Midha Launches AMP With $1.3B to Fix Underutilized AI Compute

Midha携13亿美元推出AMP，着力解决AI算力闲置问题 Anjney Midha创立AMP PBC以构建独立的算力网络，按成本价向AI研究实验室输送汇集的算力；他指出价值数十亿美元的闲置GPU是当前市场最大的瓶颈。 Anjney Midha阐述了他对科技投资的战略构想，首先提出了一项反直觉的设想，即GameStop可能收购eBay，并借助其实体零售门店为AI智能体电商提供真伪核验。随后，他介绍了AMP PBC这一独立系统运营商，该机构旨在汇集诸如xAI孟菲斯基地等未被充分利用的算力集群，并按成本价将原始算力提供给前沿研究人员。Midha还概述了AMP Foundry的实操策略，重点提及与Periodic Labs合作通过自动化与机器人实验闭环加速材料科学研发，并详细阐明了AMP的公共利益公司（PBC）架构及其由资深工程技术人员组成的精干团队。 Anjney Midha认为，Ryan Cohen掌舵的GameStop收购eBay在经济上切实可行，因为可通过削减重叠的营销支出（2025财年耗费24亿美元获取100万用户）来抵消债务利息。 (1:00) GameStop约1,600家零售门店构成了一套实体核验网络，这对欺诈问题作为主要瓶颈的高价值收藏品交易及AI智能体电商至关重要。 (3:56) Discord此前尽管GMV突破100亿美元，但因缺乏线下真伪核验的基础设施，最终终止了向实体收藏品领域的扩张。 (3:56) AMP PBC作为算力领域的独立系统运营商应运而生，致力于解决AI硬件集群普遍存在的严重低利用率问题（如xAI孟菲斯集群的MFU仅约11%）。 (7:56) AMP以公共利益公司及资本平台的架构运作，按成本价向投资组合企业输送算力，不赚取任何差价溢尼。 (7:56)

Venture investor Anjney Midha argues that the primary bottleneck in artificial intelligence is shifting from software algorithms to physical execution, specifically compute utilization, laboratory experimentation, and physical verification. To address this, his firm, AMP PBC, is structured as a public benefit utility designed to pool hardware infrastructure and incubate frontier scientific labs. Re-engineering Compute as an Open Utility (7:56) Midha compares the current AI hardware landscape to industrial England in 1885, arguing that technology companies are inefficiently hoarding compute capacity rather than optimizing its distribution. As an example, he points to xAI's Memphis cluster, where an estimated 500,000 GPUs operated at roughly 11% Model Flops Utilization (MFU) and under 60% node allocation, representing $12 billion in underused infrastructure. Midha launched AMP PBC to act as an independent system operator that leases clusters, aggregates capacity, and drives up utilization. To operate the network, AMP hired former Google engineers who previously managed the company's Borg scheduling systems at utilization rates above 95%. Operating as a Public Benefit Corporation (PBC) while transitioning to a Registered Investment Advisor (RIA), the firm provides compute to portfolio companies at cost. Midha states that the PBC governance model is specifically intended to protect management from shareholder lawsuits for passing along hardware without standard commercial markups. Automating Physical Loops in Materials Science (12:25) Contrasting broad software venture investing with deep technical incubation, Midha advocates for a hands-on venture model through AMP Foundry. Its flagship project, Periodic Labs—co-founded with Liam Fedus and Dorje Brody in a 30,000-square-foot facility in Menlo Park—focuses on discovering high-temperature superconductors by pairing AI predictions directly with automated physical synthesis and testing. The laboratory uses an automated closed loop where AI models propose candidate materials, robotic hardware synthesizes them, and X-ray diffraction equipment conducts structural validation. The empirical data is immediately fed back to retrain the models. Midha claims this system completed more physical material verifications in a recent 90-day period than the wider scientific field completed over the preceding decade, asserting that laboratory throughput rather than theory is the gating factor in materials science. Physical Storefronts as Moats for Agentic Commerce (1:00) Applying his thesis on physical bottlenecks to retail commerce, Midha defends Ryan Cohen's proposed GameStop–eBay acquisition by examining operating costs and authentication needs. He highlights eBay's FY25 10-K filing, which showed $2.4 billion in annual marketing expenditures alongside an active user increase of only one million (from 134 million to 135 million)—representing roughly $2,400 per net-new user. Midha argues this marketing expense can be drastically curtailed to service acquisition debt. Beyond cost reduction, Midha contends that physical storefront networks provide a defensible moat against competitors like Amazon in high-value, used collectibles. While autonomous AI shopping agents can locate vintage merchandise, they cannot verify authenticity in the physical world. Drawing on his experience at Discord, where physical marketplace plans were abandoned after a $10 billion digital NFT boom due to a lack of physical verification centers, Midha argues GameStop's roughly 1,600 storefronts could serve as nationwide authentication hubs for agent-driven commerce. Institutional Capital Inflows and Long-Term Adoption (14:21) Midha states that global capital markets have increasingly accepted the predictable scaling laws of modern AI, simplifying fundraising for capital-intensive infrastructure. Operating as a solo General Partner, Midha secured over $1 billion in commitments for AMP's debut fund within eight weeks, aiming to raise up to $6.5 billion more while already deploying $300 million into Anoptic. Addressing market bubble concerns, Midha argues that broader economic adoption of AI remains in an embryonic phase. Because legacy enterprise sectors and regional economies outside major technology centers have barely integrated basic large language model workflows, he expects real-world operational diffusion and productivity transformations to continue independently of venture market cycles. While Midha outlines an aggressive model for treating compute as a utility and physical infrastructure as a moat, AMP's long-term thesis relies on sustaining high operational utilization across leased hardware and proving that automated laboratory loops can generate commercially viable material discoveries.

AMP PBC 创始人 Anjney Midha 认为，当前人工智能发展的核心瓶颈并非模型概念匮乏，而是底层算力利用率低下与现实物理执行能力的严重脱节。他主张将算力视为公用事业电网进行集约调度，依托公共利益公司按成本价输送资源，并结合线下物理防伪网点与自动化实验闭环，以破解 AI 规模化商用与前沿科学发现的执行障碍。 算力电网化调度与公共利益公司架构 (7:56) Midha 将当下的 AI 硬件生态比作 1885 年工业时代的煤炭：各方急于囤积关键算力，却普遍陷入低效运转。他引述数据称，Elon Musk 旗下 xAI 位于孟菲斯的集群拥有约 50 万块 GPU，但模型算力利用率（MFU）仅约 11%，节点分配率低于 60%，意味着约 120 亿美元的算力资产处于闲置状态。AMP PBC 因此定位于算力电网的“独立系统运营商”（ISO），通过长期租赁与容量池化调度提升集群负载率。AMP 工程联创团队曾负责 Google Borg 调度系统，具备将超大规模算力维持在 95% 以上利用率的工程经验。 与加价售卖算力的传统云服务商不同，AMP 采取公共利益公司（PBC）架构，将整合的算力按成本价（at cost）输送给独立科研团队与被投企业，并将管理费与收益分成再投资于算力储备。Midha 指出，PBC 结构能有效防范因未追求算力短期利润最大化而引发的股东诉讼，从而以自律型公用事业的形式，为整个技术生态换取更长期的复合价值。 线下实体验证网络赋能智能体电商 (1:00) 针对市场关于 Ryan Cohen 推动 GameStop 洽购 eBay 的传闻，Midha 提出了不同于华尔街主流的视角。eBay 2025 财年 10-K 报告显示，其单年营销开支达 24 亿美元，但活跃用户仅从 1.34 亿微增至 1.35 亿，折合每新增净用户获客成本高达约 2400 美元。Midha 认为，通过削减这一庞大且低效的冗余营销开支，即可覆盖并购所需的债务利息成本。 更深层的战略协同则在于 AI 智能体电商（Agentic Commerce）的护城河。Midha 指出，亚马逊等标准仓储物流多年无法有效处理非标二手与稀缺藏品；当用户驱动 Claude 等智能体搜寻稀缺商品（如古董眼镜或名笔）时，交易达成的核心瓶颈在于实物真伪验证。GameStop 现存的约 1600 家线下门店恰能提供分布式的物理防伪检测网络。Midha 援引其在 Discord 的经历指出，Discord 在 2021 年 NFT 高峰期虽拥有超 100 亿美元平台交易额（GMV），但因欠缺线下鉴真基础设施，最终被迫放弃球鞋等实物收藏品市场。这证明了实体网点在智能体交易闭环中的独特壁垒。 自动化实验闭环驱动材料科学发现 (12:25) Midha 批评传统风投机构在应用层和泛化大模型领域盲目撒网，主张重回早年高确信度的深度构建模式。AMP 旗下设立 AMP Foundry 部门，深度参与技术原型的从零孵化。典型案例为其与 ChatGPT 联创 Liam Fedus 及前 DeepMind 量子物理负责人 Dorje Brody 联合孵化的 Periodic Labs。该实验室立足门洛帕克 30000 平方英尺的基地，聚焦新型高温超导体的研制，Midha 本人每周三天驻场参与早间晨会以推进执行。 Periodic Labs 建立了一套 AI 驱动的自动化物理实验回路：由 AI 智能体生成新型材料预测，经机器人自动化合成后，即刻由 X 射线衍射（XRD）设备展开物理验证，检测数据随即反哺模型进行迭代训练。Midha 表示，该实验室在 90 天内完成的材料物理验证量超过了该领域过去 10 年的总和。他强调，前沿物理与材料科学的真正短板在于物理验证的执行效率，而非概念构想，顶尖科学家亟需匹配足量的算力、资本与工程化支持。 资本扩张提速与全球真实落地差距 (14:21) 随着 Anthropic 等标杆企业的验证，资本市场已充分领会以算力扩张换取能力的“苦涩的教训”（The Bitter Lesson）。Midha 透露，AMP 成立仅 8 周便以单一关键人 GP 身份获得超 10 亿美元募资承诺，首期基金已锁定逾 13 亿美元承诺，目标年内追加至 65 亿美元，且已向光计算项目 Anoptic 投资 3 亿美元。目前 AMP 正在办理投资顾问（RIA）登记，以便在合规框架下打通一级创投、并购与二级市场。 面对外界关于 AI 行业是否存在泡沫的疑虑，Midha 认为全球采纳度仍处于极早期。尽管他在斯坦福大学合讲的 CS 153 课程吸引了数千名学员，但在科技中心之外，全球绝大多数传统行业和地域甚至尚未将 Claude 或 ChatGPT 等基础前沿工具引入日常业务。Midha 强调，大模型能力的实际生产力渗透才刚刚拉开序幕，市场仍具备跨越周期的巨大发展空间。 Midha 将算力调度、实体验证网络与自动化实验室视为破局 AI 物理限制的核心杠杆。然而，该路线要求巨量资本的持续注入，并依赖各传统产业在现实中逐步接纳前沿大模型；同时，其通过自动化实验室攻克高温超导等前沿科研难题的实际转化效果，仍需经历更长时间尺度的检验。

## Background 该内容由科技播客频道 TBPN 于 2026 年 5 月 22 日发布在 YouTube 平台。视频简介指出，本期为 AMP 负责人 Anjney Midha 做客 TBPN 的完整访谈，深入探讨了算力作为 AI 经济中最受限且最宝贵资源的投资逻辑，阐述了全行业数十亿美元闲置算力导致的巨大低效，并介绍了 AMP 如何通过构建协同“算力电网”以及“基础设施+资本”双轮驱动平台，赋能下一代前沿 AI 与科学突破。 ## TL;DR 访谈开篇，Anjney Midha 独到地拆解了 GameStop 欲收购 eBay 的底层逻辑，指出大刀阔斧削减 20 亿美元无效营销费用的财务托底价值，以及线下零售门店在未来 AI Agent 商务中不可替代的物理防伪闭环作用。随后他将焦点拉回其新创立的公共利益公司（PBC）AMP，将当前 AI 发展阶段比喻为 1885 年工业革命初期的煤炭狂潮，直指当前 GPU 集群利用率极其低下的系统性空耗。AMP 创新性地将风险资本与算力基础设施深度融合，不仅以零加价成本价向 Anthropic 及聚焦超导材料的 Periodic Labs 等前沿实验室直供稀缺算力，还将基金收益循环投入底层扩容。依靠前 Google Borg 核心团队的调度实力与 PBC 架构的长期主义保障，AMP 旨在充当自律型“算力公用事业”，为新一代顶尖科学家提供战略级动能。 ## Atmosphere & style 整场对谈节奏明快凌厉，充满了硬核工程思维与华尔街资本洞察深度碰撞的智性张力。Anjney 的表达极具穿透力且善用精妙的历史隐喻，他与主持人在并购八卦与迷因狂欢中谈笑风生，却在切入算力瓶颈时瞬间切换至严肃的“战时状态”视角。整期节目散发着一种打破传统 VC 慵懒范式的实干锐气，兼具理想主义的情怀与第一性原理落地执行的绝对自信，让听众深刻感受到 AI 基础设施前夜的巨大推力与时代脉搏。 ## Bulletin - [10:57](https://www.youtube.com/watch?v=30NHL5Jg3ek&t=657s) 前 a16z 合伙人 Anjney Midha 创立公共利益公司（PBC）AMP，将算力基础设施与风险投资融合，并以成本价向独立 AI 团队直供 GPU 算力以打破巨头垄断。 - [10:57](https://www.youtube.com/watch?v=30NHL5Jg3ek&t=657s) AMP 首期基金在 8 周内达成逾 13 亿美元承诺出资，年内目标追加筹集约 65 亿美元，并已向 Anthropic 追加注资 3 亿美元。 - [07:30](https://www.youtube.com/watch?v=30NHL5Jg3ek&t=450s) 行业算力资源错配严重，如 xAI 孟菲斯集群的模型算力利用率（MFU）仅 11%，导致高达 120 亿美元的算力资源被闲置空耗。 - [12:23](https://www.youtube.com/watch?v=30NHL5Jg3ek&t=743s) AI 材料科学初创 Periodic Labs 打造了从预测、合成到 XRD 物性检测的自动化流水线，过去 90 天完成的实验验证量已超过该领域过去十年总和。 - [02:00](https://www.youtube.com/watch?v=30NHL5Jg3ek&t=120s) Anjney 深度拆解 GameStop 收购 eBay 传闻，指出 eBay 存在 20 亿美元无效营销费用的削减空间，且 GameStop 线下网点可在 AI Agent 时代充当物理防伪验真中心。 ## Key takeaways - [00:00](https://www.youtube.com/watch?v=30NHL5Jg3ek&t=0s) 从第一性原理审视 GameStop 收购 eBay 传闻，核心在于底层商业模式能否跑通，而非短期的杠杆和融资难题。 - [02:00](https://www.youtube.com/watch?v=30NHL5Jg3ek&t=120s) eBay 2025 财年营销开支高达 24 亿美元却仅净增 100 万活跃用户，大幅削减这笔无效营销费用构成了并购的坚固财务底线。 - [03:56](https://www.youtube.com/watch?v=30NHL5Jg3ek&t=236s) 高价值二手收藏品对亚马逊自动化仓储天然免疫，与 GameStop 线下实体门店结合可建立起 AI Agent 时代的物理防伪验证壁垒。 - [05:36](https://www.youtube.com/watch?v=30NHL5Jg3ek&t=336s) Discord 内部探索实物社区电商因缺少线下验真网络而折戟，反向印证了 eBay 庞大线下交易基础设施的战略稀缺性。 - [07:30](https://www.youtube.com/watch?v=30NHL5Jg3ek&t=450s) 当前 AI 发展阶段宛如 1885 年英国工业革命初期，算力就是稀缺的“煤炭”，而各家分散自建集群造成了极高的闲置与浪费。 - [10:57](https://www.youtube.com/watch?v=30NHL5Jg3ek&t=657s) AMP 采用“基础设施调度 + 风险投资”双轮驱动模式，以零加价成本价向被投科学家团队输送关键 GPU 算力。 - [12:23](https://www.youtube.com/watch?v=30NHL5Jg3ek&t=743s) 华尔街与公开资本已被 Scaling Laws 说服，资本不再稀缺，VC 的核心壁垒已转向调度不可替代的算力与工程落地能力。 - [12:23](https://www.youtube.com/watch?v=30NHL5Jg3ek&t=743s) 重拾经典风投范式与科学家深度共创，Periodic Labs 通过 AI 结合机器人与物理仪器测试，90 天内跑完十年的材料验证量。 - [15:46](https://www.youtube.com/watch?v=30NHL5Jg3ek&t=946s) AMP 核心团队仅 5 人但彼此深度信任，工程联创曾主导 Google Borg 系统常年维持 95% 以上的极限算力利用率。 - [18:09](https://www.youtube.com/watch?v=30NHL5Jg3ek&t=1089s) AMP 注册为公共利益公司（PBC）并申请 RIA 资质，旨在解决算力与技术研发的巨大正外部性，充当自律型公用事业机构。 - [18:09](https://www.youtube.com/watch?v=30NHL5Jg3ek&t=1089s) 以成本价直供算力的策略通过 PBC 架构获得法律保护，避免被传统股东诉讼，从而依靠信任与社区构建穿越周期的长期护城河。 - [22:09](https://www.youtube.com/watch?v=30NHL5Jg3ek&t=1329s) 资本不应采取掠夺性裁员的重组方式，应推动负责任的转型，同时通过斯坦福 CS 153 讲座等平台化解外界对 AI 泡沫的误解。 - [24:41](https://www.youtube.com/watch?v=30NHL5Jg3ek&t=1481s) 现实世界具有极强的业务惯性，全球绝大多数人群尚未真正接触大模型工具，AI 技术的物理落地与扩散仍处于极为早期的阶段。 ## Detailed walkthrough ### [00:00](https://www.youtube.com/watch?v=30NHL5Jg3ek&t=0s) 开场调侃与话题转向：GameStop 欲收购 eBay 的传闻 节目一开始，主持人热情欢迎 Anjney Midha 首次以其新创立的公共利益公司（Public Benefit Corporation，简称 PBC）AMP 的负责人身份做客直播间。主持人笑称 Anjney 正处于“传奇般的上升期”（generational run），Anjney 笑着反驳说这种说法通常是送给准备退休的人的，而他才刚刚起步；不过他也承认，当意识到自己处于极佳轨道时，必须进一步提升自己（level up），否则一旦自满就会停滞不前。 在主持人准备切入正题，询问 AMP 基金的具体启动时间、投资策略以及公司构建愿景时，Anjney 却出人意料地主动打断并建议聊个更有意思的话题——GameStop 及其首席执行官 Ryan Cohen 试图收购 eBay 的轰动传闻。Anjney 表示想和主持人们就此切磋一下观点，因为 Ryan Cohen 最近在 CNBC 的专访引发了铺天盖地的争议，许多人发消息向 Anjney 吐槽这笔交易荒谬透顶。但 Anjney 在仔细翻阅了 GameStop 针对 eBay 的演示文稿（deck）以及 Michael Burry 刚发表的分析文章后，认为外界在过去 48 小时内的讨论几乎完全跑偏了。Michael Burry 确实指出了杠杆极其吃紧，CNBC 主持人也一味质问“钱从哪里来、如何为交易融资”，但 Anjney 认为核心问题根本不是 GameStop 是否买得起 eBay，而是这门底层生意在逻辑上究竟能否跑通。 ### [02:00](https://www.youtube.com/watch?v=30NHL5Jg3ek&t=120s) 20 亿美元营销沉没成本：被华尔街视作固定支出的“肥肉” Anjney 详细拆解了自己看好这笔交易底层逻辑的第一个关键——资产负债表与损益表中的巨大削减空间。他在翻阅 eBay 刚刚发布的 2025 财年 2 月份 10-K 年报时，发现了一个极其惊人的数字：eBay 在过去一年里足足花掉了 24 亿美元的营销费用（marketing expenses）。 然而，这 24 亿美元巨款换来的用户增长却微乎其微——eBay 的年活跃用户基数仅仅从 1.34 亿增加到了 1.35 亿，净增仅区区 100 万人。作为具有市场推广背景的主持人对此也大为震惊。这意味着 eBay 在每一个新增用户身上砸掉了整整 2,400 美元的营销成本，而 eBay 本身却是一个全美妇孺皆知的成熟老牌平台。这种资金流向暴露出极其严重的效率低下，本质上只是在反复花钱拉回那些早已流失的老用户。 在 Anjney 看来，Ryan Cohen 绝非盲目自大到认为自己的产品能力强于 eBay 的原班团队，而是敏锐地捕捉到了这块高达 20 亿美元、长期被华尔街习惯性计入固定成本的“肥肉”。只要大刀阔斧削减这笔无效营销费用，省下的现金流哪怕仅仅用来支付收购债务的利息都能绰绰有余，甚至直接买入美国国债（treasuries）所获得的收益都远高于当前的业务回报。这就是整个交易坚不可摧的“财务下限”（the floor）。 - **eBay 营销支出**：2025 财年投入 24 亿美元。 - **用户增长数据**：用户基数由 1.34 亿微增至 1.35 亿，净增仅 100 万。 - **获客单价（CAC）**：折合每位新增用户耗资约 2,400 美元。 - **交易托底逻辑（Floor）**：砍掉 20 亿美元无效开支以对冲债务成本或投入无风险国债。 ### [03:56](https://www.youtube.com/watch?v=30NHL5Jg3ek&t=236s) 收藏品护城河与线下实体门店：AI Agent 商业时代的物理防伪验证 确立了“财务下限”之后，作为科技投资者的 Anjney 进一步阐述了他眼中的“增长上限”（the ceiling）——收藏品市场与物理网点的协同效应。他指出，亚马逊的二手与收藏品（collectibles）业务已经连续整整 6 年停滞不前，无论是改版、做以旧换新还是推收藏板块，亚马逊统统没能做大规模。其根本原因在于物流基因的不兼容：一张价值连城的 1962 年 Mickey Mantle 稀有棒球卡，绝不可能和普通手机充电线跑在同一套高度自动化的标准化仓库传送带上。收藏品品类对亚马逊构成了结构性的防御壁垒——“亚马逊卖充电线，eBay 卖 Mickey Mantle 棒球卡”。 ```mermaid flowchart TD accTitle: Agentic Commerce Verification Loop accDescr: AI Agent search and physical verification flow through GameStop retail stores subgraph UserLayer["用户与 Agent 交互"] A["用户委托 Claude 寻宝<br/>（稀有钢笔/复古眼镜）"] --> B["Claude 在线检索锁定稀有资产"] end subgraph TrustBarrier["交易核心痛点"] B --> C{"线上防伪难题<br/>（高仿欺诈风险）"} end subgraph PhysicalMoat["物理验证闭环"] C -->|缺乏线下据点| D["线上无法核实<br/>交易被迫中断"] C -->|协同 GameStop 线下网络| E["卖家送至 1,600 家线下门店"] E --> F["实体专家防伪检测与盖章"] F --> G["Claude 识别权威验证标<br/>自动放款完成交易"] end ``` 更关键的质变发生在此刻的 AI 时代。Anjney 以自身为例：他酷爱收集万宝龙（Montblanc）稀有钢笔和 Jacques Marie Mage 复古手工眼镜，但在公务繁忙时，他会让 Claude 等 AI Agent 在全网替他寻找藏品。然而，二手高价值收藏品交易最大的痛点永远是欺诈（fraud）。当 Claude 兴冲冲地报告找到了某支极为罕见的万宝龙钢笔时，只要被要求“再三核实真伪”，整个 Agent 交易流程就会因为无法在线验证实体原件而彻底瘫痪。 如果 GameStop 与 eBay 合并，GameStop 遍布全美的 1,600 家线下零售网点便能瞬间转变为分布式的物理商品验真中心（physical verification centers）。在未来的 Agent 商务（agentic commerce）体系中，AI Agent 在判定商品已获得线下门店实体盖章防伪后即可放行交易。这种物理网络构建的信用闭环，在 AI 时代不仅不会贬值，反而会成倍增值。 ### [05:36](https://www.youtube.com/watch?v=30NHL5Jg3ek&t=336s) 从 Discord NFT 夏天汲取的教训：社区电商不可或缺的流动性与验真 Anjney 借由自身过往的创业实战经验进一步佐证了上述判断。2020 年疫情暴发前，他将自己创立的上家公司出售给了 Discord，并随后出任 Discord 的平台负责人（Head of Platform）。起初他的主要职责是构建针对游戏生态的 SDK 和 API，期间还助力了 AI 图像生成领军企业 Midjourney 的早期孵化。然而到了 2021 年所谓的“NFT 夏天”，无聊猿（Bored Ape Yacht Club, BAYC）等链上项目引爆全网，Discord 内部在不知不觉中沉淀了惊人的交易量，每月有超过 100 亿美元的商品交易总额（GMV）在服务器间流转。 面对这块巨大的蛋糕，Discord 创始人 Jason Citron 和 Stan Vishnevskiy 要求 Anjney 带领团队设法将这部分交易变现并搭建官方电商系统。经过深入调研，Anjney 领悟到了平台电商的本质： - **流动性（Liquidity）的根基**：无论是 Discord 还是 eBay，社区电商的核心价值都在于提供撮合流动性。 - **数字资产的验真机制**：NFT 能够天然在 Discord 上爆发，是因为它基于区块链（on-chain），所有权明确透明，无需任何线下实体验真环节。 - **物理资产的落寞退场**：当董事会质疑 NFT 热潮难以持久，要求团队拓展至极客钟爱的限量版球鞋（sneakers）、定制客制化机械键盘以及稀有腕带等实物藏品时，团队发现没有物理验真网络，这类交易根本无法闭环。由于 Discord 不可能凭空去建设一千多家线下门店来接收实物，该实物电商项目最终只得无奈叫停（nixed）。 正是在那次惨痛的业务探索之后，Anjney 彻底看清了 eBay 独特的线下价值，认定它是被公开市场严重低估的战略级资产。 ### [07:30](https://www.youtube.com/watch?v=30NHL5Jg3ek&t=450s) 算力瓶颈与 1885 年工业革命隐喻：创立 AMP 与算力网格愿景 主持人随后将话题拉回前沿科技领域：既然 Anjney 比市场上 90% 的风险投资人（VC）都更加坚信通用人工智能（AGI pill），那么如果站在 Agent 优先（agent first）的原生视角，究竟该如何从零构建全新的 eBay？ Anjney 直言自己目前无暇过多顾及应用层重构，因为他现在的精力已全部投身于算力基础设施之中。他创立 AMP 并将其组建为一家公共利益公司（PBC），目标是充当全美乃至全球“算力电网”（compute grid）的独立系统运营商（Independent System Operator, ISO）。 Anjney 用了一个极具画面感的历史隐喻来描绘当下的 AI 发展阶段：当前行业正处在类似 **1885 年左右工业革命时期的英国**。彼时蒸汽机刚刚被发明普及，全社会都意识到利用这项动力能够生产出钢铁、笔记本、钢笔乃至汽车等前所未有的新型工业产品；但此时工业社会面临一种极度稀缺的基础要素——**煤炭**，各大工厂都在疯狂囤积煤炭。在今天的 AI 经济体中，这种稀缺的“煤炭”就是算力（Compute）。 更荒谬的景象在于基础设施层面的极度割裂与闲置。俯瞰当年的工业英国，每家作坊都在后院自建小型发电机，并且全都在半负荷低效运转；反观当下，Anjney 观察自己投资组合中的众多 AI 初创团队，其 GPU 集群的有效利用率同样惨不忍睹。他特别点名提到，埃隆·马斯克（Elon Musk）在孟菲斯（Memphis）为 xAI 搭建的由 50 万张 GPU（如 GB100 等）构成的庞大算力集群中，其模型算力利用率（Model Flops Utilization, MFU）仅有可怜的 11%，节点分配效率更是不足 60%。 仅此一项，就意味着有价值高达 120 亿美元的昂贵算力资源正在被白白空耗。正因目睹了全行业计算资源的巨大错配，Anjney 才创立了 AMP，旨在通过集中采购算力、签订长期租赁协议并深度聚合算力集群，以基础架构与资本双轮驱动的平台模式，彻底盘活全网低效的算力网络。 ### [08:56](https://www.youtube.com/watch?v=30NHL5Jg3ek&t=536s) 算力战时状态与零售合并迷局：Ryan Cohen 的资本算盘 访谈切入时，Anjney Midha 正在描述当下算力生态的极端紧缺现状。自今年 1 月离开 Andreessen Horowitz（a16z）创办 AMP 以来，他夜以继日投入的唯一目标就是构建算力网络，通过跨集群调度容量将利用率推向极致。他预计到今年年底，AMP 协调上线的算力规模将达到数十亿美元。面对主持人抛出的关于 GameStop 与 eBay 潜在合并的商业八卦，Anjney 打趣道，自己目前根本没有闲暇研究怎么重组 eBay，“如果 Ryan Cohen（GameStop 董事长兼 CEO）打电话求助，我或许会出谋划策，但现在算力领域正处于真正的战时状态（wartime on compute）。” 主持人随后追问了这场备受关注的合并传闻背后异乎寻常的资本逻辑：当前 GameStop 市值约 100 亿美元，eBay 约 480 亿美元，若按多头逻辑合并后市值有望冲上 1000 亿美元，但令人不解的是 Ryan 似乎在资金尚未全部到位的情况下就过早公开了构想。Anjney 一针见血地分析了背后的心理战：Ryan 手握大约 90 亿美元现金，身处一家市值仅 100 亿美元的公司，这意味着除去账面现金，整个 GameStop 的品牌价值、庞大线下门店资产在二级市场上被压缩估值到仅仅 10 亿美元左右，现金资产完全没有获得任何估值溢价。Ryan 宣布这一设想时，原本很可能指望二级市场的投机情绪被瞬间点燃、股价应声暴涨，随后他便能顺理成章地登上 CNBC 阐述宏图，借势增发 200 亿美元股票完成杠杆并购。然而市场泡沫和迷因股的热潮并未如愿演化，两人顺势调侃起连 Allbirds 这种鞋履品牌都被市场炒作涉足算力基建的荒诞现状，Anjney 更是幽默地隔空喊话：“如果 Allbirds 的 CEO 在听节目，请把你们的算力全交给我们，我们全部接单！” --- ### [10:57](https://www.youtube.com/watch?v=30NHL5Jg3ek&t=657s) 解构 AMP 双轮模式：以成本价供给算力的“反云厂商”飞轮 Anjney 明确厘清了 AMP 的定位：AMP 绝不是传统的云计算提供商，而是一家集“基础设施”与“风险资本”于一体的控股型平台。其商业逻辑十分激进：基础设施业务负责在全球范围内统筹锁定顶级 GPU 算力资源，并以**零加价的成本价（at cost）**直接倾斜分配给旗下的投资组合公司。在独立 AI 研发团队被大型超大规模云厂商（Hyperscalers）卡脖子的当下，AMP 旨在为最具潜力的前沿实验室打破这一战略瓶颈。 在资本端，AMP 的首期基金在成立短短 8 周内便获得了超过 13 亿美元的认购承诺，其中已经向 Anthropic 领投并注资 3 亿美元。Anjney 表示今年还将追加募集约 65 亿美元资金。作为 Anthropic 早期最重要的天使投资人之一（被主持人戏称为“Anthropic 的 Jason Calacanis”），Anjney 过去四年里始终在疾呼：算力是 AI 时代最核心的战略资源，如果研发团队无法从 Hyperscalers 那里分到算力，前沿研究就无从谈起。AMP 借此建立起一套自我强化的经济飞轮：通过基金投资获得股权收益，再将附带权益（Carry）和管理费利润循环投入，采购更大规模的底层算力，继续以成本价扶植下一代独立模型与科学实验室。 ```mermaid flowchart TD accTitle: AMP 基础设施与资本协同飞轮 accDescr: 展示 AMP 如何通过资本募资、算力直供与收益再投资形成闭环 subgraph Capital ["资本与投资业务 (Amp Capital)"] LP["LP 机构注资<br/>(养老金/主权基金)"] --> Fund["首期基金与专项资金池"] Fund --> Invest["直投独立前沿团队<br/>(如 Anthropic / Periodic Labs)"] Invest --> Returns["股权增值与 Carry / 费率收益"] end subgraph Infra ["基础设施调度 (Amp Grid)"] Returns --> Procurement["大规模锁定采购算力"] Procurement --> ComputePool["统筹分布式算力池"] ComputePool -->|"按成本价直供 (At-Cost)"| Invest end ``` - **资本募资速度**：首期基金在启动仅 8 周内已达成逾 13 亿美元出资承诺，年内目标追加筹集约 65 亿美元。 - **Anthropic 早期渊源**：Anjney 是 Anthropic 种子轮前八位投资人之一（也是最初的天使投资人之一），首期基金已向其追加部署 3 亿美元。 - **分配机制**：基础设施团队不赚取算力差价，直接平价输送给被投团队，通过后续股权升值实现商业回报。 --- ### [12:23](https://www.youtube.com/watch?v=30NHL5Jg3ek&t=743s) 重塑古典 VC：Periodic Labs 与 AI 材料科学的验证闭环 针对主持人指出“许多主流 VC 面临五年发展断层、因迷茫瘫痪而只能在 Neo-labs 与应用层盲目下注”的痛点，Anjney 分享了截然不同的底盘逻辑。他将自己的方法论称为“回到未来”——重拾 Arthur Rock 与 Brook Byers 时代的经典风投范式。Anjney 19 岁加入 Kleiner Perkins（KP），20 岁在甚至还没考取驾照、未达到法定饮酒年龄时，便作为观察员与传奇投资家 John Doerr 同坐一个董事会。在 AMP，风险投资业务依托名为“AMP Foundry”的孵化工场，专注于与最顶尖的科学家深度共创、一砖一瓦地从零打造前沿实验室。 一个典型案例是 Periodic Labs：由 ChatGPT 联合缔造者 Liam 与前 DeepMind 量子物理研究负责人 Dorje 联手创办，专注于通过 AI 寻找全新的高温超导材料。Anjney 继去年领投种子轮后，近期再次主导了其 A 轮融资。不同于单纯停留在数字模拟层面的 AI 公司，Periodic Labs 在门洛帕克拥有一座占地 30,000 平方英尺的实体研发基地，Anjney 每周有三天在现场工作，每天早晨 8:00 至 8:30 准时参加站会制定攻坚优先级。实验室构建了一套全自动的物理验证流水线：首先由 AI Agent 预测超导新材料结构，随后由机器人自动化合成样本，再利用 X 射线衍射仪（XRD）测试材料的实际晶体结构与物性，并将物理验证结果源源不断地回传喂给训练集群。过去 90 天内，该实验室完成的材料实验验证量已经超越了该学科领域过去十年的总和。 ```mermaid flowchart LR accTitle: Periodic Labs 闭环验证架构 accDescr: AI 预测、机器人合成、物理测试与模型重训的自主研发飞轮 subgraph LabFlow ["自动化超导材料验证闭环"] A["AI 算法模型<br/>预测新材料结构"] --> B["机器人自动化系统<br/>物理样本合成"] B --> C["X 射线衍射仪 (XRD)<br/>物性检测与验证"] C -->|"实验验证数据回传"| A end ``` 谈及宏观环境，Anjney 指出 Anthropic 的崛起彻底让华尔街与公开资本市场读懂了“苦涩的教训”（The Bitter Lesson）与 Scaling Laws 的威力和确定性。当年他为 Anthropic 筹集种子轮时，给 22 个熟人好友打电话遭遇了 21 次无情拒绝；而今他只需拨出两个电话，基金额度便迎来 3 倍超额认购。主权财富基金与大型养老金争相进场，资本已不再是瓶颈，核心竞争力已经彻底转向谁能为科学家团队提供不可替代的算力、工程调度与落地执行力。 --- ### [15:46](https://www.youtube.com/watch?v=30NHL5Jg3ek&t=946s) 信任为城河：掌舵 Google 算力调度大脑的原班人马 在被问及团队如何支撑如此庞大的两栖业务时，Anjney 强调：“在资本狂潮中，信任是唯一的护城河（Trust is the moat）。”AMP 核心初创团队仅有 5 人，但全员均有极度深厚的实战交集与信任沉淀： - **工程联合创始人（Sebastian Lobo）**：Anjney 14 年前在斯坦福大学的室友。在此后的 12 年间，Seb 在 Google 与 Mihai 共同主导开发了 Borg、XBorg 以及 GQM 调度系统。该系统曾将 Google 全球内部算力资源池的利用率常年压榨在 95% 以上；在当时的 Google 内部，利用率哪怕跌至 94%，就会被内部定义为全球级别的重大运营事故（major outage）。 - **运营合伙人（Andrew Erskin）**：曾任 Orrick 律所合伙人，长期担任 Anthropic 的外部总法律顾问；同时也是 Anjney 第一次创业并成功出售给 Discord 的 Ubiquity6 公司的总法律顾问（GC）。 - **幕僚长（Rosie）**：早前在国际公关巨头爱德曼（Edelman）服务 Anjney，随后在 Anjney 担任 a16z 副总裁期间统揽其公关传播事务。 访谈最后，Anjney 澄清了外界对公司名称的误解：AMP 并不是取自他名字的首字母（并非 Anjney Midha Partners），而是源自物理学中的电流基本单位安培（Ampere）。它代表着电力、能量与算力基础设施的脉动，正如 AMP 的初心一样，旨在为整个独立前沿科学世界注入源源不断的算力动能。 ### [16:50](https://www.youtube.com/watch?v=30NHL5Jg3ek&t=1010s) 精简团队、第一性原理与拥抱 RIA 合规化 谈到当下的技术浪潮，Anjney Midha 认为人类正在步入科技史上真正的“大文艺复兴”时代。在传统的资本配置框架中，投资机构人为设立了各种壁垒分明的标签与分类——算力、资本、体育战队、杠杆收购、前沿科技等，然而在 AI 时代，这些人为界限完全不应存在。只要拥有一支彼此深度信任的核心精简团队，从第一性原理出发，再配合最先进的技术工具，所能爆发出的执行力和能量足以超越传统机构几个数量级。 当被主持人问及基金涉及公开市场建仓、同时又采用公益公司（Public Benefit Corporation, PBC）架构，是否正在注册成为注册投资顾问（Registered Investment Advisor, RIA）以兼顾更成熟的二级资产时，Anjney Midha 证实 Amp 正处于 RIA 的申请进程中。尽管 Amp 成立仅约 90 天，但他对这一套监管规范极其熟悉。此前他在 Andreessen Horowitz（a16z）担任 AI 基础设施基金的通用合伙人（GP）数年，a16z 本身就是 RIA，严苛的合规防线早已内化为他的肌肉记忆。他幽默地借用马克·扎克伯格（Mark Zuckerberg）的名言调侃道，十二年前硅谷崇尚的是“快速行动，打破常规”（move fast and break things），后来改口为“依托稳定基础设施快速行动”（move fast with stable infrastructure）；而作为一家 AI 基础设施团队，Amp 从第一天起就以兼备敏捷度与合规基底来构建。 - **组织灵活性**：摒弃传统多阶段资产分类（Multi-stage asset class firm）的繁文缛节，靠高信任度的小团队高效调度算力与资金。 - **合规进展**：团队成立 90 天内全力推进 RIA 资质落地，严格遵守投资顾问监管框架。 --- ### [18:09](https://www.youtube.com/watch?v=30NHL5Jg3ek&t=1089s) 为何采用 PBC 架构：解决正外部性与打造“自律型公共事业” 主持人好奇地追问 Amp 为何会选择 PBC 架构，并打趣道，回顾 Anjney 先前提到开启工业大建设的 1885 年，当时的金融大亨可不是什么公益公司，根本是“个人利益公司”（Personal benefit corporations），尽管造就了铁路、火车、机械与汽车，但也带来了激烈的工会斗争与社会阵痛。Anjney 对此将 PBC 的立足点拆解为“实质层面”与“调性层面”两个维度。 ```mermaid flowchart TD accTitle: 正外部性与公共事业型资本结构 accDescr: 展示风险投资与算力基础设施如何通过正外部性赋能顶尖团队，并在缺乏政府及时干预下形成自律型公用事业。 subgraph Drivers["双轮驱动业务"] VC["风险投资业务<br/>(Funding Innovation)"] Infra["算力基础设施<br/>(Compute Infrastructure)"] end subgraph Impact["系统级外溢效应"] Externality["释放巨大正外部性<br/>(Positive Externalities)"] FocusTeams["赋能小团队顶尖研发<br/>(Anthropic / Mistral 等)"] end subgraph Solution["Amp 组织形态定位"] Failure["市场失灵与消费不足<br/>(缺乏即时政策干预)"] Utility["自律型公用事业 (PBC)<br/>以成本价分发算力，跳过繁文缛节"] end VC --> Externality Infra --> Externality Externality --> FocusTeams FocusTeams --> Failure Failure --> Utility ``` 从实质层面来看，Amp 兼具风险投资与算力基础设施两大业务，二者共同的核心特征就是“正外部性”（Positive externalities）。当风险投资走在正确轨道上，能为全社会撬动巨大的创新红利；而算力若供给给像 Anthropic 这样高度专注的小团队，其产出前沿（SOTA）技术能力的效率，可达到拥有数万人巨头（如 DeepMind）的数倍甚至十倍以上。在经济学逻辑中，正外部性往往伴随着市场失灵与供给不足，通常需要政府出面纠偏，但当公共政策行动滞后时，私营部门必须率先承担责任。1885 年解决市场失灵并提供公共品的大企业，最终多被列为公用事业进行监管；而 Amp 恰恰将自己定位为一家“自律型的公用事业机构”，专门为当今最顶尖的科学家和技术领袖——新一代的 Dario Amodei（Anthropic 联合创始人）、Guillaume Lample（Mistral AI 联合创始人兼 LLaMA 核心作者）、Robin Rombach（Stable Diffusion 共同创造者）等提供 VC 资金与算力。 Anjney 回忆起本科在斯坦福大学选修彼得·蒂尔（Peter Thiel）的《从 0 到 1》（Zero to One）课程时，彼得·蒂尔当时最振聋发聩的一句话是“我们想要飞天汽车，得到的却只是 140 个字符（Twitter）”。而今，Anjney 自己回到母校主讲全校规模最大的课程 CS 153，人类正亲历飞天汽车、室温超导、治愈癌症即将成为现实的前夜。 从“调性层面”来看，PBC 架构是 Amp 最坚实的法律护城河。Amp 正以成本价直接向生态及被投企业分发数十亿美元规模的算力资源，如果换作传统上市公司或纯商业基金架构，极易被短视的股东以“破坏股东利益”为由起诉。但从长期看，这种培育生态的做法将创造数个数量级的复合价值。正如 Ben & Jerry's 冰激凌或户外品牌 REI 能够在激烈竞争的赛道中历久弥坚，AI 技术本身最终也会走向同质化，技术从来不是终极护城河，真正的壁垒在于信任、社区、文化与执行力。 - **顶尖智力的放大器**：Anjney 自谦无法企及 Dario、Guillaume 或 Robin 的科研天才，但 Amp 甘愿做他们的“实习生”，倾尽管理费与 Carry 的收益全力资助前沿突破。 - **对标百年模式**：效仿公用事业逻辑，拒绝成为暴利垄断者与“野蛮人”，平抑繁荣与萧条的极端周期。 --- ### [22:09](https://www.youtube.com/watch?v=30NHL5Jg3ek&t=1329s) 阳光下的并购实验、斯坦福 AI Coachella 与跳出泡沫叙事 在探讨技术向各行各业落地扩散时，主持人询问未来 Amp 是否会联手私募股权基金（PE）设立合资公司以推动技术下沉。Anjney 表示 Amp 的官网正是 `amppublic.com`，其核心理念正是汲取当年雷诺兹-纳贝斯克（RJR Nabisco）及《门口的野蛮人》（Barbarians at the Gate）时代的教训：传统 PE 往往充斥着嗜血裁员与掠夺性重组，而 Amp 试图探索一种“负责任的私募股权模式”。在转型过程中，资本绝不应在粗暴裁减成千上万员工后撒手不管，而必须通过教育和再培训体系让他们具备拥抱新机会的能力。 在访谈中途录音间突然传来轻微的铃声提示，引发了主持人和 Anjney 一阵轻松的玩笑互逗。随后 Anjney 敞开心扉，直言自己今年 34 岁，早已实现了财务自由，人生的下半场才刚刚开局，他唯一的野心是希望自己能作为这轮技术革命中不可忽视的“净正面力量”被历史铭记。他坦言自己早已厌倦了过去那种向别人反复证明“我早告诉过你”（I told you so）的状态： > “当他们回过头来看我的投资回报，跑来质问我：‘当初你为什么不给我打电话？’我只能回答：‘我打过了，你自己去翻翻邮件，早在 Anthropic 的 A 轮、B 轮和 C 轮我就把他们引荐给你了。’” 为了不再在沟通损耗中原地踏步，Anjney 决定直接连线有限合伙人（LP），搭建自己的基础设施平台。他主理的斯坦福重磅课程 CS 153（cs153.stanford.edu）被全校师生戏称为“AI Coachella”（AI 音乐节），旨在让不同认知层面的受众看清技术演进的真相。第一堂开幕讲座已在斯坦福官方渠道上线，后续集结了 Founders Fund 合伙人 Scott Nolan 以及英伟达创始人黄仁勋（Jensen Huang）等重磅行业领袖。 在被问及最后一个问题——“世界是否做好了‘这根本不是一场泡沫’的准备”时，Anjney 深思后感慨，现实世界的“惯性”超乎想象，大部分人至今对于 AI 真正的颠覆力一无所知。 - **理性乐观主义**：坚持 a16z 传承的理性乐观派精神，主张任何 PE 兼并或技术替代都必须在公共利益宪章监管下稳步推进。 - **CS 153 讲座阵容**：课程首讲由 Anjney Midha 担纲主讲，后续邀请包括 Scott Nolan（第八讲）以及黄仁勋（Jensen Huang）等业界核心领袖联袂授课。 ### [24:41](https://www.youtube.com/watch?v=30NHL5Jg3ek&t=1481s) AI 扩散的早期现实与节目尾声祝辞 在谈到人工智能的落地现状时，对话指出了技术渗透与大众认知之间的巨大落差。尽管业内讨论热烈，但在许多原本预期已经广泛实现 AI 扩散的现实场景中，大众实际上“几乎连 Claude 或 ChatGPT 都很少使用”。对全世界绝大多数人来说，这些大模型工具依然像天外来客一样陌生（"still alien to most of the world"）。嘉宾感慨道，哪怕今天 AI 的技术能力彻底停止迭代，甚至“AI 生态圈里有一半的人从地球上凭空消失，现实世界也不会发生什么改变”，因为整个技术普及目前依然处在极其稚嫩的早期阶段。 主持人对此深表赞同，并祝贺 Anjney Midha 顺利完成了令人瞩目的融资，赞许其将算力基础设施与投资平台深度结合的独特模式，同时期待双方的下一次对谈。Anjney 也向 TBPN 团队送上热烈祝福，称赞他们正迎来一波“时代级的爆发势头（generational run）”，并希望最近的收购案能够成为“注入更强动力的兴奋剂”，为团队的持续扩张火上浇油。他直言行业每天都需要更多像 TBPN 这样高水准的科技对谈内容，双方最终在轻松愉快的道别声中结束了整场采访。 - **AI 普及的真实鸿沟**：即便是 Claude、ChatGPT 这样家喻户晓的旗舰产品，对全球主流人群而言仍然极为遥远，技术落地才刚刚拉开帷幕。 - **行业初期特征**：当前 AI 尚未深度嵌入实体经济的核心运作中，技术能力纵使停滞也暂时不会影响社会正常运转。 - **对 TBPN 的未来期许**：赞许 TBPN 在媒体生态中的高速崛起，期待其借助资本与收购进一步扩大影响力。 ## Notable quotes > [03:56](https://www.youtube.com/watch?v=30NHL5Jg3ek&t=236s) 亚马逊卖充电线，eBay 卖 Mickey Mantle 棒球卡。 — Anjney Midha > [08:56](https://www.youtube.com/watch?v=30NHL5Jg3ek&t=536s) 如果 Ryan Cohen 打电话求助，我或许会出谋划策，但现在算力领域正处于真正的战时状态。 — Anjney Midha > [08:56](https://www.youtube.com/watch?v=30NHL5Jg3ek&t=536s) 如果 Allbirds 的 CEO 在听节目，请把你们的算力全交给我们，我们全部接单！ — Anjney Midha > [15:46](https://www.youtube.com/watch?v=30NHL5Jg3ek&t=946s) 在资本狂潮中，信任是唯一的护城河。 — Anjney Midha > [16:50](https://www.youtube.com/watch?v=30NHL5Jg3ek&t=1010s) 十二年前硅谷崇尚的是‘快速行动，打破常规’，后来改口为‘依托稳定基础设施快速行动’；而作为一家 AI 基础设施团队，Amp 从第一天起就以兼备敏捷度与合规基底来构建。 — Anjney Midha > [18:09](https://www.youtube.com/watch?v=30NHL5Jg3ek&t=1089s) 我们想要飞天汽车，得到的却只是 140 个字符。 — Peter Thiel > [22:09](https://www.youtube.com/watch?v=30NHL5Jg3ek&t=1329s) 当他们回过头来看我的投资回报，跑来质问我：‘当初你为什么不给我打电话？’我只能回答：‘我打过了，你自己去翻翻邮件，早在 Anthropic 的 A 轮、B 轮和 C 轮我就把他们引荐给你了。’ — Anjney Midha > [24:41](https://www.youtube.com/watch?v=30NHL5Jg3ek&t=1481s) 哪怕今天 AI 的技术能力彻底停止迭代，甚至 AI 生态圈里有一半的人从地球上凭空消失，现实世界也不会发生什么改变。 — Anjney Midha ## Mentioned Anjney Midha, TBPN, AMP (Amp), GameStop, eBay, Ryan Cohen, Michael Burry, CNBC, Claude, Montblanc, Jacques Marie Mage, Discord, Midjourney, Bored Ape Yacht Club (BAYC), Jason Citron, Stan Vishnevskiy, Elon Musk, xAI, Andreessen Horowitz (a16z), Allbirds, Anthropic, Periodic Labs, Arthur Rock, Brook Byers, John Doerr, Kleiner Perkins, Sebastian Lobo, Andrew Erskin, Ubiquity6, Rosie, Edelman, Mark Zuckerberg, Dario Amodei, Guillaume Lample, Robin Rombach, Peter Thiel, Ben & Jerry's, REI, RJR Nabisco, Scott Nolan, Jensen Huang, CS 153 ## Video description This is our full conversation with Anjney Midha, recorded live on TBPN. We discuss his thesis behind AMP and why he believes compute is the most constrained and valuable resource in the AI economy, how billions of dollars in underutilized compute are creating a massive inefficiency that his firm is aiming to solve by building a coordinated “compute grid,” and why he’s structuring AMP as both an infrastructure and capital platform to back the next generation of AI labs, unlock scientific breakthroughs, and accelerate progress across the entire ecosystem. Sign up for TBPN’s daily newsletter at TBPN.com Follow TBPN: https://TBPN.com https://x.com/tbpn https://open.spotify.com/show/2L6WMqY3GUPCGBD0dX6p00?si=674252d53acf4231 https://podcasts.apple.com/us/podcast/technology-brothers/id1772360235 https://www.youtube.com/@TBPNLive [Source](https://www.youtube.com/watch?v=30NHL5Jg3ek)

[3:56] Anjney Midha: 亚马逊卖充电线，eBay 卖 Mickey Mantle 棒球卡。 [8:56] Anjney Midha: 如果 Ryan Cohen 打电话求助，我或许会出谋划策，但现在算力领域正处于真正的战时状态。 [8:56] Anjney Midha: 如果 Allbirds 的 CEO 在听节目，请把你们的算力全交给我们，我们全部接单！ [15:46] Anjney Midha: 在资本狂潮中，信任是唯一的护城河。 [16:50] Anjney Midha: 十二年前硅谷崇尚的是‘快速行动，打破常规’，后来改口为‘依托稳定基础设施快速行动’；而作为一家 AI 基础设施团队，Amp 从第一天起就以兼备敏捷度与合规基底来构建。 [18:09] Peter Thiel: 我们想要飞天汽车，得到的却只是 140 个字符。 [22:09] Anjney Midha: 当他们回过头来看我的投资回报，跑来质问我：‘当初你为什么不给我打电话？’我只能回答：‘我打过了，你自己去翻翻邮件，早在 Anthropic 的 A 轮、B 轮和 C 轮我就把他们引荐给你了。’ [24:41] Anjney Midha: 哪怕今天 AI 的技术能力彻底停止迭代，甚至 AI 生态圈里有一半的人从地球上凭空消失，现实世界也不会发生什么改变。

[17:24] Anjney Midha: I don't want to get sued by shareholders for whom it's not legible why I'm giving away billions of dollars of compute at cost to portfolio companies... You could argue that shareholder value that we're destroying, but I would argue in the long term we're actually creating orders of magnitude more value.

[0:00] First time with his new fund AMP PBC. [0:04] He's here in the waiting room. He's here [0:06] with us in the TV panel [0:07] How are you doing? [0:09] >> generational run. Yes. [0:11] >> [laughter] [0:11] >> Generational run. [0:12] >> know. It's getting started, guys. No, [0:14] generational run is used for people who [0:15] are retiring. Oh, yeah. [0:17] That's true. That's true. [0:19] >> it's I think it's appropriate to I think [0:21] it's important to recognize when you're [0:22] on a generational run that you realize [0:24] you have to actually level up even [0:25] further if you want to stay on the same [0:27] trajectory. [0:28] >> Yeah. Yeah. [0:29] >> Uh cuz if you just get complacent then [0:31] >> [laughter] [0:31] >> Yeah, it's important to like count the [0:32] chickens before they hatch. Like that's [0:34] what you're saying? [0:35] >> No, no, no. The The opposite. [0:37] >> [laughter] [0:37] >> Anyway, great great to have you back on [0:39] the show. Um [0:41] Uh it's been it's been [0:42] >> Yeah. Fun to um But yeah, catch us up to [0:45] speed on [0:45] >> Yeah, what Yeah, when did the fund [0:47] launch? What do you know, what's the [0:48] strategy? Uh and uh yeah, like like like [0:52] walk us through like the current thesis [0:54] for how you want to actually develop the [0:55] firm. [0:56] No, I think we should talk about [0:57] something more interesting. Let's talk [0:58] about eBay. Let's talk about eBay. [1:00] >> Okay. Yeah, what you got? [1:02] So So here's what I'm And I want you [1:04] guys to kind of spar with me, right? So [1:05] I looked at the Ryan did this CNBC [1:07] interview and everybody's like pinging [1:09] me and saying, "Oh my god, this makes no [1:10] sense. Blah blah." And And so here's my [1:13] my take on it, okay? Which is that [1:15] like if you read the GameStop deck [1:18] like carefully for eBay [1:21] most of what's been said about the deal [1:23] in the last 48 hours is basically [1:24] totally wrong. I I read I was just [1:26] before jumping on I was reading Michael [1:28] Burry's piece on it, which you guys [1:29] should check out. Yeah. [1:30] Um and he's right that the leverage is [1:33] pretty tight. [1:35] But I think he's not answering [1:36] the wrong question. And so is Ryan on [1:38] CNBC where he's you know, they keep [1:40] asking him like [1:41] where's the cash? How are you going to [1:42] fund the deal? [1:43] >> Yeah. [1:46] I I don't think the question The [1:48] question isn't can GameStop afford eBay? [1:51] The question is whether the underlying [1:53] business actually works. And And it [1:55] does, but not for the reason I was [1:57] expecting Ryan to pitch. [1:58] >> Okay. So if you if you pull up eBay's [2:01] 10K from February, Yeah. [2:03] fiscal 25, and I did not understand this [2:05] until I read it. eBay spent 2.4 billion [2:08] dollars on marketing. Mhm. [2:11] How many new users did they get for [2:12] that? I mean, you guys are marketers, [2:13] you understand. 1 million. [2:16] Woah. [2:17] That's crazy. [2:18] >> The user base went from one 134 million [2:21] users to 135 million users after [2:25] spending 2.4 billion on marketing. [2:27] >> So and that's basically they're [2:28] basically you you would have to imagine [2:30] they're just having to reacquire all [2:32] their old users, people that have been [2:34] on the platform before, or maybe even [2:35] lost their account, and they're they're [2:37] they're coming back. [2:40] I don't know, but it's That's $2,400 of [2:42] marketing per new user on a site that [2:44] every American already knows exists. [2:46] >> Sure. [2:47] >> Yeah. [2:48] So where's all that money going? [2:49] That's rough. So I don't think Cohen is [2:52] I don't think he's buying eBay. [2:55] Like just watching Ryan, I don't think [2:57] he's [2:58] buying eBay because he thinks he's [2:59] smarter than eBay's product team. I [3:01] think he's buying eBay because he can [3:02] see 2 billion dollars of fat that Wall [3:05] Street has been pricing as fixed cost. [3:07] And so he goes, "Okay, let me cut that." [3:08] And the interest on the debt just pays [3:11] for itself. Mhm. Interesting. [3:12] >> Right? But he doesn't necessarily want [3:14] to say that cuz he could kind of [3:17] give that idea to the management. [3:18] >> to put the money together, right? But is [3:20] your thesis that like the deal is coming [3:22] together? He has investors that he's [3:24] talking to, but it's too early to say, [3:26] "Oh yeah, I actually do have a fund [3:28] that's going to give me another five [3:30] over here. I got seven over here." And [3:32] it will map out, but just give me a [3:33] week. Or is there something else going [3:35] on? [3:35] It depends on which investor he's [3:37] talking to, but if he's pitching me, [3:38] here's what I would Yeah, yeah. [3:39] underwrite, right? I'd say, "Okay, [3:41] that's the floor. The floor is Ryan's [3:43] going to cut 2 billion dollars from this [3:44] thing of fat, put that into treasuries, [3:47] and we're going to make more money Sure. [3:48] than it's currently yielding. Okay, so [3:50] that's the floor. Now, the ceiling, cuz [3:52] I'm a I'm a technology investor, right? [3:54] Yeah, the opportunity. [3:56] Okay, so Amazon's used and collect [3:58] collectibles business has been flat [4:01] for 6 years. [4:03] They tried to renew it, they tried [4:04] collectibles, they tried trade-in, none [4:05] of it scaled. [4:07] You cannot put a 1962 Mickey Mantle card [4:10] through the same warehouse as a phone [4:11] charger. That category that category, [4:14] right, collectibles is structurally [4:15] defensible against Amazon. Amazon is for [4:17] phone chargers, Mickey Mantle for eBay, [4:20] right? [4:20] eBay has the marketplace, GameStop has [4:22] GameStop has like 1,600 stores that [4:24] could physically verify the good. That's [4:27] a real moat. Yeah, yeah. And it's worth [4:30] more in the AI era than the human era, [4:32] right? Because why? Because when AI [4:34] I I collect I love collect like rare [4:36] pens, okay? This is a Montblanc. Sure. [4:39] As I get [4:40] I love it. I'm a pen guy and I I love [4:43] old vintage glasses. I don't know if you [4:45] can see my Jacque Marie Mage box in the [4:47] back, but Nice. But when I'm out of [4:50] time, what do I do? I have Claude go [4:53] look for rare pens and glasses for me [4:55] online. [4:56] The biggest problem with used rare asset [4:59] purchases is fraud. Yeah. So, I often [5:01] will tell So, Claude will be like, [5:03] "Andre, I found this amazing pen, this [5:04] Montblanc pen." And I say, "Okay, can [5:06] you triple check that it's [5:08] it's it's real, it's not fake?" [5:11] And that's where things go off rail [5:12] because there's no way for him to verify [5:13] that without messaging the agent and so [5:15] on. Yeah, yeah. [5:16] But if you have 1,600 stores [5:20] where people who have Montblanc pens can [5:22] go and physically verify those assets at [5:23] GameStop. [5:25] Now, Claude just says, "Yeah, I checked. [5:26] It has the physical verification stamp [5:28] from [5:30] Exactly. So, you can't you need physical [5:33] verification built into the system for [5:34] agentic commerce. [5:36] And look, the reason I know this is real [5:38] is cuz a few years ago, I think you you [5:40] guys we talked about this last time, but [5:42] I sold my last startup to a company [5:44] called Discord in 2020, pre-pandemic. [5:47] So, I come on as the head of platform. [5:48] My My My job is supposed to be, you [5:50] know, Ansh, go build SDKs, APIs, and so [5:53] on for gaming. Uh we helped this company [5:55] called Mid Journey get going, AI, you [5:57] know, generate [5:58] But, 12 months later, suddenly I find [6:01] myself running, without realizing it, [6:03] e-commerce business because it was the [6:05] summer of NFTs. Yeah. [6:06] >> And the the the Bored Apes are blowing [6:09] up, and suddenly we have more than $10 [6:12] billion [6:12] of GMB flowing through Discord by [6:14] selling [6:15] Yeah. And Jason and Stan are like, "Hey, [6:18] brother, your job is to like capture a [6:19] piece of that pie." Yeah. I got, "Okay, [6:22] homework [6:23] accepted." So, we started doing a deep [6:24] dive. [6:25] And we realized ultimately [6:27] the what what these users need and pay [6:29] for, like so you have liquidity, right? [6:31] Ultimately, that's what a marketplace [6:33] like eBay and Discord um provide in sort [6:36] of community commerce is liquidity. But, [6:38] you cannot provide liquidity if you [6:39] don't have physical verification. And [6:41] for Discord, that was just out of scope. [6:43] You know, you sneaker [6:45] >> Well, and it worked, you're saying it [6:46] works for NFTs because you had the [6:47] on-chain piece. [6:48] >> need any physical. There's nothing [6:49] physical to verify. [6:50] >> You owned it, and you can transact. [6:52] >> Yep. [6:52] >> Exactly. And so, we had we are [6:54] bootstrapping the e-commerce platform at [6:56] Discord with NFTs, but of course, [6:58] everyone on the board is like, "Well, [6:59] how long are NFTs going to last? It's a [7:00] fad. So, Ansh, what else is coming?" We [7:01] go look at rare sneakers, rare [7:03] keyboards, like bands, all this stuff [7:05] that nerds like me love. But, for those [7:08] you need physical verification. And once [7:09] we realized physical verification was [7:10] out of scope, we nixed the project. [7:11] >> Yeah, you're not going to go have the [7:12] 1,400 like retail locations where people [7:15] can drop things off. Yeah, that's that's [7:18] that's quite that's quite interesting. [7:20] So, that's when I realized, "Okay, eBay [7:22] is this undervalued asset." And I hope [7:25] that Ryan has figured it out as well. [7:27] Because if he hasn't, [7:29] he's giving me ideas. Yeah, yeah. Have [7:31] you Have you tried to walk through what [7:34] what uh you know, given that you're [7:37] probably more [7:38] AGI pill than than I would say 90 90% of [7:43] of VCs, have you tried to play out what [7:46] is it what how would you how would you [7:48] build eBay from the ground up today [7:51] with an agent first approach? Is that [7:53] even the right question to ask? Look, I [7:56] I have not guys cuz right now I'm in a [7:58] I'm a computer infrastructure guy, [7:59] right? We started Amp as this public [8:01] benefit corporation whose job is to be [8:03] an independent system operator of the [8:04] compute grid. We think about [8:06] We think we're roughly in like 1885 [8:08] industrial England where the steam [8:09] engine's been invented. Everybody knows [8:11] that you can make cool new products [8:12] like, you know, uh steel and and [8:16] notebooks and pens and cars. [8:18] And there's this very scarce input [8:20] called coal that everybody is hoarding. [8:23] In this case, it's compute. [8:24] >> Mhm. And when I when you fly over [8:26] industrial era England, you'll see all [8:28] these factories getting set up and [8:29] everyone's running a generator in their [8:31] backyard at half capacity. I'm going, [8:33] this makes no sense. If I'm looking at [8:34] all my portfolio companies, you know, [8:35] you these these clusters are running at [8:37] like half utilization. In fact, Elon's [8:39] Yeah. like got 500,000 GBP 100s in [8:42] Memphis at running at 11% MFU and less [8:44] than 60% their node allocation. I mean, [8:47] this is $12 billion of compute being [8:49] wasted. So, I set up Amp as an AI [8:50] infrastructure organization where we buy [8:52] a bunch of compute, we do long-term [8:54] leases, we pull that all those clusters [8:56] on the grid, we coordinate capacity, [8:58] drive up utilization. And by the end of [9:00] this year, I think we'll have, you know, [9:01] several billion dollars of compute [9:02] coming online. But that's what I've been [9:04] focused on night and day since like I [9:05] left Andreessen Horowitz in January. And [9:08] so, no, I have not had time to look at [9:09] how to redo eBay, but if Ryan called, [9:11] I'd probably help him out. But right [9:12] now, it's wartime on compute, guys. [9:14] >> So, we yeah, [9:15] I want to talk about Amp, but I also [9:17] want to talk about just last question on [9:19] the the the the combination of eBay and [9:21] GameStop. Like I get the thesis, the [9:23] bull case. [9:24] Uh GameStop is $10 billion, [9:26] eBay is like $48 billion right now. You [9:28] put them together, maybe you get to 100 [9:30] billion. I'm in for the bull case. The [9:33] question is like, [9:34] how what's going on with like the plan? [9:37] Because it feels like Ryan just doesn't [9:39] have the capital, but but then he [9:40] announced it. Like what like what what [9:42] what do you think is happening behind [9:44] the scenes because there's one thing [9:46] where you could throw it out as like oh [9:48] like these two companies should work [9:49] together. Here's a bull case. Let me [9:51] know if you want to be on the team that [9:53] does this. And then there's the other [9:54] one which is like make the offer before [9:56] the capital's lined up, but I just [9:58] haven't been through enough of these [9:59] stories to actually understand like why [10:02] it's playing out this particular way. [10:04] >> to be honest I think he's sitting there [10:05] with I think he has $9 billion of cash. [10:08] He's in a $10 billion company. I think [10:10] when he announced this, I think he [10:12] expected the stock to pop like crazy and [10:15] he'd be going on CNBC being like this [10:17] merger [10:19] could make sense. [10:20] >> Sure. And I think that [10:21] >> We'll issue another 20 billion of equity [10:23] and then and then we'll merge or [10:25] something. [10:25] >> And I think if you looked at I think if [10:26] you looked at like how [10:28] kind of frothy some things in the market [10:30] could be you could have imagined that [10:32] playing out. I mean the Allbirds thing [10:33] was I'm sure you appreciated appreciated [10:36] that from a [clears throat] meme stock [10:37] >> Direct competitor to you. [10:38] >> competitor. [10:40] Amp com Amp [laughter] versus Allbirds [10:42] will be the new horse race. Oh no. [10:45] Um anyways, that that was my that was [10:47] that was my read cuz he cuz they're [10:48] basically GameStop is basically valued [10:50] at like [10:51] the brand is and the and all the retail [10:53] locations and everything is valued at [10:54] like a billion dollars, right? He's [10:56] getting no credit for all the cash. [10:57] Yeah. So to be clear Amp is not a cloud [11:00] business. We are so I started Amp as a [11:02] holdings business and I've got an [11:03] infrastructure business and a capital [11:05] business and the infrastructure business [11:07] secures compute and passes on at cost to [11:09] our portfolio companies. We have more [11:10] than $1.3 billion in commits for our [11:11] first fund of an added eight weeks. And [11:13] so we do do venture capital investments. [11:15] We put $300 million into an optic. Yeah. [11:18] Oh, okay, cool. Yeah, but we we we need [11:20] to raise another roughly, you know, six [11:22] and a half billion dollars this year and [11:23] a lot and more is getting committed by [11:24] the day. [11:26] But we give away the computer at cost to [11:27] the independent ecosystem because my [11:28] belief is that you know, You know, the [11:30] the like sort of [11:32] the optimal unit of research today is a [11:35] like a focused out then steam outside of [11:37] the hyperscalers, you know, Anthropic [11:38] and coding which was I was the first one [11:40] of the first eight I'm certainly I think [11:42] I was the first angel investor if not [11:43] the first uh investor in the round and I [11:46] was [11:46] >> you're the Jason Calacanis of of [11:48] Anthropic. Unfortunately, J Cal I could [11:51] never top J Cal, but like if I was a J [11:53] Cal intern or something, fine, I'll take [11:55] the win, you know, but [11:57] But I I think more importantly like I [11:59] think compute is a strategic asset I've [12:01] which I've been yelling about for 4 [12:02] years and it's a primary bottleneck on [12:03] these teams and if you're not at the [12:04] hyperscalers, you just can't get access. [12:05] So we buy up that compute, we give it at [12:07] cost to the portfolio comp teams and [12:09] then we reinvest the profits of carry [12:11] and fees to buy more compute and so on [12:13] and so forth. And so I'll take as much [12:15] capacity as I can get from all birds. I [12:16] love it when new people go into the [12:17] business cuz that gives us more supply. [12:19] So if you're the Albert CEO listening to [12:20] this, please send us your compute, we'll [12:22] take it all. [12:23] That makes that makes a lot of sense. Uh [12:25] what are you excited to invest in? [12:28] You're investing in teams that need a [12:30] lot of compute. You're trying to find [12:33] things that aren't going to get [12:34] steamrolled by [12:36] Anthropic who's another big portfolio [12:38] company. [12:39] Uh there's there's you know There's [12:41] application layer things and there's a [12:43] lot of other different things that need [12:45] compute, but what what do you think so? [12:46] >> Cuz I feel like I feel like a lot of a [12:48] lot of VCs are would never say this out [12:51] loud, but a lot of [12:53] I get the sense from a lot of VCs that [12:55] they're kind of like paralyzed [12:57] where they're like they really don't [12:59] they don't have a clear sort of [13:00] understanding of where things will be in [13:02] 5 years and they feel like they need to [13:04] be active [13:06] um and so it's a mix of like doing new [13:08] neo la doing some neo neo labs, maybe [13:12] doing some application layer stuff and [13:13] just kind of praying. [13:15] But I would hope that you have a [13:18] given given your background and how [13:20] you're approaching this, you have like [13:22] a stronger thesis on on where the [13:24] opportunities to invest at the early [13:26] stages. [13:27] Look, in some sense, it's it's back to [13:29] the future. I started my career at [13:30] Kleiner Perkins when I was 19. My first [13:32] board seat was as an observer with John [13:33] Dorr when I was 20. I wasn't old enough [13:35] to drive Technic or drink, sorry. I [13:37] didn't have a I didn't have a driver's [13:38] license. I wasn't old enough to drink. [13:40] And and I got the chance to apprentice [13:42] with like the greats like you know Brook [13:43] Byers. And and look, that's the that's [13:45] the vintage of venture capitalist I grew [13:47] up admiring like Arthur Rock. And that's [13:49] my you know, our thesis at Amble on the [13:51] on the venture capital side, our [13:52] business is called the Amp Foundry where [13:53] we help create code design you know, new [13:55] labs one at a time. My current one is [13:57] called Periodic Labs. And we just [13:59] decided to lead their series A. I led [14:00] the seed round last year with Liam who [14:02] was the co-creator of ChatGPT and Dorje [14:03] was there um who led some of the quantum [14:06] physics teams at DeepMind and we're [14:07] trying to find new high temperature [14:09] superconductors there with physical We [14:10] have a 30,000 square foot facility in [14:12] Menlo Park. I spend 3 days a week there. [14:14] We do a stand-up every morning from 8:00 [14:15] a.m. to 8:30 a.m. and then we make our [14:17] priorities and then go execute. And you [14:18] know, basically we have AIs predict new [14:21] materials. We then have robots [14:22] synthesize the new materials. We then [14:23] have an X-ray diffraction machine that [14:24] tests whether the material has the [14:26] properties that robots said uh the AI [14:27] said. And then we pipe that verification [14:29] loop back into the training room like as [14:31] many times as you need for the agents to [14:32] continue predicting new superconductors. [14:34] And in the last 90 days [14:36] we've had more material verifications [14:37] than I think in the last decade in the [14:38] field. And so I'm a huge believer in [14:41] unblocking frontier progress in domains [14:43] where it the verification uh sort of [14:45] loop is clearly just like we know it's [14:48] going to work but execution is the [14:49] bottleneck. And then I like giving these [14:50] the best teams, the best scientists, the [14:52] best engineers, the capital, the [14:53] compute, and the commercial help they [14:54] need. Now, I I think that the beauty [14:57] about having Anthropic around is that [14:58] it's made this idea of the bitter lesson [15:00] and and scaling legible to the capital [15:02] markets. So now instead of me having to [15:03] call up 22 friends and getting 21 knows [15:05] which was the case with the seed round [15:06] of Anthropic, now I I make two calls [15:09] instead. We get like you know, like we [15:11] get three times oversubscribed. So [15:12] capital is no longer the bottleneck [15:14] which is phenomenal. You know, again, [15:15] I've been at it 8 weeks, have more than [15:16] a billion dollars committed for my our [15:18] first fund. I'm a solo key man GP on on [15:20] fund. And there's lots of institutions, [15:22] pension funds, sovereign funds who are [15:23] like, "Andre, how much more can you put [15:24] to work, especially in publics, [15:26] privates, buyouts?" The It's a bonanza [15:28] for people who want to be true partners, [15:29] who want to be the Arthur Rock of this [15:31] era. I think if you believe in the [15:32] bitter lesson, it's not new, it's been [15:33] around for ages. You and I The three of [15:35] us talked about this like over a year [15:36] ago at the last recent Horowitz AGM. [15:38] Yeah, still better. [15:41] I [15:41] I'm I'm more zen than I am bitter. [15:44] >> [laughter] [15:46] >> Uh how are you thinking about building [15:47] out the team on both sides? [15:50] Uh trust is the moat. So, there's five [15:51] of us on the team. My full-time [15:53] engineering co-founder was Sebastian [15:54] Lobo. We were roommates 14 years ago at [15:56] Stanford, and then he went on to build a [15:57] grid into a new Google. [15:58] >> Overnight success. [15:59] Here we go. What what was that one? [16:01] Overnight success. Overnight success. [16:03] Thank you. 12-year overnight success in [16:05] California. [16:06] >> Exactly. Um so, Seb and Mihai built the [16:08] Borg, Xborg, GQM scheduler, which kept [16:10] the Google internal capacity pool at [16:12] more than 95% utilization. There, if it [16:14] was at 94% utilization, that was [16:16] considered a major outage globally. Wow. [16:18] Um Andrew Erskin is my uh partner on the [16:20] operations side. He was a uh right a [16:22] partner at Orrick, which was the you [16:24] know, outside counsel for Anthropic. [16:25] >> Yeah. Um and he was my GC at Ubiquity 6, [16:28] which was the company I sold to Discord. [16:29] And then Rosie, who was my chief of [16:30] staff, ran comms for me from Edelman [16:32] when I was at Andreessen Horowitz, uh [16:33] you know, when I was a VP there. Um and [16:35] so, [16:36] >> the band back together. It's it's it's [16:38] it's the Andre reunion, basically. And [16:40] you know, we put we called Amp uh not [16:42] after my initials. Sometimes people [16:44] think that. It's not Andre Partners or [16:45] anything like that. It's about energy. [16:46] It's about amp, you know, ampere is the [16:48] unit of of energy. And we we want to amp [16:50] things up because we think we're [16:51] entering like the great renaissance in [16:53] technology. And [16:54] you know, if you can have a small team [16:56] that trust each other across context, [16:58] you know, compute, capital, sports [17:00] teams, buyouts, leverage technology, all [17:02] of this stuff is These are all buckets [17:04] and categories that we've all put, you [17:05] know, traditional capital allocators [17:07] have put around these asset classes that [17:09] shouldn't exist. And I think if you have [17:10] the flexibility to go back to first [17:11] principles with a small team that you [17:13] trust, you can execute, you know, with [17:15] with orders of magnitude less less size [17:17] of a team as a firm in this new era with [17:19] the right tools. I don't know if if that [17:21] makes sense. No, totally. You you [17:22] mentioned taking positions in public [17:24] companies. [17:25] The fund structure is a PBC. Are you [17:27] also an RIA? Like how are you thinking [17:30] about navigating both of those asset [17:32] classes since that's a little more [17:33] mature? [17:33] >> in process. Okay. [17:34] >> Yes, we are in process of becoming an [17:36] RIA cuz you know we founded the firm [17:38] barely Yeah. Yeah, 90 days ago, but I'm [17:40] used to that cadence because Andreessen [17:42] Horowitz was a RIA. I was a general [17:43] partner in the AI infrastructure fund [17:45] for several years as you guys know and [17:46] we were an RIA. I'm used to the [17:47] compliance and regulatory sort of [17:49] guardrails we got to follow and I think [17:50] that piece stressed us to, you know, [17:52] have that cadence from day one. And so [17:54] we're we're going to make sure that we, [17:55] you know, Zach, if you remember this [17:57] this like, you know, 12 years ago Zach [17:59] went on TV to say move fast and break [18:01] things and then you have to update the [18:02] thing to be like move fast with stable [18:03] infrastructure. And and I think we move [18:04] fast with stable infrastructure from day [18:06] one essentially cuz we are an AI [18:08] infrastructure team. Yeah. Talk about [18:09] the PBC. [18:11] Like if I'm playing back like what what [18:13] what what year were you referencing? [18:15] 1850 or something like that? 1885. [18:18] >> 1885. So if I go back to 1885 and I [18:19] think about the financiers that, you [18:22] know, created the industrial buildout, [18:25] they were not public benefit [18:27] corporations. They were [18:28] >> They were personal benefit corporations. [18:30] Yeah. [laughter] [18:31] And I mean there was a lot of good that [18:32] was created. We got railroads and trains [18:34] and, you know, machinery and cars and [18:36] all sorts of things out of the [18:37] industrial revolution. There were also [18:38] things that were rough and there was [18:40] unionization and battles and back and [18:42] forth. Like what is the PBC in service [18:44] of solving? What what what what why PBC? [18:47] Yeah, great question. So there's the I [18:49] I'll tell you the substantive answer and [18:51] then the vibes answer, right? So from a [18:54] substantive perspective if we do two [18:56] things, right? We have a venture capital [18:57] business and we have an infrastructure [18:58] business. Both things have this very [19:00] unique property called positive [19:01] externalities. When implemented [19:02] correctly, venture capital can unlock [19:04] massive positive externalities for the [19:06] ecosystem and for the world cuz you end [19:07] up funding innovation when done [19:08] correctly. And then infrastructure the [19:09] same, right? When you have compute [19:11] that's used by small focused down teams [19:13] like Anthropic that's able to produce 10 [19:15] times more soda capabilities than like [19:17] Yeah. DeepMind, which is you know 60,000 [19:18] or 160,000 people, then you're [19:20] generating positive externalities for [19:22] the world by being much more efficient [19:23] per unit of input with the output they [19:25] create. And so I was like, "Huh, well, [19:26] what happens when as an economist you [19:29] look at positive externalities?" Usually [19:31] you have market failure, you have under [19:32] consumption of that good. Well, how do [19:34] you correct the market failure? Usually [19:36] you get the government to intervene, but [19:37] if you don't have the government [19:38] intervening in time, what do you do? You [19:39] become a private sector participant. And [19:41] if you look at the arc of 1885, you [19:44] know, private sector businesses that [19:45] ended up correcting market failure by [19:48] by producing public benefits, they ended [19:50] up getting regulated as utilities. Mhm. [19:52] That's what Amp is. Amp is a [19:54] self-regulated utility of that provides [19:56] venture capital and infrastructure to [19:58] the world's leading scientific teams. [19:59] The next Dario, the next, you know, [20:02] Guillaume at Mistral who created Lama, [20:03] the next Robin who created stable [20:05] diffusion. These are the [20:07] my generation's smartest minds. I'm not [20:09] smart enough to be, you know, them, but [20:11] I can be their intern. And instead of [20:13] waiting for the rest of the space to [20:15] come up with standards and institutions [20:16] to enforce this, we're like, [20:18] "Dude, let's just do it ourselves and [20:20] show the world you can have fun while [20:21] doing it with a small team. You don't [20:22] need to be, you know, some something [20:24] called a [20:25] you know, like these words like RIA, [20:27] multi-stage asset class firm, doesn't [20:29] matter. Just let's skip ahead to the [20:30] part, the good part, and and like, you [20:33] know, use all the proceeds that we get [20:34] from management fees and carry to keep [20:36] the space like innovating at the pace [20:38] that we were promised you know, 12, 13 [20:40] years ago, and instead we got tweets and [20:41] not flying cars. You know, when I was at [20:43] Stanford, I got the chance as an [20:45] undergrad I had the chance to take Peter [20:46] Thiel's class, the Zero to One. And he [20:48] was you know, his whole moniker was, "We [20:50] want to flying cars, but we got you [20:52] know, um [20:54] 140 characters. Well, 240 characters, [20:57] thank you. And now I'm back at Stanford [20:58] teaching CS 153, which is the largest [21:00] class on campus. It's called AI [21:01] Coachella. We got thousands of people [21:03] following along. AI Coachella is a good [21:04] one. And and it's got frontier systems [21:07] cuz it's all possible now. We're [21:09] literally [21:10] in our lifetime we're going to have [21:11] flying cars. We're going to have room [21:12] temperature superconductors. We are [21:13] going to solve cancer. We just want to [21:15] do it in a way that's stable, [21:17] predictable. We want to skip all the [21:18] boom and bust cycles and the way to do [21:20] that is to lead by example and say, "Hey [21:21] guys, the public benefit, you know, is [21:24] to provide goods and services that are [21:25] utilities and make sure that we we don't [21:28] like Let's be the adults in the room and [21:30] not do the stuff where we try to be [21:31] rubber bands and monopolists and got [21:33] greedy along the way." And so, that's [21:35] the substantive answer. The vibes answer [21:36] is, "Look, I don't want to get sued by [21:37] shareholders for whom it's not legible [21:39] why I'm giving away billions of dollars [21:40] of compute at cost to portfolio [21:42] companies, right?" Because we That's [21:43] what we're doing and that shareholder [21:45] That's You could argue that shareholder [21:47] value that we're destroying, but I would [21:48] argue in the long term we're actually [21:49] creating orders of magnitude more value. [21:51] And if you look at Ben & Jerry's, you [21:52] look at REI, they've become stable [21:54] enduring businesses in categories that [21:56] are fairly crowded and eventually I do [21:57] think technology and AI will get [21:59] crowded. It will get commoditized cuz [22:00] technology's never the moat. Trust is [22:02] community is a moat. Culture is a moat. [22:04] Execution is a moat. And so, we're [22:05] trying to skip ahead to that part, but [22:07] it takes time for people to get aligned. [22:08] So, until then we'll [22:09] >> Well, we see Well, we see an Amp joint [22:12] venture with private equity [22:14] to [22:15] uh help distribute [22:17] >> diffusion. [22:19] Yeah. [22:19] >> If If we can Yeah, you know, if you go [22:21] to our website, it's called [22:22] amppublic.com because I do think if you [22:24] look back to the like vulture uh era of [22:26] private equity, you remember like RJR [22:28] Nabisco and what what what the [22:29] barbarians at the gate, we should just [22:31] learn from from their mistakes and go, [22:33] "Can we do private equity but done right [22:35] in an aligned way? Let's not rush to [22:36] like lay off, you know, hundreds of [22:39] thousands of people and then not [22:40] re-educate them and prepare them for [22:41] their new opportunities." I I'm an [22:42] optimist as you guys know. I came from [22:44] Andreessen Horowitz. So, I I I'm sort of [22:46] a rational optimist. I believe the [22:47] transition can be done in a positive [22:49] way. Um [22:51] Is that me getting kicked off? There's [22:52] like a bell ringing, but that might be [22:54] here. No, I don't know. Oh, okay, cool. [22:57] No, I'm I'm looking at [22:58] >> bell we have is this, [23:00] but no, we're [23:03] We got plenty of sounds for you. [23:04] Continue. Continue. Uh we will do [23:07] everything we everything we do is [23:09] governed by public benefit charter. So, [23:10] if we do private equity, it'll be [23:11] governed in the public benefit. If we do [23:13] education stuff, that'll be in the [23:15] public benefit. Look, I've made more [23:17] money than I know what to do in life. [23:18] I'm 34 and I'm just getting started. So, [23:20] my goal is I'll be remembered for having [23:21] been a net positive influence on the [23:23] space. I just got tired of telling [23:24] people I told you so because after a [23:26] while they start looking at my returns [23:27] and going, "Why didn't you give me a [23:28] call?" And I said, "I did. Look at your [23:29] email. I introduced you to Entropic in [23:31] the series A and the series B and the [23:32] series C." And so, at some point I was [23:34] like, "You know what? I'm just going to [23:35] go direct, talk to the LPs, set up a [23:37] platform, build infrastructure, and uh [23:40] hopefully be known as a generally like [23:43] sane, common sense, rational point of [23:45] view on stuff that can often be like is [23:48] not legible to people from different [23:49] parts of the stack. And that's what the [23:50] class is about. So, cs153.stanford.edu, [23:53] I would recommend anybody watching go [23:55] check it out. Um the the lectures are [23:57] all online and the first one went on on [23:59] Stanford's official page on Thursday. [24:01] >> Was that with Scott Nolan? [24:03] So, Scott, no, actually Scott is lecture [24:06] eight. I put up We put up lecture one, [24:08] which was mine, as a kind of the opening [24:10] act cuz, you know, Scott is one of the [24:11] main liners, the headliners of the AI [24:12] Coachella. [24:13] Scott's will be up soon as well, and [24:16] then Jensen was last week. So, I think [24:18] he followed Scott. Well, thank you so [24:20] >> last question for me. Do you think the [24:21] world is prepared for it not to be a [24:23] bubble? [24:29] That's a good question. The world if the [24:31] world prepared not for it to be a [24:32] bubble. Oh, yeah. Yeah, inertia is [24:36] Guys, inertia is a powerful thing. Most [24:37] of the world still has no idea what AI [24:39] is. It is crazy. And I've been flying to [24:41] places where I thought there would be [24:42] diffusion of AI by now, and they just [24:45] aren't [24:46] barely using Claude, ChatGPT. I mean, [24:48] these things are still alien to most of [24:50] the world. And so, if if if we stopped [24:52] capabilities today and half of us in the [24:54] AI ecosystem vanished off the planet, [24:56] nothing would change. It's still so [24:58] early. Yeah. Yeah. It is early. Very [25:00] cool. Well, great to catch up. [25:01] Congratulations on a very impressive [25:04] fund raise and and a very unique [25:05] approach and looking forward to the next [25:07] conversation. Thank you. You guys are on [25:10] a generational run and I hope the [25:12] acquisition does nothing but give you [25:14] guys more steroids and more fuel for the [25:16] fire. We need more of you every day. [25:18] Fantastic. We'll talk to you soon. [25:20] >> Have a good one. Thank you. [25:21] >> bye. Cheers.