Cerebras Files $64 Billion IPO Debut as AI Inference Demand Soars

AI推理需求飙升，Cerebras提交640亿美元IPO申请 企业级AI买家更看重推理速度而非单纯的模型智力，尽管存在静态RAM（SRAM）内存限制，仍推动了对Cerebras硬件的需求。 本期节目深入解析了Cerebras的重磅IPO：其估值翻倍至640亿美元，尽管存在底层内存限制，仍印证了市场对超快速AI推理和晶圆级芯片的强劲需求。主持人还报道了Kevin Warsh获得参议院确认担任第17任美联储主席、公众对Figure Helix 2机器人直播展示中自主性的质疑，以及Elon Musk针对OpenAI与微软1500亿美元诉讼的结案陈词。节目最后轻松探讨了风险投资人Tim Draper长达52分钟的极限冰水浴。 Cerebras 以 640 亿美元市值成功完成 IPO 上市，首日开盘价达到每股 350 美元，远超预测市场预期并使上市前估值翻倍。 (0:03) Cerebras 通过直接在硅片中构建冗余核心克服了晶圆级芯片的良率顾虑；实际业务数据显示，相较于纯粹的模型智能，企业客户更愿意为超低推理延迟支付显著溢价。 (3:53) 受制于片上 SRAM 内存微缩瓶颈，Cerebras 芯片在扩展超大上下文窗口上面临挑战；业内预计未来 AI 将迈入“编排时代”，由前沿大模型向专用极速硬件分发高频微任务。 (10:03) 美国参议院以 54 比 45 票的微弱优势确认凯文·沃什（Kevin Warsh）出任第 17 任美联储主席，其上任伊始便面临潜在滞胀与政界激进降息诉求的双重挑战。 (15:19) 马斯克诉 OpenAI 及微软案进入结案陈词阶段，马斯克团队索赔逾 1500 亿美元并要求阻止其转型为营利实体，而 OpenAI 辩方强调原告从未拿出任何具有法律约束力的出资协议证据。 (23:43)

Cerebras achieved a landmark $64 billion public listing powered by high enterprise demand for rapid inference hardware, while broader industry headlines were shaped by Kevin Warsh's contested confirmation as Federal Reserve Chair and closing arguments in Elon Musk's multi-billion-dollar lawsuit against OpenAI. Cerebras Public Debut and Valuation Growth (0:03) Cerebras surpassed public market projections during its initial public offering, opening trading at $350 per share before stabilizing near $300, delivering a $64 billion market capitalization. Pre-IPO expectations had placed an optimistic ceiling at $50 billion. The final pricing completed an aggressive upward revision schedule, rising from an initial range of $115–$125 to $150–$160 on Monday before surging at launch. Institutional demand was tightly concentrated: the top 25 institutional accounts, including Fidelity and State Street, captured 60% of the share allocation, leaving one-third of bidder order books entirely unfilled. The public offering concluded a venture-funding arc that began in 2016 with a $100 million Series A round backed by Benchmark and Eclipse. Veteran investor Pierre Lamond joined Eclipse at age 84 to champion the firm's early and repeated rounds when other premier firms declined. Cerebras subsequently secured capital from Coatue and Vy Capital, reaching valuations of $1.6 billion in 2018, $4.0 billion in 2021, and rising from $8.0 billion to $23 billion in 2025 before listing at a $48.8 billion baseline in May 2026. Inference Speed Economics and Memory Bottlenecks (3:53) Cerebras validated its wafer-scale engine architecture by overcoming early manufacturing yield skepticism, utilizing redundant silicon cores to bypass localized defects across the wafer. Commercial traction is anchored by an expansive 750 MW computing deployment with OpenAI to run models like Codex 5.3 Spark. Citing findings from SemiAnalysis, enterprise adoption patterns reveal that customers place a massive premium on real-time inference speed over incremental gains in reasoning. Anthropic's Opus 4.6 Fast mode, for example, successfully charged six times standard pricing for a 2.5-fold increase in interactivity. While researchers like Andrej Karpathy advocate running slower, more deliberative models overnight, enterprise environments prioritize lower latency to avoid workflow abandonment. Despite this momentum, Cerebras faces structural constraints in memory scaling. The architecture relies on on-chip Static Random Access Memory (SRAM) rather than off-chip DRAM or High Bandwidth Memory (HBM) setups like Nvidia's NVL72. Because SRAM density no longer scales efficiently with smaller process nodes, memory capacity grew only 10% between the WSE-2 (40 GB) and the WSE-3 (44 GB). Adding memory would require sacrificing compute real estate, pointing to an operational transition toward an 'orchestration age' where high-capacity frontier models delegate discrete, latency-sensitive sub-tasks to specialized hardware clusters. Kevin Warsh Confirmed Amid Policy Friction (15:19) The United States Senate confirmed Kevin Warsh as the 17th Chair of the Federal Reserve in a narrow 54–45 party-line vote. The divided confirmation highlights looming friction between monetary leadership and White House demands for aggressive interest rate cuts, particularly amid the policy trap of stagflation. In an environment of stagnant growth paired with high inflation, traditional monetary tools present severe trade-offs: easing rates risks compounding inflation, while raising rates threatens to depress economic output further. The hosts contrasted Warsh's contentious confirmation with Jerome Powell, who secured over 80 bipartisan Senate votes across both of his terms. While historical chairs like Paul Volcker and Ben Bernanke were tested by systemic structural collapses, the hosts debated Powell's legacy. One perspective noted that the structural baseline of the US economy in 2019 was well capitalized, limiting post-pandemic rate shock primarily to Silicon Valley valuations rather than broader institutions. Conversely, producer Tyler credited Powell's decisive monetary interventions with quietly preventing an existential market catastrophe. Musk-OpenAI Litigation and Robotics Scrutiny (23:43) Closing arguments concluded in the legal battle pitting Elon Musk against OpenAI and Microsoft, wrapping up in three weeks rather than the anticipated four. Musk's counsel demanded over $150 billion in damages, the removal of Sam Altman from the board, and an injunction against converting OpenAI into a for-profit corporate entity, alleging that leadership enriched themselves using non-profit donations. OpenAI's defense team argued that Musk failed to demonstrate an enforceable, formal written contract or produce witnesses corroborating restrictions on his early contributions. Prediction market Kalshi showed Musk's implied win probability falling from a 58% peak down to 30% by the trial's close. In concurrent robotics developments, Figure founder Bret Adcock showcased an eight-hour continuous shift featuring humanoid units running on its proprietary Helix 2 neural network, generating 3.4 million views. Viewers raised questions over potential human teleoperation after observing occasional target misses and robots touching their own heads during tasks. Adcock clarified that the head gesture was deliberate autonomy logic, ensuring robotic arms do not obstruct onboard camera sensors while scanning incoming items. The broader technology landscape continues to navigate acute operational and regulatory tensions, seen in the hardware trade-offs between SRAM density and latency, the shifting probabilities in OpenAI's governance litigation, and the political pressures confronting monetary policy under a newly confirmed Federal Reserve Chair.

本期节目聚焦于 Cerebras 凭借低延迟晶圆级芯片实现超预期上市背后的技术与商业逻辑，深入剖析了凯文·沃什出任美联储主席所面临的潜在滞胀困境，并探讨了马斯克诉 OpenAI 案结案辩论与具身智能机器人自主性引发的信任争议。 Cerebras商业化验证与低延迟溢价 (0:03) 芯片设计公司 Cerebras 的公开上市表现远超市场预期，首日交易开盘价报 350 美元并稳定在 300 美元左右，推动其市值达到 640 亿美元，彻底打破了此前 500 亿美元的乐观预测上限。其股价定价区间在挂牌前几天经历了急剧攀升，从早期的 115 至 125 美元跃升至 150 至 160 美元，最终以突破 300 美元的高位交易。本次发行的分配格局高度集中，富达（Fidelity）与道富（State Street）等排名前 25 位的大型机构投资者拿下了账簿中 60% 的配额，导致三分之一的认购订单完全落空。从早期投资历程来看，84 岁加盟 Eclipse 的传奇风投人 Pierre Lamond 在其他顶尖风投退缩时持续重注，伴随公司估值从 2016 年 A 轮的 1 亿美元、2025 年的 80 亿与 230 亿美元，一路推进至 2026 年 5 月以 488 亿美元指导估值挂牌并迅速冲破 600 亿美元。 在底层架构上，Cerebras 依靠芯片内设计的冗余核心技术，化解了外界长期以来对晶圆级引擎（WSE）良率极低、单点缺陷易毁掉整块晶圆的担忧。该公司目前已与 OpenAI 达成一项 750 兆瓦（MW）的大型电力与算力部署协议，用于驱动 Codex 中的 GPT-5.3 Spark 等模型。引用 SemiAnalysis 的行业分析，AI 经济中的企业与终端用户展现出明确的“显性偏好”，即比起纯粹的模型智能，市场更愿意为极高的推理速度与低延迟交互支付高额溢价。例如 Anthropic 的 Opus 4.6 Fast 模式以六倍价格提供仅提升两倍半的速度，仍获得企业强烈追捧；低延迟特性有效避免了用户因流式响应缓慢而走神放弃任务的现象。 晶圆级SRAM瓶颈与模型编排范式 (10:03) 尽管商业表现亮眼，SemiAnalysis 的深度分析指出 Cerebras 仍面临严峻的硬件物理瓶颈。与英伟达依托 NVL72 机架及离片 DRAM/HBM 扩展内存的互联方案不同，Cerebras 完全依赖片上静态随机存取内存（SRAM）。随着先进制程下 SRAM 难以实现有效物理微缩，其内存容量增长陷入停滞：WSE-2 拥有 40 GB SRAM，而推进完整节点制程后的 WSE-3 仅提升 10% 达到 44 GB。若要在单个晶圆上争取更多内存容量，就必须牺牲关键的计算核心面积，这使得 Cerebras 难以直接承载超大参数量模型，也难以应对复杂 Agent 流程对远超 128K 超长上下文窗口的刚性需求。 针对硬件瓶颈与应用演进，节目预测 AI 行业将由当下的“Agent 时代”迈向“模型编排时代”（Orchestration Age）。在这一复合架构下，具备超强综合思考能力的巨型模型将充当如同“高级副总裁”般的顶层指挥官，负责规划工作流并将高频、快速的模块化微任务分发给 Cerebras 这类低延迟极速硬件处理，正如现代计算体系中调用网页搜索或数据库查询的分工协作，从而在架构端规避单芯片内存扩展的物理局限。 沃什执掌美联储与滞胀货币难题 (15:19) 美国参议院以 54 票赞成、45 票反对的微弱表决结果，正式确认凯文·沃什（Kevin Warsh）为第 17 任美联储主席。这一高度分化的党派投票揭示出其任期开端充满政治博弈，沃什一方面面临来自白宫激进降息的强烈要求，另一方面又不得不应对联储委员会内部对大幅宽松持审慎态度的阻力。与前任鲍威尔（Jerome Powell）在两届任期提名中均斩获逾 80 票压倒性两党支持的历史相比，沃什的政治处境更为严峻；主持团队亦探讨了鲍威尔的任内功过，指出其虽平稳度过 2020 年疫情，但并未经历如 2008 年次贷危机或互联网泡沫破裂那般的深层结构性金融炸弹爆炸，零利率时代的流动性收紧主要冲击了高估值长周期的科技公司。 沃什即将面临的核心宏观挑战在于经济学上的恶梦——滞胀（Stagflation）。在传统政策框架中，增长停滞可通过降息刺激经济，而经济过热则可通过加息平抑通胀；然而一旦面临经济停滞与通胀走高并存的局面，降息将带来通胀脱缰的巨大风险，而加息则会进一步扼杀实体经济活力。这一两难困局不仅考验美联储传统政策工具箱的有效性，也将深度检验新任主席在多方政治与经济诉求夹击下的平衡能力。 马斯克诉OpenAI与具身智能争议 (23:43) 在司法与行业博弈层面，埃隆·马斯克针对 OpenAI 及微软的历史性诉讼正式进入结案陈词，原本为期四周的审理进程被压缩至三周完成。预测市场 Kalshi 上马斯克胜诉的概率从 28 日的 58% 峰值迅速回落至结案时的 30%。马斯克法律团队索赔逾 1500 亿美元，并要求解除萨姆·奥特曼（Sam Altman）的董事职务以及叫停其转型为营利性实体的计划；然而 OpenAI 辩护律师反击指出，原告自始至终无法出具任何具备法律约束力的书面捐赠限制契约，且没有任何联合创始人、员工或亲友证人能证实具有法律效力的协议存在。此外，法庭内外还充斥着双方就 36 英寸显示器转接线连接故障的滑稽僵局，以及 OpenAI 早期 Dota 比赛成果几乎每日被反复提及的戏剧性细节。 与此同时，机器人初创公司 Figure 创始人 Bret Adcock 发布的视频引发了具身智能领域的真实性辩论。该团队展示了在 Helix 2 神经网络控制下人形机器人实现 8 小时轮班作业的 24 小时直播，播放量突破 340 万次；但观众注意到机器人操作中偶发的 1 英寸位置偏差以及作业中途触摸头部的举动，引发了外界关于是否存在人类秘密远程操纵（Teleoperation）的广泛怀疑。Adcock 随后解释称触碰头部动作是为了在双臂跨越躯干时将肢体移开头部相机传感器视野以便扫描下一包裹，暴露出前沿具身硬件在证明其端到端模型真实自主性时面临的巨大公众信任挑战。 本期节目所梳理的科技、宏观与法律动态展现了前沿行业的关键转折：Cerebras 的资本狂欢掩盖不住 SRAM 微缩停滞的物理法则约束，沃什履新美联储必须直面降息诉求与滞胀风险的正面冲突，而马斯克对 OpenAI 的天价追索与 Figure 机器人的自主性争议，则持续拷问着人工智能时代在商业契约与技术真实性上的底线边界。

## Background 本期节目由科技播客频道 TBPN 于 2026 年 5 月 22 日发布在 YouTube 平台。页面简介显示，这是 Diet TBPN 精华系列节目，旨在用约 30 分钟浓缩主播 John Coogan 与 Jordi Hays 主持的科技谈话全景；该栏目曾被《纽约时报》誉为“硅谷的新痴迷”，近期接连专访了萨姆·奥特曼、马克·扎克伯格等科技领袖。 ## TL;DR 节目核心聚焦 AI 芯片独角兽 Cerebras 的现象级 IPO：其晶圆级芯片通过冗余设计攻克良率瓶颈并在首日冲上 640 亿美元市值，依靠与 OpenAI 达成的 750 MW 合作凸显了极速推理的巨大商业溢价，但也面临片上 SRAM 扩展受限等隐忧。宏观与行业层面，主播探讨了凯文·沃什（Kevin Warsh）以微弱优势获批出任美联储新主席所面临的滞胀两难与鲍威尔的历史功绩争议，以及黄仁勋涉华芯片博弈与 Figure 人形机器人演示引发的“遥操作”质疑。最后，节目详尽拆解了马斯克诉 OpenAI 案结案陈词阶段的 1500 亿美元天价索赔、双方在法庭上的“高管 P 图互黑”与借用显示器找不到转接头的戏剧性闹剧。 ## Atmosphere & style 整期节目节奏明快轻盈，洋溢着浓郁的硅谷极客幽默与创投圈特有的戏谑讽刺，呈现出在硬核技术拆解与荒诞吃瓜吐槽之间无缝切换的生动质感。主播们上一秒还在严谨推演半导体晶圆物理尺寸与美联储货币政策困局，下一秒便对法庭上 15 位大律师抢转接头、敲钟摆拍摄影棚以及“红毛猩猩戴 VR 遥控机器人”展开大开脑洞的锐评。这种兼具行业纵深洞察与松弛毒舌的对话状态，让原本沉重的商业诉讼与芯片技术议题充满了极具代入感的娱乐张力。 ## Bulletin - [00:06](https://www.youtube.com/watch?v=QPk38R7UaMc&t=6s) AI 芯片独角兽 Cerebras 成功上市，首日市值突破 640 亿美元，其颠覆传统的晶圆级超级芯片凭借内置冗余核心技术，打破了良率与商业化质疑。 - [05:01](https://www.youtube.com/watch?v=QPk38R7UaMc&t=301s) SemiAnalysis 调研显示用户对极速 AI 推理展现出极高支付意愿，OpenAI 已与 Cerebras 敲定 750 MW 算力协议，并将其芯片落地于 Codex 的 GPT-5.3 Spark 模型。 - [14:53](https://www.youtube.com/watch?v=QPk38R7UaMc&t=893s) 美联储人事落定，凯文·沃什以 54 票赞成、45 票反对的微弱优势获批担任第 17 任主席，在政治施压与潜在滞胀风险夹缝中走马上任。 - [23:43](https://www.youtube.com/watch?v=QPk38R7UaMc&t=1423s) 马斯克诉 OpenAI 案进入结案陈词，马斯克索赔逾 1500 亿美元并要求罢免奥特曼，但预测市场上马斯克的胜诉概率已大幅下跌至 30%。 ## Key takeaways - [00:06](https://www.youtube.com/watch?v=QPk38R7UaMc&t=6s) Cerebras IPO 首日股价最高触及 350 美元，市值冲上 640 亿美元，彻底颠覆了传统小芯片切割的晶圆级设计路线赢得资本市场狂热认可。 - [03:53](https://www.youtube.com/watch?v=QPk38R7UaMc&t=233s) 面对整片晶圆微小物理缺陷会导致报废的良率魔咒，Cerebras 依靠芯片内置的大量冗余计算核心与硬件级动态绕过，成功化解了制造与可用性难题。 - [05:01](https://www.youtube.com/watch?v=QPk38R7UaMc&t=301s) 在生产环境中推理速度比纯粹智力展现出更高的商业溢价，例如 SemiAnalysis 将 80% 的 AI 预算投入在价格贵 6 倍但速度提升不到 2 倍的快速模式上。 - [08:52](https://www.youtube.com/watch?v=QPk38R7UaMc&t=532s) 正如亚马逊发现每增加 100 毫秒延迟就会损失 1% 销售额，AI 交互中的高延迟会导致用户注意力转移，促使巨头不惜代价采购极速推理算力。 - [10:01](https://www.youtube.com/watch?v=QPk38R7UaMc&t=601s) Cerebras 未来面临严峻的物理瓶颈：台积电晶圆尺寸受限且 SRAM 缩放停滞，第三代 WSE-3 容量仅微增至 44 GB，在面对长上下文和英伟达 NVL72 机架互联时处于下风。 - [11:51](https://www.youtube.com/watch?v=QPk38R7UaMc&t=711s) AI 系统正在迈入智能编排时代，英伟达或 TPU 上的顶级大模型将扮演决策调度者，而对延迟敏感的具体子任务则交由 Cerebras 驱动的小模型协同执行。 - [13:13](https://www.youtube.com/watch?v=QPk38R7UaMc&t=793s) 84 岁高龄加盟 Eclipse 的传奇投资人 Pierre Lamond 早期重注 Cerebras，在本轮前 25 大机构包揽 60% 份额的惨烈认购中完成波澜壮阔的生涯终章。 - [14:53](https://www.youtube.com/watch?v=QPk38R7UaMc&t=893s) 凯文·沃什获批执掌美联储，参议院按党派划线的微弱票数凸显其所面临的高度政治分化，未来需在白宫降息施压与抗击滞胀两难中艰难平衡。 - [18:09](https://www.youtube.com/watch?v=QPk38R7UaMc&t=1089s) 鲍威尔的历史功绩引发激辩，一种观点认为其并未经历 2008 年式的系统性绝境，另一种观点则认为危机未爆恰恰证明其宏观调控的高效兜底。 - [20:04](https://www.youtube.com/watch?v=QPk38R7UaMc&t=1204s) 黄仁勋扛 5090 访华恶搞图走红背后，反映了英伟达在中美地缘峰会背景下，围绕 H100 销售限制与 Blackwell 架构出海监管政策所面临的高风险博弈。 - [20:44](https://www.youtube.com/watch?v=QPk38R7UaMc&t=1244s) Figure 展示的人形机器人 8 小时连续作业视频引发广泛关于 VR 远程遥操作的技术争议，官方坚称其摸头动作仅为避免遮挡摄像头。 - [23:43](https://www.youtube.com/watch?v=QPk38R7UaMc&t=1423s) 马斯克诉 OpenAI 案结案陈词演化为法庭公关战，双方不仅互秀修图丑化对方高管，马斯克团队更因缺乏书面出资限定协议与关键证人而面临法律层面的举证困境。 - [27:30](https://www.youtube.com/watch?v=QPk38R7UaMc&t=1650s) 马斯克向法庭索赔逾 1500 亿美元并要求阻止 OpenAI 营利性架构转型，但法庭现场因 36 英寸大显示器缺乏转接头而上演了荒诞的硬件插曲。 ## Detailed walkthrough ### [00:06](https://www.youtube.com/watch?v=QPk38R7UaMc&t=6s) Cerebras 震撼上市：估值翻倍与晶圆级芯片的胜利 节目开场便切入重磅科技新闻：AI 芯片独角兽 Cerebras 的 IPO 表现堪称“现象级”，公司估值几乎在一夜之间完成翻倍。主播 Brandon Gorrell 在 TBPN 的日常通讯中提到，此前外界对 Cerebras 的预期普遍较为保守，即便在周五和周一写通讯时，他认为 500 亿美元的上市估值就已经是非常乐观的估计，甚至许多预测市场上根本没有设立 500 亿美元以上的下注选项。然而 Cerebras 彻底打破了天花板，在公开交易首日市值直接冲上 640 亿美元，盘中股价一度摸至 350 美元，随后在 300 美元左右波动。在此之前，其发行定价区间经历了连续跳涨，先从 115 至 125 美元上调至 150 至 160 美元，最终在二级市场更是一路呈现“只涨不跌”（literally up only）的疯狂态势。 Cerebras 能够在资本市场上引发如此轰动，核心在于其彻底颠覆传统半导体设计的“晶圆级芯片”（Wafer-Scale Engine）。主播们强调，通常的芯片制造逻辑是将整片晶圆蚀刻出大量芯片后切割成独立的小芯片（Die），而 Cerebras 却反其道而行之，直接将一整片晶圆做成一整块超级芯片——也就是名副其实的“全球最大芯片”。这种将极致纯粹的设想严肃落地的技术路线，一度被许多业内人士质疑，如今却用惊人的资本回报证明了自身的价值。 - **市值表现**：首日交易市值达到约 640 亿美元，股价最高触及 350 美元，随后稳定在 300 美元附近。 - **发行价跳涨轨迹**：从最初预期的 115–125 美元一路调高至 150–160 美元，上市后继续暴涨。 - **核心产品定位**：摒弃传统“晶圆切割成小芯片”的做法，直接生产整块晶圆级超级芯片。 --- ### [01:27](https://www.youtube.com/watch?v=QPk38R7UaMc&t=87s) 创业者的“假装成功”片场与冰淇淋博物馆狂想 看到敲钟现场的彩带与狂欢画面，两位主播忍不住开启了一段荒诞而真实的调侃。他们联想到洛杉矶那些专门出租给网红和“导师”的假私人飞机摄影棚，并提议应该有人为创业者打造一套“假装纳斯达克敲钟”的实景摄影棚：背景挂上客户自己的公司 Logo，创业者举着木槌假装敲钟，头顶洒满漫天飞舞的金箔彩带。走出敲钟摄影棚，隔壁就是一辆租来的兰博基尼，拍完照直接拿去兜售“创业大师课”，收割群里的上千名付费学员。 这段关于“人设造假产业”的畅想，让主播回想起曾经红极一时的“冰淇淋博物馆”（Museum of Ice Cream）。与传统的自然历史博物馆或盖蒂中心（Getty）不同，这类场所并非真正的文化机构，纯粹是专为社交媒体拍照打卡而生的体验空间。虽然里面会分发一点冰淇淋，但真正的核心卖点是巨型冰淇淋雕塑和标志性的“彩色糖针海洋球池”（sprinkle ball pit）。主播调侃自己当年的点子其实是开一家“Instagram 道具博物馆”：集合假私人飞机、假兰博基尼、粉红打卡墙、人造海滩，甚至配备装满假杠铃片的健身房，让客人假装自己卧推 500 磅。顾客只需要换装穿梭于不同房间，就能拍齐从炫富、度假到结婚生子的虚构人生全套照片。当另一位主播补充指出“冰淇淋博物馆如今已经在全球开出了 7 家分店”时，两人笑称这套“打卡经济学”居然真的演化成了一门跨国大生意。 --- ### [03:53](https://www.youtube.com/watch?v=QPk38R7UaMc&t=233s) 攻克良率质疑：晶圆级芯片冗余设计与真实性能落地 欢笑过后，讨论重新聚焦到 Cerebras 扎实的技术突破上。Cerebras 处于 AI 底层供应链极其纵深的位置，长期以来围绕它的 FUD（恐惧、不确定性与怀疑）层出不穷。外界过去常常怀疑："如果 Transformer 架构被淘汰了怎么办？"或者"一旦底层算法需要完全不同的计算架构，这种专用大芯片岂不是沦为废铁？"此外，行业最为致命的质疑在于“良率陷阱”。在传统的芯片蚀刻工艺中，整片晶圆出现微小物理缺陷是不可避免的。如果一片晶圆切割成 64 颗芯片，坏了一颗只需扔掉一颗；但如果要用一整片晶圆做成一块芯片，晶圆上的任何微瑕理论上都会导致整块芯片报废，良率将低到无法商用。 ```mermaid flowchart TD accTitle: Cerebras 晶圆级芯片良率与冗余架构设计 accDescr: 对比传统切片制造与 Cerebras 冗余核心设计如何解决晶圆缺陷问题 subgraph Traditional["传统芯片制造逻辑"] A["整片晶圆蚀刻"] --> B["晶圆产生微小物理缺陷"] B --> C["切割为数十颗独立芯片"] C --> D["剔除含缺陷的坏片，保留良品"] end subgraph Cerebras["Cerebras 晶圆级设计"] E["整片晶圆保留完整"] --> F["晶圆表面必然存在局部缺陷"] F --> G["内置大量冗余计算核心"] G --> H["硬件级绕过损坏核心，不全量激活"] H --> I["实现超大单体芯片极高可用性与良率"] end ``` 主播们回顾了此前对 Cerebras 首席执行官 Andrew Feldman 的采访，透露了 Cerebras 是如何化解这一物理魔咒的：他们并未试图制造“完美无瑕的晶圆”，而是在芯片内部设计了大量的冗余计算核心（redundant cores）。在实际使用中，系统并不需要激活整片晶圆上的每一个核心，而是通过内置算法绕过缺陷核心，从而巧妙攻克了良率难题。如今，Cerebras 的芯片不仅完全能够运转，更在实际负载中展现出了惊人的计算速度，彻底粉碎了过去的怀疑论。 --- ### [05:01](https://www.youtube.com/watch?v=QPk38R7UaMc&t=301s) 速度 vs 智力：高溢价推理与 OpenAI 的 750 MW 豪赌 对于 Cerebras 而言，最具决定性的商业风向来自知名半导体研究机构 SemiAnalysis 的实测洞察：在真实生产环境中，企业和开发者展现出了对“推理速度”极高的支付意愿（willingness to pay for speed）。以 Anthropic 的 Opus 4.6 为例，其“快速模式”（fast mode）的价格高出普通模式 6 倍，尽管换来的仅仅是不到 2 倍的响应速度提升。在传统的性价比逻辑下，为 2 倍的速度支付 6 倍的价格显得极不合理，但现实中 SemiAnalysis 却将自身 80% 的 AI 支出全部砸在了快速模式上。这种用真实资金投票的“显性偏好”（revealed preference）表明，用户一旦体验到超低延迟的交互，就再也不愿倒退回缓慢等待的时代。 这一现象直接击中了 AI 领域的一场核心路线之争：用户究竟需要“更聪明的模型”还是“更快速的模型”？前特斯拉 AI 总监 Andrej Karpathy 曾主张追求更高的智能等级，表示自己完全不在意模型跑得慢，“大不了让它在后台运行一整晚”；而 OpenAI 的 Sam Altman 则坚信速度本身就是一种“魔法般的超级能力”，极速响应是推动生产力下一轮跃迁的关键。当前，OpenAI 显然坚定地押注了 Cerebras，双方已经达成了规模高达 750 兆瓦（MW）的算力合作协议。在 OpenAI 的 Codex 代码产品中，搭载 Cerebras 芯片的 GPT-5.3 已经化名为“Spark”正式上线。 主播分享了自己的实际体验：习惯了普通大模型像“打字机”一样逐字吐出 Token 并伴随震动的互动模式，乍看很可爱，但真正用于严肃工作时，立刻生成整页内容的体验要震撼得多——那种感觉就像打开维基百科页面，瞬间全盘加载完毕供人肆意滑动浏览。无论是在编程场景中瞬间输出整段代码，还是当作日常聊天机器人使用，极速推理带来的效率提升都是颠覆性的。主播用职场现实打了比方：两个能力相当的员工，如果其中一人执行速度快 5 倍，他在组织内创造的价值往往不是线性的；就像能卖出双倍业绩的销售或效率翻倍的核心骨干，在市场上拿到的薪水往往能达到普通人的 5 到 6 倍。速度带来的生产力跃迁，正在重构整个 AI 经济的定价法则。 - **SemiAnalysis 支出数据**：将其 AI 算力预算的 80% 投入在 Opus 4.6 快速模式上，宁愿为不到 2 倍的速度承担 6 倍的高昂溢价。 - **OpenAI 关键战略合作**：OpenAI 与 Cerebras 签署了高达 750 MW 的大规模供电与算力协议。 - **商用落地代号**：Cerebras 芯片目前已在 Codex 中实际驱动名为“Spark”的 GPT-5.3 模型。 ### [08:52](https://www.youtube.com/watch?v=QPk38R7UaMc&t=532s) 延迟即毒药：从亚马逊的百毫秒法则到大模型推理 电商领域流传着一句经典名言，虽不知其最初出处，却在无数行业深度分析中被反复引用：“每增加 100 毫秒的延迟，亚马逊就会损失 1% 的销售额”。当年亚马逊在扩张期曾发现，无论在前端 UI、页面排版、购买按钮位置，还是价格与折扣标签上做多少优化，一旦给页面添加了过多冗余代码导致加载变慢，转化率就会应声下跌。用户只要等上一秒钟，注意力就会瞬间涣散，直接转去干别的事情。 这种现象如今正在大语言模型（LLM）的应用场景中全面重现。用户抛出一个 Prompt，如果系统响应迟钝，人们就会想：“太慢了，我还是去刷会儿 Instagram Reels 吧”，转头就彻底忘了自己刚才到底问了什么，最终甚至连答案都没看。这种注意力流失在企业级业务场景中更为致命。当前 AI 推理（Inference）领域的格局正是如此：各大公司宁愿支付远超常规的溢价，也要换取极速推理能力，而这恰好是主打晶圆级芯片的 Cerebras 的绝佳机会。 ```mermaid flowchart TD accTitle: 延迟对用户行为的影响逻辑 accDescr: 展示从计算延迟到商业价值损失的链条机制 subgraph UserFlow["用户行为路径"] A["发起查询或加载请求"] --> B{"响应速度"} B -->|"低延迟 / 极速响应"| C["获得即时反馈，完成转化/问答"] B -->|"高延迟 / 卡顿等待"| D["注意力转移 (刷短视频/离开)"] D --> E["放弃任务，商业转化流失"] end ``` ### [10:01](https://www.youtube.com/watch?v=QPk38R7UaMc&t=601s) 晶圆级芯片的隐忧：SRAM 扩容瓶颈与 NVL72 的围堵 尽管极速推理给 Cerebras 带来了巨大的顺风，但知名半导体分析机构 SemiAnalysis 尖锐地指出了该团队未来几年必须直面的严峻挑战。目前 Cerebras 芯片在有限的片上内存中难以容纳超大模型，且多芯片互联扩展的能力尚不及英伟达——英伟达凭借 NVL72 机架将数十颗芯片紧密互联，已能从容处理当今最庞大的模型。 更大的瓶颈在于上下文窗口的演进趋势。业界正不可逆转地奔向无限上下文（Long Context），在智能体（Agentic）工作负载蔚然成风的当下，区区 128K 的上下文窗口很快就会落伍。然而 Cerebras 很难像传统方案那样轻松应对： * **晶圆物理尺寸受限**：台积电（TSMC）拥有标准化的晶圆加工规格，不可能随意将晶圆做得更大。 * **SRAM 缩放停滞**：Cerebras 的架构极度依赖直接集成在整颗晶圆上的片上静态随机存取内存（SRAM），但随着半导体工艺节点的推进，SRAM 的微缩效应已大幅放缓。 * **容量增长乏力**：从第二代 WSE-2（40 GB 内存）升级到第三代 WSE-3，人们原本期待容量翻倍甚至实现 10 倍跃升，最终却仅微增到 44 GB——跨越一个先进制程节点仅仅换来了 10% 的微弱提升。 * **算力与存储的零和博弈**：在一整张固定面积的晶圆上，想要塞入更多 SRAM，就必须牺牲用于计算核心的硅片面积，两者构成了直接的物理冲突。 ### [11:51](https://www.youtube.com/watch?v=QPk38R7UaMc&t=711s) 从单体 Agent 到智能编排时代：层级化协作与大算力生态 随着 Agent 工作流的成熟，业界可能很快就会从“用一个最顶尖模型硬抗全部任务”的 Agent 早期时代，迈入“智能编排时代（The Orchestration Age）”。未来的工作模式更像是一家公司的管理层架构：最聪明、参数量最大的顶尖模型扮演“高级副总裁”甚至“总裁”的角色，坐镇于英伟达 NVL72 或谷歌 TPU 集群上，负责全局决策与任务拆解。 正如现存系统将特定任务下发至数据库做 SQL 查询，或调用 CPU 进行网页搜索一样，顶层“老板模型”会将那些对延迟极其敏感、执行具体指令的高速子任务，直接派发给由 Cerebras 驱动的“速度型打工人”小模型。这种混合架构让各类芯片各司其职。正如前几年投资人 Daniel Gross 在分析英伟达是否被低估时所展示的产业演进逻辑，AI 时代的算力需求从最初高呼“GPU 是一切”，逐渐扩展到需要强大 CPU 乃至 Arm 和 Intel 架构协同的“巨型计算机系统（Big Computers）”。 ### [13:13](https://www.youtube.com/watch?v=QPk38R7UaMc&t=793s) 传奇风投的收官之作与 Cerebras 估值暴涨的 IPO 围绕 Cerebras 的上市，投资界的一桩传奇故事备受瞩目。投资人 Pierre Lamond 早年曾是红杉资本（Sequoia）和 Khosla Ventures 的合伙人，但最终在早期重注支持这家名不见经传芯片公司的，却是他在 84 岁高龄才加盟的 Eclipse。Pierre Lamond 出生于 1930 年——与沃伦·巴菲特同龄，在职业生涯暮年再度投出一家超级独角兽，堪称波澜壮阔的生涯终章。 本轮 IPO 的认购竞争异常惨烈，认购订单簿中有整整三分之一的投资者一股未得，分配结果被高度集中在少数顶级大户手中： * **筹码集中**：前 25 大机构投资者揽下了 60% 的份额，富达（Fidelity）和道富（State Street）等老牌巨头斩获颇丰。 * **现场氛围反差**：与去年 Klarna 上市时团队只派几个人到纽交所草草敲钟、下午就回工位当普通工作日对待的冷清场面截然不同，Cerebras 的挂牌现场狂热非凡。 * **估值跨越式暴增**： * **2016年（A 轮）**：估值 1 亿美元（由 Benchmark 与 Eclipse 投资）。 * **2016年（B 轮）**：Coatue 领投。 * **2017年（C 轮）**：VY Capital 领投。 * **后续攀升**：2018 年估值达到 16 亿美元，2019 年升至 24 亿美元，2021 年达到 40 亿美元。 * **后期飙升**：经历短暂低谷后，2025 年 Andreessen Horowitz（a16z）与富达以 80 亿美元入局，老虎环球（Tiger Global）随后以 230 亿美元入场。 * **2026年5月 IPO**：最终以 488 亿美元的惊人估值成功挂牌上市。 ### [14:53](https://www.youtube.com/watch?v=QPk38R7UaMc&t=893s) 凯文·沃什获批执掌美联储：从转笔记本名场面到滞胀危局 美联储人事尘埃落定，凯文·沃什（Kevin Warsh）被正式确认为美联储第 17 任主席。提及沃什，科技与创投圈津津乐道的是他此前与 Palantir 首席执行官亚历克斯·卡普（Alex Karp）对谈的一段经典网络名片：卡普一边若无其事地将一颗尼古丁袋塞进嘴里，一边把手指上的笔记本像篮球一样连转了四次，而打扮得像“来自 IBM 的规矩西装男”的沃什则在一旁一本正经地追问关于财务控制与产品销售的严肃问题，两人的反差感极其诙谐。 抛开花絮，参议院最终以 54 票赞成、45 票反对的微弱优势通过了对沃什的任命。这一极其分化的按党派划线表决，预示着沃什未来的任期将荆棘丛生： * **政治与委员会压力**：沃什不仅面临特朗普对降息的公开施压，还要面对美联储内部对盲目降息持极度怀疑态度的联邦公开市场委员会（FOMC）。 * **滞胀难题（Stagflation）**：如果经济停滞且通胀低迷，降息能迅速注入流动性；如果经济过热伴随高通胀，加息可以一箭双雕降温。但当通胀反弹与经济停滞同时发生时，美联储的货币工具将彻底陷入两难，绝不能贸然降息。 * **历史对比**：相较于保罗·沃尔克（Paul Volcker）或本·伯南克（Ben Bernanke）等历史传奇主席，即将于本周五卸任的杰罗姆·鲍威尔（Jerome Powell）在其两次提名确认中均斩获了参议院 80 票以上的超两党广泛支持。沃什在两党高度对立和白宫施压的夹缝中走马上任，其面临的政治与经济挑战不可同日而语。 ### [18:09](https://www.youtube.com/watch?v=QPk38R7UaMc&t=1089s) 鲍威尔历史功绩之争：卓越掌舵还是顺风行舟？ 节目围绕美联储主席杰罗姆·鲍威尔（Jerome Powell）的历史评价展开了激烈交锋。有观点认为，从 2018 年至今，鲍威尔并未真正遭遇过像 2008 年大衰退、互联网泡沫破裂或“黑色星期五”那样破坏性的系统性经济危机。即便面对全球新冠疫情的停摆冲击，美国在 2019 至 2020 年初的基本面依然非常强劲，消费者资产负债表健康、负债率低，既没有失控的影子银行体系，也没有随时可能引爆经济的“定时炸弹”。虽然短期内出现了极高的失业率并需要大规模经济刺激，但财政兜底并非美联储主席的本职工作——鲍威尔的职责是制定基准利率。至于零利率政策（ZIRP）时代的终结与后续波动，主要承受冲击和享受红利的也是硅谷那些现金流周期极长的科技企业，鲍威尔从未面临过像本·伯南克（Ben Bernanke）在 2008 年那样必须力挽狂澜挽救整个体系的绝境。 然而主播 Tyler 给出了完全不同的反向视角：没有发生超级崩盘，恰恰可能是因为鲍威尔的工作完成得太出色。如果美联储应对失误导致疫情期间经济彻底崩盘后再艰难拉回，公众反而会赞叹他力挽狂澜；正因为他平稳化解了危机，市场才觉得波澜不惊。这就像美式橄榄球里的顶级角卫（lockdown cornerback），“只要对手全场没有得分，观众就看不到惊险刺激的救球防守，因为他从一开始就把对方彻底防死了”。 - 核心争论点在于鲍威尔是否经历了类似伯南克在 2008 年金融危机级别的真正系统性挑战。 - 零利率时代（ZIRP）及其终结的阵痛，主要集中在拥有超长现金流预期的科技初创公司与风险投资生态。 ### [20:04](https://www.youtube.com/watch?v=QPk38R7UaMc&t=1204s) 黄仁勋中国行与高风险的中美芯片博弈 话题转向大洋彼岸的英伟达 CEO 黄仁勋（Jensen Huang）。科技投资人杰森·卡拉卡尼斯（Jason Calacanis）在社交媒体上发布了一张极具戏剧性的恶搞照片：画面中黄仁勋肩扛两只巨大的 GeForce RTX 5090 显卡包装箱亲赴中国推销，卡拉卡尼斯调侃其“销售永不眠（never stop selling）”。主播们笑称这张图看起来和他在阿拉斯加时的旧照如出一辙，但毫无疑问“零 AI 痕迹”。 在这场戏谑背后，是极具现实张力的半导体地缘政治博弈。随着高风险的中美峰会持续占据《华尔街日报》等各大媒体头版，围绕英伟达 H100 芯片在中国市场的出货动态以及下一代 Blackwell 架构计算芯片的出口监管政策仍处于剧烈动荡之中，全球半导体巨头正小心翼翼地在贸易监管的风口浪尖上平衡商业利益。 - 恶搞图中黄仁勋随身携带的箱子标注为最新一代消费级旗舰显卡 GeForce RTX 5090。 - 行业正密切关注中美地缘峰会对英伟达 H100 存量销售限制以及最新 Blackwell 架构芯片出海政策的具体影响。 ### [20:44](https://www.youtube.com/watch?v=QPk38R7UaMc&t=1244s) Figure 人形机器人 8 小时轮班：真全自主还是“大猩猩遥操作”？ 机器人初创公司 Figure 创始人布雷特·阿德考克（Bret Adcock）发布了一段轰动业界的视频，展示其人形机器人团队搭载自研神经网络模型 Helix 2，以人类水准连续运行完成 8 小时全班次作业，并宣称全程“完全自主运行”。该 24 小时直播引发全网狂欢，斩获了超过 340 万次播放。机器人在视频前半段的作业速度快得令人咋舌，但随后眼尖的网友开始提出质疑：机器人在分拣包裹时突然开始出现大约一英寸的定位偏差，甚至在抓取间隙做出了一个看似毫无逻辑的动作——抬起机械手臂摸了摸自己的“脑袋”。 面对“人工介入（human-in-the-loop）”或远程遥操作（teleoperation）的质疑，阿德考克给出了解释：当机械臂跨越身体向右侧抓取时，将手抬高是为了让手腕移出头部摄像机传感器的视线范围，避免遮挡镜头观察下一个待分拣包裹。不过社区显然不买账，网络上迅速流传出各种讽刺漫画，将所谓的 Helix 2 板载神经网络描绘成一个戴着 VR 头显在幕后实时遥控的操作员。主播们借机大开脑洞，提出了一个荒诞的“第三种解释”：如果阿德考克坚持声称“没有人类在环”，那完全有可能是由一只戴着 VR 头显的巨型红毛猩猩或黑猩猩在幕后操纵提线木偶，“毕竟猩猩体内也跑着生物神经网络，在字面意义上确实达成了‘零人类介入’的自主运行”。 - Figure 宣称该展示依靠在机器人机载系统上完全独立运行的自研 Helix 2 神经网络。 - 视频中引发争议的动作是机械手突然抬高至头部，官方解释为防止手部遮挡机载摄像头识别下一件包裹。 - 网络舆论与同行普遍怀疑该演示背后存在通过 VR 头显进行的遥操作（teleoperation）技术干预。 ### [23:43](https://www.youtube.com/watch?v=QPk38R7UaMc&t=1423s) 马斯克诉 OpenAI 案结案陈词：法庭“P 图大战”与多塔老梗 埃隆·马斯克起诉 OpenAI 及微软一案在联邦法院迎来了最后关头的结案陈词。这场原定长达四周的世纪诉讼被提前压缩至三周收官。在预测市场平台 Kalshi 上，押注马斯克胜诉的概率此前曾一度飙升至 58% 的高点，但截至结案当天已断崖式暴跌至 30%。《纽约时报》知名科技记者迈克·艾萨克（Mike Isaac）在现场发回了详尽的实时报道。艾萨克提到，法官正在向陪审团宣读极为严格的裁决指引，这标志着法庭公关表演的终结，法律层面对原告举证设立的门槛之高，让马斯克一方的立案诉求显得极为艰难；雪上加霜的是，法庭内的音视频设备状况百出，啸叫连连、麦克风频繁断音、视频信号卡顿，让人啼笑皆非。 庭审现场最抓人眼球的当属双方律师团撕破脸皮的“高管 P 图大战”。马斯克的代理律师对山姆·奥特曼（Sam Altman）和格雷格·布罗克曼（Greg Brockman）的人品展开狂轰滥炸，试图将奥特曼塑造成一个满嘴谎言的伪君子。为了在陪审团心中摧毁证人公信力，马斯克团队在展示屏幕上放出了两人同框对比图——奥特曼的照片被修得脸色阴郁，活像“正准备被联邦法警收监入狱的嫌犯大头照（mugshot）”，而马斯克的配图却神采奕奕，仿佛“刚刚盛装打扮准备奔赴 Met Gala 晚宴”。马斯克一方大打民粹主义牌，斥责 OpenAI 高管以慈善非营利造福人类为幌子疯狂敛财数十亿美元，并搬出谷歌创始人拉里·佩奇（Larry Page）作为反面典型，将其塑造成不顾人类存亡的技术狂魔，以此烘托马斯克投身 AI 纯粹是为了“守护人类未来”的崇高情怀。 OpenAI 方面的闭幕陈词则发起了致命反扑。辩护律师直接讽刺马斯克众叛亲离，指出其指控漏洞百出，“甚至连为他工作的员工，乃至他孩子的母亲（Grimes）都无法为他的说法作证”。与此同时，OpenAI 屏幕上展示的高管肖像则被修得极为精致柔和，画风宛如知名影楼 Olan Mills 出品的高端人像写真。整场庭审的本质被总结得十分精炼：马斯克一方坚称“这帮 OpenAI 高管富得流油且满嘴谎话”；OpenAI 一方则回击“这些全都是博眼球的场外闹剧，马斯克提出的诉求根本没有任何现行法律依据支撑”；而联合被告微软则化身经典梗图，“悄无声息地遁入了灌木丛隐身”。令人捧腹的是，作为 AI 早期里程碑的电竞对战游戏《Dota 2》（OpenAI Five）在为期三周的审讯中几乎每一天都会被律师反复提及，成为这场严肃商业对决中挥之不去的科技幽默。 - 预测市场 Kalshi 显示，马斯克打赢官司的概率在 28 日触顶 58% 后已下挫至 30%。 - 马斯克律师采取的核心诉讼策略是打碎证人信誉，公开指控奥特曼欺骗大众并以此谋求商业暴利。 - OpenAI 律师在结案陈词中强调马斯克缺乏关键人证支持，其指控在实体法层面完全站不住脚。 - 曾击败人类职业选手的《Dota 2》（Defense of the Ancients）AI 项目几乎每天都出现在法庭证据与陈述记录中。 ### [27:30](https://www.youtube.com/watch?v=QPk38R7UaMc&t=1650s) 马斯克诉 OpenAI 案结案陈词：1500 亿美元索赔与协议争议 庭审进入尾声，主持人讨论起陪审团的审议时间表。现场刚刚迎来整整一个月以来最长的一次休庭——长达 30 分钟。有前方人员打趣称终于能走出法庭买口正经饭吃，甚至琢磨着要不要去街对面的 Popeyes 买上一桶红豆米饭。彼时太平洋时间刚过下午 1 点 10 分，主持人推测当天的程序可能在下午 3 点或 4 点左右收尾。不过由于法庭通常在周五休息，双方目前的发言仅代表结案陈词阶段，并不意味着审理彻底结束，陪审团很有可能会将裁决留到下周一，迅速出结果的可能性较低。 最新庭审动态显示，OpenAI 的律师在周四就埃隆·马斯克（Elon Musk）法律团队对首席执行官萨姆·奥特曼（Sam Altman）发起的人身信誉攻击进行了全面反击。这是一场具有潜在行业地震级影响的审判，赌注极高：马斯克本人当天并未出现在法庭，而是陪同特朗普总统在海外出访；他向法庭索赔超过 1500 亿美元的赔偿金，并要求法院将奥特曼从董事会中除名，同时强行叫停 OpenAI 去年推进的向营利性企业转型的架构变更。 面对马斯克团队的猛攻，OpenAI 法律顾问团队展开了针对性抗辩： - **反驳信誉攻击与承诺指控**：OpenAI 律师团队成员萨拉·埃迪（Sarah Eddy）在结案陈词中极力削弱针对奥特曼个人诚信的攻击，并着重强调：在整场诉讼中，除了马斯克本人的口头陈述外，没有任何证人能够证实奥特曼、格雷格·布罗克曼（Greg Brockman）或 OpenAI 曾向马斯克做过任何具有法律约束力的承诺。 - **不存在资金限定协议**：休庭后，OpenAI 首席诉讼律师威廉·萨维特（William Savitt）向陪审团指出，马斯克要对这家初创公司提出权利主张，前提是必须存在一份明确规定其捐款如何使用的特定法律协议。然而萨维特直言，这样一份所谓的协议在现实中根本不存在。 ### [29:44](https://www.youtube.com/watch?v=QPk38R7UaMc&t=1784s) 法庭闹剧：36英寸显示器与“AGI时代的转接头之争” 除了严肃的法律交锋，法庭现场还上演了一出颇具戏剧性的科技设备插曲。《连线》（Wired）杂志记者马克斯·泽夫（Max Zeff）发回报道称，马斯克的律师团队搬进了一台约 36 英寸的大型显示器用于展示，OpenAI 的律师见状提出希望能共享使用该屏幕。马斯克一方当场拒绝，但法官随即介入，命令马斯克团队必须允许 OpenAI 使用这台显示器。 然而滑稽的一幕紧接着出现：OpenAI 团队发现自己的笔记本电脑根本无法直接连上这台显示器。正如现场调侃所说，人类距离通用人工智能（AGI）已经近在咫尺，但在法庭上依然被一个最原始的硬件转接头（dongle）难住了。泽夫随后更新道，终于有转接头被送进了法庭，现场一度聚集了多达 15 名西装革履的律师围在法庭中央，激烈讨论如何正确连接并点亮这台大显示器。 ### [30:22](https://www.youtube.com/watch?v=QPk38R7UaMc&t=1822s) 蒂姆·德雷珀的“监狱风”冰水推介会与节目结语 节目随后转向了硅谷知名风险投资人蒂姆·德雷珀（Tim Draper）的一则奇葩动态。德雷珀在社交媒体上发布动态宣布自己可能打破了一项世界纪录：“我在低于 40 华氏度（约 4.4 摄氏度）的温度下，在 52 分钟内听完了 52 场创业推介。欢迎来到我的办公室。”并打上了 #DraperUniversity 与 #生存训练 的标签。 主持人和制作团队就冰浴（Ice bath）的耐受极限展开了热烈闲聊： - 主持人提到普通人尝试冰浴时，水温往往其实在 50 华氏度（约 10 摄氏度）左右，并没有想象中刺骨；但如果真低于 45 华氏度，哪怕是去冲浪都会痛得让人难以忍受，通常必须穿戴潜水衣、手套或防寒鞋套。 - 团队成员补充道，播客名嘴乔·罗根（Joe Rogan）平时做冷水浸泡甚至能达到 34 华氏度（约 1.1 摄氏度），堪称冰浴界的“终极大 Boss”。 - 围绕德雷珀的照片，主持人吐槽其画风古怪得像 AI 自动生成的图：不仅四周堆满了黑色塑料垃圾袋，墙面简陋粗糙，水箱还连着外露的管道，整体视觉效果极其廉价，酷似“监狱风冰水池”。作为坐拥巨额财富的顶级风险投资家，德雷珀完全没展现出扎克伯格等顶级富豪置办奢华房产的气派，反而像是在自家后院自己胡乱拼凑、买了几只垃圾袋就直接开工。 在节目的最后，主播提醒观众在 Apple Podcasts 与 Spotify 上给予五星好评，并订阅 tbpn.com 上的独家新闻通讯，宣布明日太平洋时间上午 11 点继续准时开播。 ## Notable quotes > [00:06](https://www.youtube.com/watch?v=QPk38R7UaMc&t=6s) literally up only — Brandon Gorrell > [08:52](https://www.youtube.com/watch?v=QPk38R7UaMc&t=532s) 每增加 100 毫秒的延迟，亚马逊就会损失 1% 的销售额 > [18:09](https://www.youtube.com/watch?v=QPk38R7UaMc&t=1089s) 只要对手全场没有得分，观众就看不到惊险刺激的救球防守，因为他从一开始就把对方彻底防死了 — Tyler > [20:04](https://www.youtube.com/watch?v=QPk38R7UaMc&t=1204s) never stop selling — Jason Calacanis > [23:43](https://www.youtube.com/watch?v=QPk38R7UaMc&t=1423s) 甚至连为他工作的员工，乃至他孩子的母亲都无法为他的说法作证 — OpenAI 辩护律师 > [23:43](https://www.youtube.com/watch?v=QPk38R7UaMc&t=1423s) 这帮 OpenAI 高管富得流油且满嘴谎话 — 马斯克代理律师 > [23:43](https://www.youtube.com/watch?v=QPk38R7UaMc&t=1423s) 这些全都是博眼球的场外闹剧，马斯克提出的诉求根本没有任何现行法律依据支撑 — OpenAI 辩护律师 > [30:22](https://www.youtube.com/watch?v=QPk38R7UaMc&t=1822s) 我在低于 40 华氏度的温度下，在 52 分钟内听完了 52 场创业推介。欢迎来到我的办公室。 — Tim Draper ## Mentioned Cerebras, Brandon Gorrell, John Coogan, Jordi Hays, Tyler, Andrew Feldman, Pierre Lamond, SemiAnalysis, Anthropic, Opus 4.6, Andrej Karpathy, Sam Altman, OpenAI, Codex, GPT-5.3 (Spark), Amazon, TSMC, Nvidia, NVL72, WSE-2, WSE-3, Google TPU, Daniel Gross, Arm, Intel, Sequoia Capital, Khosla Ventures, Eclipse, Benchmark, Coatue, VY Capital, Andreessen Horowitz (a16z), Fidelity, State Street, Tiger Global, Klarna, Kevin Warsh, Alex Karp, Palantir, Donald Trump, Jerome Powell, Paul Volcker, Ben Bernanke, Jensen Huang, GeForce RTX 5090, H100, Blackwell, Jason Calacanis, Figure, Bret Adcock, Helix 2, Elon Musk, Microsoft, Kalshi, Mike Isaac, Greg Brockman, Larry Page, Grimes, Dota 2 (OpenAI Five), Sarah Eddy, William Savitt, Max Zeff, Wired, Tim Draper, Draper University, Joe Rogan, Mark Zuckerberg ## Video description Diet TBPN delivers the best of today’s TBPN episode in 30 minutes. TBPN is a live tech talk show hosted by John Coogan and Jordi Hays, streaming weekdays 11–2 PT on X and YouTube, with each episode posted to podcast platforms right after. Described by The New York Times as “Silicon Valley’s newest obsession,” the show has recently featured Mark Zuckerberg, Sam Altman, Mark Cuban, and Satya Nadella. Follow TBPN: https://TBPN.com https://x.com/tbpn https://open.spotify.com/show/2L6WMqY3GUPCGBD0dX6p00?si=674252d53acf4231 https://podcasts.apple.com/us/podcast/technology-brothers/id1772360235 https://www.youtube.com/@TBPNLive [Source](https://www.youtube.com/watch?v=QPk38R7UaMc)

[0:06] Brandon Gorrell: literally up only [8:52] 每增加 100 毫秒的延迟，亚马逊就会损失 1% 的销售额 [18:09] Tyler: 只要对手全场没有得分，观众就看不到惊险刺激的救球防守，因为他从一开始就把对方彻底防死了 [20:04] Jason Calacanis: never stop selling [23:43] OpenAI 辩护律师: 甚至连为他工作的员工，乃至他孩子的母亲都无法为他的说法作证 [23:43] 马斯克代理律师: 这帮 OpenAI 高管富得流油且满嘴谎话 [23:43] OpenAI 辩护律师: 这些全都是博眼球的场外闹剧，马斯克提出的诉求根本没有任何现行法律依据支撑 [30:22] Tim Draper: 我在低于 40 华氏度的温度下，在 52 分钟内听完了 52 场创业推介。欢迎来到我的办公室。

[3:53] Opus 4.6 fast mode... famously charges six times the price for two and a half times the interactivity... You're paying six times the price for two times the speed... but companies are paying disproportionately more for faster inference. [15:19] Alex Karp: Kevin looks like the guy from IBM

[0:03] [laughter] [0:06] >> Uh there's a ton of news. Let's start [0:08] with Cerebras. The IPO has gone [0:10] spectacularly well. Cerebras doubled [0:13] their valuation basically overnight. [0:15] Brandon Gorrell had the [0:17] good fortune of writing up some of the [0:19] details of the Cerebras news in the [0:22] newsletter today, tbpn.com. You can go [0:24] sign up. [0:24] >> Yeah, and right now it's sitting at a 64 [0:28] billion-dollar market cap. And a lot of [0:31] the prediction markets, they didn't even [0:32] have a category above 50, right? A lot [0:34] of people were just kind of [0:36] trading or or betting. [0:38] >> And when I wrote the newsletter Friday, [0:40] Monday, I I said a $50 billion IPO was [0:43] sort of being optimistic and it beat [0:46] those expectations, which is great news. [0:48] Chip design company Cerebras, if you [0:49] don't if you're not familiar, they make [0:51] a big big chip, big chip company instead [0:53] of [0:53] >> The biggest chip. [0:54] >> Instead of taking the wafer, putting a [0:55] bunch of chips on it, cutting it up into [0:57] a smaller chips, they use the whole [0:58] wafer. It's genius idea. It's one of [1:00] those simple ideas taken deadly [1:02] seriously in some ways. [1:05] But it's trading at $350 a share on its [1:08] first day of public trading, which [1:10] values the company much higher. [1:11] >> Three 300 now. [1:13] >> 300. [1:14] >> Yeah. [1:14] >> $300. Okay. The price in this IPO has [1:16] been literally up only. On Monday, the [1:18] price range was 150 to 160. Then they [1:21] raised it That was up from 115 to 125. [1:24] And today we're seeing, you know, much [1:27] higher prices. [1:27] >> to that picture. Someone should make a [1:30] set in LA. You know they have those like [1:32] fake private jet sets that [1:34] >> [laughter] [1:34] >> entrepreneurs [1:35] Imagine if entrepreneurs could have [1:36] [clears throat] a set where they put [1:37] their logo in the background, like [1:39] they're hitting it with a hammer and [1:40] there's confetti going everywhere. [1:42] >> Yeah, yeah, but it's for your course. [1:43] >> Yeah. [1:43] >> Yeah, [1:45] and and you walk right from there [1:47] to to the Lambo. [1:48] >> 1,000 students in your mastermind. [1:51] >> Yeah. No, I had this I had this idea [1:53] back in the day when Do you remember the [1:54] ice cream the ice cream museum, this [1:57] whole thing? [1:57] >> Oh, yeah. [1:58] >> So, there were there there was this tre [1:59] I mean, really bad news for the museum [2:01] industry, but they're getting eaten [2:02] alive. And so, some entrepreneurs, I [2:04] think they did very well, started at [2:06] something called the ice cream museum, [2:07] which uh was not really a museum in the [2:11] sense of like a presidential library or [2:14] uh [2:14] >> [snorts] [2:14] >> you know, the the Norton Simon or the [2:16] Getty or the you know, natural history [2:19] museum. It was more of like an [2:20] experiential place to go and hang out, [2:23] good for first dates, good for you know, [2:26] taking kids maybe. and they would maybe [2:28] give you some ice cream, but most of the [2:30] most of the museum was just like very [2:33] Instagrammable things. So, there would [2:35] be like a ball pit or a bunch of raining [2:38] confetti and stuff. And uh huge a huge [2:41] fabricated statue of ice cream that was [2:43] not a piece of art that would be sold. [2:45] Sprinkle ball pit. There you go. That [2:47] sounds real. I don't know if that's is [2:49] real, but it sounds very believable. [2:50] >> No, they have [2:51] >> They have that. Okay, yeah. And and [2:53] there were a number of other kind of [2:54] copycats that were trying to jump on and [2:57] do like, "Oh, we'll do like the waffle [2:58] museum or something or the pancake [3:00] museum." You know, cuz they just wanted [3:01] to cash in. And my idea was just the [3:03] museum of Instagrammable objects. And [3:06] so, it would have all of those. So, [3:07] there would be a private jet set and [3:08] then there would be a Lamborghini set. [3:10] And this one would fit right in. So, [3:12] it's just they have a big pink wall. So, [3:14] you can go take the pink wall photo. And [3:16] then there's a beach. And then there was [3:17] a gym with fake weights. So, you could [3:20] go and look like you're maxing out and [3:21] benching 500 lb. And so, it just says, [3:24] "Bring these clothes or we'll have them [3:25] for you." And then you move from room to [3:27] room taking the ideal dating profile [3:30] photo. And [3:31] Yeah, exactly exactly. Oh, you had kids? [3:33] Here, you in the hospital. You can live [3:35] an entire life through this fictional [3:37] museum of Instagrammable objects. Uh [3:40] more of a meme than a real business [3:42] idea, but [3:43] >> You know, John, the museum of ice cream [3:46] now has uh seven locations. [3:48] >> Okay, so they're cooking. [3:49] >> They're global. [3:49] >> They're global. They're They're doing [3:51] well. Anyway, let's go back to the [3:53] serious stuff. Cerebras, uh it's a [3:55] complicated uh company because they are [3:57] so deep in the AI supply chain, but [3:59] we'll break it all down for you. So, [4:00] there's a bunch of interesting [4:01] takeaways, some really solid positives. [4:03] Uh Cerebras chips work, which was [4:05] something people were not expecting for [4:07] a while. There was a lot of fud around [4:08] this company, just the idea of like, [4:10] "Oh, that'll never work. What if the [4:11] architecture changes? What if we go away [4:13] from transformers or something? What if [4:15] we need something quote completely [4:16] different?" Or maybe like the yields [4:18] will never work because there was this [4:19] idea that if you're using the entire [4:21] wafer, typically as you're etching the [4:24] chips onto the wafer, sometimes there's [4:26] little defects. And it's not a problem [4:28] if you're going to break up a wafer into [4:29] like 64 chips cuz you just throw away [4:31] one. But if there's one defect on [4:33] basically every wafer, well, then your [4:35] yield is going to be super low. We [4:37] talked to Andrew about how he solved [4:39] that by uh [4:40] creating redundant cores, and they don't [4:42] actually activate all the cores, and so [4:44] they sort of built in that redundancy [4:45] and got through that. But that was an [4:46] early critique of the strategy. [4:49] >> Uh so, they can use [4:51] >> You can use it. [4:51] >> use Cerebras chips today as [4:54] >> In Codex, 5.3, Spark. Uh and so, they [4:57] are very fast, and I think the most [4:59] important thing that SemiAnalysis points [5:01] out is that token consumers, customers, [5:03] businesses have shown this revealed [5:06] preference for and and a willingness to [5:08] pay for speed. And they sort of [5:09] contextualize it, and they quantify it [5:11] based on their own usage and their [5:13] experience with Anthropic's Opus models. [5:15] So, Opus 4.6 fast mode famously I like [5:19] that they use famously cuz it's like [5:21] famous to like 100,000 people. But uh [5:24] famously charges six times the price for [5:28] two and a half times the interactivity, [5:30] although it's now under two x faster. [5:32] So, effectively, you're you're you're [5:34] paying six times the price for two times [5:36] the speed. That's uh that that's [5:38] disproportionately more money for what [5:41] you're getting. You would think you'd [5:42] pay six times the price for six times [5:44] the speed, potentially. Um but there [5:46] were a lot of questions about would [5:47] people really pay for more that much [5:50] more for faster models, faster [5:52] inference? And Andrej Karpathy, Sam [5:54] Altman was was was saying like, "Do you [5:55] want faster models or smarter models?" [5:58] And uh he was like, "I think And Sam's [6:01] point was sort of like the these models [6:03] are very intelligent, but using them [6:05] faster is [6:07] sort of more of a magical superpower." [6:09] And Sam was I I felt like Sam was sort [6:11] of leading it towards like speed is [6:12] really important as the next leg up on [6:14] productivity. And Andrej Karpathy was [6:16] like, "No, I just want smarter. I'll [6:17] just let it run overnight. I don't mind [6:19] that." But that's not what everyone is [6:21] feeling. Some people, especially the [6:23] semi-analysis team, leaned more towards [6:25] interactivity or speed over raw [6:28] intelligence power. [6:29] >> Well, yeah, and then there's the [6:31] other aspect, which is just capability. [6:33] Capability, speed, and intelligence. [6:36] That's like the question people have had [6:38] is like, "Okay, what is Is there a 250 [6:42] IQ model, or is there just a much more [6:44] capable model?" [6:45] >> Yeah, the unhobbled one. [6:46] >> tools more efficiently and is really [6:48] quick? [6:48] >> Yeah. And that's actually important to [6:50] Cerebras. Semi-analysis was spending 80% [6:52] of their AI spend on Opus 4.6 fast, and [6:56] so they were willing to pay that 6x, [6:58] like 80% of their spend, [7:00] disproportionately more, even though [7:03] when their sort of expectation as they [7:05] put it was that they would always want [7:06] the smartest model. They would [7:08] they would they would be very [7:09] cost-conscious. They were in reality [7:12] saying, "I'm going to hammer fast mode. [7:13] I want to spend on fast mode." And then [7:15] I think the price was significant, and [7:17] so there's probably sort of a [7:18] renegotiation about when is the right [7:20] time to use fast versus when do you want [7:22] to leave something running overnight? [7:23] But OpenAI is clearly very peeled on [7:26] Cerebras. Cerebras has a big 750 MW deal [7:29] with OpenAI, and the chips are already [7:30] serving GPT 5.3 in Codex under the name [7:33] Spark, as we mentioned. And I've used [7:35] it, you should use it. It's a very [7:37] interesting experience because I think a [7:39] lot of people have interacted with LLMs [7:40] and chatbots, and they're sort of used [7:42] to this token streaming in. It's sort of [7:44] cute because the phone vibrates, and it [7:46] and it feels like you're talking to [7:48] someone who's typing, but it's way [7:50] better when you just land on a Wikipedia [7:51] page, the whole thing loads, and you can [7:52] just scroll however you much you want. [7:54] And that's the experience that I think [7:55] people want and will demand across [7:57] everything, especially if they're firing [7:59] off a coding task. They just want the [8:00] code immediately. You can also just go [8:02] talk to the model like it's ChatGPT. You [8:04] don't need to use Codex 5.3 Spark in a [8:06] coding context. You can ask it whatever [8:08] you want, and it will act like a normal [8:11] LLM. I personally think there will be [8:13] huge demand for faster inference across [8:15] all parts of the AI economy. There's [8:17] this old [8:18] >> Yeah, another another way to think about [8:20] it is like if you have two employees [8:22] with the same skill set, the same [8:24] capability, but one is just five times [8:27] faster, right? That person can create [8:30] way more value in the organization, [8:32] right? [8:33] >> And for a lot of things, if they're if [8:34] they're two times faster, they do [8:37] command six times the price. A sales rep [8:38] that that that sells twice as much, or [8:42] someone who's twice as effective as [8:43] their job, might actually command a [8:45] salary that's five times, six times [8:47] the the actual price. And so, there's [8:49] lots of other contexts across different [8:52] business lines that you could draw to. [8:54] There's also this old adage or saying [8:57] about e-commerce that might may or may [8:59] not be real, but it's probably been [9:00] transposed so many times in think [9:02] pieces. I don't know the real quote, but [9:03] it goes something like every 100 [9:05] milliseconds of latency costs Amazon 1% [9:09] in sales. I don't know if that's the [9:11] right the way to think about it, but [9:12] basically as Amazon was scaling, they [9:14] realized that there were a bunch of [9:16] things that they could do on the UI [9:17] side, a bunch of things they could do on [9:18] the layout side. Where does the buy [9:20] button go? Where does certain [9:22] information go? The price, the discount, [9:24] all of this stuff, the images. They they [9:26] were tweaking the front end, but as they [9:27] did that, they added bloat, and the [9:29] pages would slow down. And what they [9:30] noticed was that the slower the page [9:33] was, the lower the conversion rate [9:34] because people were waiting for [9:36] amazon.com to load, click on the page, [9:38] it takes a second, they get distracted, [9:40] they go somewhere else. And I think that [9:42] that's happening in LLM use cases all [9:44] over the place. People fire off a query [9:46] and they're like, "Ah, it's taking too [9:47] long. I'll go scroll Instagram Reels." [9:48] There's always an Instagram Reel. And [9:50] they'll be like, "Ah, I for- kind of [9:51] forgot about what I was asking about. I [9:53] didn't get my answer." And that's [9:54] certainly true in business context as [9:56] well. This is currently playing out in [9:57] AI inference. Uh companies are paying [9:59] disproportionately more for faster [10:00] inference, and this is good for [10:01] Cerebras. But SemiAnalysis does point [10:04] out a number of potential headwinds and [10:06] problems that the team at Cerebras will [10:08] have to solve or contend with over the [10:10] next few years. Uh mainly, Cerebras [10:13] chips are not currently as capable of [10:16] holding larger models in the limited [10:19] memory that they have or networking [10:20] multiple chips together to serve larger [10:23] models. We've heard about the NVL72 [10:25] racks that wire a whole bunch of Nvidia [10:27] chips together can serve these really [10:29] large models. Uh that has potentially be [10:32] been a challenge. So, SemiAnalysis says, [10:34] "Moreover, the industry is tending [10:36] trending towards larger context windows [10:40] ad infinitum. 128K context will [10:43] certainly not be acceptable for long, [10:45] especially with the prevalence of [10:47] agentic workloads. And it doesn't look [10:49] like there's a simple solution of just [10:51] scaling the wafer size larger because [10:52] TSMC is set up with a standard wafer [10:54] size or adding more memory to the [10:56] existing architecture because Cerebras' [10:59] whole design depends on a lot of SRAM, [11:03] static random access memory, directly on [11:05] the on the wafer. But SRAM is no longer [11:07] shrinking as much with each new [11:09] semiconductor node. So, the last version [11:12] of the Cerebras chip, they've done WSE1, [11:14] 2, and 3. They're on 3 now. But WSE2 had [11:18] 40 gigs of memory. WSE3, [11:21] you would expect, "Oh, we want a [11:22] doubling, right? We want a we want a 10x [11:24] or something." Uh it got 44. So, a 10% [11:27] increase over one process node, one [11:30] iteration. Is there an easy way to [11:33] double this? Is there question like how [11:35] will this scale as the models get bigger [11:38] to add more SRAM? You might have to [11:39] sacrifice compute area because [11:41] everything is being done on one wafer. [11:43] If you want computation or memory, [11:46] there's a direct trade-off because you [11:47] only have so much space on the actual [11:48] wafer. But in an agentic workflow, I [11:52] think it's [11:53] entirely possible that you want like the [11:54] biggest, most powerful model, like the [11:57] vice president delegating things that [12:00] you want the vice president to handle [12:01] >> Senior or junior? [12:02] >> Senior vice president. Uh maybe just the [12:04] president uh handling the critical work. [12:08] Uh so future models might not And that [12:09] might not be on Cerebras, that might be [12:11] on NVL-72 or TPUs or something. But uh I [12:15] imagine that we will quickly jump from [12:17] the agentic age where you're firing the [12:19] best, smartest model at the full [12:21] workload to the orchestration age where [12:24] and there will be hybrid approaches. So [12:26] the biggest and best models will [12:27] delegate certain tasks to smaller, [12:29] faster models, just like they go and do [12:31] database queries these days. Or they go [12:33] and search the web these days and that's [12:34] CPU bound. There will be certain [12:37] workloads that the larger, smarter agent [12:40] model, like the boss model, can sort of [12:42] delegate to the Cerebras speed workers, [12:45] the faster workers. A year or two ago [12:47] when uh Daniel Gross wrote AGI bets and [12:50] was sort of like is Nvidia underpriced. [12:52] I don't know if Nvidia He's managed to [12:53] say that on on on Stratechery, but you [12:56] know, we we entered the AI age and [12:57] everyone was like, "Oh, GPUs are the [12:59] future. Nvidia is the company." But then [13:01] it was like Nvidia GPUs are good and [13:03] then also CPUs are good. And and Arm is [13:06] getting into it and Intel's doing very [13:08] well. And it it worked out that [13:10] >> big computers. [13:11] >> Big computers. Big computers, for sure. [13:13] Honam says, "This IPO illustrates the [13:15] power of an individual partner over the [13:17] brand name of the firm. Pierre Lamond [13:19] was a partner at both Sequoia and [13:21] Khosla, but instead of those firm [13:22] backing Cerebras, it was Eclipse, the [13:24] firm he joined at the age of 84, that [13:28] this little known chip company multiple [13:29] times in the early days. What a way to [13:31] wrap up a career. He was born in 1930, [13:33] the same year as Warren Buffett. [13:35] >> Wow. [13:36] >> That is an awesome story. I love that. [13:38] 1/3 of the order book, the folks that [13:41] said I want shares in the Cerebras IPO, [13:44] 1/3 of the book got zero. I guess the [13:46] top 25 investors took 60%. That's [13:48] probably the big investment funds, the [13:51] Fidelity's, the State Street's, the [13:53] >> They have done quite well today. [13:56] This picture looks wildly different than [13:59] the Klarna IPO last year in which only a [14:03] handful of the team at Klarna popped [14:05] over, hit the NYSE IPO'd, went back back [14:08] home. [14:09] >> Yeah, it was very much just like another [14:11] day at the office for the team. [14:14] Yeah, that's definitely what I was [14:15] contrasting it to. The Cerebras [14:16] valuation every round, series A in 2016, [14:19] $100 foundation, Benchmark and Eclipse. [14:22] Coatue led the series B in 2016. VY [14:25] Capital led the series C in 2017. Then [14:29] 1.6 billion valuation in 2018, 2.4 in [14:33] 2019, 4 billion in 2021. That was like [14:36] maybe a little bit of a slump, but then [14:38] 2025, Andreessen Horowitz and Fidelity [14:39] come in at 8 billion. Then Tiger [14:41] >> comes in at 23 billion dollars. [14:45] >> Then in May of 2026, they IPO [14:49] 48.8 billion. Let's run through [14:53] the Kevin Warsh news because he has been [14:56] confirmed as the Fed chair. Kevin Warsh [14:59] who is most famous for interviewing Alex [15:03] Karp on CNBC [15:06] while Alex Karp appeared to have popped [15:08] a nicotine pouch and then spun a [15:10] notebook on his finger. Did you ever [15:13] find that clip, Tyler? Is that in the [15:14] >> Yeah, it should be [15:15] >> Okay. [15:16] Let's [15:17] >> Yeah, we have the the video here from [15:19] >> really put Kevin Warsh on the map. [15:22] >> I remember I showed up in your office [15:23] once. I was dressed like this and I [15:25] think you screamed at one of the guys. [15:26] You said, "Kevin's here. He looks like [15:28] the guy from IBM." And I was talking [15:30] about, well, you know, we need like [15:31] really finance controls and you know, [15:34] how you're going to sell the product and [15:36] all this stuff. [15:37] >> Okay. [15:37] >> But I would say you certainly built that [15:40] as [15:40] >> He's really spinning it. I didn't [15:41] realize he goes back to it like four [15:43] times and he keeps spinning it. [15:45] He's really good at this. [15:46] >> But somehow you grafted that on to the [15:48] to to the strange company that can [15:50] produce these products. How's that [15:51] transition been if I've got it right? [15:53] >> I have so many questions. [15:55] >> [laughter] [15:55] >> First, we have to get him to recreate [15:57] that for sure. Second, [15:59] I I thought Tyler, I thought we were [16:01] talking about that being on CNBC, but [16:02] that looks like just a podcast. Like [16:04] that doesn't have any Chiron or [16:05] >> Yeah, no. I I don't think it was [16:06] actually on. I think it was from [16:07] Palantir. Like that was a Palantir [16:09] >> Oh, okay. So, it was just like a random [16:10] podcast. And then and then I when I've [16:12] seen it on CNBC, they were playing the [16:14] clip. Got it. Okay. [16:16] >> I think so, yeah. [16:16] >> The vote was 54 to 45 in the Senate. The [16:20] divided vote signals challenges ahead [16:22] for Warsh who faces a Fed committee [16:24] skeptical of rate cuts that Trump has [16:27] demanded. And of course, we talked about [16:28] the inflation news. Typically, you don't [16:31] cut interest rates going into inflation [16:33] and potentially economic stagnation. [16:36] You definitely don't cut rates in That's [16:39] why stagflation is so difficult because [16:41] if you have stagnation and low [16:44] inflation, you can cut rates very [16:45] easily. Maybe the economy starts [16:48] overheating a little bit. You get a [16:48] little bit of inflation, but then you [16:49] can pull back. That's what we've done [16:51] historically. Vice versa, if the [16:53] economy's running hot, you're seeing [16:54] high GDP growth and high inflation. [16:57] Well, if you raise rates, you're going [16:59] to pull back on both of those. But in [17:00] stagflation, you're seeing both [17:02] inflation and [17:03] economic stagnation, harder to deal with [17:06] as a Fed chairman, which is potentially [17:08] the task he will be faced with. So, [17:10] the Senate confirmed Kevin Warsh as the [17:12] Federal Reserve's 17th chair Wednesday [17:15] in a largely party line vote that [17:17] require that reflected the how tensions [17:19] with the White House have dragged the [17:20] Fed deeper into the political fray. I [17:23] was looking back at the old Fed chairs. [17:25] There's some absolutely legends in there [17:28] cuz some of them have really long runs. [17:29] So very quickly you get back to the [17:31] black and white portrait and the [17:32] painting as you go back in time. [17:34] >> Who's your favorite Fed chair? [17:35] >> Mhm. [17:36] >> Volcker? [17:37] >> Yeah, Volcker is pretty good. And [17:39] Bernanke is pretty good. [17:40] >> dog. [17:41] >> Yeah, I don't know. Hard to pick. Hard [17:42] to pick. Chair Jerome Powell, whose [17:44] leadership tenure whose leadership [17:46] tenure ends Friday captured at least 80 [17:49] votes in Senate confirmations for each [17:51] of his two terms at top of the Fed. Wow, [17:53] Jerome Powell just fan favorite of both [17:57] teams. 80 votes in the Senate. That's [17:59] pretty significant. I'm putting him in [18:01] the conversation Jordyn, but I'm not [18:03] giving him the goat trophy. Uh [18:05] >> Yeah, I'm not saying I'm not saying [18:06] that. [18:07] >> the the challenges faced he wasn't [18:09] confronted with a great recession, a [18:11] dot-com bubble bursting, a [18:14] uh a Black Friday. Uh he like he he like [18:17] the economy from 2018 to today [18:20] >> Global pandemic doesn't you don't you [18:22] don't count a shutdown of [18:24] >> No, because No, no, no. I I I I actually [18:27] don't because the economy was pretty [18:30] strong in 2019 and it went into it went [18:33] into 2020 with uh pretty strong consumer [18:36] balance sheets, low debt. There wasn't a [18:39] shadow banking economy. There was no [18:41] bomb in the US economy waiting to [18:43] explode. And so although we saw high [18:47] unemployment briefly and we did have to [18:49] stimulate the economy, that's not his [18:51] job. His job was to set rates. There was [18:53] a little bit of like the I mean maybe [18:55] you put the inflation, you know, this [18:58] the end the ZIRP era and the end of the [19:00] ZIRP era and all of those gyrations on [19:02] him, but those the problems that were [19:04] downstream of both the ZIRP era and the [19:06] end of the ZIRP era were suffered mostly [19:10] and benefited mostly on like tech [19:12] companies and Silicon Valley companies [19:13] that had really long cash flow horizons. [19:16] And so there was not a moment where it [19:18] was a dire situation that the Fed had to [19:22] intervene in a meaningful way and like [19:24] save the economy like in 2008. It it [19:27] it's a big deal. He did great job, but [19:29] he didn't he wasn't faced with the same [19:31] challenges of a Bernanke, for example. [19:34] That's what I would say. Tyler, what do [19:35] you think? [19:36] >> Yeah, I think that [19:37] that's reasonable, but also like if [19:39] Powell was worse at his job and he saw [19:40] some crazy crash because of COVID and [19:42] then he brought it back. Like then it'd [19:44] be like, oh yeah, he did face this [19:45] massive thing. But because he did, you [19:47] know, such a good job, maybe you didn't [19:49] see any like massive crash. Nothing [19:51] super bad happening is evidence that he [19:52] was really good as a Fed chairman. [19:54] >> Uh yeah, yeah, maybe. He's a defensive [19:57] back. Uh you you know, if they don't [19:59] score, you there's no great plays cuz [20:01] he's just shut down shut down [20:03] cornerback. Anyway, Jensen Huang is over [20:05] in China. Jason Calacanis has a photo [20:08] that looks [20:09] uh extremely real, zero AI detected, but [20:12] he's bringing two huge boxes of GeForce [20:14] RTX 5090s, which are [20:15] >> This is a picture from when he was in [20:16] Alaska, too. [20:18] >> [laughter] [20:19] >> Uh Jason says never stop selling. I I [20:22] agree. Uh there is some news, which we [20:24] will cover later uh in the week around [20:27] uh the dynamics around H100 sales uh and [20:30] Blackwell's what's actually happening. [20:31] It's all in flux as the Trump China [20:34] summit uh plays out on the front page of [20:38] the Wall Street Journal every day this [20:39] week cuz it is headline news. [20:41] High-stakes US-China summit kicks off. [20:44] Watch a team of humanoid robots running [20:47] a full 8-hour shift at human performance [20:50] levels. And Bret Adcock said this is [20:52] fully autonomous running Helix 2. [20:55] >> All right, pull up pull up this post [20:58] from Pete. [20:59] >> Yes. And the the stream did [21:01] fantastically. It was 24 hours, it got [21:03] 3.4 million views. But at a certain [21:06] point during the stream, there was some [21:09] questions about whether or not the [21:11] humanoid robot was in fact [21:13] >> to the beginning. Back to the beginning. [21:15] >> Okay. Let's play this. [21:17] >> All right. So, it's cooking. I mean, the [21:18] speed is actually insane. [21:19] >> were extremely impressed by this. This [21:21] was remarkable. Remarkable. [21:22] >> Even if it's teleoperated, it's [21:24] extremely impressive. [21:25] >> Yeah. Yeah. Yeah. Like, the robot's [21:26] clearly working. This is very very [21:28] >> But, they're saying that it's not [21:29] teleop. [21:29] >> Okay. So, then the robot starts missing [21:31] things, being a little bit like an inch [21:33] off, and then reaches up and touches the [21:35] robot's head. The robot, which is [21:38] something that wouldn't normally be [21:39] necessary. It doesn't have like a [21:40] logical explanation or conclusion. And [21:43] so, a lot of people are asking [21:44] >> It does have a semi-logical conclusion, [21:47] which is that Brett is claiming when it [21:49] reaches across its body to go to the [21:52] right that it puts its hand up here to [21:54] get the hand out of the way. [21:55] >> That's what I was thinking, was that if [21:56] the hand is is halfway up, [21:59] >> To go back to the beginning. [22:00] >> the sensor, the camera sensor. And so, [22:03] even though you like you might the robot [22:04] might reach the hand up further to move [22:07] out of the view, so that the robot can [22:09] look at the next package. So, that's one [22:13] possible explanation. But, a lot of [22:15] people are asking even harder questions, [22:19] saying that potentially was there a [22:21] human in the loop? Was this [22:23] teleoperated, which is something Brett [22:26] has said it's fully autonomous. I feel [22:28] like that means no humans in the loop, [22:30] but TORC Robotics has an artist's [22:33] representation of Helix 2 figures [22:35] in-house neural network running entirely [22:37] on board. And it, of course, is a human [22:41] in a VR headset. Very [22:44] very debatable. We'll let We'll We'll [22:46] see where you stand. But, there is There [22:48] is a third option, which I have shared, [22:51] which is potentially no humans involved. [22:55] I don't know if you'd call it [22:56] autonomous, but you would call it no [22:58] humans in the loop. [23:00] >> Because you have [23:01] >> Well, it is an autonomous system, right? [23:03] It just sort of runs. [23:04] >> Yeah, I would I would consider this [23:06] autonomous. It's the It's the image that [23:08] I shared in the production chat. It's [23:10] not of a human, and it's not quite [23:12] robotic, but there's no human in the [23:14] loop. And so, this could explain the [23:16] system is running with no humans in the [23:18] loop. If you make that claim, and you [23:20] follow this, [23:21] I think this qualifies as no humans in [23:23] the loop. If you have a giant orangutan [23:25] in a VR headset puppeteering the robot [23:28] via teleoperation, you could say that [23:30] this system is does not have a human in [23:33] the loop, and you could make that claim. [23:34] >> And I could make the argument that it's [23:36] autonomous. The chimpanzee is running [23:38] its own It has somewhat of a neural [23:40] almost like a neural network. [laughter] [23:43] >> The OpenAI Elon Musk trial is in its [23:46] final day. The trial is ending. People [23:49] expected 4 weeks of trial. We only got [23:51] three. They're cutting it short. Uh what [23:54] are the prediction markets saying about [23:56] who's going to win? I want to know that, [23:58] and I want to go to Mike Isaac, the rat [24:00] king, because he has a breakdown of [24:02] what's going on. He says, "Good morning. [24:04] Closing arguments of Musk versus OpenAI [24:06] with special guest Microsoft are [24:09] happening today. [24:10] >> The Cal sheet, will Elon win his case [24:12] against OpenAI? It peaked at a 58% [24:16] chance. [24:16] >> Okay. [24:16] >> On the 28th, it's now sitting at 30% [24:19] chance. [24:19] >> 30% chance. Okay. Uh so, right now, the [24:22] judge is instructing the jury on the [24:23] criteria by which they should be judging [24:26] the outcome of the case. Important, [24:27] because if the jury listens and carries [24:29] this out, it is a very, very specific [24:30] lens through which they view all the [24:32] evidence. Ostensibly, it's where theater [24:34] ends. Listening to this and being read [24:36] out in court in the for last 23 30 [24:38] minutes is very helpful, because it's [24:39] clarifying on how high the bar is for [24:42] the plaintiff's side of proving some of [24:44] these claims. [24:45] Uh sort of feel bad for the AV guy [24:47] during this trial. There's been [24:48] feedback, been mic drops, but not in the [24:51] good way. The mics have been dropping [24:52] out, vonky video feeds. They need to [24:54] revamp this place, says Mike Isaac. [24:56] LMAO, the first joke of the tweet storm. [24:59] He says Musk counsel is going after [25:01] OpenAI execs, Altman and Brockman, and [25:03] has the mugshot style photo of Altman on [25:06] the screen again. Battle of Photoshops [25:09] of executives in this trial has been [25:10] entertaining to watch. You want to [25:12] depict your opponent in and the worst [25:15] possible light. Musk counsel going back [25:17] and forth hammering the point they made [25:18] over and over. The argument essentially [25:21] painting a picture Sam Altman, liar. [25:23] Chipping away at witness credibility has [25:24] been core strategy for the plaintiff [25:26] side. And we're back [snorts] to [25:27] everyone hates Google again. Molo is [25:30] using Larry Page who they claim doesn't [25:32] care about humanity as a foil to the [25:34] noble Musk who only who's only care with [25:37] respect to AI is the future of humanity. [25:39] Musk counsel is painting the don't don't [25:41] trust Sam picture in a bit more detail [25:43] for the jury. Also Musk's side has a [25:46] picture of Elon and Altman on the screen [25:47] now. Sam's looks like he's about to be [25:49] processed by a US Marshall. Musk's [25:51] looked like he's getting ready for the [25:52] Met Gala. LOL. Lots of Musk closing side [25:55] arguments semi-populist track of [25:57] pointing open pointing at OpenAI and [25:58] saying these billionaires are making [25:59] gobs of cash while running a charity for [26:01] the supposed good of the world. I'm [26:03] curious if jury can register this [26:05] argument even if it comes from Elon [26:06] Musk, the world's richest man. Ouch. [26:08] OpenAI counsel begins closing argument [26:10] with a broadside against Musk. Every [26:12] even the people who work for him, even [26:13] the mother of his children can't back [26:15] his story. Oh yeah, back to the war of [26:18] the Photoshops. OpenAI closing remarks [26:20] now in in the digital displays in the [26:22] monitors for exhibits, all the OpenAI [26:24] executives look like Olan Mills [26:26] photoshoots. Do you know who Olan is he [26:28] says it's complimentary. I need to get [26:30] up to speed on my photographers. Olan [26:33] Mills [26:33] >> Olan Mills is a portrait offers portrait [26:36] photography. [26:37] >> Ooh, it does look very nice if you pull [26:39] up the the Google images on Olan Mills. [26:42] Anyway, short summary of the closing. [26:43] Musk camp, all these OpenAI executives [26:45] are rich as hell and lying all the time. [26:47] OpenAI camp, all of that is a side show [26:49] and literally all the claims Musk is [26:51] bringing cannot be stood up by actual [26:53] law, the Microsoft camp disappears into [26:55] bushes. Dota got mentioned again. [26:58] >> [laughter] [26:58] >> They love mentioning Defense of the [27:00] Ancients. Incredible Photoshop from the [27:02] OpenAI camp of a calendar of events [27:04] complete with little characters and a [27:06] timeline of events. I wonder if they're [27:07] using ImageGen too. [27:09] Or if they're doing it the old-fashioned [27:10] way. I can't wait until it's entered [27:12] into evidence this afternoon so he can [27:14] show us. [27:15] Sort of want to buy this meme guitar, [27:18] but I also have two telecast is that a [27:20] completely side side note. [27:22] Gamer has entered the blog. The Dota [27:25] moment has been mentioned [laughter] [27:26] nearly every single day during this [27:28] three-week trial. AI researcher, we got [27:30] to have Mike back on the show. So good. [27:33] So it was a true breakthrough in the [27:34] technology. [27:35] >> What what is the timeline for the jury [27:38] to meet? [27:38] >> Okay. [27:39] >> Are they Is this [27:40] something they're doing today? [27:42] >> a 30-minute recess, most they've had in [27:44] a month. I might actually be able to go [27:46] outside and get real food. There's a [27:47] Popeyes across the street. Is it a bad [27:50] idea to get a bucket of red beans and [27:51] rice? That's what he's thinking about [27:52] doing. So [laughter] not much news on [27:56] when this will close. It is 1:10 Pacific [27:59] time. I imagine that they will wrap up [28:02] by What did he say? 3:00 p.m.? 4:00 [28:04] p.m.? So 30-minute break, that happened [28:07] 40 minutes ago. [28:08] So I imagine that and [28:10] >> But they've been taking Fridays off is [28:12] kind of what I'm getting at. [28:13] >> Oh, yeah. [28:14] >> Because this could very [28:15] >> this happens on Monday. This is just [28:17] closing arguments, not necessarily the [28:18] end of the trial. [28:20] >> Or the jury [28:21] >> might get the results. [28:21] >> Or the jury might might make a quick [28:24] call, but [28:25] >> Well, there was an update [28:26] >> unlikely. [28:27] >> 11 minutes ago. A lawyer for OpenAI on [28:29] Thursday defended the company's chief [28:30] executive Sam Altman from withering [28:31] character attacks by Elon Musk's legal [28:33] team as both sides delivered their [28:34] closing arguments in a trial with [28:36] potentially seismic implications. The [28:38] stakes are high. Mr. Musk, who was not [28:41] in the courtroom on Thursday because he [28:43] was in China with President Trump, is [28:45] asking for more than $150 billion in [28:47] damages. He is also asking the court to [28:49] remove Mr. Altman from the startup's [28:51] board and to stop a shift the company [28:53] made last year to operate as a [28:54] for-profit company. They pushed back. [28:56] Sarah Eddy, member of OpenAI legal team, [28:58] tried to in her closing argument to dull [29:00] the attacks on Altman's credibility and [29:02] to argue that there was never a firm [29:04] agreement among the founders that could [29:05] have been breached, not one in this case [29:07] other than Elon Musk has testified to [29:10] any commitments or promises that Sam [29:12] Altman or Greg Brockman or OpenAI made [29:14] to Mr. Musk is what she's saying. After [29:16] the recess, William Savitt, OpenAI's [29:18] lead counsel, told the jury that Musk [29:20] does not have a claim against the [29:21] startup unless there was a specific [29:23] agreement between Musk and OpenAI [29:26] describing how his donations to the [29:28] nonprofit should be spent. That [29:30] agreement does not exist, Savitt said. [29:33] So, that's where, I guess, OpenAI is [29:34] leaving it for now. [29:36] We will continue to cover the story as [29:38] it evolves. [29:39] >> the jury allowed to use Codex {slash} [29:41] gold in [29:42] >> [laughter] [29:43] >> 1 and 1/2 hours?" [29:44] >> There's other tech problems going on. [29:46] Max Zeff over at Wired has been covering [29:48] the story as well and says Musk's lawyer [29:51] brought a big monitor, maybe 36 inches [29:54] into the courtroom. OpenAI's lawyers [29:56] asked to use it. Musk's lawyer said, [29:58] "No." The judge told Musk's lawyers that [30:01] they have to let OpenAI use it. Then [30:04] OpenAI said it might not be possible to [30:06] connect their laptops [laughter] to it. [30:08] AGI is here, but we'll still need a [30:10] dongle, I suppose. A dongle has entered [30:12] the courtroom, Max Zeff says. [30:14] There's about 50 15 lawyers standing in [30:17] the middle of the room right now talking [30:18] how to use talking about how to use this [30:20] big monitor. [30:21] This is wild. [30:22] >> In other news, Tim Draper says, "I think [30:26] I broke a record. I took 52 pitches in [30:29] 52 minutes at below 40°. [30:32] Welcome to my office. #DraperUniversity, [30:35] #SurvivalTraining. [30:38] What do we think about going in the ice [30:41] tank?" [30:42] >> How cold are ice baths typically? You [30:44] you you [30:45] done ice ice baths. I feel like I did [30:47] one and it wasn't as [30:51] insanely difficult as people said, but [30:52] then I checked the temperature and I [30:53] don't think it was 40. I think it was [30:55] closer to 50. [30:57] >> can totally get closer to I I I put um [31:00] >> Cuz there's a couple companies that sell [31:02] ice baths. [31:02] >> when when if you're going surfing and [31:05] the water is below 45°, it can just be [31:08] very painful. [31:09] >> Okay. Yeah, yeah. [31:10] >> So even in a wet suit and anywhere [31:13] that's not covered, a lot of people are [31:14] putting gloves on [31:16] or booties. [31:16] >> think, Tyler? [31:17] >> So apparently Joe Rogan's at like 34. [31:19] >> 34? [31:21] >> Yeah. [31:21] >> Wow. [31:22] >> like the cold plunge, you know. [31:23] >> But he's [31:24] he's the top of the mountain when it [31:25] comes to ice baths. [31:26] >> He's the final boss. [31:28] >> Uh that's what I was going [31:29] >> yeah, this this is just a crazy picture. [31:31] I did think it was I did think it was [31:33] AI, but turns out it's real. It's just [31:35] funny because it looks like like what is [31:37] this set What is this set up? [31:39] >> What are all the trash bags there? And [31:42] the wall is like sort [clears throat] of [31:43] like [31:44] >> And there's piping in there. It looks [31:45] like kind of like a prison ice bath. [31:47] >> Yeah, this is not what you'd expect from [31:49] I mean, isn't he a billionaire investor? [31:52] You'd expect some sort of palatial, you [31:54] know, you see the the the the the the [31:56] properties that Mark Zuckerberg's [31:58] acquiring, the big investors are [31:59] acquiring. You would expect something [32:01] that would be [32:03] much more regal, [32:04] but he's doing it the old-fashioned way. [32:06] Whipped this up himself, bought some [32:08] trash bags, [32:09] >> [laughter] [32:10] >> and took some pictures. [32:12] Uh yeah, that's our show, folks. Leave [32:14] us five stars on Apple Podcasts and [32:17] Spotify. [32:18] >> Another one. [32:19] >> Sign up for our newsletter at tbpm.com. [32:21] See you tomorrow at 11:00 a.m. Pacific [32:24] time and have a great rest of your day. [32:26] Goodbye.