Etched Tapes Out AI Inference Chip, Secures $1 Billion Demand

Etched完成AI推理芯片流片，斩获10亿美元需求 推理初创公司将通过优化定制机架与集群级内存解决兆瓦级数据传输瓶颈，从而取代通用GPU。 Gavin Uberti 和 Rob Wachen 详细介绍了 Etched 的创立历程、技术战略与快速发展。作为一家为 AI 推理开发专用硬件及整机机架的初创公司，他们分享了核心架构决策——例如 Prefill 与 Decode 解耦（Prefill-Decode Disaggregation）以及低电压推理（Low Voltage Inference），旨在绕过通用 GPU 的局限并对每瓦 Token 产出进行优化。创始人还讲述了他们如何招募业内资深专家、克服班加罗尔和台湾地区的严峻工程与供应链挑战，并筹集创纪录资金以打造吉瓦级 AI 基础设施的过程。 Etched 质疑传统半导体行业的既定假设，通过剔除标准 EDA 软件中针对 0°C 等非必要极端温区的冗余约束，从芯片与电路板设计中释放出可观的性能增益。 (3:54) 其系统架构押注于“预填充与解码解耦（PD Disag）”及“低电压推理”两项核心技术，将运行电压降至常规 GPU 的一半以下以彻底消除热降频，极大提高了实际算力利用率（MFU）。 (7:52) 面对未来千兆瓦级 AI 推理的需求瓶颈，Etched 自研了全定制互连堆栈，将多芯片点对点延迟降低 5 倍以上，打造出超高速的集群级统一内存池。 (15:00) Etched 坚持全栈垂直整合与极致并行开发，不仅派驻核心工程团队常驻班加罗尔 6 个月化解供应商延误，还通过 700 多块 FPGA 提前仿真验证，将物理芯片从交付到机架推理的时间压缩至仅 40 天。 (37:30)

Etched co-founders Gavin Uberti and Rob Wachen argue that AI inference will become the dominant cost and bottleneck of future software, requiring dedicated, full-rack hardware optimized from the ground up rather than general-purpose GPUs. Challenging Legacy Semiconductor Assumptions (3:54) Uberti and Wachen argue that traditional electronic design automation (EDA) workflows and data center equipment carry historical design buffers intended for generic environments. For instance, standard EDA tools enforce timing checks at 0°C, whereas operational AI data center chips operate at or above 80°C. Stripping these non-essential constraints across power modules, printed circuit boards (PCBs), and silicon sign-offs unlocked compounding performance gains. The founders paired unconstrained young talent with veteran chip architects to challenge convention. After challenging the 21-year-old founders to produce a functional architectural simulation, former Cypress Semiconductor CTO Mark Ross joined Etched first as an advisor and later as full-time CTO, following a $5 million seed round. They also hired Brian Kelleher, former leader of NVIDIA's HGX and DGX systems engineering, to lead rack-scale development. Prefill-Decode Disaggregation and Low-Voltage Silicon (7:52) Etched designs complete rack-scale inference systems rather than standalone PCIe accelerators, segmenting large language model processing into separate prefill and decode stages. Prefill is compute-intensive and generates the Key-Value (KV) cache, while decode is memory-bandwidth-bound and produces tokens sequentially. The architecture processes prefills in one cluster and transfers the KV cache over custom interconnects to a dedicated decode cluster. To raise Model Flops Utilization (MFU) beyond the 20% to 50% range common on standard GPUs—which throttle due to thermal density—Etched applies Denard scaling through proprietary 'Low Voltage Inference.' Operating silicon at less than half the voltage of standard GPUs cuts thermal power by a factor of four ($P \propto V^2$). For decode, Etched built a custom layer-2+ interconnect that cuts point-to-point latency by more than 5x relative to the ~4,000 nanoseconds typical on Blackwell GPU clusters, creating a unified 'Cluster Scale Memory' pool. Aggressive Prefetching and Bare-Metal Software (37:30) To compress development cycles, Etched adopted an aggressive operational strategy they call 'prefetching'—spending capital to advance downstream engineering before physical silicon arrives. They simulated full reticle designs using over 700 FPGAs across a dozen models, built thermal mock chips to test liquid-cooling cold plates, and shipped pre-configured server racks directly to customer data centers. This reduced their silicon-to-inference rack bring-up time to 40 days, compared to an industry competitor's 10 months. On the software side, Etched chose a 'kernels-first' approach instead of building generalized graph compilers for PyTorch, CUDA, or ONNX. Believing that fewer than 100 core AI model architectures will dominate global workloads, the company focuses engineering effort on hand-tuning mathematical primitives like matrix multiplications and chip-to-chip collectives directly on raw silicon, drawing engineering talent from high-frequency trading firms familiar with bare-metal programming. Macro Economics and Megawatt-Scale Inference (14:15) Uberti and Wachen project that inference will become the primary cost of goods sold for software, eventually driving trillion-dollar token production facilities. Because data center power availability is severely constrained, customer evaluations increasingly center on throughput per megawatt at interactive speeds (ISO-interactivity). Etched aims to deliver an order-of-magnitude increase in user concurrency per megawatt relative to legacy clusters. To protect its Gen 1 and Gen 2 production schedules, Etched chose a TSMC 4nm process node and alternative High Bandwidth Memory configurations, avoiding direct competition with NVIDIA's 3nm and primary memory allocations. Looking forward, the founders forecast that economic productivity will shift from measuring GDP per capita to evaluating dynamic software agents per megawatt. Etched has raised $800 million and secured over $1 billion in customer demand, yet its architectural bet relies on the premise that global AI workloads will remain concentrated around a small family of model architectures. If frontier architectures shift drastically away from transformer-style primitives or data center power constraints ease, specialized inference racks could face agility trade-offs against general-purpose silicon.

Etched 联合创始人 Gavin Uberti 和 Rob Wachen 认为，人工智能的经济重心正在从模型训练转向实时推理；通过针对数据中心工作负载定制低电压芯片与集群级互连架构，专用硬件能够在能效和并发量上显著超越通用 GPU。 破除通用设计冗余与架构解耦 (3:54) Uberti 指出，传统半导体供应链与电子设计自动化（EDA）工具为了兼容极端工业环境，默认强制执行 0°C 等低热边界检查，而数据中心 AI 加速器实际运行温度始终在 80°C 以上。消除这些无关冗余后，设计团队得以在电路和供电层面重新挖掘出大幅性能空间。针对大语言模型运行特点，Etched 将推理拆分为 Prefill（计算密集型，生成 KV 缓存）与 Decode（内存带宽与延迟密集型，逐字生成 Token）两个独立阶段，并通过专用集群互连实现“Prefill-Decode 解耦”（PD Disag）。 低电压计算与集群级内存互连 (7:52) 在 Prefill 阶段，传统 GPU 受制于高门电路活动引发的热节流，实际模型算力利用率（MFU）常年在 20% 至 50% 徘徊。Etched 利用登纳德缩放定律（功耗与电压平方成正比，即 P ∝ V²），开发出运行电压不足传统 GPU 一半的“低电压推理”供电方案，以此遏制热节流并大幅提升单位面积有效算力。在 Decode 阶段，针对 NVIDIA Blackwell 系统等传统多节点跳转中高达约 4000 纳秒的点对点延迟瓶颈，Etched 自研二层以上互连协议，将点对点延迟降低 5 倍以上，使数千颗芯片能够作为统一内存池（Cluster Scale Memory）运行。 软硬件一体化与极端工程化执行 (37:30) 与行业通行的通用图编译器路线不同，Etched 坚守“内核优先”（kernels-first）策略。他们基于全球最终仅有百余种核心大模型架构的判断，选择绕过 PyTorch 或 CUDA 的通用抽象层，由工程师（包括从高频交易领域招募的底层性能专家）直接编写矩阵乘法等裸机底层代码。在硬件落地上，Etched 坚持全栈纵向整合，自研涵盖芯片、定制冷板、电路板以及机架供电的完整推理系统，并在晶圆制造前部署超过 700 块 FPGA 仿真集群，把芯片交付后至数据中心上架运行的时间从行业常见的数月压缩至 40 天。 兆瓦瓶颈与推理驱动的经济范式 (40:00) 创始人团队主张，数据中心可用电力已成为 AI 规模扩张的根本瓶颈，硬件评估核心指标正从单芯片峰值算力转向固定交互速度下的每兆瓦并发吞吐量（ISO-interactivity）。Uberti 与 Wachen 预测，推理最终将占领全球软件的边际成本结构，甚至可能在 2027 年实现知识工作智能体数量超越人类劳动者。届时经济生产力将以“每兆瓦/每吉瓦支撑的智能体规模”来衡量，催生类似半导体晶圆厂、投资规模达数百亿美元的超大型集中式“Token 工厂”。 Etched 的商业构想建立在 Transformer 底层数学结构与主流模型范式长期保持高度稳定的假设之上。尽管其已获得 10 亿美元客户意向并完成 1 亿美元以上规模融资，但在大规模商业化落地过程中，低电压技术良率、专用架构应对未来算法突变时的适应能力，以及与握有生态霸权的 GPU 巨头正面竞争的实际结果，仍有待生产环境的进一步检验。

## Background 本期节目由 Colossus 旗下的知名播客《Invest Like the Best》（第 480 期）于 2026 年 6 月 30 日在 RSS 及各大音频平台发布。节目由 Patrick O'Shaughnessy 主持，深入对话芯片初创公司 Etched 的两位年轻创始人 Gavin Uberti 和 Rob Wachen。官方节目简介指出，他们创立于 2023 年，打破了芯片行业传统教条，在首战即成功流片并成为 ChatGPT 问世后首家达成此成就的硬件公司；公司已募集 8 亿美元，手握超 10 亿美元意向订单，致力于通过整机机架全栈自研为大模型推理提供数倍于通用硬件的高性价比与低延迟算力。 ## TL;DR Etched 创始人 Gavin Uberti 与 Rob Wachen 详尽拆解了他们为何以及如何在 21 岁时选择正面挑战英伟达等半导体霸主。团队从第一性原理出发，识别出传统芯片由于兼顾物联网与边缘场景而保留的严苛温度等冗余物理约束，开创了低电压推理（Low Voltage Inference）并自研网络协议栈构建集群级内存（Cluster Scale Memory），以极度榨干每兆瓦电力下的并发吞吐。面对早期资金枯竭、芯片模拟时钟漂移等多次生死考验，他们通过“项目制招聘”吸纳了包括前 Cypress CTO Mark Ross 和前英伟达 HGX/DGX 架构主导者 Bryan Lawler 在内的传奇老将，并以极端的全栈垂直整合和前置并行开发将常规十个月的系统打通周期压缩至 40 天。他们坚信推理将成为全球最大的单一计算市场，未来的国家与商业竞争本质上是对“思考”这一核心资产的工业化规模制造。 ## Atmosphere & style 整场对话充满了一种初生牛犊式的无畏霸气与工程老将沉稳交织的硬核张力，既有年轻极客打破陈腐规则的锐气，又有生死边缘挣扎求存的真实沉重感。两位创始人语速飞快、逻辑绵密，从微观底层的皮秒级时钟抖动、算子重写，无缝切换到万亿美元数据中心与十亿并发智能体的宏大未来图景，令听众仿佛亲临战火纷飞的芯片工程一线。而在技术狂飙之外，Rob 坦承年少患骨癌命悬一线的绝境抉择，以及父母放下一切陪护求医的至善记忆，赋予了这趟激进商业冒险一层深沉真挚的情感底色，让人感到震撼且回味悠长。 ## Bulletin - `10:00` Etched 绕过传统 EDA 冗余与通用硬件包袱，推出低电压推理（Low Voltage Inference）架构，将芯片运行电压降至主流 AI 芯片的一半以下，理论功耗暴降 75% 并消除热节流现象。 - `12:12` 团队将推理自研互联协议的跳跃延迟削减 5 倍以上，把全机架 SRAM 与 HBM 打通为“集群级内存（Cluster Scale Memory）”，使 Token 生成耗时随集群规模成比例线性压缩。 - `38:01` 通过“预取（prefetching）”策略全并行推进软件、700+ FPGA 掩模仿真、散热爆破及空机架交付，Etched 在台积电晶圆回片后仅用 40 天便完成机架级推理点亮，而行业常规需耗时 10 个月。 - `01:13:58` Etched 规避与英伟达的供应链零和博弈，第一代芯片采用台积电成熟的 4nm 制程及差异化 HBM 规格，成功完成 1.03 亿美元史上最大规模 AI 芯片 A 轮融资，累计锁定超 10 亿美元客户需求。 ## Key takeaways - `02:07` 打破半导体行业年龄教条：两位 21 岁哈佛辍学创始人凭借“必要的无知”敢于质疑历史工业冗余（如无需支持 AI 机房不可能出现的 0°C 时序冗余），撬动出代差级优化空间。 - `06:50` Prefill 与 Decode 阶段解耦（PD Disag）：针对预填充阶段高 FLOPs 密度与解码阶段内存带宽瓶颈的不同物理规律，重构机架级全栈推理方案。 - `10:00` 低电压架构打破登纳德缩放诅咒：依据功耗与电压平方成正比的物理定律，借鉴比特币 ASIC 的超低电压设计，在无热节流前提下大幅提升真实模型算力利用率（MFU）。 - `12:12` 自研互联协议与集群级内存：现有 GPU 芯片间通信延迟高达 4000 纳秒拖垮并行收益，Etched 通过重写 L2 以上协议打破通信墙，将整集群打通为超大共享内存池。 - `24:12` 推理经济学终结软件零边际成本：新增用户的边际成本由推理算力决定，未来十年推理将超越训练成为万亿美元级计算市场，成为企业核心的运营支出（OpEx）。 - `28:38` “双峰人才哲学”与项目制招聘：团队将攻克过世界级难题的行业元老（如搭建英伟达 HGX/DGX 系统的 Bryan Lawler）与执行力强悍的年轻极客深度组合，实现惊人爆发力。 - `33:40` 极限攻坚扭转流片危机：在核心供应商严重延期时，创始人率十余名核心工程师亲赴印度班加罗尔常驻半年，运转 24 小时全天候开发循环强行挽回进度。 - `40:00` 台积电不可替代的壁垒在于深度工程服务：台积电不仅制程领先，更愿意自费根据客户的工艺修改建议启动专项产线验证并推行全局落地。 - `41:40` 基础设施评估重心转向“等交互延迟下的单兆瓦并发量”：随着电力成为首要物理瓶颈，纯峰值解码速度让位于在给定兆瓦功耗下所能承载的最大 Agent 并发数。 - `45:00` 墙上时钟时间（Wall-clock Time）极限压缩：极速推理能将长达一年的 Agent 复杂科研与数学证明任务压缩至一个月内完成，引发科学探索与软件工程的连锁革命。 - `50:00` 算子优先（Kernel-First）对抗图编译器：坚决不做通用 PyTorch/CUDA 编译层，针对不超过 100 个的核心主流模型极致固化算子，优先面向自主运行的 AI 代码模型设计底层工具栈。 - `52:44` “生产即产品”的垂直整合纯粹性：相比将造芯仅作为防范“英伟达税”的科技巨头，纯硬件初创公司芯片的成败直接关乎生死，这种专注构筑了极强的人才吸引力与执行力。 - `55:40` 两周生死营救 50 皮秒时钟漂移：晶圆回厂时因跨时钟域反压失效导致频繁报错，团队在被认为物理不可能的绝境下通过微频差锁相机制逆推求解成功拯救芯片。 - `01:12:15` 推理规模化三大法则与第二代极简哲学：推导出 Flops/Watt、最低互联延迟与最大量产性三大基石，第二代架构全面转向面向吉瓦级每月产能的极简可制造性设计。 - `01:15:00` 机器思维与生物大脑的天然反转：人类大脑存储便宜而数学昂贵，硅基芯片则是数学极其廉价而数据存取搬运极其昂贵，这决定了未来模型必须演进为狂暴的多专家并行计算。 - `01:20:00` 十亿并发 Agent 与万亿美元超级 Token 工厂：预计到 2027 年知识工作 AI Agent 数量超越人类劳动力，由稀疏激活 MoE 与高速互联构成的万亿美元分布式数据中心将成为支撑人类“思考”的供应链终局。 ## Detailed walkthrough ### `02:07` 初生牛犊：21 岁挑战英伟达与打破半导体教条 主持人 Patrick O'Shaughnessy 在开场时感慨，大约三年前他第一次与 Gavin Uberti 交流时便被 Etched 的愿景深深吸引。但在当时的半导体行业，投资人与从业者的绝对共识是：尖端芯片公司根本不可能由年轻人创立。传统芯片行业往往属于四五十岁、在业内摸爬滚打一辈子、经历过无数工程陷阱且多次成功流片（tape out）的资深老将；两个年仅 21 岁的毛头小子声称能击败英伟达，在所有人眼里无异于天方夜谭。面对全行业压倒性的质疑与不信，Gavin 回忆道，想要挑战全球最强的芯片并以前所未有的速度建立一家硬件公司，团队身上确实需要一种“必要的无知”（a certain level of naivety）。正是因为不受旧思维羁绊，他们才会不断追问那些在老牌工程师看来不可撼动的底层限制：“为什么这件事不可能做到？” 深入追问后，他们发现传统半导体与数据中心产业堆叠了大量的冗余与缓冲（buffer）。从 EDA 工业软件、电源模块，到印刷电路板（PCB）设计与标准单元库，整个产业链条在过去的几十年中都在致力于设计“万金油”式的通用硬件，既要兼容超大规模数据中心，又要兼顾边缘物联网（IoT）。当面对单一、极致的特定场景时，这些历史包袱就会变成巨大的优化空间。例如，芯片时序签核（timing sign-off）与主频测试中有一个被称为“工艺角”（corners）的指标，默认的 EDA 工具为了覆盖各种极端环境，往往假设芯片需要在零度以下的极寒环境中稳定运行。但现实中的 AI 数据中心从未见过结冰的情况，服务器芯片的实际工况几乎从不低于 80 摄氏度。仅仅识别出“无需支持 0°C 全速运行”这一个冗余约束，就能撬动整套芯片逻辑的优化空间，并最终在功耗与主频上累积出质的飞跃。 - 行业普遍信条：半导体初创公司必须由拥有数十年资历与多次流片经验的行业资深人士主导。 - 传统工业惯性：EDA 工具与标准单元设计过度追求通用性，为了适配边缘计算等极端场景，保留了大量面向零度或严寒环境的时序冗余。 - 复合收益：剔除无意义的物理约束后，单点 20%、50% 乃至 2 倍的性能优化在整套系统中层层复合，最终形成面向 AI 推理的代差级优势。 ### `05:00` 从怀疑到入局：传奇 CTO 马克·罗斯的加入 在面对行业质疑时，创始团队接触到了两类截然不同的人：一类人纯粹依靠经验主义下定论，认为“二十几岁的年轻人不可能战胜全球市值顶级的芯片霸主”，任凭拿出何种论据都不愿改变成见；另一类人虽同样保持深度怀疑，但骨子里属于“求真者”（truth-seekers），愿意沉下心来审视数据和底层物理逻辑。半导体领域的泰斗级专家马克·罗斯（Mark Ross，曾任赛普拉斯半导体公司 CTO，该公司后以约 100 亿美元出售）正是后者的代表。 最初，还在哈佛大学宿舍里的 Gavin 和 Rob 找到马克，直言不讳地表示想打造一款性能远超英伟达的专用推理芯片。马克的本能反应是：“这不可能，根本行不通。”但马克并没有直接关上大门，而是向他们下战书：写出一份严谨的白皮书，并构建出一套功能仿真器（functional simulation）来验证计算逻辑。在经历无数个通宵攻坚后，团队将仿真结果摆在马克面前。马克审视后承认逻辑成立，但紧接着指出了更残酷的现实挑战：芯片研发是极度重资本的游戏，启动最少需要 300 万美元。随后，团队不仅超额融到了 500 万美元，而且随着研发持续推进，马克的角色也步步升级——从半职顾问到最终全职出任 Etched 的 CTO。 ```mermaid flowchart TD accTitle: 传奇专家马克·罗斯加入 Etched 的转变历程 accDescr: 展示马克·罗斯从最初怀疑到最终出任全职 CTO 的四个阶段转化 subgraph S1["第一阶段：经验质疑"] A["哈佛宿舍面谈<br/>提出击败英伟达"] --> B["断言不可能行不通<br/>要求提供仿真验证"] end subgraph S2["第二阶段：数据说服"] B --> C["通宵完成功能仿真<br/>提出 300 万美元启动门槛"] end subgraph S3["第三阶段：资本与技术突破"] C --> D["超额募集 500 万美元<br/>出任兼职技术顾问"] end subgraph S4["第四阶段：全面下场"] D --> E["见证工程落地进展<br/>正式出任全职 CTO"] end ``` ### `06:50` 重新定义产品：机架级推理方案与 Prefill/Decode 解耦 谈及技术架构的选择，Gavin 和 Rob 强调，Etched 从第一天起就不仅仅是“造一颗芯片”，而是在打造一整套面向大规模推理的机架级全栈解决方案（rack-scale inference solution）。这一方案横跨专用芯片、主板架构、供电系统（power delivery）、芯片间互联（interconnect）乃至支撑吉瓦级部署的量产工程能力；在他们看来，“大规模量产能力本身就是产品的一部分”。 在剖析大语言模型推理的核心瓶颈时，团队将推理细分为两个差异极大的阶段：预填充阶段（Prefill）与解码阶段（Decode）。Prefill 阶段的计算核心在于吞吐海量输入文本，其目标并非预测下一个 Token，而是将模型的上下文记忆转化为正确的键值缓存（KV Cache）；随后的 Decode 阶段则是在已生成的 KV Cache 基础之上，逐步自回归生成后续 Token。基于此，业界开始走向“预填充与解码解耦”（Prefill-Decode Disaggregation, 简称 PD Disag）：由专用的 Prefill 集群完成重度计算与记忆初始化，再将 KV Cache 传输至 Decode 集群进行高速吐字。Rob 将其通俗地比喻为“装填弹药，然后扣动扳机射击”。 ```mermaid flowchart LR accTitle: Prefill 与 Decode 阶段解耦流程 accDescr: 展示输入文本经过 Prefill 集群生成 KV Cache 并移交 Decode 集群吐字的解耦架构 subgraph PRE["Prefill 预填充集群"] T["海量输入文本"] --> PF["高算力并行计算"] PF --> KV["构建模型记忆状态<br/>KV Cache"] end subgraph TRANS["高速传输互联"] KV -->|"转移 KV Cache 数据"| DC end subgraph DEC["Decode 解码集群"] DC["读取 KV Cache"] --> OUT["自回归连续输出<br/>生成目标 Token"] end ``` 市场上许多团队在设计芯片时往往陷入非此即彼的思维惰性，简单地给自己贴上“Prefill 专用芯片”或“Decode 专用芯片”的标签，又或者在底层技术上盲目押注 3D DRAM、SRAM、全板载 DDR 共享内存池或先进制程光互联。Etched 团队系统性评估了每一种前沿路径，得出的结论是“天下没有免费的午餐”：例如 3D DRAM 虽然能突破内存带宽瓶颈，却会立刻带来极其棘手的散热难题、供应链单点依赖、混合键合（hybrid bonding）良率风险以及算力（FLOPs）被压缩的代价，从而被锁死在单一的 Decode 场景。只有对 Prefill 和 Decode 两个阶段的物理极限与系统权衡都进行彻底的推演，才能设计出真正颠覆性的推理系统。 ### `10:00` 攻克热节流：低电压推理（Low Voltage Inference） 在审视 AI 芯片的设计边界时，Etched 团队发现此前几乎没有人真正探索过面向现代大模型的全新架构空间。他们从最核心的指标出发进行反思：在 Prefill（预填充）阶段，最关键的指标是 FLOPs 及其密度，但行业不能只看纸面峰值算力，而要看真实工作负载下的**模型算力利用率（Model Flops Utilization, MFU）**。在传统 GPU 上，受限于实际负载，MFU 往往只有 20% 到 50%，且在物理上根本不可能达到 100%——因为随着算力利用率提升，开闭的晶体管数量暴增，功耗急剧攀升，芯片为了防止过热会启动自我调节并强制降频，出现严重的热节流（thermal throttling）。 ```mermaid flowchart TD accTitle: 传统GPU热节流与Etched低电压路线对比 accDescr: 对比传统架构中盲目堆算力导致的功耗激增降频，与Etched通过降低电压打破热瓶颈的逻辑链路 subgraph Legacy["传统 GPU 痛点"] L1["堆叠更多 FLOPs"] --> L2["晶体管高频翻转"] L2 --> L3["功耗激增 / 产生巨大热量"] L3 --> L4["芯片触发热节流 (降频)"] L4 --> L5["实际 MFU 仅 20%-50%"] end subgraph Etched["Etched 低电压创新"] E1["依据登纳德缩放定律"] --> E2["电压减半 = 功耗降至 1/4"] E2 --> E3["开创新型供电与电压域设计"] E3 --> E4["运行电压低于市面 AI 芯片一半"] E4 --> E5["无热节流困扰，塞入极高 FLOPs 密度"] end ``` 如果只是一味地在现有 GPU 或 AI 芯片上堆砌更多算力单元，根本无法获得性能提升。问题的核心在于**登纳德缩放定律（Dennard scaling）**：功耗与电压的平方成正比。如果将运行电压提高到 2 倍，功耗就会飙升至 4 倍；反之，若能将工作电压减半，功耗就会骤降至原来的四分之一。为此，创始人休学后直奔硅谷，拜访了数十位半导体专家和芯片大厂工程师，得到的回复全都是：“这不可能，你根本无法让芯片在比 GPU 更低的电压下运行。” 然而，比特币矿机等特定 ASIC 芯片的工作电压还不到 GPU 的四分之一，这证明在物理规律上完全行得通。通过重新设计底层的供电机制、算力组织形式、电压域（voltage domains）以及电源平面（power planes），Etched 创造了被称为**低电压推理（Low Voltage Inference）**的核心架构。其第一代技术的运行电压不到当前任何主流 AI 芯片的一半，能够在不发生热降频的前提下，在相同晶圆面积内塞入高出常规数倍的真实算力。 - **登纳德平方效应**：电压每降低 50%，功耗理论上直降 75%（降至 1/4）。 - **物理先例验证**：加密货币挖矿 ASIC 的超低电压证明了传统 GPU 架构存在历史包袱，而非物理极限。 - **全栈重构**：从电压域、电源平面到芯片封装与电路板设计全面针对现代 Transformer 负载量身打造。 --- ### `12:12` Decode 阶段的关键跃迁：集群级内存（Cluster Scale Memory） 在 Decode（解码）阶段，游戏规则完全变成了内存瓶颈：谁拥有更高的内存带宽，谁就能更快地加载模型权重和 KV Cache，从而为单个用户输出更高的每秒 Token 数（tokens per second）。Etched 认为行业在此犯了方向性错误——大家总是盯紧“单颗芯片上有多少内存带宽”，而真正决定瓶颈的是**全尺度 Scale-Up 集群上的聚合内存带宽**。 ```mermaid flowchart LR accTitle: 单芯片带宽视角与集群级内存带宽视角对比 accDescr: 展现从单卡HBM瓶颈演变为通过超低延迟自研互联打通全集群内存池的架构流程 subgraph Trad["传统集群瓶颈"] T1["GPU 间点对点通信"] -->|约 4,000 ns 延迟| T2["跨卡访问开销巨大"] T2 --> T3["8x 张量并行扩展收益大幅缩水"] end subgraph CustomStack["Etched 集群级内存"] C1["全自研互联协议栈 (L2层以上)"] -->|延迟缩减 >5x| C2["跨芯片极速访问"] C2 --> C3["全集群 SRAM 与 HBM 融为统一内存池"] C3 --> C4["单 Token 耗时随集群规模成比例下降"] end ``` 在目前的 GPU 集群架构中，跨芯片跳跃的数据搬运延迟极高。以英伟达的 Blackwell 芯片为例，芯片间点对点通信的往返延迟可能高达 4,000 纳秒左右。这意味着在部署 8 路张量并行（8x TP）时，用户端获得的生成速度远达不到理想的 8 倍扩展。为了打破这一“通信墙”，Etched 自研了完整的互连协议栈，对网络物理层之上的全部协议层进行完全定制化，将点对点延迟压缩了 5 倍以上。由此，整个集群内的 SRAM 和 HBM 得以被视作单一的连续内存池（即所谓的**集群级内存，Cluster Scale Memory**），使得随着集群规模（world size）的扩展，生成每个 Token 所需的时间能够成比例直线压缩。 - **超高通信开销**：Blackwell 点对点延迟达约 4,000 纳秒，严重拖垮多卡并行的实际收益。 - **自研互联协议**：重写第二层（Layer 2）以上的所有互联逻辑，将网络跳跃延迟削减 5 倍以上。 - **现代架构分水岭**：现有 GPU 体系基本成型于 ChatGPT 爆发之前，而 Etched 专门面向当代生成式推理重新定义了芯片互联与拓扑。 --- ### `14:15` 推理经济学：万亿级代币与计算普及化 从宏观十年维度来看，提升推理吞吐量、降低每 Token 成本以及优化“每瓦特输出 Token 数量”（tokens per watt），是突破整个人类文明生产力瓶颈的决胜盘。当前的 AI 模型已经能够解决绝大多数人类无法攻克的难题，推动科学发现、普惠医疗和全民教育落地。然而，现阶段全球仅有数百万人真正为 AI 付费，渗透率尚不足全人类千分之一。若要让数十亿人同时流畅使用顶尖大模型，基础设施必须实现数量级的跨越。 推理速度的突破对两类场景具有革命性意义： 1. **压缩思考时间**：利用推理期计算（inference-time compute）的复杂自主 Agent 解决重大科研难题时，更极致的 Decode 速度能将长达一年的挂机计算（wall-clock time）压缩至一个月内完成。 2. **支撑超高并发**：目前的硬件算力无法支撑数亿人同时在线。为了不让用户遭遇降级或排队卡顿，必须依靠多芯片互联实现几万至数万 Token/秒的高并发输出。 这正如 iPhone 带来的规模经济一样——世界上最富有的亿万富翁和普通的工薪阶层使用的是完全同一款 iPhone，这种工业级复刻让高端科技产品走向了极致的普惠。而今天的 AI 产业，依然像文艺复兴时期工匠手工打磨螺丝钉一样，依赖相对小规模的通用硬件“手工捏造”每一个 Token。Etched 的野心是构建从“晶圆到瓦特、晶体管到 Token”全栈贯通的工业化基建，打破算力供给瓶颈，让最高质量的 AI 像工业制成品一样人人可享。 - **计算压缩奇迹**：高吞吐推理可将需要一整年时钟时间的 Agent 探索任务直接缩短至单月搞定。 - **并发缺口巨大**：全球仅千分之一人口体验了付费 AI 服务；要让数十亿人同时使用编程模型或顶尖智能体，必须大幅拓宽硬件的吞吐性能边界。 - **工业化规模经济**：消除手工制造 Token 的低效，让先进智能如同现代消费电子般普及且公平。 ### `20:00` 罗布·瓦琛的生死转折：从四期骨癌到 AI 基础设施的觉醒 谈及创办 Etched 的底层动机，罗布·瓦琛（Rob Wachen）回顾了自己高二时经历的一场生死劫难。当时他在一场武术比赛中受伤，第二天便无法行走。起初他以为是骶髂关节（SI joint）出了问题，在经历了数月的物理治疗和各式扫描后，医生最终在核磁共振（MRI）中发现了他背部的一个大肿块——确诊为四期骨癌，生存概率不足 30%。接下来的两年充斥着近乎疯狂的化疗、大型手术以及重新学习走路的艰难历程。罗布坦言，这种经历彻底重塑了他对人类经验的感知边界，让人真正看清什么才是至关重要的，并在心中埋下了一个执念：如果侥幸活下来，一定要做一件能够对世界产生巨大影响的事情。 进入大学后，大语言模型的爆发让罗布找到了这个方向。当具备图像上传能力的 GPT 模型推出时，罗布翻找手机相册，找到了确诊前背部肿块的照片，并向 ChatGPT 提问：“假设你是一名专业医生，有患者背部出现这样的肿块，可能是什么情况？”模型几乎在瞬间给出诊断：“这可能是肿瘤，应立即进行 MRI 检查并前往医院。”这一刻让罗布呆立在原地——当年整个医疗体系耗费了整整六个月才查出的病因，AI 只用几秒就给出了准确判断。然而，当他兴奋地想把结果展示给父母时，屏幕上却弹出了提示：“您今天的图片额度已用尽，请升级至 Pro 套餐。”这个戏剧性的瞬间让他敏锐地意识到：这项技术将颠覆一切，但当下的算力基础设施根本无法支撑它的规模化普及。 ```mermaid flowchart TD accTitle: 罗布·瓦琛从患癌到创办 Etched 的认知链路 accDescr: 展示了罗布从癌症经历、ChatGPT 诊断瞬间到发现算力瓶颈与创办 Etched 的推导过程。 subgraph Life["生死体验与价值观重塑"] A["高二确诊四期骨癌<br/>生存率不足 30%"] --> B["历经两年化疗手术与重学走路<br/>立志做具备极高影响力的大事"] end subgraph AI_Moment["技术顿悟与算力撞墙"] B --> C["GPT 图像诊断仅耗时数秒<br/>击败现实医疗耗费的 6 个月"] C --> D["额度耗尽提示弹窗<br/>暴露出底层算力基础设施的极度匮乏"] end subgraph Insight["商业洞察与 Etched 成立"] D --> E["模型层人才拥挤、晶圆代工门槛过高<br/>现有 GPU/TPU 均为预 ChatGPT 时代的妥协架构"] E --> F["创办 Etched<br/>专为现代 Transformer 推理重构硬件体系"] end ``` - **四期骨癌的生存奇迹**：高二因武术受伤后确诊，生还率低于 30%，历经长达两年的化疗、手术与艰难的康复行走训练。 - **震撼与算力瓶颈**：医疗系统用时 6 个月才确诊的肿瘤，GPT-4 视觉模型瞬间识别并警告就医，但紧接着遭遇的“额度不足”弹窗揭示了算力基础设施的严重滞后。 - **硬件代差断层**：当时市面上所有的 GPU、TPU 与 AI 芯片，在底层架构上均是在 ChatGPT 爆发之前设计的，属于针对现代架构“削足适履”的改造，迫切需要新一代专用硬件。 --- ### `24:12` 推理即命脉：从孵化器 Prod 到万亿推理市场的十年长征 在探索硬件机会的同时，罗布还创办并运营着初创孵化器 Prod。该孵化器早期孵化了诸如 Cursor（后与 Anysphere 合并）、Mercor 以及后来的 Etched 等知名项目。在 2022 年前后陪伴这些团队成长的过程中，罗布观察到一个极其残酷的现实：所有这些最聪明的技术团队，几乎将筹集到的所有资金全都砸在了算力上。当他尝试构建自己的产品构想时，惊觉仅仅为了维持推理运行，每年就将产生数千万美元的账单，这在商业模型上完全难以为继。 传统软件公司以“边际成本（COGS）近乎为零”为傲的时代彻底终结了。在 AI 原生时代，每一位新增用户带来的边际成本都极其高昂，其核心变量完全取决于推理开销；甚至未来所有企业的运营支出（OpEx），也将随着代码智能体（Coding Agents）等 AI 员工的大规模普及而直接挂钩于推理成本。罗布由此笃信，未来十年全球将见证推理成为全世界最大的单一计算市场。十年后拔地而起的大型智算中心里，几乎所有硬件形态在今天都尚未诞生，这促使他们下定决心切入推理芯片这一最核心的瓶颈赛道。 - **Prod 孵化器沉淀**：孕育了包括 Cursor/Anysphere、Mercor 等早期明星 AI 项目，深刻洞察 AI 应用公司的真实财务痛点。 - **软件经济学重构**：新增用户的边际成本不再趋零，推理算力费用直接构成了软件企业的核心营业成本（COGS）与运营支出（OpEx）。 - **十年长征的推论**：推理终将超越训练成为万亿美元级的最大计算市场，现存数据中心的基础设施面临全面推倒重构的机遇。 --- ### `26:45` 加文·乌贝蒂的少年极客之路：从底层算子优化到集群级互联 与罗布的经历不同，加文·乌贝蒂（Gavin Uberti）展示了纯粹的系统工程天才路径。在进入哈佛大学之前，加文在 17 岁时就获得了人生第一份正式工作——在边缘 AI 初创公司 Xnor.ai 担任算子（Kernel）开发工程师。由于当时未满 18 岁在法律上无法签署具有约束力的保密协议（CIIA），公司高管只能把他单独叫进办公室，非常严肃地口头告诫：“加文，千万不要把这里的信息泄露出去。”后来，Xnor.ai 被苹果以约 2 亿美元收购；加文随后又加入了 OctoML 继续从事底层算子开发，该公司后来同样被英伟达（NVIDIA）以数亿美元收购。 在长期深耕 GPU 与底层硬件算子优化的过程中，加文洞察到了一个被绝大多数人忽视的核心秘密：在推理的自回归解码（Decode）阶段，“数学计算其实相对容易，真正决定吞吐与延迟的是数据搬运（Data Movement）”。工程师绝大部分的心血并非花在矩阵乘法等算术计算上，而是在极力压榨单颗芯片内部以及多颗芯片之间的数据搬运效率。这一底层认知成为了 Etched 架构创新的基石——他们没有盲目跟风堆砌算力单元，而是开发了集群级内存技术（Cluster-scale Memory Tech），将芯片间的互联延迟降到极低水平，从而能够极快地搬运权重与 KV Cache，大幅缩短生成后续每一个 Token 的时间，为未来“将一整年的计算量压缩至一个月完成”铺平了道路。随后，对话转向了加文高中时期在 FIRST 科技挑战赛（FTC）机器人竞赛等赛场上展现出的极强好胜心。 - **17 岁天才工程师**：因未成年无法签署保密协议，由管理层口头约定后直接负责 Xnor.ai 的算子开发；该团队后被苹果以 2 亿美元收购。 - **两度被巨头并购的履历**：在 OctoML 再次验证算子开发与编译技术，后者最终被英伟达以数亿美元纳入麾下。 - **推理的核心本质是数据搬运**：高吞吐低延迟解码的瓶颈不在算术逻辑单元，而在于单芯片与跨芯片的数据搬运。 - **Etched 的互联技术路线**：专注于集群级内存与超低时延互联技术，通过压低通信耗时实现极速 Token 解码。 ### `25:00` 机器人竞赛的启示：专注产品胜过一切繁文缛节 加文·乌伯蒂（Gavin Uberti）回忆起自己在高中参加 FIRST 科技挑战赛（FIRST Tech Challenge）时的一段关键经历。当时传统的校队通常由大约 20 人组成，大家分工协作，而赛事本身非常看重与其他队伍的社交联络、详尽的技术文档撰写以及启发他人参与等外围工作。然而，加文和他的搭档桑福德（Sanford）决定摒弃这些繁复的流程，他们设立了一个极其纯粹的目标：“我们要赢，除了造出一台得分最高的机器人之外，我们什么都不做。” 于是，他们两个人脱离了大团队，组建了一支仅有两名成员的微型队伍。在绝大多数对手都有数十人规模的赛场上，他们深知自己的精力无法分摊到公关展示或文档评分上，唯一的出路就是依靠硬核的产品表现直接碾压对手晋级。他们定下了每三个月彻底重构一次机器人的激进迭代节奏。事实证明，这种极端专注带来了惊人的成绩：他们不仅一度创下了该项赛事的全球最高得分世界纪录，还在衡量机器人赛场净得分贡献的进攻实力评分（OPR）中位列全球软件开发第三名。 - **极简团队架构**：脱离 20 人的传统大团队，由 2 人小队全权负责软硬件开发。 - **高频迭代周期**：坚持每隔 3 个月彻底重新设计并制造一次机器人。 - **竞赛战绩**：刷新当时赛事的全球最高得分纪录，OPR 软件评分位列全球第三。 --- ### `26:35` 极端垂直整合与极致并行化：最好的供应商就是没有供应商 这段机器人竞赛的经历直接塑造了 Etched 如今打造全机架规模（Rack-scale）系统的底层工程逻辑。加文指出，核心法则就是“速度、速度、速度（velocity, velocity, velocity）——唯有靠交付产品才能获胜”。大公司即便坐拥两万人，也往往被繁冗的沟通与外宣拖垮，而一家初创公司只要心无旁骛地扑在产品上，并将工作流进行无情的极限并行化，完全可以用极少的人手造出世界顶尖的产品。 罗伯·瓦申（Rob Wachen）进一步提到了工程哲学上的演化：人们常说“最好的零件就是没有零件”，而对 Etched 来说，“最好的供应商就是没有供应商（the best vendor is no vendor）”。为了榨取极致性能并换取最快的迭代速度，Etched 坚决推行极端的垂直整合，从自研专用芯片、系统主板、液冷均热板（Cold plates）、互连架构，甚至到整机生产装配，全部尽可能自主完成。加文和罗伯相信，Etched 是目前唯一一家在自研芯片的同时独立打造整套机架系统的初创公司。 为了不在芯片流片制造的漫长周期中空转等待，他们引进了曾主导英伟达 HGX 与 DGX 系统架构（贡献了英伟达约 80% 营收）的行业老将布莱恩·劳勒（Bryan Lawler），在芯片交付前就并行展开了多个版本的机架设计。为了验证散热性能，他们特制了发热热点与真实芯片完全一致的模拟“热芯片”，提前装入定制液冷板进行过压极限爆破测试；因此当台积电制造的真实芯片返回时，液冷板已通过全面验证，未发生过哪怕一次漏液事故。他们在台湾设立了拥有数十名工程师的工厂，在办公室完整克隆了测试产线，并在自家办公楼层内硬生生塞入了一座 2 兆瓦（2MW）级的数据中心，工程师分白班与夜班展开 7×24 小时轮流攻坚。 ```mermaid flowchart TD accTitle: Etched 极端并行化工程流水线 accDescr: 展示芯片与机架系统并行研发、前置散热测试以及全球 24/7 全天候工程验证流程 subgraph HardwareTrack ["芯片研发线"] A["SoC 架构设计与流片"] --> B["特制模拟热芯片<br/>(模拟芯片核心热点)"] end subgraph RackTrack ["机架与液冷系统线"] C["Bryan Lawler 领衔<br/>全机架架构设计"] --> D["多版本机架迭代<br/>与自研液冷板制造"] end subgraph Validation ["前置极限验证与生产"] B & D --> E["液冷板极限过压爆破测试<br/>(确保真实回片零泄漏)"] E --> F["台湾装配工厂 + 办公区 2MW 数据中心<br/>白班/夜班 24/7 不间断点亮测试"] end ``` - **全栈自研覆盖**：涵盖专用加速芯片、主板 PCB、高规格液冷散热板、网络互连及机架制造。 - **前置热力模拟**：通过制造与量产芯片发热特性一致的模拟 Dummy 芯片，在芯片出厂前完成液冷板过压爆破验证，实现真片装配“零漏液”。 - **全天候研发基础设施**：在办公楼内自建 2 兆瓦（MW）高密度数据中心，结合台湾工厂产线与两班倒工作制实现 24 小时连续硬件调试。 --- ### `28:38` 招募行业传奇的“项目制招聘法”与双峰人才理念 面对两位年仅二十出头、缺乏深厚资历的年轻创始人如何网罗顶尖行业老兵的提问，罗伯拆解了 Etched 独特的“双峰人才哲学（Bimodal talent philosophy）”。当团队面临前所未有的工程天堑时，行业内排名第一的专家与排名第十、第一百的人之间存在着决定成败的鸿沟，因此他们必须不惜代价寻找特定领域中“全球最强的那个人”。 为此，他们建立了一套被称为“项目制招聘（Project recruiting）”的精准寻访机制。团队系统性地梳理整个人类工程史中所有同级别硬核技术攻坚案例，严格按时间线顺藤摸瓜：不仅寻找表面上名义挂帅的高管，更要揪出当年在幕后真正从 0 到 1 啃下骨头的核心攻坚者。接着，他们与大量工程大牛建立联系，并保持极其惊人的跟踪韧性。罗伯直言：“第一次聊就答应加入的人少之又少，但聊到第二十次才点头的人却多得惊人。” 在面对顶尖专家的拒绝时，加文和罗伯认为对顶级技术人才最有效的说服方式绝非画饼，而是不断兑现此前吹下的“狂妄目标”。当对方说“不”时，本质上是在等待团队拿出更扎实的阶段性成果；每一次如期撞线里程碑，都会显著击碎老将们的怀疑并建立真正的信念。正如当初决定研发高功率密度机架时，他们直接对着行业天花板反向推导——寻找那位从零构建起英伟达多代机架系统的灵魂人物，并最终成功打动了布莱恩·劳勒加盟。 - **双峰人才策略**：一端是经验深厚、攻克过世界级难题的行业元老（Legends），另一端是执行力强悍、精力旺盛的年轻顶尖工程师。 - **项目制招聘法**：深挖历史工程壮举的时间线与贡献度，区分挂名管理者与实际动手解决 0 到 1 难题的关键功臣。 - **长线攻坚与里程碑对齐**：通过多达 20 次的长期沟通磨合，用连续兑现看似激进的技术里程碑来彻底赢得行业顶尖人才的信任。 ### `30:00` 行业传奇搭配“初生牛犊”：Etched 的极致招聘哲学 在组建团队时，Etched 聚焦于寻找那些既具备初创公司拼劲、又亲眼见过大规模化交付的顶尖专家。他们系统梳理了 NVIDIA 所有机架级（rack-scale）产品相关的工程团队，锁定了三位最理想的目标人选。接触之后发现，其中两位已经退休，第三位则计划再为 NVIDIA 交付一代产品后告别职场——他就是 Brian。经过团队的极力游说，Brian 最终决定加入。Brian 当年在 NVIDIA 一手创立了 HGX 与 DGX 团队，这两个产品线如今支撑着 NVIDIA 每个季度数百亿美元的营收大头；而另外两位退休的元老后来也成为了 Etched 的投资人。行业老将的最大价值在于，他们亲历过规模化，明确知道“极致的标准是什么”；在日常工作中，Brian 常常指出某个设计细节说“这是我交过十亿美元学费换来的教训”，仅凭一句话就为初创团队省下数月的试错周期。 与 Brian 这种“传奇元老”搭档的，是充满野心、毫无思想包袱的年轻人。创始人 Gavin 高中时的世界机器人竞赛队友 Sanford 即是典型代表。当时 Sanford 刚读完大学四年级，接到电话邀请来公司帮忙看一周平台硬件问题。团队直截了当地问他：“你能不能这周把冷板（cold plate）做出来？”在传统热力学工程师看来，这种要求堪称荒谬无知，因为通常冷板研发需要数月之久。但由于 Sanford 根本不知道常规流程的阻碍在哪，他硬是在一周内手工拼装出一套机械装置，直接化解了公司在功耗散热上的核心隐患。Etched 内部将这种组合称为“心怀不忿，方能让芯片落地机房（chips on shoulders, put chips in data centers）”——一边是敢想敢干、从第一性原理不断追问“为什么”的无畏年轻人，另一边是摸透规模化陷阱、富有创业拼劲的行业传奇，两者紧密共事碰撞出惊人的工程爆发力。 由于公司目标极其反共识，招聘过程本身就形成了高度自选择的漏斗。在半导体行业，一名资深工程师放弃可预测的高薪与流动性资产，说服全家搬迁到圣何塞，住进人才公寓，去给两个年仅 24 岁的年轻创始人打工，在一个芯片未产出、供应链极度紧缺的环境下正面挑战全球芯片巨头，并且不是为了做出 10% 的改善，而是要实现 10 倍的飞跃——从纸面上看“脑子绝对有点不正常”。但正是这种严苛与不可思议，自然过滤掉了投机分子，筛选出一批被使命感驱动、渴望向信任者证明自己潜力的纯粹极客。 - **NVIDIA 核心元老 Brian 的分量**：曾创立 HGX 和 DGX 团队，主导了构筑 NVIDIA 季度数百亿美元营收的核心硬件平台。 - **神童搭档 Sanford 的惊人速度**：在短短一周内从零构建出冷板测试原型，用初生牛犊式的无畏打破常规硬件开发节奏。 - **反共识的自我筛选机制**：放弃稳定股票与舒适大厂生活、加入仅由 24 岁年轻人领导的高风险芯片公司，极大程度剔除了投机主义者。 --- ### `33:40` 远征班加罗尔：六个月极限攻坚扭转流片危机 当话题转向公司的首款产品 Sohu 时，创始人指出，单颗芯片固然关键，但 Etched 真正的底层核心是“打造一家能够像机器一样持续高效、极高质量交付各代产品的公司”。在这种崇尚极致速度与纵向整合的哲学下，团队展示出了惊人的破局冒险能力。在 Sohu 临近流片（tape-out）的关键节点，团队突然发现一家核心外包供应商的进度严重滞后。摆在面前的原本只有两个糟糕的选择：要么继续等待原供应商，整体进度延期近一年；要么换掉供应商从头再来，同样要推迟整整一年。 ```mermaid flowchart TD accTitle: 应对供应商滞后的决策路径 accDescr: 展示面对严重延期的常规方案与 Etched 采取的远征班加罗尔突击方案对比 A["供应商进度严重滞后<br/>面临流片延期风险"] --> B["传统方案 A：继续等待<br/>（延期约 1 年）"] A --> C["传统方案 B：更换供应商<br/>（重新开始，延期约 1 年）"] A --> D["Etched 第三种选择：<br/>团队肉身下场接管"] subgraph D["极限突击方案"] D1["派驻十几名核心工程师<br/>赴班加罗尔常驻半年"] D2["每天早晨第一个进驻对方办公室<br/>深夜 1 点穿行空荡街道返回"] D3["现场审计海量代码<br/>就地进行关键架构裁决"] D4["美印双向 12 小时无缝交接<br/>运转 24 小时全天候开发循环"] D1 --> D2 --> D3 --> D4 end ``` Etched 拒绝接受这两条死路，毅然选择了“第三种方案”：直接包机将十几位顶尖工程师派驻到该供应商所在的印度班加罗尔，一驻扎就是整整六个月。创始人 Gavin 本人也深入一线，在班加罗尔生活了四个半月。团队每天清晨穿过拥堵的街道第一个冲进对方办公室，不仅现场审计海量代码、开发专属加速工具，更直接就地拍板所有的架构设计分歧，彻底消除跨洋沟通产生的 12 小时时差拖延。每天深夜 1 点，大家才穿过空旷冷清的街道回宿舍歇息。与此同时，留在美国本土的工程团队与班加罗尔无缝衔接，早晚 8 点各开一次 Zoom 交接进度，建立起“我一觉醒来这个模块必须搞定”的 24 小时全天候开发循环。与 Etched 处于同阶段、依赖同一供应商的其他芯片公司，许多项目一拖数年至今仍未能流片；而 Etched 靠着极端的现场聚焦力彻底掌控了命运。 - **亲赴一线的决心**：十几名核心工程师派驻班加罗尔半年，创始人 Gavin 常驻 4.5 个月。 - **全天候双班开发机制**：美印两地早晚 8 点例会交接，软件与架构设计实现 24 小时无停歇推进。 - **决策速度压倒完美主义**：将巨大权责下放给一线团队，宁可在保证快速响应的前提下偶尔犯错，也绝不因追求完美答案让工程陷入停滞等待。 --- ### `38:01` Prefetching 哲学：用资本并发展开，将 10 个月的打通周期压缩至 40 天 在芯片开发领域，“通过实践学习（learning by doing）”往往需要巨大的资金投入与承担风险的魄力。当前全球 AI 行业每天产生逾十亿美元的算力支出，其中巨量份额流向推理领域；对 Etched 而言，推迟上市哪怕一天，都是在将数以亿计的商业机会拱手让人。因此，只要能带来清晰确定的高投资回报率（ROI），团队就极度愿意“烧钱抢时间”。他们将计算机科学中的“预取（prefetching）”概念移植到企业项目管理中：在等待上游某项关键依赖（如晶圆回片）的同时，把所有后续依赖项在没有实体芯片的情况下全部提前并行完成。 ```mermaid flowchart LR accTitle: 基于 Prefetching 的硬件并行开发流 accDescr: 展示在芯片制造期间，软件、系统、机架与生产线全面并行前置的流程 subgraph Parallel["芯片制造周期内前置推进（Prefetching）"] P1["700+ FPGA 集群<br/>全光刻掩模模拟<br/>跑通十余个推理模型"] P2["制造热仿真假芯片<br/>前置定制冷板散热模具"] P3["提前发运完整空机架至客户机房<br/>打通供电、CPU 与存储软件栈"] P4["多轮电路板迭代<br/>产线装配与测试工具就绪"] end Silicon["台积电晶圆回片"] --> Integration["芯片插入验证"] Parallel --> Integration Integration --> Result["40 天完成机架级推理点亮<br/>（行业常规耗时 10 个月）"] ``` 为了实现“芯片一旦出厂，其余一切早已就绪”的极速点亮状态，Etched 投入重金进行了全面的前置工程： - **全栈软件模拟与验证**：搭建了由 700 多块 FPGA 组成的超大集群，在芯片流片期间直接对整张满版曝光尺寸（full-reticle）芯片进行硬件级仿真，提前在其上跑通了完整的推理软件栈和十几种主流大模型。 - **提前部署机架硬件至客户机房**：在没有芯片的情况下，先把配齐所有供电、网络交换机、CPU 和存储的完整空机架直接运送到客户数据中心，提前几个月调通底层机房软件环境。 - **热特性与产线准备**：专门制造了一批模拟真实热特性的“热仿真芯片（thermal chip）”，提前完成定制冷板的开模与压力测试，并完成印刷电路板（PCB）的多轮改版，让完整产线处于待命状态。 凭借这种不惜成本的全面并行策略，当物理芯片真正从台积电产线返回时，整个配套环境已经变成了最平淡无奇的组装作业。行业内另一家声名显赫的 AI 芯片独角兽，从拿到硅片到让整机机架跑通推理整整耗费了 10 个月并公开向投资人通报；而 Etched 仅用了 **40 天** 就彻底搞定。在冲刺阶段，硬件实验室执行严格的日夜两班倒制度（早 10 点至午夜、午夜至次日早 10 点），以绝对的战术执行力完成了对传统半导体研发周期的降维打击。 ### `40:00` 芯片供应链的本质是深度协作：台积电不可替代的“客户服务” 为了保持极致的迭代节奏，Etched 超过一半的员工直接搬到了办公室隔壁居住。公司通过提供额外的薪酬补贴来鼓励这种选择，创始人打趣道：“看不见的手创造了奇迹”，创始人与员工一同日夜驻守在研发一线。当被问及当前轰动全球的半导体供应链风险与供应短缺时，两位创始人指出，如今被华尔街日报头版疯抢的内存芯片股票，在五年前不过是被视为乏善可陈的普通大宗商品（commodity）。但在硬件制造的真实世界里，供应链绝非单纯的“一手交钱一手交货”，买完就与供应商断联的模式根本行不通；能否成功，完全取决于与供应商之间能否建立极为紧密的工程协作。 以台积电（TSMC）为例，公众通常将其断层式领先归功于先进技术，但在 Etched 看来，台积电真正被外界严重低估的核心壁垒是其无可比拟的“客户服务”。如果客户向他们提出建议：“通过这样调整工艺，你们的良率（yield）可能会提高”，台积电甚至愿意自掏腰包启动专项验证实验。在 Etched 的实际案例中，当实验证明 Etched 的工艺修改建议确实有效后，台积电便果断将这一改动推行至整条生产线。这种开放且敏捷的响应在传统工业（如钢铁厂）是不可想象的，这也是台积电能够在全球代工领域稳坐第一把交椅的根本原因。 ```mermaid flowchart TD accTitle: Semiconductor Supply Chain Collaboration Model accDescr: 展现Etched与台积电等顶级供应商之间的深度工程共创与技术协作反馈循环。 subgraph HardwareDev["Etched 研发端"] A["提出工艺与架构优化方案"] D["获得良率提升与高能效芯片"] end subgraph FoundrySupport["台积电 协作生态"] B["自费建立测试验证实验线"] C["验证成功后全面推广至生产线"] end A --> B B --> C C --> D ``` - **全员紧密驻场**：公司超过半数工程师住在办公室步行范围内，公司通过薪资激励维持高密度面对面作战。 - **重新定义台积电价值**：不仅拥有全球最顶尖的半导体制程，更具备罕见的客户服务与产线协同实验机制，客户提出的工艺改进方案会直接在产线上测试并落地。 --- ### `41:40` 算力瓶颈转向电力：从追求纯速度到榨干每一兆瓦的并发量 在当下的数据中心扩张中，最大的硬约束已经转变为电力可用性（power availability）以及获电时间（time to power）。随着需求功率的剧增，电力供应的短缺呈指数级上升。正如 xAI 的 Colossus 集群为 Blackwell 芯片开出每小时 12 美元的高价，原因在于它是市场上唯一能一次性交付 2 万张芯片现货的地方；同理，500 兆瓦级的数据中心同样凤毛麟角。在新建建筑和电网接入受制于物理周期的情况下，从 100 兆瓦扩容到 1 吉瓦尚有路径，但从 1 吉瓦迈向 10 吉瓦乃至 100 吉瓦则在逼近极限。因此，整个行业正在从单纯向外堆叠机房（scale-out）转向在现有数据中心内部压榨极限密度（scale-in）。 面对机架物理空间耗尽和电力见顶的现实，客户评估算力部署的标准正在发生深刻转变。过去的基础设施热潮单纯追求解码速度，诸如基于 SRAM 的芯片能跑出每秒数千 token 的极致低延迟，但这仅解决了“速度”问题。而在真实的生产环境中，客户的核心考量是：在给定的能耗指标下，在保障特定交互速度（iso-interactivity）的前提下，系统到底能支撑多少并发用户（concurrency）？换言之，在一个 100 兆瓦的数据中心里，系统能同时并行跑多少个软件智能体（software agents）？Etched 的专用架构正是为了在同等交互延迟下提供高出一个数量级的并发处理能力，直接在混合专家模型（MoE）的大规模服务场景中，将“每瓦 token 数”与“每美元 token 数”推向全新的高度。 - **电力短缺与稀缺性溢价**：大规模机房断电成为常态，具备即时交付能力的超大规模集群（如 Colossus 集群）享有极高的定价权。 - **等交互延迟（Iso-interactivity）评估法**：生产环境不再孤立比拼单用户峰值速度，而是比拼在严格限定的交互延迟曲线上，单机架与兆瓦功耗所能承载的最大并发会话数。 - **数量级跃迁**：相比于英伟达 Blackwell 及 AMD 的交互性能曲线，Etched 在主流大模型并发服务上实现了约一个数量级（10x）的并发吞吐优势。 --- ### `45:00` 墙上时钟时间的极限压缩：从数学猜想证明到万亿智能体并行 算力性能曲线的巨幅外推，其意义远不止让聊天机器人生成文字快上 50%，而是彻底重塑解决前沿复杂任务的“墙上时钟时间”（wall clock time，即从启动运行到拿到最终结果的真实物理时间）。在顶尖实验室的研究员眼中，墙上时钟时间是研发迭代的生命线。以往需要耗费数月甚至数千年的数学猜想推导（例如大学时期曾讨论过的单位距离猜想等前沿数学难题），如今借助 AI 模型的深度推理已经实现突破。如果能将底层模型的运行速度提升 10 倍，原本长达 6 个月的实验将被压缩到 2 个月，研究人员就能在此基础上以极高的频次调整模型架构，引爆科学发现的连锁反应。 这一逻辑在长程复杂任务（long-horizon tasks）中尤为明显。此前 Cursor 曾展示过由多个代码智能体（coding agents）在一周内从零纯手工构建出一个完整的 Web 浏览器，而随着推理延迟的指数级下降，这类复杂的工程创举很快将在 1 小时内完成。原本耗时一整年的计算任务被压缩为一个月，一个月的计算被压缩至三天，三天的计算则可在七小时内收尾。超低延迟解码技术让海量智能体的大规模并行协同成为可能，使得软件开发与科学探索的时间尺度发生不可逆的坍缩。 - **压缩物理迭代周期**：从 1 年 $\rightarrow$ 1 个月 $\rightarrow$ 3 天 $\rightarrow$ 7 小时，核心关注点从短期交互体验升维至长程任务的落地周期。 - **数学与科学突破的催化剂**：通过极速推理模型跑通复杂的符号推演与多分支路径探索，原本需要上百年论证的数学猜想能在极短物理时间内完成证明。 - **智能体并行化革命**：多智能体并发协同将从“耗时数天构建浏览器”进化到“单小时内自动化交付完整工程系统”。 --- ### `46:30` 物理极限下的巨大空间：万亿级超大规模“Token 工厂”与内核工程协作 当探讨计算硬件的终极物理天花板时，两位创始人引用了著名的物理学格言“底部的空间还很大”（There's plenty of room at the bottom）。以芯片间互连延迟（chip-to-chip latency）为例，当前英伟达顶级产品的芯片间延迟约为 4,000 纳秒，而在物理光速的理论极限下，这一过程实际上只需 2 到 3 纳秒，中间存在着数千倍的工程优化裕量。在能耗方面，尽管通过大幅压低工作电压已经削减了大量功耗，但理论上仍有极其广阔的下探空间。与此同时，芯片集群的扩展域（scale-up domain）正在突破限制，从过去的 8 卡拓扑到英伟达 NVL72，未来的超大集群完全可以借鉴单体耗资 400 亿美元的半导体晶圆厂形态，构建出由单一巨型建筑容纳、专注于服务极少数主力模型的专用“超级 Token 工厂”（mega token factory），以此榨干规模经济效应。 而在微观的底层执行层面，内核工程（kernel engineering）正成为从裸机硬件中榨取极致性能的关键战场。面对矩阵乘法（matmul）、卷积以及芯片间集合通信（collectives）等基础算子，内核工程师的核心使命是通过内存重分配、计算与通信重叠（overlap）来消除停顿；哪怕是应对网络丢包重传，也要确保不让整个流水线陷入停滞。如今，这类细粒度优化虽然正演变为人类专家与 AI 系统的混合共创（hybrid collaboration），但每一次精细调优所带来的 3% 至 4% 的微小提升，层层叠加后便构成了底层性能的巨大鸿沟。Etched 在构建其全栈软件生态时，正是瞄准了这种将软硬件深度绑定的破局路径。 - **物理潜力的巨大代差**：芯片间互连延迟目前存在 4,000 纳秒到 2-3 纳秒（光速极限）的巨大压缩空间。 - **形态演进**：未来超算集群或将呈现类晶圆厂的单体巨型架构，作为高度集约的 Token 制造工厂运作。 - **内核级精细调优**：通过算子重叠与流水线防阻塞优化，单次调优可带来 3%-4% 的边际增益，目前顶级算子内核仍由“人类资深专家 + AI”的协同模式主导编写。 ### `50:00` 软件哲学与 Kernel-First 策略：为未来的 AI 编码模型打造基础设施 在软件工程的路径选择上，三年前业内主要存在两条截然相反的路线：一条路是重度投入图编译器（Graph Compilers），其优势在于“开箱即用”，无需人类工程师手动微调每个算子内核（Kernels），但缺点是硬件执行效率低下；Etched 选择了完全相反的另一条路——**算子优先（Kernels-first）编程**。这种方式虽然在一开始无法做到开箱即用，门槛极高，但只要有顶尖的算子专家介入，就能压榨出极致的性能。 这一技术豪赌契合了代码生成模型的发展趋势。随着编码大模型能力飞速进化，模型正逐步接管算子内核的编写任务，并终将达到超越人类的超人水平（Superhuman）。Etched 的系统设计直接面向这种未来：无论是性能剖析工具（Profiling tools）还是调试技术栈（Debugging stack），他们优先考虑的是“AI 模型将如何使用这些工具”，而非人类工程师的使用习惯。内部实验中，他们曾让 Codex 仅凭 Etched 的文档，就在完全无人干预的情况下通宵运行，从零独立跑通了开源大模型（GPT-OSS）。 Etched 将这种策略称为“博弈选择”（Game Selection）——既然无论选哪条赛道都必须倾注全力，就必须确保押注在正确的方向上。团队从一开始就明确做减法：坚决不构建通用图编译器，坚决不支持任意形态的 PyTorch、CUDA 或 ONNX 计算图。他们坚信，未来全球真正重要的大模型不会超过 100 个，且它们在底层数学结构上高度同质化；Etched 的使命就是利用底层物理学原理打造专用原语，将其推理加速推向物理极限，同时允许最硬核的技术客户直触硬件底层。这种不设编译器的理念最初饱受质疑，唯独得到了高频交易（HFT）群体的强烈共鸣——高频量化交易员同样痛恨编译器，全凭手写底层算子来抢夺微秒级优势。如今，已有数十位来自高频交易公司的顶级工程师加盟 Etched。 - **减法原则**：拒绝为了向下兼容通用生态而牺牲性能，明确不提供对任意计算图与框架的通用编译层。 - **目标收敛**：预判未来具有实质影响力的大模型数量在 100 个以内，依托底层数学一致性进行硬件原语极致固化。 - **工具转向**：开发环境与调试接口从“面向人类开发者”全面转向“面向大语言模型自治”。 --- ### `52:44` 垂直整合的边界：“生产即产品”与全栈存亡的纯粹性 当前 AI 芯片市场的格局异常微妙：绝大部分算力芯片被极少数头部客户垄断，而这些客户又几乎都在自主研发专用 AI 芯片（例如 OpenAI 推进的 Jalapeno 项目）。面对这种芯片买家、芯片卖家与云服务商互相渗透的局面，Etched 确立了自己的商业信条——**“生产即产品”（Production is the product）**。在他们看来，推理注定是全球最大的单一市场，未来谁能生产出最大规模的 Token，谁就将成为全球市值最高的公司。 为了以最高速度上线最大规模的 Token 吞吐产能，Etched 在产业链整合上极其克制与清醒：只做规模化所必需的关键环节，其余一概剔除。他们自主设计整机机架（Rack）而非仅卖裸片，选择代工制造模式（CM, Contract Manufacturer）而非联合设计制造模式（JDM），是因为系统级瓶颈必须由自己打破；但他们眼下绝不自建数据中心，因为客户为了获得 Etched 芯片带来的超高吞吐量，愿意主动腾出电力、调整算力集群来部署其机架。 ```mermaid flowchart TD accTitle: Etched 垂直整合边界与生态定位 accDescr: 展示 Etched 在底层模型层与上层基础设施之间的战略专注区间，凸显其在芯片到机架层面的全栈自研。 subgraph TopLayer ["顶层：模型构建层"] A["前沿基础大模型 (Frontier Models)"] end subgraph EtchedScope ["Etched 专注与整合区间"] B["深度定制通信与机架系统 (Rack / Interconnect)"] C["算子优先驱动与工具链 (Kernels & Model-first Tools)"] D["Transformer 专用推理芯片 (ASIC)"] end subgraph BottomLayer ["底层基础设施与制造"] E["晶圆代工与先进制程 (TSMC)"] F["数据中心与超大规模电力基础设施 (Hyperscalers)"] end A -.->|高吞吐 Token 需求| B B --> C C --> D D -.->|晶圆流片| E B -.->|机架交付入驻| F ``` 这种专注带来了不可思议的人才吸引力。近期，某前沿实验室负责下一代自研 AI 芯片的技术主管试图挖角 Etched 的系统架构师，Etched 反手打出一张“Uno 倒转卡”，仅用一周时间便将这位主管反向招入麾下。在最终敲定录用并散步交谈时，对方道出了离开原厂的根本原因：“对我所在的大厂而言，这款自研芯片成不成功，根本关乎不到公司的存亡。” - 谷歌即便 TPU 项目失败，搜索业务的基本盘依然稳固； - Meta 就算 MTIA 项目流产，核心广告业务照常运转； - 微软即便 Maia 折戟、OpenAI 哪怕 Jalapeno 不顺，也不会因此倒闭。 全球最顶级的芯片由一家“只造芯片”的公司（NVIDIA）做出来，绝非偶然。对大厂与超大规模云厂商（Hyperscalers）而言，造芯只求达到相近性能以逃避高昂的“英伟达税”（NVIDIA tax），其 FP8 算力密度往往连 NVIDIA Blackwell B300 都赶不上；而对 Etched 来说，上线最大化的 Token 产能是唯一的生死战。 --- ### `55:40` 惊魂两周：50 皮秒的时钟漂移危机与“假设一切皆有可能” 在将芯片推向现实的过程中，团队遭遇了创业以来最险象环生的技术悬崖。在流片前的验证阶段，Etched 搭建了规模极其庞大的 FPGA 验证集群，但 FPGA 只能验证纯数字逻辑，无法测试芯片上的模拟电路特性。当第一批流片晶圆回厂点亮时，芯片在运行数据留存应用（Retention application）时开始频繁爆出计算错误。 经过排查，团队发现跨时钟域（Clock Domain Crossing）的反压控制逻辑彻底失效。要修复这一致命问题，唯一的物理途径是将芯片上两个时钟信号的对齐误差控制在 **50 皮秒（Picosecond，即 50 万亿分之一秒）** 之内，并且在每颗芯片上以每秒 20 亿次的频率持续稳定维持这一精度。 许多资深工程师认为这在物理上根本无法做到，内部甚至有核心人员因绝望而直接辞职：“这道题无解，祝大家好运。”在至暗时刻，Etched 确立了攻坚准则：**首先无条件假设该问题在物理上必定有解，然后逆推求解路径。** 1. 团队提出微调设想：如何让时钟相位以 1 到 10 皮秒的极小颗粒度移动？ 2. 他们设计了一种时钟漂移机制，让两路时钟保持极微小的频差，静待相位漂移至完美重合的特定时间窗口； 3. 一旦相位在 50 皮秒公差内重叠，立即执行锁相锁定，彻底杜绝逻辑误判。 这场拯救芯片的攻坚战整整持续了两个极度煎熬的星期，最终方案完美生效。这一死里逃生的经历深刻塑造了 Etched 的工程哲学：“假设它是可能的”——假设单颗芯片可以塞进成倍的算力，假设芯片互联延迟能够大幅压低，假设能够构建出超高带宽的共享内存池。工程探索往往伴随着数十次彻底失败的实验，但正如 Gavin 所言：“哪怕绝大多数实验都以失败告终，我只要赌对一次就够了。” - **时间标尺**：50 皮秒是 1 秒的五百分之一微秒（50 × 10⁻¹² 秒）。 - **运行工况**：必须在 2 GHz 运行频率下、数十亿次脉冲周期中保持绝对时序锁定。 - **容错思维**：通过冗余并行方案压低系统风险，流片板级点亮实验中准备的 3 套备用硬件方案全部实现突破。 --- ### `59:20` 早期融资的至暗时刻：迈向物理设计阶段 硬件初创公司的资本消耗规模与纯软件企业存在天壤之别。对于两位年轻的初次创业者而言，要让外界相信两个刚过 20 岁的年轻人能挑战世界上体量最大的芯片巨头，融资过程经历了多轮濒临绝境的近死时刻（Near-death moments）。 在 2024 年初启动大规模 A 轮融资之前，Etched 正处于生死攸关的战略转折点： - 架构设计（Architecture）与前端逻辑验证已基本收敛，芯片核心架构的合理性已在理论与仿真中得到充分证实； - 摆在面前的硬性关卡是进入极其昂贵的**物理设计阶段（Physical Design）**； - 这一步需要签订大额商业合同与晶圆代工预订协议，没有数以千万计的真金白银作为后盾，整个造芯进程将瞬间停摆。在此阶段，两位创始人必须在没有实物硅片可供演示、外界质疑如潮的绝境下，说服顶级投资人押注这场世纪豪赌。 ### `01:00:00` 从 1500 万美元到全机架集群：陷入绝境的资本悬崖 在推进芯片设计的过程中，Etched 遭遇了极其残酷的现实冲击。单是聘请一家物理后端设计（Physical Design）供应商，开销就高达 4,000 万至 5,000 万美元。更致命的认知来自于大模型架构的演进——随着 MoE（混合专家模型）等庞大架构的爆发，团队意识到单纯造出一颗芯片根本无法解决问题，他们必须交付一整套机架级集群。这意味着除了芯片本身，Etched 还必须亲自研发电路板（Boards）、互连架构（Interconnects）、液冷冷板（Cold Plates）以及全套网络基础设施。而当时公司银行账户里仅仅躺着 1,500 万美元。面对这道巨大的资金鸿沟，恐惧感瞬间笼罩了团队，加文·尤伯蒂（Gavin Uberti）甚至半开玩笑地自嘲道，那一刻他忍不住开始琢磨：“重新回哈佛复学到底有多难？” 为了证明自身的可行性，团队在 2023 年底耗费了上百个小时撰写了一份长达 30 页的技术商业备忘录。这份文件细致入微地拆解了技术路线图、全栈工程里程碑、算力市场演化路径、端到端每 Token 成本模型以及未来海量的新应用场景。然而，带着这份备忘录走进硅谷时，几乎所有一线头部风投机构都在第一时间给出了无情的拒绝。在投资者眼中，这是两个刚从哈佛辍学的年轻人，既没有流片记录，甚至拒绝做测试芯片（Test Chip）；而且当时市场普遍迷信于模型训练，对推理市场是否足够庞大充满怀疑，伴随大模型幻觉与行业泡沫论的甚嚣尘上，没人愿意下注。彼时半导体领域史上最大的 A 轮融资记录也不过 5,000 万美元左右，而 Etched 盘算着未来 12 个月他们至少需要烧掉整整 1 亿美元。 ```mermaid flowchart TD accTitle: Etched 资本与工程扩张挑战 accDescr: 展示 Etched 从芯片设计到全机架研发所面临的资金鸿沟与全栈工程需求 subgraph Challenge ["硬件工程与资金瓶颈"] A["物理后端设计 (Physical Design)<br/>开销：$40M - $50M"] B["MoE 巨型模型爆发<br/>单芯片无法满足吞吐需求"] C["全栈集群需求<br/>电路板 + 互连 + 冷板 + 网络"] end subgraph Reality ["严峻现实与资本缺口"] D["账面可用现金：$15M"] E["未来12个月资金需求：$100M"] F["硅谷一线 VC 全员拒投<br/>被质疑无流片记录/推理市场不成立"] end A --> C B --> C C --> E D -. 产生巨大资金缺口.-> E E --> F ``` - **全机架开发范畴**：涵盖芯片、定制板卡、定制互连结构、液冷冷板及数据中心网络方案。 - **早期资本差距**：账面仅有 1,500 万美元，而满足流片及全栈工程运转至少需要 1 亿美元。 - **主流风投核心顾虑**：创始人年轻且无流片背书、缺乏测试芯片验证、质疑推理市场规模、担忧 AI 幻觉与行业泡沫。 --- ### `01:02:44` 筹资 1.03 亿美元：生死时速与供应商的豪赌 面对绝境，罗布·瓦琛（Rob Wachen）与加文深夜坐在库比蒂诺的办公室里面面相觑，逐笔计算如何压缩成本：“我们能在这里抠出 50 万美元吗？能在那边砍掉 10 万美元吗？我们到底能说服员工多久不领薪水？”但账目无论如何都算不过来。做半颗芯片或耗费数年先做一个小规模测试芯片，对于瞬息万变的 AI 爆发期而言等同于慢性自杀；他们清醒地意识到，要赢就必须孤注一掷直接做全功能全尺寸芯片。Etched 随即进入了长达数周的极端“生存模式”，两位创始人疯狂动用所有人脉，不放过任何一个可能认识投资人的电话号码，逢人便推销自己的远景。 转机如同滚雪球一般逐渐发生。从最初依靠一家债权融资机构提供资金跨过掩膜版（Mask）底线，到天使与激进投资人开出的 100 万、200 万美元支票，再到 500 万、1,000 万美元的大额进账，局面终于打开。最终在一次董事会上，当表格上的意向认购额累积跳动到 1.03 亿美元时，全员一致决定全额接纳，缔造了半导体行业有史以来规模最大的 A 轮融资之一。更为关键的是核心供应链伙伴的鼎力相助：台积电（TSMC）在 Etched 尚未融到这 1 亿美元、前景极度不明朗的至暗时刻，就毅然同意为其提供产能合作；EDA 巨头新思科技（Synopsys）则以极为优厚的多年延期付款协议，向他们提供了价值不菲的硬件仿真加速器（Emulator），这在实质上相当于为这家初创公司提供了一笔关键的巨额过桥贷款。 - **极限制造成本测算**：若仅满足 11 毫米极限掩膜版流片并全员依靠极低成本度日，最低启动资金底线为 3,000 万美元。 - **A 轮融资最终体量**：意向认购锁定在 1.03 亿美元，成为 AI 硬件初创企业有史以来规模最为庞大的 A 轮融资之一。 - **供应链核心支持**：台积电提前敲定先进制程制造支持；新思科技允许其分期数年支付仿真器费用。 --- ### `01:04:47` 晚宴上的数学草稿纸：台积电为何押注 22 岁年轻人 在被问及全球代工巨头台积电为何愿意信任这样一家雏鹰般的初创公司时，加文分享了一段戏剧性的往事。在罗布加入公司之前，加文受邀参加了一场高规格半导体行业峰会。作为极少数年轻的芯片公司 CEO，他是现场唯一一位 40 岁以下、乃至 30 岁以下的演讲嘉宾，年仅 22 岁。演讲结束后的晚宴极其奢华正式，来宾非富即贵，甚至包括 Arm 的前任 CEO，全场所有人皆身着笔挺西装。加文机缘巧合被安排坐在一位台积电核心资深副总裁身旁。交谈中两人意外发现彼此大学时都专攻纯数学，共同的学科背景瞬间拉近了距离。 加文随即拿出一张草稿纸，开始与这位副总裁在微观的“逐 Tensor 层级”（per tensor by tensor level）深度推演现代前沿 AI 模型的底层运行机制。加文详细剖析了当前通用计算芯片运行这些模型的巨大浪费，以及为何专用硬件架构结合先进制程是释放算力的终极解法。对方完全听懂了这套逻辑背后的颠覆性威力。晚宴结束后的第二天清晨，加文的邮箱里就静静躺着一封来自台积电的信函，上面直截了当地写道：“加文，我想与 Etched 携手，共同探索将这一切变为现实的路径。”这封邮件不仅打破了半导体产业对年轻创业者的刻板偏见，更为 Etched 锁定了全球最顶尖的代工生命线。 - **关键契机**：加文在半导体峰会晚宴上与台积电高管就数学背景产生共鸣，凭借扎实的模型底层算子推演赢得认可。 - **核心论据**：从每 Tensor 运算的微观视角切入，透彻阐述专用推理芯片架构相比通用硬件的能耗与吞吐优势。 --- ### `01:06:29` 逆向投资与健康无知：跨越认知偏差的战壕文化 播客主持人打破常规“第四面墙”，公开确认自己不仅是 Etched 的极早期投资人，而且在其后续融资中连续跟投了五次之多。主持人坦言，在投资领域人人都爱引用“逆向且正确”（Contrarian and Right）能带来巨额回报的黄金定律，但人们往往忽略了逆向的真正滋味——那就是在成为共识之前，全世界都觉得你是个彻头彻尾的傻瓜。当他最初为投资 Etched 进行背调时，所有传统的半导体专家都给出了极其理智且逻辑闭环的否定理由，断言其成功的概率近乎为零。主持人总结道，优秀的早期投资往往需要一种“对基础概率的漠视”（Damn the base rate）以及适度的“健康无知”（Healthy naivete）；因为如果只看行业以往的失败率，世人只需去买指数基金即可。 加文指出，这种认知偏差在硬科技投资领域屡见不鲜：绝大多数传统芯片基金错过了整整一代 AI 芯片浪潮，正如多数资深编程老手看空 AI 编程工具一样。老牌半导体从业者被过去二十年的流片经验所禁锢，见惯了流片三次都无法正常运行负载的惨痛历史，却未能察觉到底层约束条件已经彻底颠覆——如今的 EDA 工具和超大规模 FPGA 仿真集群已经能完成以往不可想象的验证深度，首次流片成功（First-pass success）不再是神话。真正敢于在早期重注 Etched 的只有两类人：一类是像主持人这样对巨大赛道与团队直觉抱有坚定信念的先锋，另一类则是像华尔街顶级高频交易机构（HFT）这样亲自下场做芯片的技术狂热者，他们派出 10 人专家团直接下场逐行审查 Etched 的微架构、RTL 代码、板卡设计与软件栈。这种在生死边缘跨越“认知偏差”（Variant Perception）的历程，最终沉淀成了 Etched 独有的工程文化——那些入职两年的老员工就像是在战壕里叼着雪茄面对炮火的百战老兵，带领着初来乍到、战战兢兢的新人们在绝境中杀出一条血路。 - **投资哲理**：逆向认知的本质是承受全世界的质疑；当外部世界的声音与前沿实践出现巨大认知鸿沟时，这种“认知差”正是超额机遇的源泉。 - **产业环境迭代**：先进的现代 EDA 工具及大规模 FPGA 前端原型验证系统，大幅打破了传统芯片流片“必失败多次”的历史约束。 - **团队战壕文化**：面对看似不可逾越的技术与资金鸿沟，老员工形成了“永远能找到破局办法”的坚定信念，以此感染新团队成员。 ### `01:10:00` 晶圆测试惊魂夜与“解谜心态” 在芯片初创公司，团队的底层信念必须是“一切皆可解决”——既然坐在这里，前提就是这件事并非不可逾越，遇到任何困难都只能咬紧牙关攻关，绝不能说“不可能”。Gavin 和 Rob 分享了一段难忘的“开荒”故事：团队引入了一位芯片验证（Silicon Validation）领域的传奇老将。芯片从晶圆厂出厂时，需要进行最初步的“晶圆测试”（Wafer Sort）。当时正是凌晨两三点，团队正隔着越洋电话与台湾台积电（TSMC）连线。测试时需要使用探针卡（Probe Card）接触尚未切割的整块晶圆，通过微小探针发送电信号测试每个裸片的良率，筛选出好芯片进行后续封装，剔除坏芯片。 屏幕上最开始显示着一片由灰色小方块组成的晶圆图案。随着测试程序开始加载，原本应该逐个转为绿色（代表通过）或红色（代表损坏）的方块，竟然在一瞬间全变成了红色。监控室里的年轻工程师们当场被吓得魂飞魄散，直呼“这下完了”。而那位经历过无数风浪的老法师却只是把身子往后一靠，云淡风轻地笑笑说：“好戏开场了，解谜正式开始（The puzzle begins）。”真正优秀的硬件工程师不仅敢于直面这种“凝视深渊”的恐惧，甚至痴迷于这种在绝境中解开谜题的快感。果然，不到一天之内，屏幕上就出现了第一个绿色方块。 ```mermaid flowchart LR accTitle: 晶圆测试排错流程 accDescr: 晶圆从出厂到探针卡测试出现全红故障，经过老将介入排查后在一天内转为良品绿色的全流程。 subgraph Fab["晶圆出厂与连线"] A["台积电代工完成<br/>整片晶圆（Wafer）"] --> B["凌晨连线台积电<br/>探针卡（Probe Card）测试"] end subgraph Test["首批测试遇挫"] B --> C["探针接触裸片<br/>发送电信号测试"] C --> D["测试图全红报错<br/>团队陷入极度恐慌"] end subgraph Resolve["解谜与突破"] D --> E["传奇验证老将介入<br/>'解谜正式开始'"] E --> F["一天内攻克故障<br/>首现合格绿色方块"] end ``` ### `01:12:15` 推理规模化的三大基石与第二代芯片的极简哲学 回顾前几年的摸爬滚打，Etched 走过了大量试错周期：从尝试把模型编译进 FPGA、尝试把权重硬烧进硅片，到探索拆分高带宽内存（HBM）分别存储 KV 缓存和模型权重等路径。经历了漫长的摸索，他们最终提炼出主导全球 AI 推理计算的“三大基本法则”： 1. **同等功率下的极致浮点算力（Flops per Watt）**：在受限的功耗预算内榨干所有能算出的吞吐量； 2. **芯片间的超低延迟互联（Lowest Latency Interconnect）**：构建全球最大的“扩大域（Scale-up Domain）”； 3. **极度可量产性（Maximum Producibility）**：制造出尽可能多的可用硬件。 在公司成立前半段，他们主要搞清楚了前两点，并据此押注于低电压推理架构与集群级共享内存（Cluster-scale Memory）。但在最近一年，第三点——**可交付能力（Availability）**——的迫切性彻底爆发。在算力极度渴求的当下，“最好的能力就是有现货（The best ability is availability）”。如果当下能交出 1000 张芯片，立刻就会被全部吃下。因此，Etched 在规划面向第二代（Gen 2）及后续芯片时，核心目标是打造在极限情况下能够达到“每月数吉瓦（Gigawatts per month）”产能的系统架构。第二代设计的核心指导思想全面转向“极简主义”：大刀阔斧地剔除冗余零部件，反复验证组装和拆卸流程以压缩生产节拍时间（Cycle Time），从底层确保系统具备极高的可靠性、可维护性以及吉瓦级别的量产能力。 ### `01:13:58` 避开零和博弈：差异化供应链与真正的垂直整合 面对先进制程产能被各大巨头瓜分的现状，Etched 从产品定义阶段就避免与英伟达（NVIDIA）陷入零和博弈（Zero-sum Game）。全球部署算力的大客户往往受限于特定制程节点或 HBM 产能的天花板，但 Etched 第一代芯片在供应链选择上精准进行了错位竞争： - **晶圆代工节点**：英伟达 Rubin 等竞品选用 3nm 制程，Etched 第一代则选用台积电成熟且产能充足的 4nm 制程； - **内存供应**：选用了与 Rubin 不同的 HBM 供应与规格； - **供给效应**：对头部云厂商而言，采购决策不再是“要 1 吉瓦 GPU 还是要 1 吉瓦 Etched”，而是“两者叠加拿到 2 吉瓦总算力”。 Gavin 强调：“如果你设计出了全球性能最强的产品，但根本造不出来，那你的产品就只是一期播客（Then you're just a podcast）。”除了芯片制造与 HBM 这类受物理限制必须深度合作的瓶颈之外，Etched 对板卡、互联器件等紧缺组件进行自主垂直整合。自主设计制造的比例越高，就越能在全球产能瓶颈之上开辟出全新增量。 ### `01:15:00` 机器思维不同于人类思维：算力极度廉价引发的模型演进 从硬件底层推演模型演进逻辑，Gavin 提出了一条核心洞察：**机器的思考方式从来都不同于人类，就如同“飞机飞行的方式与鸟类截然不同”**。 - **生物大脑与硅基芯片的天然对立**：对人类神经元而言，存储记忆和调用数据极其廉价，而做高精度的数学计算代价高昂；但在芯片架构中，这一特征恰恰反转——存储与搬运数据极其昂贵，而执行浮点数学计算却无比廉价。 - **摩尔定律与内存墙的差距演进**：由于 DRAM 物理机制的根本限制，数学计算变便宜的速率远远快于内存访问变便宜的速率。这意味着未来的模型架构必须走向“狂暴计算”——例如同时并行运行大量副本、激活极大规模的混合专家（MoE），甚至跨越多个机架同时跑超巨型专家网络。 - **超长上下文的未来**：目前学界为追求推理效率，往往回避全量加载上下文。但在未来，要想实现真正的超级智能，模型应该能够调动海量算力瞬间扫过 10 亿 token 的上下文，乃至直接在工作记忆中实时检索人类有史以来写过的所有图书。 ### `01:16:42` 动态计算（Dynamism）对老架构的降维打击 当前 AI 模型架构的最重要演进趋势是**动态性（Dynamism）**。这涵盖了根据输入的重要性在 Token 级别动态控制算力与内存分配，以及芯片根据 MoE 专家的调用实时将数据调度至其他芯片的能力。 - **非均匀分配**：并非所有参数都需要在每个 Token 上激活，也非每个 Token 都需要等长上下文。重要的 Token 应当被倾斜更多算力和更长上下文，不重要的 Token 则可以通过共享内存来减小开销。 - **旧硬件的劣势**：在这些新型动态架构出现之前设计的通用 GPU 存在极大的执行冗余与通信瓶颈，要么在应对此类动态路由时效率低下，要么只能依赖僵化笨重的块状（Blocky）架构“大力出奇迹”，强行给各异的 Token 灌注同等算力，造成巨大的能耗浪费。Etched 从底层针对这类高度动态的运算进行硬件级加速。 ### `01:18:22` 未来的计算形态：集群级协作与“每兆瓦智能体”时代的到来 关于推理成本断崖式下跌后的未来应用形态，Gavin 提到了 Noam Brown（卡内基梅隆大学知名学者、OpenAI 核心研究员，现也成为 Etched 的天使投资人）曾引发热议的一个观点：随着前沿模型推理任务的时间跨度从几分钟延伸至几个月，现实中很难真正对其进行有效评测——因为等半年测完，底座模型早就迭代换代了。 要解开这个死结，核心有两个维度： - **加速单任务周期**：利用 Etched 开发的“集群级内存（Cluster-scale Memory）”，将长达 6 个月的单体运算任务压缩到极短时间内完成； - **大规模智能体协同**：正如单个人类造不出火箭、必须依靠数千人的组织协同一样，如果让智能体去开发极度复杂的未来级软件，也必然需要 10 个乃至 100 万个 Agent 协同作战。这就要求计算架构必须同时具备海量浮点运算吞吐与超低延迟的机架集群级互联。 展望未来 10 年，AI 推理计算将稳步接管全球大部分 GDP。衡量人类社会生产力的核心指标，也将从传统的“人均 GDP（GDP per capita）”转变为“每兆瓦智能体数量（Agents per megawatt）”乃至“每吉瓦智能体数量（Agents per gigawatt）”。 ### `01:20:00` 十亿并发 Agent 与“分布式大脑”架构 人类劳动力的结构正站在剧烈重塑的悬崖边缘。虽然目前绝大多数劳动力仍是人类，但到 2027 年，从事知识工作的 AI Agent 数量就将彻底超越人类。未来的竞争甚至将演变为国家层面的算力与能源竞争：整个国家的多数能源可能都会源源不断地输送进专门处理推理任务的数据中心，而这些数据中心的能效水平，将直接决定该国能够部署多少 Agent，进而决定其“虚拟劳动力”的真实规模。如果仅仅是五到十个 Agent 组成的小组协作完成两天的项目，虽然成果惊人，但这依然停留在工具层面，未达“文明尺度”；真正不可思议的前景在于，当一个国家拥有十亿个并发运行的 Agent——相当于十亿名无需睡眠、7×24 小时全天候并行协同的超级劳动力时，人类将迎来历史上规模最庞大的一次技术大爆发。 面对如此庞大的推理需求，算力系统的构建必须依托全新的规模经济效应。人脑在运转时并非所有神经元同时激活，健康的脑组织在任一瞬间仅有一小部分处于活跃状态；混合专家模型（MoE）的运作逻辑亦是如此，在生成任意 Token 的瞬间，模型仅需调用极少一部分参数。基于硬件系统的创新，工程师可以将这种庞大的“数字大脑”切分成无数个专家模块，分别部署在不同的服务器节点上，并灌入海量的并发请求。每一股数据流都能在特定时刻精准命中对应的大脑区域，从而使单次“思考”或生成单个 Token 的单位成本大幅暴跌。这种形态最终将演化为超大规模的分布式大脑，其物理实体就是由高密芯片阵列、惊人算力峰值（FLOPS）以及超高速 Scale-Up 互连网络所构成的数据中心巨兽。 ```mermaid flowchart TD accTitle: MoE 模型与分布式硬件架构映射 accDescr: 展示海量 Token 请求如何分发到由高速互连串联的分布式专家节点以摊薄思考成本 subgraph UserTraffic["并发请求层"] T1["高并发 Token 任务流 1"] T2["高并发 Token 任务流 2"] T3["高并发 Token 任务流 N"] end subgraph DistributedBrain["分布式大脑 (数据中心规模)"] Router["动态路由与负载均衡调度"] E1["专家节点 1 (服务器阵列 A)"] E2["专家节点 2 (服务器阵列 B)"] E3["专家节点 3 (服务器阵列 C)"] E4["专家节点 N (服务器阵列 D)"] Interconnect["Scale-Up 超高速互连网络 (光互连 / 内存池)"] end T1 --> Router T2 --> Router T3 --> Router Router --> E1 Router --> E2 Router --> E3 Router --> E4 E1 <--> Interconnect E2 <--> Interconnect E3 <--> Interconnect E4 <--> Interconnect ``` - **劳动力拐点预测**：预计到 2027 年，从事知识工作的 AI Agent 规模将正式超过人类劳动力总量。 - **MoE 的稀疏激活机制**：正如健康人类大脑在任何时刻都只激活部分区域，MoE 模型在处理单个 Token 时也仅动用一小部分参数，通过将专家拆分至集群并进行高并发批处理，可将单位 Token 的计算成本降至极限。 --- ### `01:22:35` 万亿美元数据中心与“思考”供应链的终局 建造单一成本高达一万亿美元的数据中心绝非天方夜谭，这只是一个时间问题。回顾半导体制造史，行业曾经历过从 10 亿美元、100 亿美元到如今数百亿美元晶圆厂的演进，晶圆厂的经济规模绝不会在 400 亿美元的关卡戛然而止；只要持续投入资本，单片晶圆的均摊制造成本就会不断下探。这一规律在钢铁厂成立，在“制造 Token 的超级工厂”中同样适用。如果此时有一位极高智慧的外星人降临地球，询问他们所致力的事业本质是什么，答案异常纯粹：思考是这个星球上最珍贵的资产，世间所有商业帝国的运转都建立在“思考”之上。 当前人类正步入一个千载难逢的奇点：机器的思考水平已经逼近顶尖人类，并将在短时间内追平甚至彻底超越人类。这不仅意味着制造思考机器蕴藏着无可比拟的商机，更意味着当社会对智能的需求无限膨胀时，支撑这种思考运转的底层范式必须发生根本变革。面对未来承载十亿用户并发、参数量达千万亿级（Quadrillion）的模型，行业亟需一套全新的架构与路线图。由于生产智能的边际成本远远低于智能所创造的巨大经济价值，全人类将在未来数年乃至数十年内陷入长期的“Token 产能严重短缺”。在这一历史周期中，能够高效生产 Token 的芯片与系统具有近乎垄断的极高价值，而未来全球市值最大的科技巨头，必将是那些主导全球 Token 供给并掌控其供应链核心命脉的企业。更为关键的是，系统的扩展必须摆脱线性成本的陷阱：算力架构绝不能停留在“为了输出 10 倍 Token 就傻傻买入 10 倍服务器”的旧逻辑，而必须依托集群级内存等创新架构，在规模成倍扩张的同时实现显著的边际成本递减。 - **思考即资产**：世界上每一家企业的运转核心都是“思考”，机器智能的制造正在将这一抽象能力工业化。 - **Token 的长期供不应求**：智能带来的经济回报远高于 Token 的生成成本，这决定了全球将陷入持续数年甚至数十年的 Token 供给侧荒潮。 - **非线性扩展法则**：新一代算力系统必须具备规模经济效应，依靠集群级共享内存与专用芯片，使产出 10 倍 Token 时的增量成本显著低于 10 倍。 --- ### `01:25:20` 最善良的举动：十六岁那年父母的无私守护 在播客的传统收尾环节中，面对“他人为你做过的最善良的事情是什么”这一提问，Rob 动情地分享了自己少年时期与癌症抗争的生死往事。十六岁那年，他在癌症治疗过程中遭遇了人生最艰难的抉择：医生走到他面前，要求他自主决定选择手术还是放疗。医生给出的选项残酷而现实——如果选择手术，生存的概率更高，但代价是必须接受可能永远无法再次行走的现实；如果选择放疗，他保留了重新行走的希望，但生还几率大打折扣，随时可能面临死亡。面对这个沉重得令人窒息的生死抉择，父母对他说，这关乎他自己的人生，必须由他独立做出决断。在漫长而痛苦的深思熟虑后，Rob 毅然选择了接受手术。 然而手术切除肿瘤之后，医疗团队进行的坏死分析（Necrosis Analysis）却带来了坏消息。通常情况下，只有当肿瘤细胞坏死率达到 98% 至 99% 时，医学上才能判定彻底脱离危险，而 Rob 组织样本中的活细胞比例偏高，未能达到这一安全阈值，医生告知他必须立即追加放疗。雪上加霜的是，当时全球仅有极少数设备能够执行他所需的高难度特殊放疗，其中一台远在波士顿。彼时的 Rob 只能依靠轮椅代步，需要前往波士顿连续生活数月接受漫长治疗。就在那个全家最艰难的至暗时刻，父母毫不犹豫地放下了手头所有工作与个人生活，双双收拾行装搬到波士顿全程陪伴、悉心照料。这份毫无保留的爱与牺牲，成为他一生中最为感恩、永生难忘的善良之举。 - **十六岁的生死抉择**：Rob 在 16 岁患癌时面临两难：高生还率但极大概率终身瘫痪的手术，或是保留行走能力但死亡风险更高的放疗；他最终选择手术以争取活下去。 - **坏死分析测试（Necrosis Analysis）**：术后用于检测切除组织中癌细胞死亡比例的病理分析，通常需要 98%～99% 的坏死率才算安全，Rob 因未达标而被迫追加特殊放疗。 - **父母的无私付出**：面对全球仅有少数机器可用的局面，父母毅然放下一切工作，陪伴坐在轮椅上的 Rob 迁居波士顿数月接受治疗。 ## Notable quotes > `02:07` 为什么这件事不可能做到？ — Gavin Uberti > `06:50` 天下没有免费的午餐 — Rob Wachen > `25:00` 我们要赢，除了造出一台得分最高的机器人之外，我们什么都不做。 — Gavin Uberti > `26:35` 最好的供应商就是没有供应商 — Rob Wachen > `28:38` 第一次聊就答应加入的人少之又少，但聊到第二十次才点头的人却多得惊人。 — Rob Wachen > `30:00` 心怀不忿，方能让芯片落地机房 — Gavin Uberti > `52:44` 生产即产品 — Rob Wachen > `52:44` 对我所在的大厂而言，这款自研芯片成不成功，根本关乎不到公司的存亡。 > `55:40` 哪怕绝大多数实验都以失败告终，我只要赌对一次就够了。 — Gavin Uberti > `01:04:47` 加文，我想与 Etched 携手，共同探索将这一切变为现实的路径。 > `01:10:00` 好戏开场了，解谜正式开始 > `01:12:15` 最好的能力就是有现货 — Gavin Uberti > `01:13:58` 如果你设计出了全球性能最强的产品，但根本造不出来，那你的产品就只是一期播客 — Gavin Uberti ## Mentioned Etched, Gavin Uberti, Rob Wachen, Patrick O'Shaughnessy, Mark Ross, Cypress Semiconductor, Bryan Lawler, Sanford, Noam Brown, NVIDIA, TSMC, Synopsys, Apple, Xnor.ai, OctoML, Google, Meta, Microsoft, OpenAI, Cursor, Anysphere, Mercor, xAI, Colossus 集群, Sohu, Blackwell, Blackwell B300, Rubin, HGX, DGX, NVL72, TPU, MTIA, Maia, Jalapeno, ChatGPT, GPT-4, Codex, GPT-OSS, FIRST Tech Challenge, Arm, Prod ## Video description My guests today are Gavin Uberti and Rob Wachen, the founders of Etched.  A few years ago, when they set out to build a better AI chip than the largest companies in the world, almost everyone I called told me it could not be done. They have since done it, taping out a working chip on their first attempt and becoming the first hardware company founded after ChatGPT to do so. They already have more than a billion dollars of customer demand for their first product, and have raised eight hundred million dollars to build it.  Etched builds chips and systems designed to run AI models faster and at lower cost. They started the company in 2023, and that product is a complete rack for inference, the chip along with the boards, the power delivery, the interconnects, and the manufacturing to produce it all. We talk about the technical bets behind their architecture, how they hired industry legends and paired them with elite 22 year-olds, and why they believe inference will become one of the largest markets in the world. I think you will find the story of what they have built hard to forget. Please enjoy my conversation with Gavin and Rob. For the full show notes, transcript, and links to mentioned content, check out the episode page ⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠here⁠⁠⁠⁠⁠.  ----- Become a Colossus member to get our quarterly print magazine and private audio experience, including exclusive profiles and early access to select episodes. Subscribe at ⁠colossus.com/subscribe⁠. ----- ⁠Ramp’s⁠ mission is to help companies manage their spend in a way that reduces expenses and frees up time for teams to work on more valuable projects. Go to⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠ ⁠ramp.com/invest⁠⁠ to sign up for free and get a $250 welcome bonus. ----- Trusted by thousands of businesses, ⁠Vanta⁠ continuously monitors your security posture and streamlines audits so you can win enterprise deals and build customer trust without the traditional overhead. Invest Like the Best listeners get a special offer of $1,000 off Vanta when you go to ⁠vanta.com/invest⁠.  ----- WorkOS⁠ is the infrastructure B2B and AI-native companies use to sell to enterprise. It covers everything enterprise security requires: SSO, SCIM, RBAC, Audit Logs, AI governance, and more. Trusted by 2,000+ fast-growing companies, including OpenAI, Anthropic, Cursor, and Vercel. ----- Rogo is the AI platform for finance. They're building agents for Wall Street that are trained to understand how bankers and investors actually do work: from diligence and modeling, to turning analysis into deliverables. To learn more, visit rogo.ai/invest. ----- ⁠Ridgeline⁠ has built a complete, real-time, modern operating system for investment managers. It handles trading, portfolio management, compliance, customer reporting, and much more through an all-in-one real-time cloud platform. Visit⁠ ⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠ridgelineapps.com⁠. ----- Editing and post-production work for this episode was provided by The Podcast Consultant. Timestamps: (00:00:00) Welcome to Invest Like The Best (00:02:07) Gavin Uberti and Rob Wachen (00:03:54) Two 21-Year-Olds Taking on NVIDIA (00:07:52) The Two Technical Bets Behind Their Architecture (00:14:15) Why Inference Becomes the Biggest Market (00:20:23) Rob and Gavin's Origins Stories (00:28:38) How They Recruit Industry Legends (00:36:30) Moving a Dozen Engineers to Bangalore for Six Months (00:38:01) Speed Wins (00:43:58) Getting More Concurrency Out of Every Megawatt (00:52:44) Vertical Integration (00:57:43) Hardest Obstacles to Overcome (01:01:09) Raising The Largest AI Chip Series A Ever (01:06:29) TSMC (01:13:20) Designing Gen 2 for Gigawatt-Scale Production (01:16:42) Why Machines Don't Think Like People (01:20:03) A Year of Compute Compressed Into a Month (01:23:44) The Trillion-Dollar Data Center (01:26:19) The Kindest Thing [Source](https://traffic.megaphone.fm/CLS8561569864.mp3)

[2:07] Gavin Uberti: 为什么这件事不可能做到？ [6:50] Rob Wachen: 天下没有免费的午餐 [25:00] Gavin Uberti: 我们要赢，除了造出一台得分最高的机器人之外，我们什么都不做。 [26:35] Rob Wachen: 最好的供应商就是没有供应商 [28:38] Rob Wachen: 第一次聊就答应加入的人少之又少，但聊到第二十次才点头的人却多得惊人。 [30:00] Gavin Uberti: 心怀不忿，方能让芯片落地机房 [52:44] Rob Wachen: 生产即产品 [52:44] 对我所在的大厂而言，这款自研芯片成不成功，根本关乎不到公司的存亡。 [55:40] Gavin Uberti: 哪怕绝大多数实验都以失败告终，我只要赌对一次就够了。 [64:47] 加文，我想与 Etched 携手，共同探索将这一切变为现实的路径。 [70:00] 好戏开场了，解谜正式开始 [72:15] Gavin Uberti: 最好的能力就是有现货 [73:58] Gavin Uberti: 如果你设计出了全球性能最强的产品，但根本造不出来，那你的产品就只是一期播客

[2:07] Patrick O'Shaughnessy: A few years ago, when they set out to build a better AI chip than the largest companies in the world, almost everyone I called told me it could not be done. [7:52] Gavin Uberti: If I 2x my voltage, my power goes up by 4x. If I cut my voltage in half, my power goes down by a quarter... we were able to create a new mechanism of running at much lower voltages, a new type of power delivery that we call low voltage inference. [48:38] Gavin Uberti: We are kernels-first programming, and that means that, for a long time, it did not work out of the box. But if you were a kernels expert, you could get incredibly, incredibly high performance. [105:00] Patrick O'Shaughnessy: When you read about investing, everyone cites this idea of like 'contrarian and right' as the quadrant that makes all the money... But contrarian means like everyone else thinks you're stupid. And so when you go and you get an immediate no from literally everybody, it is a fascinating quadrant to exist in before you become consensus.

[0:00] Ramp is the only platform built to make your finance team leaner, faster, and better, saving businesses 5% annually on average so you can stay focused on growth. Ramp customers grew revenue 3.2 times faster than the average American business. Visa, Vercel, Cursor, Stripe, Notion, 11Lab, Shopify, and 70,000 other businesses all run on Ramp. Mine does too, and so should yours. Learn more at ramp.com slash invest. Felix by Rogo is a personal finance agent that turns a single prompt into finished client-ready work using your firm's own templates, context, and standards. Send Felix an email like, take these comments and turn them for me, or update my tracker with the context of these emails, or run the ability to pay math on this buyer, and Felix sends back finished PowerPoint decks, Excel models, and sourced research. Felix works the way your team already does, delivering work quickly and accurately around the clock. Learn more at rogo.ai slash Felix. The best AI and software companies from OpenAI to Cursor to Perplexity use WorkOS to become enterprise ready overnight, not in months. Visit WorkOS.com to skip the unglamorous infrastructure work and focus on your product. Hello and welcome everyone I Patrick O and this is Invest Like the Best This show is an open exploration of markets ideas stories and strategies that will help you better invest both your time and your money If you enjoy these conversations and want to go deeper, check out Colossus, our quarterly publication with in-depth profiles of the people shaping business and investing. You can find Colossus along with all of our podcasts at colossus.com. Patrick O'Shaughnessy is the CEO of Positive Sum, All opinions expressed by Patrick and podcast guests are solely their own opinions and do not reflect the opinion of Positive Sump. This podcast is for informational purposes only and should not be relied upon as a basis for investment decisions. Clients of Positive Sump may maintain positions in the securities discussed in this podcast. To learn more, visit PSUM.VC. My guests today are Gavin Uberti and Rob Lockett, the founders of Edged. A few years ago, when they set out to build a better AI chip than the largest companies in the world, almost everyone I called told me it could not be done. They've since done it, taping out a working chip on their first attempt and becoming the first hardware company founded after ChatGPT to do so. They already have more than a billion dollars of customer demand for their first product and have raised million to build it Etch to build chips and systems designed to run AI models faster and at lower cost They started the company in 2023 and the product is a complete rack for inference the chip, along with the boards, the power delivery, the interconnects, and all the manufacturing to produce it all. We talk about the technical bets behind their architecture, how they hired industry legends and paired them with elite 22-year-olds, and why they believe inference will become one of the largest markets in the world. I think you will find the story of what they have built hard to forget. Please enjoy my conversation with Gavin and Rob. All right, gentlemen, it's been three years or so, Gavin, since you and I last did this, which is nuts. And at the time, I was just wildly intrigued by your story and what you were going to build. I didn't know a lot about chips at the time. I was considering investing in the company, and so I was calling everyone I could conceive of that could give me an opinion or something. And at the time, basically, the consensus was these kinds of companies are not built by young people, that the semis world, the best companies are founded by 40, 50-year-old people that have had a whole career's worth of experience, have learned all the problems, have shipped multiple chips. Two 21-year-olds are like, not going to do this. It's just not going to work. It was indicative of a theme which was nobody believes in us That obviously changed a lot now Just walk the halls and talk to the people that have chosen to come work here But in the early days it felt like this was something that you had to face down What was that like facing that down where a set of incumbents and an industry worth of people and investors and everyone else sort of didn't believe in you? And what did that anneal in you to build the company the way that you are? Like, what was the impact of that? I think there's a certain level of naivety required to think that you could build a chip better than every other AI chip ever built and build a company to do it way faster than ever has been done. And we have the naivety. But there's many times where we would say, like, why isn't this possible? And, you know, really push on it. And it turns out that, like, everybody's answers are extremely siloed to a set of constraints that aren't true anymore. And the reality is the entire semiconductors and data center industry is built on buffer. And what I mean by that is every part of the stack from the EDA tools to the power modules, to the circuit boards, to the chip design and standard cells. Everything is built to be general purpose for everything, not just in the data center, but IOT on the edge and so forth. And when you have a specific use case, you're really trying to design for, you can change the constraints a lot. And I'll give you just a very simple example that we're not the only one who does, which is one of the things you care a lot about is the clock speed of your chip. It's proportionally [5:00] to the throughput of your system. When you are doing sign-off for different timing, what clock speed you're actually going to be able to run on when you tape out your chip, there's this concept called corners, which is, you know, what temperatures are you going to be able to run at this clock speed? The default configurations for a lot of these EDA tools assume that you're going to be running your chips in freezing temperatures. Now, I don't know about you, but I've never seen an AI data center with ice in it. So, you know, we can feel pretty confident that our chips don't need to run at full speed at zero degrees Celsius. In fact, like, they're never really going to be running below 80 degrees Celsius anyway. And just by knowing that that's a constraint that doesn't matter, we can make a ton of changes throughout the entire system. That's like a very simple one, but there's many more that you get 20% here, 50% there, 2x here, and these compound to a system that can be radically better for inference. I think you found two kinds of people. There are some folks who went purely on heuristics, of hey, young founders, they claim they can go beat the biggest company in the world of performance. it cannot happen. And there is no thing you could go say to me that would make me change my mind. But there's also people out there who are of course skeptical, but are willing to go ahead and say, I'll spend the time, I'll do the work, and is it actually possible? Like for example, one of our earliest earliest supporters was Mark Ross And Mark was a very prestigious semiconductor expert He used to be CTO at Cypress Semi that sold for billion And when we met him we were with a couple of guys in the dorm room, and we came to him and said, hey, do you want to go build hardware for inference? We think we can be much faster than NVIDIA. And Mark's like, no, you can't. It will not work. But he wanted to convince me, you should write a white paper. You should go ahead and build a a functional simulation and show me. And so after a lot of very long nights, went back to Mark and said, hey, here's a simulation. What do you think? And he was like, huh, this works. But to go do a company like this, you'll need a large amount of capital and at least $3 million even to get started. Ended up, went ahead and raised five, raised a lot more after that. And then he was, again, surprised, but got more involved. And then he became an advisor, a halftime advisor, and eventually our full-time CTO, as we saw more and more of the development progress. And I think in general, this skepticism has filtered really heavily. Folks who want to go ahead and be right regardless, or who want to be very truth-seeking, and say, sure, I'm skeptical, but I will go ahead and work through the numbers myself. And if I can go figure out why this is possible, well, let's go build it The specifics that you made bets on the way that you built this system are immensely interesting to me And because so many people are trying to do this now build new chips that will do a better job of serving inference at massive scale, the world is interested in the research approaches, the different architecture approaches that people are taking to building a new AI chip. And I'd love you to just start by describing what this thing is, what it does, but maybe more interestingly and more importantly, the process that you went through to decide what bets to take, what technologies to invent, and compare and contrast those with what you've seen the rest of the marketplace try to do. Yeah, I think that you can start with the product. We're not just building a chip. We're building a full inference solution. And that means a rack. That means the chip. That means the power delivery into the chip. That means the board in which it sits. That means the interconnect. I was to talk to each other. That means the production. for this mass volume of racks. Really, the production is the product. When we think about how we get our advantage, there are two key parts of running inference. There's Prefill and there's Decode. Now, we have two key tag matches in both of these things. Prefill is reading in a huge volume of text, and Decode is then using that data to generate output tokens When you go out and run Prefill your key job is not to go predict tokens You already know the text Your job is to go and get the model memory what we call is KVcache into the right state Then you can go ahead and run decode with that same KVcache So we will often go do is we call PD disaggregation, Prefile Decode Disag. You will have one cluster of servers running these pre-fills. You'll then transfer those model memories, those KVcaches, over to the decode cluster, and then go ahead and use that cluster to go generate the next tokens. So it's sort of like loading the gun and then firing it, like if I think about it in super simple terms. Yeah, you got it. It's getting the model to remember the right things, and then using those things to go do tasks. Candidly, people think about this market a bit lazily. They say, are you a pre-fill chip? Are you a decode chip? If you're a decode chip, are you an HPM chip? Are you an SRAM chip? Are you a 3D DRAM chip? Are you using optics, using copper? When we started this, we just wanted to understand why extremely smart people were working on these different directions. We seriously looked at architectures like having a bunch of DDR memory and like a shared memory pool and looking at advanced packaging to basically break out of the shoreline. We looked at things like, is there ways to put memory dies on top of compute dies? In doing so, we realized that there's no free launch. Everything is a trade-off, right? 3D DRAM, you have a thermal issue, you have a supply chain issue, you have to figure out hybrid bonding, you have to figure out the flops, and now you're a decode chip. So we went through everything, both on the pre-fill and the decode. [10:00] In doing so, we realized there's a few design spaces that nobody had seriously tried to explore because they were never done in AI chips. And we asked ourselves, what are the actual metrics that are going to matter the most? On the pre-fill side, the thing that matters is flops and flops density. And people talk about flops often as a headline number. But in reality, you should care about the flops you're getting when you're running real workloads. There's this concept called MFU, or model flops utilization, which is, you know, for every peak flop advertised, how many cents on the dollar are you actually getting? And on GPUs, you often get somewhere between 20 and 50%, depending on the workload. And actually, you can provably not run at 100% because you have a thermal issue, where as you increase the flop utilization, you have more transistors going on and off, you draw more power, and the chip will self-regulate and actually lower its clock speed to make sure it doesn't overheat. So as we looked at inference, we said, if we want way more flops, because we want to run at way higher throughputs, we fundamentally need to solve the thermal problem before we even think about adding flops to the chip. If I just add more flops to a GPU today or another AI chip, I'm not actually going to get more performance because it's just going to thermal throttle So fundamentally the essence of that is this concept of Denard scaling which is voltage is quadratically proportional to power So if I 2x my voltage my power goes up by 4x If I cut my voltage in half my power goes down by quarter So we asked ourselves how could we run voltages lower than GPS And we talked to a lot of people about this. We flew out to Silicon Valley after dropping out and basically asked dozens of people and semiconductors and all these different chip companies how they did it. And the answer we got was like, you can't. You can't run at voltages lower than GPS. And this was very dissatisfying because there was many different industries of chips that run at voltages lower than GPUs. Bitcoin miners run at under a quarter of the voltage of GPUs. So this is obviously physically possible. The question is, are there issues with GPU architectures that make it unable to run at these voltages? And when we looked at the problem for a long time, we were able to create a new mechanism of running at much lower voltages, a new type of power delivery that we call low voltage inference. And we think all AI chips in the future are going to be low voltage chips. that are going to have to cram way more flops in the same silicon area and without thermal throttling run at way lower voltages. So that's pretty so. For Decode, it is all a memory game. More memory bandwidth. You can load the model faster, load the KD cache faster, and serve more tokens per second per user. We think people ask the wrong question here. People often ask how much memory bandwidth is on your chip You should be asking how much memory bandwidth is on your full scale up cluster What we able to do is add way way more bandwidth and a much lower latency from the chip to our interconnects. It allows us to be able to go serve models at this much higher speed because you can go use the SRAM and the HBM from the full scale up cluster as a single pool. And that's our second key technical bet, what we call cluster scale memory. And on GPUs today, the cluster memory bandwidth is often very badly utilized because the time to go hop from one GPU to another is extremely long. For example, on Blackwell chips, it can be about 4,000 nanoseconds to go point to point. And that means that if you go ahead and go to an 8x TP setup, you will get way, way less than an 8x improvement in your tokens per second per user. And what we did is built our own totally custom interconnect stack. We took everything above the second layer of built a full custom. And we can go out and do far for better latencies and a balance this way too. We can go ahead and cut this by more than a factor of 5x. And that allows us to then use the memory of other chips much more effectively. As you scale the world size, your time per token goes down proportionally. Yeah and it not that surprising given all these architectures were built before ChatGPT So if we trying to build a chip for the modern workloads it going to look very different The way we organize our flops the way we do our voltage domains the way we do our power planes are going to look super different The way we do the packaging is going to look super different. The way we do the board design is going to look different. And then the decode side, the way we connect everything is going to look very different. So we're now bringing forward our first generation of this low voltage inference technology, which is running at under half the voltage of any other AI chip. If you zoom all the way out, why is this so important? Like why is the delivery of much higher throughput, much lower cost per token, better tokens per watt, like all of these metrics that the universe is going to start talking about more and more? Everyone knows the supply side of the equation is a big problem right now. Why is this in a bigger picture looking at a decade, the bottleneck in the technology world? Well, I think it comes down to productivity, where we are at this extremely interesting moment in the history of civilization, where there's real artificial intelligence. Not like sci-fi stuff, but like these models can solve problems that most humans can't. And like it's going to create new scientific discoveries. It's going to create instant access to medical care, instant access to education. And now it's just about how many people can use this at the same time, how many products can serve this at the same time, and also the speed of doing different tasks. So when you think about wallpapers, [15:00] clock time. If we can take an agent that can run at a certain model quality and could take a year to solve a certain task using inference time compute, if you have way faster decode speed, you can compress that into a month. So the amount of scientific innovation and the amount of actual proliferation of technology will happen much faster. And then the second part is concurrency, where today it's just not possible for a billion people to use these models concurrently. Ultimately, some people are going to get downgraded. Some people's model is going to be slower. Some people just won't be able to access the hardware. A few years from now, there's going to be giant models serving billions of users. We're very much in the early innings of AI today, where the paid plans, there's only a few million users in the world using paid plans of AI models. So we're at one one thousandth of the global population actually using this stuff. So if you want to serve a giant scale, a lot of things change. And one of them is the number of chips that communicate together, where people usually think about this in the context of training. You have these giant training clusters, you have Colossus with over 100,000 GPUs that are all networked together. And the inference side, today, people usually think about it as an eight-chip cluster or maybe just, you know, NVL 72, the scale-up domain. But very quickly, this is going to become thousands of chips and tens of thousands of chips. And the way to get the most performance there the time between sending data from one chip to another that primitive matters way more than is getting credit right now So when we think about optimizing memory bandwidth for the system you have to think about how fast these chips can communicate together Because if they can only communicate really quickly with themselves and very slowly with other chips, you're not going to actually be able to serve giant models at 10,000, 20,000 tokens per second. So we need multiple orders of magnitude of infrastructure built out throughout the entire stack, from the wafer to the watt, transistor to the token, to actually bring this stuff to the world. And I think that if you look at most other goods, like the iPhone, for example, they've gotten into this economy as a scale, where as a result, more money does not really buy a better iPhone. That if you're a billionaire, or if you're just the average American, you buy the same phone. And tokens aren't like that yet. We're still in the very early days, where the general purpose system, relatively small one, is kind of handcrafting these tokens, like they made screws back in the Renaissance. And I wanted to live in a world where you have the same economies of scale for token making that you do for making, say, iPhones or cars or anything else. I think that is one of the huge unlocks that allows a huge group of people to go use the best quality models. Economies of scale have all made capitalism very I don know fair I think that allows you to go ahead and have the same product in many many different hands And you able to go then serve way more users on a single scale up cluster It allows you to get closer to that point for token serving too Yeah. And also just certain products aren't usable if they're slow. So if you want to serve coding models and you want people to actually use them, like there's a certain number of tokens per second you need to hit. So the question is, while maintaining that per token speed, how many users can I serve at the same time? And you can basically decide I'm going to shut off a bunch of the world from using this stuff or everyone's going to get a worse experience. So fundamentally, you need to find ways to push out the curve. And that's why there's such a pressure for new hardware. Vanta automates security and compliance for over 16,000 fast-moving companies like Ramp, Cursor, and Harvey, keeping them audit-ready around the clock. It's the number one agentic trust platform, and it now helps companies like yours watch for the risks that show up between audits across your vendors, your AI tools, and your whole environment. Every new tool your team signs up for, every vendor that turns on AI features is an opportunity for something to go wrong. And most security programs weren't built for AI's pace of growth. The Vanta agent works like a 24-7 GRC engineer in the background, finding issues, drafting fixes for you, and cutting vendor assessment time by up to 50 Whether you a fast startup or a global enterprise Vanta helps you earn and prove trust Invest like the best listeners get a special offer for off at vanta.com slash invest. Ridgeline is the first end-to-end system of record with embedded AI for investment management firms running portfolio accounting, reconciliation, reporting, trading, and compliance on one unified platform. Firms are moving off legacy technology and onto Ridgeline because of how far ahead Ridgeline's AI features are compared to anything else in investment management software, which is why I believe that firms that come out ahead in the AI era will be the ones running on Ridgeline's unified platform. If you're serious about your firm's AI strategy, Ridgeline should be part of that conversation. You can request a demo at ridgeline.ai. I'd like to take some time to step back and hear both of your stories for how you came to this idea and this company. and then kind of walk through what it's been like to build it. Because I think in so doing, we'll understand the system that you've built for the company itself that we'll then be able to power subsequent generations of products like this one for this crazy inference feature that we're staring down. Rob, maybe starting with you, just take it however far back you want. But what I'm curious about, [20:00] And your personal story was actually the very first thing I ever heard from either one of you was your personal story many years ago now, which really kind of blew me away. I'm most interested in your motivation ultimately for being here doing this thing. It starts back actually in high school for me. I've been very unlucky and lucky at different points in life. This is one of the tougher times. At the end of my sophomore year of high school, I got injured at a martial arts tournament the next day. I couldn't walk for some reason. I thought there was something wrong with like my SI joint or something. I went through physical therapy. I did different types of scans. I couldn't figure it out. And eventually they found this big bump on my back and an MRI and told me it was a tumor. Stage four bone cancer was told I had under 30% chance of survival. It was like a two-year crazy chemotherapy, surgery, learning to walk again experience. And when you go through something like that, it changes the Overton window of human experience and makes you appreciate what actually matters. And you also ask yourself, what are you going to do if you have the chance to live? if you actually want to get through something like that, you need to be hoping for something. And I always knew I wanted to do something very impactful if I had the chance to get through it. And it took me a couple of years to figure out what that was going to be. And at the same time as I got to college and met a bunch of other people building cool tech I got extremely excited by AI models especially once GPT came out And I was like wow this is the first model that can kind of speak English And these things are going to get really smart What happened was when GPT came out there was GPT which was the first model with image uploading. So I went through my camera roll and I found a picture of my back with this bump on it before I was diagnosed. And I said, hey, chat GPT, pretend you're an expert doctor. A patient comes in and it says they have this bump on their back. What could it be? and it immediately says like, this could be a tumor. You should get an MRI immediately go to the doctor. And I just kind of sat there still. It was like, that took me six months. And yesterday, this feature wasn't there. Today, it's here. I go to show my parents and I got this notification being like, you're all out of image credits today. You need to get a pro plan. And I was like, holy crap, this is going to change everything. And we clearly don't have the infrastructure to serve it. And there's very few things you can work on that can actually bring this technology at scale to the world faster. I mean, there's like plenty of people that are super smart working on models. The fab seem maybe unreachable to work on, but it seemed like the hardware was all designed before ChatGPT. Every GPU, every TPU, every AI chip that was serving these models were just fundamentally built before this and are retrofit to serve these modern models There gonna be an entire new wave of architectures that came out And what a more exciting thing to work on than bringing this to everybody Very different angle at the same time I was running a startup incubator called Prod, which has incubated a bunch of different companies. Some of the earliest ones being Cursor and AnySphere, which merged and Mercore and Etched went through it, a handful of others. At the time, as these models were getting smarter, it was 2022. I was realizing all of these companies are spending all the money they raise on compute. and I had this realization as I was working on some of my own stuff that like, oh my God, all the products I want to build are going to cost tens of millions of dollars a year in inference. This is not going to be tenable. The cost structure of every software company, the cogs is not going to be like zero anymore for an incremental user. It's going to be quite high and it's going to be a function of inference. And then the optics of every business is going to also be inference as people use more and more coding agents. So fundamentally, it seems like inference is going to be really important. And it feels like we are on a decade march for inference to become the biggest market in the world. So when you think about that, 10 years from now, there's gonna be these giant projects where everything in that data center fundamentally hasn't been designed today. We should go pick something and work on it. That's kind of how it got started. Yeah, but I'm really excited for you to go back about as far, probably early in high school, maybe even earlier, and tell your favorite hash marks on the timeline that ultimately led to your ambition to drop out of Harvard and start this company My first job ever was at a company called Exnor where I did colonel development I was 17 and a 17 can sign illegally binding contracts. So rather than go ahead and do a traditional CIA, they went ahead and sat me down and said, Gavin, don't share this information. And Exnor was one of the only companies that saw, hey, maybe this is a good trade, and I have to work building colonels. And a number of other company since. XMAR got bought by Apple for $200 million. Did the same thing at Octo. They got bought by NVIDIA for hundreds of millions of dollars. But when you do this sort of kernel's work, what you realize is that the math is relatively easy, but to get high-speed decode, the thing that matters is data movement. Almost all the work that you do is optimizing how do you move data around a single chip or across multiple chips. That's why we went ahead and built this cluster scale memory tech. We bring that interconnect time way, way lower. You can go do way more movement And as a result, get a much faster time to generate each subsequent token and build these crazy things Rob's talking about for doing a year's worth of work in a month or more than that in the future. Can you talk about the competitive drive that's evident in some of the high school competitions that you participated in and won? We did a couple. For example, I was very active in the FTC robot. [25:00] I'd like you to have a very talented partner, Sanford. For a long time, we were part of a traditional school team, where it was about 20 guys, all working together as often typical at first tech challenge. And the goal is to go out and build a robot that scores the most points, and a bunch of other things too. And first, they put a lot of emphasis around collaborating with other teams, around trying to do really good documentation, around trying to go ahead and get others inspired to go do the same thing. Sanford and I decided, rather than go ahead and do it this way, we're going to win. and we did nothing else besides the builder about the score the most points as a two-person team as a two-person rather than a 20-person team that we were much much smaller than almost every other team in the competition we figured that if we were going to go specialize if we're going to go out and do this win the damn games really well we wouldn't need to go ahead and advance based on the quality of our documentation or of our outreach we were just going to go win and so we did we have branched off built a two-person team that built a robot and decided we were going to go and redesign it every three months. And we did. We actually had the world record for the highest four during this competition at one point. We were rated by OPR third in the world for software development and it was a damn good machine What from that episode can I translate as an analogy onto how you built Etch the company With thinking about how you want to go ahead and do a full Rackscale product like this there a couple of key ideas One of them is like velocity velocity velocity that you win by shipping You're not going to go out and win by having the best outreach or the best communications. You win by building the best product. And similarly, we think we can do it with a lot fewer folks, that if you're willing to go ahead and just focus on product, product, product, and parallelize relentlessly, you don't need 20,000 people like the big companies have. You can do the best product in the world with far fewer people. Yeah, there's a saying of the best part is no part. I think for us, it's also the best vendor is no vendor. As much as possible, we want to vertically integrate the entire product, both because we get more performance, but we can move way faster. So everything from the chips to the boards to the cold plates to the interconnects to even the production, we want to do all of it as in-house as possible. I think we're the only startup right now that's building its own rack as well as its own chips. And we did it all at the same time. A couple of years ago was the last time we were public. At that point, we just started building our rack team and we brought over Brian Loyler, who built all of NVIDIA's HGX and DGX systems, which is like 80% of their revenue. And we said, we going to build the rack at the same time We actually went through multiple iterations of the rack before the chips even came back Before the chips came back we made thermal chips that had the exact same hotspots as we expected our chips to have So we could build the cold plates We could overpressurize them and blow them up We haven had a single leak since our chips came back with the cold plates because we already validated them. We have a factory in Taiwan. We have a few dozen people out there. We built a clone of a bunch of the test stations in our office. We have a two megawatt data center on this floor and we did 24-7 development cycles. People are doing day shifts and night shifts to actually get the hardware up and running as quickly as possible. But it's that extreme vertical integration and extreme parallelization of the schedule that lets you get products to market way faster. If you think about the building of the early team and what it required as two young guys building this company, there's lots of very talented young entrepreneurs out there, maybe for the first time of this scope or magnitude in a long time, all of whom probably could benefit from the lessons that you've learned, getting very sophisticated, talented people to come join you even after careers at the other great companies. If you were teaching this as a class, like here's how to get lead talent when you're young and inexperienced and naive, what would be the syllabus? We have a pretty bimodal talent philosophy. It starts with what we call the legends, which is when we're trying to solve an incredibly hard technical problem and generally do something that hasn't been done before, we need to find the very best person in the world And often the number one guy in the world versus the number 10 guy versus the number 100 guy huge difference in whether it actually possible to solve the problem We created this system we call project recruiting where we map out all of the hardest technical problems across all industries that anyone has ever had to solve. We look at temporality. So who are the people who did the zero to one? Who is in charge, quote unquote, who actually did the work? We talk to as many people as possible, and then we just track it. And you'd be surprised by the amount of people who say yes after the first conversation is pretty low, but the amount of people who say yes after the 20th conversation is surprisingly high. You really got to keep at them. When you hear no from somebody who really is the best in the world, then that really means, hey, you should go ahead and come back when you have a few more milestones for giving out. I think it's one of the most convincing things to see is, hey, we make bold claims. And when you go ahead and hit those again and again and again, that is really belief inspiring. When we decided we wanted to build a rack and not just a chip, we were looking at this and we're saying, how many products have actually shipped that scale for a rack scale system that actually have the power density that we're trying to solve? And we just kind of said, if we were gonna wave a magic wand, what would the best possible person in the world look like? And we were like, well, if we could find somebody who started at Nvidia and built the entire rack team through all their different generation, learned all this different stuff, [30:00] scrappy, still understands the startup culture, but has seen scale, like that would be the best possible person. So we mapped all of the different teams that related to all of the different rack scale products of NVIDIA. And we found three people that we thought could fit the bill. And we talked to all of them. And two of them have just retired. And one of them was planning to do one more generation for NVIDIA and then retire. His name is Brian. And over time, we convinced him to join. Brian started the HGX and DGX team at NVIDIA, which was a majority of Nvidia's revenue, tens of billions of dollars a quarter. And the other two guys ended up investing, by the way. But when you have somebody like that, they just know what good looks like because they've seen it. And there's so many times where we would talk to Brian and he'd just point to us and be like, that's a billion dollar, like a billion dollar lesson I learned, a billion dollar lesson I learned. And like, you know, that just saves us cycles. And you pair someone like Brian with somebody like Sanford. Do you have a name for them? So Brian's a legend. What's Sanford? Yeah, we say chips on shoulders, put chips in data centers. So Sanford and Gavin in In high school, we're world robotics champions, and Stanford was finishing his senior year of college. And we called him up a couple of years ago, and we said, hey, can you come check out what we're doing? We need some help on the platform side He comes for a week and we say can you build a cold plate this week And if you asked like any thermal engineer anything like that they would think you just like totally naive right I mean, these things take months to do. And like, to be clear, they do. But you can make real progress in a week if you put your mind to it and you think it's possible. And he built a contraption in a week that like actually de-risked like a pretty key power question we had. And you put those two together and they've done incredible things. One is not possible without the other because you need the extreme. driven people that just keep asking why and don't know where the bodies are buried to like take tons of aggressive risks. And then you need the people who've seen scale and still have the startup scrappy mentality to help them along the way. So it's really the legends plus some naivete, raw first principles type talent. It's not just that you have both in the company, it's that they're working together. That's right. If I think about that funnel, anything else more interesting to say about how much better you've gotten at recruiting and like why those metrics keep getting better? One of the shocking things is actually being such a contrarian bet kind of self-selects. But like, you're the kind of person who is some opportunistic, she's going to go join whatever the hot company is, or I think go ahead and do deep diligence, you will not come work here. And it's one of the things I worry about as we announce more and more of the product and the specs we may lose some of this if we not very careful You kind of have to be sick in the head to join our company When you think about it on paper it like you a person who is probably a very accomplished engineer, making a good amount of money, it's liquid, it's predictable somewhere else. You're going to convince your family to move to San Jose and live in this apartment on this housing program for the semiconductor company run by two, what, 24-year-olds now, that's pre-product, that is going against the biggest companies in the world, and the most supply constrained environment ever created with a design that they're saying is not going to be like 10% better, but it's going to be 10x better. Something must be wrong with you to do that. People are just wired differently here that they like really want to not prove people wrong who don't believe, but prove people right who do believe. They just take it personally. And, you know, that's really fun to find those people. And frankly, just the nature of the company makes it very easy to whittle out the people who aren't like that. One of the very first things you and I talked about, Rob, was I started asking about Sohu, which is the name of the first product here. And you said, we can talk about that in great detail, but the thing you should know is that what we're really focused on is building a machine that can, at scale, produce these things and generations of them as efficiently at the highest possible quality levels We want to build like the company or the machine that is the company is the thing that will produce this thing and then subsequent things So I like to talk about a few principles or cornerstones of the company We alluded to some of them You said velocity you said vertical integration these have become more popular topics. Parallelization is something maybe that we should talk about. But I'm especially interested in your guys' willingness to take huge risk to go faster. Maybe tell your favorite story about why this is the philosophy, what it's allowed you to do that maybe other companies haven't done? There's a number of stories here, but one of my favorites is there was a time where we were getting close to taping out the chip. We realized, wait a minute, one of our vendors is way, way behind schedule. And we had two very bad options. One option is to keep the current vendor and push our timelines out by on the order of a year. Another option is to switch vendors, start over, and also you push timelines out by a year. And either of these was a good option. So we had to go look for option number three. And what that was was we figured out they're all in Bangalore. They're actually going and doing the work. We went out and shipped a dozen of our top engineers across the world to Bangalore for six months. So I was there as well. I lived in Bangalore for four and a half months personally. And every morning, we'd go ahead and walk across the crazy busy Bangalore streets into the office. We'd be the first ones in. [35:00] We'd go out and build a wide variety of tools, both things like, hey, auditing a huge amount of the code that was going in, building a bunch of tools as well to make this go even faster, making sure we're making the right design decisions on the spot right there. No 12-hour back and forth. Go ahead and decide immediately. And then at 1 a.m., we'd go walk back through the now empty Bangalore streets and do it all again the next day. We still had a bunch of the team in the U.S. we ran these 12-hour on each side handoffs where we had 24-hour development cycle. We're at 8 a.m. and 8 p.m. every day. We'd all get on the Zoom. We'd share all the data. We'd say, when I wake up, like, this must be done. Like, we must get this chip out. And it was extremely intense. At the same time, we saw other chips at the same stage as us with that same vendor that ended up taking years that still aren't out today. They still have even taped out today. And it's that level of extreme urgency that's required to bring products to market. What is the key to doing this well, this has become a trope because of Elon mostly that like his special skill and others that seek to emulate him would try to do this too, is figure out like what the binding constraint is and just flood the zone personally on that thing, which is kind of like going to Bangalore or something It seems like this is a central tenant of the business and of any business that going to do this kind of vertical integration What the key to doing that well Like again what have you learned about that specific act For me I think there are two key tricks to this The first one is that you can't build a chip alone. It's got to be a team problem. And your most important job is to go get great people to go with you and great people to go ahead and be inspired and excited to go ahead and do crazy things like this. That is a huge ask to go say, hey, guys, uproot your lives for six months or in one case, 12 months that we had sent one guy out well ahead. It sucks, but we're lucky to have team members who are in it for the right reasons. But I think the second big thing too, is being able to make decisions very fast. That one of the worst items is when there's a factory or there's a vendor who is waiting for you to go ahead and make some call and has been just stalled. And this happens all the time, even for very small things. Two, send folks, delegate a big amount of responsibility to them, and say, make a reasonable call. Okay if you're wrong every now and then, but I would much, much rather be right most of the time and give an answer immediately than wait every time for the perfect response. Speed wins. What about spending money to go faster There this learn by doing thing which has become so interesting and as the world has gone away from software and towards more hardware again in the world of technology that we outsourced so much of the learn by doing by shipping stuff overseas and effectively just being the ideal guys here in the US. It seems like that obviously is reversing. And you've adopted this way of learning by doing. You want to be in that iteration learning loop. Absolutely. And part of that is willingness to spend and take risk with dollars. Yeah. Can you talk about that a little bit? I think there's a great quote of like, the biggest risk is not taking risk. Very similar here, which is like, every day there's over a billion dollars in revenue in this category, and a lot of it's inference. So every day we don't ship, we're just leaving tons of opportunity on the table. So your willingness to spend money should be extremely high if you can get a very clear ROI out of it. So we have this concept that we call prefetching, which is when you're waiting for one thing to get done, when you know you're going to do other things once you have it, is there ways that you can parallelize the entire schedule. So for example, we know our chip is going to come back on a certain date. We want it to be that everything possible that could be done without the chip is done before the chip lands. And this costs a lot of money. This means that we want to build our entire software stack beforehand. We shipped racks to customer data centers without our chips in them with all the networking all the CPUs all the storage all set up so we could bring all that data center software up before the chips came back It meant that we took over 700 FPGAs and put the entire full reticle chip on an FPGA cluster and ran a dozen different models with our full inference stack on them before the chips came back. It means that we built a thermal chip to mock the thermal profile of our chip and built cold plates based on that before the chip came back. It means we had the entire production line ready. It means we did many revs of the circuit board. It means the entire product was ready to go before the chips came back. And this is what it gives you. There is another very famous AI chip company that took 10 months to go from getting their silicon back to having them running inference in a rack. And this was like publicly announced to their investors. And it was a really big deal. We were able to do it in 40 days. And it's because by the time the chip came back, everything was boring. The software was already written. The rack was already there. The production line was already set up. We were just go, go, go, go, go, get everything together. You don't always catch everything. You make some tweaks on the fly and then off you go. In that particular case too, that was a big part of it. But also like the shift, I think made a big difference too. We went out and literally had a day shift and a night shift. There were team members who would come in at 10 AM and leave at around midnight. She would come in at midnight and leave at 10 AM. You're running [40:00] around the clock till we get to that 40 days. Yeah. I mean, over half the company lives next to the office. So it makes it easier to do that type of thing. You pay them to do that, right? Pay them extra. Do you still do that? The invisible hand does wonders. I mean, hey, it works for me too. We're both there. I'd love to take one big step back and talk a bit about just the broader ecosystem here. The amount of shortages on the supply side, the exposure of risks in the global system and the supply chain around this stuff has become like everyday Wall Street Journal front page news, like the stocks that people are watching and investing in and excited about, about the memory stocks. These were, you know, boring commodity, like nothing burgers five years ago. And now they're the center of global attention. If you just assess because you've been building in it, the global connected supply chain that's required to make stuff like this possible, just riff on it. Like what scares you? What's working well? What needs to change? What do you hope you change by virtue of how you build this thing. What's your assessment of this story right now? I think that one of the most undervalued pieces of the supply chain story is almost none of these things are you buy them and they don talk to the vendor again You have to go collaborate That is the most important part being successful I think in ships with TSMC or with memory vendors You need that partnership. I think that for TSMC in particular, people don't understand why it is so valuable. People look at the tech, and the tech is the best in the world. But for me, the real value is all in the service. TSMC customer service is way, way better than I have seen at any other company in any other industry. It's the kind of thing where if you say, hey, you can approve your yield by making this change. You can go make them a recommendation and then we'll go run an experiment on their own dime in our case to see if they could actually get the higher yield. And when we found that we were right and the experiment worked, they moved over the rest of the line. And that kind of thing doesn't happen in most industries. If I go to like the steelworks plant and say, hey, I want you to change the competition of the steel, and they'll say, screw you. Not TSMC. It is why they are the number one and why they're going to win. One of the things that matters a ton is power availability and time to power. And the problem is the more power you want, the more shortage there is. It's actually very similar to chip clusters, which is like, why is Colossus charging $12 an hour for Blackwells? It because they the only place you can buy 20 of them at once right Why is the 500 megawatt data center so hard to find It the exact same reason One of the things we need to think about is like how do we get way more juice out of each megawatt People are looking throughout the entire stack, whether it's just improving the PUE, but also entirely new hardware to get the most tokens per megawatt to solve this problem. But fundamentally, building new buildings is hard. It's much easier to go from 100 megawatts to a gigawatt than from a gigawatt to 10 and 10 to 100. and we are pushing the limits of what's possible on these timelines. So there's a lot of people trying to scale in their data centers as much as trying to scale them out. One of the interesting things about a system like this is what it replaces. Yeah. If I think about a rack like this versus, I don't know, a set of Blackwells or something, or Rubens or whatever's coming next, how should I conceptualize that? It's not just watts, it's also physical space. The Reapers talked about this in their recent readings called that literally this is a problem that there's literally no space to put the systems. How should I conceptualize what this represents or replaces in terms of other units of compute? Here's how customers think about deploying models generally, which is when I'm building a data center or I'm building a cluster, it's not like in the abstract of like, oh, I like these chips and like, yeah, this is the power footprint and so forth. It like I have a real production workload I trying to serve And for my product to be useful there a certain speed I need to serve for that And for certain products it really fast and certain products it really slow Whatever the speed is this is my speed The question is in a given amount of power how many users can I serve while guaranteeing that speed? So another way to put it is ISO, what's called interactivity, what is my throughput? We are just finishing kind of the early innings of the AI infrastructure boom where people really just cared about speed. GPUs were not able to reach a lot of the speeds of other types of chips, like all these SRAM chips, thousands of tokens per second. And that enabled tons of new use cases that got people very excited. There's an entirely new wave of AI chips, us being one of them, that are all going to be able to hit these speeds. The question then is, if you're hitting these speeds, what is the number of users you can serve at the same time? And by proxy, if I have a hundred megawatt data center, how many software agents can I run at the same time? So when people are doing that evaluation, our hardware is going to generally be able to get you an order of magnitude more concurrency at a given level of interactivity. So that directly translates into tokens per watt, tokens per dollar, all the things people care about when they're actually serving these giant mixture of expert models at scale. There's these now famous interactivity curves. Not many people publish them, but you can see a Blackwell curve. You can see an A&B curve, which is a little bit worse than Blackwell's, and it's still an $800 billion company. [45:00] So if you think about what then the impacts are of shifting that curve, not just a little bit further out, but much further out, what are the things that most excite you about what this will enable? I mean, I want to go out and solve some of the hardest problems, and I want to go solve these in much less time. There were things growing up that I was not sure I'd be able to live to see. For example, the new district was a conjecture, one of the things that we talked about in college. And I was not sure I'd see that proven in my life. And this was done by an AI model. And it was done over a long period of time. But if you're able to then run the same model 10 times faster, you can go shrink the time to go have these breakthroughs. And there's a huge number of other problems in math like this as well. That I worry it will take a thousand years to go prove a thing like this. You can either have a much smarter model or a model of the same intelligence running much faster. You can then shrink that and I can see it. It's so cool seeing these breakthroughs get made. I am so, so excited to see much more of this happen. I think too often people think about tasks and applications and stuff in these very short time horizons. Doing a chat and it's like 50% faster is nice, but it's not like game changing. As these agents go longer and longer time horizon and the models get more and more capable you going to see gigantic bodies of work that would take months of compute And we think about this in wall clock time Like if you talk to a pre researcher at a lab they tell you that wall clock time often is one of the most important things that matters. And what wall clock time means is the time from starting your run to finishing it to actually get data back. If you can shrink this time from a six-month run to a two-month experiment, you're going to be able to do many more iterations and people will make changes on the model architectures to actually improve the wall clock time. Very similar here in terms of how we think about the use cases, which is the exciting part about super low latency decode is wall clock time on long horizon tasks becomes much shorter. So a year long compute build would now take a month and that month long compute build will now take three days. And that three day compute build will now take seven hours and so forth and so forth. So that's the thing that I think is really hard to internalize because the models are just getting capable enough to do this stuff. I thought it was really cool months ago when Cursor published that they had a bunch of coding agents build an entire browser from scratch in a week. Totally nuts. And that will soon happen in under an hour. And there's going to be many of those types of things that are going to happen with these massive parallel agents all working on a given task What are the ultimate limitations of these systems Is it just like a physics question Like how many times faster cheaper can we get theoretically There's a lot of room at the bottom, as they say. If you think about chip-to-chip latencies on an NVIDIA product, you're looking at 4,000 nanoseconds to go from one chip to another. We'll be able to do much better than that. What's the mathematical limit? Is speed of light. You can do it in just a handful, like two, three nanoseconds. And there are 4,000? 4,000 today, there is a lot of room at the bottom. The same thing for things like power efficiency that, sure, we're able to go and shave a huge amount by bringing the voltage down by so much, but you can go lower. You can go much, much lower. It's very challenging, but when I think about 20, 30 years in the future, then I think it's inevitable. And also for economies of scale, for cluster scale up, for a long time, 8 chips was the biggest scale up domain. and they even had the NL72, bringing it to, well, 72. But you can be way, way bigger. You look at like a fab, for example. You have a $40 billion single monolithic building with only a handful of lines running through it. You could have the same kind of thing for some futuristic mega cluster billion billion as a giant mega token factory serving one or a handful of models for a massive number of users to get that same economies of scale thing Same model massive number of people You mentioned kernels engineering and that being your first job. That has emerged as a thing that nobody had ever heard of in their lives to now something that you hear about all the time, the importance of it to eke more performance out of the bare raw metal. When will that just be something that AI doesn't entirely as well. Are humans still the best kernel engineers? Are they doing it with the assistance of AI systems? Like how far down will humans still be in the loop of designing these things? Like when will that go away? Today, it's all very hybrid. And the best kernels are still written by human AI collaborations. Also, any AI model is built up both these fundamental primitives like map moles, like convolutions, like chip-to-chip operations collectives. And making these overlap and making these really fast matters enormously. And it's the kernel designer's job to go ahead and figure out, or can I overlap? How do I allocate memory? How do I verify that if there's some issue, like a retransmite, it doesn't stall the whole pipeline? These things are very challenging, but they can go make your performance be, say, 3-4% better per optimization. And you can do so many. And when we thought about our software stack, we wanted to go skate it with a pocket [50:00] going to be. And three years ago, there were kind of two ways you could build software. One of them was to invest heavily in graph compilers. These things are not very performant, but they work out of the box. They don't require a human to go come in and tweak all the kernels. But we went the opposite direction. We are kernels-first programming, and that means that it, for a long time, did not work out of the box. But if you were a kernels expert, you could get incredibly, incredibly high performance. And the thing about this is that now, as the coding models get better and better, they're doing more and more of the kernel generation task. And when the models keep getting smarter, it'll eventually do all of it. It will become superhuman. So we're going to build for where the world is going. And even today, we think about our profiling tools or debugging stack. We think about it from how will the model use these tools more than we think about how will humans use these tools? We sometimes run experiments internally, and we had Codex actually get GPTOSS running from scratch just based off of our docs completely by itself. And it did it, I think, overnight. We think about game selection a lot, and what we mean by that is making sure we're investing our energy in the right bets Because regardless of what you choose to work on it will take tremendous effort And one of the things that we started with was the decision explicitly not to build an arbitrary graph compiler not to support arbitrary PyTorch not to support arbitrary CUDA not to support arbitrary Onyx graphs But instead, we envisioned a world where there was going to be under 100 models that actually mattered. And they were all going to look very similar from the underlying mathematical perspective. and that we were going to build primitives using physics that were going to accelerate these as much as humanly possible. And we were going to allow the most sophisticated customers to have direct access to the hardware and do whatever they want. And that has saved us a tremendous amount of time not having to build a compiler and that has allowed us to actually get much more performance. And funnily enough, when we started, a lot of people dismissed this idea. And the only people that took us seriously were in high-frequency trading. They all hate compilers too. They all write their own kernels. And we've had dozens of people from high-frequency trading join the team because they saw this philosophy too. What are the limits to vertical integration? How do you know where to draw the line? And I'm starting with this question to talk a bit about the broader market. The circumstances of the broader market are really interesting to me where the vast majority of chips, of AI chips, get bought by a very small set of customers. Many of those customers are themselves trying to design their own AI chips OpenAI announced jalapeno It seems like this very funny circumstance for like the most valuable thing in the world all kind of flows through a couple chip makers a couple chip buyers They all seem to be thinking about doing each other job And then you've got the circumstance where like, okay, then these things go in a data center and you've got neoclouds and inference providers and this other part of the stack. You've got model builders and providers. Like I can imagine a world where because you have the best hardware, you design models and you build data centers. You leak outside of your current vertical. So like, how do you think about where to draw the lines for the business? We have a saying that production is the product. Ultimately, what matters here is we know inference is going to be the biggest market in the world. Whoever produces the most tokens is going to be the most valuable company in the world. So all the decisions we make is how do we get the most token capacity online as possible? And part of that is building a really good product that has way more throughput, that can run at way better latencies and so forth. So we can, per chip we make, get way more tokens online. Another part of it is like not doing parts of the stack unless we absolutely have to, to get to giant scale. So there are parts that we decided to do because it was absolutely required to get to scale, like building the rack instead of just building the chips like doing a CM model instead of doing a JDM model But there are parts of it that are kind of noise to us right now Like we not going and building our own data centers today That doesn actually help us get more capacity online In general, our customers are actually making power and moving their clusters around to get our chips online because they're such high throughput. If there was a world where other things were a constraint, we would totally go and integrate with them. But the reality is we're just purely focused on getting as many tokens online as possible. I think this comes down to economies of scale again. At certain parts of the stack, there are huge economies of scale and others there aren't. For example, on designing models, huge economies of scale there. For chip fabrication, same story. But for example, if you think about building some small metal part inside of that rack, there's not that same effect. We think the natural boundaries are on the chip side, on the bottom, and the model layer at the top. And we'll fill the whole gap between. A few weeks ago, there was a guy who was running a next generation AI chip for one of the Frontier Companies and he's trying to recruit one of our architects and actually kind of did an Uno reverse card and started recruiting the person trying to recruit our guy. Then within a week we hired him and I was going on a walk as we were kind of finalizing the offer. I was like, you're leading this super important project. Why are you deciding to join? And his answer is super interesting, which was, it fundamentally is not existential for my company for this product to win. For Google with TPUs, [55:00] Revenue comes from search. Google won't fail if TPU's failure. That's right. Meta won't fail if MTIA fails. Microsoft won't fail if Maya fails. And OpenAI won't fail if Jalapeno fails. Ultimately, this is our product. It is completely unsurprising that the best chip in the world is built by a company that only builds that chip. It's NVIDIA, right? And for us, it is completely existential for us to get as much token capacity online as possible. It recruits a set of talent and recruits a support from suppliers and from customers that view it with the level of intensity that we do. Look at the raw flop stats. You compare any of these chips built by the labs or by the hyperscalers, the flop density for, say, FB8 times FB8 is lower than the Blackwell B300. And that makes sense because they don't have to go take the risk. They just have to go build a similar enough product and not pay the NVIDIA tax. As I think about you guys building the solution, the process of doing so is solving a sequence of really hard challenges. What has been the single episode that was the hardest to overcome? We were designing the chip. We built this massive, massive FPGA cluster to go out and verify that full chip worked as is. And FPGAs are digital entities. You can go test digital logic but not analog logic And it turns out that when the chip came back we began to go see issues in our retention application of incorrect results And we realized wait a minute there a problem with a back pressuring logic across a clock domain crossing is failing. And this is going to cause the chip to produce wrong results. And it is very, very hard to solve. And we realized there was one and only one way to solve it, as we had to line up two clock signals on our chip to within 50 picoseconds. That is literally 50 trillionths of a second. And we had to go get the signals aligned to this super small granularity and do it on every chip 2 billion times a second. A lot of people said this was impossible. We had people quit. Yeah. That people literally were like, this problem is unsolvable. And best of luck, guys. Well, when you have a problem like that, step one is, okay, let's assume the problem is solvable. How would it be solved? Well, first, we realized what we have to be able to do is find a way to go ahead and move our clock phase by a picosecond, 10 picoseconds. And we had an idea. What if we had these two clocks? We had them just a little bit apart from each other. We figured out that, hey, if we go out and figure out the phase, if we then go out and use a drifting mechanism to go ahead and wait for just the right amount of time to get that always lined up we can do this extremely reliably and then lock the phases exactly where they have to be We can guarantee this never happens People were, I think, blown away that this worked. It worked as well as it actually did, but we made it work. How long did that take? It was actually about two weeks. It was dark two weeks. It was a very scary two weeks. But it was the kind of thing where when that kind of thing happens, that is the most important time to go ahead and be investing effort. that is the hardest time to go do it when you feel like things are hopeless. But the sooner you solve that problem, the sooner you can get back to building and scaling up production to mass volumes. I think a lot of our story is like, as Gavin says, assume it is possible. Assume it is possible to have a chip with way more flops on it. Assume it is possible to have a system with way lower latency between chips. Assume it is possible to create a shared memory pool that can run at way higher bandwidth. How would one do it? A lot of the time when we do experiments, we will do dozens of experiments and all of them will fail. But we only need one to work. There was multiple times, I mean, Gavin, I think, was leading the charge and our chip bring up with, I think, three different board experiments and three of them worked. And all three of them are worth their weight in gold. This is one of the things that people come to me and say Gavin almost none of your experiments works I only got to get lucky once So one idea for one of these stories that I asking about you know difficult moments in the company history is around the ability to raise capital to fund the thing I think when you started it, you knew you'd need capital, but you did not know you'd need the quantum of capital that you've ultimately raised and are spending to build the solution. And you hadn't raised money before. These are all new things, right? There was moments where it was really, really difficult because I was there, I saw it. Moments where it was extremely difficult to raise the money that you did, that without which the company would not exist, it would have died. And like many great stories, there are many near death moments. But money specifically in this new world, this isn't software. You don't just need a little bit of money. Maybe you could tell the story about the true hardest part about raising money early on before you had something that you could show people and be so proud of and performance that you could show them blow their socks off. It was just you guys talking about an idea. But talk about the early difficulties raising money because it was pretty hardcore. We've had some intense moments. It reminds me of probably early 2024, before we raised our Series A. And we were at this point where we had done enough of the architecture, done enough of the design, that we knew that the chip architecture was sound. We had to go build it, there was a lot more to do. We were ready to go into what's called the physical design stage. We needed to sign an agreement [60:00] a physical design vendor, which will cost you at least $40, $50 million. And then we had this realization as the models were getting bigger and bigger, and you're seeing these giant MOE models come out, that we were going to need to build the entire cluster, not just the chip, but we were going to need to build boards. We're going to need to build interconnects. We need to build cold plates. We're going to need to figure out all of the networking and everything. And that this was going to cost a lot more than the $15 million we had in the bank. And you're like, man, that was scary. Yeah. That like you're sitting in that moment, they think, holy crap, we can't afford this. And I began looking at like, huh, how hard is it to go back to Harvard? At the end of 23, we put together this like memo. And we spent like a hundred hours on this because we're like, we have no idea how people are going to believe us when we ask for the amount of money we're about to ask for. And it was like 30 pages. It was like extremely technical and in depth of all the different things we needed to build and like all the milestones we needed to hit and how the market was going to evolve and all the new use cases and the cost per token and all this modeling. And then we went and talked to investors and every major investor in the Valley passed immediately. They were just like okay two kids that just finished Harvard haven taped out a chip no test chip inference who knows if this is going to be a big market Everything going to be training The models still hallucinate This could all be a bubble At the time the biggest semiconductor fundraiser for a Series A was around like $50 million. We were looking at this and we were just like tallying the bill. We're like, we think we're going to spend $100 million in the next 12 months. We really want to do this. If we want to actually get to scale and actually get the performance we're talking about, this is going to be extremely capital intensive. How the hell are we going to pull this off? I think that one of our key ways we got started in this process, we thought to ourselves, what is the cheapest possible way we do this? Decided, well, made almost nothing. And if I ate nothing but ramen, then we would go ahead and spend basically just the money for the mask, the 11 millimeter tape out. And that would be that. And if so, we could probably do it on $30 million, an obscenely low number. And we actually went out and got a debt provider. They wanted to go ahead and lend us the money we need to go across the chip threshold. And there, I think it was first to go catalyze a series of other, hey, maybe we can go do one more thing, one more thing, one more thing. Yeah. So we're at this moment where we're like, if we really want to build this company because we not going to half it We not going to go do a test chip and spend years on it and let the entire AI market boom while we could be building the product If we going to do it we going to go all the way We going to need to find a way to get million I mean, I remember Gavin and I were sitting down in the office in Cupertino late at night, just looking at each other. And we're like, could we cut $500K here? Could we cut $100K here? How long could we convince everyone not to take a salary? And we're like, holy fuck, the math is not going to close. We really need to solve this. and there was a period of a few weeks where you kind of just go into survival mode and you call every person that could possibly know an investor and you're like we need a hundred million dollars to do this if we do this we think this could be one of the most important companies of all time do you know somebody that wants to take an aggressive bet that wants to like believe in us like here's all the information like we're an open book here's the team it's great people we've been working super hard we've done these things in record time but we have these you know 100 things to go? Like, do you want to do this? And like, the snowball starts and you get a million here and 2 million here. And you're like, okay, we're not going to run out of money this month. You get a $5 million check, $10 million check. You're like, okay, maybe I can buy those FPGAs. And you know, the snowball happened where you know we were very lucky that we ended up putting it all together We had a board meeting I show you the spreadsheet and we look at it and it like 103 million It like these are all like soft commits And we all look at each other and we say we going to take it And that was a series A. And it's been much easier since then. And we've raised almost a half dozen rounds since then. Many of them from those investors just doubling and tripling down. That's allowed us to get to market so quickly. Like this rack would not be possible had we not have been so aggressive. I also think suppliers, too, I think deserve a little bit of a commendation here. Yeah. That TSMC was willing to work with us back before we raised any of that $100 million, back when it was still really, really scary. Synopsis actually went ahead and let us get some of their emulators on extremely favorable terms where we pay over many years. Basically a big loan. It takes a lot of belief from your partners to go do this. But at the end of that, you come out with this very strong team and all the folks who back you are not in it just out of pure financial incentive. They believe. Why didn't TSMC believe, do you think? Oh, this is a great story. Even before you joined in, there was a conference, a semi-event, and I was one of the only young CEOs of semiconductors. And I think that's kind of a novelty. They asked me to come in there and speak. I got to the semi-event, and I am the only speaker there under 40, the only person there [65:00] under 30. I was at the time 22. I go up and I speak. There was a speaker's dinner afterwards. And by pure luck, I had me sitting next to this very senior TSMC VP. It was a very nice dinner. There's like the former CEO of ARM there. It's very bougie. Everyone's in a suit. And I'm there with this VP. And it turns out we both studied math in college. We go out and get a little piece of paper. But we begin talking in great detail about how do modern AI models work at the actual per tensor by tensor level. And the guy just gets it. And we begin talking about, hey, how do you run this very effectively? Why is global such a critical technology to make this work? And the following day, I get an email from TSFC saying, Gavin, I want to work with Etched to find a way to make it happen. Crazy. They've been a great partner ever since. And it's amazing to think about some of the tropes. And obviously, like, you should break the fourth wall here. Like, I'm a big Etched investor. I've been involved for a long time. I think that's the world of you guys. So like I'm incredibly biased in this conversation. I'm trying to ask questions that are broader and interesting and could be objections to what you're doing and we'll keep doing that. But it's so interesting to me that like when you read about investing, everyone cites this idea of like contrarian and right as the quadrant that makes all the money And it sounds really nice But contrarian means like everyone else thinks you stupid And so when you go and you get immediate no from literally everybody it is a fascinating quadrant to exist in before you become consensus. What was it like for you? I'm super curious. Well, it's interesting. At the time, it was the largest by a lot first check that I had written. I used to say I'm not a math expert or a semiconductor expert or an expert really at the time. And so it was much more of a believed in the concept of this market potentially being huge, you having made very, very clear bets on how the future was going to look, having positioned the company in order to attack those things in a hardcore way. And then just the two of you and what I felt about you was the majority of the reason why we made the bet when we did in 2023 or whatever it was. But at the time, it was the biggest. And I think the same thing you said about naivete applies to investing as does to maybe building a Semiconductor startup, which is like, I didn't know what I didn't know. And when I called experts, they were basically like, this is stupid. They laid out in very logical terms, like why this wasn't going to work and why it was such a low probability bet. And I think one of the things I've learned from it, it's just like, you kind of have to damn the base rate. Like if you invested on base rates you should do something other than what we and I do There always the index fund Yeah there always an index fund exactly So it actually never been scary for me I think probably most of that is because there a lot I don know And if I knew more about what you guys have done and the difficulty, I probably wouldn't have done it. I don't know what that says about like maturing as an investor. Maybe I don't want to know a lot more and have some of that healthy naivete. I don't know. Funny, I mean, a lot of the traditional semiconductor funds missed the entire AI chip. like all the AI chip companies and like all the coding experts missed all the coding companies. I think it's very hard to realize the constraints have changed. And when you've looked at tape outs for 20 years and you've seen so many of them not work on the first try or the second try or the third try, like you couldn't even run a workload. You totally forget that EDA tools are way better and that FPGAs exist today in a way that they didn't exist before. And all the types of validation you can do today just wasn't possible for us. A lot of our believers, either they were kind of on two sides of it. They were just believers in the market and the team, or they were building chips today and extremely technical, like the high-fifty trading firms, where they would literally audit everything from the micro-architecture and the RTL to the board designs and the schedule and the software stack And we would sit down with 10 of their people who build their own chips and they asking us such detailed questions that we wondering are they going to build the chip It was really on either of those sides And if you were anywhere in the middle like you just wouldn understand it In the investing world they often talk about variant perception, something that you see or believe that others don't, right? And that perception creates the opportunity. I think I've invested five or so times in Etched. And every time when you do it, stakes are getting bigger and bigger. It does get a little scarier and scarier. And because you guys have been so quiet in the marketplace, I think it's very easy to dismiss you. As the stakes get bigger and bigger, those dismissals are harder to hear. I do think betting on something that you see when what you hear from the outside world is very different, that perception gap equals opportunity. Exactly. The last thing I would say is the accumulated evidence of your guys' and your team's ability to solve seemingly impossible problems is one of the most interesting things a company can have. It's like a binary, like companies do this or they don't. That's the thing is a big advantage of people who have been here for a long time. You get some new joiners who are scared shitless. You see a thing like this and there are old timers who've been here for all of two years. Smoking cigars in the trenches. Another one. Yeah, there's definitely a find a way, [70:00] If you're here, you're here because you assume it's possible. So like we can't be saying it's impossible. Everything is solvable and we're just going to work at it until we figure it out. A favorite story I have, this guy who's kind of a legend in Silicon Validation who joined our team and we were doing the early stages of what's called wafer sort. When your chips are coming out of fab, they go out on these wafers and you have this thing called a probe card that attaches to the wafer before you dice it with these probe pads and you send these electrical signals to basically test which chips are good and bad. So when I slice the wafer into a bunch of chips, I can package it and only package the good ones. We go through our first wafer. It's like 2, 3 a.m. because we're doing it with TSMC over the phone in Taiwan. We have the screen with the wafer. It's all gray and each chip is gray. And then as you start running the patterns, the squares are supposed to turn green or red. And they all turn red. We're like, fuck. Like, this is really bad. Like, everybody's like, guys, take a breath. He leads back and he's like, the puzzle begins. You have to have the attitude of like, yes, you will go and you will go stare into the abyss. You will go see scary things and we'll solve them. When did you see the first green square Within a day of that But in the moment it extremely scary And there a certain type of person who just like is addicted to that feeling of just feeling the fear and solving it And we are lucky to have a lot of those people here. If you think about applying all of this earned know-how from this last several years and now thinking ahead to Gen 2, Gen 3 and beyond, what will you be doing most differently as a result of everything that you've learned? just like from a conceptual standpoint, like the way that you will attack designing and producing this next one based on what you learned doing it the first time? It took us a while to get to the primitives that we think are really what matters for scaling inference. We tried a bunch of things early on from compilers that would turn different models into FPGAs to burning weights in silicon to splitting your HBM to KV cache and weights and all of these different things. And there was a lot of cycles of learning till we got to the point that we realized that like fundamentally, if you want to run a majority of tokens in the world, you need to do three things. You need to build a chip with the most flops and a given power budget. You need to build a chip that has the lowest latency between other chips, so the biggest scale-up domain possible. And you need to produce as much of it as possible. And I think probably in the first half of our journey so far, we learned the first two and that informed the design a lot And that informs a lot about the bets we making in the future with the low voltage inference and the cluster scale memory But the production part I think in the past year has become extremely obvious How much people want to deploy this stuff if you can have it available today The best ability is availability. If I have a thousand chips today, someone's going to use them. And we need to build a chip that's not just like way better than what's been built before, but it needs to be available at many gigawatt scale. We need to be able to be building a product that is producible at gigawatts per month in the limit. As we think about that, a lot of the design decisions we're making with our next gen, which you've seen already, is just about simplicity. Removing tons of parts, trying to assemble and disassemble the thing again and again, and learning how to make it as quick in the cycle times as possible in production, making sure it's going to be reliable, making sure it's going to be serviceable, and making sure it's going to be producible at gigantic scales. What about other problems in the ecosystem that are outside of your control, such as capacity at the leading nanometer at TSMC or availability of HBM4 memory or, you know, some of these other things where like everyone is fighting for a scarce unit of capacity or whatever. How do you face up against those realities when you're trying to produce as much as humanly possible? The people deploying the most compute in the world do think about supply a bit zero sum which is there only so many wafers being produced on a given nanometer node on a given fab right And there only so much memory being produced And that's why, actually, for our first-gen product, we built it on a different supply chain than the Rubens. We're on four nanometer, Rubens on three nanometer. We're on a different HPM than Rubens and so forth. So it actually is not a zero-sum thing. It's a positive something where more is more. So often when we're talking to people deploying at scale, it's not a decision between a gigawatt of a GPU and a gigawatt of us. It's two gigawatts. And I think as much as possible, thinking about supply chain early in the design decisions, because if you have the most performant product and you can't produce it, then you're just a podcast. That's the other big thing about vertical integration, too, is for certain things like for the chips and for the memory, you have to go ahead and partner. For most of the other stuff, those are also very highly in demand components. And the more that you build yourself, the more stuff you can go do on top of what the world can currently build. It is not, oh, you're taking availability with somebody else. You're adding way, way more. I think that's how you win. One of the things I realize we haven't talked at all about is the models themselves, which is kind of crazy. The things behind all of this demand. Anything interesting that you would say about the way that you see models progressing based on what we've seen so far? I guess I'm more interested in how. [75:00] You as thinkers about hardware think hardware might impact where the models themselves go in the future. One of the most important ideas that we believe in is that machines don't think like people think. You look at airplanes, for example. Airplanes don't fly like birds fly. That when you think about how mechanical devices have to work, it's often very different. And in much the same way, for people, storing data and loading memory is very cheap for neurons, and doing math is relatively expensive. and it is the exact opposite for chips. Generally, one data is very expensive and doing math is very cheap. And as time goes on, you'll end up finding that math gets cheaper at a rate that is faster than a memory gets cheaper due to this fundamental limit on any kind of DRAM device. You should go ahead and think about how can I make my model use a huge, huge amount of compute? What if I had, for example, many copies running at the same time? What if I activated a huge number of experts? What if I had gigantic experts that I could go ahead and run on multiple server racks at the same time? That is how I think you'll build models that are the next generation of intelligence. And context too There been a lot of work on a very efficient inference What if I don load the full context in the memory And most of the time I think that makes a lot of sense You want to build a super intelligence. Why can't you go look at a billion tokens of context? Why can't you spend a huge amount of compute to go ahead and read all in a super fast? I would love to be able to talk to a machine that was able to go attend to every book ever written and short-term memory. And I think that's going to get to a point where you can. A theme in models right now is this focus on something called dynamism, which is this ability to control the level of computation and memory spent at a per token or per user level when doing attention, as well as this ability to dynamically in your chip on the fly, send data to other chips for different MOE models doing certain types of operations. And the reason is fundamentally, as we are scaling context length, as we're scaling model size, as we're scaling the amount of computation per user, we're looking for ways to be more efficient. So the first thing is, like Gavin says, a mixture of experts architecture is where maybe we don't need every parameter being used for every token. But maybe there are things where even at a token level, we can say, well, this token needs this context and this other token. They can share that memory so we don have to have overhead of using the memory as much Maybe this token is really important so we should spend more compute We should have longer context on that token So hardware that really accelerates these types of very dynamic computations is extremely important And as you can imagine, current hardware that was designed before those types of architectures have lots of overheads in doing them. So you basically end up in these really bad worlds where you have inefficient hardware at doing this dynamism, so therefore you can't run it very well, or you have these very blocky architectures that are kind of applying blunt force to many different tokens that all need more or less computation. I have two questions about the future. We've talked a lot about what you've built so far and how you've built it. The first is about the new ways that people might start using these systems, the raw technology, log of run times, things of this nature. When inference gets much cheaper, faster, more accessible, there's more tool supply and it's better. What are the things that you think people will use that capacity to do that are the most interesting and exciting to both of you. Yeah, there was a viral tweet by Noam Brown, where he said that as these models are having longer and longer time horizons, they can do tasks that take, say, six months. And there's often not enough time to go and evaluate them for such a long period of time because by that point you have a new model app You want to go evaluate instead And with tech like we built our cluster scale memory you can go ahead and run that six job much faster But there's a second piece of this too. I've talked to Noah about it. He's now an angel as well, where it's not just the time. It's also the number of people or agents who are working on this. If you're trying to go and evaluate, can a human build a rocket? You will find that the answer is no. No one person can go build a rocket. Instead, you have to go put a team together. And I believe the same thing will be true of agents too. If you want to go ask, can an agent go out and build some crazy futuristic piece of software, you will probably need a very large team. Maybe that's 10, maybe that's a million. You have to go and have this enormous amount of both cluster scale memory to go ahead and have that very short time per token and a huge amount of flops to be able to go run that whole fleet. I'm going to be a little futuristic. I firmly believe we are on a global march of inference becoming majority of global GDP. It may take more than 10 years, but it's going to happen. And right now we measure productivity as a society as GDP per capita, but really it's going to look much more like agents per megawatt or maybe agents per gigawatt by then. And while we're being futuristic, I think this is the second to last year where I'm [80:00] majority of the workforce is going to be human. I think in 2027, you're going to see there's going to be more agents doing knowledge work than humans. And it's going to be extremely interesting to see what happens. You could imagine a world where for countries, a majority of their energy ends up going into data centers doing inference. And the energy efficiency of those data centers basically governs how many agents and therefore how big their workforce is. So you're going to see, like as Gavin is saying, one agent or a team of five to 10 agents working on group projects for a couple of days. So you can do pretty cool stuff because they're smart, but it's not going to be civilization scale. What happens when you have countries that can have literally a billion concurrent agents, like a billion people in the workforce working 24 seven concurrently on the same stuff? It's just kind of unfathomable what's going to happen. And it's going to be the biggest proliferation of technology humanity's ever seen. I think as well, like when you have these huge, huge amounts of demand, you get this idea of economies of scale again. If you think about people, I have a brain, I'm not using the whole thing all at the same time. That's only part of it's going to be active, and this is the way healthy brains work. MOE models it works much the same way On an MOE model only a small fraction of the parameters is being used for any given token at any given moment But if you have a large number of users on a piece of hardware you can go kind of take that brain cut it up into many different experts on many different servers, and run a huge amount of volume through it. So you'll have a bunch of different pieces of traffic. You'll have many of them using each part of the brain at any given point in time. And you'll also make the cost per thought, cost per token, way, way lower. So I think you're going to end up with these giant scale distributed brains. The form factor of this is a big data center with a bunch of chips, a huge amount of flops, and a huge amount of scale up interconnect. You think we'll see a trillion dollar individual data center? Absolutely. It is a matter of time. It's like asking, well, you see a billion dollar fab or a 10 billion dollar fab or a hundred billion dollar fab. It is inevitable that the economies of scale don't stop at, oh, 40 billion dollars is the magic number for fabs. No, the coffee wafer keeps going down as you keep spending more money. And the same thing will be true of plants that go out and make steel or plants that go out and make tokens. A very smart alien lands on Earth and wants to know from each of you how you would frame up this opportunity that you guys are tackling. What do you say to them? What do you mean to us? Thinking is really valuable that every company in the world runs on thinking And we are entering this really unique moment in time where you have machines that can go think almost as good and as soon as good and as soon better than the best humans can. Building these machines is going to be a huge opportunity, but more important than that, the way in which you go ahead and run this kind of thinking is going to be very, very different as demand goes higher and higher and higher and higher. There's a unique moment right now to go build a new set of solutions, a new roadmap. for how do you run the future quadrillion parameter models for a billion people all at the same time on a gigantic scale-up cluster? We are in a new era of intelligence where the cost of producing intelligence is dramatically so much cheaper than the value of the intelligence that we are in a many-year, probably many-decade supply shortage of these tokens. and basically any chip or any system that can produce tokens is likely to be extremely valuable and you should find some part of the supply chain of the token. It can be everything from model training down to what we're doing in silicon and otherwise to spend time on and push the frontier and that the companies that are the largest are frankly going to be the companies that produce most of the global supply of tokens and own a majority of the supply chain of that token And importantly it people who build systems that as they get more and more chips put together get cheaper The way you want this to scale is not that, oh, if I want to go serve 10 times more tokens, I buy 10 times more servers. It must be some solution where if I want to go serve 10 times more tokens, then I get some economies of scale benefit with my, say, cluster scale memory tech. that allows me to then not charge as much as 10 times more for those NAC tokens. What a ridiculously exciting future that you guys are building to enable. When I did this with Gavin last time, I asked him my traditional closing question, so this time I'll ask you, what is the kindest thing that anyone's ever done for you? During my cancer treatment, there was a big decision I had to make. The doctors came to me and said, it's time for you to decide. Do you want to get surgery or do you want to get radiation? Here's the trade-off. If you get surgery, you're more likely to live, but you have to assume you'll never be able to walk again. If you get radiation, you'll be able to walk again, but it's not the same probability that you'll live. You may die. What do you want to do? And I was 16, and my parents said, you have to make this decision for yourself. I thought a lot about it for a long time and decided I'm gonna do the surgery. I get the surgery. One of the things you do when you get a tumor resection is they do something called a [85:00] necrosis analysis, where they look at all the different cells and say, is a cell dead or alive? Because if you have a bunch of cancer cells that are alive, you have a problem. And they looked at it and they said, you know, you usually want 98, 99% necrosis for us to say you're in the clear. You're below that. You should go get radiation. And there was only a few machines in the world that actually could do the type of radiation I needed. One of them was in Boston. I was in a wheelchair and I needed to move to Boston for multiple months. And both of my parents decided to move out and drop everything they were doing and live with me. And I'm eternally grateful. Peace. Beautiful. Thanks, guys. Thank you and risk. Learn more at vanta.com slash invest. The best AI and software companies from open AI to cursor to perplexity use WorkOS to become enterprise ready overnight, not in months. Visit workos.com to skip the unglamorous infrastructure work and focus on your product. Ridgeline is redefining asset management technology as a true partner, not just a software vendor. They've helped firms 5x in scale, enabling faster growth, smarter operations, and a competitive edge. Visit ridgelineapps.com to see what they can unlock for your firm. Every investment firm is unique and generic AI doesn't understand your process. Rogo does. It's an AI platform built specifically for Wall Street, connected to your data, understanding your process and producing real outputs. Check them out at rogo.ai slash invest.