或明确将分配优先级从 Ineffable 转走;且 Ineffable 60 天内没有替代算力安排 | 升级至董事会;将 90 天内落实第二算力备选设为契约条件;若未解决,下调信念 |
| RL 扩展平台期 | Ineffable 和前沿竞争者的同行评议 RL 泛化结果;RL scaling laws 学术文献;内部里程碑披露(如有);arXiv / NeurIPS / ICML | Ineffable 训练启动后 30 个月内,任何前沿来源都未证明 open-world RL 泛化进展,且 Ineffable 没有披露足以反驳趋势的里程碑 | 将论点下调为投机;要求 6 个月内披露技术里程碑;委托独立 RL 专家评估 |
| 安全 / 对齐治理缺位 | 发布 RSP 或同等文件;宣布安全负责人任命;确认 AISI 接触;签署 Seoul Summit 承诺 | Series A 交割后 12 个月内,没有 RSP、没有安全负责人、也没有 AISI 接触记录;或监管执法行动启动 | 将安全治理包(RSP + 安全领导 + AISI 接触)设为 Series A 交割条件;若执法开始,则构成重大不利变化 |
| 关键人物——David Silver 离开 | LinkedIn / 公开活动;媒体报道;Companies House 董事变更;每月检查董事登记 | Silver 宣布离任 CEO、长期健康缺席,或转投受竞业限制覆盖的竞争对手 | 立即重估论点;10 个工作日内召集特别投资者委员会;在继任计划明确前暂停进一步资本投放 |
| 下一轮融资风险 | 月度估算 runway(算力发票 + 人员成本);Sequoia / Lightspeed Series A 信号;前沿 AI 私募市场估值 | runway 降至 12 个月以下且没有估值 ≥$4 B 的已承诺 Series A 条款清单;或 Sequoia、Lightspeed 公开拒绝领投 Series A | 论点破裂;启动过桥或二级流动性流程;按估算 mark-to-market 重估持仓 |
否决标准是面向投资者的触发器,不是法律义务。所有阈值均基于公开信息设定;一旦数据室开放,并披露 burn rate、商业条款和技术路线图,阈值可能需要调整。监测指标应按季度审查。
[CR039, CR040, CR041, CR042]将 Ineffable Intelligence 的所有重大风险放在“可能性 × 严重性”网格中。 右上象限(高可能性、关键严重性)包含最急需缓释或终止标准监控的风险。 算力集中、关键人物、安全治理缺失和 RL 扩展不确定性都集中在高风险区。
[CR009, CR011, CR013, CR017, CR025, CR032]有向无环图展示主要风险如何在组织内传导,并影响收入轨迹、投资者信心和最终投资论点的完整性。 三条最强的传导路径是:(1)Silver 离职 → 人才流失 → 研究停摆 → 投资者信心崩塌;(2)NVIDIA 延误 → 训练停摆 → 里程碑落空 → 下一轮融资失败;(3)RL 停滞 → 没有能力证明 → 商业化失败 → 投资论点破裂。
[CR009, CR011, CR032, CR039, CR040, CR041]图中列出所有关键外部依赖。Ineffable Intelligence 位于中心;第一圈每个节点都代表一个依赖,一旦失效会直接传导到研究计划。 NVIDIA 和 Google Cloud 是集中度最高的两个单点依赖。David Silver 是唯一的人类依赖节点。 UK Government 提供合法性和监管入口,但直接运营杠杆有限。
[CR009, CR017, CR023, CR024, CR032]7.7 附录
08估值
8.1 投资论点与反论点
Ineffable Intelligence 提供了 2026 年科技版图中风险最高的一类投资命题:一家由顶尖单一创始人领导的强化学习实验室,以 $5.1 billion 种子轮估值获得一线风险资本支持,却没有收入、没有产品,也没有披露商业时间表。投资论点建立在四根支柱上——创始人质量、范式独特性、资本与基础设施结构、主权顺风——每一根都有可信反论点。 基础支柱是 David Silver 的卓越履历。作为 Google DeepMind 中 AlphaGo、AlphaZero、AlphaStar 和 AlphaProof 的主要架构师,Silver 比任何在世研究者都更能证明:强化学习可以从学术理论转化为改变范式的成果。2016 年 3 月,AlphaGo 以 4-1 击败 18 次世界冠军李世乭,吸引 2 亿人观看;AlphaFold 改写结构生物学;AlphaProof 推进了自动数学推理。Ineffable 的共同领投方之一 Lightspeed Venture Partners 明确把这段履历作为投资理由,称 Silver 用「近二十年把强化学习从一个研究想法变成该领域其他人赖以构建的成果」。反论点是集中风险:Silver 就是整个论点。没有具名联合创始人,没有披露继任计划,也没有公开介绍研究领导团队。一旦 Silver 离开,结果概率分布会急剧塌缩。 第二根支柱是范式独特性。以公开信息看,Ineffable 的无数据体验式 RL 路径,是全球唯一达到种子轮规模的同类尝试。OpenAI、Anthropic、Google DeepMind 和 xAI 都依赖人类生成训练数据和 RLHF;Ineffable 的系统则设计为自行生成并评估经验。反论点是竞争速度:Google DeepMind 拥有除 Ineffable 之外最深的 RL 人才池,也是 Silver 的职业大本营,可能在估计 12–36 个月内把资源转向同一范式。MIT Technology Review 的分析已发布证据,提示无数据 RL 在开放式环境中面临根本障碍;同行评审文献也越来越多地指出基于奖励系统的扩展上限。 第三根支柱——由一线领投方与 NVIDIA / Google Cloud 共同投资构成的 $1.1 billion 资本结构——同时给出多年 runway 论据和高质量背书信号。反论点是:没有披露烧钱速度,资本充足性无法验证。按 Epoch AI 分析,前沿 RL 算力成本每年约增长 2–5×;公司的基础设施伙伴关系虽部分缓解 capex,却不能完全抵消这一结构性压力。 第四根支柱——UK Sovereign AI Fund 和 British Business Bank 共同投资——提供政治合法性和政策顺风。英国国务大臣称这是「以风险投资的速度和一个国家的力量支持他们」。独立治理专家明确提出的反论点是,政府约 2% 持股和少数股权地位并未确保 IP 权利、治理控制权,或对发现成果的可执行主张。投资买到的是政治关联,不是结构性主权。 表 TV002 将这四根支柱对应到支持证据和反论点条件,并列出会强化或击穿各项论证的转折信号。 [CV013, CV014, CV015, CV016, CV017, CV018]
| 论点支柱 | 支撑证据 | 反论点 | 会改变判断的证据 |
|---|---|---|---|
| 创始人质量 | David Silver 共同打造 AlphaGo、AlphaZero、AlphaStar、AlphaProof——这是史上最有分量的 RL 作品组合;Lightspeed 将其列为主要投资理由 | 关键人物高度集中;未具名联合创始人;未披露继任计划 | 具名联合创始人或同等 RL 级别 CTO;披露董事会继任计划 |
| 范式独特性 | 截至 2026 年 6 月,没有已知可比实验室以如此资本规模押注无数据体验式 RL;范式差异化独特 | 既有巨头(尤其 DeepMind)可在 12–36 个月内复制;开放式环境中的 RL 扩展限制已有论文证据 | Ineffable 发表经同行评议的论文,证明开放式 RL 可扩展到受限环境之外 |
| 资本充足性 | $1.1B 欧洲最大种子轮;Tier-1 Sequoia + Lightspeed;NVIDIA 和 Google Cloud 作为经济共同投资者 | 前沿 RL 算力每年增长 2–5×;未披露 burn rate;算力胃口可能在里程碑前耗尽资本 | burn rate 确认低于 $100M/yr 且 runway 多年;或宣布上调估值的后续过桥融资 |
| 基础设施锁定 | NVIDIA Grace Blackwell 和 Vera Rubin 联合设计;Google Cloud AI Hypercomputer 为首选提供商;两家公司均持股 | 每项都依赖单一提供商;没有第二算力备选;NVIDIA 分配未在已披露条款之外获得合同保证 | 在主合作伙伴之外,同时宣布多云或本地部署的主权算力安排 |
| 主权 / 政策顺风 | UK Sovereign AI Fund + British Business Bank 共同投资;国务大臣亲自背书;AI 政策采用「促创新」框架 | 政府约 2% 股权不带来 IP 权利、治理控制权或对发现成果的可执行主张;批评者指出,少数股权买到的是政治关联,不是主权 | Ineffable 与英国政府之间签署正式 IP 许可、股权换发现成果,或优先访问协议 |
| 退出 / 商业化路径 | 长期 AGI 可触达市场以 $trillions 计;NVIDIA、Google 或主权买家存在战略收购可选性 | 近期没有退出路径;没有产品路线图;商业验证还在 7–15 年后;未见二级市场交易报道 | 首份已签署商业合同,或与具名交易对手签署 $100M+ 战略 LOI |
反论点基于截至 2026 年 6 月 22 日可获得的公开证据。每个支柱都可单独评估;六项反论条件目前全部成立。尚未披露能够改变判断的证据。
[CV013, CV014, CV016, CV020, CV025, CV026]8.2 融资背景与估值批判
Ineffable Intelligence 的 $5.1 billion 投后估值——由 2026 年 4 月种子轮确立——是有记录以来欧洲种子阶段公司最高估值,约为 $1.1 billion 融资额的 4.6 倍。公司没有收入,无法构建传统收入倍数;没有提交财务账目可供交叉验证;也没有独立分析师发布目标价或折现现金流模型。这个估值完全建立在三件事上:(a)一线投资人的质量信号,(b)美国前沿 AI 实验室估值先例,(c)使命范围中内含的预期。 融资结构把 Ineffable 放进美国前沿 AI 实验室的估值惯例参照组,而不是欧洲初创生态;在欧洲,即便最大型公司通常也需要收入或近期产品可见性才能支撑这种价格。Lightspeed Venture Partners 发布的投资理由称,Silver 独特地有能力取得「重新定义 AI 系统能力边界」的科学突破——这更像研究实验室承销逻辑,而不是产品驱动增长。Sequoia Capital 曾公开分析 2026 年环境下的 AI 资本配置,指出收入前前沿实验室估值中的集中风险。 优先权堆栈和稀释画像都是重大未知项。以 $5.1 billion 投后估值融资 $1.1 billion,隐含投前估值约 $4 billion。种子阶段投资人合计持有约 21.6%,且尚未计入未来轮次。考虑到达到商业证明所需资本强度——几乎必然需要多轮后续融资——种子投资人在整个投资生命周期内面临显著稀释风险。优先权结构(清算优先权、反稀释、按比例认购权)未披露,可能显著影响真实回报结果。 UK Sovereign AI Fund 和 British Business Bank 合计约 $20 million 的持股(低于本轮 2%)无法为公共资金资本损失提供有意义的财务保护。本轮融资时被引用的 AI 治理专家明确警告,公共资金在这一阶段应该「买到的不止一篇新闻稿和少数股权」——Sovereign AI UK 或 British Business Bank 公告中没有披露任何 IP、治理或控制条款,直接印证了这一判断。 Companies House 文件 16865241 确认,截至 2026 年 6 月,Ineffable Intelligence Ltd 尚未提交年度账目,符合其种子阶段状态和 2025 年 11 月较早成立日期。没有任何类型的财务数据可供公开独立验证。以公司年龄看,这种缺失并不罕见,但没有直接数据室访问,标准财务尽调无法开展。图 FV003 展示牛、基准和熊三种情景隐含的估值区间。 [CV001, CV002, CV003, CV004, CV005, CV012]
8.3 可比公司组
标准估值方法——收入倍数、EBITDA 倍数、DCF——不适用于 Ineffable Intelligence,因为公司没有收入、没有披露财务数据,也没有商业时间表。唯一可用的基准框架,是其他前沿 AI 实验室的可比私募轮次。表 TV004 列出主要可比样本,并注明所有重大限制。 OpenAI 的 $157 billion 投后估值(2024 年 10 月)代表前沿 AI 实验室定价上限。不过,OpenAI 当时年化收入约 $4 billion——这与 Ineffable 有关键区别。$157 billion 估值意味着约 39× 收入倍数,按历史科技基准本身也很激进,但至少锚定实际商业牵引力。Ineffable 收入为零,因此 OpenAI 可比样本只能作为市场环境指标,不能作为直接定价参照。 Anthropic(约 $60 billion,2025 年 1 月)和 xAI(约 $80 billion,2025 年 3 月)都已从已部署产品产生收入——分别是 Claude 和 Grok——且都不是在可比的收入前种子阶段融资。Anthropic 的安全优先使命与 Ineffable 的研究取向有结构相似性,但 Anthropic 在大额估值轮之前已部署商业 API 产品。xAI 是一家营利实体,并拥有已部署消费者产品;其公司页面已确认这一点,这与 Ineffable 的纯研究实验室商业姿态根本不同。 Mistral AI 2024 年 6 月约 €6 billion(约 $6.4 billion)的估值,是最相关的欧洲前沿 AI 可比样本。不过,Mistral 当时已部署 Le Chat,并通过公开定价页产生 API 收入,意味着其每美元收入估值低于 Ineffable 隐含倍数。作为拥有公开可见定价结构和产品的欧洲实验室,Mistral 在地域上更接近,但业务阶段不同。 公开信息中没有任何可比样本是一家种子阶段、零收入、估值超过 $5 billion 的纯研究 AI 实验室。Ineffable 的价格位于美国前沿实验室建立的层级,而那些实验室已经部署产品。收入前实验室是否配得上这一层级,完全取决于技术赌注能否兑现、商业化能否跟上——今天的公开证据无法做出这个判断。表 TV004 及配套枚举范围文件记录了这一可比组的限制。 [CV007, CV008, CV009, CV010, CV011, CV043]
| 可比公司 | 估值时阶段 | 轮次 / 募资额 | 投后估值 | 轮次发生时收入 | 与 Ineffable 的相关性 | 关键限制 |
|---|---|---|---|---|---|---|
| OpenAI(2024 年 10 月) | AGI 前阶段;$4B+ ARR;ChatGPT 每周用户 200M+;o1/o3 已部署 | 融资 $6.6B | $157B | 年化约 $4B(估计) | 前沿 AI 实验室;AGI 使命;截至当时美国最大风险投资轮 | 已产生收入;已与 Microsoft 集成;不是收入前种子轮;收入倍数约 39× |
| Anthropic(2025 年 1 月) | 已部署 Claude 3 系列;企业 + API 收入;Amazon 合作 | Amazon 支持的融资 | ~$60B | 未披露;估计 ARR $1–2B | 以安全为核心的前沿 AI 实验室;使命表述相近;商业化前安全研究 | 已部署商业 API 产品;治理结构不同(公益公司) |
| xAI(2025 年 3 月) | 已部署 Grok 消费产品;Colossus 100K GPU 集群;重算力使命 | 融资 $6B | ~$80B | 未披露;Grok 订阅收入 | 前沿 AI 实验室;重算力投入;使命接近 AGI | 已部署消费产品;Elon Musk 创始人效应抬高估值信号 |
| Mistral AI(2024 年 6 月) | 已部署 Le Chat;公布 API 定价;欧洲实验室 | 融资 €600M | ~€6B (~$6.4B) | 未披露;API + 企业收入已产生 | 欧洲前沿 AI 实验室;地理上最接近 Ineffable 的可比对象 | 有商业产品和收入;规模更小;技术范式不同(LLM) |
| Ineffable Intelligence(2026 年 4 月) | 纯研究;从隐身到种子轮;零收入;零产品 | $1.1B 种子轮 | $5.1B | 零 | 标的公司;可比基准测试的参照点 | 无收入;无产品;无商业时间线;价格完全押注信号和叙事 |
| 前沿 AI 层级平均(2024–2026) | 已部署产品阶段各不相同 | 多种轮次 | 观察区间 $5B–$80B | 各不相同;所有具名可比公司都已产生收入 | 投资者风险偏好的市场环境背景 | 异质性高;区间内没有直接等同的收入前种子轮 |
估值来自公开新闻报道;若有更准确的一手数据,全部可能修订。未公开披露的收入数字均为估计。可比集合明确不完美:所有具名可比公司在所引估值轮之前都已产生收入。
[CV001, CV007, CV008, CV009, CV010, CV011]8.4 牛、基准与熊三种情景
Ineffable 的情景规划需要基于里程碑承销,而不是收入预测,因为公开证据看不到收入轨迹。以下三种情景都锚定可观察的技术和商业触发器,而非投机性收入条目。 牛市情景假设 David Silver 在 2027–2028 年发布可验证 RL 里程碑,证明开放式环境扩展能力;Ineffable 在 2028–2029 年拿下首个主权或企业商业合同;Series B 估值达到种子轮的 2× 或更高。在这种情景下,价值创造路径从约 2029 年开始的许可和 API 收入,延伸到 2032 年的战略收购或 IPO 候选资格。隐含企业价值区间为 $30–75 billion。这个情景要求连续取得独立胜利——先技术里程碑,再商业证明,再融资里程碑——每一步都以前一步为条件。概率为低到中。主要敏感性驱动因素是已发布 RL 里程碑的规模和可信度。 基准情景假设到 2028 年研究进展可见且可信,但商业部署推迟到 2030–2031 年。公司以温和估值上调完成一到两轮过桥融资,初始商业牵引力来自小众主权或科学许可细分市场。2032 年退出时隐含企业价值区间为 $8–20 billion。这个情景最符合公开证据:技术进展先于即时商业化,正好匹配每一家前沿 AI 研究实验室已观察到的发展模式。 熊市情景假设到 2027–2028 年出现 RL 扩展平台期证据,且 Ineffable 没有反证;Silver 离开或改变公司方向;资本在商业证明前耗尽。在这种情景下,退出选项收缩为清仓出售、以低于种子轮估值的人才收购,或只剩残余 IP 价值。隐含企业价值区间为 $0.5–4 billion。如果 Ineffable 在种子轮完成后 24–30 个月内没有产出公开可验证里程碑,这一情景的概率会上升。 Stanford HAI AI Index(2025)记录,2024 年全球 AI 私人投资超过 $100 billion,其中前沿模型投资占主导——这确认了 Ineffable 种子轮定价所处的市场环境。表 TV003 将三种情景对应到假设、价值路径和概率信号。图 FV002 分离关键敏感性驱动因素。图 FV003 绘制各情景下的企业价值区间。 [CV021, CV022, CV023, CV024, CV025, CV035]
| 情景 | 关键假设 | 价值创造路径 | 隐含企业价值(2032) | 概率信号 |
|---|---|---|---|---|
| 牛市 | 2027–28 年发布技术里程碑(开放式 RL 扩展);2028–29 年拿下首个主权 / 企业合同;Series B ≥2× 种子轮($10.2B+);NVIDIA 或 Google 表达战略收购兴趣 | 约从 2029 年起产生许可 + API 收入;在体验式 RL 赛道取得先发优势;到 2032 年具备战略收购或 IPO 候选资格 | $30–75B | 中低;至少需要三个连续、独立的里程碑全部跑通 |
| 基准 | 到 2028 年研究进展可信;商业化推迟到 2030–31 年;以温和估值上调完成 1–2 轮过桥融资;在小众主权 / 科学许可中获得牵引 | 向主权和科学场景做小众许可;企业分发有限;以温和溢价实现战略收购或结构化退出 | $8–20B | 中;最符合已观察到的前沿 AI 实验室开发时间线 |
| 熊市 | 2027–28 年确认 RL 扩展平台期,且 Ineffable 拿不出反证;Silver 离职或调整使命;商业验证前资金耗尽 | 减记;贱卖;以低于种子轮估值被 acqui-hire;只剩残余 IP 价值 | $0.5–4B | 若种子轮交割后 24–30 个月内没有可验证里程碑,则为中高 |
所有企业价值都是说明性情景估算,不是财务预测。Ineffable 没有公开收入或成本数据;估算基于可比前沿 AI 实验室轨迹和按里程碑承销的原则。各情景并非互斥;里程碑确认时间是主要区分变量。
[CV021, CV022, CV023, CV024]以 $5.1B 基准为参照,示意 Ineffable 企业价值对八个最关键正负变量的方向性敏感度,单位为估算的 $B 变化。 上行驱动主要来自技术里程碑兑现;下行驱动主要来自 Silver 离职和 RL 停滞。该图不是财务预测——所有数值都是分析师基于情景分析和可比实验室案例作出的估算。
数值仅为方向性分析师估算。Ineffable 没有公开收入、成本或融资数据。敏感度幅度依据 TV003 中的情景价值区间和可比实验室估值波动校准。 柱状条展示各驱动因素对基准估值假设的加性影响;驱动因素之间的相互作用未建模。
[CV021, CV022, CV024, CV025, CV026, CV027]牛市、基准和熊市情景下,2032 年退出时低 / 基准 / 高企业价值估算,单位为十亿美元。 图中展示潜在结果的完整光谱;区间很宽,反映缺少公开收入、成本或融资数据。熊市情景与种子轮估值重叠($5.1B 基准情景下界), 说明种子投资者在稀释和优先权之后,现实中可能获得零回报甚至负回报。
数值为情景估算,来自基于里程碑的承销和可比 Frontier AI 实验室案例。不是财务预测。 所有数字均为稀释前,且未计入优先权堆叠;条款未披露,可能实质影响投资者层面的回报。
[CV021, CV022, CV023, CV024]8.5 建议、信心与风险评级
对 Ineffable Intelligence 的建议是有条件的监测性兴趣——持观察仓位,而不是承诺出手。这不是温和对冲。证据清楚显示,正面因素极其突出(创始人质量、范式、资本、基础设施),但负面因素在现阶段对高确信仓位构成结构性否决:零收入、零客户、无烧钱速度、无商业路线图、无安全治理,且公共共同投资人没有 IP 锁定。下一轮若要领投,投资人必须押注至少两个相互独立的不确定事件按顺序兑现:技术前提被证明成立,商业赌注在可融资时间表内跟上。 对任何确定性估值判断的信心都低。公开信息中没有提交的财务报表、没有独立审计账目、没有披露资金用途,也没有商业证明。投资人承销的叙事风险不亚于业务风险。Sequoia 的「Act Two」分析明确警告,近期没有产品—市场匹配的 AI 实验室,会在算力成本膨胀和投资人耐心消退时面临融资风险——这一动态直接适用于 Ineffable 的轨迹。 整体风险评级为极高。主要风险维度——David Silver 的关键人集中、范式级 RL 扩展不确定性、对 NVIDIA 和 Google Cloud 的单一来源算力依赖、安全或对齐治理框架缺位,以及完全财务不透明——现阶段都无法单独管理。它们会相互叠加。任一风险触发,投资论点都会迅速恶化。 估值立场偏紧,但符合先例。$5.1 billion 种子轮价格在欧洲风险投资史上前所未有,并按美国前沿 AI 实验室惯例定价;而这些惯例通常要求真实收入和已部署产品。没有收入时,价格完全依赖叙事和背书信号价值。这个信号确实非同寻常——但叙事风险真实存在,市场对收入前 AI 实验室的重估可能迅速且严厉。 即便在牛市情景下,合适投资周期也是 7–15 年。采用标准基金生命周期的投资人与 Ineffable 的商业化周期存在结构错配。长期资本工具——延续基金、主权财富结构、家族办公室——比传统十年期 VC 结构更适合这个时间跨度。目前没有可见或披露的近期退出路径(IPO、M&A、二级交易)。表 TV001 以结构化形式总结建议。 图 FV001 追踪建议逻辑链。图 FV004 给出可供 IC 使用的关键投资维度评分。 [CV013, CV022, CV039]
| 维度 | 数值 | 依据 | 行动含义 |
|---|---|---|---|
| 建议 | 有条件持续关注(观察) | 收入前;零客户;大规模 RL 尚未验证;创始人和资本信号异常强 | 不领投;承诺前先设定 Series A 里程碑门槛 |
| 置信度 | 低 | 没有收入、没有已提交账目、没有 burn rate、没有商业路线图;所有模型都依赖技术假设 | 任何投委会材料都应承认模型不确定性很高 |
| 风险评级 | 很高 | 关键人物;RL 扩展不确定性;算力集中;治理缺位;财务不透明 | 五项独立足以触发否决标准的风险同时存在 |
| 估值立场 | 偏紧 / 符合先例 | $5.1B 是欧洲史上最大种子轮;处于美国前沿 AI 实验室层级;无法用收入锚定 | 跟踪 Series A 定价,作为验证或降级的首个市场信号 |
| 投资期限 | 7–15 年(耐心资本) | 研究 → 能力 → 商业证明周期;未披露产品时间线 | 仅适合长久期资本结构(continuation funds、主权、family office) |
建议和置信度仅基于截至 2026 年 6 月 22 日的公开证据。没有数据室访问;收到私有财务和技术材料后,所有判断都可能修订。
[CV031, CV032, CV033, CV034]从证据输入(市场规模、创始人质量、资本、范式风险、零商业证明、治理缺口、估值语境)一路推导到最终建议:有条件的跟踪关注(观察仓位)。 图中展示正负因素如何合在一起,形成一个信心较低、风险极高的观察判断。
[CV013, CV016, CV039]面向 IC 的评分,对 Ineffable Intelligence 的八个投资维度逐项打分,每项按 1–10 分计,依据截至 2026 年 6 月 22 日的公开证据。 市场规模和创始人质量得分很高,但商业证明和治理得分极低,最终只能形成观察级信念。所有评分均不基于私人或非公开信息。
[CV013, CV022, CV039]8.6 论点破裂触发器与最终尽调要求
六类论点破裂事件会在任何计划内监测复盘节点之前,把建议从观察转为退出审查或减记。最紧迫的是 David Silver 在没有披露且可信继任者的情况下离开 CEO 职位。整个论点就是 Silver:若他离开——且没有具名、声望相当的继任者——所有情景概率都必须立即重估。第二类是同行评审论文或可信技术发布确认 RL 扩展进入平台期,而 Ineffable 没有反证。这会击穿范式前提,把熊市情景从低概率推成基准情景。第三类是 NVIDIA Vera Rubin 硬件配额实质减少或撤回,或 Google Cloud 合作终止——任一事件都会让研究时间表延后数年。 第四类论点破裂是 Series A 估值低于或不高于 $5.1 billion 种子轮投后估值。持平或下行轮说明市场已下调能力主张评级,里程碑证据不足以支撑估值上调,且跟融资风险升高。第五类是安全治理失败——监管调查、对齐事故,或归因于 Ineffable 系统的对抗测试失败——它会同时固化监管、声誉和共同投资人风险。第六类是英国政府澄清,其主权投资并不相对外资投资人享有 IP 权利、治理权或发现成果优先访问权。 监测这些触发器需要:持续观察 David Silver 的公开露面和 Ineffable 公告;按季度复盘 ArXiv 与同行评审场所的 RL 扩展文献;监测 NVIDIA 和 Google Cloud 合作新闻;追踪任何监管问询或安全事故归因。表 TV005 定义每个触发器的阈值、传导路径和行动含义。 最终数据室要求,是把观察转为未来轮次任何仓位的有条件承销所需最低证据。它们覆盖五个阻断性证据缺口:$1.1 billion 种子资金对应的烧钱速度和 runway;经董事会批准的具体技术里程碑定义;完整股权结构和优先权结构,包括清算优先权与反稀释;内部安全治理和对齐框架文件;以及英国主权共同投资人持有的 IP 权利精确范围。这些问题没有一个能靠公开来源解决。表 TV006 按负责人、路径和优先级列出各项要求。没有数据室访问,现阶段对 Ineffable 建立任何高确信仓位都只能靠信号。 [CV025, CV026, CV027, CV028, CV037, CV040]
| 触发项 | 阈值 / 事件 | 对论点的传导 | 监测路径 | 紧迫性 |
|---|---|---|---|---|
| David Silver 离职 | Silver 退出 CEO 职位,且没有具名、可信、声望相当的继任者 | 整个 RL 论点坍塌;团队凝聚力承压;范式独特性的主张被削弱 | 监测 Ineffable 新闻稿、董事会公告、LinkedIn、Companies House 董事备案 | 立即 — 任何时间 |
| RL 扩展平台期得到确认 | 经同行评议或可信技术出版物确认,无数据 RL 无法在开放式环境中扩展;Ineffable 在 6 个月内没有反证 | 范式押注失败;熊市情景变成基准情景;所有情景估值被压缩 | 按季度监测 ArXiv RL 扩展文献;追踪 Ineffable 论文 | 12–30 个月 |
| 算力访问中断 | NVIDIA Vera Rubin 配额被实质削减,或 Google Cloud 合作终止;未宣布替代方案 | 核心研究基础设施不可用;时间线延后数年;烧钱速度加快 | 监测 NVIDIA 和 Google Cloud 合作新闻;关注算力重新分配信号 | 12–24 个月 |
| Down-round 或平轮 Series A | 下一轮融资定价等于或低于 $5.1B 投后估值 | 市场下调能力主张定价;后续融资风险升高;投资者情绪转向 | 追踪所有融资公告;监测二级市场数据 | 18–30 个月 |
| 安全治理失败 | 公开安全事件、英国 AI 监管调查,或对抗测试失败被归因于 Ineffable 系统 | 监管、声誉和投资者风险同时兑现;主权共同投资者承受政治压力 | 监测 UK AISI 和 ICO 通讯;追踪媒体中归因于 Ineffable 的事件 | 任何时间 |
终止触发项基于可公开监测的事件。内部信息(董事会纪要、安全报告、算力配额协议)会提供更早预警,但无法获取。紧迫性指触发项可从公开证据中得到判断的时间窗口,不是发生概率。
[CV025, CV026, CV027, CV028]| 主题 | 缺失证据 | 重要性 | 负责人 / 尽调路径 | 优先级 |
|---|---|---|---|---|
| 财务烧钱速度和 runway | 月度烧钱速度;按类别拆分的资金用途;基于已融 $1.1B 的 runway 估算 | 没有烧钱数据,就无法判断资本是否足以支撑到可展示里程碑,也无法判断近期是否存在融资悬崖 | CFO;data room;董事会批准预算;索取 24 个月现金预测 | 关键 |
| 技术里程碑定义 | 对「superlearner」研究里程碑给出具体、可衡量、有时限、经董事会批准的 OKR 定义 | 没有里程碑,任何情景都没有承销基础;技术押注无法追踪 | CEO;技术团队;董事会纪要;投资者更新材料 | 关键 |
| 股权结构和优先权结构 | 完整股权表;清算优先权堆栈深度;反稀释条款;各投资者的 pro-rata 权利 | 种子轮投资者的优先权堆栈直接决定基准和熊市情景下的实际回报;深度未知 | 法律顾问;term sheet;股权表管理工具(如 Carta) | 关键 |
| 安全与对齐框架 | 内部安全治理文件;red-teaming 计划;任何负责任扩展政策或等效文件;AI Safety Institute 接触状态 | 没有安全框架,监管暴露(EU AI Act GPAI)、声誉风险和责任都无法量化 | CEO;CISO;外部安全审查;UK AI Safety Institute 接触文件 | 关键 |
| IP 权利和主权治理 | UK Sovereign AI Fund 和 British Business Bank 持有 IP 权利的精确合同范围;任何排他性、优先访问或版税条款 | 主权 AI 顺风论点需要实质内容;若没有正式权利,政府持仓只是声誉背书,且可能逆转 | 律师;政府资助协议;Sovereign AI UK term sheet;BBB 投资协议 | 高 |
五个主题都是阻断性证据缺口:要对任何未来轮次仓位形成有信心的承销判断,必须逐一解决。没有一项能从公开来源补齐。全部需要直接接触管理层或访问 data room。优先级含义为:关键 = 缺少则无法推进;高 = 显著影响情景概率权重。
[CV040]8.7 附录
免责声明
本评估仅基于公开证据;没有管理层访谈、数据室、客户访谈或非公开财务资料。
证据索引
| 编号 | 陈述 | 可信度 | 来源 |
|---|---|---|---|
| CO001 | INEFFABLE INTELLIGENCE LTD was incorporated in England and Wales on 19 November 2025 as a private limited company with company number 16865241. | 高 | SO002, SO025 |
| CO002 | The registered office of INEFFABLE INTELLIGENCE LTD is at 3rd Floor, 1 Ashley Road, Altrincham, Cheshire, United Kingdom, WA14 2DT. | 高 | SO002, SO003 |
| CO003 | Ineffable Intelligence's operating headquarters is London, as consistently cited in investor announcements, government press releases, and independent news coverage. | 高 | SO004, SO010, SO016 |
| CO004 | The official website of Ineffable Intelligence is https://www.ineffable.ai/. | 高 | SO001, SO010 |
| CO005 | The company's SIC code is 74909 (Other professional, scientific and technical activities not elsewhere classified), consistent with a pre-revenue research and development entity. | 高 | SO002, SO025 |
| CO006 | Ineffable Intelligence's mission is to "make first contact with superintelligence" by creating a superlearner that discovers all knowledge from its own experience without relying on human data. | 高 | SO001, SO004 |
| CO007 | The company was launched publicly and exited stealth on 27 April 2026, the same day as the seed round announcement. | 高 | SO004, SO005 |
| CO008 | Ineffable's technical approach is reinforcement learning — agents that learn through experience and trial-and-error rather than training on human-generated data — which Ineffable describes as placing fundamentally different demands on compute infrastructure than LLM pretraining. | 高 | SO001, SO013, SO016 |
| CO009 | David Silver is CEO and Founder of Ineffable Intelligence. | 高 | SO016, SO004, SO012 |
| CO010 | David Silver is a Professor of Computer Science at University College London (UCL). | 高 | SO012, SO004, SO010 |
| CO011 | David Silver was formerly Head (Vice President) of Reinforcement Learning at Google DeepMind, where he spent approximately two decades. | 高 | SO004, SO011, SO012 |
| CO012 | David Silver was a principal architect or lead contributor on AlphaGo, AlphaZero, AlphaStar, AlphaFold, and AlphaProof at Google DeepMind. | 高 | SO011, SO012, SO010 |
| CO013 | David Silver was appointed as director of INEFFABLE INTELLIGENCE LTD on 16 January 2026. | 高 | SO003, SO025 |
| CO014 | Alfred Lin was appointed as director of INEFFABLE INTELLIGENCE LTD on 17 April 2026; his registered address is 2800 Sand Hill Road, Menlo Park, California (Sequoia Capital headquarters). | 高 | SO003, SO025 |
| CO015 | Ravi Mhatre was appointed as director of INEFFABLE INTELLIGENCE LTD on 17 April 2026; his registered address is 2200 Sand Hill Road, Menlo Park, California (Lightspeed Venture Partners address). | 高 | SO003, SO025 |
| CO016 | George Samuel Rose, a Canadian national resident in Canada, is listed as a director in Companies House records for INEFFABLE INTELLIGENCE LTD; his formal role or title has not been publicly disclosed. | 高 | SO003, SO026 |
| CO017 | OAKWOOD CORPORATE SECRETARY LIMITED (company number 7038430) was appointed as company secretary of INEFFABLE INTELLIGENCE LTD on 19 November 2025, the date of incorporation. | 高 | SO003, SO002 |
| CO018 | Sesamers (citing tech.eu) reports that Wojciech Czarnecki, Lasse Espeholt, and Junhyuk Oh — three additional ex-DeepMind researchers — are members of the Ineffable founding team; this is secondary-source reported and has not been confirmed in official company or Companies House filings. | 中 | SO017, SO027 |
| CO019 | Ineffable Intelligence announced a $1.1 billion seed round on 27 April 2026. | 高 | SO004, SO005, SO014 |
| CO020 | The seed round has a $5.1 billion post-money valuation. | 高 | SO004, SO014, SO006 |
| CO021 | The seed round was co-led by Sequoia Capital and Lightspeed Venture Partners. | 高 | SO004, SO005, SO014 |
| CO022 | Named participating investors in the seed round include NVIDIA, Google, Index Ventures, DST Global, EQT Ventures, Flying Fish Ventures, Evantic Capital, and BOND Capital. | 高 | SO011, SO004, SO007 |
| CO023 | The UK Wellcome Trust is reported as a participant in the seed round by EU-Startups and Hotminute. | 中 | SO007, SO022 |
| CO024 | The British Business Bank invested $20 million (£14.8 million) in the seed round, as confirmed in its own press release. | 高 | SO011, SO018 |
| CO025 | The UK Sovereign AI Fund co-invested in the seed round alongside the British Business Bank; the exact amount is "commercially sensitive" and has not been publicly disclosed; the government states typical investments are in the £1–10 million range. | 高 | SO010, SO019 |
| CO026 | The seed round is described as the largest European seed round in history by multiple independent sources including the company, its legal adviser (Cooley), the British Business Bank, and tech media. | 高 | SO014, SO011, SO006 |
| CO027 | Cooley LLP (London partner Eric Davison) advised Ineffable Intelligence on the seed financing. | 中 | SO014 |
| CO028 | David Silver stated that "any money that I make from Ineffable will go to high-impact charities that save as many lives as possible." | 高 | SO005, SO022 |
| CO029 | Sesamers reports that Silver's Founders Pledge commitment covers 100% of any personal proceeds and is described by Founders Pledge as the largest pledge in the organisation's history; this is secondary-source reported. | 中 | SO017, SO022 |
| CO030 | UK Science and Technology Secretary Liz Kendall endorsed the Ineffable investment, stating it would support a company "at the very frontier of AI, with the potential to transform entire sectors." | 高 | SO010, SO004 |
| CO031 | AI Minister Kanishka Narayan issued a formal statement endorsing the investment, calling Silver "one of the world's foremost AI leaders" and pledging full support of the British state. | 高 | SO010, SO012 |
| CO032 | Josephine Kant, Head of Ventures at the Sovereign AI Unit, stated that "very few founders in the world could credibly set out to build a superlearner" and endorsed Silver's two-decade RL track record. | 高 | SO007, SO010 |
| CO033 | NVIDIA and Ineffable Intelligence announced an engineering-level collaboration on RL infrastructure on 13 May 2026, with engineers from both companies co-designing the training pipeline. | 高 | SO013, SO023 |
| CO034 | The NVIDIA-Ineffable RL-infrastructure work is starting on NVIDIA Grace Blackwell and will be among the first to explore the upcoming NVIDIA Vera Rubin platform. | 高 | SO013, SO023 |
| CO035 | Jensen Huang, NVIDIA CEO, stated: "We are thrilled to partner with Ineffable Intelligence to codesign the infrastructure for large-scale reinforcement learning as they push the frontier of AI and pioneer a new generation of intelligent systems." | 高 | SO013, SO023 |
| CO036 | Google Cloud and Ineffable Intelligence announced a strategic partnership on 16 June 2026, with Google Cloud named as Ineffable's preferred cloud provider. | 高 | SO016, SO013 |
| CO037 | Under the Google Cloud agreement, Ineffable will deploy one of the largest clusters of A5X powered by the NVIDIA Vera Rubin NVL72 on Google Cloud, using the AI Hypercomputer architecture. | 高 | SO016, SO023 |
| CO038 | Google Cloud CEO Thomas Kurian stated: "We are honored that the Ineffable Intelligence team has chosen Google Cloud to power its mission." | 高 | SO016, SO004 |
| CO039 | The Google Cloud partnership was announced at Google Cloud Summit London '26 on 16 June 2026. | 高 | SO016, SO004 |
| CO040 | Newspage.news published commentary from AI consultants who questioned whether the UK government's small equity stake in Ineffable buys meaningful sovereignty or governance rights, warning of a potential "governance gap." | 中 | SO019, SO020 |
| CO041 | KYC Digital CTO Katrina Young warned that UK public capital should come with enforceable conditions — UK anchoring of capability, independent safety evaluation, transparency on outputs, and defined rights over downstream use — and that sovereignty is not achieved through presence on a cap table. | 中 | SO019, SO020 |
| CO042 | Electronics Weekly reported that as part of its investment, the UK government receives a London address, a small equity stake, and first refusal on the next round, but no structural guarantee that Ineffable's discoveries, IP, or commercial value remain in the UK. | 中 | SO020, SO019 |
| CO043 | Ineffable Intelligence has no disclosed revenue, ARR, gross margin, customer count, or headcount as of the run date; the company is operating as a pre-revenue research entity. | 高 | SO001, SO004 |
| CO044 | David Silver published a personal founding rationale note on the Ineffable blog dated 15 January 2026, describing the mission as his "life's work" and the decision to open a window for ambitious RL research "without bending to the demands of incremental products and near-term profits." | 高 | SO001, SO005 |
| CO045 | TechCrunch noted that the Ineffable seed round fits the pattern of so-called "coconut rounds" — large first-money-in rounds for months-old ventures founded by star researchers — alongside AMI Labs ($1.03B at $3.5B pre-money) and Recursive Superintelligence (~$500M–$1B reported). | 高 | SO005, SO004 |
| CO046 | Alfred Lin's address (2800 Sand Hill Road, Menlo Park, CA) is Sequoia Capital's headquarters address, confirming his role as the Sequoia partner representing the co-lead investor on the board. | 高 | SO003, SO004 |
| CO047 | Ravi Mhatre's address (2200 Sand Hill Road, Menlo Park, CA) is consistent with Lightspeed Venture Partners' Sand Hill Road office, confirming his role as the Lightspeed partner representing the co-lead investor on the board. | 中 | SO003, SO006 |
| CO048 | The Ineffable+NVIDIA blog post published on the official Ineffable site on 13 May 2026 is co-authored by Ineffable and NVIDIA. | 高 | SO023, SO013 |
| CO049 | UCL confirmed that Ineffable Intelligence is an independent company and is not a UCL spinout; it was founded by an individual with a current academic affiliation at UCL. | 高 | SO012, SO004 |
| CO050 | The UK Wellcome Trust is reported by EU-Startups and Hotminute as a participant in the seed round; this is not confirmed in the BBB, UK government, or Cooley press releases. | 中 | SO007, SO022 |
| CO051 | EU-Startups reports the seed round in Euro terms as €937 million at a €4.3 billion post-money valuation, implying a EUR/USD rate of approximately 1.18 at the announcement date. | 中 | SO007 |
| CO052 | David Silver described his reason for selecting Google Cloud over simple GPU-renting alternatives as requiring a resilient and scalable environment — specifically a systems-level integration of hardware, software, and networking rather than raw compute alone. | 高 | SO016, SO023 |
| CM001 | Ineffable Intelligence does not operate in a single defined market; its addressable universe spans frontier AI training infrastructure, AI-for-science applications, and sovereign/strategic AI capacity — three overlapping but distinct spend categories. | 高 | SM012, SM013 |
| CM002 | The frontier AI training infrastructure market is distinct from the broad AI applications market; it encompasses the hardware, software, and orchestration used to train the world's most capable models and should not be conflated with AI software subscriptions or consumer AI products. | 高 | SM006, SM007 |
| CM003 | Status-quo substitutes for a superlearner system include large language models trained on human-curated data (the dominant supervised-learning paradigm), human scientific experts and domain-specific simulation software, and the internal RL research programs of hyperscalers and frontier labs such as Google DeepMind, OpenAI, and Anthropic. | 中 | SM013, SM011 |
| CM004 | AI-for-science applications — including drug discovery, protein folding, materials science, and genomics — represent a key adjacency to Ineffable's superlearner thesis, with demonstrated commercial demand from pharmaceutical companies and national labs. | 中 | SM009, SM013 |
| CM005 | General AI applications such as enterprise chatbots, NLP-as-a-service, and recommendation systems are excluded from Ineffable's addressable market; these are downstream use cases and do not involve procurement of frontier RL training infrastructure. | 高 | SM013, SM014 |
| CM006 | The broad analyst market figures for "global AI market" — ranging from $200 billion to over $1 trillion depending on methodology — conflate infrastructure, applications, and services, and cannot be adopted as Ineffable's TAM without explicit boundary logic. | 高 | SM003, SM004 |
| CM007 | NVIDIA reported data-centre segment revenue of $39.1 billion in Q1 FY2026 (quarter ending 27 April 2025), up 73% year-on-year and 10% quarter-on-quarter, representing the dominant proxy for global AI compute infrastructure spend. | 高 | SM006, SM027 |
| CM008 | NVIDIA provided Q2 FY2026 revenue guidance of approximately $45 billion (total), despite an ~$8 billion headwind from US export controls on H20 products for the Chinese market. | 高 | SM006, SM027 |
| CM009 | Goldman Sachs Research forecast that global AI-related investment would approach $200 billion by 2025, with the US positioned as the market leader and earlier adoption by larger firms in information and professional/scientific services. | 高 | SM003, SM005 |
| CM010 | Goldman Sachs Research explicitly notes that AI productivity impact is likely delayed to the second half of the current decade, raising concerns about the pace of discretionary AI spend beyond hyperscalers and early adopters. | 高 | SM003, SM005 |
| CM011 | Epoch AI's research concludes that training compute for frontier AI models has grown approximately 4–5x per year between 2010 and 2024, and that this trend is consistent across OpenAI, Google DeepMind, and Meta AI's top models. | 高 | SM007, SM006 |
| CM012 | Epoch AI identifies AlphaGo Master and AlphaGo Zero as compute outliers: they are RL game-playing systems that consumed far more compute than typical deep learning models of their era, and when included in trend analysis they "single-handedly warp the trend of compute in frontier models." | 高 | SM007, SM008 |
| CM013 | The Kimi k1.5 paper (January 2025, arXiv) reports that RL-trained LLMs can match OpenAI's o1 on multiple reasoning benchmarks — including 77.5 on AIME, 96.2 on MATH 500, and 94th-percentile on Codeforces — without relying on Monte Carlo tree search, value functions, or process reward models. | 中 | SM008 |
| CM014 | The Kimi k1.5 paper explicitly states that "prior published work has not produced competitive results" in RL scaling for LLMs, framing their results as a new demonstration rather than confirmation of an established trend. | 中 | SM008 |
| CM015 | AlphaFold (DeepMind, published in Nature 2021) predicted protein structures with atomic accuracy in CASP14, with a median backbone accuracy of 0.96 Å r.m.s.d.95 — competitive with experimental structures — demonstrating that deep learning can solve transformative scientific problems. | 高 | SM009, SM011 |
| CM016 | AlphaFold was developed at Google DeepMind, where David Silver (Ineffable's CEO and Founder) served as VP of Reinforcement Learning; AlphaFold's success is cited by multiple sources as validation of RL-adjacent AI applied to scientific discovery. | 高 | SM009, SM013 |
| CM017 | Hyperscalers (Google, Microsoft, Amazon, Meta) represent the primary near-term buyer tier for frontier RL training infrastructure, as they control the world's largest AI training clusters and have the budget and strategic incentive to acquire advanced RL capabilities. | 高 | SM006, SM024, SM011 |
| CM018 | Google Cloud has been named Ineffable's preferred cloud provider and is deploying one of the largest clusters of A5X powered by the NVIDIA Vera Rubin NVL72, confirming that a hyperscaler is already an activated infrastructure partner and prospective payer. | 高 | SM024, SM011 |
| CM019 | The British Business Bank confirmed a $20 million investment in Ineffable Intelligence, and the UK Sovereign AI Fund also invested an undisclosed amount, establishing UK sovereign programs as confirmed payers in the current round — not hypothetical buyers. | 高 | SM016, SM017 |
| CM020 | The UK AI Opportunities Action Plan (January 2025), led by Matt Clifford CBE, contains 50 recommendations to grow the UK AI sector including frontier AI infrastructure, confirming that UK government policy is structurally aligned with supporting frontier AI development. | 高 | SM010, SM017 |
| CM021 | NVIDIA's announcement at ISC 2026 of 35 new AI supercomputers across Europe, including JUPITER (Europe's first exascale supercomputer at Forschungszentrum Jülich, Germany), confirms that sovereign compute infrastructure investment is materialising at scale across the continent. | 高 | SM027, SM010 |
| CM022 | Scientific R&D organisations — pharmaceutical companies, genomics institutes, and national laboratories — represent a secondary buyer tier for AI-for-science applications, with demonstrated demand validated by AlphaFold adoption and broader drug-discovery AI programmes. | 中 | SM009, SM013 |
| CM023 | Statista's AI Market Outlook projects the global AI market to reach multi-hundred-billion dollar scale through 2030, driven by healthcare AI, chatbot/virtual assistant adoption, AI chips and edge computing, and IoT integration; this figure is not specific to frontier RL training and should be used for directional context only. | 中 | SM004 |
| CM024 | The adoption trigger for scientific R&D organisations as buyers of superlearner outputs would require a published RL demonstration of autonomous scientific discovery analogous to AlphaFold — a milestone Ineffable has not yet achieved as of the run date. | 中 | SM009, SM012 |
| CM025 | Enterprise R&D departments represent a speculative long-term buyer tier that requires both a commercial product and a validated track record before procurement can begin; Goldman Sachs notes that enterprise AI adoption by large non-tech firms is likely delayed to the second half of the decade. | 中 | SM003, SM014 |
| CM026 | Jensen Huang, NVIDIA CEO, stated in the Q1 FY2026 earnings release that AI inference token generation "surged tenfold in just one year" and that "Countries around the world are recognizing AI as essential infrastructure — just like electricity and the internet," confirming strong structural demand for AI compute infrastructure. | 高 | SM006, SM027 |
| CM027 | UK Science and Technology Secretary Liz Kendall and AI Minister Kanishka Narayan both issued formal endorsements of Ineffable's raise, framing it as evidence that the UK can be an "AI maker, not taker," confirming the sovereign buyer narrative is actively supported by government policy. | 高 | SM017, SM016 |
| CM028 | AlphaFold's adoption across global pharmaceutical research and the broader AI-for-science movement confirms there is demonstrated willingness-to-pay from scientific R&D organisations for AI systems that deliver transformative research productivity gains. | 中 | SM009, SM018 |
| CM029 | The EU AI Act's GPAI (general-purpose AI model) obligations became effective on 2 August 2025, requiring providers of high-capability GPAI models to conduct safety assessments, disclose training data, and mitigate systemic risks — obligations that would apply to frontier RL systems operating in or toward the EU market. | 高 | SM002, SM025 |
| CM030 | The EU AI Act extended transition periods to 2027–2028 for certain high-risk AI systems embedded in regulated products, but GPAI obligations are already active (August 2025), creating near-term regulatory pressure specifically for frontier AI model providers including potential future Ineffable deployments. | 高 | SM002, SM025 |
| CM031 | US export controls on NVIDIA's H20 products for the Chinese market caused NVIDIA to incur a $4.5 billion charge and lose approximately $2.5 billion in unable-to-ship revenue in Q1 FY2026 alone, demonstrating that geopolitical supply shocks can materially disrupt AI infrastructure procurement plans. | 高 | SM006, SM027 |
| CM032 | Epoch AI's analysis shows that RL training runs, specifically AlphaGo Master and AlphaGo Zero, are compute outliers relative to the frontier model trend — implying that a pure RL superlearner system could face training costs substantially above the baseline projection for equivalent supervised models. | 高 | SM007, SM008 |
| CM033 | Goldman Sachs Research explicitly notes that despite growing AI mentions in earnings calls (16%+ of Russell 3000 companies as of 2024) and rising investment, AI productivity benefits are likely to materialise only in the second half of the decade, implying that enterprise and non-hyperscaler buyers will defer large RL commitments. | 高 | SM003, SM005 |
| CM034 | Buyer concentration is extreme: approximately five hyperscalers (Google, Microsoft, Amazon, Meta, and to a lesser extent Apple and Alibaba) control the largest AI training clusters globally, making Ineffable dependent on a thin payer pool and limiting its negotiating leverage for compute access and IP terms. | 中 | SM006, SM024, SM011 |
| CM035 | Newspage.news and cited AI consultants including KYC Digital CTO Katrina Young warn that the UK government's minority stake in Ineffable provides no enforceable IP-anchoring or sovereignty conditions, raising questions about whether UK public capital achieves its stated strategic objectives. | 中 | SM022, SM023 |
| CM036 | Ineffable Intelligence has no disclosed commercial product, revenue, customer, or published research milestone as of 27 April 2026 (the run date); the entire adoption pathway from research to commercial payer is speculative at this stage. | 高 | SM012, SM013 |
| CM037 | The Kimi k1.5 paper notes that "prior published work has not produced competitive results" in RL scaling for LLMs before their work, suggesting that while RL scaling is advancing rapidly, it remains an unsettled research frontier with no commercial deployment at scale as of early 2025. | 中 | SM008, SM007 |
| CM038 | UK Frontier AI Safety Commitments signed at the AI Seoul Summit 2024 require leading AI organisations to conduct safety evaluations, share findings with governments, and not deploy models posing unacceptable risk — commitments that would apply to Ineffable as a UK-domiciled frontier lab at the relevant capability threshold. | 高 | SM025, SM017 |
| CM039 | Multiple independent analyst sources (Goldman Sachs, Statista, OECD AI Observatory) provide AI market size estimates ranging from ~$200 billion (investment flows) to multiple hundreds of billions (broad market revenue), with incompatible methodologies that preclude direct comparison; no source isolates RL training as a measurable sub-category. | 高 | SM003, SM004, SM005 |
| CM040 | The Goldman Sachs AI investment forecast article projects that over the long term, AI-related investment could peak at 2.5–4% of US GDP and 1.5–2.5% of GDP in other major AI-leading economies, if AI growth projections are fully realised, implying a very large long-term market but with significant uncertainty and timing risk. | 中 | SM003 |
| CM041 | OECD's AI Policy Observatory tracks AI investment across member countries and provides data showing growing government involvement in AI infrastructure, supporting the sovereign buyer thesis across OECD economies including EU member states. | 中 | SM005, SM010 |
| CP001 | OpenAI operates today as a public benefit corporation (OpenAI Group) governed by the OpenAI Foundation nonprofit, with a stated mission to ensure that artificial general intelligence benefits all of humanity. | 高 | SP001, SP002 |
| CP002 | OpenAI's published "Planning for AGI and beyond" post acknowledges that successfully transitioning to a world with superintelligence is "perhaps the most important — and hopeful, and scary — project in human history," directly positioning OpenAI as a paradigm-level competitor to Ineffable on the AGI mission. | 高 | SP002, SP001 |
| CP003 | In October 2024, OpenAI completed a $6.6 billion funding round at a $157 billion post-money valuation — at the time the largest single private technology raise in history — making it the most highly valued standalone AI startup by reported terms. | 中 | SP009 |
| CP004 | OpenAI's RL investment is concentrated in chain-of-thought reinforcement learning and process reward models (the o1/o3 family), which rely on human-generated training data as the foundation — making it paradigmatically distinct from Ineffable's proposed data-free experiential RL superlearner approach. | 高 | SP015, SP016 |
| CP005 | OpenAI publishes a public API pricing page at openai.com/api/pricing listing per-token rates for its full model family, establishing the market reference price for frontier AI API access — a commercial benchmark Ineffable cannot currently match or compare against. | 高 | SP003, SP017 |
| CP006 | OpenAI's safety infrastructure includes published system cards, usage policies, and safety evaluations; its safety page lists outputs including o3/o4-mini system cards and red-team findings — a formal safety apparatus that Ineffable does not yet have and will need to develop before any commercial deployment. | 高 | SP004, SP001 |
| CP007 | Anthropic was founded in 2021 by Dario Amodei, Daniela Amodei, and colleagues who departed OpenAI, with an explicit corporate mission centred on AI safety research and responsible deployment of frontier AI systems. | 高 | SP005, SP007 |
| CP008 | Anthropic's published "Core Views on AI Safety" explicitly states the company believes it "may be building one of the most transformative and potentially dangerous technologies in history" yet presses forward because safety-focused labs need to be at the frontier — a mission posture that directly overlaps with Ineffable's focus on superintelligence, but through an alignment-first LLM lens rather than a data-free RL lens. | 高 | SP007, SP005 |
| CP009 | As of January 2025, Anthropic was reportedly raising capital at a valuation of approximately $60 billion, representing strong continued investor demand for safety-focused frontier AI labs and more than 10x Ineffable's concurrent seed valuation. | 中 | SP011, SP013 |
| CP010 | Anthropic's published API pricing page lists per-token rates for the Claude 3 and Claude 4 model families including Haiku, Sonnet, and Opus tiers — commercially available pricing that Ineffable cannot match, as it has no commercial product or pricing as of the run date. | 高 | SP006, SP005 |
| CP011 | Anthropic's Constitutional AI and Responsible Scaling Policy (RSP) represent the industry's most systematized published safety framework for frontier AI deployment; Ineffable has no equivalent public safety architecture, creating a gap that must be addressed before any deployment under the AI Seoul Summit commitments. | 高 | SP007, SP005 |
| CP012 | Google DeepMind is the consolidated AI research organization of Alphabet Inc., formed from the 2023 merger of Google Brain and DeepMind — the institution where Ineffable's founder David Silver spent approximately two decades and led the AlphaGo, AlphaZero, AlphaStar, AlphaFold, and AlphaProof programs. | 高 | SP008, SP018 |
| CP013 | Google DeepMind's official about page states its team is "harnessing our unparalleled computing infrastructure to create the next wave of research breakthroughs" — explicitly positioning Alphabet's TPU compute scale as a primary competitive differentiator, a resource that Ineffable can only partially access via Google Cloud commercial agreements. | 高 | SP008, SP023 |
| CP014 | Google DeepMind's pioneering RL work on AlphaGo, AlphaZero, AlphaFold, and AlphaProof — all conducted under or alongside David Silver prior to his departure — represents the deepest extant track record in reinforcement learning at any scale, providing DeepMind with the institutional knowledge, data infrastructure, and talent base required to reconstitute a competitive RL program rapidly. | 高 | SP008, SP018 |
| CP015 | As a division of Alphabet, Google DeepMind has access to TPU Pod supercomputing infrastructure at an organizational scale that no standalone startup, including Ineffable, can independently replicate without a hyperscaler partnership agreement. | 中 | SP008 |
| CP016 | Google DeepMind's Gemini model family, delivered through Google Cloud, Google Search, and the Android operating system, gives it consumer and enterprise distribution at a scale that dwarfs any independent AI lab — including Ineffable, which has zero distribution. | 中 | SP008 |
| CP017 | xAI, founded by Elon Musk in 2023, raised $6 billion in a funding round valuing the company at approximately $80 billion in March 2025, making it one of the most heavily capitalised frontier AI labs after OpenAI and Anthropic. | 中 | SP012 |
| CP018 | xAI's Grok chatbot is integrated into the X (formerly Twitter) platform with approximately 500 million registered users, giving xAI a direct consumer distribution channel that Ineffable does not possess and cannot replicate through any currently disclosed partnership. | 中 | SP012 |
| CP019 | xAI's Colossus GPU cluster reportedly comprised approximately 100,000 NVIDIA H100 GPUs at the time of reporting, representing one of the largest single-purpose AI training clusters outside Google and Microsoft and substantially larger than Ineffable's disclosed Google Cloud cluster. | 低 | SP012 |
| CP020 | xAI has not published a formal safety framework or signed frontier AI safety commitments equivalent to the AI Seoul Summit 2024 standards signed by OpenAI, Anthropic, Google DeepMind, and other frontier labs — creating a regulatory posture divergence from Ineffable's UK/EU-anchored model. | 中 | SP012, SP025 |
| CP021 | Mistral AI, a French frontier AI lab founded in 2023 by ex-Google DeepMind and Meta Fundamental AI Research alumni, raised approximately $640 million in a funding round valuing the company at approximately $6 billion in June 2024. | 中 | SP010 |
| CP022 | Mistral's deployed product portfolio includes Le Chat (consumer and enterprise AI assistant with on-premises and sovereign cloud options), Mistral Large, Mistral Small, Codestral, Pixtral, and Devstral — a substantially more commercially deployed product set than Ineffable holds as of the run date. | 高 | SP014, SP010 |
| CP023 | Mistral is the leading European frontier AI lab by valuation and deployed product portfolio as of 2024–2025, making it the primary adjacent competitor to Ineffable for EU sovereign AI positioning, European regulatory alignment, and European enterprise buyer narratives. | 中 | SP010, SP014 |
| CP024 | Mistral offers commercial API access, enterprise Le Chat subscriptions, and optional on-premises sovereign deployment — pricing tiers that give European enterprise customers a viable EU-anchored commercial alternative that overlaps with Ineffable's potential future sovereign-buyer segment. | 中 | SP014 |
| CP025 | The five frontier AI labs most directly competitive with Ineffable's stated mission (OpenAI, Anthropic, Google DeepMind, xAI, and Mistral) collectively raised more than $20 billion in disclosed private rounds through 2025 and carry aggregate reported valuations exceeding $300 billion — representing a capital density and organisational scale that dwarfs Ineffable's $1.1 billion seed round. | 中 | SP009, SP011, SP012, SP010, SP019 |
| CP026 | None of Ineffable's named competitors pursues the specific technical paradigm of experiential RL without human data at the scale Ineffable proposes; the closest analogues are DeepMind's pre-2026 RL programs (AlphaGo, AlphaZero) which were confined to constrained game domains, not open-domain knowledge discovery. | 高 | SP008, SP015, SP018 |
| CP027 | OpenAI's o1/o3 family, Anthropic's Claude reasoning capabilities, and Google DeepMind's AlphaProof all employ variants of reinforcement learning applied to language model reasoning chains — but all are founded on human-generated training data, making them paradigmatically distinct from Ineffable's proposed data-free approach. | 高 | SP015, SP016, SP007, SP008 |
| CP028 | The dominant commercial model for frontier AI labs is token-based API pricing with tiered consumer subscription products; Ineffable has no commercial product, no API, and no pricing, and will need to define a novel commercial model if and when its superlearner system produces deployable output. | 高 | SP003, SP006, SP014 |
| CP029 | AMI Labs and Recursive Superintelligence are cited in commentator coverage as potential peers pursuing AGI or self-improvement paradigms, but neither entity has sufficient public disclosure (no official websites, funding rounds, or published research identified in this run's searches) to profile as a competitor with confidence. | 低 | SP018, SP019 |
| CP030 | The aggregate consumer distribution of Ineffable's competitors — ChatGPT (OpenAI, 200 M+ weekly users), Claude.ai (Anthropic), Gemini (Google Search + Android), and Grok (xAI, ~500 M X users) — creates an ecosystem moat that cannot be replicated by a research lab without a product, generating a severe deployment asymmetry if Ineffable eventually needs to commercialise. | 中 | SP001, SP005, SP008, SP012 |
| CP031 | Google DeepMind could reconstitute an experiential RL without-human-data research program internally, drawing on the same published RL research corpus, the remaining members of Silver's former team, and compute infrastructure that vastly exceeds Ineffable's — representing the most credible fast-follower risk with materially superior compute access. | 中 | SP008, SP014, SP018 |
| CP032 | OpenAI's structural shift to a public benefit corporation in 2025 removed one governance differentiation that previously distinguished nonprofit-anchored labs; with OpenAI now operating commercially at scale, the "pure research" positioning Ineffable articulates is more crowded as a narrative than it was at the time of Ineffable's founding. | 中 | SP001, SP009 |
| CP033 | Newspage.news commentary and cited AI experts explicitly question whether Ineffable's UK government minority stake prevents IP, talent, or breakthrough discoveries from flowing to US-headquartered investors who control the board — a governance gap that has direct implications for Ineffable's ability to maintain independent strategic direction and its competitive positioning as a sovereign AI asset. | 中 | SP021 |
| CP034 | The compute requirements for experiential RL training at superlearner scale exceed any amount independently accessible to Ineffable without its Google Cloud and NVIDIA partnerships; any competitor — particularly Google DeepMind or a US hyperscaler — that self-funds a similar program would have structural compute access superiority, not parity. | 中 | SP008, SP022, SP023, SP024 |
| CP035 | Frontier AI labs including OpenAI and Anthropic have invested heavily in reinforcement learning from human feedback (RLHF) as a core alignment technique, creating an institutional knowledge base in hybrid RL/supervised approaches that could be redirected toward purely experiential RL with sufficient motivation and compute — lowering the paradigm-switching cost for well-funded incumbents. | 中 | SP015, SP007 |
| CP036 | All five named direct and adjacent competitors hold at minimum one major hyperscaler partnership (OpenAI/Microsoft, Anthropic/Amazon+Google, DeepMind/Google inherently, xAI's own cluster, Mistral/Azure+Google); Ineffable's Google Cloud and NVIDIA partnerships partially offset this but do not provide the same multi-cloud optionality, contractually committed compute, or revenue integration that hyperscaler-backed competitors enjoy. | 中 | SP022, SP023, SP001, SP005, SP008 |
| CP037 | Mistral's European regulatory positioning — designed for EU AI Act compliance and offering open-weight models with sovereign deployment options — gives it a distinct go-to-market avenue for EU sovereign and enterprise customers that partially overlaps with Ineffable's UK/EU positioning, creating a potential near-term commercial collision risk if Ineffable pivots toward European enterprise deployment. | 中 | SP014, SP010 |
| CP038 | Ineffable Intelligence has no disclosed pricing model, commercial product, or API offering as of the run date; any pricing comparison for this chapter is therefore between Ineffable's undefined future commercial model and the existing published API pricing of OpenAI, Anthropic, and Mistral. | 高 | SP019, SP020 |
| CP039 | The combination of David Silver's two decades of primary-source RL expertise, a team of ex-DeepMind RL researchers, and the first large externally funded pure experiential RL research program constitutes a legitimate paradigm-leadership moment — but AI paradigm leadership has historically lasted 12–36 months before well-funded incumbents replicate the core approach, as evidenced by the rapid convergence after AlphaGo (2016) and transformer-based language modelling (2017–2020). | 中 | SP018, SP008, SP024 |
| CP040 | Incumbent frontier labs have demonstrated the ability to shift technical direction rapidly when a new paradigm proves commercially viable: DeepMind pivoted from game RL to protein folding (AlphaFold), achieving a Nobel Prize-winning result; OpenAI pivoted from language modelling to RL-based reasoning (o1 family) within approximately 24 months of observing commercial viability; Anthropic has iteratively invested in novel alignment-relevant RL research; these precedents suggest low paradigm-switching cost for well-resourced incumbents. | 高 | SP008, SP015, SP007 |
| CI001 | Ineffable Intelligence has no disclosed revenue, customers, or commercial products as of 22 June 2026; it operates as a pre-revenue research entity. | 高 | SI010, SI015 |
| CI002 | Ineffable's stated mission explicitly rejects near-term commercial output, seeking 'a window where ambitious research can thrive, without bending to the demands of incremental products and near-term profits.' | 中 | SI010 |
| CI003 | Ineffable Intelligence raised $1.1 billion in a seed round announced on 27 April 2026 at a $5.1 billion post-money valuation, co-led by Sequoia Capital and Lightspeed Venture Partners. | 高 | SI015, SI016 |
| CI004 | The $1.1 billion seed is the largest European seed round in history and one of the largest first-money-in rounds in global AI history. | 高 | SI015, SI016 |
| CI005 | No debt instruments, credit facilities, venture debt, project finance, or secondary transactions have been publicly disclosed for Ineffable Intelligence. | 中 | SI009, SI015 |
| CI006 | The British Business Bank confirmed a $20 million (£14.8 million) investment in Ineffable Intelligence as part of the seed round. | 高 | SI014, SI020 |
| CI007 | The UK Sovereign AI Fund co-invested in the seed round; the exact amount is described as 'commercially sensitive' and has not been disclosed. | 高 | SI014, SI020 |
| CI008 | Companies House records show a Statement of Capital (SH01) filed on 7 April 2026 in connection with an allotment of shares, approximately three weeks before the public seed round announcement. | 中 | SI009 |
| CI009 | Companies House records show a Statement of Capital (SH01) of GBP 350 nominal value filed on 12 March 2026, indicating a share allotment prior to the seed round close. | 中 | SI009 |
| CI010 | Companies House records show a Statement of Capital of GBP 320 nominal value filed on 21 January 2026, associated with the allotment of shares when David Silver was appointed director and person with significant control. | 中 | SI009 |
| CI011 | Ineffable Intelligence's directors approved an employee share option plan (with a US sub-plan) on 5 February 2026, prior to the seed round close. | 中 | SI009 |
| CI012 | No annual accounts or financial statements have been filed at Companies House as of 22 June 2026; the company was incorporated in November 2025 and first accounts are not yet legally due. | 中 | SI009 |
| CI013 | Google Cloud is Ineffable's preferred cloud provider, deploying one of the largest A5X clusters powered by NVIDIA Vera Rubin NVL72, as announced at Google Cloud Summit London on 16 June 2026. | 高 | SI011, SI004 |
| CI014 | The NVIDIA–Ineffable partnership involves engineers from both companies co-designing RL training pipelines, initially on Grace Blackwell and planning to be among the first to use the Vera Rubin platform. | 高 | SI012, SI013 |
| CI015 | Reinforcement learning workloads require continuous act-observe-score-update loops that 'put pressure on interconnect, memory bandwidth and serving in ways that pretraining doesn't,' according to the Ineffable/NVIDIA blog. | 高 | SI013, SI012 |
| CI016 | Epoch AI analysis finds that amortised hardware and energy costs for the final training run of frontier AI models have grown at a compound annual rate of 2.4× per year since 2016. | 中 | SI002 |
| CI017 | Epoch AI analysis estimates that hardware (AI accelerators, servers, interconnect) accounts for 47–67% of total frontier AI model development cost, R&D staff for 29–49%, and energy for 2–6%. | 中 | SI002 |
| CI018 | Epoch AI projects that if the trend of growing training costs continues, the largest training runs will cost more than $1 billion by 2027, making frontier AI training 'too expensive for all but the most well-funded organizations.' | 中 | SI002 |
| CI019 | NVIDIA Q2 FY2026 (quarter ended 27 July 2025) revenue was $46.7 billion, up 56% year-on-year, with Blackwell Data Center revenue growing 17% sequentially. | 中 | SI007 |
| CI020 | NVIDIA GAAP gross margin in Q2 FY2026 was 72.4%, and Jensen Huang described demand for Blackwell as 'extraordinary' with production ramping at 'full speed.' | 中 | SI007 |
| CI021 | Google Cloud's AI Hypercomputer architecture is described as a systems-level integration of GPU, networking, and storage specifically designed for frontier AI training workloads. | 中 | SI011, SI004 |
| CI022 | David Silver stated: 'We evaluated the space and chose Google Cloud as the best fit for our reinforcement learning infrastructure. We aren't just looking for processors; we are building a resilient and scalable environment to make first contact with superintelligence.' | 高 | SI004, SI011 |
| CI023 | No pricing model, commercial offering, or customer-facing product exists for Ineffable Intelligence; the company has no pricing page, no product page, and no disclosed commercial contracts. | 高 | SI010, SI015 |
| CI024 | Four plausible future revenue paths for Ineffable are: (a) API access to a deployed superlearner; (b) weight/capability licensing to enterprises and governments; (c) B2G sovereign AI contracts; and (d) royalties from scientific IP generated by the system. | 低 | SI010, SI004 |
| CI025 | Frontier AI labs Anthropic and OpenAI operated with multi-hundred-million-dollar annual compute-plus-staff budgets before reaching comparable capital levels; Ineffable's $1.1B seed places it in comparable capital territory to their early high-spend phases. | 中 | SI015, SI006 |
| CI026 | The NVIDIA Vera Rubin platform is described as the next-generation GPU successor to Grace Blackwell; being 'among the first to explore' it is an explicit goal of the NVIDIA–Ineffable engineering collaboration. | 高 | SI012, SI013 |
| CI027 | Sequoia Capital's 'Act Two' analysis warned that AI labs had entered 'an unsustainable feeding frenzy of fundraising, talent wars and GPU procurement' and that 'a lot of AI companies simply do not have product-market fit or a sustainable competitive advantage.' | 中 | SI006 |
| CI028 | UK AI Growth Zones policy, announced April 2025, aims to unlock investment in AI-enabled data centres, creating a potential non-dilutive funding pathway for qualifying AI infrastructure projects. | 高 | SI008, SI003 |
| CI029 | No customer acquisition cost (CAC), lifetime customer value (LTV), or payback period metric is publicly available for Ineffable Intelligence, which has no customers. | 高 | SI010, SI015 |
| CI030 | No ARR, GMV, unit volume, active users, or other revenue traction metric is publicly available for Ineffable; the company is explicitly pre-revenue. | 高 | SI010, SI015 |
| CI031 | No disclosed COGS, gross margin, operating leverage, or cost-of-delivery data is publicly available for Ineffable Intelligence. | 高 | SI010, SI012 |
| CI032 | The DSIT Annual Report 2023–24 describes £38.7 billion in total departmental expenditure, including grants to research councils, the Alan Turing Institute, and digital infrastructure programmes. | 中 | SI003 |
| CI033 | Governance critics have argued that UK government co-investment provides 'roughly one percent influence over systems that could generate strategically significant knowledge' and that 'sovereignty is not achieved through presence on a cap table.' | 中 | SI018, SI019 |
| CI034 | Google Cloud GPU pricing for accelerator-optimised instances (A3/A4 with H100/H200) runs at several dollars per GPU-hour on-demand; frontier RL training at Vera Rubin NVL72 scale would require thousands of GPUs operating for weeks to months per training run. | 低 | SI001, SI002 |
| CI035 | Arxiv research (2022) found that large-scale ML models required 10–100× more compute than standard deep learning, and that training compute in the Deep Learning era doubled approximately every 6 months. | 中 | SI005 |
| CI036 | A Vera Rubin NVL72 compute cluster of the scale described in the Google Cloud partnership represents one of the largest deployed AI training installations worldwide, with running costs estimated in the multi-million-dollar-per-month range before partnership subsidies. | 低 | SI001, SI007 |
| CI037 | Ineffable's research mission implies multi-year cycles before any commercial capability reaches the market; the company has explicitly declined to tie its work to near-term profits or incremental products. | 高 | SI010, SI016 |
| CI038 | The employee share option plan adopted by Ineffable on 5 February 2026 will generate non-cash share-based compensation charges under IFRS 2 once the company begins formal financial reporting. | 中 | SI009 |
| CI039 | The most plausible triggers for Ineffable's next equity round are: (a) a landmark scientific milestone that validates the superlearner capability thesis; (b) approaching exhaustion of seed capital; or (c) onset of a commercial deployment phase. | 低 | SI010, SI006 |
| CI040 | Any financial model of Ineffable's future revenue requires explicit scenario assumptions across four inputs: time to commercial deployment, target customer segment, pricing mechanism, and gross margin — none of which are currently disclosed. | 中 | SI010, SI015 |
| CI041 | UK government co-investment via the British Business Bank and Sovereign AI Fund reduces first-round financing risk for Ineffable but does not establish revenue sustainability or commercial viability. | 中 | SI014, SI020 |
| CE001 | Ineffable Intelligence has no commercial product, API, customer, or deployed AI system as of 22 June 2026; the company operates as a pure frontier research entity. | 高 | SE001, SE005 |
| CE002 | The company's stated mission is to build a "superlearner" — an AI system that discovers all knowledge from its own experience using reinforcement learning, without relying on human-generated data. | 高 | SE001, SE003 |
| CE003 | David Silver described the superlearner ambition as "a scientific breakthrough of comparable magnitude to Darwin: where his law explained all Life, our law will explain and build all Intelligence." | 高 | SE001, SE005 |
| CE004 | The core technical approach uses reinforcement learning, where agents learn through trial and error by acting in environments, observing outcomes, and updating their policies based on reward signals — no human-curated training data is used. | 高 | SE001, SE002, SE007 |
| CE005 | Unlike LLM pretraining on static datasets, RL workloads generate training data on the fly: the system must act, observe, score, and update continuously in tight loops. | 高 | SE002, SE003 |
| CE006 | RL-based training at frontier scale "puts pressure on interconnect, memory bandwidth and serving in ways that pretraining doesn't," per the co-authored NVIDIA-Ineffable blog. | 高 | SE002, SE003 |
| CE007 | NVIDIA and Ineffable announced an engineering-level collaboration in May 2026 to co-design RL infrastructure starting on Grace Blackwell and extending to the Vera Rubin platform. | 高 | SE002, SE003 |
| CE008 | The NVIDIA Vera Rubin NVL72 rack integrates 72 Rubin GPUs and 36 Vera CPUs connected via NVLink 6 and trains large mixture-of-experts models with one-quarter the GPU count of Blackwell, at up to 10x higher inference throughput per watt. | 高 | SE008, SE004 |
| CE009 | The NVIDIA Vera CPU Rack (256 Vera CPUs per rack) is purpose-built for RL and agentic workloads that require large numbers of CPU-based environments to test and validate GPU- generated model results. | 高 | SE008, SE002 |
| CE010 | The NVIDIA-Ineffable collaboration is described as "engineering-level," with engineers from both companies working together to build the RL training pipeline — it is not a standard cloud or hardware resale agreement. | 高 | SE002, SE003 |
| CE011 | Google Cloud was selected as Ineffable's preferred cloud partner in June 2026 following what Ineffable described as "a rigorous evaluation of the infrastructure market." | 高 | SE004, SE006 |
| CE012 | Under the Google Cloud partnership, Ineffable will deploy "one of the largest clusters of A5X, powered by the NVIDIA Vera Rubin NVL72 on Google Cloud." | 高 | SE004, SE006 |
| CE013 | Google Cloud's AI Hypercomputer architecture combines performance-engineered GPUs with Jupiter high-efficiency networking and optimised storage for tightly integrated training and inference workloads. | 高 | SE004, SE006 |
| CE014 | David Silver explicitly rejected a "box of chips" approach to GPU procurement and selected Google Cloud specifically for its AI Hypercomputer systems-level integration. | 高 | SE004, SE006 |
| CE015 | Ineffable's system will train on "rich forms of experience quite distinct from human language and other human data" and may require "novel model architectures and training algorithms," per the NVIDIA-Ineffable blog. | 高 | SE003, SE002 |
| CE016 | The NVIDIA Vera Rubin NVL72 delivers up to 10x higher inference throughput per watt and trains large MoE models with one-fourth the number of GPUs compared with Blackwell. | 高 | SE008, SE009 |
| CE017 | No model card, safety evaluation framework, alignment research agenda, red-teaming protocol, or responsible-scaling-policy equivalent has been published by Ineffable Intelligence as of 22 June 2026. | 高 | SE001, SE014, SE018 |
| CE018 | Ineffable Intelligence is not among the signatories of the Frontier AI Safety Commitments agreed at the Seoul AI Summit (May 2024) as of the February 2025 update to the signatory list. | 高 | SE018, SE019 |
| CE019 | No model evaluation framework, safety testing protocol, or capability-progression governance structure has been publicly disclosed by Ineffable Intelligence. | 高 | SE001, SE007 |
| CE020 | Ineffable has published no technical papers, preprints, or research blog posts describing any trained AI system or capability output under the Ineffable banner as of 22 June 2026. | 高 | SE001, SE015, SE016 |
| CE021 | The superlearner research platform is at concept and infrastructure build-out stage; no trained RL agent, benchmark result, or capability demonstration has been produced. | 高 | SE001, SE005, SE007 |
| CE022 | Silver described the intended scope of the superlearner as learning "from elementary motor skills through to profound intellectual breakthroughs" — an unbounded, open-ended capability goal without defined milestones. | 中 | SE004, SE007 |
| CE023 | IMPALA (Espeholt et al., 2018), co-authored by Ineffable co-founder Lasse Espeholt, established scalable distributed RL with importance-weighted actor-learner architectures — a direct technical precedent for the infrastructure Ineffable is building. | 高 | SE010, SE013 |
| CE024 | The A3C framework (Mnih et al., 2016), co-authored by David Silver, established asynchronous parallel RL training — one of the foundational techniques underpinning the distributed training pipelines Ineffable is building. | 高 | SE011, SE016 |
| CE025 | The "Era of Experience" position paper by Silver and Sutton formally argues that agents learning from experience are the paradigm required for superhuman AI, in contrast to LLMs bounded by human knowledge. | 中 | SE016, SE007 |
| CE026 | Simulation-to-real-world transfer is an unsolved research challenge: RL systems trained purely in simulation may fail when applied to out-of-distribution or real-world conditions, particularly in open-ended domains with ambiguous reward signals. | 中 | SE012, SE011 |
| CE027 | Ineffable's compute programme depends on timely delivery of NVIDIA Vera Rubin NVL72 at scale; NVIDIA supply constraints, priority allocation to competing hyperscalers, and software-stack maturity are external risks outside Ineffable's control. | 中 | SE008, SE002 |
| CE028 | The commercial terms of both the NVIDIA engineering collaboration and the Google Cloud partnership — pricing, compute credits, exclusivity provisions, and service-level agreements — are not publicly disclosed. | 高 | SE004, SE002 |
| CE029 | If either NVIDIA's Vera Rubin production ramp or Google Cloud's A5X cluster availability is delayed, Ineffable's training timeline is directly and materially affected, as no alternative compute supply has been disclosed. | 中 | SE008, SE004 |
| CE030 | The UK AI Regulation White Paper (2023) takes a pro-innovation, sector-specific approach with no specific binding frontier AI rules; Ineffable faces no specific regulatory obligations under UK law as of the run date. | 高 | SE019, SE025 |
| CE031 | The UK Sovereign AI Fund co-invested alongside the British Business Bank, providing not only equity capital but also access to UK supercomputing resources, visas, and the "unique levers of the British state." | 高 | SE013, SE023, SE025 |
| CE032 | No EU AI Act conformity assessment, designated EU representative, or GPAI compliance statement has been filed or published by Ineffable Intelligence as of 22 June 2026. | 高 | SE019, SE001 |
| CE033 | Ineffable has no public GitHub repository, open-source code, public API endpoint, developer documentation, or active external developer community as of 22 June 2026. | 高 | SE001, SE015, SE016 |
| CE034 | David Silver's personal blog lists his publications through 2025, including papers in Nature (2025) on RL algorithm discovery and Olympiad-level mathematical reasoning — none attributed to Ineffable Intelligence. | 高 | SE016, SE013 |
| CE035 | David Silver's DeepMind portfolio includes leading or centrally contributing to AlphaGo, AlphaZero, AlphaStar, AlphaFold, and AlphaProof — the most significant body of published reinforcement learning results in the field's history. | 高 | SE013, SE016, SE024 |
| CE036 | NVIDIA confirmed that the Vera Rubin platform is "in full production" as of GTC (June 2026), with seven new chips in full production for the platform. | 高 | SE008, SE004 |
| CE037 | David Silver cited Google Cloud's "Jupiter networking" and "AI Hypercomputer" integrated architecture as the decisive selection criterion over competitors offering "a box of chips" approach. | 高 | SE004, SE006 |
| CE038 | Ineffable's stated ambition is to rediscover and then transcend "the greatest inventions in human history, such as language, science, mathematics and technology." | 高 | SE004, SE001 |
| CE039 | The quality and diversity of simulation environments used to generate RL training experience is a central unknown: no simulation engine, domain coverage, or reward model design has been disclosed by Ineffable. | 中 | SE002, SE003 |
| CE040 | Anthropic's Responsible Scaling Policy (v3.3, May 2026) sets capability-progression thresholds and governance gates before training proceeds at each tier; Ineffable has published no equivalent safety governance framework. | 高 | SE014, SE018 |
| CU001 | Ineffable Intelligence has zero paying customers as of 22 June 2026. | 高 | SU018, SU015, SU016 |
| CU002 | Ineffable Intelligence has generated zero commercial revenue as of 22 June 2026. | 高 | SU014, SU013, SU015 |
| CU003 | Ineffable Intelligence operates in an explicitly pre-commercial research phase, with David Silver stating the company will not "bend to the demands of incremental products and near-term profits." | 高 | SU018, SU016 |
| CU004 | NVIDIA and Ineffable Intelligence announced an engineering-level co-development collaboration to design reinforcement learning training infrastructure, beginning on Grace Blackwell and extending to the NVIDIA Vera Rubin platform. | 高 | SU019, SU012 |
| CU005 | Google Cloud announced a preferred-partner agreement with Ineffable Intelligence in June 2026, under which Ineffable will deploy one of the largest A5X GPU clusters on Google Cloud's AI Hypercomputer. | 高 | SU011, SU002, SU006 |
| CU006 | The British Business Bank invested $20 million in Ineffable Intelligence as part of the $1.1 billion seed round, representing approximately 1.8% of the total round. | 高 | SU013, SU014 |
| CU007 | The UK Sovereign AI Fund co-invested alongside the British Business Bank in Ineffable Intelligence and additionally provides access to the UK's largest AI supercomputers and visa support. | 高 | SU001, SU014 |
| CU008 | The UK Science and Technology Secretary stated Ineffable Intelligence has "the potential to transform entire sectors," representing the government's stated rationale for its investment. | 高 | SU014, SU001 |
| CU009 | David Silver described the superlearner's intended output as discovering "new forms of science, technology, government or economics" and stated the mission is "making first contact with superintelligence." | 中 | SU016, SU018 |
| CU010 | The UK Sovereign AI Fund's support for Ineffable includes "the unique levers of the British state" but does not include disclosed governance rights over IP, downstream use, or commercialisation outcomes. | 中 | SU001, SU010 |
| CU011 | Independent AI governance experts characterised the UK government's stake in Ineffable as "participation, not sovereignty" and argued that a small equity cheque does not buy meaningful governance over AI outputs. | 中 | SU010 |
| CU012 | Governance consultant Katrina Young stated the current Ineffable investment model means governance terms are "not visible" and that "sovereignty is not achieved through presence on a cap table." | 中 | SU010 |
| CU013 | Five prospective buyer segments are identifiable for a commercialised superlearner platform: sovereign AI labs, hyperscalers, science research organisations, enterprise R&D, and defence and national security agencies. | 中 | SU001, SU014, SU004 |
| CU014 | Procurement timelines for government sovereign AI buyers and national AI programme contracts typically range from 18 to 36 months after capability demonstration, based on standard public sector IT and research procurement patterns. | 低 | SU004 |
| CU015 | No pilot programmes, letters of intent, or pre-commercial agreements with any customer have been disclosed by Ineffable Intelligence. | 高 | SU018, SU015, SU016 |
| CU016 | No NRR, GRR, customer count, ARR, churn, satisfaction score, or any other commercial retention metric exists for Ineffable Intelligence. | 高 | SU018, SU013 |
| CU017 | NVIDIA's relationship with Ineffable is a technology co-development and infrastructure supplier partnership; NVIDIA is not a customer of any Ineffable product or service. | 高 | SU019, SU012 |
| CU018 | Google Cloud's relationship with Ineffable is as a preferred infrastructure provider; Ineffable is Google Cloud's paying customer for compute, not the reverse. | 高 | SU011, SU002 |
| CU019 | The UK Sovereign AI Fund and British Business Bank are equity investors in Ineffable Intelligence; neither entity is a product customer or has disclosed product use rights. | 高 | SU013, SU014, SU001 |
| CU020 | Ineffable Intelligence's stated target applications include scientific discovery across medicine, engineering, science, and mathematics, suggesting that research institutions and enterprise R&D labs are the intended product users. | 中 | SU018, SU014 |
| CU021 | David Silver described the superlearner as intended to "rediscover and then transcend the greatest inventions in human history, such as language, science, mathematics and technology." | 中 | SU011, SU018 |
| CU022 | The OECD reported that more than one-third of individuals across OECD member countries used generative AI tools in 2025, indicating broad awareness and early adoption of AI among the institutional buyer segments Ineffable targets. | 中 | SU004 |
| CU023 | The OECD noted that initial evidence on AI's potential shows a 20-40% improvement in specific workplace tasks, with long-term macroeconomic gains dependent on widespread adoption — indicating commercial pressure on institutional buyers to demonstrate AI value. | 中 | SU004 |
| CU024 | Ineffable Intelligence's entire disclosed compute infrastructure is concentrated on a single cloud provider (Google Cloud) and a single hardware platform (NVIDIA Vera Rubin NVL72), creating critical single-vendor concentration risk. | 高 | SU011, SU019 |
| CU025 | No public evidence exists of any product demonstration, product beta programme, public benchmark, model card, or technical evaluation of any Ineffable Intelligence system. | 高 | SU018, SU015 |
| CU026 | Ineffable Intelligence has not disclosed any pricing, commercialisation roadmap, go-to-market strategy, or target customer revenue model in any public filing or announcement. | 高 | SU018, SU016 |
| CU027 | AI governance expert Rohit Parmar-Mistry stated "who gets to govern what happens next?" as the central governance question for Ineffable, noting that the current investment structure leaves downstream IP ownership and commercialisation control unresolved. | 中 | SU010 |
| CU028 | No information is publicly available about what contractual or governance conditions — if any — the UK government attached to its equity investment in Ineffable Intelligence; the Department for Science, Innovation and Technology described investment terms as "commercially sensitive." | 高 | SU010, SU014 |
| CU029 | Ineffable Intelligence's path to first commercial customer requires at minimum: a demonstrated scientific breakthrough, product packaging of that breakthrough, and a commercial procurement cycle — a sequential process that is unlikely to complete in under four years. | 中 | SU003, SU004 |
| CU030 | Based on comparable frontier AI research lab timelines (DeepMind AlphaFold to commercial deployment: approximately five years), a five-to-seven year horizon to first paying customer for Ineffable Intelligence is a plausible base-case estimate. | 低 | SU009, SU004 |
| CU031 | Sovereign AI labs and national AI programmes represent the most institutionally aligned near-term buyer segment for Ineffable Intelligence, given the existing Sovereign AI Fund backing and the government's stated desire to be an "AI maker not taker." | 中 | SU001, SU014 |
| CU032 | Ineffable Intelligence's open job roles (as of June 2026) focus on research engineering and machine learning science; no sales, customer success, business development, or account management roles are visible in the public listings. | 中 | SU003 |
| CU033 | Ineffable Intelligence has not disclosed any channel partners, resellers, distribution strategy, or OEM licensing arrangements with any third party. | 高 | SU018, SU015 |
| CU034 | Google Cloud CEO Thomas Kurian stated that Ineffable is "leveraging our full-stack AI Hypercomputer, from Jupiter networking to our optimized storage," confirming Ineffable's role as an infrastructure consumer rather than a product supplier. | 高 | SU011, SU002 |
| CU035 | David Silver stated that Ineffable "evaluated the space and chose Google Cloud as the best fit" after a rigorous market evaluation, indicating deliberate infrastructure procurement rather than a commercial customer relationship. | 高 | SU011, SU002 |
| CU036 | Neither the NVIDIA partnership nor the Google Cloud infrastructure agreement constitutes product customer adoption; both are supply-side relationships where Ineffable receives technology and compute, not product relationships where Ineffable sells a service. | 高 | SU019, SU011 |
| CU037 | The Sovereign AI Fund's investment in Ineffable was described as its "second direct investment in just a couple of months," indicating the fund is early in its operations with limited track record of translating investment into customer relationships. | 中 | SU001, SU014 |
| CU038 | The British Business Bank noted its AI portfolio includes nine AI investments in the last twelve months, including Wayve and PolyAI, indicating Ineffable is not uniquely advantaged by government backing relative to other portfolio companies. | 高 | SU013, SU014 |
| CU039 | Enterprise AI procurement for frontier research capabilities typically involves multi-year vendor evaluation, proof-of-concept phases, and legal review of IP and governance terms before any commercial commitment, creating structural procurement friction. | 中 | SU004, SU023 |
| CU040 | The UK government obtained "first refusal on the next round" as a condition of its investment in Ineffable Intelligence, which is an investment option — not an evidence of product access, customer rights, or technology deployment commitment. | 中 | SU010, SU014 |
| CU041 | An AI governance expert quoted in newspage.news asked "who audits outputs that sit beyond existing human understanding?" — flagging the absence of any oversight mechanism for superlearner outputs as a material governance and downstream-customer risk. | 中 | SU010 |
| CR001 | Ineffable Intelligence's superlearner, if deployed or marketed in any EU member state, is highly likely to qualify as a General Purpose AI (GPAI) model under Article 3(63) of the EU AI Act and may qualify as a GPAI model of systemic risk under Article 51 due to training compute scale; however, Ineffable has not disclosed any EU AI Act compliance programme as of June 2026. | 中 | SR009, SR018 |
| CR002 | The UK had no binding AI-specific legislation as of June 2026; the 2023 AI Regulation White Paper adopted a principles-based pro-innovation approach relying on existing regulators, with no dedicated AI Act and no AI Bill introduced to Parliament. | 高 | SR009, SR022 |
| CR003 | The National Security and Investment Act 2021 designates artificial intelligence technology as one of 17 mandatory notification sectors; any acquisition of material influence over Ineffable Intelligence by a non-UK acquirer must be notified to the Investment Security Unit for review. | 高 | SR023, SR002 |
| CR004 | ICO guidance on AI and data protection (updated March 2023) requires organisations developing or deploying AI to apply UK GDPR principles including data protection by design, purpose limitation, transparency, and accountability; these obligations apply to Ineffable's RL pipeline to the extent it processes personal data. | 高 | SR024, SR009 |
| CR005 | Ineffable Intelligence has not publicly confirmed the appointment of a Data Protection Officer, the existence of a data processing agreement, or any engagement with the ICO as of June 2026. | 中 | SR001, SR004 |
| CR006 | The Seoul Summit Frontier AI Safety Commitments (May 2024) are voluntary; Ineffable Intelligence has not publicly confirmed it has signed these commitments or any equivalent voluntary safety framework as of June 2026. | 中 | SR010, SR004 |
| CR007 | David Silver held a professorial role at UCL and was employed at Google DeepMind for approximately two decades; UCL's intellectual property policy applies to inventions made by staff, and no IP assignment agreement between Silver, UCL, DeepMind, and Ineffable has been publicly disclosed. | 中 | SR001, SR012 |
| CR008 | Ineffable Intelligence Ltd is registered as a standard private limited company under the Companies Act 2006 and has no disclosed public-benefit covenant, CIC designation, or equivalent governance structure that would legally bind it to public-interest obligations. | 高 | SR001, SR002 |
| CR009 | Ineffable Intelligence's entire training pipeline is co-designed around NVIDIA's Grace Blackwell and Vera Rubin NVL72 hardware deployed on Google Cloud; no alternative compute arrangement capable of equivalent RL loop performance has been disclosed. | 高 | SR007, SR001 |
| CR010 | NVIDIA announced the Vera Rubin NVL72 platform in May 2026; as of June 2026 the platform had not yet reached general availability, and production capacity allocation commitments to Ineffable specifically have not been publicly disclosed. | 高 | SR007, SR031 |
| CR011 | RL's landmark achievements — AlphaGo, AlphaZero, AlphaStar — operated in bounded game environments with precisely defined reward signals; scaling RL to open-ended knowledge discovery in open-world environments with loosely specified rewards has not been demonstrated by any laboratory as of June 2026. | 中 | SR018, SR030 |
| CR012 | Reward hacking and specification gaming are documented failure modes of RL at scale: systems optimise the proxy reward function rather than the intended goal. Ineffable has disclosed no reward specification framework, evaluation methodology, or red-teaming programme to mitigate these failure modes. | 中 | SR019, SR021 |
| CR013 | Ineffable Intelligence has no publicly disclosed responsible scaling policy, safety framework, alignment research agenda, head of safety, red-teaming programme, or external safety advisory body as of 22 June 2026. This distinguishes Ineffable from every other frontier AI lab at comparable funding scale, all of which have published at least a safety commitment document. | 高 | SR001, SR021 |
| CR014 | Google Cloud's preferred partnership with Ineffable is described as non-exclusive in public announcements; Google Cloud simultaneously serves Google DeepMind, Microsoft Azure (indirectly via OpenAI's infrastructure), and its own internal AI workloads, which may compete for capacity. | 中 | SR008, SR026 |
| CR015 | Frontier RL training-inference tight loops require tighter latency and higher memory bandwidth than standard supervised pre-training workloads, resulting in higher per-effective-FLOP energy costs; total power and cooling requirements for a superlearner training run at frontier scale are not publicly quantified by Ineffable. | 中 | SR007, SR029 |
| CR016 | The simulation-to-real gap — the documented failure of RL systems trained in simulated environments to transfer capabilities to open-world settings — is a recognised research challenge; Ineffable's superlearner concept depends on open-world generalisation that would require solving this gap. | 中 | SR018, SR019 |
| CR017 | Sequoia Capital (Alfred Lin) and Lightspeed Venture Partners (Ravi Mhatre) hold board seats as directors of Ineffable Intelligence Ltd, co-led the seed round, and are expected by market convention to lead or signal the Series A round. | 高 | SR003, SR006 |
| CR018 | The UK Sovereign AI Fund and British Business Bank have invested in Ineffable Intelligence; the specific governance rights, information rights, and public-benefit obligations attached to these investments are not publicly disclosed in any government or company document. | 中 | SR013, SR004 |
| CR019 | NVIDIA is simultaneously an equity investor in Ineffable and the sole disclosed hardware provider for its training infrastructure; this dual role creates a potential conflict of interest in supply allocation decisions that is not addressed by any publicly disclosed conflict-of-interest policy. | 中 | SR007, SR003 |
| CR020 | Google Cloud is simultaneously an equity investor in Ineffable and its preferred cloud infrastructure provider; no conflict-of-interest policy governing supply allocation or preferential treatment has been disclosed. | 中 | SR008, SR003 |
| CR021 | UCL provides a talent pipeline and academic credibility anchor through David Silver's ongoing professorial affiliation; however, UCL affiliation also creates publishing pressure and an institutional expectation that research discoveries will be published rather than retained as proprietary IP. | 中 | SR012, SR001 |
| CR022 | No contingency plan, secondary compute arrangement, or business continuity plan has been publicly disclosed by Ineffable in the event that the NVIDIA or Google Cloud partnerships are disrupted. | 高 | SR001, SR007 |
| CR023 | Ineffable confirmed NVIDIA and Google Cloud infrastructure partnerships in May and June 2026 respectively; the commercial terms, pricing, exclusivity provisions, SLA guarantees, and capacity allocation commitments in both partnerships are not publicly disclosed. | 高 | SR007, SR008 |
| CR024 | No secondary or fallback compute provider (Lambda Labs, CoreWeave, AWS Trainium, Microsoft Azure) has been disclosed by Ineffable; the company has no publicly documented multi-cloud strategy. | 高 | SR001, SR031 |
| CR025 | Ineffable Intelligence raised $1.1 billion at a $5.1 billion post-money valuation in April 2026 and has no disclosed revenue, products, customers, or commercial contract as of June 2026. | 高 | SR003, SR006 |
| CR026 | Ineffable has disclosed no burn rate, runway estimate, headcount, use-of-funds breakdown, management accounts, auditor appointment, or financial reporting framework as of June 2026 — unusual for a company that has raised $1.1 billion. | 高 | SR001, SR003 |
| CR027 | Sequoia Capital's "Generative AI: Act Two" analysis (2023) documented a dangerous gap between AI compute spending and AI revenue, warning that labs burning billions on compute without product market fit face existential financing risk; this thesis applies directly to Ineffable's pre-revenue posture. | 高 | SR014, SR015 |
| CR028 | Based on Epoch AI modelling and Wired's reporting on frontier training costs, a single frontier-scale RL training run could cost in the hundreds of millions of dollars; at a comparable burn rate, $1.1 billion provides an estimated 18–36 months of runway before a next round is required. | 中 | SR017, SR029 |
| CR029 | Ineffable has no disclosed commercial product, pricing model, customer, API, or revenue stream as of June 2026; the company has explicitly stated it seeks a window for research "without bending to the demands of incremental products and near-term profits." | 高 | SR001, SR011 |
| CR030 | If the $1.1 billion seed proves insufficient for a full superlearner training run before demonstrable capability, Ineffable will require a follow-on round with no commercial milestone to anchor valuation — a structurally difficult fundraising scenario in a potentially changed AI market. | 中 | SR014, SR025 |
| CR031 | The UK government's combined investment through British Business Bank and the Sovereign AI Fund represents approximately $30–40 million — less than 4% of the $1.1 billion seed — giving public funders limited economic leverage over commercial terms or governance outcomes. | 中 | SR006, SR004 |
| CR032 | David Silver is the sole named founder and CEO of Ineffable Intelligence; no co-founder, Chief Technology Officer, Chief Operating Officer, Chief Financial Officer, or Chief Safety Officer has been publicly confirmed as of June 2026. | 高 | SR003, SR012 |
| CR033 | David Silver's career prior to founding Ineffable was entirely as a research scientist at UCL and Google DeepMind; there is no publicly available evidence of prior company-building, operational leadership, commercial partnership development, or product management experience. | 中 | SR012, SR011 |
| CR034 | Ineffable Intelligence's public job board on Ashby shows open roles including senior RL researchers, ML infrastructure engineers, and research scientists, as of June 2026 — indicating the research team is still being assembled approximately twelve months after incorporation. | 高 | SR001, SR011 |
| CR035 | Google DeepMind, OpenAI, and Anthropic all maintain highly competitive frontier RL research programmes and recruiting pipelines; researchers with AlphaGo or AlphaZero-level experience are among the rarest and most in-demand in the industry. | 中 | SR026, SR011 |
| CR036 | No co-founder, named scientific co-lead, board-level scientific committee, or disclosed succession plan for David Silver has been published by Ineffable or any of its investors as of June 2026. | 高 | SR001, SR016 |
| CR037 | Ineffable has published no responsible AI principles, ethics governance framework, internal review board, or external advisory body documentation as of June 2026. | 高 | SR001, SR021 |
| CR038 | UCL's academic culture creates a structural publishing pressure for affiliated researchers; there is an inherent tension between the open-science norms of an academic lab and the proprietary commercialisation model that would maximise Ineffable's IP value. | 中 | SR012, SR021 |
| CR039 | The thesis-break triggers identified for Ineffable are: (1) David Silver's departure from the CEO role; (2) NVIDIA Vera Rubin NVL72 delay or reallocation of more than twelve months; (3) no RL generalisation milestone from any frontier source within thirty months without Ineffable counter-evidence; and (4) failure to close a Series A at or above $4 B valuation within thirty-six months. | 中 | SR014, SR025 |
| CR040 | Monitorable leading indicators for the primary risk clusters include: NVIDIA earnings call language on NVL72 ramp; AISI engagement announcements; Ineffable technical milestone disclosures; frontier RL benchmark publications; Companies House director changes; and Sequoia/Lightspeed Series A signalling. | 中 | SR007, SR025 |
| CR041 | RL scaling limits in open-ended environments are documented in peer-reviewed literature; if these limits prove fundamental rather than engineering-solvable, Ineffable's technical thesis breaks regardless of capital adequacy or team quality. | 中 | SR018, SR030 |
| CR042 | Independent commentators in Newspage and Electronics Weekly, writing in April 2026, provided specific adverse analysis: public money without governance rights, IP covenants, or public-benefit enforcement mechanisms provides legitimacy to the company without accountability to the public — a structural governance failure at the seed stage. | 中 | SR004, SR005 |
| CV001 | Ineffable Intelligence raised $1.1 billion at a $5.1 billion post-money valuation in its April 2026 seed round. | 高 | SV011, SV012, SV013 |
| CV002 | The $5.1 billion seed valuation is the largest for any European seed-stage company in recorded history at the time of announcement. | 高 | SV012, SV013 |
| CV003 | The $5.1 billion post-money valuation represents approximately 4.6 times the $1.1 billion capital raised, an unprecedented ratio for a zero-revenue seed-stage entity. | 中 | SV011, SV012 |
| CV004 | Sequoia Capital and Lightspeed Venture Partners co-led the April 2026 seed round and both hold board seats at Ineffable Intelligence. | 高 | SV011, SV013 |
| CV005 | Lightspeed Venture Partners published an investment rationale for Ineffable stating that David Silver spent nearly two decades turning RL from a research idea into the results the rest of the field builds on. | 中 | SV010, SV001 |
| CV006 | The Sovereign AI UK head of ventures described the investment as backing a founder who could credibly build a superlearner that discovers new knowledge from its own experience. | 高 | SV001, SV024 |
| CV007 | OpenAI closed a $6.6 billion funding round in October 2024 at a $157 billion post-money valuation, with approximately $4 billion in annualised revenue at the time of the round. | 高 | SV014, SV012 |
| CV008 | Anthropic was reported to be raising capital at approximately $60 billion valuation in January 2025, with Amazon having committed up to $4 billion in investment. | 中 | SV015, SV030 |
| CV009 | xAI raised $6 billion in a March 2025 funding round valuing the company at approximately $80 billion; Grok, its AI assistant, was deployed to consumers at the time. | 中 | SV016, SV004 |
| CV010 | Mistral AI raised approximately €600 million at an approximately €6 billion (roughly $6.4 billion) valuation in June 2024, with Le Chat deployed and API pricing published. | 中 | SV022, SV009 |
| CV011 | OpenAI's $157 billion valuation implies an approximately 39× revenue multiple against its estimated $4 billion annualised revenue — a revenue anchor entirely absent from the Ineffable comparables. | 中 | SV014, SV023 |
| CV012 | No independent analyst, research house, or investment bank has published a standalone price target, discounted cash flow model, or formal valuation opinion for Ineffable Intelligence as of 22 June 2026. | 高 | SV011, SV019 |
| CV013 | David Silver co-authored AlphaGo, AlphaZero, AlphaStar, and AlphaProof at Google DeepMind, representing the most commercially and scientifically consequential body of reinforcement learning research in history. | 高 | SV028, SV010 |
| CV014 | Ineffable Intelligence's data-free experiential RL approach is, to public knowledge, the only seed-scale pursuit of this paradigm globally as of June 2026, creating genuine paradigm distinctiveness that no named incumbent has matched. | 中 | SV021, SV028 |
| CV015 | NVIDIA and Google Cloud, as strategic infrastructure partners and equity co-investors in Ineffable, have economic incentives to support its research progress independent of commercial product outcomes. | 中 | SV006, SV013 |
| CV016 | Ineffable Intelligence has no disclosed revenue, customers, product, commercial timeline, or burn rate as of 22 June 2026, making standard financial valuation frameworks (revenue multiples, EBITDA, DCF) inapplicable. | 高 | SV021, SV019 |
| CV017 | The $5.1 billion seed valuation implies that investors expect commercial proof equivalent in scale to OpenAI or Anthropic within a 7–15 year horizon, for which no public evidence exists. | 中 | SV023, SV011 |
| CV018 | Sequoia Capital's Generative AI Act Two analysis warned that AI labs without near-term product-market fit face financing risk as compute costs balloon and investor patience erodes. | 高 | SV017, SV023 |
| CV019 | Sequoia's AI capital allocation analysis warned that the concentration of pre-revenue AI lab valuations at seed stage creates systemic financing risk if milestone evidence is delayed. | 中 | SV023, SV017 |
| CV020 | MIT Technology Review (2025) published analysis identifying fundamental barriers for data-free RL in open-ended environments, including reward specification, distributional shift, and catastrophic forgetting as unresolved research problems. | 中 | SV025, SV020 |
| CV021 | The bull-case scenario for Ineffable assumes a verifiable RL milestone by 2027–28, a first commercial contract by 2028–29, and a Series B at 2× or more the seed valuation, implying a $30–75 billion enterprise value by 2032. | 低 | SV023, SV011 |
| CV022 | The primary upward sensitivity driver in the bull case is the scale and credibility of the published RL milestone; the primary downward sensitivity driver is David Silver's departure, estimated at a $40 billion valuation impact. | 低 | SV023, SV028 |
| CV023 | The base-case scenario assumes research progress visible by 2028 but commercial deployment delayed to 2030–31, with 1–2 bridge rounds at modest step-up, implying a $8–20 billion enterprise value at a 2032 exit. | 低 | SV023, SV017 |
| CV024 | The bear-case scenario assumes RL plateau confirmation by 2027–28 without Ineffable counter-evidence, Silver departure or pivot, and capital exhaustion before commercial proof, implying a $0.5–4 billion outcome (fire sale or acqui-hire). | 低 | SV025, SV017 |
| CV025 | The single greatest threat to Ineffable's bull case is confirmation that the fundamental technical premise — experiential RL scaling to superintelligence without human data — fails to generalise beyond constrained environments. | 中 | SV025, SV020 |
| CV026 | David Silver's departure from the CEO role without a disclosed successor would constitute a thesis-break event, collapsing the key-person thesis that underpins all positive scenario probabilities. | 中 | SV028, SV010 |
| CV027 | A Series A at or below the $5.1 billion post-money seed valuation would signal market reassessment of capability claims and materially increase the probability of terminal financing failure. | 中 | SV023, SV017 |
| CV028 | Sequoia Capital's 2026 AI capital allocation analysis highlighted concentration risk in pre-revenue AI lab valuations, warning that AGI-narrative pricing has historically stretched multiples beyond defensible fundamentals. | 中 | SV023, SV017 |
| CV029 | No public secondary market trade, tender offer, structured liquidity event, or Series A announcement for Ineffable Intelligence has been reported as of 22 June 2026. | 中 | SV005, SV011 |
| CV030 | The UK Sovereign AI Fund and British Business Bank combined contributed approximately $20 million to Ineffable's seed round, representing less than 2% of the total, with no disclosed IP rights, governance veto, or enforceable claim on discoveries. | 高 | SV001, SV018 |
| CV031 | The recommendation for Ineffable Intelligence is conditional monitored interest (watch): the positive signals — founder quality, paradigm uniqueness, capital — do not override the structural disqualifiers — zero revenue, absent commercial proof, unresolved RL risk — for a conviction lead position at the seed valuation. | 中 | SV023, SV017 |
| CV032 | Confidence in any definitive valuation judgment for Ineffable is low: no filed accounts, no audit, no commercial proof, no disclosed burn rate, and no use-of-funds exist as public evidence. | 高 | SV019, SV021 |
| CV033 | The overall risk rating for Ineffable is very high, reflecting the simultaneous presence of key-person concentration, paradigm-level technical uncertainty, compute concentration dependency, governance and safety framework absence, and financial opacity. | 中 | SV025, SV020 |
| CV034 | The valuation stance for Ineffable is stretched but precedent-consistent: $5.1 billion at seed is unprecedented in European venture history but sits within the pricing tier established by US frontier AI labs, none of which were pre-revenue at their comparable rounds. | 中 | SV023, SV014 |
| CV035 | Stanford HAI's 2025 AI Index documented that global private investment in AI exceeded $100 billion in 2024, with frontier model investments comprising the dominant share — confirming the investment-wave environment in which Ineffable's seed was priced. | 中 | SV002, SV026 |
| CV036 | McKinsey and Sequoia Capital analysis of the 2025–2026 AI investment landscape found no established public methodology for benchmarking pre-revenue frontier AI lab valuations against each other. | 中 | SV023, SV026 |
| CV037 | NVIDIA's fiscal year 2026 annual results showed total revenue of approximately $130.5 billion, with data-centre revenue alone at approximately $115 billion, underscoring the scale of the compute market Ineffable depends on. | 高 | SV008, SV020 |
| CV038 | Goldman Sachs forecast that global AI investment would approach $200 billion annually by 2025, consistent with the investment environment in which Ineffable's seed was priced but not a direct valuation anchor. | 中 | SV027, SV026 |
| CV039 | No proximate exit route — IPO, strategic M&A, or structured secondary liquidity — for Ineffable is currently visible or disclosed; a realistic exit horizon for seed investors begins no earlier than 2030, and more plausibly 2032+. | 中 | SV029, SV023 |
| CV040 | The minimum data-room asks for Ineffable before conditional underwriting cover: monthly burn rate and runway, board-approved technical milestone OKRs, full cap table and preference structure, safety governance documents, and precise IP rights held by sovereign co-investors. | 中 | SV019, SV023 |
| CV041 | Lightspeed Venture Partners published an investment rationale for Ineffable citing Silver's creation of AlphaGo and AlphaZero as proof of ability to achieve scientific breakthroughs that redefined what AI systems could achieve. | 中 | SV010, SV001 |
| CV042 | Google Cloud's blog post on the Ineffable partnership states it is deploying one of the largest A5X clusters powered by NVIDIA Vera Rubin NVL72, framing the collaboration as a systems-level AI infrastructure investment. | 中 | SV006, SV005 |
| CV043 | xAI's official company page confirms it is a for-profit entity pursuing the true nature of the universe with a deployed consumer product (Grok), contrasting with Ineffable's pure research, pre-revenue posture. | 中 | SV004, SV016 |
| CV044 | UK Sovereign AI's official announcement described its Ineffable investment as backing a homegrown AI lab with potential to transform entire sectors, without specifying IP rights, governance provisions, or discovery-access terms. | 高 | SV001, SV024 |
| CV045 | Companies House filing 16865241 confirms Ineffable Intelligence Ltd was incorporated in November 2025 and has filed no annual accounts as of June 2026, consistent with its pre-revenue seed-stage status. | 高 | SV019, SV021 |
| CV046 | The frontier AI lab comparable set for benchmarking Ineffable includes OpenAI ($157B, Oct 2024), Anthropic (~$60B, Jan 2025), xAI (~$80B, Mar 2025), and Mistral (~€6B, Jun 2024) — all of which had deployed commercial products at the time of their cited rounds, making none a true like-for-like pre-revenue comp. | 中 | SV007, SV014, SV015, SV016 |
| 编号 | 出版方 | 标题 | 引文 |
|---|---|---|---|
| SO001 | Ineffable Intelligence | Ineffable Intelligence — Official Homepage | We are creating a superlearner that discovers all knowledge from its own experience, from elementary motor skills through to profound intellectual breakthroughs. |
| SO002 | UK Companies House | INEFFABLE INTELLIGENCE LTD — Company Overview (16865241) | Incorporated on 19 November 2025; Private limited Company; SIC 74909; Registered office 3rd Floor 1 Ashley Road, Altrincham, Cheshire, WA14 2DT. |
| SO003 | UK Companies House | INEFFABLE INTELLIGENCE LTD — Officers (16865241) | 5 officers listed: Oakwood Corporate Secretary Limited (secretary), Alfred Lin (director, appointed 17 April 2026), Ravi Mhatre (director, appointed 17 April 2026), George Samuel Rose (director), David Silver (director, appointed 16 January 2026). |
| SO004 | CNBC | Former Google DeepMind researcher's AI startup raises record $1.1 billion seed funding to pursue superintelligence | The startup is pursuing superintelligence and was founded in late 2025 by UCL professor and former lead of DeepMind's reinforcement learning team, David Silver. The seed round is the largest ever in Europe, amounting to a valuation of $5.1 billion. |
| SO005 | TechCrunch | DeepMind's David Silver just raised $1.1B to build an AI that learns without human data | Silver also told Wired that "any money that I make from Ineffable will go to high-impact charities that save as many lives as possible." |
| SO006 | SiliconAngle | Ineffable Intelligence raises $1.1B at $5.1B valuation to build an AI 'superlearner' | |
| SO007 | EU-Startups | Ineffable Intelligence lands historic €937M ($1.1B) seed round at $5.1B valuation | London-based Ineffable Intelligence has come out of stealth with €937 million ($1.1 billion) in Seed funding at a €4.3 billion ($5.1 billion) post-money valuation — Europe's largest Seed financing to date. Participating investors included Wellcome Trust. |
| SO008 | Unite.AI | Ineffable Intelligence Closes $1.1B Seed at $5.1B Valuation | |
| SO009 | TechFundingNews | David Silver's Ineffable Intelligence closes Europe's largest seed at $5.1B valuation from Sequoia, Lightspeed | |
| SO010 | UK Government (DSIT) | UK backs company building breakthrough AI that can discover new knowledge | "This investment in Ineffable will support a company at the very frontier of AI, with the potential to transform entire sectors, underlining our determination to ensure that the UK isn't just an AI taker but an AI maker." — Liz Kendall, Science and Technology Secretary. |
| SO011 | British Business Bank | British Business Bank and Sovereign AI invest in AI superintelligence company Ineffable Intelligence | The British Business Bank has invested $20m in Ineffable Intelligence as part of a $1.1bn seed funding round. |
| SO012 | UCL (University College London) | UCL researchers lead two of Europe's largest-ever AI funding rounds | Ineffable Intelligence and Recursive Superintelligence are independent companies. Neither is a UCL spinout. Both were founded by individuals with current academic and research affiliations at UCL. |
| SO013 | NVIDIA (blogs.nvidia.com) | Ineffable Intelligence and NVIDIA Partner on Reinforcement Learning Infrastructure | "The next frontier of AI is superlearners — systems that learn continuously from experience. We are thrilled to partner with Ineffable Intelligence to codesign the infrastructure for large-scale reinforcement learning." — Jensen Huang. |
| SO014 | Cooley LLP | Ineffable Intelligence Announces $1.1 Billion Seed Financing | Cooley advised Ineffable Intelligence on its $1.1 billion seed financing led by Sequoia Capital and Lightspeed Venture Partners at a $5.1 billion post-money valuation, Europe's largest ever seed financing to date. |
| SO015 | BusinessCloud | £1.1 billion: UK's Ineffable Intelligence raises largest European seed round in history | |
| SO016 | Google Cloud Press Corner | Ineffable Intelligence Selects Google Cloud To Power Its Superintelligence Mission | Ineffable Intelligence has selected Google Cloud as its preferred cloud partner, utilizing Google's world-class AI-optimized technology to advance the next frontier of artificial intelligence. Ineffable Intelligence will deploy one of the largest clusters of A5X, powered by the NVIDIA Vera Rubin NVL72 on Google Cloud. |
| SO017 | Sesamers | Ineffable Intelligence $1.1B seed round — David Silver's DeepMind alumni team sets European record | Ineffable Intelligence was founded in 2025 by David Silver... He is joined by three further DeepMind alumni: Wojciech Czarnecki, Lasse Espeholt and Junhyuk Oh. |
| SO018 | UK Tech News (UKTN) | Ineffable Intelligence secures £814M ($1.1B) seed round | The Bank has additionally invested $20m (£14.8m) in the emerging tech company. |
| SO019 | Newspage.news | Sovereign AI backs Ineffable Intelligence but experts warn public money at seed stage should buy more than a press release and a minority stake | "Public money at seed stage should buy more than a press release and a minority stake. If the Government is backing frontier AI alongside major international venture capital, the minimum conditions should include clear governance rights, transparency on downstream use and credible safeguards around where the benefits, control and risks actually land." — Rohit Parmar-Mistry, Founder at Pattrn Data. |
| SO020 | Electronics Weekly | UK government backs $5bn AI startup Ineffable Intelligence | |
| SO021 | Intelligent CIO | British Business Bank backs Ineffable Intelligence in US$1.1bn AI funding round | |
| SO022 | Hotminute | AlphaGo architect David Silver raises £880M for Ineffable in UK record | Silver has committed to donating 100% of any personal proceeds from his equity in Ineffable Intelligence to charity via Founders Pledge. |
| SO023 | Ineffable Intelligence (official blog) | NVIDIA + Ineffable Intelligence: Building the Future of Reinforcement Learning Infrastructure | Together, we are building the reinforcement learning infrastructure that unlocks new levels of intelligence. Engineers from both companies have teamed up to explore the best way to create this training pipeline. |
| SO024 | Founders Pledge | Founders Pledge — About | |
| SO025 | UK Companies House | INEFFABLE INTELLIGENCE LTD — Filing History (16865241) | |
| SO026 | UK Companies House | Companies House Search — Ineffable Intelligence | |
| SO027 | tech.eu | Ineffable Intelligence closes Europe's largest seed at $5.1B valuation from Sequoia, Lightspeed | |
| SM001 | Stanford HAI (Human-Centered AI) | AI Index | Stanford HAI | |
| SM002 | European Commission — Digital Strategy | AI Act | The AI Act puts in place rules for providers of such models [GPAI]. This includes transparency and copyright-related rules. For models that may carry systemic risks, providers should assess and mitigate these risks. The AI Act rules on GPAI became effective in August 2025. |
| SM003 | Goldman Sachs | AI investment forecast to approach $200 billion globally by 2025 | Business surveys suggest that [AI investment] is likely to start having an investment impact in the second half of this decade, with earlier adoption by larger firms in information and professional, scientific, and technical services. |
| SM004 | Statista | Artificial Intelligence — Worldwide | Market Forecast | |
| SM005 | OECD AI Policy Observatory | Live data from OECD.AI — Investment and Industry | |
| SM006 | NVIDIA | NVIDIA Announces Financial Results for First Quarter Fiscal 2026 | Global demand for NVIDIA's AI infrastructure is incredibly strong. AI inference token generation has surged tenfold in just one year, and as AI agents become mainstream, the demand for AI computing will accelerate. |
| SM007 | Epoch AI | Training compute of frontier AI models grows by 4–5x per year | We tentatively conclude that compute growth in recent years is currently best described as increasing by a factor of 4–5x/year. AlphaGo Master and AlphaGo Zero are compute outliers — they single-handedly warp the trend of compute in frontier models. |
| SM008 | arXiv (Kimi Team) | Kimi k1.5: Scaling Reinforcement Learning with LLMs | Scaling reinforcement learning (RL) unlocks a new axis for the continued improvement of artificial intelligence, with the promise that large language models can scale their training data by learning to explore with rewards. |
| SM009 | Nature | Highly accurate protein structure prediction with AlphaFold | AlphaFold structures had a median backbone accuracy of 0.96 Å r.m.s.d.95, demonstrating accuracy competitive with experimental structures in a majority of cases and greatly outperforming other methods. |
| SM010 | UK Government (DSIT) | AI Opportunities Action Plan | |
| SM011 | NVIDIA | NVIDIA and Ineffable Intelligence Team Up to Build the Future of Reinforcement Learning Infrastructure | |
| SM012 | Ineffable Intelligence | NVIDIA and Ineffable Intelligence Team Up to Build the Future of Reinforcement Learning Infrastructure | |
| SM013 | TechCrunch | DeepMind's David Silver just raised $1.1B to build an AI that learns without human data | |
| SM014 | CNBC | DeepMind vet raises $1.1B to build 'AI superlearner' at record seed valuation | |
| SM015 | SiliconAngle | Ineffable Intelligence raises $1.1B at $5.1B valuation to build AI superlearner | |
| SM016 | British Business Bank | British Business Bank and Sovereign AI invest in AI superintelligence company Ineffable Intelligence | |
| SM017 | UK Government | UK backs company building breakthrough AI that can discover new knowledge | |
| SM018 | UCL News | UCL researchers lead two of Europe's largest ever AI funding rounds | |
| SM019 | EU Startups | Ineffable Intelligence lands historic €937M seed round at €4.3B valuation | |
| SM020 | UK Tech News | Ineffable Intelligence secures £814m seed round | |
| SM021 | TechFunding News | David Silver's Ineffable Intelligence raises $1.1BN seed: Europe's largest | |
| SM022 | Newspage | Sovereign AI backs Ineffable Intelligence but experts warn public money at seed stage should buy more than a press release and a minority stake | Sovereignty is not achieved through presence on a cap table; UK public capital should come with enforceable conditions — anchoring of capability, independent safety evaluation, transparency on outputs, and defined rights over downstream use. |
| SM023 | Electronics Weekly | UK government backs $5bn startup | |
| SM024 | Google Cloud Press Corner | Ineffable Intelligence Selects Google Cloud To Power Its Superintelligence Mission | |
| SM025 | UK Government | Frontier AI Safety Commitments, AI Seoul Summit 2024 | |
| SM026 | Hotminute | AlphaGo architect David Silver raises £880 million for Ineffable in UK record | |
| SM027 | NVIDIA News | NVIDIA News — AI for Science, Sovereign Infrastructure, and EU AI Supercomputers (June 2026) | Europe Unveils a Record 35 New NVIDIA AI Supercomputers; NVIDIA Vera Rubin Delivers World-Class Supercomputers for Science; NAIRR Science Program expands AI for research. JUPITER shows what exascale science looks like. |
| SP001 | OpenAI | About OpenAI — Mission, structure, and team | "OpenAI consists of the nonprofit OpenAI Foundation and the for-profit OpenAI Group. The Foundation governs the Group, which operates as a public benefit corporation." |
| SP002 | OpenAI | Planning for AGI and beyond | "Successfully transitioning to a world with superintelligence is perhaps the most important—and hopeful, and scary—project in human history." |
| SP003 | OpenAI | OpenAI API Pricing — model pricing overview | |
| SP004 | OpenAI | Safety — OpenAI safety overview and system cards | |
| SP005 | Anthropic | Anthropic — Company overview and mission | |
| SP006 | Anthropic | Anthropic API Pricing — Claude model tiers | |
| SP007 | Anthropic | Core Views on AI Safety — Anthropic's published safety stance | "We believe we may be building one of the most transformative and potentially dangerous technologies in history, yet we press forward anyway." |
| SP008 | Google DeepMind | Google DeepMind — About page and history | "As Google DeepMind, our world-class talent is harnessing our unparalleled computing infrastructure to create the next wave of research breakthroughs and transformative products." |
| SP009 | SiliconAngle | OpenAI completes $6.6B funding round at $157B valuation | |
| SP010 | SiliconAngle | French AI startup Mistral raises $645M in funding | |
| SP011 | SiliconAngle | Report: Anthropic raising new money at $60B valuation | |
| SP012 | SiliconAngle | Elon Musk's xAI raises $6B funding round valuing company at $80B | |
| SP013 | CNBC | Anthropic raising money at a $60 billion valuation | |
| SP014 | Mistral AI | Le Chat — Mistral AI enterprise and consumer assistant | |
| SP015 | OpenAI | Learning to reason with LLMs — OpenAI o1 release | |
| SP016 | OpenAI | OpenAI o3-mini release and capabilities | |
| SP017 | OpenAI | OpenAI Research — overview of research agenda and teams | |
| SP018 | TechCrunch | DeepMind's David Silver just raised $1.1B to build an AI that learns without human data | |
| SP019 | SiliconAngle | Ineffable Intelligence raises $1.1B at $5.1B valuation to build AI superlearner | |
| SP020 | CNBC | DeepMind vet raises $1.1B to build 'AI superlearner' at record seed valuation | |
| SP021 | Newspage | Sovereign AI backs Ineffable Intelligence but experts warn public money at seed stage should buy more than a press release and a minority stake | Experts warn that UK government's minority stake provides no enforceable IP-anchoring or sovereignty conditions — public capital at seed stage may buy no more than a press release and a minority stake. |
| SP022 | NVIDIA | NVIDIA and Ineffable Intelligence Team Up to Build the Future of Reinforcement Learning Infrastructure | |
| SP023 | Google Cloud Press Corner | Ineffable Intelligence Selects Google Cloud To Power Its Superintelligence Mission | |
| SP024 | Epoch AI | Training compute of frontier AI models grows by 4–5x per year | |
| SP025 | UK Government | Frontier AI Safety Commitments, AI Seoul Summit 2024 | |
| SP026 | Stanford HAI | AI Index Report — Stanford Human-Centred AI annual report | |
| SI001 | Google Cloud | VM Instance Pricing — Google Cloud Compute Engine | Accelerator-optimized: Ideal for massively parallelized CUDA compute workloads, such as machine learning and high performance computing. |
| SI002 | Epoch AI | How much does it cost to train frontier AI models? | The amortized hardware and energy cost for the final training run of frontier models has grown rapidly, at a rate of 2.4x per year since 2016. |
| SI003 | UK Department for Science, Innovation and Technology | DSIT Annual Report and Accounts 2023 to 2024 | |
| SI004 | UK Tech News (UKTN) | Ineffable Intelligence and Google Cloud partner on new frontier AI lab | Experience-based learning places different demands on computing infrastructure than training on static datasets, requiring enormous computational scale, high-performance networking, and tightly integrated training and inference systems. |
| SI005 | arXiv / International Joint Conference on Neural Networks | Compute Trends Across Three Eras of Machine Learning | Since the advent of Deep Learning in the early 2010s, the scaling of training compute has accelerated, doubling approximately every 6 months. |
| SI006 | Sequoia Capital | Generative AI's Act Two | We found ourselves in an unsustainable feeding frenzy of fundraising, talent wars and GPU procurement... a lot of AI companies simply do not have product-market fit or a sustainable competitive advantage. |
| SI007 | NVIDIA Corporation | NVIDIA Announces Financial Results for Second Quarter Fiscal 2026 | NVIDIA Q2 FY2026 revenue was $46.7 billion, up 56% year-on-year; Blackwell Data Center revenue grew 17% sequentially. |
| SI008 | UK Department for Science, Innovation and Technology | AI Growth Zones (open for applications) | AI Growth Zones will unlock investment in AI-enabled data centres and support infrastructure by improving access to power and providing planning support. |
| SI009 | Companies House (UK) | INEFFABLE INTELLIGENCE LTD Filing History | Statement of capital following an allotment of shares on 7 April 2026. |
| SI010 | Ineffable Intelligence | Ineffable Intelligence — Company Website | A window where ambitious research can thrive, without bending to the demands of incremental products and near-term profits. |
| SI011 | Google Cloud Press Corner | Ineffable Intelligence Selects Google Cloud To Power Its Superintelligence Mission | |
| SI012 | NVIDIA Corporate Blog | NVIDIA, Ineffable Intelligence Team Up to Build the Future of Reinforcement Learning Infrastructure | We are thrilled to partner with Ineffable Intelligence to codesign the infrastructure for large-scale reinforcement learning as they push the frontier of AI and pioneer a new generation of intelligent systems. |
| SI013 | Ineffable Intelligence | NVIDIA, Ineffable Intelligence Team Up to Build the Future of Reinforcement Learning Infrastructure (Ineffable Blog) | The system has to act, observe, score and update continuously in tight loops, which puts pressure on interconnect, memory bandwidth and serving in ways that pretraining doesn't. |
| SI014 | British Business Bank | British Business Bank and Sovereign AI invest in AI superintelligence company Ineffable Intelligence | |
| SI015 | CNBC | DeepMind's David Silver raises $1.1B for AI startup Ineffable Intelligence | |
| SI016 | TechCrunch | DeepMind's David Silver just raised $1.1B to build an AI that learns without human data | |
| SI017 | SiliconAngle | Ineffable Intelligence raises $1.1B at $5.1B valuation to build an AI superlearner | |
| SI018 | Newspage.news | Sovereign AI backs Ineffable Intelligence but experts warn public money at seed stage should buy more than a press release and a minority stake | Sovereignty is not achieved through presence on a cap table — it is secured through control, leverage and accountability. |
| SI019 | Electronics Weekly | UK government backs £5bn startup | |
| SI020 | UK Government (DSIT) | UK backs company building breakthrough AI that can discover new knowledge | |
| SI021 | HotMinute | AlphaGo architect David Silver raises £880 million for Ineffable in UK record | |
| SI022 | NVIDIA Corporation | NVIDIA Announces Financial Results for First Quarter Fiscal 2026 | |
| SI023 | EU-Startups | Ineffable Intelligence lands historic $1.1 billion seed round at $5.1 billion valuation | |
| SI024 | University College London | UCL researchers lead two of Europe's largest ever AI funding rounds | |
| SI025 | Cooley LLP | Ineffable Intelligence Announces $1.1 Billion Seed Financing | |
| SE001 | Ineffable Intelligence | Ineffable Intelligence — Official Website and Mission Statement | "A place where the deep question of intelligence is faced head on: how to discover new knowledge from experience in the environment." |
| SE002 | NVIDIA | NVIDIA Blog — Ineffable Intelligence and NVIDIA: Building the Future of Reinforcement Learning Infrastructure | "The system has to act, observe, score and update continuously in tight loops, which puts pressure on interconnect, memory bandwidth and serving in ways that pretraining doesn't." |
| SE003 | Ineffable Intelligence | Ineffable Blog — NVIDIA, Ineffable Intelligence Team Up to Build the Future of Reinforcement Learning Infrastructure | "The system will train on rich forms of experience that are quite distinct from human language and other human data, and may require novel model architectures and training algorithms." |
| SE004 | Google Cloud | Ineffable Intelligence Selects Google Cloud To Power Its Superintelligence Mission | "Ineffable Intelligence will utilize Google Cloud's high-performance computing capabilities to accelerate its mission of developing a 'superlearner'. This partnership will also see Ineffable Intelligence deploy one of the largest clusters of A5X, powered by the NVIDIA Vera Rubin NVL72 on Google Cloud." |
| SE005 | TechCrunch | DeepMind's David Silver just raised $1.1B to build an AI that learns without human data | "Ineffable Intelligence plans to skip that step [pretraining]. Additionally, it will place its AI models in simulations that will enable them to learn from one another." |
| SE006 | UKTN | Ineffable Intelligence and Google Cloud partner on new frontier AI lab | "We evaluated the space and chose Google Cloud as the best fit for our reinforcement learning infrastructure. We aren't just looking for processors; we are building a resilient and scalable environment to make 'first contact' with superintelligence." |
| SE007 | Wired | David Silver on Ineffable Intelligence, Reinforcement Learning, and the Superlearner | "Human data is like a kind of fossil fuel that has provided an amazing shortcut. You can think of systems that learn for themselves as a renewable fuel — something that can just learn and learn and learn forever, without limit." |
| SE008 | NVIDIA | NVIDIA Vera Rubin Platform — GTC Announcement | "Reinforcement learning and agentic AI workloads rely on large numbers of CPU-based environments to test and validate the results generated by models running on GPU systems. The NVIDIA Vera CPU Rack delivers dense, liquid-cooled infrastructure built on NVIDIA MGX, integrating 256 Vera CPUs." |
| SE009 | NVIDIA | NVIDIA Blackwell Platform Arrives to Power a New Era of Computing | "Powering a new era of computing, NVIDIA today announced that the NVIDIA Blackwell platform has arrived — enabling organizations everywhere to build and run real-time generative AI on trillion-parameter large language models at up to 25x less cost and energy consumption than its predecessor." |
| SE010 | arXiv (Espeholt, Soyer, Munos, Simonyan et al.) | IMPALA: Scalable Distributed Deep-RL with Importance Weighted Actor-Learner Architectures | IMPALA — Importance Weighted Actor-Learner Architectures — enables distributed deep RL at scale. Lead author Lasse Espeholt is a co-founding team member of Ineffable Intelligence. |
| SE011 | arXiv (Mnih, Badia, Mirza, Graves, Lillicrap, Harley, Silver, Kavukcuoglu) | Asynchronous Methods for Deep Reinforcement Learning | A3C establishes asynchronous parallel RL as a scalable training paradigm; David Silver is a co-author and the approach underpins distributed RL infrastructure of the kind Ineffable is building. |
| SE012 | arXiv (Adaptive Agent Team, Google DeepMind) | Human-Timescale Adaptation in an Open-Ended Task Space | AdA demonstrates large-scale multi-task RL agents that adapt across hundreds of tasks with human-timescale in-context learning — the closest published predecessor to Ineffable's superlearner concept. |
| SE013 | Sovereign AI | Sovereign AI Invests in Ineffable Intelligence | "Very few founders in the world could credibly set out to build a super learner… David is one of them. From AlphaGo to AlphaZero to Alpha Proof, he has spent nearly two decades turning reinforcement learning from a research idea into the results the rest of the field builds on." |
| SE014 | Anthropic | Responsible Scaling Policy — Current and Prior Versions | RSP v3.3 (effective May 26, 2026) sets governance thresholds for capability progression; Ineffable has published no equivalent framework. |
| SE015 | David Silver (personal) | David Silver — About and Research Overview | "I build AI that learns for itself to solve problems that humans can't. I am the CEO of Ineffable Intelligence. Until recently, I led the reinforcement learning team at DeepMind." |
| SE016 | David Silver (personal) | David Silver — Publications List | Publications include: "Discovering state-of-the-art reinforcement learning algorithms" (Nature 2025); "Olympiad-level formal mathematical reasoning with reinforcement learning" (Nature 2025); "Welcome to the Era of Experience" (Silver, Sutton). |
| SE017 | SiliconANGLE | Ineffable Intelligence raises $1.1B at $5.1B valuation to build AI superlearner | |
| SE018 | UK Government (DSIT) | Frontier AI Safety Commitments — AI Seoul Summit 2024 | |
| SE019 | UK Government (DSIT) | AI Regulation — A Pro-Innovation Approach (White Paper) | |
| SE020 | Epoch AI | How Much Does It Cost to Train Frontier AI Models? | |
| SE021 | Epoch AI | Training Compute of Frontier AI Models Grows by 4–5x per Year | |
| SE022 | CNBC | DeepMind co-founder David Silver launches Ineffable Intelligence with record $1.1B seed | |
| SE023 | British Business Bank | British Business Bank and Sovereign AI Invest in Ineffable Intelligence | |
| SE024 | Google DeepMind | AlphaGo — Inventing Winning Moves | |
| SE025 | UK Government (DSIT) | UK Backs Company Building Breakthrough AI That Can Discover New Knowledge | |
| SU001 | Sovereign AI (UK Government Venture Programme) | Sovereign AI is backing Ineffable Intelligence | Alongside capital, Sovereign AI is supporting startups, including Ineffable, with access to the UK's largest AI supercomputers, visas, and the unique levers of the British state. |
| SU002 | UKTech News | Ineffable Intelligence and Google Cloud enter strategic partnership for frontier AI lab | |
| SU003 | Ineffable Intelligence (via Ashby HQ) | Ineffable Intelligence Jobs | |
| SU004 | OECD | Artificial Intelligence: Key Policy Insights and Data | More than one-third of individuals across the OECD used generative AI tools in 2025, highlighting how rapidly AI is becoming part of everyday life. |
| SU005 | Artificial Intelligence News | Ineffable Intelligence and Google Cloud partner to develop superlearner AI | |
| SU006 | CNBC | Ineffable Intelligence selects Google Cloud as preferred infrastructure partner | |
| SU007 | SiliconAngle | Ineffable Intelligence teams up with Google Cloud to advance superlearner research | |
| SU008 | TechFunding News | Ineffable Intelligence and Google Cloud partner for superlearner AI infrastructure | |
| SU009 | DeepMind (Google) | Artificial General Intelligence Research at DeepMind | |
| SU010 | NewsPage | Sovereign AI backs Ineffable Intelligence but experts warn public money at seed stage should buy more than a press release and a minority stake | Public money at seed stage should buy more than a press release and a minority stake. If the Government is backing frontier AI alongside major international venture capital, the minimum conditions should include clear governance rights, transparency on downstream use and credible safeguards around where the benefits, control and risks actually land. |
| SU011 | Google Cloud | Ineffable Intelligence Selects Google Cloud to Power Its Superintelligence Mission | Ineffable Intelligence will utilize Google Cloud's high-performance computing capabilities to accelerate its mission of developing a 'superlearner.' |
| SU012 | Ineffable Intelligence | NVIDIA and Ineffable Intelligence Team Up to Build the Future of Reinforcement Learning Infrastructure | The system has to act, observe, score and update continuously in tight loops, which puts pressure on interconnect, memory bandwidth and serving in ways that pretraining doesn't. |
| SU013 | British Business Bank | British Business Bank and Sovereign AI invest in AI superintelligence company Ineffable Intelligence | The British Business Bank has invested $20m in Ineffable Intelligence, the UK-headquartered AI superintelligence company, as part of a $1.1bn seed funding round. |
| SU014 | UK Government (DSIT) | UK backs company building breakthrough AI that can discover new knowledge | This investment in Ineffable will support a company at the very frontier of AI, with the potential to transform entire sectors. |
| SU015 | CNBC | DeepMind's AlphaGo creator raises record $1.1 billion to build AI that learns without human data | |
| SU016 | TechCrunch | DeepMind's David Silver just raised $1.1B to build an AI that learns without human data | |
| SU017 | SiliconAngle | Ineffable Intelligence raises $1.1B at $5.1B valuation to build AI superlearner | |
| SU018 | Ineffable Intelligence | Ineffable Intelligence — Company Homepage | |
| SU019 | NVIDIA (Blogs) | Ineffable Intelligence and NVIDIA Partner to Advance Reinforcement Learning | The next frontier of AI is superlearners — systems that learn continuously from experience. We are thrilled to partner with Ineffable Intelligence to codesign the infrastructure for large-scale reinforcement learning. |
| SU020 | UKTech News | Ineffable Intelligence secures $814m seed round | |
| SU021 | Tech.eu | Ineffable Intelligence closes Europe's largest seed at $5.1B valuation from Sequoia, Lightspeed | |
| SU022 | Electronics Weekly | UK Government backs $5bn startup building frontier AI | |
| SU023 | Intelligent CIO | British Business Bank backs Ineffable Intelligence in US$1.1bn AI funding round | |
| SU024 | EU Startups | Ineffable Intelligence lands historic $1.1 billion seed round at $5.1 billion valuation | |
| SU025 | Sesamers | Ineffable Intelligence $1.1B seed round — David Silver, DeepMind, European record | |
| SU026 | Hot Minute | AlphaGo architect David Silver raises $880 million for Ineffable in UK record | |
| SR001 | Ineffable Intelligence | Ineffable Intelligence — Official Website and Mission Statement | We are creating a superlearner that discovers all knowledge from its own experience, from elementary motor skills through to profound intellectual breakthroughs. |
| SR002 | UK Government / DSIT | UK Government Backs Company Building Breakthrough AI That Can Discover New Knowledge | The UK Government is backing Ineffable Intelligence through the British Business Bank and the UK Sovereign AI Fund. |
| SR003 | CNBC | DeepMind co-founder raises $1.1 billion in record seed round for Ineffable Intelligence | Ineffable Intelligence has raised $1.1 billion at a post-money valuation of $5.1 billion, co-led by Sequoia and Lightspeed, with NVIDIA and Google among participants. |
| SR004 | Newspage / Expert Commentators | Sovereign AI backs Ineffable Intelligence but experts warn public money at seed stage should buy more than a press release and a minority stake | Experts warn that public money at seed stage should buy more than a press release and a minority stake — governance rights, IP protections, and public benefit covenants are absent. |
| SR005 | Electronics Weekly | UK Government Backs £5bn Startup | Questions have been raised about whether the UK government's minority stake provides any meaningful sovereignty over the AI being developed. |
| SR006 | British Business Bank | British Business Bank and Sovereign AI Invest in Ineffable Intelligence | The British Business Bank and the UK Sovereign AI Fund have today announced a combined investment in Ineffable Intelligence. |
| SR007 | NVIDIA | NVIDIA Blog — Ineffable Intelligence and NVIDIA: Building the Future of Reinforcement Learning Infrastructure | The system has to act, observe, score and update continuously in tight loops, which puts pressure on interconnect, memory bandwidth and serving in ways that pretraining does not. |
| SR008 | Google Cloud | Ineffable Intelligence Selects Google Cloud to Power Its Superintelligence Mission | Ineffable Intelligence has selected Google Cloud as its preferred cloud provider to deploy its superlearner on the AI Hypercomputer platform. |
| SR009 | UK Government / DSIT | AI Regulation — A Pro-Innovation Approach (White Paper) | The government has decided not to introduce a single, central AI regulator or introduce new primary legislation at this stage. |
| SR010 | UK Government | Frontier AI Safety Commitments — AI Seoul Summit 2024 | Frontier AI developers have committed to share information with governments, conduct safety evaluations, and not deploy models deemed too dangerous. |
| SR011 | TechCrunch | DeepMind's David Silver just raised $1.1B to build an AI that learns without human data | Silver's Ineffable Intelligence has raised $1.1 billion to build an AI superlearner that will discover all knowledge from its own experience. |
| SR012 | Wired | David Silver Is Building an AI That Learns Without Human Data | Silver wants to build an AI that learns from its own experiences, not from the vast troves of human-generated data that underpin today's large language models. |
| SR013 | UK Sovereign AI Fund | Sovereign AI Invests in Ineffable Intelligence | The UK Sovereign AI Fund has made an investment in Ineffable Intelligence as part of the government's commitment to AI leadership. |
| SR014 | Sequoia Capital | Generative AI: Act Two | There is a dangerous gap between AI revenue and AI spending; companies burning billions on compute without product market fit face existential risk. |
| SR015 | Sequoia Capital | AI Scorecard | The gap between compute spending and AI revenue has not narrowed at the pace investors initially hoped; capital discipline is increasingly important. |
| SR016 | The Register | Ineffable Intelligence raises $1 billion seed round — eyebrows raised | Not everyone is convinced that a $5.1 billion valuation for a pre-product company with a single named founder is justified by the technical thesis alone. |
| SR017 | Wired | AI Frontier Labs Are Spending Billions on Compute | Frontier AI labs are spending at a rate that raises fundamental questions about whether any of them can reach commercial scale before their capital runs out. |
| SR018 | MIT Technology Review | RL Scaling at the Frontier: How Far Can It Go? | RL has shown remarkable results in bounded domains but faces unresolved challenges when applied to open-ended knowledge discovery tasks. |
| SR019 | MIT Technology Review | Reinforcement Learning from Human Feedback Is a Mess | RL reward functions are notoriously difficult to specify correctly; systems routinely find unexpected ways to maximise reward that were not intended by designers. |
| SR020 | MIT Technology Review | Ineffable Intelligence and David Silver — The AGI Competition | Silver's bet is that reinforcement learning without human data can generalise beyond game-playing — a claim that remains to be proven in open-ended environments. |
| SR021 | Epoch AI | The Current State of AI Safety | Most frontier AI labs have now published responsible scaling policies; the absence of safety governance frameworks at pre-product labs is a growing concern among the AI safety community. |
| SR022 | UK Government | National AI Strategy | The UK government will position the UK as a global leader in AI, investing in compute infrastructure and establishing appropriate governance frameworks. |
| SR023 | UK Government | National Security and Investment Act — Government Collection | The National Security and Investment Act 2021 gives the UK government powers to scrutinise and intervene in acquisitions in 17 sensitive sectors including advanced AI. |
| SR024 | Information Commissioner's Office (ICO) | Guidance on AI and Data Protection | Organisations developing or deploying AI systems must apply data protection by design, ensure fairness in automated processing, and maintain robust accountability and governance frameworks under UK GDPR. |
| SR025 | Sequoia Capital | AI Capital Allocation 2026 | The frontier AI labs that will survive the next five years are those that achieve a demonstrable commercial return on their compute investment before capital markets lose patience. |
| SR026 | Wired | OpenAI, Anthropic, and Google DeepMind — The AGI Race in 2025 | The competition for frontier RL talent is intense; DeepMind, OpenAI, and Anthropic are all bidding aggressively for researchers with the skills required to advance general intelligence. |
| SR027 | BBC News | AI Start-Up Raises $1.1 Billion at $5.1 Billion Valuation | Ineffable Intelligence has raised a record seed round from investors including Sequoia, NVIDIA, and the UK government. |
| SR028 | Financial Times | Ineffable Intelligence — AI Investment | |
| SR029 | Wired | AI Training Costs at the Compute Frontier | Frontier model training runs now routinely cost hundreds of millions of dollars, and RL at frontier scale is more expensive per effective FLOP than supervised pre-training. |
| SR030 | MIT Technology Review | Reinforcement Learning Scaling Limits | There is growing evidence that RL scaling faces fundamental limits in open-ended environments that are qualitatively different from the scaling laws seen in supervised language modelling. |
| SR031 | Ineffable Intelligence (NVIDIA blog) | Ineffable Intelligence and NVIDIA Team Up to Build the Future of Reinforcement Learning Infrastructure | Working closely with NVIDIA engineering teams to co-design the infrastructure for a new class of RL system. |
| SV001 | Sovereign AI UK | Sovereign AI invests in Ineffable Intelligence | Very few founders in the world could credibly set out to build a super learner — a system that discovers new knowledge from its own experience, rather than ours. David is one of them. |
| SV002 | Stanford HAI | AI Index 2025 — Stanford Human-Centered AI | |
| SV003 | MIT Technology Review | We need to talk about AI and energy — what experts say | The energy cost of AI training is not a static parameter — it scales with model ambition, and the most ambitious frontier labs face compounding energy and infrastructure cost pressures. |
| SV004 | xAI | xAI — Company: Accelerating Scientific Discovery | |
| SV005 | SiliconAngle | Ineffable Intelligence teams up with Google Cloud to advance its superlearner AI research | |
| SV006 | Google Cloud | Ineffable Intelligence selects Google Cloud to power its superintelligence mission | |
| SV007 | Anthropic | Anthropic Research — Overview | |
| SV008 | NVIDIA Investor Relations | NVIDIA Announces Financial Results for Fourth Quarter and Fiscal Year 2026 | |
| SV009 | Mistral AI | Mistral AI — Pricing | |
| SV010 | Lightspeed Venture Partners | Lightspeed's investment in Ineffable Intelligence | David Silver has spent nearly two decades turning reinforcement learning from a research idea into the results the rest of the field builds on. Very few founders in the world could credibly set out to build a superlearner. |
| SV011 | SiliconAngle | Ineffable Intelligence raises $1.1B, valued at $5.1B, to build an AI superlearner | |
| SV012 | TechCrunch | DeepMind's David Silver just raised $1.1B to build an AI that learns without human data | |
| SV013 | CNBC | DeepMind co-creator David Silver raises $1.1B for Ineffable Intelligence AI startup | |
| SV014 | SiliconAngle | OpenAI completes $6.6B funding round at a $157B valuation | |
| SV015 | SiliconAngle | Report: Anthropic raising new money at $60B valuation | |
| SV016 | SiliconAngle | Elon Musk's xAI raises $6B funding round valuing company at $80B | |
| SV017 | Sequoia Capital | Generative AI's Act Two | A lot of AI companies simply do not have product-market fit or a sustainable competitive advantage, and the overall ebullience of the AI ecosystem is unsustainable. |
| SV018 | NewsPage | Sovereign AI backs Ineffable Intelligence but experts warn public money at seed stage should buy more than a press release and a minority stake | Public money at seed stage should buy more than a press release and a minority stake — experts question whether the UK government's position secures meaningful sovereignty over Ineffable's discoveries. |
| SV019 | Companies House | Ineffable Intelligence Ltd — Company Filing 16865241 | |
| SV020 | Epoch AI | Training compute of frontier AI models grows by 4–5× per year | |
| SV021 | Ineffable Intelligence | Ineffable Intelligence — Official Website | |
| SV022 | SiliconAngle | French AI startup Mistral raises €645M in funding at €6B valuation | |
| SV023 | Sequoia Capital | AI Capital Allocation 2026 | |
| SV024 | British Business Bank | British Business Bank and Sovereign AI invest in Ineffable Intelligence | |
| SV025 | MIT Technology Review | The limits of reinforcement learning scaling at the frontier | Reinforcement learning in open-ended environments faces fundamental barriers that pure scaling of compute and data does not resolve — reward specification, distributional shift, and catastrophic forgetting remain open research problems. |
| SV026 | Stanford AI Index | AI Index Report — Stanford HAI | |
| SV027 | Goldman Sachs | AI investment forecast to approach $200 billion globally by 2025 | |
| SV028 | Wired | David Silver left DeepMind. Now he wants to build an AI that can think for itself. | |
| SV029 | Epoch AI | How much does it cost to train frontier AI models? | |
| SV030 | CNBC | Anthropic is raising money at a $60 billion valuation |