CoreWeave Signs 5-Year GPU Deals as Power Bottlenecks Limit AI Growth

电力瓶颈制约AI增长，CoreWeave签订5年期GPU长约 由于扩展定律依然有效，且通电数据中心容量取代芯片供应成为增长的主要制约因素，AI实验室正锁定不可取消的五年期算力合同。 CoreWeave首席开发官（CDO）Brannin McBee详细阐述了企业客户在面临账单冲击并转向多年期基础设施承诺之际，AI算力格局正在发生的转变。对话指出，尽管原生GPU供应紧缺已有所缓解，但目前主要的运营瓶颈是获取配备变压器和高压电网接入的通电数据中心建筑空间。McBee解释了客户需求为何依然压倒性地集中在NVIDIA硬件上、推理工作负载如何主导算力使用，以及CoreWeave如何利用结构化SPV获取长期债务融资，同时对算力的大宗商品化保持怀疑态度。 企业客户正因未加节制的AI使用面临账单冲击，这推动其转向模型路由机制，将更简单的查询分发给成本更低的小模型。 (0:00) CoreWeave需求强劲，正在金融服务领域扩大业务版图，并为中国以外全球前十大AI实验室中的9家提供支持。 (5:00) AI硬件的保值时间长于此前预期，Nvidia A100和H100等老款GPU的使用寿命可维持6至8年。 (5:00) 随着各实验室寻求长期容量保障，云算力合同期限已从3年延长至5年期不可取消协议。 (15:00) Nvidia的技术路线图正从风冷Hopper机架转向液冷Blackwell NVL72机箱以及即将推出的Vera Rubin系统。 (20:00)

CoreWeave Chief Development Officer Brannin McBee asserts that enterprise AI compute demand is accelerating rather than slowing, leading AI labs and enterprises to commit to unprecedented five-year infrastructure contracts. Rather than chip availability or raw real estate, physical power delivery and operational complexity now dictate industry bottlenecks, keeping high-performance compute resistant to financial commoditization. Enterprise Demand and Multi-Year Infrastructure Commitments (5:00) Despite corporate alarm over surging AI expenses—exemplified by Uber exhausting its 2026 AI budget in four months and an enterprise client racking up a $500 million monthly bill—McBee contends these events reflect authentic, expanding consumption rather than budgetary retrenchment. CoreWeave reports a financial services backlog approaching tens of billions of dollars, alongside serving 9 of the top 10 global AI labs outside China. To manage costs and preserve access, client behavior has shifted from short-term experimentation to locking in long-term capacity. Standard take-or-pay cloud contract lengths have extended from three years to four, with frontier AI labs now committing to five-year non-cancelable agreements without hardware refresh requirements. Furthermore, optimal model routing—diverting routine queries away from frontier models to smaller instances—has countered early market fears of rapid GPU obsolescence, extending the operational lifespan of hardware like Nvidia A100s and H100s to an estimated six to eight years. Persistent Nvidia Dominance Across Diverse Workloads (22:30) Despite growing industry interest in custom ASICs and hyperscaler silicon, CoreWeave operates under a demand-driven capital expenditure model where procurement is dictated entirely by client requests. McBee states that customer demand remains almost exclusively concentrated on Nvidia hardware, driven by the maturity and efficiency of its 15-year-old CUDA software ecosystem. While industry observers often argue that alternative chips can adequately service inference workloads, CoreWeave's operational data shows that Nvidia chips dominate both training and inference. Inference workloads now account for well over 50% of CoreWeave's active capacity, yet customers deploy the same fungible Nvidia GPU clusters interchangeably across both training and inference tasks rather than migrating to specialized non-Nvidia hardware. Electrical Supply Chains as the Primary Bottleneck (28:00) Securing physical land and sourcing individual GPUs are no longer the primary operational hurdles for AI expansion. Instead, the critical constraint is the availability of "powered shells"—data centers fully provisioned with high-voltage electricity, utility interconnections, and integrated cooling systems ready to energize server racks. Expansion timelines are dictated by severe supply chain delays for electrical equipment, notably high-voltage transformers and backup battery arrays, as well as a severe shortage of certified high-voltage electricians, who require more than five years of apprenticeship training. Capital markets have increasingly favored operators who can translate power into active compute; CoreWeave has raised over $21 billion to support more than one gigawatt of energized, operational capacity. Structured SPV Financing and Compute Non-Fungibility (40:00) To fund massive capital expenditures without overloading the corporate balance sheet, CoreWeave structures its hardware deployments into bankruptcy-remote Special Purpose Vehicles (SPVs). These entities package GPU hardware, data center leases, and non-cancelable five-year take-or-pay customer contracts together, matching predictable revenues directly against debt amortization schedules. This structure recently enabled an investment-grade-rated, non-recourse debt facility (DDTL4) priced at SOFR plus 225 basis points. McBee expresses skepticism toward near-term attempts to create standardized financial futures or liquid commodity markets for compute capacity. Because hardware performance metrics like Goodput and Model FLOPs Utilization (MFU) vary wildly depending on network topology, liquid cooling setups, and proprietary orchestration software, large-scale compute clusters running hundreds of thousands of GPUs remain fundamentally non-fungible. While structured debt markets have unlocked the capital necessary to fund large-scale AI infrastructure, physical electrical constraints and proprietary orchestration software continue to separate high-end performance from commoditized compute. Whether emerging custom hyperscaler silicon and standardized software frameworks will eventually break Nvidia's ecosystem advantage remains the central unresolved tension in compute market structure.

CoreWeave联合创始人兼首席开发官Brannin McBee指出，近期企业面临的算力账单冲击并非需求放缓的迹象，而是实际应用激增的证明；随着前沿AI实验室将基础设施采购锁定为长达五年的不可取消合同，行业的首要瓶颈已从GPU芯片供应转移至具备高压供电和液冷能力的带手机房（powered shells）及复杂的电气供应链。 企业算力支出剧增与长期合同锁定趋势 (5:00) 针对企业因Token成本飙升而引发的“算力账单冲击”——例如优步（Uber）在四个月内耗尽2026年全年的AI预算，以及有企业因未设置API调用上限而在单月产生5亿美元账单，McBee反驳了有关AI普及率放缓的猜测。他认为，成本激增恰恰反映了终端用户和企业真实的巨大消费需求。目前CoreWeave在金融服务领域的积压订单已接近数百亿美元，且服务于全球（不含中国）前十大AI实验室中的九家，量化投资巨头Jane Street也直接与其合作构建专有模型。 此外，硬件基础设施的采购模式发生了根本性转变。两年前客户标准合同仅承诺三年期限，随后在Hopper与Blackwell阶段延长至四年，目前前沿AI实验室普遍要求签订长达五年的不可取消“照付不议”（take-or-pay）合同，以锁定前沿计算资源。同时，模型路由技术的分流使较小任务得以分配至低算力芯片，这也打破了华尔街此前关于GPU在两年内贬值归零的担忧，使Nvidia A100和H100等老代际硬件的使用寿命延长至6到8年。 CUDA生态壁垒与推理负载主导现状 (22:30) 尽管科技巨头与量化交易机构正积极探索定制ASIC芯片（如微软的Maia 200芯片、谷歌TPU和亚马逊Trainium），但McBee表示CoreWeave采取的是完全由客户需求驱动的“基于成功的资本支出”（success-based CapEx）模式，目前客户几乎所有明确承诺的资金仍完全投向英伟达基础设施。他将此归因于英伟达历经十五年构建的CUDA软件生态系统，该生态在效率、稳定性和扩展性上构筑了坚实壁垒。 在工作负载分布上，McBee透露CoreWeave平台上活跃的基础设施利用率中，推理任务占比已远超50%，且训练与推理工作流是在同一套可互换的物理英伟达GPU集群上无缝交替运行。从硬件迭代来看，虽然前沿实验室正争相部署液冷NVL72机架的Blackwell并测试下一代更高Token能效的Vera Rubin架构，但广大企业客户仍对Ampere、Hopper和Blackwell等前代或当前代际芯片保持着旺盛需求。 算力扩张约束转向带手机房与电气供应链 (28:00) McBee明确指出，当前限制全球算力扩张的核心瓶颈已不再是获取土地或采购GPU裸片本身，而是“带手机房”（powered shells）的交付能力——即完成实体建设、具备高压电力接入、集成液冷系统并能立即向机柜供电的机房空间。构建此类设施受到严苛且周期缓慢的电气设备供应链限制，包括高压变压器和后备电池储能系统等核心组件的漫长交付期。 除硬件设备外，专业高压电工等熟练技术劳动力的短缺亦成为难以在短期内扩充的关键制约，其资质通常需要五年以上的学徒期培养。尽管弗吉尼亚北部等地的数据中心面临地方政府审查与社区抵制，但实际投产延误的主因仍是电网并网与物理建造周期。目前CoreWeave已运营超过1吉瓦（GW）已部署GPU的活跃数据中心算力，且年内已筹集超过210亿美元融资，市场正给予能将通电空间快速转化为计费算力小时的运营商显著的执行溢价。 结构化SPV资产融资与算力非同质化特征 (35:45) 为支持大规模资本扩张并降低资本成本，CoreWeave建立了双层资本架构：母公司（ParentCo）负责企业级股权与资产负债表运营，资产公司（AssetCo）则通过独立的特殊目的载体（SPV，俗称“盒子”）运作。每个SPV内打包了具体的GPU机架、数据中心租约、债务以及与客户签署的五年期不可取消照付不议合同；合同产生的稳定现金流直接用于偿还固定本息摊销并覆盖运营支出。这种破产隔离的非追索权结构赢得了固定收益市场的认可，其DDTL4项目获得了投资级评级，定价为SOFR+225个基点，显著收窄了信用违约互换（CDS）利差。 针对未来是否会出现类似芝加哥商品交易所（CME）挂牌的流动性GPU算力期货市场，McBee基于其大宗商品交易经验持怀疑态度。他强调算力并非像黄金或天然气那样标准化的同质化商品，不同云环境下的模型FLOPs利用率（MFU）和有效产出（Goodput）差异悬殊；在数十万张GPU互联的超大规模集群中，依赖底层自研编排软件的故障自愈与热插拔能力构成了核心差异，使得大规模算力在短期内无法实现标准化金融合约交易。 本期节目总结表明，AI算力基础设施的演进本质上是一场关于新型市场结构的探索，其核心矛盾在于高度定制化的工程编排与标准化大宗商品化之间的张力。虽然企业正通过模型路由等手段控制调用开支，但只要扩展定律（Scaling Laws）在前沿模型研发中持续生效，对十万卡级集群的确定性锁定仍将推动资本以长周期结构化工具持续注入硬件底座。

## Background 本期节目由彭博社（Bloomberg）旗下的知名播客《Odd Lots》于 2026 年 6 月 8 日在 RSS 及各大网络平台发布。根据官方简介，节目邀请了算力云服务领军企业 CoreWeave 的联合创始人兼首席开发官 Brannin McBee，共同探讨在公司上市并大规模发行债券后的最新算力市场格局，内容深入覆盖了 GPU 租赁模式、英伟达 Vera Rubin 等新一代硬件架构、训练与推理需求的演进，以及变压器与电力紧缺等物理基础设施瓶颈。 ## TL;DR 节目全面破除了近期市场关于企业缩减 AI 算力支出的悲观论调，指出账单休克背后恰恰印证了推理需求的爆发，顶尖 AI 实验室正通过长达 5 年的“照付不议”长约抢占锁定算力。CoreWeave 分享了其年内完成逾 210 亿美元融资、并依托 SPV 隔离结构开创首个投资级 GPU 基础设施贷款的资本运作机制。同时，对谈深入剖析了物理现实对算力扩张的严酷制约——真正制约交付的瓶颈已从单纯的显卡短缺转向带电厂房（Powered Shell）、大宗变压器交付及高压持证电工的极度匮乏。最后，从大宗商品交易视角冷峻论证了 GPU 算力因底层运维、液冷工程与调度软件（Goodput/MFU）的巨大差异，短期内根本无法被标准化为大宗期货资产。 ## Atmosphere & style 整场对话兼具华尔街金融宏观分析与硅谷硬核工程实操的复合质感，节奏紧凑、信息密度极高。主持人 Joe 与 Tracy 以轻松敏锐的日常打趣切入，随后与出身大宗商品交易员的嘉宾展开深度博弈，交流风格冷静、理性且直击要害。对话从资本市场对算力违约风险的恐慌叙事一路穿透至数据中心现场泥泞的电工学徒制与高压变压器断供危机，既有揭秘百亿级金融工程创新的智力快感，又充盈着直面物理实体供应链壁垒时的冷峻现实感。 ## Bulletin - `06:05` CoreWeave 驳斥了企业削减 AI 开支导致算力需求放缓的传闻，透露其年内融资已超 210 亿美元，实际运行的 GPU 电力规模突破 1 GW，排名前十的顶尖 AI 实验室中有 9 家在使用其基础设施。 - `16:34` 顶尖 AI 实验室对长期算力短缺的焦虑加剧，正普遍要求签订长达 5 年期的“照付不议”（Take-or-pay）长约，锁定固定费率且不可解约或调换硬件。 - `25:00` CoreWeave 平台超 50% 的算力利用率来自推理任务，但算力扩张的最大物理瓶颈已从单纯的 GPU 芯片短缺转向“带电厂房”（Powered Shell）、大宗变压器短缺及高压持证电工匮乏。 - `42:47` 由于不同云厂商在网络拓扑、液冷工程及调度软件上的差异导致 Goodput 和 MFU 性能分化，GPU 算力缺乏同质性（Fungibility），短期内难以形成标准化的算力期货交易市场。 ## Key takeaways - `01:05` 企业界因未做用量配额管控导致算力账单爆炸（如 Uber 4个月烧光全年预算），推动智能模型路由（Model Routing）加速落地以分流任务。 - `06:05` CoreWeave 前线数据表明市场对顶级算力的需求毫无限缩迹象，手握超 10 家单笔 10 亿美元级客户，金融服务业订单储备达数十亿美元。 - `10:00` 模型路由与任务分层彻底推翻了“GPU 两年报废”的悲观论调，轻量任务使 A100 保持高负载运转，H100 真实服役寿命延长至 6 到 8 年。 - `12:45` 金融巨头（如量化做市商 Jane Street）正绕过 OpenAI、Anthropic 等大模型实验室中间层，直接与 CoreWeave 签约并自主调度裸算力集群。 - `16:34` AI 实验室签约周期从早期的 3 年演进至当前 5 年期“照付不议”锁死合约，为基础设施服务商通过债券和资产证券化融资提供了极具确定性的现金流。 - `20:00` 大模型扩展定律（Scaling Laws）依然有效，英伟达架构从风冷为主的 Hopper 42U 机架迅速跃迁至全液冷 Blackwell NVL72 机架，CoreWeave 亦已入驻 Vera Rubin 测试机架。 - `25:00` 数据中心扩建的核心瓶颈已转向具备完整供电、变电与冷却系统的“带电厂房”（Powered Shell），大型变压器交期与 5 年学徒期的高压电工极度短缺。 - `30:00` 算力供给的核心壁垒在于迅速将纸面电力规划兑现为可产生账单的 GPU 运行工时，而非画饼式的远期电力签约。 - `32:15` 随着 CDS 违约利差大幅收窄，CoreWeave 借鉴现代银行业风控逻辑，主动将富余算力定义为“待分配”而非“滞销”，坚决拒绝被单一巨头包揽以分散集中度风险。 - `37:00` 通过设立独立的 SPV 载体隔离硬件与债务风险，CoreWeave 将 5 年期长约与债务摊销挂钩，以 SOFR + 225 bps 完成了首个针对 GPU 的投资级非追索权贷款（DDTL4），成功引入保险长线资金。 - `42:47` GPU 算力缺乏同质性（Fungibility），两家相同配置的 H100 集群因调度软件与底层集成差异会导致有效吞吐量（Goodput）和利用率（MFU）悬殊，阻碍了算力期货标准化合约的落地。 - `47:35` 算力市场呈现严重两极分化：单张或小规模 GPU 供应在长尾市场随手可买，但具备极高运维门槛、数万至数十万张 GPU 互联的超大规模连片集群极其稀缺。 - `50:00` 全球算力供需缺口在中短期内无法弥合，尽管数据中心选址面临能耗和环保争议，但工程建设供应链的硬性工期仍是比审批舆论更关键的制约因素。 - `54:40` 超大规模云厂商与高频量化机构加速自研专有芯片（如 TPU、Trainium、Maia），但英伟达依托 CUDA 生态构建的 15 年软件与能效壁垒依然牢固统治着主流需求。 ## Detailed walkthrough ### `01:05` 推理市场的剧变与企业 CFO 的“算力账单休克” 节目开场，主持人 Joe Weisenthal 和 Tracy Alloway 打趣道，鉴于当前人工智能推理市场的演进速度实在太快，无论是模型的底层调用逻辑还是具体的应用落地场景，几周之前讨论的内容往往很快就会过时，以至于他们甚至有必要像追踪股票市场一样，每月甚至每周专门录制一期节目来追踪算力与推理市场的最新动态。他们半开玩笑地提议把播客分支命名为“大量推理”（Lots of Inference）。两年前那种仅仅让 AI 写一首打油诗的尝鲜与玩具阶段早已彻底翻篇，如今全球企业正在向算力投入天文数字般的真金白银，以至于各大跨国公司的首席财务官（CFO）们普遍遭遇了严重的“账单休克”（sticker shock）。 各大公司因算力账单超标而紧急踩刹车的消息接连见诸报端。最具代表性的是 Uber 近期的举措：管理层不得不明确给员工设立硬性配额，每人每月在 AI token 上的开销不得超过 1500 美元。有报道指出，Uber 仅仅用了四个月时间就彻底烧光了原本为 2026 年全年规划的 AI 算力预算，促使其首席运营官（COO）开始严肃质询这些支出到底有没有带来对应的生产力飞跃。Axios 甚至报道过一个更为极端的案例：一家未具名的企业客户由于没有设置用量上限与配额管控，在一个月内竟产生了高达 5 亿美元的惊人账单。这种混乱的根源在于，普通员工往往并不理解不同模型之间的成本差异，即便只是为了查询一个嘉宾头衔之类的微小需求，也会直接调用最昂贵的尖端前沿大模型。 针对这一痛点，Joe 提出了关于模型路由（optimal model routing）的行业演进预判，并回顾了此前与高盛首席信息官 Marco Argenti 的讨论。由于轻量级模型与顶级前沿模型在单次查询成本上可能相差近百倍，企业未来必将大力投资于智能模型路由机制，让系统自动将任务分流到性价比最高的模型上。与此同时，Tracy 补充道，英伟达等芯片制造商也在推出旨在降低 token 开销的新架构。而正如 Hudson River Trading 的 Ian Dunning 在近期的现场对谈中所指出的，当前的瓶颈正在从早期的单纯争抢 GPU 芯片，转移到物理基础设施层面——寻找能够接纳高密度 GPU 负载并具备充足电力供应的数据中心地产，其难度如今已经不亚于甚至超过了采购芯片本身，加之社会层面对数据中心建设的阻力也日益增加，整个市场的博弈维度已变得空前复杂。 - **Uber 算力预算危机**：仅用 4 个月便耗尽全年的 AI 预算，随后出台每人每月最多 1500 美元的 token 限制令。 - **极端账单事故**：Axios 披露有企业客户因未做用量限制，单月算力账单飙升至 5 亿美元。 - **模型路由（Model Routing）需求激增**：前沿模型与专用/轻量模型单次调用成本相差多达 100 倍，倒逼企业加速布局调用优化架构。 - **物理基础设施短缺**：可供接入算力集群的机房地产与充沛电力，已取代单纯的芯片短缺成为最突出的瓶颈。 ### `04:55` 重逢 CoreWeave：从加密挖矿旧时代到 AI 新云霸主 在这一背景下，节目迎来了 CoreWeave 的联合创始人兼首席开发官（CDO）Brannin McBee。Joe 回忆道，他们第一次接触 CoreWeave 仿佛是在很久以前的另一个时代——当时整个科技界和资本市场还在热衷于将英伟达的 GPU 芯片用于加密货币挖矿。而如今，CoreWeave 已经脱胎换骨，完成了公开上市，成为当红的“新一代云服务商”（NeoCloud），专门为全球最核心的人工智能工作负载提供训练（Training）与推理（Inference）算力集群。 作为陪伴英伟达技术架构一路演进的核心算力平台，CoreWeave 见证了算力需求的彻底质变。在过去的三年里，算力基础设施的商业模式、客户群体以及融资机制均发生了颠覆性的变化，CoreWeave 不仅在公开市场募资，还深度绑定了芯片制造龙头英伟达，身处整场 AI 算力军备竞赛的最核心旋涡。 ```mermaid flowchart LR accTitle: CoreWeave 业务演进历程 accDescr: 展示 CoreWeave 从早期的加密货币挖矿硬件基础设施，逐步转型升级为服务全球顶尖 AI 实验室的 NeoCloud 算力平台。 subgraph EarlyStage["早期探索阶段"] A["GPU 挖矿机队"] --> B["算力基础设施积累"] end subgraph PivotStage["战略转型阶段"] B --> C["转向通用 AI 算力"] C --> D["构建专用低延迟网络集群"] end subgraph ExpansionStage["当前领军阶段"] D --> E["NeoCloud 领头羊"] E --> F["服务全球 90% 顶级 AI 实验室"] end ``` ### `06:05` “没有放缓，只有对顶级算力的无休止需求” 针对近期市场上关于企业 CFO 们紧急缩减开支、AI 投资是否会放缓的普遍疑虑，Joe 向 Brannin 抛出了核心问题：在经历了去年底到今年初由编码模型（如 Claude Code 等工具）所引爆的第一波推理需求高潮之后，随着第二季度财务层面的刹车声频出，CoreWeave 是否在前线观察到了客户放慢算力采购节奏的迹象？ Brannin 对此给出了斩钉截铁的反驳，认为媒体和外界完全误读了 CFO 们的账单焦虑。在他看来，企业在支出上的惊慌失措，恰恰从反面印证了市场对 AI 的真实需求（authentic and foundational demand）是何等澎湃。在一两年前，整个行业还在反复质疑“AI 真正的杀手级应用到底在哪里”、“大家天天挂在嘴边的推理需求究竟何时兑现”。而当今年年初新一代模型集中亮相时，市场几乎是在一夜之间达成了共识，认定了这就是大家期待已久的技术突破。 Brannin 进一步分析道，这一轮让企业账单爆表的现象，目前其实主要集中在极少数专业群体身上——例如高强度的软件工程师和部分金融专业人士。仅仅依靠这小部分极客在超大规模基础设施上的密集消耗，就已经制造了惊人的数据，而目前各大企业粗暴的“token 限额”手段只是一种短期阵痛，长远来看并不可持续。真正的重头戏在于 AI 正以惊人的速度渗透到传统大企业和金融服务业的核心流程中。 针对 CoreWeave 自身的运营数据，Brannin 透露了几个极具分量的指标，证明了需求侧不仅没有萎缩，反而处于前所未有的加速期： - **超大体量客户规模**：CoreWeave 目前已经拥有超过 10 家合同金额超过 10 亿美元级别的客户。 - **金融行业订单储备**：来自金融服务行业客户的在手订单储备（backlog）已经积累到了数百亿美元的级别，显示出传统高价值行业对专有算力的饥渴。 - **顶级实验室覆盖率**：如果排除中国市场，在全球排名前十的顶级 AI 实验室中，有 9 家都在使用 CoreWeave 的基础设施来支撑其核心业务。 - **市场前线观察**：无论是前沿模型的训练还是日常推理，前线看不到任何支出回撤的迹象，客户对市场上最优质的 AI 算力基础设施依然保持着“无休止的极度渴求”（unrelenting demand）。 ### `10:00` 客户结构多元化与 GPU 真实折旧寿命的重塑 CoreWeave 在首次公开募股（IPO）期间曾因客户集中度过高而遭受外界诸多质疑。批评者当时指出，其营收过度依赖少数几家超大规模云厂商（Hyperscalers）以及一两家顶尖 AI 实验室。然而，目前的客户版图已发生深刻裂变，清晰地分化为三大支柱：第一类是保持稳步增购的超大规模云厂商；第二类是 AI 实验室，全球排名前十的顶尖 AI 实验室中有九家均采用 CoreWeave 的算力；第三类则是爆发式增长的企业级客户。尽管企业级客户签署的单笔合同金额无法与数十亿美元的巨额大单相比、较少登上新闻头条，但其扩张势头极猛——仅在第四季度，CoreWeave 新增的企业客户数量就达到了以往任何单季度的两倍之多。 与客户结构扩散相伴而生的，是市场对算力生命周期认知的重要转变。随着“模型路由”（Model Routing）概念的普及，业界逐渐意识到并非所有应用场景都需要调用最新、最昂贵的前沿旗舰大模型，轻量化和特定领域的模型同样能够精准解决具体任务。这种需求的分层映射到底层算力基础设施上，打破了 2025 年一度盛行的“GPU 硬件两年内就会被淘汰归零”的悲观论调。通过将不同体量的工作负载与不同规格的 GPU 算力矩阵相匹配，上一代乃至更早的芯片找到了极具经济效益的生存空间。在实际运营中，包括 A100 在内的老旧集群至今依然在全负荷运转，而 H100 的服役生命周期被重新评估为完全可以延续 6 至 8 年，彻底推翻了过往关于算力资产快速折旧的假设。 ```mermaid flowchart TD accTitle: Model Routing and GPU Workload Matching accDescr: Diagram showing how diverse queries are routed to appropriate models and matched with different generations of GPU hardware. subgraph Routing ["模型路由层 (Model Routing)"] Q["用户提示词 / 任务请求"] --> R["智能路由分发"] R --> M1["轻量/垂直领域模型"] R --> M2["主流推理/微调模型"] R --> M3["前沿最先进大模型 (SOTA)"] end subgraph Infrastructure ["算力硬件矩阵"] M1 --> H1["A100 集群 (长效服役)"] M2 --> H2["H100 集群 (生命周期延长至 6-8 年)"] M3 --> H3["Blackwell / 新一代集群"] end ``` ### `12:45` 金融巨头的直接入场与自研模型算力需求 针对企业客户的拓展，主持人特别探讨了诸如花旗集团（Citigroup）等金融机构的合作模式：这些机构究竟是作为 Anthropic 或 OpenAI 的企业软件终端用户间接消耗算力，还是以独立实体的身份直接向 CoreWeave 采购裸算力？这一界限对于理解当前企业自研模型的真实渗透率至关重要。 对此，CoreWeave 明确界定了两者的区别：如果大型银行仅持有 AI 实验室的企业软件许可，那该笔业务归入 AI 实验室客户名下；但当前更为显著的趋势是金融机构越过中间层，直接与底层算力平台签约。目前直接面向金融服务业客户的在手订单（Backlog）规模正逼近 10 亿美元大关。典型的公开案例便是知名量化做市商简街资本（Jane Street）。Jane Street 并非通过任何大模型实验室转道而来，而是直接接入 CoreWeave 平台并自主调度管理底层的 GPU 基础设施。出于金融交易的高度保密性，外部无法获知其确切是在进行特定交易模型的训练还是生产级实时推理，但这种“机构绕过通用 AI 实验室直接掌控算力流水线”的转变，标志着企业客户已经开始将算力作为独立战略资源进行直接采购。 - **金融直客在手订单规模**：正逼近 10 亿美元（接近 billion-dollar backlog）。 - **合作模式本质差异**：客户直接在 CoreWeave 平台上交互、管理与调度底层算力基础设施，而非通过主流 AI 实验室的 SaaS/API 企业协议间接调用。 - **代表客户**：量化巨头 Jane Street 独立签约，自主掌控模型构建或实时推理的硬件资源。 ### `16:34` 推理需求爆发与五年期“照付不议”长约锁定 在经历商业广告插播后，讨论聚焦于大模型研发从纯粹的“前沿探索”（SOTA Frontier）向工具链与工程外壳（Harness，例如 2025 年底备受瞩目的 Claude Code 等工具）转移的行业动态。主持人询问，随着推理需求在算力饼图中占据越来越大的份额，AI 实验室是否开始放缓追求极致参数规模的步伐，转而将资本投向能够从现有模型中榨取更高生产力的工程工具。 CoreWeave 指出，作为基础设施提供商，他们并不干预也无法窥探 AI 实验室内部的技术路线权衡，但从交易对手的行为变化中，能够清晰观察到一个极其激进的确定性趋势：**AI 实验室要求锁定的算力规模越来越大，合同周期也越来越长**。回顾算力租赁市场的合约演进： 1. **一至两年前**：主流签约期限为 3 年期承诺。 2. **2025 年前后**：客户普遍要求将合约延长至 4 年，明确锁定对应周期的 Hopper（如 H100）或 Blackwell 芯片独占使用权。 3. **当前态势**：各大实验室主动寻求签订长达 5 年期的“照付不议”（Take-or-pay）合同。 客户在签约时要求整个 5 年周期内算力服务绝对不可中断，愿意全程锁定同一商业条款与固定费率，同时严格接受不可升级、不可更换底层硬件以及不可单方面解约的苛刻约束。这种长周期、高约束力的硬性承诺，不仅反映出顶尖实验室对长期算力短缺的极度焦虑，也为 CoreWeave 大规模通过债务和资产证券化筹资建设超大型数据中心提供了极度稳固的现金流保障。 ### `20:00` 扩展定律依然成立与英伟达架构演进：从 Hopper、Blackwell 到 Vera Rubin 在 AI 实验室的基础设施决策中，客户不仅要求长达 5 年的长期租约，而且部署规模也越来越大。行业内的工具链优化固然重要，但“扩展定律”（Scaling Laws）依然牢不可破——通过更大规模的基础设施来推进前沿模型能力的路径依然有效，且预计在下一代 Vera Rubin 架构推出时仍将延续，看不出任何放缓的迹象。关于英伟达的架构迭代，当前 CoreWeave 正在大规模部署的是 Blackwell 架构，主要以 NVL72 机架配置为主。这与上一代 Hopper 形成了根本性的工程变革：在 Hopper 时代，数据中心普遍采用 42U 机架，单个服务器机箱通常容纳 8 颗 GPU，基本采用风冷即可稳定运行；而 Blackwell 部署形态是整台机架整合了 72 颗 GPU，宛如一座巨大的立柜机箱（正如在社交平台 X 上疯传的照片那样），数据中心内部需要全面升级为重度液冷系统（Liquid Cooling）。 随着 CoreWeave 已经开始接收用于早期测试的 Vera Rubin 架构机架，整套系统的能效比（如每单位电力消耗产出的 Token 数）进一步跃升。不过，这并不意味着所有客户都在盲目追逐最新硬件。CoreWeave 平台对 Ampere、Hopper、Blackwell 等各代算力均保持着极其庞大的需求。前沿 AI 实验室通常是第一批排队锁定最新一代架构的买家，而大型企业客户则往往更侧重于当前成熟商用的 Hopper 和 Blackwell 架构。主持人 Joe 幽默地承认，过去诸如 286、386、486、Pentium 那样简单清晰的数字命名序列已成往事，如今各种以科学家命名的架构代号让人目不暇接。针对微软发布 MAI Thinking One 模型并在自研的 Maia 200 芯片上进行优化、甚至像 Hudson River Trading 这样的高频量化机构也开始试水定制硬件的趋势，Brannin 指出：CoreWeave 的商业逻辑是完全由客户需求驱动的“基于成功的资本支出”（Success-based CapEx），绝不进行盲目投机的先期建仓；而在实际业务中，除英伟达外，客户压根没有向他们提出部署其他定制芯片的需求。这背后是英伟达依托 CUDA 架构深耕 15 年以上所构建的生态壁垒，其在能效、扩展性与可靠性上依然是整个市场难以企及的标杆。 - **Hopper 架构部署**：传统 42U 机架，单机箱配置 8 颗 GPU，大部分依赖风冷系统。 - **Blackwell 架构部署**：主要采用 NVL72 整机柜架构，单机柜集成 72 颗 GPU，以液冷系统为主。 - **Vera Rubin 架构规划**：最新下一代架构，目前 CoreWeave 已开始接收测试机架（Testing Racks）。 - **资本支出策略（CapEx）**：采用需求驱动型模式，仅在客户给出明确规格并签署承诺后才采购硬件，当前客户指定采购的几乎全部为英伟达系统。 ```mermaid flowchart TD accTitle: Nvidia Architecture Evolution and Deployment Formats accDescr: A comparison showing the evolution from Hopper to Blackwell and Vera Rubin architectures alongside their cooling and chassis changes. subgraph Hopper["Hopper 架构（上一代）"] H1["标准 42U 机架"] H2["每机箱 8 颗 GPU"] H3["风冷为主（少量液冷测试）"] H1 --- H2 --- H3 end subgraph Blackwell["Blackwell 架构（当前主力）"] B1["NVL72 巨型机柜架构"] B2["整柜集成 72 颗 GPU"] B3["全面采用重度液冷技术"] B1 --- B2 --- B3 end subgraph Rubin["Vera Rubin 架构（下一代新平台）"] R1["已进驻早期测试机架"] R2["极致 Token/Watt 能效比"] R3["延续扩展定律（Scaling Laws）"] R1 --- R2 --- R3 end Hopper --> Blackwell Blackwell --> Rubin ``` --- ### `25:00` 推理市场的真实生态与新型瓶颈：“带电厂房”（Powered Shell） 针对业内常有一种说法认为“英伟达锁定了模型训练，但在推理环节自研芯片或竞品（如 Cerebros 的巨型芯片）完全可以平替”，Brannin 用 CoreWeave 的真实业务数据给出了有力回应。在 CoreWeave 最近的财报沟通中，CEO 明确指出，其平台上**超过 50% 的算力利用率实际上来自于推理工作负载**。由于训练与推理在基础设施层面具有高度的通用性（Fungible），客户在实际生产环境中依然首选英伟达芯片来跑推理。尽管行业内时常传出其他硅芯片的探讨，且部分机构会在小规模层面做尝试，但只要客户未发生根本性转向，英伟达就依然是市场上最可靠、最高效的“默认选项”（De Facto Choice）。 与三年前显卡极度短缺的局面相比，算力市场的核心约束已经发生了转移。土地本身并非制约数据中心扩张的核心瓶颈，真正的最大瓶颈是所谓的**“带电厂房”（Powered Shell）**。所谓带电厂房，是指一座已经接通高压电网、配备了完整电力传输变电组件、内置成熟冷却系统、能够直接向机架输送电力的空置数据中心。这一环节面临着层层供应链的严重卡点：不仅需要电力容量与土地指标，还受制于大型变压器（Transformers）、备用电池储能系统（UPS/BESS）的长交付周期。更严峻的是专业技术工人的匮乏——为超大规模高压数据中心施工的持证电工需要经历长达 5 年以上的学徒培训期，这种无法依靠工厂流水线快速扩充的高技术工种，已成为限制数据中心交付速度的关键实体阻碍。 - **推理工作负载占比**：在 CoreWeave 平台总利用率中，推理任务占比远超 50%，均运行在英伟达硬件之上。 - **带电厂房（Powered Shell）核心组件**：具备电网通电条件（Energized）、完整配电系统、数据中心内置液冷/风冷基础设施。 - **实体供应链卡点**：大宗变压器短缺、备用电源供给受阻，以及需要 5 年以上正规学徒认证的高压电工严重短缺。 ### `30:00` 纸上承诺与现实交付：执行力如何击穿融资瓶颈 CoreWeave 在算力供应链上占据着得天独厚的优势。Brannin McBee 谈到，公司与英伟达（Nvidia）建立了极佳的合作关系，并通过无可挑剔的实操证明了自己是“这个星球上运营此类基础设施最优秀的团队”。而此前长期制约公司扩张的融资瓶颈如今也彻底烟消云散。McBee 直言，今年迄今为止 CoreWeave 已经完成了超过 210 亿美元的融资。这种从零到数百亿规模的跨越绝非凭空产生，而是多年扎实执行积累的结果。目前 CoreWeave 已经拥有超过 1 吉瓦（GW）处于活跃运行状态的数据中心电力，全部实打实地接入了 GPU 并向客户交付计算服务。 市场上长期存在一种严重的误解，即把“纸面承诺”等同于“物理交付”。McBee 犀利地指出，签约承诺在 2030 年交付电力，与真正把电力转化为可产生账单的 GPU 运行工时（billable GPU hours），两者之间隔着巨大的执行力鸿沟。CoreWeave 的商业核心与核心壁垒恰恰就在于填补这条鸿沟——迅速将数据中心落地转化为源源不断的账单收入。正是这种惊人的兑现能力，赢得了债权人和股权投资者的信任，并极其激进地拉低了公司的资金成本。 - CoreWeave 年内已累计完成超 210 亿美元融资，验证了强劲的资本市场认可度。 - 实际投入运营并向客户稳定提供 GPU 算力服务的电力规模已突破 1 吉瓦（GW）。 - 算力市场的核心壁垒在于将纸面电力规划高效转化为可计费的 GPU 机时，而非单纯圈地画饼。 ### `32:15` 算力中心的“类银行业务”：CDS 降温与客户集中度风险 谈及融资环境的变化，主持人提到半年前市场上充斥着以 Michael Burry 为代表的做空悲观论调，当时推特等社交平台上疯传 CoreWeave 的信用违约互换（CDS）利差飙升图。然而正如金融市场一贯的规律——“CDS 暴涨时人人都想发图，收窄时却无人问津”——如今 CoreWeave 的 CDS 利差已大幅回落，市场的违约恐慌明显消退。主持人联想到硅谷银行（SVB）暴雷事件，指出数据中心与现代银行业存在高度相似的逻辑：银行业惧怕存款人单一以及行业过度集中导致的踩踏式提款，而数据中心同样面临租户集中度过高的风险敞口。 McBee 对此深表赞同，并坦言去年筹备上市时客户集中度确实是外界最大的质疑点之一，但公司在此方面已取得突破性进展。他特意澄清，公司当前的可用算力绝非“未售出产能”（unsold capacity，暗指缺乏需求），而是“尚未分配的产能”（unallocated capacity，意味着需求极度旺盛，公司正在筛选分配路径）。虽然眼下完全可以把这批未分配算力一次性全部分配给某几个单一顶级巨头（Single-name clients），但 McBee 坚称这违背了 CoreWeave 的长期战略——他们要打造的是一个服务全球顶尖 AI 生产者与消费者的多元化中立云平台，绝不做仅绑定一两家巨头的专属工具。 - 伴随市场疑虑消退，反映违约风险预期的 CoreWeave 信用违约互换（CDS）利差显著大幅收窄。 - 算力资产的资产负债管理机制与银行业类似，必须严密防范租户集中度过高带来的系统性违约与周期震荡。 - 公司将剩余算力定义为“待分配”（Unallocated）而非“滞销”（Unsold），拒绝被单一大客户完全包揽，坚决推进多租户云生态布局。 ```mermaid flowchart TD accTitle: CoreWeave 算力资产隔离与融资架构 accDescr: 展示 ParentCo 与 AssetCo SPV 之间的隔离结构及照付不议合同如何闭环偿债 subgraph Parent["母公司层 (ParentCo)"] CorporateDebt["母公司股权及综合融资"] end subgraph SPV["资产独立载体 (AssetCo / SPV 结构)"] direction TB GPUAssets["GPU 硬件资产 & 数据中心设施"] TakeOrPay["5年期照付不议长约 (Take-or-Pay)"] DebtService["项目专属债务摊销 (Amortization)"] TakeOrPay -->|"稳定现金流流入"| DebtService DebtService -->|"按计划逐步偿还本息"| GPUAssets end CorporateDebt -.->|"资本支撑与运营管理"| SPV ``` ### `37:00` 解密融资机器：SPV 隔离架构与照付不议长约 面对“为何债券市场突然对 AI 大举打开闸门”的提问，McBee 揭示了背后两大催化剂。其一是需求结构的本质转变：从去年底至今年初，推理（Inference）需求的爆发性拐点让投资者彻底确信 AI 的商业化落地前景，打消了纯模型训练泡沫的疑虑。其二则是 CoreWeave 自身成熟的资产证券化与结构化融资体系，极大地降低了资本市场的风险溢价。 在具体操作层面，CoreWeave 将融资严格划分为两大板块：母公司融资（ParentCo financing）与资产公司融资（AssetCo financing）。所有的 GPU 硬件采购、数据中心成本以及终端客户合同均被装入独立的特殊目的载体（SPV，即金融语境中的项目资产隔离“盒子”）。在 SPV 内部，客户签署的长达 5 年的“照付不议”（Take-or-pay）合同与债务的本息摊销计划（Amortization schedule）实现严格锁定与现金流精准匹配。客户付款优先支付数据中心日常运营成本并按期偿还债务本息，在极大隔离违约风险的同时，仍能为平台贡献丰厚的现金回报。 - 2024 年末至 2025 年初推理（Inference）算力需求的井喷，是驱动二级固收与债券投资者下场押注的核心拐点。 - 融资架构划分为母公司层（ParentCo）与底层资产层（AssetCo），利用独立 SPV 载体隔离硬件与债务风险。 - 依靠长达 5 年的“照付不议”（Take-or-pay）长约与债务偿还周期精准挂钩，确保现金流完全覆盖机房运维与利息支出。 ### `40:00` SPV 债务融资创新与开拓资本市场新阵地 CoreWeave 在资本运作上开创了全新路径。他们通过设立独立于母公司的特殊目的载体（SPV），将 GPU 算力设施与下游具有真实算力需求的高信用客户签署的长期租赁合同打包注入其中。这些底层资产具有极其丰厚的边际利润率（contribution margin），母公司以此为基础向信贷市场发行结构化融资工具。 ```mermaid flowchart TD accTitle: GPU 算力资产的 SPV 融资结构架构图 accDescr: 展示下游客户算力租赁合同与数据中心资产如何注入 SPV，进而获得评级并引入保险资本等多样化资金池。 subgraph SPV_Stack["SPV 资产池 (离岸/独立实体)"] A["高利润 GPU 算力长期合同"] --- B["数据中心与硬件资产"] end SPV_Stack -->|"打包资产 / 现金流支撑"| C["DDTL4 (首创投资级非追索权 HPC 债务)"] C -->|"SOFR + 225 bps"| D["保险资金等长期大型机构"] C -->|"利润边际向上贡献"| E["CoreWeave 母公司"] E --> F["可转债 / 无抵押债券 / 战略股权"] ``` Brannin McBee 透露，CoreWeave 最近完成了被称为“DDTL4”的重大融资安排。这是业内首个获得投资级评级（Investment Grade）的 GPU 非追索权（non-recourse）高性能计算（HPC）基础设施信贷项目。该笔借款的定价成本极低，仅为 SOFR + 225 个基点，不仅为 CoreWeave 锁定了极为惊艳的资金成本，更成功撬动了此前鲜少涉足这一领域的保险资金——这类手握庞大长期资本体量的顶级机构。过去在传统超大规模云厂商（Hyperscalers）的建设过程中，几乎从未有过专门为 GPU 建设提供此类专项贷款的先例。CoreWeave 不仅开拓了可转债、无担保债券市场，还吸纳了战略股权投资，为整个新兴 AI 云算力行业探索并打通了前所未有的资产证券化通道。 - **DDTL4 融资亮点**：业内首个针对 GPU 的投资级非追索权 HPC 基础设施贷款，利率达到 SOFR + 225 个基点。 - **资金池扩张**：成功引入体量庞大、要求极严苛的保险机构资本（Insurance Capital）。 - **资本工具矩阵**：覆盖 SPV 项目贷款、可转换优先票据（Convertible Notes）、无抵押信用债（Unsecured Market）以及战略股权投资（Strategic Equity）。 --- ### `42:47` 算力期货与标准化交易的障碍：GPU 算力的非同质化本质 主持人谈及 Brannin 早年从事大宗商品交易员的职业背景，并提出当前资本市场高度关注的一个热点：算力交易（Compute Trading）是否会成为一种标准化金融资产？近期 Compute Exchange 等机构正积极筹备在芝加哥商业交易所（CME）上市算力期货合约。然而，针对市场上试图通过对冲与投机生态来建立高流动性算力衍生品市场的设想，Brannin 给出了冷峻的行业判断：“短期内绝对不可能，长期也只能打一个‘也许’的问号。” 其核心根源在于大宗商品最底层的逻辑——**同质性（Fungibility，可替代性）**。黄金之所以能成为全球交易的大宗商品，是因为其由绝对一致的化学成分构成，毫无歧义；而 GPU 算力在当下根本不是同质化资产： 1. **硬件一致并不等于产出一致**：即使是同一款由英伟达生产的 H100 芯片，部署在 A 云厂商与部署在 B 云厂商，实际跑出来的性能可能存在巨大差距。 2. **专业评估指标的严苛分化**：如 SemiAnalysis 等独立机构评估算力时，依赖的是“有效吞吐量”（Goodput）与“模型浮点运算利用率”（MFU, Model Flops Utilization）。 3. **模糊性阻碍定价**：如果两朵云中的 H100 在训练同一个大语言模型时实际输出效率截然不同，算力就缺乏统一、透明的基准数据面（Data Plane），无法被大宗商品化（Commoditized）。 --- ### `45:00` 软硬件集成壁垒与日趋复杂的物理工程 当主持人追问究竟是什么造成了两台物理配置相同的 H100 会有如此大性能差异时，Brannin 详细剖析了系统工程的深层秘密。他解释说，CoreWeave 虽然全量严格遵循英伟达官方最高性能准则的 DGX 参考规范（DGX Reference Spec）进行构建，但这仅仅是准入的“及格线”，真正的核心竞争力来自于上层运维与调度软件栈。 在数万张 GPU 互联的大规模训练集群中，单点硬件失效（如 GPU 损坏、光模块故障、网络拥塞）是家常便饭。如果系统无法做到实时预测故障，并在毫秒级内自动热替换备用算力节点，客户的分布式训练任务就会直接中断，导致巨大的算力闲置与资源浪费。CoreWeave 自研的专有基础设施管理与调度软件栈，正是保证其 Goodput 和 MFU 业界领先的关键。 主持人 Tracy 随即在线查证了行业术语向听众解释： - **Goodput**：训练任务实际从硬件理论峰值算力中压榨并提取出的有效算力比例。 - **MFU (Model Flops Utilization)**：大语言模型（LLM）训练中衡量真实世界硬件计算效率的核心硬件指标。 结合以往十余年交易天然气、电力和农产品的经验，Brannin 强调大宗商品化需要技术走向极度成熟与平民化。但当下的算力世界恰恰在**背道而驰**：技术不仅没有变简单，反而在急剧复杂化——行业正从相对容易的 42U 风冷 Hopper 机架，飞速迈向极端复杂的水冷 Blackwell 机架，未来还将演进到下一代 Vera Rubin 架构。算力的物理交付、工程构建和集群运维难度在指数级上升。只要基础设施无法实现傻瓜式运维，企业就必须依赖顶尖的工程运营商，算力现货与期货的标准化就依然遥遥无期。 --- ### `47:35` 算力市场的两极分化：长尾碎片与十万级集群的鸿沟 主持人 Joe 分享了自己的亲身经历：作为一个自制小模型（基于 Google 2018/2019 年开源的 BERT 进行微调）的推理与轻量训练用户，他使用 Modal 等新兴云平台接入 Cloud Code，只花了 10 分钟注册，花费不到 5 美元，就轻而易举地调用到了 GPU。在市场铺天盖地宣称“算力极度紧缺、利用率全线爆满”的背景下，为什么普通个人开发者却能随手买到便宜的算力？ Brannin 揭示了算力市场的残酷分层： - **零散算力（Ones or Tens）**：对于获取单张、几张或几十张 GPU 的个人开发者或初创团队，长尾云市场上存在大量零碎的空闲配额，供应相对宽裕且易于获得。 - **超大规模集群（Hundreds of Thousands）**：CoreWeave 的核心客户索求的是规模达到数万乃至数十万张顶级 GPU 的超大规模连片集群。在全球范围内，能够交付并稳定运行此等规模算力阵列的厂商屈指可数。这种规模效应彻底打破了算力的标准化属性，让顶级 AI 算力演化成极度稀缺、不可替代的定制化重资产工程。 在讨论临近尾声时，主持人预告了他们即将启程前往中国香港，并引出下一阶段的重点探讨方向——大洋彼岸的中国 AI 生态格局。 ### `50:00` 中国市场需求与本土数据中心的选址政治阻力 当 Tracy Alloway 询问当前美国与中国算力市场的差异，并希望为未来的实地考察寻求提问灵感时，Brannin McBee 直言 CoreWeave 目前并未直接参与中国市场。尽管公司已经在加拿大和欧洲设立业务，且向东扩张符合战略逻辑，但管理层坚持稳扎稳打的节奏。不过在 McBee 看来，中国市场的情况与美国本质上并无二致——两地都面临着对 AI 算力“永无止境且无法满足的狂热需求”。无论供应链如何加班加点，“在短期甚至中期内，根本看不到任何能够彻底填补这一需求缺口的解决方案”。 随后 Joe Weisenthal 将话题引向数据中心建设面临的现实阻力。2023 年双方初次交流时，美国民间对数据中心的抗议还微乎其微；而如今，“反数据中心情绪”已在政界和环保团体中蔓延，成为激烈的政治争议焦点，北弗吉尼亚等数据中心重镇甚至出现因无法与当地政府达成协议而被迫叫停的项目。面对围绕水资源消耗等环保争议的提问，McBee 承认这一问题确实成了热门政治话题，CoreWeave 正在积极主动地推进本地尽调以澄清误解。然而从实际运营层面来看，政治阻力并非扩张的头号障碍，眼下最致命的瓶颈依然是硬核工程进度：变压器等重型电气元件的交期、现场建筑施工节奏以及将所有精密硬件按时组装入场的严苛供应链周期。 - **国际布局与市场供需**：CoreWeave 现阶段将业务重点放在北美及欧洲，对进一步向东扩张保持审慎；McBee 强调算力极度短缺是全球普遍现象，供需鸿沟在未来数年内无法闭合。 - **选址与舆论挑战**：数据中心面临诸如冷却水消耗、电网负荷等外界争议，部分区域项目遭遇地方审批受阻甚至搁浅。 - **核心工程瓶颈**：比起审批流程与社区阻力，制约数据中心交付速度的核心痛点仍在于变电组件交付迟滞与复杂的现场机电集成。 --- ### `52:15` 算力期货的冷思考：定制化工程与大宗商品化之争 在访谈告一段落后，Joe Weisenthal 与 Tracy Alloway 针对算力衍生品市场展开了深入复盘。即便跳出人工智能的技术语境，算力期货的尝试本身也是一场极具吸引力的“金融市场微观结构实验”，直面如何凭空定义并建立全新交易市场的深层哲学命题。讨论的核心聚焦于 Brannin McBee 与此前嘉宾 Lewis Hart 共同指出的“可替代性”（fungibility）硬伤——既然被称为“大宗商品”（commodity），其底层资产就必须具备标准化的可替代性；但如果算力无法标准化，商品市场便无从谈起。 ```mermaid flowchart TD accTitle: 算力商品化与定制化工程的矛盾 accDescr: 展现算力究竟能否像石油或电力一样大宗商品化的两种不同演进路径与技术分歧。 subgraph StandardPath["标准化假说（商品市场路径）"] S1["英伟达官方参考规范架构"] --> S2["统一机架与网络标准"] S2 --> S3["算力期货与标准化合约"] end subgraph CustomPath["工程差异假说（不可替代性障碍）"] C1["高压供电与定制液冷架构"] --> C2["极端网络拓扑与低延迟优化"] C2 --> C3["深度调优榨取算力最大性能"] C3 --> C4["集群间性能离散，无法对等替换"] end ``` 两位主持人剖析了算力商品化的现实矛盾：一方面，CoreWeave 等算力云厂商在部署 GPU 集群时，很大程度上遵循英伟达官方制定的参考规范（Reference Spec），这在理论上为全行业提供了一个可供对齐的硬件基准；但另一方面，为了从昂贵的算力集群中榨取极致性能（“get the maximum juice”），不同运营商在光纤互联拓扑、液冷工程控制和集群调度软件层面的投入差距巨大。如果最终推理服务和模型训练的高度依赖底层特定工程优化，导致不同数据中心的 GPU 输出性能千差万别，那么算力就极难被封装为同质化的金融合约。这一市场未来究竟会走向高度标准化的统一规格，还是深陷高度碎片化的工程定制，目前依然存在极大的博弈空间。 --- ### `54:40` 自研芯片的算盘：超大规模云厂商的硬件分化路径 对话最后延伸至英伟达生态之外的定制芯片版图。Joe Weisenthal 指出，整个计算架构正在迎来巨头林立的定制化浪潮：Google 拥有迭代多年的 TPU，亚马逊 AWS 部署了专用的 Trainium，微软也在打造自研硬件（如 Maia），甚至像 Jane Street 或 Hudson River Trading（HRT）这类对微秒级延迟有着极致追求的顶尖高频量化交易机构，都可能或已经在研发专有计算硬件。 两位主持人探讨了这一现象背后的核心驱动力：这些巨头斥巨资自研芯片，究竟能在特定工作负载下带来多大的实际性能提升与能耗优势？如果将模型算法的数学逻辑与硅片架构进行端到端的深度协同设计，能够释放出何种量级的优化红利？与此同时，更现实的商业考验在于，终端客户与开发者在面对非英伟达（Non-Nvidia）硬件体系时，究竟能产生多大程度的实际采购意愿与软件迁移决心。这种硬件垂直整合模式对整个云计算产业格局的深远影响，也成为节目未来重点追踪的核心议题。 ## Notable quotes > `06:05` 没有放缓，只有对顶级算力的无休止需求。 — Brannin McBee > `06:05` 如果排除中国市场，在全球排名前十的顶级 AI 实验室中，有 9 家都在使用 CoreWeave 的基础设施来支撑其核心业务。 — Brannin McBee > `10:00` 在实际运营中，包括 A100 在内的老旧集群至今依然在全负荷运转，而 H100 的服役生命周期被重新评估为完全可以延续 6 至 8 年。 — Brannin McBee > `12:45` 机构绕过通用 AI 实验室直接掌控算力流水线，标志着企业客户已经开始将算力作为独立战略资源进行直接采购。 — Brannin McBee > `16:34` 各大实验室主动寻求签订长达 5 年期的“照付不议”（Take-or-pay）合同。 — Brannin McBee > `20:00` 平台对 Ampere、Hopper、Blackwell 等各代算力均保持着极其庞大的需求。前沿 AI 实验室通常是第一批排队锁定最新一代架构的买家，而大型企业客户则往往更侧重于当前成熟商用的 Hopper 和 Blackwell 架构。 — Brannin McBee > `25:00` 在其平台上超过 50% 的算力利用率实际上来自于推理工作负载。 — Brannin McBee > `25:00` 土地本身并非制约数据中心扩张的核心瓶颈，真正的最大瓶颈是所谓的“带电厂房”（Powered Shell）。 — Brannin McBee > `30:00` 签约承诺在 2030 年交付电力，与真正把电力转化为可产生账单的 GPU 运行工时（billable GPU hours），两者之间隔着巨大的执行力鸿沟。 — Brannin McBee > `32:15` 公司当前的可用算力绝非“未售出产能”，而是“尚未分配的产能”。 — Brannin McBee > `42:47` 短期内绝对不可能，长期也只能打一个‘也许’的问号。 — Brannin McBee > `50:00` 在短期甚至中期内，根本看不到任何能够彻底填补这一需求缺口的解决方案。 — Brannin McBee ## Mentioned Joe Weisenthal, Tracy Alloway, Brannin McBee, Ian Dunning, Marco Argenti, Lewis Hart, Michael Burry, CoreWeave, Bloomberg, Odd Lots, Nvidia, Uber, Axios, Goldman Sachs, Hudson River Trading, Jane Street, Citigroup, Anthropic, OpenAI, Google, Amazon AWS, Microsoft, SemiAnalysis, Compute Exchange, CME (Chicago Mercantile Exchange), Modal, Claude Code, BERT, Hopper, Blackwell, NVL72, Vera Rubin, A100, H100, Ampere, DGX, CUDA, TPU, Trainium, Maia, MAI Thinking One, Cerebras ## Video description <p data-prosemirror-content-type="node" data-prosemirror-node-name="paragraph" data-prosemirror-node-block="true" data-pm-slice="1 1 []">When we last spoke to Brannin McBee, the co-founder and chief development officer of cloud company CoreWeave, his business was not yet public and sourcing GPUs was a key constraint on growth. But three years later, things look pretty different. CoreWeave IPOed and has been raising money in the bond market too, as well as signing more deals with chipmaker Nvidia. In fact, investors have basically been throwing money at all-things-AI. But there are persistent bottlenecks to further growth. Chip supply is still scarce, but so are transformers and electricity. In this episode, we catch up with Brannin on everything he's seeing in the market for compute right now, including leases, Nvidia's new Vera Rubin systems, demand for training versus inference, and the possibility of standardizing the market for compute.</p> <p data-prosemirror-content-type="node" data-prosemirror-node-name="paragraph" data-prosemirror-node-block="true">Read more:<br data-prosemirror-content-type="node" data-prosemirror-node-name="hardBreak" data-prosemirror-node-inline="true"><a href="https://www.bloomberg.com/news/articles/2026-06-05/trump-officials-worry-us-loophole-let-chinese-firms-buy-nvidia-blackwells?utm_medium=referral&utm_source=podcast&utm_campaign=odd_lots&utm_content=article" data-prosemirror-content-type="mark" data-prosemirror-mark-name="link">Trump Officials Worry US Loophole Let Chinese Firms Buy Nvidia Blackwell Chips</a><br data-prosemirror-content-type="node" data-prosemirror-node-name="hardBreak" data-prosemirror-node-inline="true"><a href="https://www.bloomberg.com/news/articles/2026-06-03/broadcom-outlook-disappoints-investors-seeking-ai-fueled-blowout?utm_medium=referral&utm_source=podcast&utm_campaign=odd_lots&utm_content=article" data-prosemirror-content-type="mark" data-prosemirror-mark-name="link">Broadcom Slides Most Since January 2025 on AI Outlook Miss</a></p> <p data-prosemirror-content-type="node" data-prosemirror-node-name="paragraph" data-prosemirror-node-block="true">Only <a href="http://Bloomberg.com" data-inline-card="" data-card-data="">http://Bloomberg.com</a> subscribers can get the Odd Lots newsletter in their inbox each week, plus unlimited access to the site and app. Subscribe at&nbsp; <a href="https://www.bloomberg.com/subscriptions/oddlots?in_source=oddlotspodcast" data-prosemirror-content-type="mark" data-prosemirror-mark-name="link">bloomberg.com/subscriptions/oddlots</a></p> <p data-prosemirror-content-type="node" data-prosemirror-node-name="paragraph" data-prosemirror-node-block="true"><a href="http://bloomberg.com/subscriptions/oddlots" data-prosemirror-content-type="mark" data-prosemirror-mark-name="link">Subscribe to the Odd Lots Newsletter</a><br data-prosemirror-content-type="node" data-prosemirror-node-name="hardBreak" data-prosemirror-node-inline="true"><strong data-prosemirror-content-type="mark" data-prosemirror-mark-name="strong">Join the conversation:</strong> <a href="https://discord.gg/oddlots" data-prosemirror-content-type="mark" data-prosemirror-mark-name="link">discord.gg/oddlots</a></p><p>See <a href="https://omnystudio.com/listener">omnystudio.com/listener</a> for privacy information.</p> [Source](https://podtrac.com/pts/redirect.mp3/tracking.swap.fm/track/UVBrz8bN8aM2Xe47PEPu/traffic.omny.fm/d/clips/e73c998e-6e60-432f-8610-ae210140c5b1/8a94442e-5a74-4fa2-8b8d-ae27003a8d6b/450514c3-2648-49a8-9794-b46000d1738a/audio.mp3?in_playlist=982f5071-765c-403d-969d-ae27003a8d83)

[6:05] Brannin McBee: 没有放缓，只有对顶级算力的无休止需求。 [6:05] Brannin McBee: 如果排除中国市场，在全球排名前十的顶级 AI 实验室中，有 9 家都在使用 CoreWeave 的基础设施来支撑其核心业务。 [10:00] Brannin McBee: 在实际运营中，包括 A100 在内的老旧集群至今依然在全负荷运转，而 H100 的服役生命周期被重新评估为完全可以延续 6 至 8 年。 [12:45] Brannin McBee: 机构绕过通用 AI 实验室直接掌控算力流水线，标志着企业客户已经开始将算力作为独立战略资源进行直接采购。 [16:34] Brannin McBee: 各大实验室主动寻求签订长达 5 年期的“照付不议”（Take-or-pay）合同。 [20:00] Brannin McBee: 平台对 Ampere、Hopper、Blackwell 等各代算力均保持着极其庞大的需求。前沿 AI 实验室通常是第一批排队锁定最新一代架构的买家，而大型企业客户则往往更侧重于当前成熟商用的 Hopper 和 Blackwell 架构。 [25:00] Brannin McBee: 在其平台上超过 50% 的算力利用率实际上来自于推理工作负载。 [25:00] Brannin McBee: 土地本身并非制约数据中心扩张的核心瓶颈，真正的最大瓶颈是所谓的“带电厂房”（Powered Shell）。 [30:00] Brannin McBee: 签约承诺在 2030 年交付电力，与真正把电力转化为可产生账单的 GPU 运行工时（billable GPU hours），两者之间隔着巨大的执行力鸿沟。 [32:15] Brannin McBee: 公司当前的可用算力绝非“未售出产能”，而是“尚未分配的产能”。 [42:47] Brannin McBee: 短期内绝对不可能，长期也只能打一个‘也许’的问号。 [50:00] Brannin McBee: 在短期甚至中期内，根本看不到任何能够彻底填补这一需求缺口的解决方案。

[0:00] Tracy Alloway: An AI consultant told Axios that one of their clients recently spent half a billion dollars in a single month after failing to put usage limits on... [5:00] Brannin McBee: Users are going to need to find the way to use the appropriate model for their prompts... But that just further enables this concept that infrastructure is going to be used longer. [22:30] Brannin McBee: The client isn't asking for anything but NVIDIA infrastructure... people are still using NVIDIA infrastructure. They're committing to NVIDIA infrastructure for five plus year contracts in these billion, multi-billion dollar commitments because they know that that is going to be a critical part of how they scale their business. [35:45] Brannin McBee: You're able to pair this like five year take or pay contract to an amortization schedule on the debt. And you have the revenue come into the box, pay down the amortization schedule, pay down the operating costs of the data center... [50:00] Joe Weisenthal: Even if you're not that interested in AI, this is a really interesting market structure story... It's basically the creation of a brand new market and poses all these interesting philosophical questions about how you do that.

[0:00] Whatever your goal, trade show giveaways, client gifts, or team gear, 4imprint has the promo products to match. With thousands of options, from apparel and drinkware to tech and totes, it's easy to find the right fit for your brand and budget, with standout choices at every price point. And with their 360-degree guarantee, you can be 4imprint certain your order will show up just right, right on time. Explore more at 4imprint.com. 4imprint. 4certain. So there's a lot of noise about AI. But time's too tight for more promises. So let's talk about results. At IBM, we work with our employees to integrate technology right into the systems they need. Now, a global workforce of 300,000 can use AI to fill their HR questions, resolving 94% of common questions. Not noise. Proof of how we can help companies get smarter by putting AI where it actually pays off, deep in the work that moves the business. Let's create smarter business. IBM. Wise is the smart way to manage the currencies you need around the globe. When you send money abroad using your bank, you could get hit with hidden fees and exchange rate markups. There's a better way. Try Wise. Wise uses the exchange rate you usually find on Google with no unwelcome surprises Plus most transfers happen in under 20 seconds which means your money arrives in less time than you been listening to me It simple and free to sign up when you download the WISE app. Be smart, get wise. T's and C's apply. Bloomberg Audio Studios. Podcasts, radio, news. Hello and welcome to another episode of the Odd Lots podcast. I'm Joe Weisenthal. And I'm Tracy Alloway. Tracy, I'm envisioning this future where like we have to do a state of the sort of AI inference market episode like once a month, you know, Like where it's like things are moving so rapidly and there's so much change either in terms of what models are using or what they're being used for, et cetera, that in the same way we would do like, you know, the occasional regular stock market episode or whatever, we would just do, OK, what are we seeing right now in inference trends? Because it just feels like the moment we do an episode a few weeks later it may be out of date We should just bite the bullet and do a weekly episode Transform lots more into a market update on compute We could do inference I don know We have to workshop Odd inference? No. No. We'd have to. But anyway. Lots of inference. Lots of inference. This is like the story of the moment. And we know that, you know, a couple years ago, everyone was sort of dabbling around with various things and experimenting and using AI. I had like, oh, like write a poem for me about this, et cetera. That phase of AI is long over. And we know that companies specifically are spending a ton on compute, so much so that CFOs around the world are getting sticker shock about their compute budgets. And there was even a headline of like Uber saying like, OK, like fifteen hundred dollars of max per employee, like don't spend more than that in a month on tokens. So like this is a very fast moving area. Yeah, you're starting to get headlines about, I guess, a corporate reckoning with AI as more people experiment and spend money on it. The Uber headline that you mentioned, apparently Uber burned through its entire 2026 AI budget in four months, basically. And like, what's more important is the COO was actually asking whether or not that was worth it like whether they saw productivity gains or whatever as a result of that the other very amusing headline that i saw and it was citing an unnamed source it's from axios so you know oh yeah not entirely sure it's true but reportedly a great headline it was a great headline an ai consultant told axios that one of their clients recently spent half a billion dollars in a single month after failing to put usage limits on yeah it's Because everyone is like, oh, I just have a simple question. I want to look up our guest's title. I'm going to use the most advanced model to do that, et cetera. I have a theory, and we will get into this with our guests, that one of the things that will – and we've talked about this with Goldman's Marco Argenti. But one of the things I predict is that companies are like clearly – they're going to keep using it more and more would be my guess. but there were probably a lot of investment made in sort of like optimal model routing because some models are like a hundredth per query of what a frontier model is probably a lot of people don't know like what is the sort of like efficient frontier model usage and so actually routing the query to the sort of most uh efficient model i have a feeling we're going to see a lot of investment in that area specifically well there's also just the question of whether or not the models get cheaper overall as they advance right and we have a lot of investment in that area [5:00] I've seen some, I think NVIDIA has a new system or chip out or something that is supposed to reduce token usage. We can get into that as well. And, you know, we did that live episode recently with Ian Dunning of Hudson River Trading, and he said a lot of interesting things in that. But one of the things he said is that the scarcity is increasingly like just the real estate component. finding a suitable place to plug in your GPUs, at least from his perspective right now, is as much, if not more so of a challenge than securing GPUs themselves. Which is different to what it was like three years ago. Yeah. Yeah. So just like where you plug it in, we know there's all the like the anti-data center politics out there. So it's like, yeah, we got to take the pulse of this market. All right. Consider this our inference update. Yeah. Well, I'm really excited to say we really do have the perfect guest. Someone we spoke to truly feels like eons ago. I think the first thing we ever connected with this company, they've always had a lot of chips. But I think the first time we ever linked up with this company was still in the era where people were excited about NVIDIA chips being used for crypto mining and stuff like that. But we are now in this very different era. And this is truly one of the companies of the moment And that is of course CoreWeave one of the so NeoClouds offering both training and inference services for all sorts of different AI workloads I'm very excited to say back on the show, we have Brandon McBee, CoreWeave's co-founder and chief development officer. So, Brandon, thank you so much for coming on OddLots. Appreciate being invited back, guys. And that was a fantastic intro. We look forward to hitting these topics today. All right. Here's my question. So, we know that at the tail end of last year. And then in the first quarter of this year, everyone started using cloud code and just this clearly a key inflection moment for sort of like overall AI demand. And then we get into Q2 and suddenly the CFO is like, oh my gosh, we're spending this much on inference. We got to like figure things out just straight up. Like in the last month, whatever, do you see any signs of that happening yet of these companies, which are all like still AI, eager AI adopters trying to get a little bit of a handle and maybe slowing the rate of the rate of growth? Is that happening yet Yeah I think you see headlines there that there are surprises of spend etc I say our interpretation of it is entirely look at the authentic and foundational demand that is out there All we're really doing is talking about how much consumption there is of AI and use for it. And I think that that was a real question in the market 12, 18, 24 months ago is, will there be demand for AI? Where is this inference demand that everyone's been talking about? And I think you're absolutely correct. January or so with this kind of like next group of models that were coming out everyone all of a sudden and all at once said this is what we've needed. Like this is the real product breakthrough. But I think it's worth keeping in mind that product breakthrough was like for a limited set of people at the end of the day, right? We're talking like coding professionals, some finance professionals, but it's a relatively small group of people that are using infrastructure at this enormous scale. And so where we see this moving towards next is broader enterprise use, like likely not seeing this whole token mapping approach And I think that that is unsustainable But do we see adoption in other sectors and how this can continue to spread out Absolutely I mean you know on our end I think we have 10 over billion clients at this point And our financial services client backlog is in the tens of billions of dollars at this point And so we're now talking about things outside of AI labs, outside of hyperscalers. And look, as you guys know, we support nine of the top 10 AI labs on the planet. If you exclude China and everything that's going on over there, like we have a lot of visibility into what people are doing. And we're not seeing any pullback on what they're doing on inference today. If anything, it just remains this unrelenting demand for access to the best technology solution in the market for running artificial intelligence. And that's core weak solution in the market. Wait, say more about the customer mix now versus, say, three years ago. So you have hyperscalers, you've got startups, you've got various businesses. How has that, I guess, composition shifted over time? Yeah, it's. [10:00] It's shifted enormously towards a more diverse customer base, right? We got a lot of flack for this in our IPO, right? People were noting that we only had a handful of large clients, that our clients were just the hyperscalers and AI Lab or two. And I think that we have made tremendous progress in driving diversification. So I'd say it's probably across three buckets today, right? We have hyperscaler clients who continue to grow with us. We have AI lab clients. As I said, nine of the top 10 AI labs on the planet choose CoreWeave. And then we have this enterprise base. And the enterprise base just doesn't grab as many headlines as you would expect. This is not these massive multibillion dollar contracts that are being signed. But I think in Q4 alone, we added twice as many logos to our client base as we had ever done versus any previous core. And that enterprise base is the one that's growing so much. And there was a point you guys hit on in the intro that I think is really worth acknowledging And it was this concept of model routing and the idea that like not everyone needs just the latest model that it's different types of models. I can hit different use cases. And this is something we've been talking about for a while, right? As it relates to the infrastructure side of things as well, right? Because you don't need that latest model for everything. And accordingly, you don't need the latest piece of infrastructure to support every single inference or training query that's out there. You can kind of conceptualize this matrix of different sizes of workloads relative to different sizes of GPUs. And all of a sudden that tells you, my God, like H100s could last six, seven, eight years. A100s are going to last longer. And it totally changes the entire conversation around depreciable life of infrastructure as that was a really popular topic during 2025. People were saying like, oh, this stuff will last two years. It's worth zero afterwards. And like, we've never seen any semblance of that because of the point you guys are accurately making, which is users are going to need to find the way to use the appropriate model for their prompts And that be solved by ModelRabbit to your point But that just further enables this concept that infrastructure is gonna be used longer. And we see that every day in our portfolio, extending all the way back to A100s. I just wanna ask a specific question about the broadening out of the customer base. And you mentioned, for example, financial services clients. When you talk about, say, a financial services client as being distinct client from one of the major AI labs, does that mean what you're saying? So it's like I'm just making it up. Let's just say I don't know if these relationships exist. Let's say a city group has an enterprise license with an anthropic. Does that count as anthropic as a customer or city as a customer? And when you talk about this broadening out, are there essentially more types of entities who are building some type of model, not necessarily an LLM per se, but some type of internal house specific model from which they want to run inference? It's a great question. The scenario you presented anthropically be our client. OK, got it. So what I highlighting I want to correct a number I said earlier are financial service clients And this is direct to those financial services they approaching billion in backlog So this would be a good example of this and that something we made recently is with Jane Street That's not Jane Street coming through OpenAI or Anthropik to get to us. That is Jane Street coming directly to us and using our platform. For a model that they're building. So it's a Jane Street. Or inference, right? It's training. No, no, no. I'm not setting aside training, but it would be inference of a model that is Jane Street's model of something rather than Jane Street's contract and enterprise relationship with one of the major labs. At the end of the day, we don't know what exact workloads these entities are running, especially for entities like Jane Street. I would imagine that's highly secretive. But the point I would say is more that this is not them coming through an AI lab to us. they are interfacing with and managing the infrastructure directly on our platform. And that's a really important distinction as we grow this diversified client base. And again, I think that we've just done a wonderful job of executing on that over the past year. [15:00] Whether you're planning a big tech event, launching a new campaign, or just stocking up on team gear, finding the right promotional products makes all the difference. 4imprint offers thousands of options. From on-trend apparel and premium drinkware to tech, totes, and giveaways, so you can find the right fit for any audience, purpose, or budget. You can customize it all. Your logo, your message, your look, and many items come with no setup charge to help you save. And if you're really watching the bottom line, you'll find standout choices at every price point so you can make a real impact while staying on budget. Plus, you'll get expert help, fast turnaround times, and their 360-degree guarantee. So you can be 4imprint certain your order will arrive on time and look exactly right. Whatever your goal, 4imprint makes it easy to find your perfect promo match. Explore the possibilities today at 4imprint.com. 4imprint. 4certain. buying the dip, manually sweeping idle cash, putting on a hedge. On public, you can now create AI agents that handle all these tasks on your behalf. Just describe what you want to do in plain English. Like, if the VIX hits 25, buy a put option on the S&P 500. Or, if my cash balance goes above $20,000, move the excess into my direct index. You approve of the workflow, and your agent handles the rest. Monitoring the market, watching for your conditions, and executing your strategies exactly as defined. An investing platform driven by your intent, not just your clicks. You can also get full read and write access to your account via the public API. Go to public.com slash market and fund your account in five minutes or less. That's public.com slash market. Paid for by Public Investing. Brokered services by Open to the Public Investing, Inc., Member FINRA, and SIPC. Advisory services by Public Advisors, LLC, SEC Registered Advisor. Complete disclosures available at public.com slash disclosures. The thing about AI for business, it may not automatically fit the way your business works. At IBM, we've seen this firsthand. But by embedding AI across HR IT and procurement processes we reduced costs by millions slash repetitive tasks and freed thousands of hours for strategic work Now we helping companies get smarter by putting AI where it actually pays off deep in the work that moves the business. Let's create smarter business, IBM. As you've talked about including in earnings releases, and as you can just tell from these huge token budgets, inference demand is booming, but model training is still important. But in addition to model training, so you say, okay, if you have a pie chart, the part that's inference is getting bigger, but I assume the training is also growing as well. But I'm curious from the perspective of like, say, the AI labs, when they think about growth, has there been a subtle shift from investing to push the pure model frontier, having the absolute best state-of-the-art model versus investing in, say, better harnesses? Because a big reason we're excited and all talking about AI right now is really the excitement that happened over with Cloud Code in the final quarter of 2025 And it like oh this harness has really unlocked a bunch of capabilities Has there been a shift in investment from rather than just the purest most advanced model to let invest more in tooling capacity and other things that allow companies and clients to get more juice from an advanced model I don't think that we're exposed to that decision making with the AI labs as counterparties to us. The observation I would make in a behavior change for the AI labs is they want access to more infrastructure for longer duration. And I'll qualify it a little bit, which is a year or two years ago, we were signing three-year committed contracts. The type of contracts we sign are basically like take-or-pay contracts, which is the best way to finance the infrastructure that we are building for our clients. last year, it was four year contracts, right? They were saying we want explicit access to Hopper for four years or Blackwell for four years. Now they're coming and saying, well, actually we want it for five years. We don't want any interruption of use. We'll commit to the exact same economics throughout the full duration of the contract. You can't upgrade or change the infrastructure within it. You cannot cancel the contract. We [20:00] want it for five years and they want it at more scale, right? The deployments are getting larger and larger. So that's probably the best characterization we can offer on decision-making that AI labs are going through right now as they look from an infrastructure perspective. It absolutely seems like tooling is important, but scaling laws are still holding, right? Like your ability to advance your frontier model through accessing more infrastructure at scale holds. and that will hold through Vera Rubin, we expect. And seemingly it's not stopping anytime soon. Oh yeah, what's the deal with Vera Rubin? Can you explain that to us? Which aspect of it? What is it? Oh, yeah, yeah, yeah, yeah. Basically. Yeah, so it's just NVIDIA's next architecture that's coming out, right? Like the current architecture that we're deploying today is Blackwell. Blackwell comes, we deploy predominantly in a NBL 72 configuration, which was an entire architecture change. from deployment, right? If you recall, Hopper came before Blackwell. Hopper, you could deploy these 42U racks, which was typically like eight GPUs in a server case. You would take it plug it in largely air as well right We ran some liquid cooling just so we understood the requirements of liquid cooling because Blackwell for our deployments is overwhelmingly liquid in its deployment configuration. And instead of eight GPUs in a 42U configuration, it's in this larger 72 GPU rack. It's like an entire chassis that's being brought in, And it just looks entirely different in the data center. It's like this giant tower thing that you've seen in pictures floating around on X. So Verirubin will be the next architecture that comes out. And we've started receiving testing racks for Verirubin. But the basic idea is like the new configuration makes the whole system more efficient, like more tokens per energy use and that sort of thing. Yes. Yeah. I think that's kind of where you're getting to with it. But that doesn't necessarily mean, going back to the point earlier, that everyone only wants the latest generation of GPU. Right. We have massive demand for Ampere, Hopper, Blackwell, etc. And it just varies by use case model and type of client as well Like I would qualify that AI labs are probably the ones who are lining up first to secure access to the latest generation GPUs whereas enterprise clients might be very focused on current generation like Hopper and Blackwell right now I'm going to be honest for a second. You know, I try to keep up on a lot of things AI related. I really do. And every single day, the one thing I do not keep like in my mind, if you asked me, like I liked it in the old days when it was like 186, 286, 386, 486, Pentium, and then like Pentium 2, et cetera. There was just this numerical sequence that I could keep track of in my head. And so if someone asked me like Joe, like Vera Rubin, Hopper, Blackwell, what was the sequence? I'd be like, I got to be honest with you. I don't exactly remember. And I will prioritize that at some point. But speaking of silicon, so yesterday Microsoft came out with a big, they're really, they wanna be in the game too. They don't wanna just be connected to the labs. They wanna have advanced models too. And apparently it's a good model. And they announced the MAI Thinking One model but they said it optimized on the Maya 200 chip which is their own chip And this is a thing which is even again going back to our recent conversation we had, even a place like Hudson River Trading is thinking about getting into the customized hardware game. How much juice for the squeeze is there of aligning the model with custom silicon from your vantage point what what we could offer is what we hear from our clients yeah on that and it's important to keep in mind we can run any type of silicon on our platform okay right we are entirely customer-led in what we build like we don't go commit to capex and speculatively hope people come and use infrastructure right like we wait until a client says we want you to go do this specific build, here's what we want it to look like. And then we go commit to that CapEx, right? It's more like a success-based CapEx approach. And the client isn't asking for anything but NVIDIA infrastructure. And I think a large contributor to that is, I mean, they built this incredible ecosystem around their chipset. They have been dedicated to that for, I think, over 15 years at this point through the CUDA architecture. [25:00] And NVIDIA, from what we hear from our clients, that platform just remains the most efficient, the most scalable, the most reliable set of infrastructure that is in the market. So I think others, there's always been, I mean, if you think over the past few years, there's always been talk like, what is it? But yeah, this other silicon and these other chips. And at the end of the day, like people are still using NVIDIA infrastructure. They're committing to NVIDIA infrastructure for five plus year contracts in these billion, multi-billion dollar commitments because they know that that is going to be a critical part of how they scale their business. We really don't see demands on a material basis for anything but that NVIDIA compute. And that's what we are building today. Obviously, just to push back on this a little bit, and I'm not really in any position to push back. I can only relay what past guests have said in my own reading. So what one of our guests said is that absolutely NVIDIA has the lock on model training that if you want to train a model that yes NVIDIA chips are the only game in town But then for inference there really his view this is Ian Dunning again his view is there really were options. And then, of course, we had someone who was much more biased. We interviewed the CEO of Cerebros, the company that makes the gigantic plate and, or sorry, the gigantic chip. The gigantic plate. And of course, he did. But I mean, of course, he was going to say, yeah, the CUDA moat is vastly overrated for inference. It barely exists. Now, of course, of course, he's going to say that. So like, you know, he's in a competitor. But we've also heard it from a user of inference. And intuitively, it makes sense. Like training is very complicated and all that stuff. But what you're saying is that from the customer standpoint, you see the demand for NVIDIA on both the training and the inference as being steady and that you perceive that advantage to be consistent through both aspects? So I believe in our last quarterly report, our CEO might qualify that inference workloads represent well in excess of 50% of infrastructure utilization on our platform. Okay. It's the exact same infrastructure that we use for training as well right Going back to my comment of like it very fungible between those different types of workloads Those customers are choosing NVIDIA to work with on inference I think what you going to see is people will want to try at small scale other types of silicon, but the reliable, proven and remains from our perspective, most efficient infrastructure to use is NVIDIA today. Does that change over time? Who really knows? But I think we've seen NVIDIA battling this concept for years and every year they show up and like they remain the de facto choice for AI infrastructure. I think we're going to be one of the first people in the market to see it because that will be a tone shift change from our clients asking us to run something else. That hasn't happened. Okay. So have the constraints on your business changed at all? So three years ago, we were talking about GPUs and how hard they were to actually get. I imagine securing GPUs is still competitive, to say the least. But are you seeing other constraints emerge, like Joe mentioned in the intro, just land usage just places to actually build data centers Land usage specifically I wouldn say is as much of a concern Having a powered shell is the bottleneck today. And let me qualify a powered shell. Powered shell is effectively an empty data center that is energized. It has all the power and associated components. I can come into it and deliver electrons into a rack. has the cooling system built within it. Like it has a whole thing, right? A powered shell is the industry term for it. That is the bottleneck because of all of the supply chains that come into that, right? Like not only do you have electricity, do you have the land, et cetera, but you have the backup battery supplies, you have the transformers, you have personnel, right? Let's just think about the electricians for these sites and getting the accreditation on the electrician side to be able to participate in these bills. I mean, I think it's a five-year plus apprenticeship to be able to go through that program, right? We can't just make new electricians leveraging a supply chain, right? Like that's a trade that you can't really scale efficiently. So that is absolutely the bottleneck for us. And I think our peer set, [30:00] that's out there right now. Access to chips. I think we have a phenomenal relationship with NVIDIA where we've just proven to be the best operator of this infrastructure on the planet. You know, a bottleneck that existed for us previously, I think, was access to financing. Yeah. Doesn't seem to be an issue anymore. I would agree with that broadly, but that's years of work and execution that has delivered that ability for us. I mean, year to date, we've raised over $21 billion of financing for our business. Like you don't get to do that and just go from zero to 20, right? Out of nowhere. And I think that's largely driven by our track record of execution, right? Our investors, our creditors can see this deep set of experience over the years of consistently delivering on these builds. I mean, we have over a gigawatt in active power at this point, right? Like a gigawatt like at the data center level with GPUs you know delivering into clients And I think that there has been kind of a misunderstanding in the market where people are conflating the concept that like you know something on paper is the equivalent to being physically done and delivered. And all I can say is there's an enormous gap between signing for power delivery in 2030 versus actually delivering that into billable GPU hours. And that gap of execution is what has driven down our cost of capital so aggressively. That gap is where our business sits and why it's been so successful. I mean, that's the secret sauce is our ability to take these data center deployments and these customer relationships and deliver billable GPU hours into them. You know, speaking of financing, I just want to say, you know, during last year, like maybe six months ago, that might have been the sort of near peak of the Michael Burry inspired. These chips are like in the last two years stuff. And one of the viral charts that you would see on Twitter was the core we've CDS chart. Those have come way in So it is it is you know I haven seen those charts in a while yeah that right that the thing about cds no one ever posts charts of credit default swaps when they coming in they people love to post them when they're blowing out they have um come in so you know that does speak to some of this point about these anxieties uh having been alleviated at least uh somewhat since the start of the year you know it occurred to me like we're talking about credit default swaps So we're talking about financing. I'm sort of gearing up to write a big thing maybe, but I'm writing it in my head currently, that there really are a lot of analogies between the business of data centers and the business of banking. And one of the things in banking, as we all learned from SVB, was the risk of industry and depositor concentration. That if you have all your depositors are either in like one depositor gets too big or all your depositors are in the same industry, then you have this risk of like correlated withdrawals. And that's what obviously did in SVB. When you think about planning and you think about, OK, here's an investment, et cetera. How much does this come up sort of like thinking about I guess tenant diversification as something that you think about in your multi planning It a critical aspect of it right as i said earlier too like this was a key criticism of us coming into our idea last year right where we had that customer concentration in our revenue and we have made enormous progress there and i think the best way to think about it is we could take all of our unallocated capacity and i say that very specifically it's not unsold capacity implying that there's no demand for it It's unallocated. There's intense demand for it. We're figuring out where it should go. And that customer piece of it, I think, honestly, we could allocate all that capacity to single-name clients. There is a pretty significant number of single-name clients we can go allocated out into. But I don't think that is the business we are supposed to be building here. I think the business we are supposed to be building is a diversified cloud that is supporting the leading AI consumers and producers on the planet. I don't think we're supposed to be supporting just one or two companies. [35:00] Thank you. watching the bottom line, you'll find standout choices at every price point so you can make a real impact while staying on budget. Plus, you'll get expert help, fast turnaround times, and their 360-degree guarantee so you can be 4imprint certain your order will arrive on time and look exactly right. Whatever your goal, 4imprint makes it easy to find your perfect promo match. Explore the possibilities today at 4imprint.com. 4imprint. 4certain. Support for this show comes from public.com. If you're actively involved in your portfolio, you probably catch yourself repeating the same actions, buying the dip, manually sweeping idle cash, putting on a hedge. On public, you can now create AI agents that handle all these tasks on your behalf Just describe what you want to do in plain English Like if the VIX hits 25 buy a put option on the S 500 Or if my cash balance goes above $20,000, move the excess into my direct index. You approve the workflow and your agent handles the rest. Monitoring the market, watching for your conditions, and executing your strategies exactly as defined. An investing platform driven by your intent, not just your clicks. You can also get full read and write access to your account via the public API. Go to public.com slash market and fund your account in five minutes or less. That's public.com slash market. Paid for by Public Investing. Brokered services by Open to the Public Investing, Inc., member FINRA, and SIPC. Advisory services by Public Advisors, LLC, SEC Registered Advisor. Complete disclosures available at public.com slash disclosures. The thing about AI for business, it may not automatically fit the way your business works. At IBM, we've seen this firsthand. But by embedding AI across HR, IT, and procurement processes, we've reduced costs by millions, slash repetitive tasks, and freed thousands of hours for strategic work Now we helping companies get smarter by putting AI where it actually pays off deep in the work that moves the business Let create smarter business IBM When it comes to financing, can you say a little bit more about what changed to make the market more comfortable with this? Because this is the big story in markets, just how much AI is now being issued through the corporate bond market. The equity market, as we know, is basically all big tech at the moment. What changed on the part of investors? Was it just pure return and performance or were there, I guess, efforts to make the contracts more robust or increase visibility into demand and that sort of thing? Seeing the inference aspect of it really emboldened investors. But that was really just January or maybe late Q4, where you started seeing there's just massive inflection of demand driven by inference. For us, right, it's tough for me to speak about other companies, but for us, why have we been underwritten at such scale and at a decreasing cost of capital I think it goes back to that track record of execution right It just the market has watched us execute and watch us deliver on these contracts And the way tell me if I going into too much detail here but the way that we finance our business, you kind of break it into two broad buckets, right? You have parent co-financing and asset co-financing. And asset co-financing is where all of the GPUs get financed, right? It's where all of our client contracts sit. And we can take these financings and put them into SPDs, or we'll just call it a box, so to say. And you... Sorry, I'm sorry. Keep going. Mentioning boxes is dangerous. Boxes on lots of connotations, but keep going. We put them into SPDs. And these these STVs, they have the infrastructure, they have the data center costs and they have the the debt agreements within them. And so you're able to pair this like five year take or pay contract to an amortization schedule on the debt. And you have the revenue come into the box, pay down the amortization schedule, pay down the operating costs of the data center. and it still contributes a, it has a 25% cost. [40:00] contribution margin of profit up to the parent code, right? Like these are highly profitable agreements down in the SPV stack. And so you take that SPV out to the credit market and say, look at this instrument. It's a discrete set of contracts with counterparties, like entities who want to consume GPU compute. You have the data centers within it, et cetera. and you know one of the latest ones we did was uh as we call ddtl4 this was a investment grade rated first of its class no one had done this before for gpu financing non-recourse hpc infrastructure financing and got done it so for plus 225 like that is a phenomenal cost of capital for us and importantly we were able to bring in uh the insurance charge of capital which is a massive Chantra Capital out there that is looking to do allocations into the space. So we're kind of continuously making progress through these different stacks of capital, unlocking access to more and more types of investors. It's why you've seen us move into the convertible note market, into the unsecured market as well, along with taking direct strategic equity investments But for us it really important for the entire investor space to understand this business because this business largely didn't exist before, right? Like people weren't making loans into the hyperscalers to go credit these build-outs, right? It's on core weave, honestly, to be building this path into how do you finance the AI hyperscaler effectively. And I think we've just done a terrific job of it over the past few years. You used to be in a prior lifetime, a trader, right? Yes, I was a commodity trader. So I'm curious, like, you know, there's a lot of interest in and I don't know if it's going to materialize in GPU capacity trading and there's going to be a new contract. We recently interviewed the CEO of Compute Exchange and they're very close to having something listed on the CME. from your perspective, because I don't have a view on this yet. On the one, like you see like, okay, a big AI company does a five-year contract. As you say, the duration is lengthening. We're going to lock this in. I don know like what the need is for tradable compute in that environment et cetera What your guess Like do you anticipate that there will be a sufficient ecology of hedgers and speculators such that there will be a liquid market for tradable compute i think it a it a very much a timeline question that's out there short term no let me offer why no short term and then I'd say maybe in the long term. And it all comes back to fungibility, right? If you think about gold, gold is defined by its chemical composition, right? And there's no question of what is gold and not gold, et cetera. Compute really isn't, right? And especially GPU compute. GPU compute today is not fungible. And I think that this is well understood by our client base, by our suppliers, by third-party consultants like Sini Analysis. And it's this idea that an H100 deployed in one cloud doesn't have the same performance of an H100 deployed in another cloud. And the metrics that people use are things like good put or model flop utilization, MFUs, and there are these measurements of how much more performant is the exact same GPU by the versus another gpu deployed in another facility and so in order for something to be commoditized it has to be fungible right otherwise there just too much you know murkiness and there isn't like uh an exact data plan in there can i push on that a little bit further so yeah i mean i think that seems like a reasonable view is the non-fungibility related to configuration of like how they literally like the configuration of the GPUs within physically? Like, what is it? Is it about power? I mean, I think they all have like, you know, there are plenty of places that will say, you know, we have nine nines or however many nines you need in your industry or whatever. What is it in your view that would cause significant changes in the performance of an H100 in one cloud versus another? It could be in some part configuration, We build everything to DGX reference spec, which is the most, outlined by NVIDIA, it's the most performant way to build, operate, and deliver GPUs. But the rest of it, honestly, is just how you operate the GPUs. And that is the core weave software stack that is... [45:00] keep these GPUs online, right? Like what happens if a GPU fails? Can you predict if a GPU is about to fail and swap in other infrastructure so that the client doesn't have downtime on that component? And there's an immense suite of software solutions and infrastructure management solutions that we have built to have the best good put, to have the best MFUs in the industry. And none of that is off the shelf, right? And so I wouldn't say it comes down to the strict components. That's kind of like a bare minimum starting point, right? Like you have to start in DJX reference spec, but where's differentiation come from there? I mean, that's the core product you're describing right there. By the way, Tracy, I'm just looking up terms of art. Good put measures the fraction of peak hardware performance that the training job can extract. This is according to Google in MFUs, model flops utilization, hardware metric for evaluating real world efficiency of LLM training. So two new terms. I actually hadn't heard of MFUs or Goodput before this. So I just learned two new terms today. We got to create a glossary. A glossary. Yeah, we do. Brandon when Joe asked you that question about compute markets earlier you said it was a timeline question which in my mind implies that it inevitable Like it just a question of how long it takes But then when you describe the fungibility problem it seems like this is an actual issue that will be very difficult to solve Yes, I think that characterization is absolutely correct, right? Like if you just take general commodity theory, and I traded natural gas, electricity, agriculture products for over a decade, like it suggests that it should become that at some point. But what is the reality today? The reality is this stuff isn't getting easier to operate, right? We've moved from these kind of relatively simple 42U air-cold racks of Hopper to these immensely complex Blackwell deployments moving into Vera Rubin following that. It's not getting easier to build, operate, provision, deliver these GPUs. It's getting more difficult. And I think until it starts becoming easier, you don't really have a path to commoditization. You will have to continue to prioritize working with the world and world operators of infrastructure That where we sit First of all this is helpful And I like that we getting multiple perspectives because I do think this is gonna be like one of the big questions for financial markets because let's say if they took off, then you could imagine that might even improve financing conditions because then the lender can hedge against. Yeah. So like there would probably be some good things for the industry if this took off. So I appreciate it's good to have your perspective on this. Why is it, you know, I'm an inference user, by the way. So I made a little machine learning model in one of my hobby projects. And I provide inference over at havelock.ai, or I'm a user of inference or whatever. I have a model, whatever. Why is it that I... It'd be impressive if you were providing inference. I'm trying to, I guess I'm a consumer of inference. I use a, anyway, why is it that I'm actually very easily able to get, now not a huge allocation of like GPU access? So I was like how do I train this model It a model called BERT that Google released in 2018 or 2019 I fine tuned it for my purposes And then literally using Cloud Code I was able to in 10 minutes sign up I started using this company called Modal and I was able to start training a model I was surprised that there was like and it didn cost me very much and I have like no volume. But nonetheless, evidently, there was a little GPU capacity out there that I could get, and it cost me like $5 or something for the whole thing. Given what you always hear about like a utilization is slammed, why is it actually not that hard to find GPU capacity for someone like myself? You know, I think it's the scale difference right there. Finding ones or tens of GPUs, I think that's way more accessible out there. Okay. Our clients are focused on the hundreds of thousands of GPUs. I'm not there yet, but I'm not there yet. Not yet. I'm sure you'll get there. Yes. And that's where it kind of decommoditizes itself with scale as well, right? Like as you're in the hundreds of thousands component, there's just not that many deployments. It's handfuls of deployments at that size. But getting access to ones of GPUs, I think that there is a lot more ability to go secure that sizing in the market. So Joe and I are heading to Hong Kong very soon, and I expect that AI in China is going to be a big topic of conversation. [50:00] How would you characterize, I guess, the difference between the U.S. and the Chinese market at the moment? I'm sure this is something you think about, even though you don't participate in the Chinese market directly. Yeah, that's like... Tracy's asking for questions. Yeah, that's basically it. It's like questions that we can ask people when we're over there. Yeah, that's likely going to be my response, Tracy, is like we just do not participate in that market. I think that there's opportunity for us to be expanding. As you guys know, we operate in Canada, Europe. I think moving further east makes a lot of sense for us, but we're trying to be very methodical in the way that we expand. So unfortunately, I'm not going to be able to help you with specific questions in that market. But I would imagine you're going to encounter a lot of the same things that you're seeing in the U.S., which is just insatiable, unrelenting demand for AI. And like, we just kind of keep coming back to this. It's like, there is no solution in sight for being able to satiate demand, right? There's just too many supply chain. There's no path to solving demand in the near term or even the medium term frankly You mentioned so Tracy asked you about land use You said that really was an issue But like the first time we talked to you in 2023 or whenever that was there was not a major growing movement of people who are just like anti centers in America Maybe there were a few fringe people, but it was not something that was on the minds of politicians and activists and so forth. And you do see these headlines, you know, about some projects really having been shelved. It was like a big one. Northern Virginia is a huge hotspot for it. And there was a big project that was they pulled the plug on due to some they couldn't get an agreement with the local government. That must affect you. What are you seeing in terms of like your capacity to build? How has it changed specifically in light of or have you seen a change? Would you be able to build faster in a world where this had never become a political hot button issue? I believe it has become that hot button issue. It's something that we're quite proactive about in market. And I think you just kind of go through the checks on the diligence process to make sure you're going through it correctly. I think that there's misconceptions out there. like water usage. Yeah, setting aside the misconception, like setting aside I know setting aside the whole debate about but just in terms of like operationally what it changed for you in terms of your plan No I would say our greatest challenge is still just getting that delivery of our like the construction and all the components and getting everything in there Like that is truly more of the bottleneck that's in the market today. Brandon, thank you so much for coming back on OddLots. We'll have you back next month for another market. No, or at least, or maybe in three years. Not three years. Yeah, not three years, but really. That's an eternity. Yeah, I know. Thank you so much. Thanks, guys. Appreciate it. I'm very excited about whether compute features will take off. I think that's an exciting story. You know, it's not the biggest story in the world, but it is actually a very exciting story. I've said this before. Even if you're not that interested in AI, this is a really interesting market structure story, right? It's basically the creation of a brand new market and poses all these interesting philosophical questions about how you do that. And I thought Brandon point about fungibility I mean that is a real issue And it does seem like it a challenging one to fix at the moment I don't know if it's inevitable in the future, but who knows? No, no. I mean, it makes a lot of sense. This was also Lewis Hart's point that it's like, you know, it's in the word commodity, right? If it's not a commodity, you're not going to get a commodity market for it. And of course, a number of entities are betting that it will be commoditized. But if the if it's true that, like, you know, they're getting more difficult to work, that the technical demands on the inference provider on the data center company are getting greater in order to get the maximum juice. Yeah. Then maybe it doesn't become commoditized. But I think that's like a fascinating question. like if you do see those efficiency improvements and new designs and things like that, you could imagine that like the demand is there for a standardized GPU as well. So I don't know, like I'm really torn. It feels like it could go either way. Well, and even in his answer, he talked about how they configure their own GPUs to a spec largely that NVIDIA itself has come up with. So in theory, like there is aspect [55:00] that everyone can match to. So that's like a really interesting, that's a really interesting question. I also really want to do more on all of these. So Google has TPUs, Amazon has Tranium, Microsoft has its own hardware, maybe even Jane Streets and the Hudson River tradings will have their own hardware. If they're not like, I want to understand better why, right? Because like, they presumably have some reason and they at least like the Microsoft will say, well, this will run better on our customized hardware. I want to understand why that would be, how much difference in performance is there? And then the degree to which demand materializes from users for non-NVIDIA silicon is like a really big question. Yeah. Why custom chips? Yeah. And what can you get out of that if you align model and chip to optimally work together I have no idea but I feel like it an episode I would like to do Yeah we should All right Shall we leave it there in the meantime Let leave it there All right This has been another episode of the All Thoughts Podcast I Tracy Allaway You can follow me at Tracy Allaway And I'm Joe Weisenthal. You can follow me at The Stalwart. Follow our guest, Brandon McBee at Brandon McBee. Follow our producers, Carmen Rodriguez at Carmen Armand, Dashiell Bennett at Dashbot, Kale Brooks at Kale Brooks, and Kevin Lozano at Kevin Lloyd Lozano. And for more Odd Lots content, go to Bloomberg.com slash Odd Lots, or the daily newsletter and all of our episodes. And you can chat about all these topics 24-7 in our Discord, discord.gg slash oddlots. And if you enjoy Oddlots, if you want us to do an episode on custom chips, then please leave us a positive review on your favorite podcast platform. And remember, if you are a Bloomberg subscriber, you can listen to all of our episodes absolutely ad-free. All you need to do is find the Bloomberg channel on Apple Podcasts and follow the instructions there. Thanks for listening. Thank you Music Music Thank you. and sixty degree guarantee you can be for imprint certain your order will show up just right right on time explore more at for imprint dot com for imprint for certain when you're running a business the best days are the ones where priorities stay on track for midsize and large companies risk can affect multiple parts of the organization at once from property and liability to cyber and regulatory challenges at that level managing risk becomes an ongoing discipline We see you next time insurance company and its property and casualty affiliates, Hartford, Connecticut. These days, it seems like AI agents are just about everywhere you turn, every field and every function. But without identity, you can't trust they'll serve your business instead of jeopardizing it. Fortunately, Okta helps you get identity right by securing your AI agent's identities, giving you a single layer of control, a single standard of trust. So whether an AI agent supports a single user or your entire enterprise, with Okta, you'll turn risk into opportunity. Secure every agent. Secure any agent. Okta secures AI.