Cerebras CEO Says Nvidia CUDA Dominance Fades as AI Shift Accelerates

Cerebras CEO称随着AI转型加速，英伟达CUDA主导地位逐渐减弱 Cerebras首席执行官Andrew Feldman指出，英伟达的软件护城河在AI推理中已无关紧要，且在模型训练中正失去优势，这使得晶圆级芯片能够以更低成本绕过GPU内存瓶颈。 Cerebras首席执行官Andrew Feldman做客节目，探讨了公司640亿美元的IPO估值，以及其Wafer-Scale Engine（晶圆级引擎）的技术突破——该芯片绕过内存带宽瓶颈，实现了比传统GPU快15倍的AI推理速度。Feldman详细阐述了GPU token成本与晶圆级芯片效率之间的经济权衡，认为高速执行对于复杂的智能体AI（agentic AI）工作流至关重要。对话还探讨了英伟达CUDA软件护城河的瓦解、企业向开源模型的转向、《芯片法案》（CHIPS Act）下美国本土半导体制造面临的挑战，以及国家安全出口管制议题。 Cerebras 完成重磅 IPO，以 640 亿美元估值募资 55 亿美元，其独特晶圆级芯片架构造就了这一公开上市里程碑。 (0:00) Cerebras Wafer-Scale Engine 的体积是普通单芯片的 58 倍，攻克了长达 75 年的半导体工程难题，在 AI 推理任务上的运行速度比商用 GPU 快 15 倍，特定专用工作负载下最高可快 1000 倍。 (5:00) GPU 虽能以低成本生成慢速 Token，但在高吞吐加速时能耗与成本会急剧攀升；而 Cerebras 架构解耦了高速 Token 生成与失控成本，为智能体（Agentic）多步骤工作流提供关键的复利经济优势。 (10:00) Cerebras 依托台积电 5 纳米工艺，成功避开了 HBM 与 CoWoS 封装等行业制造瓶颈，但目前整个行业正面临带电数据中心机房等物理基础设施的严重短缺制约。 (20:00) 开源模型与闭源模型的性能差距已收窄至仅 3% 至 5%，推动企业转向成本更优的开源方案；同时 NVIDIA CUDA 护城河在推理端毫无用武之地，在前沿大模型训练端的软件市场份额也已缩水约 70%。 (26:40)

Following Cerebras's public debut, CEO Andrew Feldman argues that wafer-scale silicon solves the memory and cost bottlenecks of AI inference, positioning the company to challenge traditional GPU dominance as compute demand shifts toward agentic workloads. Wafer-Scale Architecture and Overcoming Memory Bottlenecks (5:00) Traditional GPUs rely on dense but slow memory, which creates data transfer bottlenecks during model generation. Cerebras bypassed this limitation by building a single chip across an entire silicon wafer, resulting in a design 58 times larger than standard chips. This expansive surface area accommodates vast amounts of ultra-fast memory, allowing Cerebras to run inference roughly 15 times faster than standard commercial GPUs, and up to 50 to 1,000 times faster on specific specialized workloads. Achieving wafer-scale integration required a 10-year development effort, including $500 million spent in the first five years to co-develop new lithography techniques with TSMC, custom packaging, liquid cooling, and proprietary software. The Economics of Speed in AI Inference (10:00) While GPUs produce slow tokens at a low cost, attempting to push GPU clusters to higher generation speeds leads to exponential spikes in power consumption and expense. Feldman asserts that Cerebras decouples speed from runaway operational costs by generating fast tokens with significantly lower power per token. High-speed generation is increasingly critical for agentic AI tasks, where autonomous models carry out complex, multi-step operations like writing software; in these workflows, latency directly dictates economic productivity. Market appetite for speed has already been demonstrated by instances such as Anthropic selling out a premium tier that offered double the token speed for six times the standard price. Supply Bottlenecks, CUDA Moat, and Open Source (20:00) Unlike competitors throttled by shortages in High Bandwidth Memory (HBM), TSMC's CoWoS packaging, and cutting-edge 3-nanometer capacity, Cerebras utilizes mature 5-nanometer nodes and bypasses HBM entirely. Instead, Feldman identifies the primary industry gating factor over the next 15 to 18 months as the physical scarcity of powered data center real estate. Feldman also argues that Nvidia's CUDA software moat is rapidly eroding, claiming CUDA plays no role in inference and pointing out that frontier models like Google's Gemini and Anthropic's Claude train and serve on non-CUDA hardware like TPUs and Trainium. Concurrently, high-performing open-source models—trailing closed options by only 3% to 5% in capability—allow enterprises to avoid paying steep closed-model markups when run on high-speed hardware. Commercial Traction, Financial Strategy, and Geopolitics (40:00) Cerebras went public raising $5.5 billion at an approximate $64 billion valuation, trading around 67 times forward sales. Commercial traction includes a reported $20 billion deal with OpenAI, a deployment deal on AWS Bedrock paired alongside AWS Trainium, and extensive data center operations for Abu Dhabi-based AI champion G42. Feldman addressed past scrutiny regarding G42 and investor 1789 Capital, stating that CFIUS national security reviews were resolved by March 2025 prior to the round. On industrial policy, Feldman notes that domestic semiconductor fabs require $30 to $40 billion, five to six years to construct, and a sustained 20- to 25-year commitment to build supporting ecosystems, while maintaining that strict export controls are necessary even if they exclude firms from foreign markets. As enterprise users encounter high API costs, deployment strategies are expected to bifurcate: routine queries will migrate to cheaper open-source models and standard GPUs, while latency-critical agentic workflows will justify high-performance compute engines like wafer-scale chips.

Cerebras首席执行官Andrew Feldman指出，公司通过晶圆级芯片克服了制约传统GPU数十年的内存瓶颈，在AI推理领域实现数量级提速并摆脱指数级能耗攀升；随着开源模型崛起及主要前沿实验室转向替代芯片，NVIDIA的CUDA生态护城河正面临迅速瓦解。 晶圆级架构对内存瓶颈的物理突破 (5:00) 传统GPU依赖高密度但速度缓慢的内存架构，在向计算核心传输数据时形成严重的传输瓶颈，导致AI逐Token生成内容时延迟高企。Feldman介绍，Cerebras历经10年工程研发（仅前5年即耗资5亿美元），与台积电深度合作重构了光刻、单晶圆供电封装及新型液冷散热系统，制造出面积达普通商业芯片58倍的单一餐盘级晶圆芯片，成功攻克了自1980年代Gene Amdahl的Trilogy项目失败以来困扰半导体行业75年的晶圆级集成难题。 由于在巨大晶圆表面直接铺设超高速内存，Cerebras芯片在AI推理任务中的处理速度比顶尖商业GPU快15倍，在特定计算负载下提速可达50至1000倍。这一技术路径已取得重大商业落地，包括与OpenAI签订的价值逾200亿美元的合作协议，以及与亚马逊AWS达成合作将其系统直接部署至数据中心并集成入AWS Bedrock服务。 Token经济学与Agent工作流的速度红利 (15:00) 针对AI模型在2025至2026年跨越实用门槛后爆发的推理需求，Feldman坚决驳斥了“智能体自动化工作流不需要极致速度”的论调。他类比互联网拨号上网的发展史强调，在代码编写与自动化多步骤业务流程中，速度直接决定产出效率；如果竞品能在20分钟内执行5到10倍的推理工作，响应迟缓的系统就会在商业竞争中落败，速度因此构成了随时间持续复利的经济优势。 在底层生成成本方面，Feldman解释了GPU与晶圆芯片的物理经济学差异：GPU极擅长以低成本低速产出Token，但一旦强行推高生成速度，其功耗与单Token成本将呈指数级激增，如同汽车以极速行驶时油耗骤增；而Cerebras架构则使高速Token生成与失控的电力与硬件成本脱钩。市场上消费者与企业对高速响应存在极高付费意愿，例如Anthropic推出提速2倍但定价6倍的Token服务便迅速售罄。 避开封装瓶颈与电力基础设施制约 (20:00) 当前AI算力行业受到高带宽内存（HBM）、台积电CoWoS先进封装产能及3纳米先进制程的三重严重限制。Feldman说明，Cerebras不采用HBM与CoWoS工艺，且基于成熟的台积电5纳米节点制造，因此完全避开了制约竞争对手的核心代工瓶颈，台积电能够持续满足其晶圆供应需求。 然而，Feldman坦言整个行业正面临极具讽刺意味的硬约束：在突破前沿芯片工程后，业务扩张真正的制约因素变成了具备电力供应的实体数据中心建筑，他预计这种物理电力机房的短缺将在未来15至18个月内持续存在。至于中东地缘紧张对氦气等光刻原材料供应链的潜在扰动，他表示日常化学品管理由台积电全权负责，公司无法控制也未受直接波及。 CUDA护城河松动与开源模型成本优势 (26:40) 针对业内普遍认为NVIDIA依赖其专有软件栈CUDA建立垄断的观点，Feldman提出了尖锐反驳。他断言CUDA在推理环节“毫无立足之地”，开发者仅需敲击约10次键盘修改API指向即可将模型平移至Cerebras硬件。在模型训练端，前沿机构对CUDA的依赖亦大幅萎缩：一年前顶尖模型全系捆绑CUDA，如今Google Gemini完全使用自研TPU，Anthropic的Claude则使用AWS Trainium和TPU搭配GPU，仅OpenAI GPT仍依赖CUDA环境，意味着NVIDIA已流失领先实验室约70%的软件份额。 与此同时，企业正加速从高溢价专有模型转向开源生态。Feldman指出，顶级开源模型（如万亿参数的Kimi K2与Qwen）与闭源模型的性能差距已缩减至3%到5%，但其单位智商成本大幅降低，企业仅需承担芯片折旧与电费。在Cerebras云平台上，Kimi K2的推理速度可达常规平台的10至15倍，进一步加速了企业替换专有模型的趋势。 地缘合规审查、资本运作与硬件周期 (45:00) 在公司运营与资本层面，Cerebras于2026年5月完成IPO，募集55亿美元，估值约640亿美元，远期市销率达67倍，使800余名员工成为百万富翁。针对阿联酋大客户兼少数股东G42曾引发的美国外国投资委员会（CFIUS）审查，Feldman澄清相关国家安全审查早在2025年3月已完全解决，随后才引入包括1789 Capital在内的多元背景资本，未寻求任何政治优待。 谈及本土半导体制造政策与地缘管控，Feldman指出建设单座先进制程晶圆厂耗资300亿至400亿美元、历时5至6年，美国必须具备跨越政府周期的20至25年产业生态耐心以应对碎片化的地方规章。同时，他高度支持美国商务部产业与安全局（BIS）实施严格出口管制，并明确表示接受部分海外市场因此对Cerebras关闭的商业代价；相比软件行业“快速试错”的迭代文化，硬件工程需要数倍的高额资本开支与精准容错周期。 播客主持人在总结中指出，面对高昂API账单带来的“Token休克”，企业算力采购正从盲目调用闭源专有模型转向分级部署：常规任务将交给低成本开源模型与标准硬件处理，而多步骤Agent协作等延迟敏感型复杂任务，则将流向Cerebras等具备超低延迟的高性能推理引擎；在硬件创新的多周期投入与软件快速迭代的冲突下，如何平衡当期财务回报与万亿级算力基础设施竞赛仍是行业长期博弈的核心。

## Background 本期节目由彭博社（Bloomberg）旗下的知名播客《Odd Lots》于 2026 年 5 月 21 日发布在 RSS 及网络平台上。节目单简介显示，正值 AI 芯片初创巨头 Cerebras 上市当周，创始人兼首席执行官 Andrew Feldman 做客节目，深入探讨其面积比普通芯片大 58 倍、如餐盘大小的巨型晶圆级芯片（WSE）如何打破算力瓶颈以提供极速 AI 推理，并分享了 GPU 发展史、开源与闭源生态竞争以及与台积电的供应链合作等关键议题。 ## TL;DR 节目围绕 Cerebras 颠覆传统 GPU 架构的晶圆级芯片展开，深入解析了其如何通过超大片上高速内存彻底突破“内存墙”，并在规避 HBM 紧缺与 CoWoS 封装瓶颈的同时实现比传统架构快 15 倍以上的超低延迟推理。CEO Andrew Feldman 驳斥了“智能体不需要极致速度”的论调，阐明速度在复利式生产力竞争中的决定性作用，并指出前沿模型正加速与 CUDA 解耦。对话还进一步探讨了开源与闭源模型的单位智能成本权衡、算力资产金融工程化、数据主权防线以及硬件研发必须经受的长周期考验与工程师情怀。 ## Atmosphere & style 对谈在《Odd Lots》标志性的机智诙谐与深度财经硬核分析之间自如切换，主播 Joe 与 Tracy 以“AI 精神病”的幽默自嘲开场，随后迅速切入严肃的半导体与商业架构探讨。嘉宾 Andrew Feldman 展现出资深硬科技创业者的笃定与犀利，既有反驳分析师观点时的直言不讳，也有讲述带领团队十年磨一剑、创造 800 位百万富翁时的真情流露。整场对话既充满高密度的技术洞见与商业逻辑推演，又洋溢着对“赋予硅片生命”的浪漫工程师情怀，节奏紧凑且引人入胜。 ## Bulletin - `05:00` Cerebras 在上市周公开市场估值达 640 亿美元，其制造的餐盘大小单片晶圆级芯片（WSE）面积是常规芯片的 58 倍，直接集成海量片上高速内存以解决“内存墙”，推理速度相比传统顶级 GPU 提高 15 倍至数十倍。 - `20:00` 公司通过差异化架构全面避开了当前半导体供应链的三大卡脖子环节（垄断且紧缺的 HBM 内存、台积电 CoWoS 先进封装与 3nm 产能争夺），改用成熟的 5nm 制程；当前全行业真正的物理瓶颈已转移至具备电力供应的数据中心机房建设。 - `12:15` CEO Andrew Feldman 驳斥了“智能体（Agentic）工作流对推理速度不敏感”的观点，指出智能体效率倍差会带来指数级复利差距，且传统 GPU 在强行提速时能耗与成本会急剧飙升，而晶圆级架构大幅压低了高速 Token 的边际电费与计算成本。 - `27:30` 英伟达 CUDA 护城河在推理时代大幅贬值，API 迁移仅需几行代码；而在前沿模型训练端，三大前沿模型中已有 Google Gemini（TPU）和 Anthropic Claude（Trainium）脱离 CUDA 生态，其垄断统治力正在被自研 ASIC 蚕食。 ## Key takeaways - `05:00` 突破内存墙：Cerebras 打破了芯片切片封装的传统路径，将整块晶圆做成单枚芯片，通过将数十倍面积全部用于集成片上超高速内存，根本性消除了 GPU 外部数据搬移的延迟瓶颈。 - `05:00` 商业化破局：在克服材料、供电、高密度液冷与编译软件等全栈工程难题后，公司与 OpenAI 敲定了价值超 200 亿美元的合作，并与 AWS 达成战略落地部署。 - `10:00` 训练与推理的分化：训练是制造 AI，推理是使用 AI；2025 年大模型跨越实用拐点后引发了推理需求的全面爆发，专用硅片架构开始取代修补型 GPU 成为刚需。 - `12:15` 智能体工作流对极速的刚性需求：智能体代表单位时间产出，20 分钟内完成数倍工作量的能力会在商业竞争中产生复利式领先，“慢速智能体依然可用”是认知误区。 - `14:40` GPU 的“油耗悖论”：传统 GPU 擅长以低成本输出慢速 Token，但强行追求低延迟时单 Token 耗电与成本呈指数级恶化；晶圆级架构则能以极低能耗持续制造超高速 Token。 - `20:00` 供应链差异化突围：Cerebras 绕开了 HBM 内存采购周期、台积电 CoWoS 先进封装与 3nm 产能挤兑，选用 5nm 成熟制程获得了充沛的代工配额。 - `20:00` 基础设施新瓶颈：半导体制造的工程极限已被攻关突破，全行业未来 15 至 18 个月最核心的硬约束是“带电的数据中心实体地产（Powered Buildings）”的供应短缺。 - `23:45` 单位智能平准化成本：闭源前沿模型相比开源模型仅维持 3% 到 5% 的增量智能优势，却包含高昂研发摊销；开源模型在高速推理芯片上仅需支付电费和硬件折旧，具备极端成本优势。 - `27:30` CUDA 护城河瓦解：推理层对 CUDA 零依赖，迁移只需重定向 API；在三大代表性前沿模型中，Gemini 基于 TPU、Claude 主力基于 AWS Trainium，英伟达在前沿训练阵地的统治力已松动约 70%。 - `30:00` 算力资产金融工程化：从 CoreWeave 以物理 GPU 作为底层抵押进行债务融资开始，算力市场正朝标准化推进，未来势必催生类似大宗商品的套期保值与金融衍生品。 - `32:48` 核心客户与主权算力：阿联酋国家级 AI 领头羊 G42 是 Cerebras 早期最大客户与投资方，在北美多地部署晶圆级集群用于训练 Jais 双语大模型与支撑石油巨头智能化转型。 - `37:26` 垂直专有数据主权：掌握数十年真实临床实验数据的制药和医疗巨头（如 GSK、Mayo Clinic），为防止私有资产反哺通用大模型，必须依赖物理隔离的独立私有推理栈。 - `40:00` 美国本土建厂现实阻碍：单座先进晶圆厂建设耗资 300 至 400 亿美元且历时 5 至 6 年，跨政党政治周期的不确定性与碎片化的区域建筑规范远比资金补贴本身更难克服。 - `43:24` 出口管制路线分歧：英伟达等厂商主张持续供应以绑定海外开发者的软硬件生态惯性，而 Feldman 坚定支持以严格技术阻断来延缓对手的代际演进节奏。 - `45:00` AWS 合作与监管合规澄清：Cerebras 将通过 Bedrock 提供与 Trainium 混合的解耦推理服务；公司早于 2025 年 3 月即彻底解决 CFIUS 安全审查，否认了依靠政治关联通关的传言。 - `50:00` 硬科技的研发哲学：芯片研发属于“两次度量一次下刀”的极高容错成本领域，不能套用互联网软件敏捷试错思维，需要团队对赋予硅晶以生命怀抱长期主义。 - `52:32` 推理需求跑步机效应：随着长链智能体调用不断累加单步延迟，用户对速度的阈值将不断提高；而在收到高额 Token 账单后，企业也将加速开启算力分级部署。 ## Detailed walkthrough ### `00:00` “AI 精神病”与估值狂热：Odd Lots 走进晶圆级芯片巨头 播客开场，主播 Joe Weisenthal 与 Tracy Alloway 展开了一场幽默的互侃。Joe 坚称自己并未患上所谓的“AI 精神病”（AI psychosis），但不得不承认，关于大模型的思考正像滚雪球般侵占他脑中的全部空间——从哪个模型表现最强、推理的经济学效益如何，到模型训练中预训练与后训练的比例究竟占多少。Tracy 调侃道，整天琢磨 AI 本身可能就是一种精神病症状，而“承认自己有问题是康复的第一步”。不过，两人一致认同，如今资本市场乃至实体经济的巨大体量都在围绕 AI 运转，过度关注实属必然。在 AI 产业链的各项细分领域中，芯片始终是播客最热衷探讨的核心基石。 本期节目的主角正是当周刚完成重磅 IPO 的 AI 芯片独角兽 Cerebras。Tracy 指出，这家公司不仅以极高估值完成了巨额融资（按前瞻市销率计算高达约 67 倍，估值极为火热），更有别于传统硬件巨头模块化拼接、网络堆叠的方案，打造出了颠覆行业认知的“餐盘大小”的巨型晶圆级芯片（Wafer-Scale Engine）。如果说业界通常依靠网络将成千上万枚独立芯片互连来换取算力和内存，Cerebras 则选择缩短所有计算单元之间的物理距离，直接把一整张晶圆做成单枚芯片。 - **市场背景**：Cerebras 在其上市周引发市场轰动，虽然尚未实现盈利，但资本市场给出了极高的前瞻估值倍数。 - **差异化路线**：行业主流是采用模块化网络连接（将多个小芯片拼接成集群），而 Cerebras 押注于单颗超大尺寸晶圆芯片。 --- ### `05:00` 盘子一样大的芯片：突破内存墙与超越 GPU 的底层逻辑 Cerebras 创始人兼首席执行官 Andrew Feldman 做客播客，并在镜头前举起了他们标志性的巨型芯片。这一视觉冲击力极强的硬件实物甚至比普通餐盘还要大，面积是历史上制造过的普通芯片的 58 倍。针对为何要做如此庞大的芯片，Andrew 从底层架构进行了剖析：更大的芯片能在更短时间内处理更多信息，从而输出更快的计算结果。过去五六年间，英伟达（NVIDIA）也将芯片面积从 400 平方毫米逐步推向 800 平方毫米的物理极限，而 Cerebras 则彻底突破了切片限制，直奔整张晶圆。 Andrew 指出，这颗巨型芯片从根本上解决了 AI 推理的核心瓶颈——“内存墙”（Memory Wall）。在计算体系中，内存主要分为两种：一种单位容量大但访问极慢，另一种单位容量小却快如闪电（片上高速内存）。历史上的所有图形处理器（GPU）为了装下模型参数，普遍依赖高容量但速度缓慢的外部内存，导致计算单元在等待数据搬运上耗费了大量时间；这就是用户在使用各类 AI 对话时总能明显感受到延迟卡顿的原因。而 Cerebras 走向晶圆级架构后，虽然无法提高高速内存本身的单位存储密度，却直接将芯片表面积放大了数十倍，得以“塞入满满当当的超高速内存”。这也使得其推理速度比当前最顶级的 GPU 快出 15 倍，在特定计算任务上甚至能实现 50 到 1000 倍的性能飞跃。 ```mermaid flowchart TD accTitle: GPU与晶圆级芯片架构对比 accDescr: 对比传统GPU的外部慢速大容量内存架构与Cerebras晶圆级芯片片上超高速内存架构的数据传输与性能差异 subgraph GPU["传统 GPU 架构"] A["慢速/大容量外部内存"] -->|"数据搬运产生高延迟瓶颈"| B["计算核心"] B --> C["推理速度慢、响应时延高"] end subgraph Cerebras["Cerebras 晶圆级架构 (WSE)"] D["58倍面积超大晶圆芯片"] --> E["直接集成海量高速片上内存"] E -->|"极低通信延迟与极高带宽"| F["超大规模集成计算核心"] F --> G["推理速度提升 15x 至 1000x"] end ``` 在半导体行业 75 年的历史中，尝试做“晶圆级集成”的先驱几乎全军覆没，其中包括计算机体系结构先驱、Amdahl 定律提出者 Gene Amdahl 在 1980 年代创立 Trilogy Systems 时的惨烈失败。即便在 Cerebras 取得突破后，试图效仿其做法的后来者也无一成功。Andrew 表示，为了实现这一构想，团队攻关了整整十年，耗时五年并投入约 5 亿美元才交付了第一代芯片。期间克服了横跨半导体全栈的极端工程挑战： - **光刻与晶圆制造**：与台积电（TSMC）深度联合研发，解决了跨光罩视场（reticle）的互连技术。 - **先进材料与封装**：彻底重构将整张硅片贴合到母板、进行超大尺寸电气互连的封装工艺。 - **供电架构**：研发全新的供电系统，以安全输入并分配远超常规邮票大小芯片的数万瓦级超高功率。 - **散热与冷却**：设计前所未有的高密度液冷方案，均匀带走巨大的热负荷。 - **底层软件生态**：完全重写适配这种单芯片非冯·诺依曼巨型阵列的编译与运行时软件。 凭借这一工程壮举，公司迎来了商业爆发期。Andrew 透露，他们此前已与 OpenAI 敲定了一项价值超 200 亿美元的重大合作协议（硅谷历史上金额最大的合同之一），并随后与 AWS 达成战略合作，将其晶圆级系统直接部署到 AWS 数据中心内部，赋能全球前沿大模型的高速计算与推理。 ### `10:00` 训练与推理的分野：Cerebras 的底层构想与从零重构 在经历漫长的研发不确定期后，Cerebras 迎来了里程碑式的时刻，公司成功上市，在 IPO 初期的市场估值高达 640 亿美元。面对主持人关于“Cerebras 的超大芯片究竟是专攻推理还是兼顾训练”的疑问，首席执行官安德鲁·费尔德曼（Andrew Feldman）给出了极为清晰的界定：“训练是我们创造 AI 的方式，而推理是我们使用 AI 的方式。” 费尔德曼回顾道，在 2025 年上半年，业界训练出的模型终于跨越了可用性门槛，真正具备了实用价值，这随即引发了终端使用量的井喷。使用 AI 的过程本质上全在跑推理，因而产生了一场巨大的推理需求海啸，这种势头在 2026 年依然狂飙猛进，并预计将在未来数年内持续。而在 2015 年 Cerebras 创立之初，团队就预见到 AI 终将吞噬海量的算力。基于这一判断，他们做出了两项根本性押注：第一，AI 必须拥有专属定制芯片——正如图形渲染孕育了 GPU、移动计算催生了 ARM 架构一样；第二，修修补补现有的 GPU 架构是行不通的死胡同，必须从零开始、在一张白纸上重新设计。 尽管 Cerebras 打造的全新计算架构在训练和推理两端都实现了数量级的速度提升，但由于当前市场上推理需求的爆发式增长，公司当下的绝大部分业务重心落在了推理上，即便它在训练速度上同样以绝对优势碾压传统 GPU。 ```mermaid flowchart TD accTitle: AI工作负载与硬件进化路径 accDescr: 展示从AI计算需求到Cerebras白纸重构架构的技术分流与应用落地 subgraph Hardware_Evolution["硬件演进逻辑"] A["专用计算需求"] --> B["图形渲染 -> GPU"] A --> C["移动计算 -> ARM 处理器"] A --> D["AI 极端算力需求 -> 必须专用硅片"] end subgraph Cerebras_Approach["Cerebras 白纸重构"] D --> E["拒绝修改 GPU 架构"] E --> F["晶圆级芯片全新架构"] end subgraph AI_Workloads["AI 实际落地"] F --> G["模型训练 (创造 AI)"] F --> H["模型推理 (使用 AI)"] H --> I["2025-2026 爆发式海啸需求"] end ``` - **市场估值**：在 IPO 初期的交易中，公开市场对 Cerebras 的估值达到了 640 亿美元。 - **两大底层押注**：必须使用专用硅片（Dedicated Silicon）；坚决放弃修补 GPU，坚持从零设计（Clean sheet of paper）。 - **训练 vs 推理**：训练是“制造 AI”，推理是“消费与使用 AI”；推理业务在 2025 年初模型达到实用拐点后迎来了全面爆发。 --- ### `12:15` 驳斥 Ben Thompson：“智能体推理”对速度更加饥渴 主持人引述了知名科技商业分析师本·汤普森（Ben Thompson）在通讯专栏中的观点：推理可以划分为“回答型推理”（Answer inference，如格式化简历、撰写短文、问答互动）与“智能体推理”（Agentic inference，如后台自主调度、多步骤执行实际工作与服务）。汤普森曾推论，在无需人类死守屏幕等待即时视觉输出的智能体流程中，速度可能并没那么重要。 费尔德曼对这一论调予以了直截了当的驳斥，明确表示“这种认为智能体流程不需要速度的看法完全是错的（dead wrong）”。费尔德曼承认回答型与智能体推理在形态上存在差异，且 Cerebras 的芯片能完美兼顾两者，但他强调速度在两种场景下的重要性完全等同。在人机直接交互场景中，没人愿意忍受延迟——正如慢速搜索引擎的市场规模是零、拨号上网的市场规模是零一样，等待是用户体验的死敌。而在智能体场景中，速度则直接决定了生死存亡的商业竞争力。 费尔德曼生动地解释道，智能体代表着生产力：如果竞争对手的 AI 智能体在 20 分钟内完成的工作量是你的 3 倍、5 倍甚至 10 倍，“你绝对会被打得落花流水（you're going to get smoked）”。这种效率差距会随时间不断复利放大——当对手做完 1 个单位的工作时，你能做完 3 个；当下一次他们再做 1 个单位时，你已经完成了 6 个。在任何行业竞争中，单位时间内完成更多工作的能力，都是能够不断累积的终极优势。因此，作为 Cerebras 看家本领的高吞吐超低延迟，在所有生产力场景下都具有决定性意义。 - **反驳核心**：速度在智能体（Agentic）工作流中不仅重要，更是核心竞争壁垒。 - **经典类比**：慢速搜索与拨号上网的市场份额皆为零，印证了等待对人类专注度的扼杀。 - **复利优势**：智能体执行任务的倍速差异（3x、5x、10x）会在商业运营中产生指数级的生产力差距。 --- ### `14:40` Token 经济学：GPU 的“油耗悖论”与极速推理的成本真相 针对主持人提出的“巨型晶圆对 Token 经济学意味着什么”（是否类似于在 CVS 买单支牙膏与去 Costco 批量采购的大包装成本差异），费尔德曼揭示了芯片底层功耗与定价之间的反直觉逻辑。他指出，当前市场愿意为“极致速度”支付极高的溢价，例如 Anthropic 曾推出一项速度快 2 倍的高级服务，其收费高达常规价格的 6 倍，结果依然迅速售罄且供不应求；而 Cerebras 芯片的生成速度比 Anthropic 这个“2倍速”还要再快上 15 倍。 费尔德曼指出，更快的 Token 并不意味着更高的制造成本，传统 GPU 架构存在一个致命的物理特征与“油耗悖论”： - **GPU 架构的局限**：GPU 非常擅长且极具性价比地生成“极其缓慢的 Token”。如果你对延迟毫不在意，GPU 制造 Token 的单价非常低廉。然而，一旦试图强行提高 GPU 的推理速度以满足人类的交互需求，其单 Token 的耗电量和运行成本就会呈指数级飙升，就像汽车开得越快、每加仑油跑的里程数（MPG）越低一样。 - **Cerebras 的颠覆点**：为了达到能抓住用户注意力、具备真正实用价值的响应速度，GPU 集群的部署会变得极其昂贵且耗电惊人。相比之下，Cerebras 芯片制造“高速 Token”的成本远远低于 GPU，并且只需耗费后者一小部分的电力。 - **市场溢价案例**：Anthropic 提速 2 倍的 Token 售价高出 6 倍仍被疯抢；Cerebras 的实际速度相比该加速版仍快 15 倍。 - **GPU 的速度与能耗曲线**：在极慢速度下 GPU 极其廉价；但强行加速时，Token 的计算边际成本与功耗急剧恶化。 - **晶圆级优势**：以极低能耗与低廉边际成本提供超高速 Token，打破了延迟与功耗相互掣肘的技术瓶颈。 --- ### `18:30` 推理市场未来的版图之争 在技术原理与速度价值达成共识后，对话转向了未来的商业竞争与市场渗透节奏。主持人提出，假定全行业都认同 Cerebras 的一体化超大晶圆技术代表着终极解决方案、所有人都渴望拥有最快速度的推理体验，那么展望明年及随后的几年，Cerebras 究竟能在多大程度上将技术优势转化为对庞大推理市场的实际占有率，其扩张的上限又将由哪些关键市场变量所决定。 ### `20:00` 台积电产能与芯片制造瓶颈：Cerebras 如何绕开三大供应枷锁 在谈及芯片代工产能是否成为公司增长的掣肘时，Andrew Feldman 坦言台积电（TSMC）确实是整个供应链中举足轻重的一环，但 Cerebras 却拥有极为独特的竞争优势。当前限制各大厂商构建 AI 算力的瓶颈主要集中在三个领域：其一是高带宽内存（HBM）。HBM 虽然容量大，但读写速度慢，全球主要由三星（Samsung）、SK海力士（SK Hynix）和美光（Micron）三家把持，供应极度紧张、交货周期冗长且成本高得令人咂舌——而 Cerebras 的晶圆级芯片压根不用 HBM；其二是台积电内部极为紧俏的 CoWoS 先进封装工艺，这是英伟达（NVIDIA）等一众 GPU 厂商不可或缺的工序，Cerebras 同样完全不依赖它；其三则是台积电负荷最重、产能最为吃紧的 3 纳米先进制程，而 Cerebras 选用的则是成熟度与产能更稳固的 5 纳米工艺。 正因如此，Cerebras 避开了当前半导体行业中最致命的几道供应枷锁。谈到与台积电的合作，Feldman 盛赞对方是“迄今为止地球上最伟大的制造企业”，并将台积电的晶圆厂比作“现代金字塔”，甚至热心向听众推荐去台湾一定要参观台积电的创新馆（Museum of Innovation）。台积电为 Cerebras 提供了充足的晶圆配额。Feldman 感叹，行业当前真正的瓶颈其实充满了黑色幽默：工程师们突破了计算机科学 75 年未解的物理制造极限，打造出比传统巨头快得多的硬件，并编写了非凡的软件，然而最终全行业都被“带电的建筑物”（即配有充沛电力的数据中心机房地产）硬生生卡住了脖子，这种由电力与机房短缺导致的瓶颈在接下来的 15 到 18 个月内都无法缓解。谈到霍尔木兹海峡局势导致的氦气（用于光刻工艺）短缺风险时，Feldman 表示团队虽在日常监控，但这属于制造伙伴应对的宏观范畴，并未对 Cerebras 造成任何实际冲击。 - **半导体行业三大卡脖子供应链**：HBM 内存（三大存储厂垄断且交期长）、台积电 CoWoS 先进封装、台积电 3 纳米生产线。Cerebras 全面绕开了这三项。 - **制程工艺选择**：Cerebras 采用台积电 5 纳米工艺，获得了完全满足业务需求的晶圆配额。 - **物理基础设施短缺**：全行业当前最核心的瓶颈不在晶圆代工，而在“有电力供应的数据中心实体建筑（Powered Buildings / Real Estate）”。 ```mermaid flowchart TD accTitle: AI硬件制造约束与Cerebras技术路径对比 accDescr: 对比传统GPU与Cerebras在内存、封装工艺、芯片制程及当前行业共同瓶颈上的差异。 subgraph IndustryPressure ["传统 GPU 供应链三大瓶颈"] A["HBM 内存供应紧缺<br/>(三星 / SK海力士 / 美光)"] B["台积电 CoWoS<br/>先进封装产能吃紧"] C["台积电 3 纳米<br/>尖端制程排期过载"] end subgraph CerebrasPath ["Cerebras 差异化架构路径"] D["晶圆级片上 SRAM<br/>(规避外部 HBM)"] E["整片晶圆互连<br/>(无需 CoWoS 封装)"] F["台积电 5 纳米成熟制程<br/>(获得充分配额保障)"] end subgraph RealWorldConstraint ["全行业共同的终极现实瓶颈"] G["数据中心基础设施<br/>(电力供应 + 机房地产限制，预计持续15-18个月)"] end A -.->|规避| D B -.->|规避| E C -.->|替代| F D --> G E --> G F --> G ``` ### `23:45` 开源与闭源模型的单位智能成本之争 除了销售硬件系统，Cerebras 还搭建了自营的云服务平台（Cerebras Cloud），界面体验与 OpenRouter 类似，为开发者提供主流开源模型的超高速 API 访问。针对主持人提出的“每单位智能平准化成本”（Levelized Cost of Intelligence，即每个智商点数的实际花销）这一构想，Feldman 明确断言：开源模型的单位智能成本远比闭源模型便宜，而且“便宜非常多”。在闭源世界中，用户为了获取最后那 3%、4% 或 5% 的增量智能表现，必须支付极其高昂的溢价。虽然开源模型在绝对能力上仍略逊于最前沿的闭源模型，但两者差距已缩小到个位数的百分比。 当开发者在 Cerebras 平台上运行拥有上万亿参数的开源模型（如国内顶尖的月之暗面 Kimi 或阿里通义千问 Qwen）时，推理速度可以达到传统方案的 10 到 15 倍。最关键的经济学逻辑在于：用户在这里只需为运行消耗的电力和硬件计算时间买单，完全不必分摊前期高达数亿美元的模型训练研发成本。当下市场正呈现两大阵营的对抗：一方是拥有闭源编程助手的 OpenAI 和 Anthropic，另一方则是基于开源生态蓬勃发展的 Cursor 和 Cognition 等新兴独角兽。Cerebras 既为 OpenAI 提供底层推理支持，同时也赋能 Cognition 等团队。展望未来二十年，Feldman 认为 AI 领域绝不可能被单一霸主垄断，正如企业软件领域有 Salesforce 和众多垂直玩家，芯片领域存在 x86（Intel/AMD 争雄）、ARM 生态与定制自研 ASIC 共存的格局一样，闭源大厂与开源模型将长期并存。 - **性能与价格权衡**：前沿闭源模型在能力上通常严格领先开源模型约 3% 到 5%，但用户获取这部分边际性能的溢价极高。 - **成本构成拆解**：调用开源模型推理仅需支付电费与硬件折旧算力成本，彻底免去了巨额的预训练摊销费用。 - **生态与客户支撑**：Cerebras 目前既为 OpenAI 供能，也为 Cognition 等采用开源架构的头部 AI 应用提供算力支持。 ### `27:30` 破除 CUDA 护城河神话：推理时代的范式转移与前沿格局重塑 谈及初创芯片企业挑战英伟达时避不开的“CUDA 软件生态护城河”，Feldman 给出了一针见血的剖析。他毫不掩饰对英伟达的尊敬，称其为“本世纪前二十五年最伟大的公司”，并赞扬黄仁勋（Jensen Huang）与博通（Broadcom）的陈福阳（Hock Tan）、AMD 的苏姿丰（Lisa Su）并列为当代最杰出的科技 CEO。他肯定了 CUDA 在开创 AI 计算版图时的历史功勋，但随即指出：“CUDA 在今天已经不再具有决定性意义，在推理领域更是毫无立足之地。” Feldman 透露，如今开发者如果想把运行在 GPU 上的模型迁移到 Cerebras 架构上，只需要“敲击十次键盘（10 keystrokes）”，将 API 指针重定向即可完成。而在决定行业风向的前沿大模型（Frontier Models）训练端，英伟达的统治力也已出现剧烈松动： 在一年前，全球所有顶尖前沿实验室的大模型全部建立在 CUDA 基石之上；而如今，全球最具代表性的三大前沿模型中，已有两个完全脱离了 CUDA 生态： 1. **Google 的 Gemini**：全部在自研 TPU 上训练并提供服务，完全没有 CUDA 的参与； 2. **Anthropic 的 Claude**：主力在 AWS Trainium 芯片上训练，并在 TPU、Trainium 及 GPU 等多元硬件上提供推理，对 CUDA 的依赖已被大幅解耦； 3. **OpenAI 的 GPT 系列**：仍主要在英伟达 GPU 与 CUDA 环境下训练。 在 Feldman 看来，三大前沿模型中有两家已经弃用，意味着 CUDA 在最核心的阵地失去了近 70% 的份额，堪称“市场份额大出血（hemorrhaging of share）”。三至五年前 CUDA 牢不可破的中心地位，在训练端正被自研 ASIC 加速蚕食，在推理端更是直接沦为了无足轻重的次要因素。正当话题转向近期资本市场对算力基础设施金融化的新动作时，对话进入了下一层深入探讨。 ### `30:00` 算力资产的金融工程化：从 GPU 抵押贷款到衍生品市场 面对主持人关于“推理服务商未来是否需要类似大宗商品的金融衍生工具来对冲算力价格波动，还是仅靠与数据中心签订双边长期协议即可”的疑问，Andrew Feldman 首先坦言自己并非金融工程师，但他建议从近年算力行业的融资演化史中寻找线索。他重点提到了新兴算力云提供商 CoreWeave 的破局之道——CoreWeave 在早期极富创造性地采用了以物理 GPU 作为底层抵押品的债务融资模式。这种金融工程层面的激进创新，让 CoreWeave 成功撬动了庞大的资本杠杆，在“新型云”（neocloud）赛道中一举夺得先发优势。 Feldman 进一步指出，在 CoreWeave 推进这一模式的过程中，像 Coatue 这样的投资机构敢于成为首批接受以 GPU 为抵押物放贷的机构，双方共同开辟了这一全新赛道。如今，以算力硬件和数据中心建设为标的的债务融资市场已经变得极其庞大且活跃。按照金融市场发展的普遍规律，一旦底层市场规模足够大、交易足够频繁，多空双方的博弈需求就会催生资产的标准化与规范化；而只要市场完成了标准化并确立了公允定价机制，衍生品与风险对冲工具的出现就只是时间问题。虽然 Feldman 对此类对冲工具目前是否是个好主意持保留意见，但他认为算力市场正在不可逆转地走向金融成熟期。 ```mermaid flowchart TD accTitle: 算力资产金融工程化演进路径 accDescr: 展示从早期 GPU 抵押贷款创新到未来算力金融衍生品市场的逐步演进过程。 subgraph Phase1 ["第一阶段：抵押创新"] A["CoreWeave 等新型云厂商"] -->|"以物理 GPU 为抵押"| B["Coatue 等先锋投资机构"] B -->|"提供巨额债务融资"| C["抢占 Neocloud 算力先发优势"] end subgraph Phase2 ["第二阶段：市场规模化与规范化"] C --> D["数据中心建设与硬件采购债务市场爆发"] D --> E["资产风险与现金流逐步标准化、易于定价"] end subgraph Phase3 ["第三阶段：金融衍生品涌现"] E --> F["算力多空双向对冲需求"] F --> G["算力容量远期合约与衍生金融工具"] end ``` --- ### `32:48` 招股书背后的神秘巨头：阿布扎比 G42 的算力布局 针对 Cerebras 提交的 IPO 招股书中高度集中的营收来源，主持人询问了其最大客户兼重要少数股权投资方——阿布扎比人工智能巨头 G42 的具体运作细节。在 OpenAI 达成合作之前，G42 为 Cerebras 贡献了绝大部分业务收入。Feldman 解释称，G42 是阿拉伯联合酋长国（UAE）国家级的人工智能领头羊（National AI Champion），其战略定位是为整个阿联酋构建全国性的算力与 AI 云底座。 G42 部署的 Cerebras 算力集群直接服务于阿联酋顶尖的科研高校、阿布扎比国家石油公司（ADNOC）等国家支柱型企业，以及 G42 旗下的九家专业运营子公司。值得注意的是，为了满足高性能与能耗需求，这些算力基础设施迄今为止多数部署在美国境内，包括加州圣克拉拉（Santa Clara）、明尼苏达州明尼阿波利斯（Minneapolis）、德克萨斯州达拉斯（Dallas），并且即将在加拿大多伦多（Toronto）落地新数据中心。G42 在这些集群上同时跑训练与推理任务，不仅研发了顶尖的阿语-英语双语大模型（如 Jais 模型），还在基因组学及模型在线服务（Model Serving）领域全面发力，其算力服务不仅辐射阿联酋本土生态，也正在向全球企业开放。 - **G42 身份定位**：阿联酋国家 AI 领头羊、Cerebras 核心战略客户及少数股权投资方。 - **算力部署地点**：位于美国本土的圣克拉拉、明尼阿波利斯、达拉斯，以及筹备中的加拿大多伦多。 - **业务覆盖领域**：石油巨头 ADNOC 工业智能化、高校科研、阿英双语基础大模型预训练、基因组大数据分析及企业级云端推理。 --- ### `37:26` 企业数据主权防线：为何专有领域巨头拒绝向通用大模型低头 随着 OpenAI 和 Anthropic 等模型厂商不断推陈出新——正如主持人所言，Anthropic 几乎每隔几天就会发布一个能直接颠覆报税或特定垂直行业应用的新提示词模板或功能，导致一众依赖其 API 的创业公司应声倒下——主持人提出一个核心商业博弈：企业级客户是否会越来越倾向于选择完全独立于基础模型厂商之外的第三方数据中心和推理服务商，以防自己的核心数据在被调用的同时“哺育”了未来的潜在竞争对手？ Feldman 深入剖析了受 AI 冲击的工种差异。他指出，目前最直接遭受 AI 威胁的是那些基于现成公开知识体系提供中介服务的传统白领职业，例如普通会计师和常规律师：“他们所做的本质上是站在普通民众与复杂的国税局（IRS）税法或判例法之间充当知识中介”，而这恰恰是当前通用大模型最擅长吞噬和替代的领域。但与此截然相反的是拥有极高“数据壁垒”的垂直硬核领域，如药物研发、遗传学和基因组学： - **独特数据资产的价值**：像葛兰素史克（GlaxoSmithKline, GSK）以及 Cerebras 的大型客户梅奥医学中心（Mayo Clinic）这类机构，手中掌握着历经二三十年、耗资数百亿美元真实实验积累下来的独家临床与生物分子数据。 - **严防通用模型反哺**：这些行业巨头绝不可能将高价值私有数据拱手交给基础模型厂商去直接调用 API，除非有绝对的技术架构保证其专有数据不会被用来强化对方的通用模型。 - **对独立基础设施的需求**：高价值垂直行业必须依赖底层物理隔离的独立算力栈与专用推理架构，以确保在挖掘自身数据商业价值的同时，将企业核心竞争壁垒牢牢锁在自己手中。 ### `40:00` 专有数据价值与美国本土芯片制造的真正阻碍 在企业级 AI 应用中，数据壁垒始终是最坚固的护城河。制药公司和医疗保健巨头每年投入数十亿美元采集病患护理记录或药物设计测试结果，这些专有数据蕴含着极具价值的深刻见解。由于洞见直接衍生自他们独有的数据，这种竞争壁垒相比开源基础模型而言受到更严密的保护。 谈及制造端，主持人感叹此前去台北未去参观台积电晶圆厂颇感遗憾。话题随后转向美国的《芯片与科学法案》（CHIPS Act）及产业政策。针对“为何在美国建立芯片制造产能如此艰难”的核心问题，Andrew Feldman 直言，首要原因是建厂本身就是一项极其艰巨且昂贵的工程：一座先进晶圆厂耗资高达 300 亿至 400 亿美元，建设周期长达 5 到 6 年。 这种极高的资金规模与超长建设周期直接导致跨届政府问题——在政权更迭频繁的政治体制下，很难维持政策的长期连贯性。此外，美国各地区错综复杂、碎片化的建筑规范（building codes），让晶圆厂建设团队在合规谈判中耗费大量精力。尽管台积电在亚利桑那州投入数百亿美元、三星在德州投资数千亿美元，但 Feldman 强调，芯片制造并非只建一座厂，而是需要打造一个持续 20 至 25 年的完整生态圈，确保不仅能满足今天的先进制程，还能在未来 10 年持续保持全球前沿竞争力。 - 建造单座先进晶圆厂的成本高达 300 亿至 400 亿美元，工期需 5 至 6 年。 - 台积电（亚利桑那州）与三星（德克萨斯州）已投入数百亿美元并承诺后续追加投资。 - 真正难点在于维持长达 20 至 25 年的产业生态与政策耐力，而不仅是建起单一工厂。 --- ### `43:24` 出口管制的分歧：技术封锁与“绑定客户”路线之争 半导体不仅是商业产品，更被上升为战略资产，出口限制和许可证审查正深刻影响芯片企业的日常经营。Feldman 坦言，如果三四年前被问到这个问题，他会觉得出口管制对公司完全不重要，但如今它已至关重要。他曾与两届美国商务部及负责出口管制的工业与安全局（BIS）领导层深度打交道，他感叹那里的公务员聪明勤勉，领着远低于私营领域的微薄薪资，肩负着极其复杂的国家安全使命。 然而，在“如何防止最顶尖技术流入产业对手（如中国）手中”这一战略目标上，业内存在根本性的策略分歧： ```mermaid flowchart TD accTitle: 半导体出口管制战略路线对比 accDescr: 对比限制出口与保持供应两种对华芯片战略 subgraph NvidiaRoute["路线 A：英伟达主张（生态绑定）"] A1["持续向中国市场供应定制/合规产品"] --> A2["防止中国自主研发替代品"] A2 --> A3["迫使对手持续依赖美国软硬件生态"] end subgraph FeldmanRoute["路线 B：Feldman / Cerebras 观点（技术物理阻断）"] B1["严格限制前沿核心技术外流"] --> B2["接受部分海外市场彻底关闭的代价"] B2 --> B3["在物理上延缓对手技术代际迭代节奏"] end ``` 英伟达等公司认为，正确的打法是继续对中国提供产品，让他们的开发者始终绑定并习惯使用美国架构与软硬件生态；而 Feldman 则坚定支持另一方立场，他认为限制顶尖技术的扩散是合理的，即便这意味着企业必须主动放弃某些庞大的海外市场。 --- ### `45:00` AWS 合作落地方式与破除政治“通关”传言 谈及与亚马逊 AWS 的合作机制，Feldman 透露客户目前还不能直接在 AWS 上调用 Cerebras 芯片，但该服务很快将在 Bedrock（AWS 的托管 AI 平台）中上线。届时，用户可以在下拉菜单中直接选择超高速推理服务。 该服务底层采用了名为“解耦方案”（disaggregated solution）的架构设计： - 一部分推理工作负载运行在 AWS 自研的 Trainium 芯片上； - 另一部分关键加速环节由 Cerebras CS-3 超级计算系统接管驱动。 关于定价，Feldman 预计客户会为这种极致的低延迟付费，但最终定价权完全由亚马逊自主决定。 随后话题转到公司于 2026 年 5 月正式启动的 IPO。此前，Cerebras 的上市之路多次因阿布扎比 AI 巨头 G42 的密切合作而引发美国外国投资委员会（CFIUS）的国家安全审查。针对外界质疑——即 Donald Trump Jr. 旗下基金 1789 Capital 参投公司 G 轮融资是否帮助公司“政治开绿灯”——Feldman 坚决驳斥了这种推论。他明确澄清，公司在 2025 年 3 月就已彻底解决了所有 CFIUS 审查合规问题，时间远早于 1789 Capital 的注资；公司接纳资本只看其能否协助企业发展，管理层从未、也绝不会寻求政治利益输送或走政治后门。 --- ### `47:37` 一夜身家过亿的真实心境：比起个人财富，更骄傲创造了 800 位百万富翁 当主持人问及“在一夜之间成为十亿美元级富翁（Billionaire）是什么感觉”时，Feldman 坦承这对他的心态来说几乎是“毫无波澜的无感状态”（a big nothing）。他直言，做科技公司 CEO 如果纯粹为了个人发财，那简直是最痛苦、最艰难的赚钱途径；唯有发自内心热爱这项事业、热爱团队，才可能坚持下来。 对他而言，IPO 最具成就感的时刻绝非个人账户资产的跳动，而是让并肩作战的团队兑现回报： - 在他创办的上家公司（SeaMicro），上市与并购为团队创造了 100 多位百万富翁（Millionaires）； - 而在 Cerebras 此次 IPO 中，公司成功诞生了超过 800 位百万富翁。 Feldman 动容地表示，这才是能让他每天早晨醒来都感到由衷自豪的事情。节目尾声，主持人指出硬科技研发往往具备超长的研发周期与巨额资金消耗，并进一步向这位新晋上市公司掌门人提问：在走向公开市场后，面对华尔街季度财报的严苛审视，公司究竟该如何在短期业绩压力与长远硬核研发投资之间求得平衡。 ### `50:00` 硬件研发的慢工细活与赋予硅片生命 关于芯片架构创新与资本支出（CapEx）的未来演进，Andrew Feldman 直言，基于晶圆级芯片（Wafer-Scale Engine, WSE）的架构创新“最精彩的成果仍在前方”。他坚信团队未来几年所能实现的飞跃，将完全不逊于、甚至超越最初打造出“地球上最大芯片”时的突破。对于硬件研发周期漫长且试错成本极高的现实，Feldman 展现出坚定的硬件工程师情怀：做硬件绝不可能在一周、一个月甚至一年内完成，这就需要创业者从骨子里接受这种挑战。对于那些习惯于软件领域“快速行动、快速失败、迅速写代码扔给市场验证”的人，他虽然送上祝福（“Godspeed”），但明确表示这绝非硬件之道。 ```mermaid flowchart TD accTitle: 芯片研发哲学对比 accDescr: 对比传统互联网软件与硬科技芯片研发理念的本质区别 subgraph SW["软件开发思维 (Software)"] S1["快速迭代与试错"] --> S2["极低单次错误成本"] S2 --> S3["几周内发布并交付市场"] end subgraph HW["晶圆级芯片研发 (Cerebras Hardware)"] H1["深思熟虑: 两次度量一次下刀"] --> H2["错误代价极其高昂"] H2 --> H3["多年磨砺: 赋予整片硅晶生命"] end ``` 在芯片硬件领域，原则必须是“三思而后行，两次度量一次下刀”（Measure twice before we cut once）。工程师必须适应并热爱“一次失误就价值连城”的高压环境，更要着迷于“赋予一块硅片生命、让它完成前所未有壮举”的成就感。Feldman 强调，如果这项事业在一周内就能轻松完成，他反而不会如此热爱；他身边的工程师团队之所以投身于此，从来不是单纯为了赚大钱，而是真正渴望亲手打造出极度困难的硬核产品。主持人由此联想到身为物理学家的父亲常提的一句话：碳与硅在元素周期表上紧紧挨在一起，碳构成了地球所有有机生命的基石，而人类创造“人工生命”与智能的载体正是硅。两人还打趣谈到硅下方的锗（Germanium），笑言未来是否会轮到锗来大放异彩。 - **硬件创新的核心信条**：做底层硬件必须接受高成本、长周期的规律，容错率极低，与软件的敏捷试错风格截然相反。 - **工程师的内在驱动力**：优秀芯片团队的凝聚力源于挑战极端工程难题，而不是短期的财务回报。 - **碳与硅的哲学共鸣**：碳承载了生物生命，硅承载了人工智能；两者在元素周期表上的相邻，为计算硬件赋予了浪漫的哲学隐喻。 --- ### `52:32` 主持人复盘：推理经济学、速度溢价与大模型算力配置 访谈结束后，主持人 Joe Weisenthal 与 Tracy Alloway 展开了深度复盘。他们感叹，AI 芯片领域关于“推理（Inference）的经济学、容量与速度市场”其实才刚刚拉开序幕。造出像餐盘一样巨大的晶圆级芯片，听起来甚至带有一丝《洋葱报》（The Onion）式的荒诞黑色幽默——“这家公司靠造出全世界最大的芯片解决了推理问题”；但结合此前与 SemiAnalysis 分析师 Ray Wang 讨论的行业背景，Cerebras 绕开了高带宽内存（HBM）的集成瓶颈，且不必在先进制程纳米节点上头破血流地争抢产能，确实为自身赢得了宝贵的供应链呼吸空间。 ```mermaid flowchart LR accTitle: AI推理场景对推理延迟的需求分化 accDescr: 简单查询与长链Agent工作流在速度敏感度上的巨大差异 subgraph LowSpeed["低敏感场景：单次静态问答"] Q1["日常查询 (例: 翼手龙为什么不是恐龙)"] --> R1["等待数秒甚至数分钟完全可接受"] end subgraph HighSpeed["高敏感场景：Agentic 智能体系统"] Q2["多步自主推理 / 编程 / 长链调用"] --> R2["毫秒级累加成数十秒"] R2 --> R3["极度依赖高吞吐与极低时延"] end ``` 针对推理速度的价值，两位主持人展开了辨证讨论： 1. **速度边际效应与场景分化**：如果是为了解答“翼手龙为什么不是恐龙”这种日常闲聊问题，用户多等几秒乃至几分钟无伤大雅，为极致速度支付高额溢价显然不划算；但如果是面向智能体编码（Agentic coding）等需要数十次、数百次链式推理调用的长流程场景，单次 Token 生成的延迟会成倍累加，极速推理就具有不可替代的生产力价值。 2. **“预期的跑步机效应”**：技术进步会不断推高用户的心理阈值。原本需要 30 分钟的任务若缩短至 30 秒，用户很快就会对这 30 秒失去耐心，进而渴望 10 秒完成——对时间的极致压缩是永无止境的竞争。 3. **AI 估值的核心谜题与 Token 账单冲击**：目前资本市场面临的最大问号在于：市场究竟愿意为“略优于开源模型的闭源模型”以及“略快于普通算力的极速推理”支付多高的溢价？随着各企业在收到巨额“Token 账单冲击”（Token shock）后逐渐清醒，盲目把所有负载扔给顶级闭源模型的做法势必改变，企业未来将更擅长根据场景进行算力分级，市场格局与估值逻辑也将随之重塑。 ## Notable quotes > `10:00` 训练是我们创造 AI 的方式，而推理是我们使用 AI 的方式。 — Andrew Feldman > `12:15` 这种认为智能体流程不需要速度的看法完全是错的（dead wrong）。 — Andrew Feldman > `12:15` 你绝对会被打得落花流水（you're going to get smoked）。 — Andrew Feldman > `27:30` CUDA 在今天已经不再具有决定性意义，在推理领域更是毫无立足之地。 — Andrew Feldman > `27:30` 敲击十次键盘（10 keystrokes） — Andrew Feldman > `27:30` 市场份额大出血（hemorrhaging of share） — Andrew Feldman > `47:37` 做科技公司 CEO 如果纯粹为了个人发财，那简直是最痛苦、最艰难的赚钱途径；唯有发自内心热爱这项事业、热爱团队，才可能坚持下来。 — Andrew Feldman > `50:00` 三思而后行，两次度量一次下刀（Measure twice before we cut once）。 — Andrew Feldman ## Mentioned Cerebras, Andrew Feldman, Joe Weisenthal, Tracy Alloway, NVIDIA, TSMC, OpenAI, AWS, Ben Thompson, Anthropic, Samsung, SK Hynix, Micron, Gene Amdahl, Trilogy Systems, Kimi, Qwen, Cursor, Cognition, Jensen Huang, Hock Tan, Lisa Su, Broadcom, AMD, Google, Gemini, Claude, TPU, Trainium, CoreWeave, Coatue, G42, ADNOC, GlaxoSmithKline, Mayo Clinic, CFIUS, 1789 Capital, Donald Trump Jr., SeaMicro, Ray Wang, SemiAnalysis ## Video description <p data-renderer-start-pos="126">Size is the name of the game for the AI chipmaker Cerebras: Their chips are truly massive, about the size of a dinner plate. According to Andrew Feldman, CEO and founder of Cerebras, that is about 58 times larger than the average chip. That sheer size enables blazing fast inference for AI queries. Feldman joins us on the week of his company's IPO to talk about his core product and how it fits into the AI boom. We discuss the history of the GPU, competition between open-and closed-source models, the company's relationship with with TSMC, and more.</p> <p data-renderer-start-pos="680">Read more:<br><a href="https://www.bloomberg.com/news/articles/2026-05-20/nvidia-gives-disappointing-forecast-as-chip-competition-mounts?utm_medium=referral&utm_source=podcast&utm_campaign=odd_lots&utm_content=article" data-renderer-mark="true" data-is-router-link="false" data-testid="link-with-safety">Nvidia Tells Skeptical Investors That AI Is Ready to Go Mainstream</a><br><a href="https://www.bloomberg.com/news/articles/2026-05-21/trump-set-to-sign-ai-cybersecurity-directive-as-soon-as-thursday?utm_medium=referral&utm_source=podcast&utm_campaign=odd_lots&utm_content=article" data-renderer-mark="true" data-is-router-link="false" data-testid="link-with-safety">Trump Set to Sign AI Cybersecurity Directive as Soon as Thursday</a></p> <p data-renderer-start-pos="824">Only <a href="http://bloomberg.com/" data-testid="inline-card-resolved-view">Bloomberg - Business News, Stock Markets, Finance, Breaking & World News</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-renderer-mark="true" data-is-router-link="false" data-testid="link-with-safety">bloomberg.com/subscriptions/oddlots</a></p> <p data-renderer-start-pos="996"><a href="http://bloomberg.com/subscriptions/oddlots" data-renderer-mark="true" data-is-router-link="false" data-testid="link-with-safety">Subscribe to the Odd Lots Newsletter</a><br><strong data-renderer-mark="true">Join the conversation:</strong> <a href="https://discord.gg/oddlots" data-renderer-mark="true" data-is-router-link="false" data-testid="link-with-safety">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/4945a340-783e-4fba-9782-b4510023ea71/audio.mp3?in_playlist=982f5071-765c-403d-969d-ae27003a8d83)

[10:00] Andrew Feldman: 训练是我们创造 AI 的方式，而推理是我们使用 AI 的方式。 [12:15] Andrew Feldman: 这种认为智能体流程不需要速度的看法完全是错的（dead wrong）。 [12:15] Andrew Feldman: 你绝对会被打得落花流水（you're going to get smoked）。 [27:30] Andrew Feldman: CUDA 在今天已经不再具有决定性意义，在推理领域更是毫无立足之地。 [27:30] Andrew Feldman: 敲击十次键盘（10 keystrokes） [27:30] Andrew Feldman: 市场份额大出血（hemorrhaging of share） [47:37] Andrew Feldman: 做科技公司 CEO 如果纯粹为了个人发财，那简直是最痛苦、最艰难的赚钱途径；唯有发自内心热爱这项事业、热爱团队，才可能坚持下来。 [50:00] Andrew Feldman: 三思而后行，两次度量一次下刀（Measure twice before we cut once）。

[5:00] Andrew Feldman: The Cerebras Wafer-Scale Engine is 58 times larger than any other single chip ever built. [10:00] Andrew Feldman: In agentic workflows like coding or automated business processes, speed dictates output; if a competitor's system completes 5 to 10 times more work in 20 minutes, the slower user gets 'smoked.' [15:00] Andrew Feldman: GPUs produce slow tokens cheaply, but become astronomically expensive and power-hungry when pushed to deliver high speeds. [26:40] Andrew Feldman: CUDA has no role whatsoever in inference. [26:40] Andrew Feldman: NVIDIA has effectively lost 70% of its software market share among the industry's leading frontier labs. [50:00] Andrew Feldman: To the extent we can make artificial life... the intelligent part will be founded on silicon.

[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. Health care doesn always work great If you ever waited on a refill or couldn schedule an appointment you get it That the kind of stuff Optum is changing They using data and technology to integrate patient care pharmacy and everything else So health care is connected not complicated What that look like Cheaper prescriptions that are easier to get and care that looks at the whole person how you need it Optum is helping make healthcare work as one for everyone. Learn more at business.optum.com. Bloomberg Audio Studios. Podcasts. Radio. News. Hello and welcome to another episode of the Odd Lots podcast. I'm Jill Weisenthal. And I'm Tracy Alloway. Tracy, I have to say, unfortunately, I don't have AI psychosis. I'm certain of that. Debatable. I'm pretty sure I don't have AI psychosis. I do have to say unfortunately like the amount of time now where it like it feels like AI related questions and there are many of them are sort of like swallowing up the other thoughts that I have in my head of whether it questions about which models best and why and what are the economics of inference and how much training is pre versus post for each model Like it just sort of like this blob that growing that taking up more and more of my thoughts What is your definition of AI psychosis Because one would argue that maybe thinking about AI literally all the time would be a form of psychosis. Well, let's just say, like, I'm not the type who thinks that, like, I don't think that the AI is a friend, for one thing. I'm not in love with the AI models. I don't think that in collaboration with chat GPT that I'm stumbling on unified theory of physics and things like that. So like. But you do spend a lot of time inputting instructions, pressing the button. Yeah. Oh, yeah. What comes out. And see what comes out. I'm just saying I think I'm aware that I'm talking to a machine and that we're not establishing any great breakthroughs of which we are collaborators and partners and friends. Recognizing you have a problem is the first step towards healing, Joe. seriously, though, there's a good reason to think about AI more and more, which is that a huge chunk of not just the market, but the real economy is now revolving around AI, right? Totally. So anyway, again, within the AI conversation, there are a lot of subcategories. One of the subcategories happens to be another of Lot favorite topic which is chips Of course chips are used in multiple different ways Chips are used in different parts of the AI supply chain different types of chips have different roles And so we have to learn more We have to learn more And I have to say I'm particularly interested in the company we're about to speak to, partly because the two things I know about them are, number one, they just had a huge IPO, right? Raising something like $5.5 billion dollars at kind of insane multiple. I can't even do a price to earnings multiple because they're not profitable yet. But I think just on a sales basis, it was like 67 times forward earnings, which is pretty juicy, pretty hot. And the second thing I know about the company is they make giant wafers, which is just a fun image to have in your head. That's right. So if you were thinking it's like, OK, there is a hot entrant in this space, what is their differentiator? Well, one fact about them is their chips are just enormous about the size of the dinner plate. One might think you're reading an Onion article, but in fact, it's real. And apparently it actually has some real technical advantages. And it's different to what everyone else is doing. So everyone else is, I guess, doing this sort of like modular networking thing where you get together a bunch of chips and you connect them together and that's how you get. [5:00] More compute, more memory, more power, basically. But this company has done something different in the form of the giant wafer. The giant wafer. And if you figure that to get maximum performance, you sort of want to lessen the distance between things, then put it all on one wafer. Anyway, we're going to learn a lot more. I'm very excited to say about giant wafers and more. I'm very excited to say we do have the founder and CEO of Cerebros on the podcast, Andrew Feldman, truly the perfect guest. So, Andrew, thank you so much for coming on the podcast on the week of your IPO. Well, thank you so much for having me. What a pleasure. Absolutely. Why don't you just start us off the big giant chip? They're apparently real. They're as big as a dinner plate. What is the technical reason why this actually makes sense as a superior form of architecture for at least some aspect of AI? I think larger chips process more information in less time. Okay. And that produces faster results. And everybody had gone to bigger chips. NVIDIA had moved from 400 square millimeters to 800 square millimeters over the course of five or six years for this exact reason And in the compute industry wafer scale which is building a chip this big There it is By the way for those who are just listening Andrew is now holding up the chip And yes, it actually looks bigger than its interplate, to be honest. But that is a big chip. That's a big chip. That's a big chip. It's beautiful. It's 58 times larger than any other chip that had ever been built. Wow. And what it did was it allowed us to use a different type of memory. Okay. A type of memory that, at the beginning, there are two types of memory. There's memory that can store a lot, but it's really slow. Okay. And there's memory that can't store very much per square millimeter, but it's blisteringly fast. Okay. And historically, all graphics processing units use this memory that could store a lot, but was really slow. And that's the reason they do inference so slowly. So if you're using Claude right now or you're using anything but ChatGPT, what you'll frequently feel is you'll enter your prompt and you'll wait for an answer. Right. And that's because the memory is slow and they have to move a ton of information from memory to compute. Now by going to wafer scale we could use this fast memory Now we couldn make that memory store more information per square millimeter but we could add square millimeters And so by building this big chip, we were able to stuff it to the gills with this fast memory. And that's why we're 15 times faster than the fastest GPU. That's why on some problems, we're 50, 100, even 1,000 times faster than graphics processing units. Wait, can you explain how you actually managed to do this? Because I know there have been previous attempts to do wafer scale, and I seem to remember there was even like an early attempt in the 1980s or something to do it. How were you able to pull this off? Yeah, it was an ambitious undertaking, that's for sure. Every previous effort in the 75-year history of our industry had failed, including Gene Amdahl, who's sort of on the Mount Rushmore of compute in our industry. He failed sort of spectacularly in the mid-80s at a company called Trilogy. Not only that, but after we succeeded, people who had visited us, who'd been in our labs, tried to copy us and they also failed And so what we were able to do is solve a set of really fundamental problems And those problems cut across a wide swath of technology They cut across lithography so we had to collaborate closely with TSMC and they turned out to be a great partner. We had to make inventions in material and packaging. That's how you put a processor, how you put a piece of silicon on a motherboard, deliver power and I-O to it. We had to make inventions in power delivery. When you build a giant chip, you're going to deliver way more power to it than if you do a chip the size of a postage stamp. We had to invent ways to cool it. We had to write new types of software that ran on it. All of these had never been done before. And it was a decade-long process. It took us five years and about $500 million to deliver the first one. And it's been an extraordinary run since. In December, we signed a deal with OpenAI north of $20 billion, one of the largest contracts ever signed in Silicon Valley. And then in March, we signed a deal with AWS, where they would deploy our systems in their data centers, in their AWS data centers. And so it's just been an extraordinary run, but it took a long time. It took extraordinary engineering efforts. [10:00] And there were certainly long periods of time when it wasn't clear we were going to make this work. Obviously, you've hit this remarkable milestone. You have, in fact, IPO'd and so forth. And right now, markets valuing your company at $64 billion early days of the IPO. Just for the listener to understand, the chips, are they solely an inference as opposed to, you know, training? When we think about AI, I think about, OK, there's training, training the model and then answer giving. That's the inference. Are the chips for just for inference? So a couple of things. I think you framed it exactly right. Training is how we make AI. Yeah. And inference is how we use AI. And so what happened was that in sort of 2025, in the first part of 2025, the models we made were smart enough to be useful. And there was an explosion of use. And we use AI by doing inference. So there was this sort of tidal wave of demand on inference. And that has continued in 2026. And we think it will continue for years and years to come. And so that's what had happened. In 2015 when we began thinking about the company we knew that AI was on the horizon and it would eat a huge amount of compute right And we made sort of two fundamental bets We bet that it would need dedicated silicon and graphics had needed dedicated silicon. That's how you got the graphics processing unit. Mobile compute had needed dedicated compute. That's where you got ARM processors. We made that bet, and we made a bet that modifying the GPU architecture wouldn't be right. You needed to start with a clean sheet of paper. And so what we started with was a new vision, and that vision could do training, and it could do inference, and it was orders of magnitude faster at both. But right now, what we're seeing is such an explosion in demand for inference that a lot of the business this minute is inference, even though we're just as fast at, you know, the same amount faster than GPUs on training. That's interesting. Maybe we'll get more to the theoretical training market a little later. Just real quick on inference, Ben Thompson, who writes a newsletter about tech, he wrote a piece in which he distinguishes between answer inference and agentic inference So answer inferences like you know format by resume or whatever or write me an essay on X or Y or answer some questions And then agentic inference is like OK here this thing that going to go around Do you distinguish and do services for you, not producing visual answers? Do you distinguish between those two? Is that a real divide in your view? And can your chips do both? Our chips can do both. I think it is a divide. OK. I think speed matters equally in both. I think if you are engaged with the AI, if you're writing code, which is agentic, if you're writing code or you're doing work, nobody wants to wait. I mean, we could just turn the question around and say, well, how big is the market for slow search? Zero. How big is the market for dial-up internet? Zero. Why is that? Because nobody wants to wait. right so if you're engaged with the ai speed is of the essence but if the ai is doing agentic work and your competitor gets three times five times ten times as much work done in 20 minutes than you do you're going to get smoked and so this notion somehow that ben proposed that speed isn very important in agentic flows is dead wrong that speed is important in all aspects of productive work And that your ability to get more done in less time is a fundamental advantage that accrues over time, right? If while your competitor is doing one unit of work, you can do three. And in the next time they do one unit of work, you do six, right? This adds up over time and you beat them in any line of work. And so speed, which is sort of our specialty, is important across the board. What do giant wafers and speed in general actually mean for, I guess, the economics of tokens? Because one way I think about it, I have this sort of vision in my head, like, OK, if I'm out shopping for toothpaste, I know I need toothpaste every once in a while. And I go into like a CVS store, I get one thing of toothpaste and then maybe a week later I get some more toothpaste. Or I could go to Costco and buy a giant thing of toothpaste and take it home probably at a cheaper cost. And that's sort of how I think of the giant wafers. Maybe it's a bad analogy. But what does speed actually mean for the cost of tokens? Well, I think there are a couple observations. [15:00] I think people have chosen so far to price speed a little higher. For example, Anthropic offered a premium service in which they offered tokens twice as fast and charged six times as much. And they sold it out and they couldn't meet the demand. Now, just to give you an idea, we're 15 times faster than they're twice as fast. And so people value speed because it allows them to do more work and they value their time. And when you can do more work in less time, you are making people more productive. That's why people have chosen to price them at a premium. They don't cost more to make. In fact, in the GPU architecture is an extremely good architecture and extremely efficient at building very slow tokens. And if you don't mind slow, the cost per token on a GPU is extremely low. But the GPU has a characteristic that as you try and go faster, the cost and the power used per token increase. Sort of like as you go faster in your car your miles per gallon decrease right So what happens is as you try and get fast enough to be useful fast enough to be interesting fast enough to keep users' intelligence focused on this product, they become extremely expensive and extremely power hungry. And so the question is not just what people are paying for a token, what people are choosing to price them at, but what they actually cost to make. And GPUs make very slow tokens very cheaply, and they're unbelievably expensive at fast tokens. We make fast tokens vastly less expensive than GPUs, and we use a tiny fraction of the power. 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 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. For certain. The U.S. Virgin Islands invites you to discover the stunning beauty of St. Thomas, St. Croix, and St. John. The U.S. Virgin Islands is truly a magical place, one of a kind, and unlike any other destination, with incredible food, warm hospitality, vibrant culture, and breathtaking beaches. It's everything you could want in a vacation. And the best part? No passport is required when traveling from the U.S. No hassle, just easy travel. Whether it's the world-famous Trunk Bay in St. John that was voted the best beach in the world, the picturesque Megan Bay Beach or the stunning Point Udall America easternmost point USVI will take you to a place that no other Caribbean destination can match For most the U Virgin Islands is just a quick nonstop flight or a simple connection away Start planning your getaway to America's Caribbean at visitusvi.com. That's visitusvi.com. USVI, naturally in rhythm. Get the news you need in just 15 minutes. Start your day with Bloomberg Daybreak, the podcast with a global view on the stories that matter. I'm Nathan Hager. And I'm Karen Moscow. Join us each morning for curated stories on current events, politics, business and foreign relations. Plus one conversation on the day's biggest developments, all in just 15 minutes. Subscribe to Bloomberg Daybreak for a precise, thoughtful take on the stories that matter. Listen to Bloomberg Daybreak each morning on Apple, Spotify or anywhere you listen. Let's say we stipulate that this is all true and everyone wants the fastest and everyone's like, you know what? This is the solution that the Cerebras technology, one big chip, this is really where it's at. How much of like your market share for the inference market when you look out next year, the year after, et cetera, how much is your market share going to be dictated by the market [20:00] your ability to get capacity at TSMC fabs. How much is that a gating mechanism for growth? TSMC is a huge part of the supply chain. But we have some real advantages. There are three areas right now that are limiting vendors and building AI compute. Number one is HBM memory. It's this memory we described earlier that can store a lot, but it's really slow. That's made by three companies approximately, Samsung, Hynix, and Micron. And it's under unbelievable supply pressure. It's extremely difficult to get. They're very long lead times. It's unbelievably expensive right now. We don't use it. The second part that's limiting is a process inside of TSMC called CoOS. And this is a process that NVIDIA and other GPUs use. We don't use it. The third thing is that at TSMC, the factory that is under most pressure is their three nanometer factory. We don't use it. We use five nanometer. So we have managed to avoid some of the most binding supply constraints Now TSMC still has to give us a meaningful allocation And they been an extraordinary partner from the get And they are the greatest manufacturing company on earth by far. A fab is sort of a modern pyramid. It's an unbelievable thing. And I highly recommend you or any of your listeners, if you get a chance to go to Taipei, go and see them. They are just extraordinary. Can you do fab tours? You can't. Oh, I didn't know that. You can do fab tours. You can go and they have a museum of innovation, and it is an extraordinary thing. They are sort of the national champion of Taiwan. But I think today, TSMC has given us as many wafers as we've needed. Business today is constrained by data centers, and that's the grand irony, right? You invent technology that has been unbuildable, never been invented for 75 years in the history of compute. You write software that is extraordinary. You built a product that is vastly faster than the incumbent. And what are we all constrained by? Buildings. Data centers right now are everybody's constraint in the entire industry. Powered buildings, so real estate. It is an amazing thing right now. And that is true sort of across the board And that will not change for the next 15 or 18 months for sure I mean since we talking physical constraints I guess I should ask you we did an episode about helium recently, a helium shortage given the situation in the Strait of Hormuz. And one of the things that helium is used for is lithography on semiconductor chips. Has that affected you at all? Or is that something that you're monitoring? We monitor, but there's not a lot we can do. And there's plenty of stuff to worry about that we can't affect. We obviously are in communication every day with TSMC. We're in communication with our entire supply chain every single day. And we stay abreast of the various issues. but it has had no impact on us. And we put that in the bucket of things that our manufacturing partners worry about also, and that we can't help. You know, so in addition to manufacturing these chips, you actually, I didn't realize this, you have your own cloud. We do. And, or you have your own cloud services. I have a bunch of questions about that, but you have your and cloud services through which a user can actually get access to various open source models and so forth It looks a little bit sort of visually it looks a lot like the open router interface roughly the same environment except it all like the open source What I'm, something I'm curious about, and maybe you can speak to this, you know, in traditional software, open source, one nice thing about open source is you don't have to pay for it. So it's free. It's a little bit different when we're talking about, there's no really such thing as like free AI software, Because even if it's like free, you still have to like pay for the depreciation of the chips and you have to pay for the electricity to run them. So there's no real such thing as like free open source software. But what I am curious about in your experience as a cloud vendor, are the open source models cheaper on a per unit of intelligence basis? if we had some way of saying levelized cost of intelligence, which I don't know if the industry has yet, are open source models cheaper per IQ point, however we want to measure intelligence? Yes, by a lot. Really? Yeah. I think in the closed source world, you're paying a lot for that extra little bit of intelligence, right? The open source models, there are no open source models that are as good as the closed source models. Okay. Think of it as three, [25:00] 4%, 5% different. Something in that range. It could be a little more, it could be a little less. But the cost to you using them, you can jump up right now and run KimiK2. It's a 1 trillion parameter model. It's an open source model on Cerebris. We're 10 or 15 times faster than others. And what you're paying for is the cost of our power and some cost of the compute that took to calculate it. What you're not paying for was the cost to train it. Right. And that's a battle that is underway in the market. You have OpenAI with their coding software. You have Anthropic with their coding software. And you've got companies like Cursor and Cognition that are using open source. We power OpenAI and we power Cognition. You have a battle underway between closed source and open source. And I think that the winners of that battle is yet to be determined. What is clear is that the closed source is strictly better by a little bit, by how much varies, and it's more expensive. Yeah I think we talked about this before but I heard of a lot of big companies in the who have been like very quietly shifting from some of the closed source models to the open source models like the Chinese ones like Kimi Is that what it's called? Kimi and Quinn. I'm sorry to press you on this point, but if you had to make a bet, like in 20 years, is the dominant aim AI model going to be a cheap open source thing or a more expensive, incrementally better closed source model? I don't think there's going to be one. There's not one SaaS software. There's some big dogs. There's Salesforce. There's some other sort of giant players, and there are lots of other specialists. I can't think of many markets where we've sort of settled onto one player. If you look at the semiconductor market, you've got x86 where you've got two major players in AMD and Intel. And then you've got a whole adjacent market owned by ARM and the companies that build ARM parts. And then you've got custom silicon around that. I think that's the way you're going to have this. We're going to have – OpenAI is going to continue to do extraordinary things. They will be competitors to them, and they will be open source. I don't think any of those go away. Since we on the topic of software one of the things you often hear when talking about you know new chip entrants going up against NVIDIA is this idea that well you know like NVIDIA chips they great and all but the real moat of NVIDIA business is CUDA right That software stack that goes with it. What's your take on that? Like, is that a realistic concern for someone who's trying to go up against a company as big and I guess as embedded in the software system as NVIDIA currently is? NVIDIA is probably the greatest company in the first part of this century, right? You know, Jensen's one of the great CEOs of our era, along with Hawk Tan at Broadcom and maybe Lisa at AMD. Just extraordinary. And CUDA was really important in the creating of the AI landscape. But it's not important now. And it has no role whatsoever in inference. If you want to move from running a model on GPUs today to running it on us, we can move it in 10 keystrokes. Just move, point to our API. So that's the first part. The second part is that a year ago every major Frontier Lab model had been built on a CUDA foundation And today two of three haven So they lost 70 market share There are three leading frontier models Gemini Claude and GPT Gemini, built by Google, on TPUs, trained on TPUs, served on TPUs, no CUDA. Anthropics models trained on Tranium, no CUDA, served on TPUs, on Tranium, and on GPUs. And OpenAI's GPT trained on GPUs in the CUDA environment. So two of the three leading models today use no CUDA. That's a hemorrhaging a share. And so I think what was true three or five years ago in which CUDA had a dominant position with Central has shrunk significantly and not important at all in inference and shrinking in its role in training. You know, since we're talking about the economics of inference and all this stuff, I would love to get your take. One of the things that literally in the last couple of weeks, there's been this flurry of announcements of these attempts to financialize the market foreclosures. [30:00] compute. And so it's like, oh, you're going to buy some capacity, the H100 benchmark, etc. And people want maybe theoretically hedging it. I'm not entirely convinced. It still seems to me like it's not like maybe, but on the other hand, an inference provider can lock in a very long-term relationship bilaterally with a data center and so forth and no need for these spot hedging markets. Do you think the market is going to evolve in such a way that there will be significant demand for financial instruments that allow inference providers to hedge their price exposure? I don't know. I'm not a financial engineer is the first thing. But we can look a little bit at history. The guys at Corweave were enormously innovative in how to fund some of their massive deployments. They were some of the first to use a debt instrument that had a backstop with the GPU. And this enabled them to really leap out and sort of have first mover advantage in the neocloud space. And that was an innovation in financial engineering and extremely creative Others followed and now there a big and active debt market in funding the building and the fit-out of data centers. When you have a market that is that big and that active, you have people who want to make bets on either side. And I think over time, those bets normalize and regularize, and you can wrap them up, and you can make it easy to make the bet. when sort of CO2 was one of the first to loan money against GPUs for CoreWeave, this was really innovative. And not only does CoreWeave get credit for the creating of the instrument, but so does the other side of the deal for doing it and making a successful innovative bet. And as sort of more and more people jumped in and these could be regularized, they could be more easily priced. And then once it's regularized and you have a market, then derivatives of that market are easy to make historically. And that's sort of the way I see this unfolding, that as this market for data centers and compute matures, there'll be people making bets on either side and financial instruments will be created to do it Whether it a good idea or not I have no opinion at this time Since we brought up finance I was looking through the IPO filing and looking at some of the actual numbers in there. And I know you have the OpenAI deal now, but a huge chunk of your revenue comes from this company called G42 in Abu Dhabi. And I think they're both like your biggest customer and also a major investor. What does G42 actually do with all these chips? Sure. Last year, they were a really important chunk of our business, a lot of it. They're a minority investor. They are the national champion, the national AI champion of the UAE. And they built a cloud that is used across the UAE's ecosystem. So it's used by leading universities there. It's used by leading companies there, companies like Adnock. They're a leading oil company. It's used by G42's nine operating companies. The deployments to date have been in the U We have massive data centers that run equipment for G42 here in Santa Clara but also in Minneapolis in Dallas Texas soon in Toronto And so they doing training and they doing inference The training they're doing, they have pioneered some of the leading English-Arabic models. They've done genomic work. They are doing serving of models and they're operating as a cloud. particularly for the UAE ecosystem, but also for global companies. 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. [35:00] 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. The U.S. Virgin Islands invites you to discover the stunning beauty of St. Thomas, St. Croix and St. John. The U.S. Virgin Islands is truly a magical place, one of a kind, and unlike any other destination with incredible food, warm hospitality, vibrant culture, and breathtaking beaches. It's everything you could want in a vacation. And the best part, no passport is required when traveling from the U.S. No hassle, just easy travel. Whether it's the world-famous Trunk Bay in St. John that was voted the best beach in the world, The picturesque Megan's Bay Beach or the stunning Point Udall, America's easternmost point. USVI will take you to a place that no other Caribbean destination can match. For most the U Virgin Islands is just a quick nonstop flight or a simple connection away Start planning your getaway to America Caribbean at visitusvi That visitusvi USVI, naturally in rhythm. Have you ever wondered how Jesse Cole took the Savannah Bananas from this? We had a $6 million failure last year. We're going to have bigger ones as we go. To this? We've got shareholders, investors that reach out to us regularly, and the answer is always no. or why L. Duncan would say this about a Netflix sports broadcast. Sometimes we're going to take really big swings and we're going to freaking whiff. Then the deal is the show for you. It's a Bloomberg podcast hosted by me, Alex Rodriguez. And me, Jason Kelly. We talk to the biggest names in the world of sports and business, including NBA Hall of Famer Tracy McGrady on one of his biggest blunders. I think I've created something magical. Well, I struck out. And you'll even get some of my baseball hot takes. I've had owners tell me it doesn't matter. The game has to be fixed. It's broken. If we have to lock out the whole year, we will. New episodes air every Thursday. Don't miss out. Do you think that over time corporate users and perhaps individual users but corporate users will want inference served from a company that separate from the model maker such that they can be certain that they are not revealing and thus training the company that might replace them I mean, look, Anthropic every couple of days announces some new thing. Oh, we have a new markdown file that could do this for tax or they could do this for whatever. And then a bunch of companies fall. Like, are companies that use AI increasingly going to want to use data centers and inference providers that aren't the model themselves? Well, first, I think there is a type of professional, a type of job that is most directly under threat from AI. Okay. And they're almost always white collar. and they required you to have expertise over a body of knowledge, right? That's what an accountant is, right? You have expertise over a body of knowledge of rulings, of previous examples of tax case law, et cetera. That exactly what AI is good at right now Exactly So lawyers accountants these professionals who have stood between sort of the ordinary person who doesn't know anything about IRS tax rules and the tax rules, that is under threat. And that is something that it will be very easy for companies like OpenAI and Anthropic to chew through. There are other areas, like, say, drug design, genetics, genomics, where companies like GlaxoSmithKline have remarkable and unique data sets. This is true for one of our large customers, Mayo Clinic. It's true for GlaxoSmithKline and other of our pharma customers. they have unique data. And they will be able to find insight in that data. And they will be able to get value from that data. And they will certainly not want to share that data with the foundation model makers unless they are guaranteed that it will not sort of make the general model smarter. And these are companies that have spent 20 or 30 years spending tens of billions, [40:00] dollars a year gathering data, right? Patient care records or test results for drug design. They're going to mine the insight in this work, and they're going to find extraordinary things. And those are much more protected because the insights in the data, and they have the data. You know, you were talking about fabs in Taiwan earlier, and I'm now regretting not going on a FAP tour when I was in Taipei, but it just didn't cross my mind at that time. Next time. Next time. Yeah, hopefully. And there have been various efforts under the Chips Act and some other industrial policies to try to build more chip making capacity in the US. In your view, what's the big, I guess, impediment to actually doing? Yeah. A, is it happening? And then B, why does it seem so difficult to actually make happen? Right. The first thing is difficult because it's a difficult problem. They're hard. They cost 30 or 40 billion dollars and take five or six years to build. So that amount of money in that amount of time cuts across administrations Right And that a problem with the politics in the US is it hard to make policy that durable across administrations and across time It's the first thing. The second thing is these are remarkably complicated buildings. And we have a sort of a hodgepodge, a sort of strange latticework of local regional building codes that a fab maker has to negotiate. Third is we're trying. TSMC has dedicated tens of billions of dollars to their fabs in Arizona and have committed hundreds of billions more. Samsung has dedicated tens of billions of dollars and committed hundreds of billions more to their fabs in Texas. But they take a long time, and we have to remain committed to building not just the fab, but the surrounding ecosystem. them, not just for three or five years, but for 20 years or 25 years, because you want not just one fab, but you want a whole trajectory of fabs. You want them working at today cutting edge but tomorrow and next year and in 10 years cutting edge as well And those are things that have proven really challenging in the U And I think we need it They strategic assets And I think we need to find ways to collaborate with those that have the expertise and to find ways to build policy that is durable over a length of time that can build a vibrant ecosystem in the fab and the associated elements. So the other big political economy theme, I guess, when it comes to semiconductors is this idea that they are, in fact, a strategically important technology. And so the U.S. should place some limitations on their use abroad. And so we've seen things like export controls, export restrictions. You're an actual chip company. And so I'm very curious at an operating level what your experience of these kind of export controls has actually been. Like, how much time does that take up for you? And then also, given that one of your biggest customers is an international firm in Abu Dhabi, like, how important is the trajectory of those export controls to your future business? I think you know three or four years ago I would have said not important at all I think today they really important In the last administration I got to know the leadership in the Department of Commerce and in the BIS Division of Commerce which oversees the licensing. I think this is an extraordinarily difficult job. And we saw really hardworking, smart people doing a job that is very, very difficult. I got to know the people in this administration and I found the same. Every single one of them is earning a tiny fraction of what they could earn in the private sector, and is doing this because they believe that this is an important mission. The problem is that there are differing views about the right way to do this. And there are differing views on the right way to achieve the goal, which is to not give your most precious technology to your industrial enemy. I think we can agree that today, in today's environment, China is an industrial enemy. Good, well-meaning people can disagree on whether the right strategy is to limit them from gaining access. Others argue, as those at NVIDIA have argued, is that the right strategy is to give them access and to keep them working on our product, on US-made, on US sort of designed product. [45:00] I come down on the other side of that argument. I understand they're good arguments in both directions. I think limiting the distribution, the diffusion of our most precious technologies makes sense. And I think we have to do it thoughtfully and we have to recognize that means some markets will be foreclosed to us. And I'm OK with that. Just quickly on the sort of like current business stuff, you mentioned the deal with AWS. How does that work? Can customers right now, like, can customers of AWS pay them to have inference service specifically on one of your chips? Not yet, but soon. Okay. It will be served in Bedrock, which is their AI as a service offering. And they will yes be able to go down the click down menu and get super fast inference which will be delivered via a combination of what called a disaggregated solution which is using some tranium for some of the inference work and using the cerebrous technology in our systems called the CS3 for other parts of the work And presumably someone who scrolls down and selects that they would pay some premium for that ultra inference I think they will pay a premium We will see. This is entirely as Amazon wishes to price it. This is their product. So you IPO'd this week. It's May 2026. This is not the first time that you've tried to or looked towards going to the IPO market. There were headlines going back to 2024 about wanting to try for the IPO market. And then there were headlines last year, especially because of the relationship with G42 about CFIUS and some of the national security concerns. And maybe that was an issue with the IPO. And then but also last September, you got one of your looks like G round G round. One of the participants in the G round investor was 1789 Capital, which is, of course, the firm that's associated with Donald Trump Jr., which is a lot of things. And then the IPO happens. I'm a cynic. So I wonder if the participation, if Donald Trump Jr.'s investment in your company made it easier to get the green light from these national security concerns to do an IPO I wish it were that easy No it had no role at all We resolved all CFIUS issues in March of 2025 I believe that was before we took money from 1789 Moreover, I wouldn't ask. That's not who I am, and that's not the way we roll. We took money because they're a thoughtful venture firm. We don't believe that there's only one point of political view. There are lots of political views. They all have some merit. They all have some weaknesses. And so we have right-leaning political, some investors. We have left-leaning. The fact that this firm had some right-leaning investors, we were looking only at their ability to help us build an extraordinary company. And we have asked, and we have never asked, nor will we ever ask for political access or anything of the kind. What's it like to become a billionaire in a single day? This is something I assume will never happen to me, so I might as well ask you No I think the honest truth is it was a big nothing for me I had some wealth before and have some wealth after I think this is a very difficult way to make money being a tech CEO I think what you have to do is you have to love the work you have to love the people, and you have to think every day about how to make your team rich. And far more important than some change in my wealth was we made more than 800 millionaires. Nice. And that's something I'm proud of every minute of every day. And at my last company, we made 100 millionaires. And at this company, through our IPO, we made more than 800. And that's something that you wake up feeling good about yourself every single day. That was going to be my last question. But actually, you just reminded me in that answer. You know, the idea that getting here, I said you became a billionaire in a day. But obviously, this was the outcome of years and years and years of work. And if we think about technological hardware, one of the things most people associate it with is really long lead times and really big research and development budgets. Now that you're a public company, how do you sort of balance that quarter to quarter financial performance pressure with the idea that you still need to be investing? [50:00] in CapEx, in new ways of designing chips, new improvements to the existing ones? First, we think the opportunity for innovation based on our way for scale engine, the best work is still ahead of us, number one. We see an opportunity for extraordinary innovation in the years ahead to make leaps every bit as big and often bigger than what we made by building the largest chip on Earth. When you love building hardware, the fact that it takes time is part of the deal, right? That what we do can't be done in a week or a month or a year. And that's what you sign up for. And that's true in every profession. You sign up for the good and the challenging. And you have to sort of make peace with that if you're a person that wants to dive in and sort of begin iterating right away and fail quickly and code up something and look at it and throw it out in the market and see if it wins. Godspeed. That's great. And that's not for me. In our business we measure twice before we cut once And you have to put that in your soul and you have to like it You have to like that mistakes in our business are really expensive And you have to like the fact that you breathe life into a chunk of silicon. And you get it to do things that nobody else has ever been able to make a chunk of silicon do. and if that's for you then this process that that takes time and money um you love that too and and so i i think i would love it less if you could do it in a week and i i think the people that that i love to work with they feel the same way and they like being engineers not because it's a path to money they like being engineers because they like building things and they like building hard things. And I like working with them for exactly that reason. Yeah, you mentioned breathing life into a chunk of silicon. My dad, who's a physicist, always likes to point out how carbon and silicon are right next to each other on the periodic table. They are. And they're sort of like, here are the two things that we have closest to life, and they're literally touching each other. Maybe there's something deep in that. I think that's a really thoughtful thing your father said. Thank you And I think that really cool And nobody pointed that out to me though We stared at periodic tables for a long time But I think to the extent we can make artificial life we need silicon Yeah. And they're right next to each other. Right. Carbon. Carbon's the heart of all other life. And artificial life will be we've founded, at least the intelligent part will be founded on silicon. Right below silicon is germanium. Maybe the next. I don't know. What does that mean? Ask your dad. Yeah, let's keep an eye on Germanium next. Andrew, thank you so much for coming on Odd Lots. Fascinating conversation right in the sweet spot of what we're interested. Really appreciate you taking your time. Hey, thank you guys for having me, and I really appreciate it. Look forward to seeing you again soon. That was really fun. I'm super interested in this topic. And it does feel to me like the economics of inference in particular and the market for inference, inference, capacity, speed. Like it's still day one. You know what I'm saying? I just like looking at the giant. It's so cool. It really does seem like an onion thing, doesn't it? It like company solves inference with a giant chip By building the biggest chip in the world But it is interesting we did that episode of course with uh ray wang from semi and talking about the role like memory as being this really important part of this sort of cutting edge chipsets and it's interesting to think it's like okay well here is a bottleneck that doesn't run into the that they don't have and the idea that at least as he described it they're not fighting to get the smallest nanometer chips and so maybe that gives them a little bit of breathing room on capacity there too yeah i i mean i do imagine there are some downsides to having giant chips in you know just as there are upsides that andrew laid out the other thing i was wondering i know he made the case for the reason speed is very important but like i can also imagine a world where maybe it's not that important you know like i think at some point like the incremental speed factor just starts to become less important when weighed against like the incremental cost of generating that extra speed i think it really this is like one of those things where it probably really depends what you're what you're using it for right so it's like if you're like you know what i'm really curious uh why pterodactyls are not [55:00] actually dinosaurs, can you explain it to me? Then it's like, I don't care about that. Like that fraction of a second is not that important. I would wait five minutes for the chat bot to tell you you're wrong, Joe. You just, buddy, you just don't really care that much. But if you're doing some sort of like a gent decoding thing or whatever, et cetera, then like, yeah, that definitely adds up. And I will say like, as you use it more, like, it's just like everything else, the treadmill of expectations. Here's some tasks that you can do in 30 seconds. which maybe several years ago would have taken you 30 minutes. And you get a patient in that 30 seconds and you want it in 10 seconds. And that's just like that competition to shave down seconds. I think it's always going to be there. So no one ever gets satisfied with this is my point. It always eventually becomes like it feels like waiting. But to me, this feels like this is the crux of the A.I. valuation argument, which is like how much of a premium are we going to place on a model that may be a closed source model that is maybe slightly better than an open source model? How much premium are we going to place on compute that is slightly faster than this other type of compute or like other use of compute Like that to me it an unanswered question And Andrew was pretty up front about closed versus open source but i think on the speed question too like we gonna find out we're gonna find out and you know the i think one of the things that is going to happen and there have been all these stories about sort of like token shock like how much companies are spending on tokens my guess is one of the things that will happen at some point is there's going to be a lot more discussion about why are we using this ultra premium model when we could have done this. Like there is a lot of just like throw it at the AI, rack up those bills, et cetera. And at some point there's going to be this like, OK, what really needs to be served fast? What really needs to be served on the most premium closed source models? And companies are probably going to get a lot more skilled at allocating from, you know, different forms of inference depending on the need. Yeah, I think that's exactly it. And at that point, like we could well see some of the dynamics in the market start to change in terms of valuation. Shall we leave it there? Let's leave it there. This has been another episode of the All Thoughts podcast. I'm Tracy Alloway. You can follow me at Tracy Alloway And I Joe Wazenthal You can follow me at The Stalwart Follow our producers Carmen Rodriguez at Carmen Armand Dashiell Bennett at Dashpot Kale Brooks at Kale Brooks and Kevin Lozano at Kevin Lloyd Lozano And for more Odd Lots content go to Bloomberg slash Odd Lots We have a daily newsletter in all of our episodes. And you can chat about all these topics 24-7 in our Discord, discord.gg slash Odd Lots. And if you enjoy Odd Lots, if you like it when we talk about giant wafers, then please leave us a positive review on your favorite podcast platform. And remember, if you're 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. 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 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. The U.S. Virgin Islands invites you to explore the stunning islands of St. Thomas, St. Croix, and St. John. USVI is truly a one-of-a-kind destination with mouth-watering food, the warmest hospitality from the locals, vibrant culture, and pristine beaches. USVI has everything you're looking for in a getaway. And the best part? No passport is required when traveling from the U.S. Go to visitusvi.com to start planning your next vacation. That's visitusvi.com. USVI, naturally in rhythm. Bloomberg Green returns to New York during Climate Week on September 22nd and 23rd, bringing together industry leaders, policymakers, and climate innovators to explore what's next in the transition to a more resilient and future-ready world. Powered by Bloomberg's editorial coverage and data-driven insights, we'll examine how climate intersects with the forces shaping our future, from technology and the economy to politics. [60:00] and community. Learn more at bloomberglive.com slash greennyradio.