Cerebras and Planet Labs Say IPO Regimes Sharpen Corporate Focus

Cerebras与Planet Labs表示IPO制度让企业更加专注 实时卫星数据与太空计算的结合将在十年内推动AI从盲目文本模型向行星智能转变。 该研讨环节探讨了2026年IPO市场的强劲复苏，重点突出了Cerebras与Planet Labs截然不同的公开市场历程。讨论深入剖析了向上市公司转型过程中的运营现实，包括监管阻碍以及旨在逐步释放股东流动性的创新“细水长流式锁定期”（dribble lockups）。对话还概述了在发射成本骤降和在轨全天候太阳能战略优势推动下，算力向太空转移的重大趋势。最后，与会嘉宾详细阐述了消除内存瓶颈的AI芯片架构革命，并展望了卫星数据与AI未来融合成“行星智能”的前景。 2026年的IPO市场正在强势回归，多项预测表明其公开募股规模可能创下历史新高。 (0:00) 尽管在拜登政府下面临有关其阿联酋投资的监管阻碍，Cerebras仍成功完成IPO，定价为185美元，开盘报约320美元，最终收于230美元左右。 (2:16) 上市提供了关键的信誉与稳定性，这对于向大型商业、民用和国防客户保证公司的长期存续必不可少。 (8:01) Planet Labs的股价在12个月内从5美元暴涨10倍至50美元，表明公开市场仍可汇聚巨额价值，而不仅限于私募融资阶段。 (8:01) 现代太空革命正由两大并行趋势推动：火箭发射成本降至原先的四分之一至五分之一，以及卫星的大幅小型化。 (10:07)

Tech founders and investors argue that venture-backed startups are returning to public markets earlier to gain operational discipline and capture post-IPO upside, even as computing architectures undergo radical shifts toward specialized silicon and potential orbital data centers. Reversing the Private Model for Public Discipline (29:33) The panel highlights a structural retreat from the decade-long philosophy of staying private indefinitely, noting a renewed push for IPOs in the $1 billion to $5 billion valuation range. Public markets impose rigorous quarterly accountability, which speakers describe as sharpening execution while providing crucial counterparty legitimacy. Will Marshall emphasizes that going public in 2021 at a $2 billion valuation gave Planet Labs the credibility required to secure long-term government, agricultural, and defense contracts, with roughly 90% of the company's value created three to four years post-IPO as its shares surged from $5 to $50. Andrew Feldman similarly challenges the notion that private markets capture all significant value. Cerebras priced its IPO at $185, opened near $320, and settled around $230, utilizing a structured 'dribble lockup' to distribute shares gradually based on performance hurdles. Feldman maintains that historically, the absolute majority of investment dollars are made following an IPO rather than in early private rounds, because public markets provide the scale necessary to deploy massive capital into mature execution. Silicon Redesign to Break the Memory Bottleneck (21:31) Feldman argues that silicon market share shifts only when entirely new computing workloads emerge, drawing parallels to GPUs for graphics, ARM for mobile devices, and specialized hardware for internet networking. To capture AI workloads, Cerebras bet that hardware could not rely on legacy GPU architectures, which have already exhausted low-hanging architectural optimizations. To bypass the primary constraint in AI—the latency of shuttling data between separate compute chips and memory—Cerebras built dinner-plate-sized wafer chips that place memory directly adjacent to processing cores. Feldman claims this approach delivers speeds 15 to 18 times faster than standard GPUs when running OpenAI workloads, asserting that near-instant latency is essential because end users will not tolerate slow AI query responses. The Economics and Hurdles of Space Compute (13:16) Marshall outlines an economic thesis for non-terrestrial data processing, citing a joint study with Google showing that space data centers become cheaper than Earth facilities once launch costs drop to $200–$300 per kilogram. Current costs hover just above $1,000 per kilogram, but Marshall forecasts that SpaceX's Starship could reach the economic tipping point within two to three years. In sun-synchronous dawn-dusk orbits, solar panels operate continuously without batteries or terrestrial grid constraints, collecting five times more energy per panel than on Earth. Planet Labs has already orbited Nvidia GPUs and is testing Google TPUs for space deployment, with Marshall projecting space compute could become a multi-trillion-dollar industry within a decade. Feldman offers a cautionary counterpoint, warning that establishing high-speed communication clusters among chips in orbit remains a formidable technical bottleneck. He likens orbital clustering to the protracted final phase of autonomous vehicle development, arguing that while falling launch costs will enable necessary trial-and-error experimentation, reliable space-based clustering remains an unsolved engineering challenge. From Language Models to Planetary Intelligence (31:11) Marshall contends that existing Large Language Models are fundamentally limited because training solely on internet text leaves them blind to dynamic physical conditions on Earth. To bridge this gap, Planet Labs operates approximately 200 satellites capturing daily imagery of Earth's entire landmass across agriculture, climate monitoring, and defense sectors. By streaming continuous, real-time physical observations into neural networks, Marshall envisions moving beyond text-based models toward 'Large Earth Models.' This integration of continuous satellite telemetry and AI compute aims to establish 'planetary intelligence,' enabling autonomous systems to track and respond to environmental events, supply chain shifts, and security developments in real time. While the panel anticipates a durable return to public offerings as companies seek liquidity and enterprise credibility, the realization of orbital computing and planetary intelligence remains contingent on drastic reductions in launch costs and solving intricate chip-clustering physics in space.

科技行业正经历从长期维持私有化向积极重返公开市场的转变，上市不仅能提供关键流动性并强化业务执行纪律，绝大部分商业价值亦是在上市后阶段实现；与此同时，AI大模型正驱动底层芯片架构的颠覆式重构，并与商业航天深度结合以探索太空算力及“行星智能”。 公开上市的现实与价值创造周期 (2:16) Cerebras创始人Andrew Feldman指出，尽管IPO流程充斥着繁杂的低效行政会议和文档修辞调整，但敲钟之后核心业务的执行本质并未改变。Brad Gerstner回顾称，Cerebras历经十年才走向公开市场，曾因阿联酋投资而在拜登政府时期遭遇监管审查；最终其IPO定价两次上调至185美元，开盘报约320美元并收于230美元左右，达成百亿美元级市值。关于上市时机，Feldman幽默地归因于“过去十年屡屡犯错”的摸索，并肯定了公开市场对老员工长期投入的回报。 与过去十年风投机构倡导的“永久私有化”理念不同，讨论嘉宾认为当前估值在10亿至50亿美元之间的初创企业正加速寻求上市，因为公开市场的严苛审查能够有效倒逼企业提升运营专注度。以Planet Labs为例，该公司于2021年以20亿美元估值通过SPAC上市，其首席执行官Will Marshall指出，公司约90%的价值增量是在上市后第三至第四年创造的，过去12个月股价从5美元上涨至50美元，且最大单一股东Google至今未减持一股。此外，Cerebras采用为期六个月、基于业绩门槛分批解禁的“滴灌式锁定期”（dribble lockup），佐证了历史上绝大多数投资回报实际上是在企业进入公开市场后产生的论点。 算力架构变革与晶圆级芯片设计 (21:31) Feldman阐述了半导体发展史的底层规律：唯有全新工作负载的诞生，才会打破既有的芯片市场份额格局，例如图形计算催生了Nvidia的独立GPU、移动端计算确立了ARM对Intel和AMD的颠覆，以及网络基础设施浪潮孕育了Cisco和Arista。2015至2016年兴起的AI技术打破了计算机以往仅擅长处理数字而无法深度理解图像与自然语言的局限，开启了全新的应用维度。Cerebras当时确立的核心论点是：若要实现比行业龙头高出20倍的性能飞跃，绝不能沿用传统GPU架构，因为既有路线的低垂果实已被采摘殆尽。 当前AI计算的核心瓶颈在于内存与算力核心之间的数据传输延迟。针对该问题，Cerebras放弃了传统的邮票大小芯片方案，转而研发出餐盘大小的单颗晶圆级芯片，将超高速内存直接集成于计算核心旁。Feldman声称，在运行OpenAI的工作负载时，其实际处理速度比传统GPU快15至18倍；他强调AI交互在响应延迟上极其严苛，如同现代搜索引擎不容忍任何卡顿一样，算力速度将直接决定终端用户的留存率。 太空数据中心的经济模型与工程瓶颈 (13:16) Will Marshall提出了建设轨道数据中心的前瞻构想。Planet Labs在八至九年前与Google开展的经济学联合测算表明，当地球发射成本降至每千克200至300美元区间时，在太空部署数据中心的综合成本将低于地面。目前火箭发射成本略高于每千克1000美元，但Marshall预测，随着SpaceX星舰（Starship）的推进，行业将在未来两至三年内突破这一成本临界点。太空算力的核心优势在于能源供给：在晨昏太阳同步轨道上，太阳能电池板可实现24小时不间断采光，单块板的能量收集效率是地面的五倍，从而完全规避了地面电网容量限制与间歇性能源所需的昂贵储能系统。目前Planet Labs正与Google合作测试将TPU送入太空，此前亦已成功发射过Nvidia GPU。 然而，Feldman对该设想的时间表提出了审慎质疑。他认为在轨道中构建多个芯片协同作业所需的高速通信集群（clustering），是极其复杂的未解工程难题，其研发难度堪比自动驾驶技术的“最后10%”，耗时长久且充满不确定性。他指出，目前业内在地面集群技术上仍未尽善尽美，在空间环境中实现互联更为严峻；唯有持续大幅压低发射成本，工程师才能开展低成本试错与高频硬件迭代，进而攻克轨道互连难题。 行星智能：对地遥感数据与大模型结合 (31:11) 卫星硬件的极端轻量化与发射成本在过去十年下降四至五倍，彻底重塑了商业航天生态。Marshall指出，卫星已从以往20吨重、耗资10亿美元的笨重系统演变为数十至数百公斤的模块化设备。Planet Labs目前运营着由约200颗卫星组成的对地观测星座，实现了对全球所有陆地板块的每日全覆盖成像，其高频时间序列数据广泛应用于防务情报、农作物长势监测及森林火灾等民政应急场景，为企业与主权客户提供了验证公司存续稳定性所需的关键透明度。 展望未来AI与空间技术的融合，Marshall指出当前主流大语言模型（LLM）的本质缺陷在于过度依赖互联网文本训练，因而对物理世界的实时动态处于“失明”状态。通过将每日更新的全球卫星遥感图像与连续环境监测数据输入AI系统，大模型有望演进为“大地球模型”（Large Earth Models），从而催生覆盖地表全要素的“行星智能”（planetary intelligence），这将成为继单纯数字计算之后下一阶段极具经济规模的前沿领域。 与会者一致认为，硬件架构与商业航天的双重突破正将AI应用拓展至物理现实与空间基建，公开市场亦重获科技巨头的青睐。然而，太空计算的实际落地仍受制于轨道高速互联等尚未攻克的硬件工程瓶颈，发射成本的下降速度与大地球模型的商业化兑现周期仍有待市场长期检验。

## Background 本期视频由知名播客《All-In Podcast》于 2026 年 6 月 27 日发布在 YouTube 平台上。内容记录了在 All-In 流动性峰会（All-In Liquidity Summit）上的一场高规格圆桌对谈，Cerebras 创始人兼 CEO 安德鲁·费尔德曼（Andrew Feldman）与 Planet Labs 联合创始人兼 CEO 威尔·马歇尔（Will Marshall）同台，围绕硬科技企业重返 IPO 的现实挑战、太空数据中心的时间表、AI 芯片底层变革以及创始人流动性博弈等核心议题展开深入探讨。 ## TL;DR 对谈深入拆解了深科技企业从私募深水区迈向公开市场的实战得失与战略抉择。刚完成高光 IPO 的 Cerebras 揭示了上市后核心研发与供应链挑战依旧的清醒现实，并详细阐述了其凭借餐盘大小的晶圆级芯片（WSE）攻克 GPU 内存墙瓶颈、实现 15 至 18 倍推理加速的架构创新；Planet Labs 则通过自身从上市初期遇冷到实现十倍暴涨的轨迹，力证成为上市公司对构筑国家级与跨国信任的关键价值，并详尽论证了当发射成本下探至每公斤 200 至 300 美元时晨昏轨道太空数据中心的巨大能源套利与经济可行性。此外，嘉宾们集体反思并抨击了过去十年“永久维持私有化”的硅谷教条，认为公开市场的严格审视犹如“铁磨铁”般淬炼企业的专注力，并展望了实时地球遥感与在轨计算融合下的“行星智能”新范式。 ## Atmosphere & style 讨论在老友间熟络的插科打诨中拉开帷幕，达沃斯的凡尔赛式轶事与调侃迅速炒热全场气氛。但随着话题切入硬核商业与技术细节，两位经历过近十年漫长周期考验的深科技掌门人展现出极度务实、冷峻且不加粉饰的工程师底色，毫不留情地刺破 IPO 的虚假繁华。整场对话兼具对太空算力与大地球模型的宏大技术浪漫，以及在资本流动性、地缘审查与芯片架构博弈上的刀刃向内，信息密度极高，充满了硅谷顶峰智识交锋的现实感与张力。 ## Bulletin - [04:39](https://www.youtube.com/watch?v=jLICvWE7w2Q&t=279s) AI 芯片先锋 Cerebras 克服长达十年的监管与市场波折，成功完成市值达 500 亿至 600 亿美元的 IPO；其晶圆级芯片（WSE）突破内存墙瓶颈，在实际推理中达到传统 GPU 速度的 15 至 18 倍。 - [08:02](https://www.youtube.com/watch?v=jLICvWE7w2Q&t=482s) 商业航天企业 Planet Labs 凭借运营约 200 颗卫星的每日全球地表成像网络，过去一年股价实现 10 倍跃升，且公司约 90% 的企业价值增长均在公开上市后的第三至第四年实现。 - [13:18](https://www.youtube.com/watch?v=jLICvWE7w2Q&t=798s) 太空数据中心商业可行性正在加速显现，测算显示当发射成本降至每公斤 200–300 美元时，轨道太阳同步数据中心凭借 24/7 直面阳光、5 倍于地面的太阳能发电效率及零储能电池架构，将在经济上全面超越地面设施。 - [24:47](https://www.youtube.com/watch?v=jLICvWE7w2Q&t=1487s) 针对上市流动性管理，Cerebras 创新推出了挂钩业绩与股价门槛的渐进式解禁机制（Dribble Lockup）；与会投资人对硅谷长达十年的“永久维持私有化”理念提出批评，认为公开市场的严苛检验更能淬炼深科技企业的核心竞争力。 ## Key takeaways - [02:16](https://www.youtube.com/watch?v=jLICvWE7w2Q&t=136s) 公开发行敲钟并不能凭空解决企业的现实经营挑战，上市次日企业依然必须面对未变的芯片销量、研发进度与供应链瓶颈，商业本质依然立足于工程研发与客户交付。 - [04:39](https://www.youtube.com/watch?v=jLICvWE7w2Q&t=279s) Cerebras 的上市之路历经严苛的 CFIUS 外资审查与十年技术长征，最终发行定价两次上调至 185 美元，开盘冲高至 320 美元，稳居 500 亿至 600 亿美元大体量市值区间。 - [08:02](https://www.youtube.com/watch?v=jLICvWE7w2Q&t=482s) 公开上市能够为深科技公司提供无可替代的长期公信力与财务透明度背书，这对于依赖长期业务存续保障的农业巨头、跨国政企以及国防情报等核心客户至关重要。 - [10:18](https://www.youtube.com/watch?v=jLICvWE7w2Q&t=618s) 航天生态正经历类似计算机从“大型主机”向“台式机”的演变：除了发射成本骤降外，卫星从数十吨巨型设备微型化为百公斤级轻量设备，使每日全球动态监测成为现实商业应用。 - [13:18](https://www.youtube.com/watch?v=jLICvWE7w2Q&t=798s) 随着 SpaceX 星舰推进，火箭发射成本预计在未来 2 至 3 年内下探至每公斤 200 至 300 美元的临界点，届时部署太空数据中心的综合经济性将彻底超越地面数据中心。 - [13:18](https://www.youtube.com/watch?v=jLICvWE7w2Q&t=798s) 太空数据中心可部署于晨昏太阳同步轨道，享有无日夜更迭、无云层遮挡的持续太阳能供给，单块太阳能板发电量达地面的 5 倍且免去昂贵的化学储能与电网调峰负担。 - [17:58](https://www.youtube.com/watch?v=jLICvWE7w2Q&t=1078s) 太空算力目前仍面临超大规模计算集群在轨互联等极具挑战的底层物理障碍，其技术成熟过程可能类似于自动驾驶的长尾攻坚，需要借助廉价发射成本进行硬件的快速迭代验证。 - [20:22](https://www.youtube.com/watch?v=jLICvWE7w2Q&t=1222s) 在已有成熟生态中模仿 GPU 架构击败 Nvidia 的概率几乎为零；Cerebras 转向专用芯片（ASIC）并采用餐盘大小的晶圆级芯片设计，将极速内存紧邻计算核心以打破内存墙瓶颈。 - [20:22](https://www.youtube.com/watch?v=jLICvWE7w2Q&t=1222s) AI 推理属于没有“慢速妥协”的用户体验战场，低延迟是复杂大模型推理与实时交互的核心前提，Cerebras 架构在头部模型实测中实现了对传统 GPU 达 15 至 18 倍的交付加速。 - [24:47](https://www.youtube.com/watch?v=jLICvWE7w2Q&t=1487s) Planet Labs 的实践打破了风投过早分红的逻辑，早期投资机构如 Google 与 Capricorn 的长期坚守验证了绝大多数商业回报来自公开市场长期增值的主升浪。 - [24:47](https://www.youtube.com/watch?v=jLICvWE7w2Q&t=1487s) Cerebras 与投资银行设计了为期六个月、与多项业绩及股价门槛绑定的渐进式解禁机制（Dribble Lockup），有效平滑了传统六个月到期后的大规模集中抛压。 - [27:39](https://www.youtube.com/watch?v=jLICvWE7w2Q&t=1659s) 如果超级独角兽在私募市场过度透支百倍估值溢价，二级市场将难以承载其数万亿美元以上的增值空间；更早进入公开市场有助于让大众投资者分享成长红利。 - [27:39](https://www.youtube.com/watch?v=jLICvWE7w2Q&t=1659s) 公开市场的公众监督与财报问责如同“铁磨铁，刃更锋”，不仅不会扼杀硬科技创新，反而倒逼管理层剔除怠惰，极大淬炼了团队的战略聚焦与工程执行力。 - [31:06](https://www.youtube.com/watch?v=jLICvWE7w2Q&t=1866s) 人工智能未来范式正从仅吞吐文本的大语言模型（LLM）转向理解物理现实的“大地球模型（LEM）”，在轨感知与轨道算力的融合将催生万亿美元级的“行星智能”经济体。 ## Detailed walkthrough ### [00:00](https://www.youtube.com/watch?v=jLICvWE7w2Q&t=0s) 达沃斯轶事与 AI、太空新贵的登台 在 2026 年可能创下美股 IPO 历史纪录的宏大背景下，峰会迎来了两位引领科技前沿的掌门人：刚刚完成备受瞩目的“年度 AI 芯片 IPO”的 Cerebras Systems 创始人兼 CEO 安德鲁·费尔德曼（Andrew Feldman），以及致力于“像谷歌索引互联网一样索引整个地球”的 Planet Labs 联合创始人兼 CEO 威尔·马歇尔（Will Marshall）。Altimeter Capital 创始人兼 CEO 布拉德·格斯特纳（Brad Gerstner）在开场短片中难掩兴奋，直言这股 AI 浪潮是科技史上最具变革意义的机遇。当几位嘉宾走上沙发坐定，主持人杰森·卡拉卡尼斯（Jason Calacanis）立刻开启了老友间的插科打诨，聊起大家不久前在达沃斯世界经济论坛上的碰面。 杰森在台上绘声绘色地讲起一段关于“蹭大牌名头”的幕后插曲：原本他正打算按惯例去日本滑雪，甚至为此推掉了与知名媒体人塔克·卡尔森（Tucker Carlson）的行程，结果大卫·萨克斯（David Sacks）突然来电，急迫地表示美国总统（POTUS）在达沃斯需要“全球最棒的主持人”压阵。杰森在三天内赶到瑞士，领到了一枚象征极高安保权限的专属绿色通行证，而当穿过安检闸机、看到电子屏幕上醒目地打出“杰森·麦凯布·卡拉卡尼斯对话唐纳德·J·特朗普”时，他只觉得滑稽又戏剧化，随即便在达沃斯连续录制了多场高质量的深度对话。这段自带凡尔赛色彩的调侃瞬间点燃了全场气氛。 - **Cerebras 的行业地位**：Andrew Feldman 强调 Cerebras 在其所涉足的每一个 AI 硬件细分赛道上，运算速度均位列“毫无争议的第一（fastest bar none）”。 - **Planet Labs 的核心愿景**：Will Marshall 提出太空与 AI 是天作之合，Planet Labs 的目标是每日监测并索引全地球的地表变化，使其具备可检索性。 ### [02:16](https://www.youtube.com/watch?v=jLICvWE7w2Q&t=136s) 上市背后的繁文缛节与商业本质 玩笑过后，话题迅速切入硬核的商业现实。杰森指出，Cerebras 和 Planet Labs 分别代表了当下最具战略意义的两个赛道——AI 算力芯片与太空数据中心。面对“敲钟上市究竟是怎样一种体验”的提问，刚带领 Cerebras 上市约三周的 Andrew 展现出极其清醒且冷峻的务实态度。他先是打趣称听完杰森的达沃斯经历后，自己得好好提升一下“报出名人名字（name-drop）”的功力，随后毫不留情地揭开了 IPO 过程中繁复沉重的官僚主义面纱。 在 Andrew 看来，上市过程中充斥着难以想象的、耗费精力的“垃圾杂务”：开不完的百人 Zoom 大会，各方专业人士反反复复在法律招股书里斟酌逗号的位置，几乎没有为公司创造任何实质价值。更残酷的是，上市仪式结束后的第二天清晨，一切喧嚣散去，现实毫无改变——公司的芯片不会因为敲了钟就多卖出一颗，工程师的研发进度也不会凭空向前跨越半步。账面上虽然多了资金，但如果供应链原本存在瓶颈，瓶颈依然在那里；如果与供应商的关系紧张，紧张也分毫不减。除了需要应对公开市场的投资者这一全新利益群体，公司的核心业务依然要求创始人和团队低下头、击个掌，然后立刻回到车库和工位上继续死磕。 ```mermaid flowchart TD accTitle: 硬科技企业 IPO 前后现实对比 accDescr: 展示上市过程中的繁琐流程与上市后企业不变的核心商业与工程本质 subgraph Prep["IPO 冲刺阶段"] A["繁复的流程消耗"] --> B["130+人 Zoom 协调会"] A --> C["SEC 法律文件反复核对"] end subgraph Reality["上市次日现实"] D["公开发行敲钟"] --> E["银行账户资金增加"] D --> F["产品销量并未凭空增加"] D --> G["工程研发进度保持原样"] D --> H["供应链与供应商关系依旧"] end Prep --> Reality ``` ### [04:39](https://www.youtube.com/watch?v=jLICvWE7w2Q&t=279s) Cerebras 的十年长征：从 CFIUS 阻碍到 600 亿美元高光 早期投资了 Cerebras 并在董事会任职的 Brad Gerstner 随即补充了更深层的背景。他感慨 Andrew 是一位经历过长达十年残酷风浪的硬汉战士。Cerebras 的上市之路极其坎坷，期间因早期投资人中包含阿联酋（UAE）背景资本，甚至在拜登政府任期内引发了美国外资投资委员会（CFIUS）的长时期审查与问询。正如 Brad 所观察到的，整个过程展现出极端的非对称性：“长达九年半的时间里，每一天都步履维艰，直到最后的十二个月，势头突然变得无比顺畅，所有资本都争先恐后想要挤上车。” 最终，市场给予了这家 AI 芯片领军者惊人的反馈： - **发行定价上调**：IPO 招股价区间两度上调，最终以 185 美元定价； - **二级市场表现**：开盘价冲至 320 美元，目前稳定在每股 230 美元左右； - **企业市值**：公司市值稳稳迈入 500 亿至 600 亿美元的超大体量区间。 谈及外界赞叹的“精准时机把握”，Andrew 幽默地反驳道：“很多人问我们怎么把上市时间点掐得这么准，答案其实是：我们足足搞错了十年。”他分享了敲钟当天最触动人心的细节：公司邀请了所有工龄超过 9 年的老员工及其家属前往交易所。Andrew 惊讶地发现，这群平时不修边幅的硬件工程师不仅真的拥有领带，戴上领带后也并未感到不适；更重要的是，这一刻让员工的家人们感受到了如成人礼（Bar Mitzvah 或 Quinceañera）般的巨大荣耀。一位核心高管的华裔移民父亲甚至在现场幽默地感叹道：“我还以为这事儿本该发生得更快一些呢！”这种来自外部世界的正向认可极大地振奋了团队，但庆祝过后，这群骨子里热衷于解决极限硬核难题的工程师们只留下了一句“现在该干嘛？回去干活吧”，旋即重返研发战壕。 ### [08:02](https://www.youtube.com/watch?v=jLICvWE7w2Q&t=482s) Planet Labs 的破局：从冷遇、10倍暴涨到国家级信任背书 随后，聚光灯转向了走过不同公开市场轨迹的 Planet Labs。与 Cerebras 甫一上市便炙手可热不同，Planet Labs 在上市初期遭遇了公开市场长达数年的认知盲区。Will Marshall 回忆道，作为市场上最早一批纯正的商业航天概念股，当时的华尔街根本无法理解太空中正在发生的技术剧变，普遍带着“这到底是个什么玩意儿”的困惑与冷淡。然而在过去的 12 个月里，市场认知发生根本性逆转，Planet Labs 的股价从每股 5 美元飙升至 50 美元，实现了惊人的 10 倍涨幅。 Will 指出，上市固然为早期员工和天使投资人带来了流动性，并为公司资产负债表补充了现金，但对深科技企业而言，成为公众公司最无可替代的红利在于**构建跨国级与行业级的顶级公信力**。Planet Labs 的核心客群包括全球跨国农业巨头、多国民事政府部门以及顶级国防与情报机构，甚至有主权国家在关键决策上完全依赖其提供的遥感与卫星情报： - **业务存续保障**：核心政企客户最忌讳供应商中途倒闭或技术断档，上市公司的公开财务透明度向市场证明了企业“有资本支撑、将长期存在”； - **摆脱短期股价波动**：Will 强调团队并不关心每日的股价涨跌，而是聚焦于长期的股东价值创造； - **卫星与太空的无形渗透**：公众往往忽视了太空数据早已无处不在——人们每一次使用智能手机通信、依赖 GPS 导航定位或查看地图，底层都紧密交织着卫星系统与空间数据的支持。随着 SpaceX 等标志性企业的催化，资本市场终于开始深刻意识到太空正在重塑整个人类社会的运行底层。 ### [10:18](https://www.youtube.com/watch?v=jLICvWE7w2Q&t=618s) 地球观测与军工争议：Planet Labs 的双重驱动力 Planet Labs 首席执行官 Will Marshall 阐述了公司的核心定位与业务实质。针对外界是否会因 Anduril 等军工科技企业的崛起而将其归为“军工股”的提问，Marshall 指出这种定性需要更加细致的辨析。Planet Labs 目前运营着全球规模最大的地球观测卫星群，拥有约 200 颗在轨卫星，能够实现每天对整个地球表面进行全覆盖成像。这相当于 Google 地图上的卫星图层，但彻底告别了以往三至五年才更新一次的滞后性，变成了“今日最新画面”，并且积累了完整的历史时间序列。这种持续监控能力不仅为农业估产、能源基础设施监测、市政防汛防火提供了底层数据支撑，也自然而然地延伸到了国家安全领域。Marshall 强调，AI 正在显著降低各行各业利用遥感数据的技术门槛，但其核心价值始终取决于训练数据的质量。 面对主持团队对军工业务营收占比的直接追问，Marshall 坦言，虽然安全业务从创立之初就是规划的一部分，但在当前地缘政治剧烈震荡的背景下，军工安全领域的实际需求确实超出了最初的预期。军方与情报机构极度依赖这种宏观动态感知来“洞察转角处的潜在威胁”，从而获得数周甚至数月的早期预警，争取通过威慑或外交手段制止冲突爆发。当被问及是否介意被贴上“军用科技公司”标签时，Marshall 表示并不排斥，但坚决认为 Planet 的格局远不止于此——客户群体不仅包括五角大楼，还涵盖 NASA、各国市政机构以及大量商业企业。 谈及航天底层生态的巨变，Marshall 指出外界往往只关注到过去十年火箭发射成本降低了四到五倍，却忽略了一个更为关键的飞跃：卫星的极端微型化。传统间谍卫星动辄耗资 10 亿美元、重达 20 吨，而如今功能相当甚至更强的卫星已被压缩至几公斤、几十公斤或数百公斤级别。这一演进完美复刻了昔日计算机从庞大笨重、仅供少数人使用的“大型主机”（Mainframe）向个人“台式电脑”（Desktop PC）的普及革命，软硬件协同降本正成倍释放此前无法想象的全新商业应用。 - **卫星在轨规模**：约 200 颗卫星组成全球最大商业地球成像网络，实现全球地表每日无死角扫描。 - **降本驱动双引擎**：过去十年火箭发射成本下降约 4 至 5 倍；卫星重量从数十吨骤降至几十至上百公斤，造价缩减数个数量级。 --- ### [13:18](https://www.youtube.com/watch?v=jLICvWE7w2Q&t=798s) 太空数据中心：能源套利与星际算力时代 随着主持团队将话题引入“将地面数据处理基础设施重构至天空”的宏大叙事，Marshall 详细拆解了空间计算的商业前景与太空数据中心（Space-based Datacenters）的经济逻辑。短期内，仅结合大语言模型与实时地球图像的应用生态（覆盖农业、能源审计、建设许可等），就将开拓出 750 亿至 1000 亿美元的地球观测新市场。而放眼长远，算力上天正从科幻变为必然。Planet Labs 早在八九年前就曾与 Google 团队联合进行可行性测算，对比地面与轨道的建设及运营全生命周期成本。 测算模型显示，只要火箭发射成本下探至每公斤 200 至 300 美元区间，在太空部署数据中心在经济上就将彻底击败地面。目前商业航天的平均发射成本在每公斤 1000 美元左右（过去十年已下降十倍），随着 SpaceX 星舰（Starship）的技术推进，这一经济临界点预计在未来两到三年内就会实现。更重要的是太空无可比拟的能源与系统优势：地面数据中心本质上受制于电力供应，依赖太阳能面临间歇性发电痛点，搭配昂贵的储能电池、燃气电站或审批严苛的核电将导致成本飙升。反观太空，数据中心卫星可以部署在晨昏太阳同步轨道（Sun-synchronous dawn-dusk orbit），实现“365天、每天 24 小时无遮挡直面太阳”，单块太阳能板的能量捕获效率是地面的五倍以上，且无需任何储能电池。 ```mermaid flowchart TD accTitle: 太空数据中心与地面数据中心架构及成本对比 accDescr: 对比地面与太空数据中心在能源获取、发射临界成本以及核心基础设施构成上的差异 subgraph Ground ["地面数据中心瓶颈"] G1["受制于电网与选址容量"] --> G2["太阳能间歇性发电"] G2 --> G3["需高昂储能电池 / 燃气 / 核电调峰"] end subgraph Space ["太空数据中心极简架构"] S1["发射成本临界点: $200-$300 / kg"] --> S2["晨昏太阳同步轨道部署"] S2 --> S3["24/7 无间断直面阳光"] S3 --> S4["单板发电效率为地面 5 倍且零电池缓冲"] S4 --> S5["极简硬件: 太阳能板 + 芯片 + 射频传输"] end ``` 太空数据中心的物理架构异常纯粹，剔除了冷却塔、复杂的备用发电机与储能电网，精简至“太阳能板 + 算力芯片 + 射频通信上行与下行链路”。Marshall 透露，Planet 已经在轨道上测试运行了部分英伟达 GPU，目前正在携手 Google 将其 TPU 芯片送入太空进行早期技术验证。他大胆断言，尽管工程上仍有大量细节需要攻克，但在未来十年内，全球绝大部分新增算力设施都将被推向太空，催生一个规模超越当下所有航天细分赛道、高达数万亿美元的超级产业。 - **发射成本临界值**：当前发射成本刚跨过约 $1000/kg，距离 $200–$300/kg 的经济爆发点预计仅剩 2 至 3 年。 - **能源收集效率**：在晨昏太阳同步轨道下，无日夜更迭与云层遮挡，单位太阳能板发电量达到地面的 5 倍，摆脱对化学储能的依赖。 - **测试落地进程**：已在轨运行英伟达 GPU，正与 Google 合作发射 TPU 进行天基集群计算验证。 --- ### [17:58](https://www.youtube.com/watch?v=jLICvWE7w2Q&t=1078s) 硬件极限与现实阻力：从太空互联困境到 AI 芯片的新范式 面对太空算力狂想，Cerebras 创始人兼首席执行官 Andrew Feldman 表达了务实而清醒的技术质疑。Feldman 坦承，除了将裸芯片发射升空这一简单动作之外，太空数据中心还面临着一两个极其致命且尚未解决的硬核工程障碍，其中最核心的就是超大规模集群互联通信。他直言不讳地指出：“我们在地面上搭建超大规模计算集群都还搞得一头雾水，到了太空环境更是难上加难。” Feldman 将太空算力的工程成熟度类比为自动驾驶技术的发展路径——往往“最后的 10% 难题却需要耗费整整 80% 的研发时间”。自动驾驶行业曾让全行业深陷长达十年的拉锯战，直到近期才算勉强翻越山头，太空计算是否也会深陷类似的“长尾陷阱”仍是未知数。不过他也承认，Planet 正在推进的先锋探索意义重大，而火箭发射成本的指数级下降正是关键杠杆，它赋予了工程师低成本试错、在太空中交学费并快速迭代硬件的底气。 随后，对话切入 Feldman 的核心战场——地面芯片与硅基算力架构的范式颠覆。被问及 Cerebras 颠覆传统芯片规则（制造出整整一张披萨大小的晶圆级芯片 WSE）背后的底层考量时，Feldman 回顾了 2015 年前后 AI 带来的算力历史性拐点。他指出，计算机自诞生以来的绝大多数岁月里，只在“数字运算与转换”上展现出神迹，但在处理图像和自然语言时却极度笨拙——几十年来人类计算机对图像和文字除了简单的存取和搬运之外几乎束手无策。而以深度学习为代表的 AI 浪潮，不仅在底层重新定义了计算机能够解决的问题边界，更彻底动摇了传统由晶体管密度主导的摩尔定律，倒逼专用领域架构（Domain-Specific Architectures）在硅片设计与封装形式上掀起一场脱胎换骨的全新革命。 > “我们在地面上搭建超大规模计算集群都还搞得一头雾水，到了太空环境更是难上加难。” —— Andrew Feldman ### [20:22](https://www.youtube.com/watch?v=jLICvWE7w2Q&t=1222s) 算力革命与晶圆级芯片：Cerebras 为何放弃 GPU 架构 Andrew Feldman 回顾了人工智能在 2015 至 2016 年间带来的技术拐点。彼时，计算机的视野被彻底打开，计算系统不再局限于对文本进行简单的存储与机械复述，而是开始具备理解语言、自主生成文本以及深入解析图像内容的能力。与此同时，来自地面与太空卫星的图像数据量呈爆发式增长，全新的非结构化数据领域向计算工具敞开大门，直接推动了 Nvidia 等硬件厂商的爆炸式增长。作为芯片架构师，Feldman 在 2015 年便敏锐地意识到：AI 将成为计算资源前所未有的超级消耗者。 从计算机体系结构的历史演进来看，每当全新的计算负载出现时，行业格局与市场份额就会发生颠覆性洗牌。当年 PC 图形处理需求的爆发催生了专用 GPU，从而造就了 Nvidia；手机计算的兴起让手握晶圆厂与顶尖架构师的 Intel 和 AMD 几乎丢失了全部移动端份额，成就了 ARM 架构；90 年代末数据网络爆发时，未能跟上步伐的 Nortel 等老牌巨头迅速陨落，Cisco、Juniper 和 Arista 等新势力借机崛起。面对 AI 浪潮，Cerebras 在创立之初便下了两项坚定的赌注：第一，AI 必须依靠专用定制芯片（ASIC）；第二，这颗芯片绝不能走 GPU 的老路。Feldman 直言，如果你想在性能上实现对对手 20 倍的飞跃，就不能在架构层面模仿对方，因为 Nvidia 已经把 GPU 架构下的低垂果实采摘殆尽，"如果你制造的是 GPU，在我们看来，击败 Nvidia 的概率几乎为零。" 基于此，Cerebras 颠覆了行业传统的芯片堆叠思路，专注于解决 AI 计算最底层的阿喀琉斯之踵——数据在内存与计算核心之间的传输瓶颈（内存墙问题）。主流芯片通常只有邮票大小，需要依赖片外高带宽内存；而 Cerebras 选择制造一块面积与餐盘相当的单晶圆级巨型芯片（Wafer-Scale Engine），将内存直接放置在计算单元身旁。这种架构引入了速度远超普通显存的超高速片上内存，在 OpenAI 等实际运行场景中，其推理速度比传统 GPU 快了 15 至 18 倍。Feldman 打了个形象的比方：正如今天拨号上网与慢速搜索引擎的市场份额为零、用户等待网页加载超过 3 到 5 秒就会关掉窗口一样，没有人会为缓慢的 AI 响应买单，实时交付是 AI 的唯一生存之道。 ```mermaid flowchart TD accTitle: GPU与Cerebras芯片架构及数据传输瓶颈对比 accDescr: 对比传统GPU片外显存传输瓶颈与Cerebras晶圆级芯片将超高速内存紧挨计算核心的设计差异。 subgraph GPU["传统 GPU 架构：分离式瓶颈"] A["计算核心 (邮票大小芯片)"] B["片外高带宽内存 (HBM)"] A <-->|"慢速片外总线传输 (内存墙瓶颈)"| B end subgraph Cerebras["Cerebras 晶圆级架构：近存计算"] C["计算核心 (餐盘大小晶圆级芯片)"] D["片上极速内存 (SRAM)"] C <-->|"内部近距离极速互连 (15-18倍加速)"| D end ``` - **历史变革类比**：图形时代成就 Nvidia，移动端成就 ARM 并击败 x86，网络数据时代成就 Cisco、Juniper 与 Arista。 - **性能实测差距**：在针对 OpenAI 等头部模型的测试中，Cerebras 硬件的响应与交付速度达到传统 GPU 的 15 到 18 倍。 - **用户体验逻辑**：AI 推理不存在"慢速市场"，低延迟是满足高复杂度推理和即时人机交互的核心前提。 --- ### [24:47](https://www.youtube.com/watch?v=jLICvWE7w2Q&t=1487s) 早期投资人与流动性博弈：Planet Labs 与创新解禁机制 圆桌话题随后转向资本市场与流动性退出。Brad Gerstner 指出在场的许多有限合伙人（LP）都在急切寻求流动性，并询问 Planet Labs 早期风投的持股情况。Will Marshall 介绍，Planet Labs 于 2021 年通过 SPAC 方式上市，估值为 20 亿美元，其早期风投阵容包括 Steve Jurvetson（原 DFJ）、Capricorn Investment Group、Peter Thiel 的 Founders Fund 以及 Yuri Milner 旗下的 DST Global。引人深思的是，Planet Labs 上市四年以来，约 90% 的企业价值实际上是在上市后第三至第四年才被创造出来的，而大多数早期风投选择长期守仓，从而完整吃到了这一波丰厚红利。 Marshall 补充道，Google 作为 Planet Labs 最大的单一股东至今未卖出一股，Capricorn 也是直到最近才开始部分套现，事实证明这种坚守极其明智。Gerstner 借此痛陈了风险投资行业的通病：许多 LP 往往在公司刚上市时就急迫要求分配股票（Distribution），而这往往会错失最丰厚的主升浪。例如 Altimeter 在十年前曾以 10 亿美元估值投资某家 Pre-IPO 独角兽，随后在 30 亿至 40 亿美元估值上市时顺应 LP 的呼声将股票分出，结果该标的在接下来的 24 个月内狂飙至 500 亿美元市值，LP 们反过来抱怨基金为何未能继续持有。 面对解禁期引发的市场抛售与流动性管理难题，Andrew Feldman 透露 Cerebras 与华尔街投行共同设计了一套创新的"渐进式分批解禁机制"（Dribble Lockup）。与以往六个月锁定期一到便形成集中抛压的模式不同，该机制设定了长达六个月的周期，并挂钩一系列严格的业绩与股价触发门槛，允许股份根据经营指标逐步释放。Gerstner 评价这一解禁创新极具建设性，并预测后续备受瞩目的 SpaceX 上市很可能会借鉴这一架构。 - **Planet Labs 上市路径**：2021 年通过 SPAC 上市，初始估值约 20 亿美元；90% 的市值增量在上市后第 3 至 4 年实现。 - **战略投资人定力**：Google 作为最大股东上市以来一股未抛；早期风投机构如 Capricorn 也保持了长期持股。 - **Cerebras 解禁创新**：与投资银行合作设计基于业绩目标（Performance Hurdles）的渐进式解禁机制（Dribble Lockup），平滑市场抛压。 --- ### [27:39](https://www.youtube.com/watch?v=jLICvWE7w2Q&t=1659s) 走出“长期私有化”迷思：公开市场锤炼与价值重构 在探讨 VC 机构是否应在解禁期满后继续持股时，Andrew Feldman 明确指出，历史统计数据已经反复证实了一个客观规律：企业在 IPO 之后创造的财富总量，无论是在涨幅百分比上还是在实际绝对金额上，都远远超过了 IPO 前的私募阶段。究其原因，除了极个别极端异常值外，大部分初创企业在未上市前能够容纳的资本体量十分有限，而在公司完成公开上市、商业模式被全面验证后，公开市场才能支撑起数额极其庞大的资金池，带来更具爆发力的资本复利。 Will Marshall 随即抛出了对当前万亿级巨头上市的深刻质问。过去大部分老牌科技巨头都是以数亿到数十亿美元的体量上市，公开市场投资者享受了从几十亿跨越到数万亿美元的完整红利；而如果像 SpaceX、OpenAI 这样已在私募市场膨胀至数千亿美元甚至数万亿美元的超级巨头上市，如果要复制过往科技股百倍千倍的财富效应，其市值目标难道要瞄准"千万亿"（Quadrillion）级别？这无疑对二级市场承接力提出了极其苛刻的假设。 ```mermaid flowchart LR accTitle: 两种上市路径与价值分配模式对比 accDescr: 展示以Planet Labs为代表的早期上市路径与以万亿超级巨头为代表的长期私有化路径在价值分配上的差异。 subgraph EarlyIPO["健康模式：早期公开上市 (如 Planet Labs)"] direction TB E1["10-20亿美元小体量上市"] --> E2["公开市场承接5-10倍成长"] E2 --> E3["散户与二级市场获得主要红利"] end subgraph LateIPO["晚期模式：长期维持私有 (过去十年常见)"] direction TB L1["在私募市场膨胀至数千亿美元"] --> L2["绝大部分溢价留给私募风投"] L2 --> L3["公开市场缺乏后续倍数空间"] end ``` 针对这一分歧，Gerstner 旗帜鲜明地批评了过去十年由 Andreessen Horowitz 等硅谷风投极力倡导的"永久维持私有化"（Stay Private Forever）理念。他认为资本钟摆正在急剧回调，像 Planet Labs 这样较早进入二级市场、在公开市场完成 10 倍跃升的范例正在被重新认可。如果 Cerebras 能在 18 个月前以 100 亿美元而不是后来的 500 亿美元上市，这期间的 5 倍增长就能回馈给二级市场大众。Chamath Palihapitiya 最终总结指出，过度的私募保护容易滋生怠惰，"铁磨铁，刃更锋"（Iron sharpens iron），更早将企业推上公开市场的聚光灯下，接受严谨的财报审查、经营问责与公众监督，非但不会扼杀企业，反而能极大淬炼核心团队的战略专注度，激发更具实效的突破性创新。 - **上市前后收益对比**：绝大多数商业案例证明，IPO 之后的绝对收益体量显著高于私募阶段早期投资。 - **估值天花板挑战**：若万亿体量的独角兽在私募市场耗尽所有溢价，公开市场投资者将很难再获取数十倍的超额回报。 - **公开市场的淬炼效应**：上市带来的透明度和业绩压力如"铁磨铁，刃更锋"，倒逼管理层聚焦主业，加速底层技术创新。 ### [31:06](https://www.youtube.com/watch?v=jLICvWE7w2Q&t=1866s) 行星智能与大地球模型：AI 与太空融合的未来展望 在对谈进入尾声之际，主持人向两位嘉宾致谢，并感叹接下来的几年科技行业将会变得“极其精彩刺激（really spicy）”。Planet Labs 首席执行官 Will Marshall 随即借此机会要了三十秒时间，分享了他对未来几年的愿景与前瞻。他认为，最令人兴奋的变革将来自于 AI 与太空两大领域的深度融合，这种融合将催生出海量前所未有的应用爆发。 Will Marshall 指出，当前大语言模型（LLM）所展现出的各种酷炫能力，本质上都建立在对全网文本数据的吞吐与吸收之上。尽管这已经展现出惊人的力量，但现有的模型对真实的物理世界仍然处于“盲目（blind）”状态——它们并不知道具体农田的作物长势、突发的洪水灾情，或是转角处正在发生的安全局势。如果能将真实物理世界的感知数据灌注给这些模型，AI 才能真正开始解决现实世界的问题。因此，未来的范式转变将不再局限于处理文本的大语言模型，而是演变为“大地球模型（Large Earth Models）”；人们追求的也不再仅仅是常规意义上的 AI，而是“行星智能（Planetary Intelligence）”。 为了构建这一愿景，人类需要将在轨的行星级遥感监测系统与部署在太空的行星级算力系统紧密结合。Will Marshall 坦言，虽然大家对于技术全面落地的具体时间表可能存在分歧，但这无疑是必然会发生的大趋势，并将由此催生出一个规模极其庞大的太空经济体。随着全场观众的热烈掌声，主持人和全场嘉宾互道感谢与祝贺，圆满结束了本次圆桌讨论。 ```mermaid flowchart TD accTitle: 从大语言模型到行星智能的范式演进 accDescr: 对比传统 LLM 基于互联网文本数据的局限性与结合太空遥感算力的行星智能架构。 subgraph Current["当前范式：大语言模型 (LLM)"] A["互联网文本数据"] --> B["LLM 预训练模型"] B --> C["局限：对真实世界盲目<br/>(未知灾害/农田/安全局势)"] end subgraph Future["未来范式：行星智能 (Planetary Intelligence)"] D["太空行星级感知系统<br/>(实时地球观测)"] --> E["大地球模型<br/>(Large Earth Models)"] F["太空行星级计算系统<br/>(在轨处理与推断)"] --> E E --> G["解决现实物理世界问题<br/>(催生万亿级太空经济)"] end C -. 范式演进.-> E ``` - **从 LLM 到 LEM**：跳脱单纯依赖互联网文本的局限，打造能够实时理解物理世界的“大地球模型（Large Earth Models）”。 - **行星智能（Planetary Intelligence）架构**：由部署在太空轨道的行星级感知网络（Sensing Systems）与行星级算力网络（Compute Systems）共同支撑。 - **现实应用爆发**：赋予 AI 对农田状况、洪涝灾害、地缘安全等现实物理环境的实时洞察能力，从而打开无数高价值的商业与公共治理应用场景。 ## Notable quotes > [00:00](https://www.youtube.com/watch?v=jLICvWE7w2Q&t=0s) fastest bar none — Andrew Feldman > [04:39](https://www.youtube.com/watch?v=jLICvWE7w2Q&t=279s) 长达九年半的时间里，每一天都步履维艰，直到最后的十二个月，势头突然变得无比顺畅，所有资本都争先恐后想要挤上车。 — Brad Gerstner > [04:39](https://www.youtube.com/watch?v=jLICvWE7w2Q&t=279s) 很多人问我们怎么把上市时间点掐得这么准，答案其实是：我们足足搞错了十年。 — Andrew Feldman > [04:39](https://www.youtube.com/watch?v=jLICvWE7w2Q&t=279s) 我还以为这事儿本该发生得更快一些呢！ — 核心高管家属 > [04:39](https://www.youtube.com/watch?v=jLICvWE7w2Q&t=279s) 现在该干嘛？回去干活吧 — Cerebras 工程师团队 > [10:18](https://www.youtube.com/watch?v=jLICvWE7w2Q&t=618s) 洞察转角处的潜在威胁 — Will Marshall > [17:58](https://www.youtube.com/watch?v=jLICvWE7w2Q&t=1078s) 我们在地面上搭建超大规模计算集群都还搞得一头雾水，到了太空环境更是难上加难。 — Andrew Feldman > [17:58](https://www.youtube.com/watch?v=jLICvWE7w2Q&t=1078s) 最后的 10% 难题却需要耗费整整 80% 的研发时间 — Andrew Feldman > [20:22](https://www.youtube.com/watch?v=jLICvWE7w2Q&t=1222s) 如果你制造的是 GPU，在我们看来，击败 Nvidia 的概率几乎为零。 — Andrew Feldman > [27:39](https://www.youtube.com/watch?v=jLICvWE7w2Q&t=1659s) 铁磨铁，刃更锋 — Chamath Palihapitiya > [31:06](https://www.youtube.com/watch?v=jLICvWE7w2Q&t=1866s) really spicy — Jason Calacanis ## Mentioned Andrew Feldman, Will Marshall, Brad Gerstner, Jason Calacanis, David Sacks, Chamath Palihapitiya, Tucker Carlson, Donald J. Trump, Cerebras Systems, Planet Labs, Altimeter Capital, All-In Podcast, CFIUS, Anduril, Nvidia, Google, SpaceX, NASA, Intel, AMD, ARM, Cisco, Juniper Networks, Arista Networks, Nortel, OpenAI, Steve Jurvetson, Capricorn Investment Group, Founders Fund, Peter Thiel, DST Global, Yuri Milner, Andreessen Horowitz, Wafer-Scale Engine (WSE), Starship, TPU, GPU, Large Earth Models (LEM), Planetary Intelligence ## Video description (0:00) CEOs Andrew Feldman (Cerebras) and Will Marshall (Planet Labs) join the Besties! (2:05) Both CEOs on going public: Impact on employees, customers, and business operations (13:18) Timelines for datacenters in space (19:28) Cerebras business breakdown, AI's impact on the silicon market (24:45) How Founder/CEOs think about liquidity on the road to going public Thanks to our partners for making this possible! EY - Great tech starts with a big idea. From startup to scale, EY helps tech founders get financials right early so they can focus on what’s next. https://www.ey.com/en_us/tech-sector/tech-startups?WT.mc_id=3501317&AA.tsrc=sponsorship NYSE - Thank you to our partner, the New York Stock Exchange - a modern marketplace and exchange for building the future. It all happens at the NYSE. https://www.nyse.com Plaud - Never miss a moment. Plaud, our official wearable AI note-taking partner at All-In Liquidity Summit, captured every insight. https://www.plaud.ai Follow Brad Gerstner: https://x.com/altcap Follow Andrew Feldman: https://x.com/andrewdfeldman Follow Will Marshall: https://x.com/Will4Planet Apply for Summit 2026: ⁠https://allin.com/events⁠ Follow the besties: ⁠https://x.com/chamath⁠ ⁠https://x.com/Jason⁠ ⁠https://x.com/DavidSacks⁠ ⁠https://x.com/friedberg⁠ Follow on X: ⁠https://x.com/theallinpod⁠ Follow on Instagram: ⁠https://www.instagram.com/theallinpod⁠ Follow on TikTok: ⁠https://www.tiktok.com/@theallinpod⁠ Follow on LinkedIn: ⁠https://www.linkedin.com/company/allinpod⁠ Intro Music Credit: ⁠https://rb.gy/tppkzl⁠ ⁠https://x.com/yung_spielburg #allin #tech #news [Source](https://www.youtube.com/watch?v=jLICvWE7w2Q)

[0:00] Andrew Feldman: fastest bar none [4:39] Brad Gerstner: 长达九年半的时间里，每一天都步履维艰，直到最后的十二个月，势头突然变得无比顺畅，所有资本都争先恐后想要挤上车。 [4:39] Andrew Feldman: 很多人问我们怎么把上市时间点掐得这么准，答案其实是：我们足足搞错了十年。 [4:39] 核心高管家属: 我还以为这事儿本该发生得更快一些呢！ [4:39] Cerebras 工程师团队: 现在该干嘛？回去干活吧 [10:18] Will Marshall: 洞察转角处的潜在威胁 [17:58] Andrew Feldman: 我们在地面上搭建超大规模计算集群都还搞得一头雾水，到了太空环境更是难上加难。 [17:58] Andrew Feldman: 最后的 10% 难题却需要耗费整整 80% 的研发时间 [20:22] Andrew Feldman: 如果你制造的是 GPU，在我们看来，击败 Nvidia 的概率几乎为零。 [27:39] Chamath Palihapitiya: 铁磨铁，刃更锋 [31:06] Jason Calacanis: really spicy

[2:16] Andrew Feldman: not a damn thing changes in the important parts of your business [2:16] Andrew Feldman: by getting it wrong for a decade. [2:16] Andrew Feldman: engineers owned ties [8:01] Will Marshall: indexing the earth and making it searchable [8:01] Will Marshall: here to stay. [10:07] Will Marshall: see threats around the corner [16:57] Will Marshall: within 10 years most compute will be putting in space. [18:08] Andrew Feldman: one or two hard problems still left [18:08] Andrew Feldman: We're not good at doing it on the ground. We're really not good at doing it in space. [21:31] Andrew Feldman: it couldn't look like a GPU. [21:31] Andrew Feldman: eaten all the lowhanging fruit. [22:54] Andrew Feldman: 15 or 18 times faster than a GPU. [22:54] Andrew Feldman: How big is the market for slow search today? Right. It's zero. [26:39] Andrew Feldman: historically more money's made after IPO than before [29:33] stay private forever [29:33] sharpens the focus [29:33] iron sharpens iron. [31:11] Will Marshall: blind [31:11] Will Marshall: Large Earth Models [31:11] Will Marshall: planetary intelligence

[0:00] Hey, 2026 could be an all-time [music] [0:03] record for IPOs. [0:04] >> The AI IPO of the year so far. That [0:08] company is Cerebras. [0:09] >> Cerebras Systems founder and CEO Andrew [0:11] Bubman. We are participating [music] in [0:13] something extraordinary on everything we [0:15] do. We are the fastest bar none. [0:17] >> Well, Marshall is the co-founder and CEO [0:19] of Planet [music] Labs. [0:21] >> Space and AI are really um a match made [0:24] in heaven. They're getting married. In [0:25] fact, just like Google figured out how [0:28] to index the internet and make it [0:29] searchable, we are indexing the earth [0:32] and making it searchable. [music] He's [0:33] got his glasses, the famous red glasses. [0:35] Brad Gersner is here, founder and CEO of [0:37] Autoimmet [music] Capital, a leading [0:39] tech investment firm. [0:40] >> I believe that the wave is the biggest [0:42] wave in the history of technology, [0:43] [music] will be incredibly beneficial [0:45] for America. I'm rooting for all of them [0:47] because I'm rooting for America. Ladies [0:50] and gentlemen, please welcome Brad [0:53] Gersonner, Will Marshall, and Andrew [0:56] Feldman. [1:00] [music] [1:01] >> On the couch, we switch on the couch. [1:05] >> Nice to see you, my friend. [1:07] >> Hey, [music] big boy. [1:09] >> Nice to see you. [1:09] >> Last last time I saw you, we were in [1:11] Davos. [1:12] >> Yes, [1:12] >> we were in Davos causing another drop. [1:15] Another JL. Do you hear that little [1:16] Davos? [1:17] >> We were just, you know, it was preipo. [1:19] We're chopping it up [1:20] >> with Davos. [1:21] >> We're in Davos. [1:22] >> Hanging out at Davos. [1:23] >> Well, no. Listen, I was Everybody knows [1:25] the story. I'm supposed to go on my [1:27] yearly Japan ski trip. Sax calls me [1:29] >> with Tucker. [1:30] >> Yeah. Well, anyway, we don't drop that [1:32] name, but [laughter] I'll pick it up for [1:34] you. Put it over here. [1:35] >> Tucker. Anyway, so I cancel on Tucker. I [1:37] cancel because Sax calls me. He says, [1:39] "Listen, Pus needs you, the world's [1:41] greatest moderator in Davos." I said, [1:43] "No problem." [1:44] >> I said, "Saxs, POTUS, and Davos." So I [1:48] said, "When?" He says, "In 3 days." I [1:51] say, "You got it." I go and they give me [1:53] a badge. And it's like the special green [1:55] badge and they buzz you through the [1:58] security and I look at the monitor and [2:00] it says Jason McCabe Calacanis with [2:03] Donald J. Trump. [2:05] >> Oh wow. How did you feel? [2:08] >> I thought it was hilarious. [laughter] [2:10] So then I went and we did a great [2:12] interview there and we did like six or [2:14] seven of these great all-in interviews [2:16] and it was fun. [2:16] >> Let's start this because uh the two of [2:19] you guys run two of the most interesting [2:22] and consequential newly public companies [2:24] in the stock market. Andrew Feland is [2:26] the founder and CEO of Cerebrus. Will [2:27] Marshall is the founder and CEO of [2:28] Planet Labs. But you are also the [2:30] insight and a gateway for all of us to [2:32] understand these two big trends. One is [2:35] in AI silicon, the other one is in space [2:38] data centers. I think it would be a [2:39] really interesting thing to [2:41] >> and emerging. [2:42] >> And emerging. Yeah. Um but let's just [2:45] take one step back. Uh you just heard [2:48] the last conversation about being [2:50] public, going public early. Let's just [2:52] talk about that cuz I'm just very [2:53] curious. How's it been? It's been 3 [2:55] weeks or so for you. It's been about a [2:57] year and a half or two years for you. Uh [3:00] >> it's more fresh. [3:01] >> Was it Was it everything that you [3:02] thought it would be? Like [3:04] >> what's clear so far is I need to upgrade [3:07] my namerop game. I mean that that was a [3:09] tour to force. [laughter] [3:11] >> But by the way, you were you were in [3:13] Davos with [3:13] >> J. I was I was there but that um uh [3:17] >> tour [laughter] to force [3:20] uh look I I I think you do all this work [3:23] and I I think it's really difficult to [3:26] to overestimate the amount of garbage [3:28] that's involved in in going public. the [3:31] number of meetings where you you look on [3:35] the the Zoom and there are 130 attendees [3:38] and the amount of times you review these [3:41] documents and the commas move and and [3:43] just no value added. You go there and [3:47] you have this enormous event and the [3:51] next morning you've sold no more stuff. [3:53] your engineering projects have made no [3:55] progress since the day you weren't [3:57] public. And you go back to work and um [4:02] you you have some new constituents that [4:04] that you have to to to address and [4:07] communicate with, but the core parts of [4:10] your business, you have more money in [4:11] the bank. Um but not a damn thing [4:15] changes in the important parts of your [4:18] business. um if you still if you need [4:21] new supply or if your relationships with [4:23] your vendors are bad, they're still bad. [4:25] If they're good, they're still good. And [4:27] and so I I think what what we've seen is [4:31] um you your employees have a party, [4:34] everybody's really excited, and you put [4:36] your head back down, you high five, and [4:38] you go back to work. Can [4:39] >> can I can I just give a little context [4:41] and then I want to hear from Will. You [4:43] know, if if I can, Andrew, you know, we [4:45] were investors in Cerebrus. I was on the [4:47] board a year earlier where we were [4:49] trying to go public. Um, and you know, [4:53] aside from just being a warrior who [4:55] weathered a decade worth of storms that [4:58] would have taken out any normal human [5:00] being, the path to going public for [5:03] Cerebras was a particularly challenging [5:05] one. One of their investors was the UAE. [5:08] So there was questions about CPHAS, you [5:12] know, in in the prior under the Biden [5:14] administration challenging to get [5:15] public. My observation outside looking [5:18] in is everything was really hard until [5:23] it got really easy like n 9 and a half [5:26] years of really hard and then 12 months [5:29] you know of of of really easy where [5:31] everybody wanted to get in. [5:33] >> They priced the IPO at 185 which was up [5:36] the range was taken up two times. Okay. [5:39] The stock opened at $320 a share I [5:43] think. Today it's at 230 bucks a share, [5:46] 5060 billion dollars in market cap for a [5:49] business like you know and andrew is [5:52] just one of these people. Let's let's [5:53] get back to work and build But my [5:55] my just add-on question to that is from [5:58] an employee morale perspective like [6:00] distraction perspective etc. has the [6:04] last 3 weeks you got a lot more capital [6:06] you got a lot more profile presumably [6:08] it's easier to sell to enterprise [6:09] customers today. net net if you were [6:12] advising me if I was in a similar [6:14] position would you say go public? [6:17] >> I I I think the first thing is a lot of [6:19] people asked us about how we got the [6:22] timing right. [6:23] >> Right. [6:23] >> And I I think the answer is by getting [6:25] it wrong for a decade. I mean that's [6:27] really the right way to get timing [6:29] right. Um I I think um first I we we [6:34] we've been at this for for more than a [6:36] decade and and we we brought everybody [6:39] who'd been with the company more than [6:41] nine years to share and we brought their [6:42] families. [snorts] And first I learned [6:45] that engineers owned ties. I didn't [6:46] actually know that. [laughter] Um and [6:49] they didn't die when they wore them. And [6:51] second I was surprised at how big a deal [6:54] it was for them and their family. they [6:56] they were really proud in a way that [6:59] sort of their parents [clears throat] [7:00] might have heard of it or uh that that [7:03] somehow this was like a a bar mitzvah or [7:06] kinera or something. [7:08] >> Um [7:09] >> and then you you had these sort of the [7:11] the children of immigrants, one of our [7:14] uh one of our leaders, her father, [7:17] Chinese immigrants said, "I thought it [7:19] would have happened faster." [7:22] [laughter] [7:23] >> Right. And but I I think um we are sort [7:29] of by nature [7:31] uh [7:33] sort of in the uh in the trenches [7:37] people. And so um we we love solving [7:41] hard problems. And so when when we had [7:43] this excitement, everybody went and they [7:45] were so excited and we had a party and [7:47] um I I think it it gave external [7:50] validation. And then everybody turned [7:52] around and said, "Now what are we now [7:54] back to work?" And and so I I I think um [7:59] >> and so you you started off kind of bang [8:02] right out of the gates. Will you had a [8:04] little bit different experience in terms [8:05] of you know the entry to the public [8:08] markets, but over the last 12 months [8:10] your stock has gone from five bucks a [8:13] share to 50 bucks a share, some 10x move [8:15] in the public markets. So talk us [8:17] through the other side of this where you [8:19] come public, nobody really notices until [8:21] they notice. [8:22] >> Well, we were one of the first space [8:24] stocks and and I think people just had [8:26] no idea what earth is going on in space, [8:29] how it was changing everything and they [8:31] were just like what the heck is that? [8:33] And and but but you know, I have similar [8:36] opinions. I mean, in the end, you've [8:37] just got to get on with executing the [8:39] business. uh the going public gives you [8:41] access to liquidity uh for early [8:44] shareholders whether that's the employ [8:46] early employees or early investors and [8:48] that's great it gives you cash for the [8:51] company that's great and I do think um [8:54] it helps your business as well because [8:57] the maturing event gives you more [8:59] credibility to various customers and for [9:02] us we work with biggest agricultural [9:04] customers big governments civil [9:06] governments defense and intelligence all [9:08] of those sort of actors they want to [9:09] know you're going to be around. [9:10] >> Exactly. [9:11] >> And not going to disappear. I mean, we [9:12] have countries that are fully dependent [9:14] on us giving them information. They [9:16] don't want to just disappear. So, they [9:18] really care that we're going to be [9:19] around and and being a public company [9:21] gives you the kind of force in the world [9:23] that people go, "Okay, you're here to [9:24] stay and you have access to capital if [9:27] you need it and so on, right? It's [9:28] legitimizing [9:29] >> and you know, you you know where the [9:31] stock is at any one day. You know, we're [9:33] not focused on that dayto-day. We're [9:35] focused on how we build long-term value [9:37] for our shareholders, right? And um you [9:41] know the market is I think started to [9:42] really understand where space is going, [9:44] why it is changing the the the world. [9:47] You know people forget how space is part [9:50] of your everyday life. Every time you [9:51] use a phone you're using communications [9:54] using satellites or GPS using satellites [9:57] or satellite data in some way or another [9:59] that's sort of integrated in your lives. [10:01] you may not realize it um but it's just [10:05] booming now there's there's a [10:07] >> and the the story's changed as well [10:09] obviously with SpaceX going public but [10:11] has the framing of [10:14] planet gone from like a data source for [10:17] people who need data from space and maps [10:18] to hey this is a tool to accomplish [10:21] tasks and military like post Andrew's [10:25] success like you probably would have [10:27] been bucketed into Andrew as a military [10:30] tech company. So is that framing what's [10:33] driving a lot of [10:33] >> I think it's a bit more nuance than [10:34] that. I mean firstly for the audience's [10:36] benefit what planet does we have [10:38] satellites doing earth imaging. We have [10:39] the largest earth imaging fleet about [10:41] 200 satellites. They image the entire [10:43] earth every day. So think of it like the [10:45] Google uh satellite lay on Google maps [10:48] that you can look at except it's today's [10:50] date rather than 3 years old and we have [10:52] every day going back. So it's a time [10:54] series analysis of everything going on [10:56] on the earth. That's useful for farmers. [10:58] It's useful for energy companies. It's [11:01] useful for civil governments, flooding [11:03] and fires. It's useful for security [11:06] applications like you're getting at. And [11:08] it's a wide variety of use cases. Um, I [11:11] think that where we're seeing this is [11:13] that AI is now enabling the it's [11:18] basically reducing the barrier to entry [11:20] so that more people can get access to [11:22] this, right? And uh, you know, there's a [11:24] lot more to say on that, but AI is only [11:26] as good as the data it's trained upon. [11:28] percentage is military. I'm curious. [11:30] Sorry. [11:30] >> What percentage of revenue customer bas [11:34] um uh security is part of the initial [11:36] thing that we said we would do um out of [11:39] the gate, but it's true there's a bigger [11:41] fraction today than perhaps we would [11:43] have guessed. Um but the needs of the [11:45] geopolitical situation right now demand [11:48] what we're doing. Um you know, just as [11:50] an example, what this does is enable [11:52] them to see threats around the corner. [11:54] >> Yes. and then uh you know give them [11:57] weeks or months advanced warning of [11:59] things and then that enables them to [12:00] more likely do things that stop [12:02] conflict. So we believe this is you know [12:05] really better for the world. [12:06] >> Are you reticent to be perceived as a [12:09] military company? [12:10] >> Not really. I but I wouldn't say we're [12:12] limited to being perceived like that, [12:14] right? um we we are helping farmers, we [12:16] are helping uh you know uh energy [12:19] companies, civil governments, we work [12:21] with NASA, we work with what have you [12:24] and and so um it's a it's a bigger it's [12:27] a bigger play than that but back to the [12:28] space uh piece of it what has changed [12:31] obviously rocket costs have come down [12:35] about four or 5x over the last uh 10 [12:37] years which has helped tremendously but [12:40] a thing that people don't know that is [12:42] actually perhaps more important is that [12:44] we've had a miniaturization of [12:46] satellites. So that the same satellite [12:48] that used to cost a billion dollars and [12:50] weigh 20 tons now cost a few kg or a few [12:54] tens or hundreds of kg [12:56] >> and can do just as much stuff if not [12:58] more. It's it's the same as the sort of [13:00] mainframe computer to desktop revol [13:03] computer revolution for space and it's [13:05] unlocking just like mainframes to [13:07] desktops unlock loads of applications. [13:10] This is unlocking loads of applications [13:12] and it's so both go in combo the launch [13:15] cost coming down and this [13:16] >> let's build let's build on this. So I [13:18] think I I'd like first you maybe take a [13:20] few minutes and then I I want to talk to [13:22] Andrew the same question. Both of you [13:24] guys are at the foot of what are [13:26] probably huge secular trends in [13:29] technology. How I would frame this is we [13:32] are rebuilding the data processing [13:35] infrastructure that has existed on the [13:38] earth in the sky. [13:40] And first you do the satellites, but I [13:43] would love for you to explain [13:44] space-based data centers because I think [13:46] everybody's hearing about that. [13:48] >> Are they really viable? What are they? [13:49] How will they work? etc. And then [13:51] Andrew, this is the rebirth of silicon. [13:54] We're going to find the next version of [13:56] Moore's law, which I think is more [13:58] timebounded, not transistor density [14:00] bounded. We now hear a lot about domain [14:03] specific architectures. we hear I mean [14:06] your chip was just a complete [14:07] transformation in terms of the design [14:10] principles that you know like at Grock [14:12] we took a very different approach [14:14] Nvidia's taken a very different approach [14:16] you took a big pizza shaped die and said [14:18] it yolo this is it and you were [14:20] right just explain where we're going in [14:23] silicon so maybe will you start and then [14:25] Andrew you start [14:26] >> I mean what we're seeing firstly in [14:28] space is is all these new applications [14:30] based on data and AI so you know we [14:33] we're collecting vastly more data about [14:35] the planet. And with SpaceX and Starlink [14:38] and One Webb, they're they're [14:40] transporting far more data around the [14:42] planet. As you say, we're sort of [14:43] changing um uh the nature of data using [14:46] satellites. And that's basically doing [14:49] what was once the province of [14:50] governments only and giving everyone [14:52] else access to uh satellite [14:53] capabilities. And that's going to I I [14:55] mean I I estimate there's a 75 to$100 [14:58] billion market just on Earth [14:59] observation. this kind of data we [15:01] collect and AI on top of that unleashing [15:04] all that application. So that's the [15:05] near-term thing. Applying large language [15:07] models to earth imagery data, unlocking [15:10] agriculture, you know, um energy, civil [15:14] government applications, permitting, you [15:16] name it. This is going to make [15:17] everything more efficient. Um and then [15:21] where we're going is indeed space is is [15:25] we did a study with uh our partners at [15:28] Google about eight or nine years ago uh [15:30] looking at what are the costs of data [15:32] centers on the ground, what are the [15:33] costs that it would take to put them in [15:35] space and when might it make sense to uh [15:38] do it non-aterrestrially. [15:40] And we figured out that when launch [15:42] costs come down to about $200 to $300 a [15:44] kilogram um uh it would be cheaper, just [15:48] simply cheaper to put the data centers [15:50] in space. Now we're about $1,000 a [15:52] kilogram, just over that in right today. [15:55] Uh but that's come down about 10x in the [15:57] last 10 years. Um on the current [16:00] trajectory with Starship in particular, [16:02] I would expect it the launch cost come [16:04] down there in 2 to 3 years. Elon might [16:07] say it's next week, but at least [16:10] realistically a couple of years. So [16:11] we're not far away from it literally [16:13] just being cheaper. Then in addition and [16:16] the intuition there that is that helps [16:18] people understand that is you would [16:20] naturally use solar panels for doing uh [16:23] the the data centers are a power [16:25] problem. It's a power game and you would [16:27] normally use solar panels. That's the [16:29] cheapest way to get a watt uh today by [16:31] far. But you you don't want intermittent [16:34] power. So then you have to have [16:35] batteries or you then you have to have [16:36] gas or then you have to have nuclear and [16:39] then it gets really expensive. Um in [16:42] space you can put a solar panel in a [16:44] suns synchronous uh dawn dusk orbit [16:46] where you're 24/7 looking at the sun. So [16:49] you can have a solar panel that collects [16:51] and gathers five times more energy per [16:55] solar panel than on the ground and you [16:57] don't have to have batteries or anything [16:58] else. Uh so the infrastructure for [17:01] comput in space is literally just solar [17:03] panels and the chips and then the RF [17:06] signals up and down. So it's actually [17:09] really quite simple. It was just a [17:10] question of when it's going to be [17:11] cheaper to launch all those solar panels [17:13] and chips into space than putting on the [17:16] ground. And it turns out that's going to [17:17] be in a few years. So we're partnering [17:19] uh with Google to launch some of their [17:21] TPUs into space. We've already launched [17:23] Nvidia's uh uh some of Nvidia's GPUs [17:27] into space. We're launching Google's [17:28] TPUs into space on an early test. [17:31] There's lots of technology to figure [17:33] out. [17:34] >> Let's have a conversation. Um, but it's [17:37] an early it's early days, but I think no [17:40] question within 10 years most compute [17:43] will be putting in space, which to give [17:46] you a sense is a lot of money, [17:48] [laughter] [17:48] like trillions, and um will be bigger [17:52] than any of the other space businesses [17:54] today. comes Earth imaging. This is why [17:57] we're getting into this. [17:58] >> Do you believe this? Do you believe [18:00] sending data centers to space makes more [18:02] sense or is it just the regular [18:05] >> Can you have him explain the business [18:06] first and then [18:07] >> Oh, yeah, of course. Yeah. So, I think [18:09] they're, you know, with all due respect, [18:11] one or two hard problems still left be [18:14] beyond putting putting GPUs in in in [18:17] space right now. I I think um we we [18:21] we're not super good yet at [18:25] building the clusters in space necessary [18:28] for the communication [18:30] >> between [18:30] >> between Exactly. between [18:33] >> We're not good at doing it on the [18:34] ground. [18:34] >> We're not good at doing it on the [18:35] ground. We're really not good at doing [18:36] it in space. I I think this is an [18:38] extraordinarily important and [18:39] interesting problem and one we should be [18:41] spending money and attacking. I've got [18:43] it in a slightly different time frame, [18:45] but one that certainly will occur. And [18:48] the the hard part is is it is it one of [18:51] those problems where uh the last 10% is [18:55] 80% of the time. [18:57] >> Now, self-driving was a problem like [18:58] that, right? Where the last 10% proved [19:01] to be a decade's worth of work and just [19:04] now we're over the hump. And we don't [19:06] know yet, but I think the interesting [19:08] work they're doing uh at Planet is [19:10] really important. And I think the [19:12] fundamental driver to experiment to even [19:14] get insight into whether I'm right or [19:16] not is to get down the cost of launch [19:19] vehicles. Then you can start doing [19:21] experiments and getting it wrong and [19:23] fixing it and figuring it out. And until [19:26] then it was unpaid. [19:27] >> For the foreseeable future, you're going [19:28] to be terrestrial. explain your business [19:30] and how you made these critical [19:32] decisions that kind of took you on a [19:35] different path and you know you versus [19:37] Nvidia versus AMD and what you think the [19:39] future of AI silicon looks like. I I I I [19:42] think there were two parts your your [19:44] first question was around sort of the [19:48] rise of silicon in general and I I think [19:50] what AI did and it's it's rarely sort of [19:53] framed this way but it allowed computers [19:56] to address a class of problems that [19:58] before AI computers were bad at. [20:02] We were bad at images for almost the [20:05] entire history of compute. We could [20:07] store them and that's about it. We were [20:10] bad at language. We could store it but [20:12] that's about it. We could transform [20:15] numbers. We were magical with numbers. [20:18] And what AI did starting in about 2015 [20:22] 16 is it opened the door the aperture to [20:25] say maybe we could use computers on [20:27] images. [20:29] All right. Maybe we could find insight [20:31] in images. Maybe not only could we store [20:34] language but we could generate it. All [20:37] right. maybe we could understand it [20:40] rather than storing it and regurgitating [20:42] it. And what this did is it it opened up [20:46] sort of to compute [20:48] huge areas that were previously [20:50] foreclosed and at the same time we were [20:53] adding to those areas. we were taking [20:55] vastly more images, [20:57] all right, terrestrially [21:00] in satellites. And what this did is it [21:02] it simultaneously opened up this entire [21:05] area and allowed compute to attack it. [21:09] And this is what's underpinning both [21:10] Nvidia's growth and and sort of all the [21:12] growth you're hearing about in in AI [21:15] compute is as a as a processor builder [21:18] as a hardware builder suddenly our tools [21:22] could attack more and different parts of [21:25] knowledge [21:27] and and that was sort of the first part [21:29] to to to answer your question. Now how [21:32] you do that there are lots of different [21:34] strategies tons of different ways to [21:35] skin cats. What we saw in 2015 were [21:40] several things. First, we saw that AI [21:43] would be an enormous consumer of [21:46] compute. [21:48] All right. And historically for computer [21:50] architects, [21:52] new workloads were the opportunity for [21:55] share to change, [21:57] right? Share changed when the rise of [22:00] graphics emerged and you got the [22:01] dedicated GPU. That's how Nvidia was [22:03] born. Share changed when cell phone [22:06] compute emerged and Intel and AMD who [22:09] had fabs and the best architects got [22:11] zero share and it all moved to ARM. [22:14] Right? Share changed in the late '9s [22:17] when Nortell and all these companies [22:19] we've forgotten about couldn't build [22:22] chips and couldn't do uh uh data [22:25] networking and and what you got with [22:28] Cisco and Juniper and Arist and this [22:30] collection of new companies. So we knew [22:32] that um this new problem [22:37] would present an opportunity for massive [22:40] change. [22:41] So we saw that we we made two bets. Um [22:45] the first was dedicated silicon would be [22:47] the answer. The second was it couldn't [22:51] look like a GPU. [22:54] And our view as computer architects is [22:57] if you want to be 20 times better than [22:59] somebody, right? Your architecture can't [23:01] look like them, [23:03] right? It it it can't. They have they [23:07] have enjoyed and and eaten all the [23:08] lowhanging fruit. So, if you build a [23:11] GPU, the odds that you're better than [23:12] Nvidia in our view are approximately [23:14] zero. That led us to a fundamentally [23:16] different architecture. All right. The [23:19] hard part here, the hard part is moving [23:23] data from memory to compute. [23:26] >> This is the fundamental problem in AI. [23:30] And we solved it with a way that that [23:33] very few others had even attempted, [23:35] which was to build a very big chip and [23:37] to put memory right next to compute. And [23:40] by building a big chip, a chip the size [23:42] of a dinner plate, whereas most chips [23:44] are the size of a postage stamp, we [23:47] could use a different type of memory. [23:50] And by using a different type of memory, [23:51] a memory that was vastly faster, we [23:54] opened up all sorts of opportunity. [23:57] So when OpenAI uses us, we're 15 or 18 [24:00] times faster than a GPU. [24:03] That means your answers are delivered [24:06] more quickly. It means your engagement [24:08] with with the AI is more enjoyable. It [24:11] means you can use the AI to solve harder [24:13] problems and not wait. [24:16] And the way to think about this is sort [24:17] of to ask yourself the the counter [24:20] factual question. How big is the market [24:22] for slow search today? [24:25] Right. Right. Is zero. How big is the [24:28] market for dialup? It's zero. H how long [24:32] do you wait for a website to resolve [24:34] before you click away? 3 seconds, 5 [24:37] seconds. You will not wait for AI. We [24:40] have to deliver it to you in a in in [24:43] real time. And that's what we saw. [24:46] That's what we built. [24:47] >> So the panel's on going public. A lot of [24:49] LPs in the room. They need to get [24:51] liquid. I'm curious about the journey [24:53] for your investors. Yeah. [24:56] >> Okay. So, well, you guys went public [24:57] what year? [24:58] >> Uh 2021. [24:59] >> 2021 by way of a spa. [25:01] >> Correct. [25:02] >> Okay. And your VCs were who? [25:06] >> Um Drake Professor Jverson was one of [25:09] the earliest Capricorn. Um Peter Te's [25:12] founders fund. Then we got uh Yuri [25:15] Milner's DST. [25:17] >> Okay. So your your investors come in, [25:19] you go public at 2 billion via a spa. [25:24] >> Now we're four years later. Really, it [25:26] wasn't until year three or four that 90% [25:29] of the value was created. Okay. So did [25:33] those early investors capture this 90% [25:37] move? Did they stay in it? [25:38] >> Most of them did. Yeah, most of them [25:40] did, which is really smart on their [25:42] part. Obviously, I think they should [25:44] hold on even more. Uh, if I didn't think [25:47] that, you should. [25:48] >> I'm a little bit self-interested. [25:50] >> What's interesting about this is [25:51] >> No, but really they did. And and I mean, [25:53] Google hasn't sold a share. They're our [25:55] largest single investor. Um, uh, [25:58] Capricorn didn't until very recently. [26:00] So, basically, most of them stayed [26:02] really well in and they got all of that [26:04] upside. And good for them. And the [26:07] reason I think this is so important [26:09] >> is that there are a lot of LPs in this [26:10] room [26:11] >> who they're like when a company goes [26:12] public give us the shares. [26:14] >> No no [26:15] >> give us give us the shares. This is a [26:16] counter example right? This happened to [26:19] us in 10 years ago. We invested [26:22] preipo at a billion dollars. We [26:24] distributed the shares I think at three [26:26] or four billion and then it went to 50 [26:29] billion over you know the course of the [26:30] next 24 months. And we had people who [26:32] called us who said well why didn't you [26:34] hold on to the shares? And we're like [26:35] because you're pounding on us, right, to [26:37] distribute to the shares. So you're an [26:38] example. Now [26:40] >> in your case, Andrew, [26:43] >> you have an innovation, right? You're [26:45] just now uh public. So all of your [26:49] investors are still under lockup like [26:51] like Altimeter, but you guys have [26:53] innovated with the banks on what I call [26:55] a dribble lockup. So over 6 months the [26:58] shares can be dribbled out according to [27:01] a bunch of performance hurdles which [27:03] SpaceX is going to have a very similar [27:05] >> when did we start this process of the of [27:08] the dribble [27:08] >> the dribble [laughter] [27:11] >> concept you started it years ago but [27:13] >> yes [27:13] >> but u we're all of that with respect to [27:16] the lockup I think this is the the most [27:19] innovative and I think SpaceX is going [27:21] to have a very similar innovation but [27:23] Andrew for your investors if you were [27:25] talking to ILPs right in the room. [27:28] Should Alimter be distributing the [27:30] shares when they come out of lock? How [27:33] do you think about you know your VCs [27:36] holding on to the shares kind of post [27:38] lock? [27:39] >> Well, I I I think historically more [27:40] money's made after IPO than before. [27:43] >> Yeah. [27:43] >> I I I think every single study shows [27:46] that uh there is more money to be made [27:50] both in percentage and in in what we [27:52] care about which is absolute. [27:54] >> Yes. and and and so uh the amount of [27:57] money that it's possible to put to work [27:59] in most venture companies is very [28:01] modest. I mean there are two or three or [28:02] five outliers but for the most part you [28:05] can only put a relatively little bit of [28:06] money to work. Um by the time we get [28:09] public there's a lot more money there if [28:11] things are going well and the [28:12] opportunity to make vastly more is after [28:15] IPO not before. Hey, if I could just add [28:18] on that, one interesting question is [28:20] what's going to happen with SpaceX on [28:22] this because you know a lot of the value [28:25] is is in the future, right? But I mean [28:28] most of the big tech companies went [28:30] public at a few billion, [28:32] >> not a few trillion. Like there's a lot [28:34] of zeros in between those, right? And [28:36] you got all this upside afterwards. Now, [28:39] for the equivalent liftoff, SpaceX would [28:41] have to be aiming at quadrillion [28:44] valuations. Now, I know Elon has those [28:47] sort of ambitions, but you really have [28:48] to believe in that uh to get, you know, [28:51] >> this is this is kind of the point I'm [28:53] getting to, right? We have three mega [28:55] IPOs, you know, we keep talking about [28:58] that are multi- trillion. [28:59] >> Yeah. [29:00] >> All of that value acrewed to private [29:02] market investors. [29:03] >> Planet Labs is a great example of [29:07] venture capital in the public markets [29:09] where the 10X has occurred in the public [29:12] markets. We're all advocates of these [29:14] companies coming public sooner. Had [29:16] Andrew had his way, he would have been [29:18] public 18 months ago, probably at $10 [29:21] billion rather than $50 billion. And [29:23] that 5x over the course of the last two [29:26] years would have gone to public market [29:27] investors. So go ahead. [29:29] >> Way better to be lucky than good. [29:31] >> Yeah. [29:33] >> So, so I think that I hear a lot of [29:36] people thinking that anthropic open AI [29:38] and SpaceX are the new normal. I [29:41] actually think the public markets may be [29:43] shifting back in this direction and a [29:45] lot of the companies in our portfolios [29:47] are now thinking about going public at a [29:49] billion or three billion or 5 billion. [29:52] >> We had this period of a decade where [29:54] Andrees was really pushing stay private [29:56] forever and I see the pendulum swinging [29:59] back to companies are like man I want to [30:02] be like Planet Labs and get public, [30:04] >> right? and and have to play in the big [30:06] leagues and do it in the public markets [30:08] like [30:09] >> here's what I'll say maybe just like to [30:10] the two of you guys both of you guys [30:12] have had enormous pressure because [30:14] there's visible competition that's [30:16] always sort of in your periphery [30:19] but I do think that getting public [30:21] sooner having the scrutiny of public [30:24] markets having the scrutiny of having to [30:25] deliver sharpens the focus it steel [30:29] sharpens steel iron sharpens iron [30:31] >> and I think innovation tends to get [30:33] better And so the idea that you allow [30:36] everybody to participate but you also [30:38] put yourself in the spotlight to me is [30:40] where great things happen. [30:41] >> I agree. [30:42] >> Um and so anyways I just wanted to say [30:44] to both of you um just as we wrap you [30:48] guys are an incredible testament to [30:49] entrepreneurship both of you. I mean [30:51] we've been talking literally since day [30:53] one me and Andrew because we went in [30:55] different paths and then we kind of [30:57] recon converged and then Will same with [30:59] you. [30:59] >> I'm happy it worked out for you Jimoth. [31:01] >> Well it's worked out for both of us so [31:02] it's [laughter] fine. Um you guys are [31:04] incredible testament to entrepreneurship [31:06] and uh I just want to say thank you for [31:08] everything you guys are doing and the [31:09] next few years are going to be really [31:10] spicy. [31:11] >> Yeah, if I could just spend a 30 seconds [31:13] on the next few years because I think [31:14] it's going to be so exciting with um as [31:17] I mentioned AI and space merging [31:19] together. We're going to see a take off [31:20] our applications. I like to say like all [31:23] the cool stuff that we're doing on the [31:25] with LLMs now is really based on just [31:28] the text of the internet being absorbed [31:30] into these models which is incredibly [31:32] powerful already but they don't know [31:35] about the real world. I call them [31:36] blind to you know they don't know about [31:38] that farm field that flood that security [31:40] situation around the corner. If you give [31:43] them real world data, then they can [31:45] answer real world problems. And that's [31:47] going to open up gazillions of [31:49] applications for these uh AI models. I [31:52] call them instead of having large [31:53] language models, large earth models or [31:56] uh instead of AI, planetary intelligence [31:59] where you have planetary sensing systems [32:00] in space, planetary compute systems in [32:03] space. And we can disagree or agree on [32:06] on exact time frame, but I think it's [32:07] going to happen. and and then that's [32:10] going to enable a huge economy. So, it's [32:13] an exciting time in the next few years. [32:15] >> Will Andrew, thank you guys very much. [32:17] Well done. [applause] [32:18] >> Thanks, guys. Okay. [music] [32:22] Thanks, buddy. Appreciate you. Really [32:24] appreciate it. Great seeing you, [32:25] brother. Congrats.