Atreides' Baker Sees AI Stock Sell-Off Masking Compute Price Surge

Atreides的Baker：AI概念股抛售掩盖了算力价格飙升 公开市场AI股票的下跌反映出市场对开源模型普及的误读，而底层的算力现货费率、DRAM价格和Token需求实际上正在加速增长。 Gavin Baker深入剖析了7月份AI概念股抛售潮背后的动态，认为公开市场误读了暂时的信贷与开源叙事恐慌，同时低估了实体GPU利用率、现货租赁价格以及长期供应协议的加速势头。他探讨了模型路由的经济学、支出从人力薪酬向Token消耗的转变，以及Claude算法如何像现代“Walter Cronkite”一样驱动市场情绪循环。此外，Baker还讨论了软件的防守壁垒、围绕数据中心的监管风险，以及SpaceX大规模算力扩张与轨道基础设施这一新兴前沿领域。 尽管 7 月公开市场 AI 概念股遭遇 40% 至 60% 的剧烈抛售，但硅谷实地的底层核心指标——包括 GPU 可用性、租赁现货价格和 Token 生成量——均在全速加速增长。 (2:35) 此前长约锁定的存量 GPU 算力价格目前大幅低于现货费率；随着旧合约陆续到期并按更高市价续约，将自然拉动超大规模云厂商的营收加速与利润率提升。 (5:06) 市场对科技行业利差走阔与债务违约风险的恐慌脱离了实际；若 Blackwell 和 Rubin 架构实现合理的商业化变现，各大云厂商将获得约 2 万亿美元的运营现金流，基本无需依赖外部债务融资。 (11:54) Claude 等 AI 总结工具已成为股市中类似“沃尔特·克朗凯特”的权威声音，通过算法统一了投资者的情绪与新闻解读，导致市场波动与供应链叙事周期被极度压缩。 (21:35) 顶尖 AI 原生初创公司正将 20% 至 50% 的人力薪酬预算转向 Token 消耗，表明直接由算力替代高薪知识劳动的潜在市场规模已达数万亿美元。 (30:51)

Gavin Baker argues that public market sell-offs in artificial intelligence reflect an algorithmic sentiment disconnect rather than fundamental deterioration, as physical compute pricing, cash flow generation, and token demand continue to accelerate. He contends that expanding operating cash flows and open-source inference routing will self-fund infrastructure expansion, leaving regulatory and political pushback as the primary threat. Market Sentiment Disconnect Versus Hardware Fundamentals (0:00) Baker observes that public AI equities fell 40% to 60% during a sharp July drawdown, driven by narrative shifts rather than deteriorating fundamentals. He attributes this heightened volatility partly to algorithmic homogenization, noting that institutional investors feeding news into models like Claude have turned the tool into an algorithmic 'Walter Cronkite.' This shared consensus compresses multi-year supply-demand cycles into weeks, as seen when Japanese capacitor equities completed a three-year cycle in six weeks without changes to baseline operations. In contrast to public pessimism, Baker reports that physical compute metrics in Silicon Valley are accelerating. Legacy contracted compute currently trades at a discount to the spot market, meaning expiring multi-year contracts will reprice higher. Spot rates for Nvidia Blackwell B200 GPU clusters have jumped 50% to 60% over a six-month period, while inference cloud providers report expectations of paying up to 100% more upon contract renewals, supporting revenue and margin growth. Hyperscaler Self-Funding and Novel Compute Financing (10:51) Baker disputes Wall Street fears that rising real yields and wider credit default swap spreads will derail the capital-intensive AI build-out. Consensus financial models assume future Blackwell and Rubin hardware will monetize at the muted rates of two-generation-old Ampere chips. Baker calculates that if next-generation silicon monetizes anywhere near current spot values, hyperscalers will generate roughly $2 trillion in operating cash flow, eliminating approximately $700 billion in projected debt requirements. Adjusting for one-off European Union regulatory fines, hyperscaler operating cash flow growth already accelerated from 28% to 35% quarter-over-quarter. To bridge capital deployment lags without taking debt onto its balance sheet, Nvidia has structured credit wrappers that provide downside protection to third-party lenders while securing cloud revenue-share royalties when spot prices exceed a baseline floor. Concurrently, long-term supply agreements (LTAs) for components like high-bandwidth memory (HBM) have become non-cancellable strategic assets; Baker notes that in a competitive race among Amazon, Google, AMD, and Nvidia, defaulting on an LTA during a temporary lull risks permanent loss of supplier allocations. Open-Source Models, Routing, and Software Defensibility (23:55) Baker characterizes open-source models as market 'dark matter' that public investors misinterpret. When silicon token indices dipped due to a mix shift toward cheaper open-source models like Kimi, markets treated the shift as a demand drop. However, Baker argues that 'a token is a token' requiring identical raw compute, memory, and wattage. By reducing nominal token costs, open-source adoption expands volume elasticity and transfers margin from frontier model developers directly to underlying compute providers. This shift supports software defensibility via multi-model router architectures. AI-native applications route 30% to 60% of execution volume to low-cost, fine-tuned open-source models, querying expensive frontier models like Claude or Grok only for complex planning and edge-case verification. Baker notes that this architecture can lower blended token expenses by up to 15 times, enabling enterprise software startups to protect gross margins while capital-efficient inference clouds like Together, Modal, and BaseTen scale rapidly. Power Deployments, SpaceX, and Regulatory Risks (50:35) Baker identifies political narratives and regulatory intervention—such as proposed data center moratoriums in New York—as the single greatest threat to the AI ecosystem. He views this as an industry public relations failure, arguing that popular claims regarding data centers draining community water supplies stem from a debunked academic citation error, whereas modern behind-the-meter energy deals often reduce local utility rates and create durable, high-paying infrastructure jobs. Regarding physical infrastructure execution, Baker highlights SpaceX, CoreWeave, and Crusoe as the only non-hyperscalers to successfully deploy over 500 megawatts of power within a single year. While Baker considers public reports of SpaceX targeting 8 gigawatts over 18 months to be very implausible due to power grid constraints, he emphasizes SpaceX's unmatched speed in deploying physical clusters. Furthermore, venture investments by Benchmark into StarCloud suggest that space-based orbital compute using Starlink's optical laser mesh is becoming a viable medium-term alternative. Baker concludes that operational AI fundamentals remain insulated from public equity volatility, supported by rising compute contract rates, cash-generative hyperscaler balance sheets, and architectural efficiency. However, the ecosystem's momentum depends heavily on whether the technology sector can mount a cohesive public relations effort to defeat local utility misconceptions and regulatory data center bans before they restrict physical power deployment.

Gavin Baker指出，公开市场在2024年7月因宏观信用担忧和叙事驱动出现类似2022年的暴跌，但硅谷底层的算力租赁价格、DRAM现货及Token生成量等关键基本面指标仍在持续加速，二级市场与产业真实动能之间存在严重脱节。 公开市场恐慌与底层算力加速的背离 (0:00) 公开市场AI标的在7月份单月下跌40%至60%，但Baker强调，硅谷底层的各项量化指标——包括GPU可用性、GPU租赁费率、DRAM现货价格以及整体Token增长——都在全面加速。华尔街的模型存在巨大盲区，主要因为其忽视了OpenAI、Anthropic等未上市巨头以及Together、Modal、BaseTen和Fireworks等开源推理云的惊人现金流与使用量。例如，Blackwell B200 GPU集群的现货租金在六到七个月内飙升了50%至60%，部分推理云服务商甚至预计新硬件合约续签价格将上涨100%。 此外，早先市场普遍预期旧款GPU在长期合约到期后会面临租赁费率的下调，但现实中早期锁定的长期算力合约价格远低于当下的现货水平。随着旧合约陆续到期并按更高市价重签，算力资产正迎来一轮重新定价。剔除监管罚款等偶发项目后，微软、Meta和亚马逊的整体经营现金流增速已从28%加速至35%，表明底层ROI正在改善，且这一加速发生在英伟达下一代Rubin架构商用之前。 宏观信用误读与自给自足的超大规模资本开支 (6:53) 7月份科技板块出现真实收益率上升、信用利差扩大以及Meta与英伟达信用违约互换（CDS）走阔的现象。市场传言Meta计划出租闲置算力，加剧了投资者对资本开支削减和产能过剩的担忧。Baker反驳称，Meta出租算力仅是效仿SpaceX高溢价变现高频算力集群的战术举措，其随后发布的Muse 1.1模型及内部遥测数据表明其基础设施投资从未放缓。 华尔街当前的资本开支模型过度悲观地假定Blackwell和Rubin芯片的变现能力仅相当于两代前的Ampere架构。Baker分析认为，如果下一代芯片能按接近当前的现货价格变现，超大规模云服务商（Hyperscalers）将产生约2万亿美元的经营现金流，这足以完全覆盖约7000亿美元的潜在债务融资需求，从而消除行业陷入传统信用挤压型资本周期的风险。 开源暗物质与智能路由重塑软件单位经济 (15:22) Baker将开源模型比作推动算力需求的“暗物质”。中国模型Kimi等开源工具的发布引发市场对前沿模型降价的担忧，甚至导致硅数据Token价值指数走低。但Baker解释称“Token就是Token”，生成开源Token所消耗的FLOPs、内存和能耗与闭源模型完全一致。开源的普及通过价格弹性放大了总Token需求，只是将原本属于前沿模型软件层的极高利润转移到了底层硬件和基础设施层。 在应用架构上，原生AI企业正在转向智能路由（Routers）方案。例如Cursor和Harvey等企业利用Fireworks Nexus等工具，将30%至60%甚至80%的常规流量导入低成本的开源微调模型（毛利率约30%），仅将复杂边界情况交由Claude或Grok等高价值前沿模型验证（毛利率约90%）。这一架构不仅将名义Token成本降低最高达15倍，还使AI应用摆脱了单纯“套壳”的质疑，显著增强了商业防御壁垒。 Token代际替代、长期供应协议与算力博弈 (30:51) 在企业微观层面，顶尖AI原生企业正将高达20%至25%（极端情况下达到30%至50%）的人力薪酬预算转向采购Token算力。在全球知识工作约25万亿美元的总体有效市场（TAM）中，即使只有20%被Token和算力替代，也意味着5万亿美元的直接市场空间。与以往苹果可以单方面压榨供应链的历史周期不同，当前亚马逊、谷歌、AMD与英伟达处于激烈的四方竞逐中，高带宽内存（HBM）的长期供应协议（LTA）构成了生死攸关的战略资产，任何买方若在短期波动中违约，都会被供应商永久重新分配产能给对手。 为进一步巩固生态，英伟达开创了一种新型融资架构，通过提供“信用包装”（Credit Wrapper）和在现货价格高于底线时的收入分成，联手黑石（Blackstone）与阿波罗（Apollo）等私募信贷机构为GPU集群融资，以此降低自身资产负债表风险并挤压初创芯片对手。同时，OpenAI与xAI在帕累托前沿的激进军备竞赛，使得Anthropic等原本对资本开支持谨慎态度的实验室也不得不重新加码计算资源投入。 SpaceX算力进局与数据中心的公共叙事风险 (50:35) 非超大规模厂商中，仅有CoreWeave、Crusoe和SpaceX在一年内成功上线了超过500兆瓦的电力基础设施。针对有行业报告称SpaceX计划在18个月内上线8吉瓦算力的说法，Baker虽然认为鉴于电网限制该目标在期限内极不可能完全实现，但指出SpaceX在部署速度和单位成本上具备极强优势。此外，Benchmark投资的StarCloud正联合SpaceX尝试利用Starlink光学激光网格进行轨道计算，使太空算力逐步脱离纯科幻设想。 但Baker警告称，AI生态当前面临的最大威胁并非技术或资本瓶颈，而是严重的监管与公关失败。针对纽约等地提出的数据中心暂停建设提案，他指出公众舆论普遍被水资源耗费（源于被辟谣的学术计算小数点错误）和电网负担等错误叙事误导。实际上大型机房通常能带动电价下调并创造大量高薪蓝领就业，AI产业亟需发起协同的公共宣传以扭转政治阻力。 综上，Gavin Baker认为短期市场动荡主要源于算法驱动的共识恐慌（如投资者过度依赖Claude解读资讯产生的从众行为）和对信用指标的误判，而非物理算力需求的衰退。然而，该推演成立的前提依赖于Blackwell及后续芯片能维持目前的变现溢价，且行业能否成功克服电网并网限制与日益收紧的数据中心区域监管政策，仍是尚未解决的核心变数。

## Background 本期节目由 Colossus 旗下的知名播客《Invest Like the Best》于 2026 年 8 月 4 日发布，主持人 Patrick O'Shaughnessy 与 Atreides Management 创始合伙人兼 CIO Gavin Baker 展开了第七次深度对谈。节目简介指出，在公开市场 AI 概念股历经惨烈回撤的背景下，双方聚焦于公开市场情绪恐慌与硅谷前线产业强劲需求之间的巨大脱节，深入剖析了 GPU 租金重定价、内存长约博弈论、Claude 对股市的认知单一性影响、SpaceX 算力野心以及整个行业面临的真正监管风险。 ## TL;DR Gavin Baker 深度复盘了 7 月份公开市场 AI 板块惨遭血洗、而硅谷实地需求全方位逆势狂飙的戏剧性割裂。他指出，随着旧款 GPU 长期折价合约到期并按高昂现货价格重定价，巨头们的自由现金流将大幅造血，直接消解数百亿美元的信贷缺口隐忧，而开源模型的爆发实质上是在降本提效并加速底层算力消耗。此外，针对英伟达创下十年新低的远期估值、内存厂商的长期供应协议博弈、推理架构的软硬件解耦，以及 SpaceX 被低估的数据中心交付能力，Gavin 均给出了极其硬核的推演，并强调阻碍这场技术革命的真正黑天鹅并非 ROI 破灭，而是致命的公众叙事失控与行政监管壁垒。 ## Atmosphere & style 整场对话在硅谷风投 Benchmark 著名的晚宴长桌旁录制，充溢着资深投资人之间高度默契、语速极快且信息密度爆炸的智力交锋。Gavin 的叙事兼具宏观对冲基金经理的冷峻纪律与技术极客的澎湃激情，面对二级市场的集体恐慌，他既展现出“翻遍硅谷每一个角落却找不到一项看空量化证据”的从容笃定，又对信贷周期和公关危机保持着敏锐的敬畏感。听感宛如置身于一场既有华尔街博弈论深度、又饱含前沿物理算力一手情报的高端闭门会谈，在一片风声鹤唳的市场迷雾中为听众剥离噪音、锚定理性坐标。 ## Bulletin - `02:35` 7 月美股 AI 概念股回撤达 40% 至 60%，但硅谷一线需求全方位加速，老款 GPU 租金非但未降反而垂直拉升，同配置 B200 集群租金半年内上涨 50% 至 60%。 - `11:54` 华尔街卖方模型按两代前的 Ampere 费率保守测算超大规模厂商现金流，但实地以 Blackwell 现货价测算的现金流可达 2 万亿美元，凭空抹去约 7,000 亿美元债务融资需求。 - `17:35` 英伟达远期 P/E 跌至 10 年最低位，同时通过联合贷款机构推出“信用增级”分成模式，以算力采购为筹码快速搭建经常性收入业务。 - `57:15` 大模型推理正走向预填充、注意力与前馈网络的物理分拆，SRAM 加速芯片有望摆脱对稀缺 HBM 的依赖，显著释放存量与新建集群的推理 ROI。 - `50:35` 加文·贝克强调当前 AI 产业最大的黑天鹅是监管与公关缺失，数据中心被不实传言污名化，行业急需通过超级碗等大众媒介扭转舆论。 ## Key takeaways - `02:35` 二级市场 7 月份经历“一个月走完 2022 年整年”的惨烈暴跌，但硅谷底层在 GPU 租金、DRAM 现货价格及 Token 生成量等指标上全线加速。 - `04:08` 华尔街缺乏对 OpenAI、Anthropic 及开源推理云真实运营数据的能见度，将这些非上市生态视作资产负债表外的“暗物质”。 - `05:06` 存量算力大多锁死在早期低价长约中，随着合同到期以高昂现货价格重签，重定价红利将大幅改善超大规模云厂商的经营现金流。 - `06:53` Meta 出租闲置算力的传闻被市场误读为削减 CapEx，实则是借鉴 SpaceX 将优质集群打包溢价转租的高 IRR 商业模式，其实际算力投入仍在加码。 - `08:22` Kimi K3 与 GLM 5.2 等顶尖开源模型的崛起压缩了闭源前沿的高暴利，但由于生成物理 Token 的 FLOPs 与能耗完全等价，反而直接利好底层硬件。 - `10:00` 实际收益率上升与科技巨头 CDS 费率走阔虽然反映了宏观信贷承压，但算力集群自身的强劲造血能力极大削减了对外部发债的依赖。 - `21:35` 投资人普遍使用 Claude 进行财报与新闻解读，导致金融市场分析高度趋同、认知单一化，反而放大了二级市场的极端波动。 - `23:55` 持续学习与高样本效率技术的突破长期看不会扼杀算力需求，因为半导体开支重心正不可逆地由预训练全面转向推理端。 - `26:38` 模型路由器（Router）配合定制化监督微调的开源模型，使得企业以极低成本并发调用 Token，最终反而拉升了底层 GPU 消耗时长。 - `30:00` 全球深度使用智能体 AI 的核心人群仅数十万就已引发算力荒，未来知识工作支出向 Token 转移将撑起数万亿美元的替代市场。 - `33:37` 内存原厂（SK 海力士、美光等）通过长期供应协议（LTA）锁定价格上下限，在算力军备竞赛中单方面毁约将面临永久丧失芯片配额的灭顶之灾。 - `36:41` 英伟达通过“信用增级包装”与算力分成模式绑定客户，同时以数十亿美元现金参股生态初创企业，极大拉升了单吉瓦算力的综合变现率。 - `43:28` 中国自主 DUV 光刻机实现从无到有的“相变”，但半导体制造遵循严苛的“干中学”经验积累规律，短期内难以彻底跨越工艺代际差距。 - `46:10` 开源模型能力提升叠加 Fireworks 等推理微调平台，让 Harvey、Cursor 等原生 AI 公司仅需极低代码量即可自建私有模型防御壁垒，摆脱对闭源 API 的依赖。 - `50:00` 以 Modal、Together、BaseTen 为代表的专业推理云呈现高增速且极少烧钱的高资本效率，正打破大模型实验室对超额价值的单一垄断。 - `57:15` 大模型推理正解耦为预填充（无 HBM 低成本芯片）、注意力（高带宽 HBM 芯片）与前馈网络（SRAM 加速芯片），全面优化算力投资回报率。 - `01:00:00` SpaceX 在 18 个月内计划上线高达 8 GW 电力算力，其凭借全行业最低交付成本与最高交付速度，极大地被公开市场低估为纯航天通信公司。 - `01:03:00` 随着星舰可回收发射与星链空间激光互联技术成熟，Benchmark 等非马斯克系顶级资本对轨道计算的注资印证了太空计算已具备现实落地土壤。 ## Detailed walkthrough ### `02:35` 七月的剧烈波动：一个月走完 2022 年整年的暴跌行情 距离加文·贝克（Gavin Baker）上次做客播客仅仅过去了两个月，帕特里克·奥肖内西（Patrick O'Shaughnessy）打趣道，如今播客更新的间隔已经缩短到了人工智能前沿大模型的迭代频率。加文自嘲地回应，之前一直有人批评说他们每次录播客都恰好撞上公开市场的“局部高点”，但这次显然没人能这么说了——整个 7 月的美股 AI 板块简直就像“把 2022 年整整一年的下跌浓缩进了一个月里”。大量公开市场的人工智能概念股从近期高位直线下跌了 40% 到 60%，市场情绪陷入极致恐慌。 然而，公开市场的哀鸿遍野与加州湾区的真实产业体感构成了极其刺眼的反差。加文询问帕特里克整个夏天在硅谷的见闻，帕特里克证实：在硅谷第一线，根本听不到哪怕一项关于 AI 需求放缓的负面量化数据，没有任何指标在减速，相反，所有指标都在全方位加速。这种乐观绝非信徒的盲目吹捧，而是从业者拿出各自后台的真实数据进行印证的结果。加文进一步指出，无论从哪个维度切入分析——是 GPU 的可用性、GPU 租赁单价、当月 DRAM 内存现货价格，还是底层 Token 生成量的飙升速度，现实情况全都在猛烈加速。 - **AI 概念股回撤幅度**：公开市场主流 AI 标的在短短一个月内直线回撤 40% 至 60%。 - **底层指标全面加速**：GPU 租赁价格、GPU 紧缺程度、DRAM 现货价格以及全网 Token 生成量均呈现逆势加速态势。 --- ### `04:08` 公开市场看不见的“暗物质”：非上市大模型与开源推理云 加文认为，公开市场之所以产生如此严重的误判，关键在于公开市场的投资者根本看不清非上市公司的数据。华尔街目前极度缺乏对 Anthropic、OpenAI 以及美国本土开源推理云（如 Fireworks、Baseten、Modal、Together AI 等）核心运营数据的能见度。近期开源大模型生态由于中国团队推出的 GLM 5.2、Kimi K3 以及英伟达自身的 Nemotron 等模型的发布而迅猛扩张，而这些推理计算全都在上述推理云上疯狂跑动并变现。OpenAI 的业务在加速，Anthropic 不仅维持着极其强劲的增长，而且“几乎可以肯定正在源源不断地创造巨额自由现金流”。 华尔街目前最流行的一张恐慌图表，是将半导体行业剧烈攀升的资本支出与超大规模云厂商（Hyperscalers）平缓甚至下滑的自由现金流做对比，以此得出“ROI 无法闭环”的悲观结论。加文指出，这张图表完全漏掉了 OpenAI 和 Anthropic 这类非上市巨头正在吃下的大量价值。更重要的是，它忽视了一个反常识的算力重定价机制。 --- ### `05:06` 旧款 GPU 租金暴涨与长期合约的重定价红利 在 2024 年和 2025 年初，不管是极度看好多头还是空头，市场形成了一个共识预期：随着时间推移和新架构芯片问世，老款 GPU 的租赁价格必然会下降，区别仅在于多头认为会温和走低，而空头认为会悬崖式崩跌。然而事实是，几乎没有任何人预料到老款 GPU 的市场价格竟然会垂直拉升暴涨。 当初大量新型算力云服务商（Neoclouds）为了给昂贵的 GPU 采购进行债务融资，必须提前锁定长期承购协议（Offtake Agreements）。这就导致存量计算集群的合同锁定价格远远低于当前的现货市场价格。随着这些老合同逐步到期，算力资源正在以大幅跳涨的溢价重新签约。即便未来现货价格出现回落，整体加权平均的算力重定价依然在向上重估，这将直接在财报中填平所谓的“ROI 鸿沟”。本季度微软、Meta 和亚马逊等巨头的经营现金流（OCF）增长率已经从 28% 提速到了 32%；如果剔除掉数额异常巨大的欧盟反垄断罚款等一次性法律支出，巨头们的实际经营现金流增速是从 28% 跃升到了 35%。这种体量的现金流在如此体量下加速，是在下一代 Rubin 架构芯片以更高溢价开售前就已确立的趋势。 ```mermaid flowchart LR accTitle: GPU 合约重定价与现金流加速机制 accDescr: 展示老旧长期合约到期后向高现货价重定价并推动超大规模厂商现金流增长的逻辑链条 subgraph ContractStage ["前期合约锁定"] A["Neoclouds 早期融资<br/>签署长期固定承购协议"] --> B["算力以大幅折价锁定<br/>(远低于当前现货价格)"] end subgraph RepricingStage ["重定价红利释放"] B --> C["早期长期合同陆续到期"] C --> D["按高昂现货价格重定价<br/>(老卡租金垂直上涨)"] end subgraph FinancialImpact ["财务数据兑现"] D --> E["巨头经营现金流加速<br/>(调整后 28% 升至 35%)"] E --> F["解答市场 ROI 闭环疑虑<br/>(Rubin 芯片铺开前已加速)"] end ``` - **经营现金流反弹**：Meta、微软、亚马逊的合计经营现金流增速从 28% 加速至 32%，剔除欧盟罚款等一次性项目后加速至 35%。 - **算力长协折价差**：当前运行中的存量算力大多锁死在早期低价长约中，到期重签的提价弹性远大于现货边际波动。 --- ### `06:53` 复盘七月恐慌：Meta 出租算力引发的虚惊与真相 加文梳理了引发 7 月大抛售的导火索链条。最开始的恐慌源于一条传闻：Meta 计划对外出租自己的闲置算力。市场瞬间将其解读为极其可怕的看空信号——认为 Meta 内部算力过剩、即将削减资本支出（CapEx）。然而 Meta 随后发布的财报彻底打破了这一谣言，CapEx 不仅毫无缩减，反而继续加码。 加文深入解释了 Meta 当时这一动作的真实动机：Meta 看到 SpaceX 等公司利用自己庞大的内部算力底座，将其中的量化交易优化集群以极其夸张的溢价打包转租给市场，获取了巨额回报。Meta 显然也盯上了这个机会，内部可能打算通过拿出一小部分算力出租，向市场证明其自建算力具备惊人的内部回报率（IRR），进而更有底气通过股权融资来进一步扩张算力。无论如何，根据行业供应链的深度遥测数据，Meta 的算力投资计划自始至终没有出现一丝一毫的动摇，反而在不断变得更加激进。几乎在同一时期，Meta 发布了 Muse 1.1，其性能表现是该团队两年来最强的一次，哪怕被同期的 Grok 4.5 夺走了一部分注意力，也充分证明扎克伯格绝不可能在 AI 竞赛中松开油门。 --- ### `08:22` 开源大爆发、Token 构成切换与黄仁勋的开源逻辑 随后引发市场第二次恐慌的，是 Kimi K3 与 GLM 5.2 的爆发，以及第三方监测机构 Silicon Data 的 Token 指数突然出现放缓和平缓震荡。华尔街将这两件事关联起来，产生了“AI 需求撞墙”的恐惧。加文解释称，该指数的波动纯粹是因为 Token 的“构成配比切换（Mix Shift）”，而非总量减少。 由于 GLM 5.2、Kimi 等极具性价比的顶尖开源模型崛起，市场上的调用量迅速从昂贵的闭源前沿模型转移到了开源模型上。前沿模型的推理毛利率此前高达 80%、90% 甚至 95%，当工作负载转向开源模型时，模型服务商的毛利额被压缩了，但“一个 Token 就是一个 Token，生成它所需要的算力分毫不差”。不管使用哪家模型，生成相同数量的 Token 所需的算力 FLOPs、显存带宽和电量消耗是完全一样的。开源的崛起凭借价格需求弹性（Elasticity）极大地拉动了 Token 的整体消费量，它削减的仅仅是闭源前沿模型层的超高暴利，反而把更多毛利真金白银地输送给了底层的算力基建层。 加文指出，这也是为什么英伟达 CEO 黄仁勋（Jensen Huang）会成为全世界最大的开源支持者： > “黄仁勋是个极具理想主义情怀的爱国者，他总是在做正确的事。但如果开源真的对他的生意有害，他还会成为全球头号开源支持者吗？” 开源打破了一两家前沿大模型垄断行业并收取 90% 暴利的局面，从人类社会和技术生态的角度来看至关重要；而在商业逻辑上，开源带来的 Token 廉价化与用量爆炸，最终都会原封不动地转化为对英伟达算力底层更为渴求的需求。 ### `10:00` 信用利差走阔与资本周期的真正隐忧 市场在消化完一系列事件后，又传出中国在 DUV 光刻机领域的突破，导致持有相关股票投资组合的机构集体抛售半导体设备股。加文·贝克（Gavin Baker）指出，在一连串利空噪音中，真正值得严肃对待的是宏观信贷层面的异动：实际收益率（Real Yields）显著上升，信用利差（Spreads）全面走阔。Meta 上周刚完成一笔债券定价，其发行成本和息差完全不符合一家巨头应有的优质待遇；与此同时，英伟达乃至全行业的信用违约互换（CDS）费率都在飙升。尽管许多精明的私募市场资本人士安抚称，这不过是投行与承销银行在为自己的贷款承诺做常规对冲，但盘面表现无疑极为难看。 加文强调，如果 AI 这轮超级基建真的需要完全仰仗债务融资来推进，那眼下的信贷信号确实足以令人背脊发凉——这正是金融史上屡见不鲜的“经典资本周期陷阱”：过度依赖杠杆激进扩张，而债务必须按期刚性兑付，一旦供需在短期内出现丝毫失衡，资金链和资产价格就会以极快速度全面崩盘，就像当年电信与互联网泡沫破裂时的惨状一样。正因如此，加文过去一个月里不断对信贷风险进行极限压力测试。 - **信贷市场预警信号**：实际收益率走高，科技巨头信用利差扩大，Meta 新发债券定价不及预期，英伟达及全行业 CDS 费率大幅飙升。 - **历史资本周期的警示**：由举债推动的基建扩张天生脆弱，一旦供需关系稍有脱节，刚性偿债需求将引发灾难性的流动性踩踏。 ### `11:54` AI 算力基建到底需要借债吗？ 要判断信贷恶化是否会终结 AI 周期，核心就在于厘清现有算力装机容量的“合约价”（Contract Pricing）与“现货价”（Spot Pricing）之间的巨额剪刀差，以及未来半年到数年的真实融资需求。加文详细拆解了华尔街主流卖方的一致预期模型：目前分析师在测算即将上线的数十吉瓦（GW）黑盒（Blackwell）及下一代鲁宾（Rubin）芯片时，所给出的变现回报率竟然大致等同于两代之前的 Ampere 架构（A100 时代），甚至完全无视了 Hopper（H100）的定价力。按照这种极度保守的假设，超大规模云服务商（Hyperscalers）只能产生 1 万亿至 1.1 万亿美元的经营现金流，这确实会产生数百亿美元的信贷缺口。 然而在加文看来，随着硅谷前线真实需求爆发，新一代算力绝不可能以 Ampere 那样低劣的水平变现。哪怕按当前 Blackwell 现货价打个折扣测算，超大规模云服务商的未来经营现金流也会高达约 2 万亿美元。这意味着系统将凭空多出近 1 万亿美元的内部造血能力，瞬间抹去 7,000 亿美元的借债需求，并迅速修复所有信贷比率。加文回顾了几个月前讨论过的“Blackwell 空气口袋”（Air Pocket）风险——即行业投入千亿美元购买芯片主要用于模型训练，而训练本身不直接产生收入的断层期。此前 Anthropic 的惊人突破让市场在 4 到 6 月选择忽略这一断层，但在 7 月多重情绪交织下，市场又突然重拾恐惧；而最具讽刺意味的是，恐惧蔓延的同时，巨头们的真实造血能力正在狂飙，例如微软在 6 月份上线了海量算力集群，而这笔新增产能带来的巨大回报甚至还没反映在第二季度的财报之中。 ```mermaid flowchart TD accTitle: AI 基建现金流与信贷需求模型对比 accDescr: 对比华尔街卖方保守模型与加文实地调研的现金流及债务需求测算 subgraph WallStreet["华尔街卖方一致预期"] WS_Model["假设按两代前的 Ampere 费率变现"] WS_CF["经营现金流仅约 1 万亿至 1.1 万亿美元"] WS_Debt["巨额基建依赖发债融资<br/>信贷指标恶化并承压"] WS_Model --> WS_CF --> WS_Debt end subgraph Reality["实地产业真实变现节奏"] R_Model["Blackwell / Rubin 算力现货高价且供不应求"] R_CF["经营现金流跃升至约 2 万亿美元"] R_Debt["减少约 7,000 亿美元信贷需求<br/>几乎完全依靠自身现金流造血"] R_Model --> R_CF --> R_Debt end ``` - **卖方模型的严重滞后**：市场普遍按照 Ampere（落后两代）的变现能力来预测 Blackwell 和 Rubin，严重低估了现金流。 - **现金流对信贷的替代效应**：若按实际算力定价测算，云服务商经营现金流可达 2 万亿美元，凭空消除约 7,000 亿美元的债务融资需求。 - **产能释放的时滞**：微软等巨头在 6 月单月部署的巨量算力，其产生的现金流增量在二季度财报中尚未完全显现。 ### `15:22` 没有明确“反派”的无名抛售与暴涨的算力租金 二级市场的这轮暴跌在技术分析师眼中显得极不寻常且令人发毛。2022 年市场大跌时，敌人明明白白摆在桌面上：通胀飙升、美联储激进加息、衰退阴影；哪怕是在 DeepSeek 冲击或是“解放日”（Liberation Day）闪崩时，大家也心知肚明是在为什么恐慌。但这一次，除了信贷指标的短期扰动外，其他利空由头在加文看来都近乎荒谬。“在投资中，最危险的往往是你看不见的那颗子弹，”加文引用道，“最重要的三个词从来不是‘安全边际’（Margin of Safety），而是‘我不知道’（I Don't Know）。”市场在缺乏坚实反向基本面的情况下持续下跌，让技术派神经紧绷。 然而，一旦深入硅谷一线，所见所闻与公开市场的悲观预期截然相反。就在录制播客的当天早晨，加文与硅谷一家极炙手可热的头部 AI 初创企业交流。该公司在 7 个月前租下了一个由数千张 Blackwell（B200）组成的超大集群，当时的合约价格在每 GPU 小时 2 美元中段；如今 7 个月过去，当他们试图续租或新增一个规模、配置完全一致的 B200 集群时，他们唯一的奢望居然是“能抢到每小时略低于 4 美元的报价”。在摩尔定律主导的芯片历史上，硬件价格正常应随时间温和阴跌，但眼下 B200 算力租金在短短半年多里逆势暴涨了 50% 到 60%。某知名推理云服务商负责人在播客中同样直言不讳：现有长期合约到期后，他们已做好心理准备为 Blackwell 算力支付比原来高出 100% 的价格。加文直言，这表明目前几乎所有超大规模云服务商都在其锁定的存量算力上“严重低估了真实盈利能力”（Under-earning）。 - **无名抛售引发的心理恐慌**：相比 2022 年逻辑清晰的宏观去杠杆，本轮抛售缺乏明确的主线利空，加剧了市场的不安。 - **算力租金倒挂飙升**：顶级创业公司同等规模的 B200 集群租金，在 7 个月内从 2 美元中段涨到逼近 4 美元，涨幅达 50%–60%。 - **云巨头的“低盈利状态”**：由于早期长协锁定了低价，随着旧合约陆续到期，重新按市价定价将带来收入与利润的大幅跳升。 ### `17:35` 开源作为“暗物质”与英伟达十年最低远期估值 加文此行来到硅谷的核心任务，就是带着极度的审慎去寻找“哪怕一个真实的负面量化指标”。但他向所有人抛出这个问题后，唯一被反复提及的负面数据，仅仅是某第三方机构监测到的 Anthropic 增长曲线似乎稍微脱离了原有的陡峭斜率。然而，这个数据在 Anthropic 的现有股东群体中遭到了极为激烈的驳斥——投资人们一个个心急如焚地想向加文证明真实情况依然好得惊人，但又因为极端害怕触犯公司保密协议而丧失未来 IPO 的配售份额，只能在欲言又止中憋得抓耳挠腮。 与此同时，真实生态正在全面加速：OpenAI 增长势头强劲，Cursor 和搭载 Grok 4.5 及 Grok Builds 的 xAI 在 7 月迎来爆发式跨越，而开源阵营更是凭借智谱 GLM 5.2 和 Moonshot Kimi K3 等实现了能力飞跃，英伟达自研的 Nemotron 也在快速逼近前沿闭源水平。加文形象地将开源生态比作二级市场眼中的“宇宙暗物质”——公募基金的财务模型根本无法直接追踪和审计开源世界的活跃度，但如果倾听各个推理云服务商的一线反馈，Token 的真实消耗量正在以不可思议的速度扩张。在这种基本面狂飙与公开市场杀跌的巨大割裂下，英伟达的远期市盈率（Forward P/E）在录音当天已经跌至过去 10 年来的最低水平。在整个半导体板块历史上，比当前更便宜的时刻唯有“解放日”闪崩与 DeepSeek 恐慌的那两个极端 V 型底部。二级市场显然在用脚投票，笃定巨头们当前处于“严重超额盈利”的顶点，但加文在硅谷翻遍每一个角落，依然找不到任何能支撑这种崩溃论调的底层量化证据。 - **Anthropic 股东的微妙心理**：面对第三方数据关于增速放缓的质疑，知情股东急于反驳又深恐泄密失去未来 IPO 配额。 - **开源生态的“暗物质”效应**：GLM 5.2、Kimi K3 以及英伟达 Nemotron 驱动推理需求暴增，公募市场无法在财报上量化，导致认知严重脱节。 - **英伟达的历史估值冰点**：远期 P/E 触及 10 年最低位，估值水平仅次于历史上两次极端恐慌底，公开市场预期与产业基本面出现巨大剪刀差。 ### `20:00` 算力短缺的自愈逻辑与市场的“去风险”狂潮 过去一个月对公开市场投资者而言极度煎熬。Gavin Baker 提到一位在富达（Fidelity）工作的朋友总结的投资心得：“过去三年生存的最佳法则，就是以最快速度做出最蠢、最肤浅的反应，然后在这些反应之间来回轮动。”在当前环境下，这意味着只要市场充斥着关于巨额资本支出不可持续的叙事，投资者就会本能地全面削减风险敞口（cut risk）。然而 Gavin 指出，这些引发恐慌的宏观叙事除了信贷利差扩大这一项属实之外，从基本面上看几乎全都不成立。 更深层的经济学逻辑在于，算力建设甚至可能根本不需要大规模依赖信贷。退一步说，就算建造所需浮点运算能力（FLOPs）确实依赖信贷支持，如果信贷市场不买单，其直接后果也只是让市场上现有的存量算力变得极其稀缺和昂贵。算力租金上涨会立竿见影地改善相关基础设施公司的财务指标，进而让信贷重新变得触手可及。只要当下的全球算力短缺依然存在——且 Gavin 强调自己竭尽全力寻找“算力过剩”的蛛丝马迹却一无所获，短缺反而在与日俱增——这种“短缺本身就会反向修复财务指标”的自愈机制就会持续生效。 ### `21:35` Claude 成为股市的 Walter Cronkite 与认知单一性崩溃 Gavin 谈到迈克尔·莫布森（Michael Mauboussin）的经典理论：系统多样性的崩溃（breakdown in diversity）往往是资产泡沫与踩踏崩盘的直接导火索。如今在二级市场投资圈（无论是散户还是专业对冲基金机构），几乎所有人都会把研报、财报以及新闻喂给 Claude、Claude Code 或其 Agent 进行解读。然而大语言模型的推断本质上是概率性的，导致成千上万的投资人得到的往往是高度趋同的认知输入，差异性几乎被抹平。 过去人们常讨论媒体碎裂化，怀念冷战时代著名主持人沃尔特·克朗凯特（Walter Cronkite）作为“全美唯一真理之声”的一统天下。具有讽刺意味的是，现在的金融市场仿佛退回到了那个时代——Claude 俨然成了股票市场的“沃尔特·克朗凯特”，所有人对它的解读深信不疑。然而 Claude 尽管聪明，其逻辑并非次次正确，面对股票市场本质上的贝叶斯概率预测更是如此。这种高度同质化的交易行为导致了荒诞的市场波动，正如 X 平台上著名的匿名半导体博主 TBU 分享的日本电容器股票走势图：基本面订单甚至还没真正落地，整个板块已经垂直拉升后又暴跌，短短六个星期就透支走完了一个通常需要三年的完整行业周期。 ```mermaid flowchart TD accTitle: Claude 驱动的认知单一性与市场波动传导机制 accDescr: 展示市场新闻如何通过 Claude 产生同质化解读并引发极端交易波动的流程图。 subgraph Input ["信息输入"] A["市场新闻与研报数据"] end subgraph AI_Processing ["认知归拢 (单一声音)"] B["Claude / Claude Code / Agents"] C["高度趋同的概率性逻辑解读<br/>(金融界的 Walter Cronkite)"] A --> B B --> C end subgraph Market_Impact ["同质化交易与极端周期"] D["买卖双方蜂拥执行同质策略"] E["周期极端压缩<br/>(例如日本电容股 6 周透支 3 年周期)"] C --> D D --> E end ``` ### `23:55` 持续学习与样本效率：短期冲击还是算力跃迁？ 在讨论前沿实验室（如 Ilya Sutskever 创立的 SSI，声称将在 8 月推出模型，以及 Black Forest Labs 等）的技术创新时，Gavin 观察到业界在“持续学习”（continual learning）与“高样本效率学习”（sample-efficient learning）方面似乎正无限接近突破。如果模型不再需要一口气吞下 300 万亿 Token，而是只需 10 万亿 Token 就能在与真实世界的互动中持续高效进化，这是否会引发算力需求的短期断崖式下跌？ 尽管从理论上讲，样本效率的暴增可能会阶段性冲击预训练算力需求，但 Gavin 认为预训练在整体半导体算力需求中的占比本就会逐步边缘化，最终大头必然是推理。这种算法范式的跃升对人类文明而言是天大的好事，且英伟达也深度参与并投资了这些初创公司。纵观历史，效率提升从来都会激发更庞大的应用广度，很难真正压垮底层的算力基建需求。 - 传统的巨型模型训练需要吞吐数百甚至 300 万亿 Token，而高样本效率模型有望在极小数据基数（如 10 万亿甚至更低）下启动并自我迭代。 - 长期来看，芯片需求的大盘重心不可逆转地从预训练转向推理应用，预训练算力占比会无限趋近于一个极小值。 ### `25:19` 究竟什么情境会真正引发恐慌？ 当被问及什么样的现实情况才会迫使他彻底推翻多头逻辑时，Gavin 直言真正的风险点在于现金流引擎的熄火： 1. **核心实验室造血能力停滞**：如果超大规模云厂商与头部实验室（Anthropic、OpenAI、xAI 的 Grok、Cursor 等）的经营性现金流（OCF）停止加速甚至开始下滑，迫使基建必须转向全额债务融资。 2. **GPU 价格剧烈且持续地坍塌**：如果 GPU 现货与长约价格出现剧烈暴跌，且市面上开始出现“买不到芯片的窘境完全消失”、“人人都说自己手头算力过剩”的现象（现实中目前恰恰相反，抢算力“就像抢毒品一样疯狂”）。 3. **闭源与开源总需求双双见顶**：如果模型调用总量不再扩张，而非仅仅是开源替代闭源的分流。 ### `26:38` 智能路由与多模型架构下的算力悖论：成本腰斩与算力消耗暴增 面对很多传统大厂“AI 预算在短短三个月内烧光 20 倍”的痛点，业界正在快速向“模型路由器”（Router）驱动的多模型架构转移。然而，许多公开市场分析师误以为“企业使用开源模型配合路由器把开销砍半”意味着对算力需求的利空，Gavin 严正反驳了这一浅层认知。 当企业通过推理云（Inference Cloud）进行监督微调（SFT）和强化学习（RL），定制开源模型作为第一道防线，仅在关键步骤才将请求路由至 Claude 或 Grok 等前沿闭源旗舰模型时，用户账面支出的下降只是因为 Token 的毛利率发生了转移——从闭源巨头 90% 毛利的昂贵 Token 变成了微调开源模型约 30% 毛利的廉价 Token。对于底层硬件而言，制造相同数量的 Token 所需的物理算力毫无差别。更重要的是，成本大幅下降让企业敢于在业务管线中并发调用多得多的 Token，最终底层 GPU 的算力消耗小时数（GPU compute hours）不仅没有萎缩，反而随着开源 Token 的放量而加速膨胀。 与此同时，硅谷的 AI 原生初创企业（AI Natives）展现出了完全不同的经济模式：他们几乎不招聘传统员工，将绝大部分资金预算直接倾注到 Token 的消费与生成中，其对算力的饥渴程度不仅没有丝毫降温，反而在全力加速。 ### `30:00` AI 采用浪潮与知识工作支出的代换博弈 当前全球范围内对 AI 的采纳呈现出极度不均衡的“多波次”特征：美国西海岸的技术极客公司正飞速推进，而东海岸的传统企业才刚开始触及皮毛，至于美国非沿海地区的公司采纳速度更慢；欧洲则陷入了典型的“还没开始用，就先忙着出台法规监管”的状态。然而，这背后有一个令人震撼的巨大反差：全球目前真正深度使用智能体 AI（Agentic AI）的人口可能仅有 25 万到 50 万人，但整个行业却已经陷入了极其严峻的算力短缺。全球有七八十亿人口，一旦使用人数从几十万跳升至 1 亿乃至 5 亿，底层的算力挤兑将难以想象。 针对市场上“超大规模云厂商（Hyperscalers）营运现金流加速的资金到底来自哪里、终端客户在何处”的质疑，逻辑上只有两种路径：要么源于生产力爆发带来的整体经济加速增长，要么来自劳动力替代（Labor Substitution）。在 AI 原生初创公司中，劳动力替代早已切实发生，其表现并不是大规模裁员，而是人效（每全职员工毛利额 / Gross Profit per FTE）呈现垂直攀升——它们可以用极少的人手支撑庞大的业务。从全球知识工作（Knowledge Work）高达 25 万亿美元的盘子来看，即便只将 20% 的人工支出转移至模型代币（Token）支出，也是一个高达 5 万亿美元的巨额市场。人类社会显然更渴望通过生产力扩张来实现增长，而 Cognition、Ramp 和 Stripe 等机构的数据也印证了这一点：AI 支出越高的企业，其业务增长速度明显更快；甚至在 Cognition Index 中，连水管工、暖通空调（HVAC）承包商等蓝领行业都在因 AI 工具而显著受益。 - **深度用户基数与算力缺口**：全球仅约 25 万至 50 万人在使用智能体 AI，却已引发严重的硬件与算力荒。 - **Token 支出占人力薪酬的比例**：在极端拥趸型公司中，Token 支出已经占到总员工薪酬支出的 20% 到 25%；SemiAnalysis 的 Dylan Patel 公司这一比例达到 30%，业内传闻甚至已有公司飙升至 50%。 - **创始人控制型企业的用工韧性**：扣除疫情期间的过度招聘后，顶尖科技企业并未出现大规模失业潮，多数创始人和管理层更倾向于“劳动力囤积（Labor Hoarding）”配合 AI 工具赋能，而非粗暴裁员。 --- ### `33:37` 抢电危机与长期供应协议（LTA）的博弈论 除了算力本身，电力已经成为算力中心点亮（Energize）的最大瓶颈。资本主义市场正在展现惊人的解决能力：政策法规在积极调整，涡轮机与柴油喷气引擎制造商甚至将退役老飞机的航空涡轮拆解翻新，改装为数据中心应急发电装置。但在半导体供应链端，行业正在经历结构性转向：尤其是内存芯片（HBM/DRAM）厂商，不再单纯追求单季度的业绩爆发，而是选择用短期超额利润去换取带有预付款、价格底线与上限的长期供应协议（LTA）。 ```mermaid flowchart TD accTitle: LTA 违约与供应链淘汰的博弈链条 accDescr: 展示四巨头若在下行周期单方面撕毁内存 LTA，将在随后的周期轮换中失去配额并断送核心业务的因果关系。 subgraph Trigger ["周期下行：价格诱惑"] A["2027-2028 周期下行<br/>现货价格暴跌"] --> B["超大规模买家动摇<br/>企图撕毁 LTA 违约"] end subgraph Retaliation ["内存原厂反制与产能出清"] B --> C["内存厂取消长期供货保底<br/>将产能转配给竞对"] C --> D["内存全行业缩减资本开支<br/>产能急剧收缩"] end subgraph Catastrophe ["周期反转：不可逆淘汰"] D --> E["下一轮严重的供不应求周期到来"] E --> F["违约巨头失去配额与芯片分配<br/>整个 AI 算力底座与业务彻底出局"] end ``` 这种 LTA 机制构建了一场高风险的“权力的游戏”。当前真正具备规模化算力影响力的四家巨头分别是：研发 Trainium 的亚马逊、研发 TPU 的谷歌、AMD，以及体量远超其余三家总和的英伟达。增加单位算力（FLOP）所匹配的内存容量，是提高 Token 产出、摊薄单位成本最关键的杠杆。假设在 2027 或 2028 年行业出现下行，某家巨头（例如谷歌）出于压价诱惑试图单方面撕毁 LTA，其博弈后果将是毁灭性的： - 内存行业具有极强的周期性，一旦出现违约，内存原厂会顺理成章地缩减资本开支、出清产能，并将宝贵的配额转移给对手。 - 当下一轮严重的供给短缺再次来临时，议价权将彻底倒向 SK 海力士（SK Hynix）或美光（Micron），撕毁合约者将被剥夺配额分配，从而在 AI 军备竞赛中直接出局。 - 过去苹果可以凭借绝对的单一采购量随意压榨供应商而不承担后果（就算苛待 SK 海力士，美光也会争相接盘）；但在四巨头争霸与 Etched 等各类芯片初创企业蜂拥的当下，买家只要毁约，失去的不仅是短期价格，而是整个企业的生死存亡。 --- ### `36:41` 英伟达的“信用增级”分成模式与内存原厂的自救 针对英伟达目前处于历史相对低估值的现象，其强大的生态垄断与金融创新能力被严重低估。在当前宏观环境下，没有任何资产比英伟达的 GPU 更容易获得机构信贷支持；同时英伟达在土地获取、电力撮合上展现了极高的掌控力。更具颠覆性的是英伟达新推出的类“信用增级包装（Credit Wrapper）”商业模式：英伟达联合第三方贷款机构为 GPU 采购方提供融资支持，同时在 GPU 市场价格高于预设底线时，英伟达直接从中抽取收入分成（Revenue Share）。 这种模式巧妙地打破了资本支出与现金流不匹配的死结：英伟达在赚取巨额硬件利润的同时，不直接承担违约放贷的全部风险，反而借助权利金（Royalties）在极短时间内变相搭建起一个极其庞大的云服务经常性收入业务。尽管英伟达在投资协议中名义上会要求资金不可直接用于采购自身芯片，但在资本充沛且具有流动性的环境下，此举实质上盘活了整个客户生态的现金流循环。 面对这一产业变局，内存巨头（如 SK 海力士）如果想摆脱周期性惨败宿命，最合理的战略就是全盘复制英伟达的这套金融玩法： - **参与信用增级体系**：主动找到 GPU、Trainium 等算力采购方，加入类似的信用增级包装中。 - **前期现金出资与增信**：利用当前暴增的营运现金流在前期出资作为债务增信与保证金（Surety），以此安抚黑石（Blackstone）、阿波罗（Apollo）等在信贷风险面前犹豫不决的华尔街私募债权人。 - **绑定长期收入分成**：放弃对部分当期现金暴利的贪婪，以承担部分底线风险为筹码，换取后续 AI 收入的长期抽成，将传统的 LTA 演进为兼具抗周期韧性与长期现金流捕获能力的深度利益绑定。 ### `40:00` 英伟达的股权版图与算力军备竞赛的博弈论 英伟达正在执行一项极具前瞻性却被资本市场深度误解的资本运作策略：通过为生态伙伴提供算力与信贷支持，换取经常性收入的分成（Revenue Share）乃至股权期权。Gavin Baker 指出，黄仁勋（Jensen Huang）之所以对 AI 抱有几乎无限的信心，是因为他身处行业的最顶层，几乎所有顶尖 AI 实验室（包括最新创立的 Safe Superintelligence）都必须与英伟达深度绑定。黄仁勋目睹了所有持续学习（Continual Learning）与模型架构的前沿突破，所见所闻赋予了他无与伦比的牛市视野。英伟达手握数百亿美元的充沛自由现金流，利用这些资本参股 Anthropic 等关键前沿玩家，并在各家初创公司因前期高额资本开支陷入自由现金流为负的夹缝期时提供支撑，既分享了未来股权的巨大上涨空间，又以极高的效率拉升了英伟达每吉瓦（per gigawatt）算力基础设施的创收能力。 这一打法同时构筑了极其坚固的竞争壁垒。对于那些自研 AI 芯片的初创公司而言，它们在台积电代工采购晶圆的价格高于英伟达，采购 HBM 内存的价格也高于英伟达，在芯片融资成本和供应链地位上更是处于全面劣势。这种结构性成本重负使得除了架构路线彻底迥异的极少数玩家外，几乎没有人能绕开英伟达的引力场。 与此同时，前沿大模型实验室之间的算力博弈正在上演经典的囚徒困境。此前 Anthropic 联合创始人 Dario Amodei 曾理性反思：如果在大模型规模化扩张中购买过多算力，企业可能会在极其沉重的折旧与开支下破产；但如果算力储备不足，就会彻底被前沿淘汰。然而现实是残酷的：OpenAI 与依托 SpaceX 及海量集群的 xAI（Grok 4）已经全力重返算力军备竞赛的最前线。当看到竞争对手在帕累托前沿（Pareto Frontier）不断发力狂奔，处于同一牌桌上的任何前瞻性实验室都绝对不敢松开油门。 ```mermaid flowchart TD accTitle: AI Frontier Compute Game Theory accDescr: The feedback loop of compute investment between leading AI labs and hardware providers. subgraph Labs["前沿实验室博弈"] OAI["OpenAI / xAI 等激进扩容"] -->|"争夺帕累托前沿"| Anthropic["竞争对手跟进扩张"] Anthropic -->|"算力不足即出局"| CapEx["巨额算力投入 (CapEx)"] end subgraph Nvidia["英伟达生态赋能"] CapEx -->|"采购芯片与集群"| NV["英伟达 (NVIDIA)"] NV -->|"自由现金流转投资"| Equity["股权绑定 + 营收分成"] Equity -->|"强化资金周转与每吉瓦创收"| Labs end ``` --- ### `41:46` 硅谷的极度乐观：一张 H100 年租金 25 万美元的推演 当被问及在硅谷拜访期间是否遇到比自己更乐观的人时，Gavin 直言：“在这里，几乎所有人都比我更乐观。”公开市场与硅谷一线从业者之间的认知撕裂已经到了令人匪夷所思的地步。在华尔街，面对 AI 股票的回撤与质疑，Gavin 感觉自己像个盲目乐观的傻子；但一旦踏入旧金山湾区的大模型实验室与算力生态圈，他发现自己瞬间变成了全场最悲观的人。 Dwarkesh Patel 等科技分析师甚至提出了极为激进的算力超额需求推演——预测未来租用单张英伟达 H100 显卡一年的成本可能飙升至 25 万美元，相当于当前现货价格的 15 倍左右。Gavin 坦言，这甚至超出了自己贝叶斯概率模型中考虑过但已排除的最极端边界。这类极度乐观者的底层逻辑在于三重相乘效应的飞轮：算力的物理总量正在几何级暴增、算力硬件与托管的毛利率正在持续扩张、推理端（Inference）的商业化毛利率同样在不断攀升。这三者的乘数效应推动了前沿实验室收入的疯狂加速，也解释了为何从业者完全无视公开市场的短期恐慌情绪。 - 华尔街在 7 月经历了如 2022 年熊市一般的惊弓之鸟心态，但硅谷实体的算力需求毫发未损。 - 前沿从业者认为即使开源模型在快速普及，其极低的调用成本反而刺激了整个行业推理总量的几何倍数增长。 - 乐观论调建立在“算力需求总量 × 算力毛利 × 推理毛利”的指数级共振假说之上。 --- ### `43:28` 中国 DUV 光刻机突破：技术相变与股市的剧烈波动 对于近期关于中国本土自主研发出深度紫外（DUV）光刻机的传闻与讨论，市场反应呈现出剧烈分化：有人将其贬低为不过是 ASML 在 2001 年的技术水平，不足挂齿；另一些人则将其视作全球前沿算力供应重组的惊雷。Gavin 认为这两种观点在不同维度上都有其道理。 如果将成熟的先进光刻机比作现代喷气式飞机涡轮发动机，EUV 极紫外光刻机就像是科幻中的曲速引擎，而早期的 DUV 机型可能只是螺旋桨飞机。虽然中国目前据传实现的设备在代际上落后了行业最先进水准近 20 至 25 年，但从“无”到“有”的跨越本身是一场不可忽视的技术“相变”（Phase Transition）——就像物质从液态凝固为固态一样，它建立了自主闭环的工业起点。半导体制造本质上是深度的“干中学”（Learning by doing），工业经验与工艺良率的积累无法通过技术瞬移来跨越，必须经历一代代晶圆流片的真实试错循环。 资本市场的反应往往具有高度投机性：一有风吹草动便极度恐慌性抛售 ASML 等半导体设备巨头；但实际上即使中国的技术替代在未来真正冲击到 ASML 的订单，也至少是五年后的事情。在此期间，金融市场往往会经历无数次“惊恐、遗忘、再次恐慌、再度遗忘”的情绪钟摆。尽管美中科技脱钩正成为一种难以逆转且相互强化的客观现实，且任何单方面的战略企图都无法被轻易打断，但断言中国能在短时间内从 2001 年的工艺跨越至 2030 年同样脱离了半导体物理制造的客观规律。 --- ### `46:10` 软件业与 AI 原生应用的曙光：从“套壳”到私有强化学习 在过去一个月中，当 AI 硬件标的遭遇剧烈估值杀跌时，美国传统软件板块却迎来了近乎垂直向上（Vertical）的强劲反弹。这一转折的背后，是开源大模型能力快速逼近前沿水平，以及以 Fireworks AI 为代表的推理与模型微调平台的成熟，这为软件产业与 AI 原生公司带来了巨大的转机。 两年前，风投界曾惊叹许多 AI 创业公司在创立短短 9 个月内就将年经常性收入（ARR）从零做到 5,000 万美元并实现正向现金流；然而当时市场普遍质疑其护城河，将其轻蔑地称为“ChatGPT 套壳”（Wrappers）。如今，这一格局正在被彻底颠覆。AI 创业公司在特定垂直领域深耕的过程中，沉淀了大量极其独特的专有业务数据。借助如 Fireworks 发布的 Nexus 这类工具，开发者仅需在 Claude Code、OpenAI Codex 或 Grok 编写的现有代码中加入区区“三行代码、大约二十个单词”，就能自动将私有业务数据接入底层平台，通过强化学习（RL）微调出专属模型，并配合智能路由器（Router）分发请求。 这种架构革新让 Harvey（在法律垂直赛道深耕）、Cursor（在 AI 编程领域领跑）等原生 AI 企业找到了构筑防御纵深的绝佳解法： - 摆脱对单一前沿闭源模型的绝对依赖，将原本 100% 的闭源 API 依赖度降至仅 30% 到 60%。 - 将大量高频任务分流至企业自有的强化学习微调开源模型，大幅削减算力代偿成本。 - 从随时可能被底座模型吞噬的“套壳工具”，演变为兼具专有数据沉淀与成本结构优势的垂直壁垒型企业。 --- ### `48:16` 前沿模型分流、RSI 假说与开源生态的护城河 Cursor 等团队揭示了一种极其高效的模型协同范式：AI 正在以惊人的速度“速通”人类社会的组织分工逻辑——让能力最强的前沿大模型（Frontier Model）专注于顶层的架构规划与任务分解，随后将具体的繁重子任务分派给参数规模更小、响应更轻快、成本更低廉的模型去执行。这种协同架构不仅使系统执行效率暴增 15 倍，而且其调用主力变成了单价极低且能力足够胜任的开源模型。 面对这一趋势，部分硅谷核心圈层（尤其是 Anthropic、OpenAI 及 Grok 的极简主义/最大主义信徒）持有截然不同的观点：他们认为，一旦闭源前沿模型触发“递归自我改进”（Recursive Self-Improvement, RSI），头部实验室就能够通过超级模型的自研蒸馏机制，在任何智力层级上以不可思议的极致低成本向外界输出服务，届时所有的开源模型和独立推理生态都将被瞬间降维打击、荡然无存。 Gavin 对此保持了高度的敬畏与审慎，但他指出 RSI 最大主义者的消灭论在商业现实中发生的概率并不高。核心原因在于： 1. **私有领域数据的不可替代性**：海量垂直领域的 AI 原生企业已经圈定并积累了大量深潜于行业内部的专有数据，这些数据根本不会对公有前沿模型开放。 2. **摆脱服务条款与算力盘剥的渴望**：在开源模型尚未崛起、推理云与路由调度技术未成熟前，企业只能无奈接受大型前沿模型厂商苛刻的 API 协议与不断变动的条款；而拥有自主蒸馏与微调能力后，各垂直领域的领军者正在果断跳下闭源厂商的“跑步机”，夺回战略自主权。 ### `50:00` 开源 Token 的繁荣与低烧钱率的推理云崛起 在模型生态的发展路径上，更便宜的开源 Token 并不会扼杀前沿模型的溢价，反而可能极大推高最顶尖的前沿 Token 价值。打个比方，如果市面上充斥着跑起来极其便宜的“120 智商”开源模型，那么一个能充当总指挥、能够协调并编排这些子模型的“160 智商”前沿模型，其战略价值与商业定价反而会水涨船高。这就像企业内部的薪酬分配逻辑：CEO 的薪资往往是中位数员工的数十倍，因为前沿指挥官的调度能力决定了整个系统产出的上限。 与此同时，AI 价值捕获的版图正在悄然分化。此前业界普遍惊讶于前沿大模型实验室拿走了绝大部分经济收益，但如今以 Together、Modal、BaseTen 为代表的专业推理云（Inference Clouds）正在打破这一垄断。最令人震惊的是它们的商业模式效率——这些推理平台在业务增速上几乎紧咬早期前沿实验室的爆发节奏，却以极高的资本效率运作，几乎不烧多余的现金。如果用软件行业传统的“Rule of 40”（40法则）等 SaaS 财务指标来衡量，它们展现出的经营数据高得惊人。 从算力底层来看，前沿 Token 或许依然会攫取大部分经济利润，但开源 Token 可能会在处理总量（Token 吞吐量）上占据绝对多数。无论 Token 来自哪种模型，在物理底层上“一个 Token 就是一个 Token”——生成它所必需的浮点运算次数（FLOPs）、电力消耗（瓦特数）、机房占地面积与散热冷却需求都是完全等价的，因此开源生态的爆发对整体算力基础设施的需求同样是巨大的利好。 ```mermaid flowchart TD accTitle: AI模型分层与底层算力承接结构 accDescr: 展现高智商前沿模型作为编排调度层与高吞吐低成本开源模型结合，共同拉动底层算力消耗的系统架构。 subgraph Orchestration["协调指挥层 (高价值密度)"] Frontier["160 IQ 前沿模型<br/>(Frontier Tokens: 捕获绝大部分超额经济价值)"] end subgraph Execution["执行与高吞吐层 (高资本效率)"] OpenSource["120 IQ 开源模型<br/>(低成本、高并发处理日常推理)"] InfClouds["高效推理云生态<br/>(Together / Modal / BaseTen: 高增速且极少烧钱)"] end subgraph Physical["物理算力承接层 (等价物理消耗)"] Hardware["算力基础设施<br/>(FLOPs 算力 / 瓦特功耗 / 机房空间 / 水冷散热)"] end Frontier -->|"调度与任务编排"| OpenSource OpenSource -->|"承载与分发"| InfClouds InfClouds -->|"底座消耗<br/>(A Token is a Token)"| Hardware Frontier -->|"底座消耗"| Hardware ``` --- ### `50:35` AI 最大的黑天鹅风险：监管与致命的公关缺位 当探讨 AI 行业可能遭遇的最坏情景时，真正的最大风险既不是技术停滞，也不是资本开支过剩，而是毫无疑问的**监管风险**（Regulatory Risk）。在二级市场上，AI 相关股票经历了惨烈的回调、估值不断降低，而 7 月份硅谷前沿实地的基本面甚至比 6 月份还要显著好转；但纽约州等地方政府推动的数据中心建设禁令（Data Center Moratorium）表明，来自政策层面的行政阻力正变得无法忽视。 AI 行业在面向公众沟通和公共关系（PR）上的表现堪称灾难。目前许多美国普通民众脑海中被灌输的叙事充满负面色彩——“数据中心会推高电价、吸干所有的水、最后夺走你的饭碗”。但现实情况恰恰相反： - **电价不仅没有上涨，往往还会下降**：现代超大规模数据中心通常采用“表后交易”（Behind-the-Meter Deals）或配套大规模清洁能源项目，并与当地政府达成类似特朗普呼吁签署的“数据中心倡议”（Data Center Pledge）。开发商为了落地，不仅承诺自行消纳电力增量并降低居民电费，甚至还会出资为社区新建医院、学校、警察局与消防站。 - **创造持久的高薪蓝领岗位**：数据中心绝非仅仅带来短期建筑工人，日常运行必须依赖大量的管道工、电工、暖通空调（HVAC）维护人员及后续硬件返修（RMA）与定期架构升级。数据中心建设实际上是近年来推动蓝领工人薪资上涨的最强驱动力之一；讽刺的是，自诩代表蓝领工人的民主党政客却在试图叫停这些项目。 - **被极度放大的“喝水”谎言**：坊间盛传的数据中心海量耗水谣言，源头竟是一位作者在出版书籍中犯下的低级数学错误，将数据中心耗水量凭空夸大了整整 10,000 倍（足足 4 个数量级）。尽管作者本人多次公开承认错误并进行辟谣澄清，但虚假信息早已飞遍全球，宛如经典的“大力水手吃菠菜补铁”效应（源于 80 年前学术著作点错了小数点，导致几代人误以为菠菜含铁量奇高）。 --- ### `53:44` 重塑叙事：用超级碗级的公关讲好救人与繁荣的故事 正如谚语所言：“真相还在穿鞋，谎言已经走遍半个世界”（A lie can go around the world faster than truth gets out of bed）。AI 行业极度需要成立专门的基金会或政治行动委员会（PAC），在“超级碗”、“四强赛”（Final Four）、NFL 职业橄榄球赛和 MLB 世界大赛等全美顶级体育赛事的黄金时段投放广告，主动正本清源： 1. **社区红利直达**：告诉普通选民，签署了协议的数据中心能带来更便宜的电费、更富裕的社区设施、大量长期稳定的高薪蓝领就业，且对当地水源与生态环境没有负面冲击。 2. **生命拯救的现实力量**：在刚刚结束的 ASCO（美国临床肿瘤学会）等顶级医学年会上，AI 辅助发现的科学突破数量创下历史之最。对于患有罕见病或重症的家庭而言，AI 正在实打实地提高患者生存康复的几率。 硅谷科技圈的菁英往往觉得这些益处“显而易见”，因而无法理解为何绝大多数普通美国人对此持怀疑排斥态度。即便是极度支持经济增长的深红州（保守派共和党执政州）官员，也在对业界发出警告：“如果你们自己都不站出来把故事讲明白，我们哪怕想帮你们背书也无能为力。”纽约的停建令只是一个危险的苗头，行业必须立即抢回话语权。 --- ### `57:15` 推理架构解耦与被低估的黑马 在算力架构的演进前沿，一个被市场普遍忽视的关键变量是**基于 SRAM 的新型加速芯片**。这类芯片不依赖紧俏且昂贵的高带宽内存（HBM DRAM），通常采用相对成熟的工艺制程制造，因而不必与英伟达最顶级的 GPU 争抢先进制程产能。 从计算流水线来看，整个大语言模型推理过程正在发生深度解耦（Disaggregation）。业内通常将推理划分为预填充（Pre-fill）与解码（Decode）两个阶段，但解码阶段实质上包含两个计算性质完全不同的部分： 1. **注意力机制（Attention）**：计算与带宽密集型； 2. **前馈网络（Feed-Forward Network, FFN）**：参数访问密集型。 推理架构的终极形态是将三者拆解到最合适的物理硅片上：用无 HBM 的低成本芯片做预填充，用挂载高容量 HBM DRAM 的超强计算芯片处理注意力，而将前馈网络完全卸载到专属的 SRAM 加速芯片上。无论硬件设计者如何在单颗主芯片上微调算力、HBM 与 SRAM 的容量比例，面对层出不穷的新模型工作负载都无法做到完美匹配；而通过将这三部分物理分离，现有算力存量和新建集群的推理投资回报率（ROI）将被彻底释放。 在巨头林立的“权力的游戏”棋局中，也有黑马正在崛起： - 比如芯片界传奇人物李立游（Leo Li）领衔的团队以及 Fireworks AI 的执行力备受赞誉； - 打造了 Devin 的 Cognition 联合创始人兼 CEO Scott Wu 也是不容小觑的顶级选手。 谈及资本市场对新晋上市龙头的认知，以 SpaceX / xAI 为代表的平台在公开市场的定价尚未被完全消化。尽管基本面在持续狂飙——比如发布了 Grok 2（乃至筹备中的 Grok 4.5）、收购了增速惊人的 AI 编程神器 Cursor——而且过去几年证明了其以全行业最快速度、最低价格部署超大规模算力的能力。它们在现货算力高点大举切入，一口气向市场释放了庞大算力储备，而整个市场瞬间将其完全吸收，未对运力与算力交易产生丝毫负面冲击。根据科技博主预测，其雄心勃勃的算力规划甚至可能在未来直接冲击 8 吉瓦（GW）级别的电力规模。 ### `01:00:00` SpaceX 的算力野心与被低估的数据中心潜能 在探讨 SpaceX 时，Gavin Baker 重点分析了其庞大的算力扩张规划与市场预期之间的巨大反差。市场上流传着一份公开的研究报告指出，SpaceX 计划在未来 18 个月内上线高达 8 吉瓦（GW）的电力算力。虽然 Gavin 直言自己“绝不看空埃隆·马斯克（Elon Musk）”，但在 18 个月内实现 8 GW 依然是一个近乎不可思议的壮举。自他们签订上一批合同以来，算力租赁费率不降反升，目前的变现效率约为每吉瓦产生 500 亿美元的收入。而市场对明年的普遍收入预期仅为 730 亿美元，如果算力部署真能接近这个数字，哪怕只点亮一小部分，其带来的体量都将彻底重塑公司的基本面。这甚至无需计入 Starlink V3、Starlink 手机直连（Direct to Cell）业务，或是 Grok 4.5 与 Cursor 结合后可能迅速冲破 100 亿美元的 ARR，更不用提成熟的 Starlink 核心订阅业务。 回顾历史数据，迄今为止全球仅有几家机构能在一年内上线超过 500 兆瓦（MW）的电力：除几大超大规模云厂商（Hyperscalers）之外，就只有 CoreWeave、Crusoe 以及 SpaceX。在这些选手中，SpaceX 往往以最低的成本、最快的速度交付了最大规模的算力基础设施，且客户对其实际集群性能的满意度极高。然而，公开市场对 SpaceX 的定位仍停留在航天与卫星通信公司，纽约某些对冲基金甚至正在建立庞大的做空逻辑，赌算力现货价格将暴跌 90%，认为大举建成的算力根本无法带来预期的收入回报。对此，Gavin 引用了马斯克的一句名言：“我们擅长的是让不可能的事情延期实现（We specialize in making the impossible late）”，幽默地指出哪怕计划延期，其最终兑现的物理交付能力依然惊人。 - **8 GW 算力部署目标**：公开报告称 SpaceX 计划在未来 18 个月上线 8 吉瓦电力容量，按行业现行变现水平测算极具颠覆性。 - **500 MW 俱乐部**：全球仅极少数超大规模云服务商及 CoreWeave、Crusoe、SpaceX 能在单年内实现 500 兆瓦以上的电力交付。 - **市场认知错配**：网络上流传“SpaceX 其实是一家数据中心公司”的梗图，但资本市场的一致预期模型中几乎完全没有定价这一重算力资产。 ### `01:03:00` 太空轨道计算的真实性与基准验证 随着星舰（Starship）近期的成功回收与着陆，太空基础设施的成本结构正在被彻底改写。Gavin 提到自己曾在得克萨斯州博卡奇卡的星舰基地（Starbase）待了很长时间，切身感受到“轨道计算（Orbital Compute）”概念已经不再是科幻小说，而是变得越来越真实。更具说服力的是外部聪明钱的背书：硅谷老牌风险投资机构 Benchmark 领投了轨道计算初创公司 StarCloud，而 SpaceX 也正在与其展开合作，计划向其开放关键的星链激光通信技术（Starlink laser technology），该技术正是轨道集群间高速互联的核心基础。 Gavin 坦言，当评估这样一个看似天方夜谭的赛道时，需要一个理性的参照物来进行思维校准。Benchmark 的合伙人们极其敏锐，且并不属于所谓的“马斯克生态圈”信徒，但他们愿意在没有 SpaceX 内部极低发射成本优势的前提下，以可观的估值下注轨道计算公司。这说明“也许我疯了，也许埃隆疯了，但如果 Benchmark 和 SpaceX 的顶尖工程师们也都认同这套逻辑，那它大概率绝非无稽之谈”。 ```mermaid flowchart TD accTitle: 轨道计算落地可行性验证体系 accDescr: 展示 Starship 低成本运载能力与星链激光通信技术如何结合，由 Benchmark 外部投资完成理性验证 subgraph SpaceX_Infra["SpaceX 核心底座支持"] Starship["星舰突破: 发射运力暴增与成本暴跌"] Laser["Starlink 空间激光互联技术"] end subgraph Ecosystem["轨道算力创新生态"] StarCloud["StarCloud 等轨道计算初创企业"] end subgraph Validation["资本与技术理性校准"] Benchmark["Benchmark 等非马斯克系顶级 VC 注资"] RealityCheck{"'我们都疯了吗？' 还是确定性趋势？"} end Starship --> StarCloud Laser --> StarCloud StarCloud --> Benchmark Benchmark --> RealityCheck ``` - **激光互联技术**：星链的高速空间激光网络被视为解决太空轨道算力节点间超低延迟互联的关键拼图。 - **估值与逻辑闭环**：独立顶尖 VC 对无垂直发射优势的初创企业下注，为整个太空计算赛道的可行性提供了强有力的第三方交叉验证。 ### `01:04:12` 对话尾声：基准餐桌旁的共识与时间裁判 节目的尾声洋溢着轻松而幽默的气氛，主持人与 Gavin 揭晓了本次录制的特殊地点——正是在 Benchmark 办公室那张见证过硅谷无数传奇晚宴的著名大餐桌旁。双方特别向 Benchmark 的合伙人 Sheila 以及协助牵线协调的 Eric 致以谢意。 面对当下资本市场的剧烈波动与对 AI 投资回报率的剧烈争论，Gavin 总结认为，未来是充满概率性的，但所有这些持仓与争论在一年之内都将由时间给出明确裁决，究竟谁在裸泳、谁站在历史正确的一边很快就会水落石出。考虑到基础模型以周为周期的狂飙迭代速度，主持人打趣道，如果按照新模型发布的节奏来更新播客，他们俩恐怕过两周就得再次坐在这张桌子前继续复盘了。 ## Notable quotes > `02:35` 把 2022 年整整一年的下跌浓缩进了一个月里。 — Gavin Baker > `08:22` 黄仁勋是个极具理想主义情怀的爱国者，他总是在做正确的事。但如果开源真的对他的生意有害，他还会成为全球头号开源支持者吗？ — Gavin Baker > `15:22` 在投资中，最危险的往往是你看不见的那颗子弹，最重要的三个词从来不是‘安全边际’，而是‘我不知道’。 — Gavin Baker > `20:00` 过去三年生存的最佳法则，就是以最快速度做出最蠢、最肤浅的反应，然后在这些反应之间来回轮动。 — Gavin Baker > `08:22` 一个 Token 就是一个 Token，生成它所需要的算力分毫不差。 — Gavin Baker > `41:46` 在这里，几乎所有人都比我更乐观。 — Gavin Baker > `53:44` 真相还在穿鞋，谎言已经走遍半个世界。 — Gavin Baker > `01:00:00` 我们擅长的是让不可能的事情延期实现。 — Elon Musk (Gavin Baker 引用) > `01:03:00` 也许我疯了，也许埃隆疯了，但如果 Benchmark 和 SpaceX 的顶尖工程师们也都认同这套逻辑，那它大概率绝非无稽之谈。 — Gavin Baker ## Mentioned Gavin Baker, Patrick O'Shaughnessy, Atreides Management, Nvidia, Jensen Huang, Anthropic, Claude, Claude Code, OpenAI, Meta, Mark Zuckerberg, Microsoft, Amazon, Google, SpaceX, Elon Musk, xAI, Grok, Cursor, Fireworks AI, Baseten, Modal, Together AI, Kimi K3, GLM 5.2, Nemotron, Blackwell (B200), Rubin, Hopper (H100), Ampere (A100), SK Hynix, Micron, ASML, TSMC, CoreWeave, Crusoe, StarCloud, Benchmark, Ilya Sutskever, Safe Superintelligence (SSI), Black Forest Labs, Michael Mauboussin, Dylan Patel, SemiAnalysis, Dwarkesh Patel, Harvey, Cognition, Scott Wu, Devin, Leo Li, Starlink, Starship ## Video description My guest today is Gavin Baker, founding partner and CIO of Atreides Management. This is our seventh conversation, and just two months after Gavin's last appearance. It's about the gap between what the market is doing and what companies are seeing. It's been a tough month or so for public AI names, but there's no sign of a slowdown on the ground in Silicon Valley. We discuss the latest moves, contracted vs. spot GPU prices, the game theory of memory supply agreements, and why Claude has become the Walter Cronkite of the stock market. We close on SpaceX, orbital compute, and what Gavin sees as the single biggest risk to all of it. Please enjoy this conversation, from the famous table at Benchmark, with my friend Gavin Baker. For the full show notes, transcript, and links to mentioned content, check out the episode page here. ----- Become a Colossus member to get our quarterly print magazine and private audio experience, including exclusive profiles and early access to select episodes. Subscribe at colossus.com/subscribe. ----- Ramp's&nbsp;mission is to help companies manage their spend in a way that reduces expenses and frees up time for teams to work on more valuable projects. Go to&nbsp;ramp.com/invest to sign up for free and get a $250 welcome bonus. ----- Trusted by thousands of businesses,&nbsp;Vanta continuously monitors your security posture and streamlines audits so you can win enterprise deals and build customer trust without the traditional overhead. Invest Like the Best listeners get a special offer of $1,000 off Vanta when you go to&nbsp;vanta.com/invest.&nbsp; ----- WorkOS&nbsp;is the infrastructure B2B and AI-native companies use to sell to enterprise. It covers everything enterprise security requires: SSO, SCIM, RBAC, Audit Logs, AI governance, and more. Trusted by 2,000+ fast-growing companies, including OpenAI, Anthropic, Cursor, and Vercel. ----- Rogo&nbsp;is the AI platform for finance. They're building agents for Wall Street that are trained to understand how bankers and investors actually do work: from diligence and modeling, to turning analysis into deliverables. To learn more, visit&nbsp;rogo.ai/invest. ----- Ridgeline&nbsp;has built a complete, real-time, modern operating system for investment managers. It handles trading, portfolio management, compliance, customer reporting, and much more through an all-in-one real-time cloud platform. Visit ridgeline.ai. ----- Editing and post-production work for this episode was provided by&nbsp;The Podcast Consultant. Timestamps: (00:00:00) Welcome to Invest Like The Best (00:02:35) First Question: July Was 2022 in a Month (00:04:08) The Private Companies Public Markets Can't See (00:05:06) Old GPUs Repricing Higher (00:06:53) Walking Through the Month (00:08:22) Kimi, GLM 5.2 & the Open Source Freak-Out (00:10:51) Real Yields, Spreads & CDS (00:11:54) Does the Build-Out Need Credit? (00:15:22) A Sell-Off With No Clear Villain (00:17:35) Open Source as Dark Matter (00:18:39) Nvidia's Lowest Forward PE in 10 Years (00:21:35) Claude as Walter Cronkite for the Stock Market (00:23:55) Continual Learning & Sample Efficiency (00:25:19) What Would Actually Scare Him (00:26:38) Routers & the Multi-Model Future (00:30:51) Tokens as a Percent of Comp Spend (00:33:37) The Game Theory of Breaking an LTA (00:36:41) Nvidia's Credit Wrapper & Revenue Share (00:37:45) What He'd Do If He Ran Hynix (00:41:46) Who's More Bullish than Him (00:43:28) China's DUV Machine (00:46:10) Bull Case for Software (00:48:16) The RSI Maximalist View (00:49:31) Inference Clouds Growing Without Burning Cash (00:50:35) The Biggest Risk Is Regulation (00:53:44) Telling the Story Better (00:57:15) Dark Horses (00:58:02) SpaceX in the Public Markets [Source](https://traffic.megaphone.fm/CLS6565465184.mp3)

[2:35] Gavin Baker: 把 2022 年整整一年的下跌浓缩进了一个月里。 [8:22] Gavin Baker: 黄仁勋是个极具理想主义情怀的爱国者，他总是在做正确的事。但如果开源真的对他的生意有害，他还会成为全球头号开源支持者吗？ [15:22] Gavin Baker: 在投资中，最危险的往往是你看不见的那颗子弹，最重要的三个词从来不是‘安全边际’，而是‘我不知道’。 [20:00] Gavin Baker: 过去三年生存的最佳法则，就是以最快速度做出最蠢、最肤浅的反应，然后在这些反应之间来回轮动。 [8:22] Gavin Baker: 一个 Token 就是一个 Token，生成它所需要的算力分毫不差。 [41:46] Gavin Baker: 在这里，几乎所有人都比我更乐观。 [53:44] Gavin Baker: 真相还在穿鞋，谎言已经走遍半个世界。 [60:00] Elon Musk (Gavin Baker 引用): 我们擅长的是让不可能的事情延期实现。 [63:00] Gavin Baker: 也许我疯了，也许埃隆疯了，但如果 Benchmark 和 SpaceX 的顶尖工程师们也都认同这套逻辑，那它大概率绝非无稽之谈。

[6:53] Gavin Baker: Jensen is the world's largest supporter of open source... But does it really stand to reason that Jensen would be the world's biggest supporter of open source if it was bad for his business? [46:10] Gavin Baker: If you can go from just using one, two, or three frontier models to using those frontier models for whatever it is, 30 to 60 percent of your token consumption and then use your own RL model, all of a sudden you're not a wrapper. You're way more defensible. [60:02] Elon Musk: We specialize in making the impossible late.

[0:00] Ramp is the only platform built to make your finance team leaner, faster, and better, saving businesses 5% annually on average so you can stay focused on growth. Ramp customers grew revenue 3.2 times faster than the average American business. Visa, Vercel, Cursor, Stripe, Notion, 11Lab, Shopify, and 70,000 other businesses all run on Ramp. Mine does too, and so should yours. Learn more at ramp.com slash invest. Felix by Rogo is a personal finance agent that turns a single prompt into finished client-ready work using your firm's own templates, context, and standards. Send Felix an email like, take these comments and turn them for me, or update my tracker with the context of these emails, or run the ability to pay math on this buyer, and Felix sends back finished PowerPoint decks, Excel models, and sourced research. Felix works the way your team already does, delivering work quickly and accurately around the clock. Learn more at rogo.ai slash Felix. The best AI and software companies from OpenAI to Cursor to Perplexity use WorkOS to become enterprise ready overnight, not in months. Visit WorkOS.com to skip the unglamorous infrastructure work and focus on your product. Hello and welcome everyone I Patrick O and this is Invest Like the Best This show is an open exploration of markets ideas stories and strategies that will help you better invest both your time and your money If you enjoy these conversations and want to go deeper, check out Colossus, our quarterly publication with in-depth profiles of the people shaping business and investing. You can find Colossus along with all of our podcasts at colossus.com. Patrick O'Shaughnessy is the CEO of Positive Sum, All opinions expressed by Patrick and podcast guests are solely their own opinions and do not reflect the opinion of Positive Sum. This podcast is for informational purposes only and should not be relied upon as a basis for investment decisions. Clients of Positive Sum may maintain positions in the securities discussed in this podcast. To learn more, visit PSUM.VC. Gavin, it's only been two months. Like the model release cycles, the gap between our podcast episodes are shortening. We're basically, you and I are basically on a model release cadence at this point. Well, I was sensitive to criticism that I think somebody pointed out that our podcasts were coincident with like local market peaks. And nobody can say that after this What on your mind It been a crazy Yeah I would describe July as 2022 in a month Yeah There are some fundamental negatives which we should talk But on the whole the ballots of fundamentals, I think, is improving significantly. Loads of AI names are down 50, 60 percent from their highs. We'll call it 40 to 60 percent in a month in a straight line. And I asked you before we started, you've been out here for the summer. Have you heard a single negative quantitative metric about AI? A single instance of deceleration? Nothing. Nothing. In fact, every metric is accelerating. And to your point, not just blind optimism from people excited about AI, but like, here's some data that they can show you from their different vantage points. Absolutely. I mean, however you cut it, whether you cut GPU availability, whether you cut GPU rental pricing, whether you cut the spot price of DRAM this month, token growth, everything is actually accelerated. And I do think a big part of the problem is one the market does not have visibility into anthropic open AI And then I would say these open source inference clouds that monetize inference here in America Fireworks, Base 10, Modal together. And the picture looks very different when you see that. Because open source has accelerated massively because of GLM 5.2, Kimi K3. And then Nematron continues to kind of chug along. We had a great, very small American open source model release. OpenAI has accelerated. Anithropic continues to grow really strongly and is almost certainly pumping out significant amounts of free cash flow. And I just think if there's this chart that everybody looks at of semiconductor cash flow going like this and hyperscale free cash flow going like that, and you're missing these private companies. But I also think that that chart misses something very important, which is just that you have everyone in 24 and 25. Even if you were really bullish, you thought that GPU prices, if you're really bullish, you thought they would decline slowly. If you're bearish, you thought it would decline precipitously. I don't think anyone in 24 or 25 thought that the prices of old GPUs [5:00] would be going vertical. Everybody thought, hey, we're going to be smart. We're going to sign these long-term contracts. And to some degree, like a lot of the neoclouds had to do that because they needed an offtake agreement to finance the GPUs. And so essentially you have the contracted base of installed compute trading at a massive discount to the current spot market. And as those contracts roll off and compute gets repriced higher, and SPOC can decline and compute will still get repriced higher. I think you're going to see a lot of acceleration that's going to answer these ROI questions. You've started to see that this quarter. If we look at operating cash flow, not free cash flow, operating cash flow from Microsoft, Meta, and Amazon has reported accelerated from 28 to 32. There are some actually pretty big unusual items now, like these hyperscalers, they always seem to have billions of dollars of legal expenses that are unusual, mostly fines to the EU. But there is an unusual amount of one-timers this quarter. And if you adjust for that we went from 28 to 35 That a material acceleration at this scale And that really before they start to light up the Rubens which will come at a meaningful premium before these contracts reprice It been a challenging month. Is it helpful to kind of like walk through the month, how we got here? Yeah. So first, Meta is going to rent out compute. And this is seen as like very bearish. They have excess capacity. They're going to cut CapEx. This is a disaster. This is not at all what it was. They just reported they didn't cut CapEx. What it was is they saw SpaceX have a big installed base of compute and sell some big trading optimized clusters into the market at a truly massive premium to these contracted rates. And at least analysts like that, they saw an opportunity. There's a lot of speculation they're going to raise capital. So maybe what they're thinking is like, hey, we will show on a small chunk of capacity that we can generate really strong IRRs, then we're going to raise equity capital and we'll be off to the races and probably raise CapEx. It doesn't look like that's what they're doing. But nonetheless the market sold off because it interpreted this very negatively And I was really sure it wasn negative A lot of telemetry into Metis CapEx plans None of that telemetry had shifted at all If anything they continuing to get more aggressive And then shortly after that, they released their best model in a long time, Muse 1.1, which is actually a very good model. It was overshadowed by GROC 4.5, but it was a good model. Way better than anything in two years. So just no chance they're taking their foot off the gas. Then Kimi comes out. And then there's this huge freak out about open source. And at the same time, this silicon data token index kind of dips and flattens. And the two are connected. What the silicon data token index captures is mix. And they don't see all the tokens. But because of GLM 5.2 and then Kimmy, although it took a while to layer in, there's kind of a mix shift in this data from more expensive frontier tokens, which probably have an inference margin. We can debate whether it's 80, 90, or 95, but super high towards open source tokens. And for whatever reason, the market thought this was negative, but the reality is a token is a token and you need the exact same amount of compute to make a token all else equal It takes the same amount of flops the same amount of memory the same amount of watts Tokens are not equal but broadly speaking all open source taking share does is take margin dollars out of the frontier model layer There is elasticity, thereby driving token demand. You need more demand for compute. And the margins, you know, Anthropic and open source, they all run on the same underlying cloud providers who charge the same amount of compute. So you're literally just taking margin from frontier models and essentially driving more margin dollars into the AI infrastructure layer. That was the catalyst. This combination of things. Well, yeah. Jensen is the world's largest supporter of open source. He is like a super idealistic guy. He's a patriotic American. I think he always does what's right. But does it really stand to reason that Jensen would be the world's biggest supporter of open source if it was bad for his business? You know, he'd still support if it was the right thing for the world. And by the way, I think open source is really important to worlds where there's just one or two dominant frontier models that charge like 90% margins. It's not good for humans. It might not be good for society. And I think we want a lot of models [10:00] as we've discussed before. So then it's like, okay, the market digests that and comes to turn with it. Then China has a DUV machine. Everybody's in these baskets. This causes a huge sell-off in semi-cap equipment. And then we get to what I think is, in a lot of ways, the real concern, which is real yields have gone up, which makes sense. We're investing a lot to fund this investment. And for sure, credit is an increasing part of it, even if the majority is still funded out of operating cash flows. So real yields go up and spreads widened. Meta priced a bond last week, and it did not price where you would think a Meta bond would price. And this just shows that the credit market- NVIDIA CDS was blowing out. All of the CDS, CDS for everybody is blowing out. And very smart private capital people just like, hey, this is just exactly what you'd expect. These are just banks hedging their commitments, but nonetheless, it doesn't look good. And these are undeniable facts. CDS is up, spreads are widened, real yields are up. That would be really, really scary if we needed debt to finance this buildout And that where I think it this differential between spot and contract pricing for the installed base of compute is so important It so important to understand what the financing will be like for the next six months or something The degree to which this build out is going to require credit. Right. Which would be the classic capital cycle. Absolutely. Overextend ourselves with debt. And that's where things get 100 percent. And then debt fueled build outs. They demand immediate repayment. So if supply and demand get a little bit out of whack, things can unwind very, very quickly. That's what happened to the Internet. If one believes, as I do, rightly or wrongly, after this month, I'm super open. You know, I'm looking like I've been pressure testing all of these. And like I really went deep on credit because, hey, this is real. It's undeniable. And if we need credit to fund this build out, this is a significant negative. And if you model it out, if you look at the amount of gigawatts that are supposed to come on and consensus estimates for hyperscalers, they're effectively modeled. And these are gigawatts of Blackwell and Rubin, Rubin being NVIDIA's next chip, Blackwell being the current chip. They are essentially modeled to monetize roughly at the rate of Ampere which is two generations behind not at Hopper but Ampere So there 1 to 1 trillion in hyperscale operating cash flow If you just assume I think it's very unlikely they monetize at the rate of Ampere. We could go into why. Some of it comes from just seeing what is happening on the ground with demand here for real quantitative metrics. But let's just say they monetize at a discount to current Blackwells. Then it's more like $2 trillion of operating cash flow. And that kind of takes $700 billion of credit demand out, ironically, as that improves all the credit ratios. As these installed bases of compute reprice, we're going to continue accelerating. Because since this is modeling at a deceleration, which I think is unlikely, then the credit metrics look better. And then all of a sudden, it gets easier to finance with credit. Now, whether they choose to do that or not, we'll see. This is all a little bit, you know, I think we spoke. Two months ago. No, but the time before that about kind of the risks of a Blackwell air pocket. Oh, yes. Where you're spending hundreds of billions of dollars on Blackwells. They're mostly being used for trading initially. Trading does not generate a return This could be a risk We actually really saw that in the first quarter I think one reason to the podcast two months ago I got comfortable with that risk was just that you were seeing such incredible things out of Anthropic. And then it's like, okay, well, the market's kind of gonna look past this. And it did look past it in April, in May, in June. And then in July, because of this kind of confluence of things, stopped looking past it. just as the operating cash flow started to really accelerate. And this is just a fact. It is accelerating at big scale. Like Microsoft, they brought on a huge slug of capacity in the month of June. That didn't even show up in the second quarter. So essentially, what this all comes down to is, do you believe that the quantitative demand signals seeing on the ground here in Silicon Valley from private companies are going to continue such that the installed base of compute reprices higher as contracts roll off, operating cash flows go up, and you could fund most of this out of operating cash flows, maybe all of it. Like if it reprices at current rates, you could probably fund all of it for the next several years. It has been a very unusual episode [15:00] in the market. We should talk about what the fundamentals are that are getting better that I'm talking about. Technicians would say it's actually in 22. Okay. The market is worried about a recession rates going up inflation. That's what the market was worried about in 22. You knew exactly what it was. Okay. Deep seek. You know what it's worried about liberation day. You know what it's worried about. There's something very clear and in a weird way, that's comforting. Sure. And here, you know, we talked about a lot of specific things, but it just feels all those specific things with the exception of credit are just kind of ridiculous. And so the fact that it is still going down, a technician would say, hey, that's a little scary. It's definitely the bullet you don't see that gets you. I think we've talked before about how I think the three most important words in investing aren't margin of safety, but I don't know. Oh, but just you've been out here for two months. I've been out here. I literally spoke to a company this morning who rented a cluster of several. And this is one of the sexiest startups that people want to be in business with. And they had rented a cluster of several thousand black wells And we just call it somewhere in the mid per GPU hour They renting the exact same size cluster essentially identical in every way B200s no differences. And they're hoping seven months later to pay just under $4 today. That's pretty crazy because, again, you would expect a really gentle decline in prices would be bullish. Instead, we're up, depending on the starting point, 50 to 60 percent in six or seven months. There have been so many anecdotes like that. Like I think one of the inference clouds, I think it was based in, I'm not sure, they went on a podcast and they essentially said, we are planning to pay 100 percent more for Blackwell's when our contract expires. And that just means that essentially all the hyperscalers are under-earning. my main kind of mission out here this week is like pressure test. Yeah. Yeah. Tell me something negative. Like, you know, the question I asked you, is there one negative quantitative metric you've heard has been what I've been asking everyone. The main thing people are saying is the third party data suggests that the anthropic curve started to go off of its trajectory a little bit That like the only thing that I think that may very well be true But then you have open AI and open source massively accelerating Yeah the complex You look at the sum It is net accelerating. Like, I think open source is a little bit of a, you know, they talk about dark matter in the universe, like open source is kind of dark matter to the public markets. It's hard for public markets to measure it. But like, if you just track what these inference clouds are saying, people saying things on podcasts or people saying things in meetings. They're not audited financials. Demand is clearly accelerating, which makes sense because you have this huge capability leap with GLM 5.2 and Kimi K3, which I think we're going to see continue. I think you're going to see NVIDIA bring Nebotron steadily closer to the frontier. But man, it has been a humbling, challenging month. But just it's also like, wow, I've kind of pressure tested every assumption. The underlying fundamentals are improving. NVIDIA is actually, as we record this, at its lowest forward PE of the last 10 years. Crazy. The only time the SIMIs have been cheaper were Liberation Day and DeepSeek, and those were kind of B bottoms. And that means to you just that the market thinks they significantly over Yeah the market 100 thinks they significantly over And we need to be humble Maybe they are. Maybe they are. But my kind of mission out here this week was to look for negative data points as hard as I could. And normally you come to Silicon Valley and there's a mixture of here's something negative, here's something positive, da-da-da, unbalance it's positive. If you know tech, it creates value over time. But I haven't been able to find one that is like a quantitative metric. That anthropic third-party data, I would say that seems to be hotly contested by the anthropic shareholders who are chomping at the bit to tell you what they know. We're also very scared they're not going to get an IPO allocation if it gets back to the company that they're the ones who said, actually, things are great. You know, you can just see Anthropic shareholders. Like they want to be like, it's not true. You know? It's hard for me to believe that open source and open AI have accelerated to the extent they did. But yeah, Anthropic is clearly in the pole position. And oh, by the way, Grok and Cursor have also, you can see from third party data, like July was a pretty transformational month with Grok 4.5, Grok builds coming out. [20:00] So it has been a tricky month. And I have a friend, a friend of fidelity, who just says the way to have navigated the last three years is just do the dumbest, most superficial thing as quickly as possible and just cycle between them. What is that now? Yeah. Well, that's just that has been to cut risk all month in response to these narratives that factually, except for credit, are not true. And the work we've done makes me think that credit just isn't going to matter as the street prices. Let's just say you do need credit to, like, build the flops we need. Well, if credit's not there, it just means the flops that are there are going to be even more valuable. And then eventually that will improve the metrics. And then it's like credit is there. So as long as we're in a compute shortage, which I'm just like desperately trying to find a single sign that we're not in one and that it's not actually getting worse almost by the day. It's almost like the problem becomes the solution. And then this company Black Forest Labs I think that their name I hope I got it right Because there is an interesting essay that got sent to me I think we talked before about Mike Momison theory that breakdown in diversity is kind of what leads bubbles and crashes And essentially everyone I know in the public equity investment business, whether retail or institutional, every piece of news gets fed into Claude. And Claude, Claude code, sometimes a Claude agent. And it's probabilistic. There's probably not that much variation in the way it's interpreting this news. It's almost like we're back to, in stock market terms, there's never really been this way in the stock market before, but people talk about the fragmentation of media and how it used to be like Walter Cronkite, only voice of truth, and now we don't have that anymore. It's like Claude is kind of Walter Cronkite for the stock market, and everybody just believes whatever it says. By the way, It's really smart, but it's not always right. It's interpretation isn't always correct. And with the stock market, you are fundamentally dealing about, you know, a probabilistic Bayesian interpretation of the future. It feels like in the market here this piece of news It gets fed through Claude Claude interpreted this way It a huge chunk of people trade on claude view and so you seen stuff there this guy tbu he like a part of the anonymous semiconductor mafia on x yeah actually very smart guy i know him in real life but he posted this amazing chart of japanese capacitor stocks and he said we've had an entire capacitor cycle in six weeks and it's true you know the stocks like whether they double triple or quadruple I don't know, but like vertical and then whoosh, like the actual fundamentals haven't even hit. And yet you've already had what probably would have normally been a three-year cycle in like six weeks. Vanta automates security and compliance for over 16,000 fast-moving companies like Ramp, Cursor, and Harvey, keeping them audit-ready around the clock. It's the number one agentic trust platform, and it now helps companies like yours watch for the risks that show up between audits across your vendors, your AI tools, and your whole environment. Every new tool your team signs up for, every vendor that turns on AI features is an opportunity for something to go wrong. And most security programs weren't built for AI's pace of growth. The Vanta agent works like a 24 seven GRC engineer in the background finding issues drafting fixes for you and cutting vendor assessment time by up to 50 Whether you a fast startup or a global enterprise Vanta helps you earn and prove trust. Invest like the best listeners get a special offer for $1,000 off at vanta.com slash invest. Ridgeline is the first end-to-end system of record with embedded AI for investment management firms, running portfolio accounting, reconciliation, reporting, trading, and compliance on one unified platform. Firms are moving off legacy technology and onto Ridgeline because of how far ahead Ridgeline's AI features are compared to anything else in investment management software, which is why I believe that firms that come out ahead in the AI era will be the ones running on Ridgeline's unified platform. If you're serious about your firm's AI strategy, Ridgeline should be part of that conversation. You can request a demo at ridgeline.ai. What's your sense of being out here, especially it makes me especially curious about this, the innovation that is going on here to improve the efficiency and every aspect of serving inference of training models, et cetera, and how that will affect like public markets over time. Like, have you learned anything interesting about the long lead time innovation type stuff that has you especially excited or excited? [25:00] Curious? Yeah, I am very curious. A lot of people seem to feel like they are very close to solving continual learning and sample efficient learning, which we've talked about before. And it is possible that if those are solved, that could that be like a temporary discontinuity in demand if instead of I was trained on effectively 20 billion tokens and that it's like these models are trained on 300 trillion tokens. And if you can trade something on 10 trillion tokens and then let it out into the world and learn sample efficiently, that doesn't sound good for trading demand. But trading has a percentage of semiconductor demand to compute is going to asymptote to something not approaching zero, but very small. But I would say that is the most interesting. And who knows if it's long horizon or short horizon. SSI says that they're going to come out with their model in August. There's this whole generation of new labs that are focused on this. And this would be good for the world. This would be amazing. To be clear, yeah. This would be awesome for the world. We all want this. Yeah, we want this. It would be amazing for the world And it just it hard for me to believe that that would actually be negative for AI infrastructure demand But again trying to be really really open I would say that was probably like the biggest scientific or technical takeaway. You know, it's also- I just don't know. Well, yeah. And also like NVIDIA is heavily involved with all of these startups. If you were just forced to come up with the set of circumstances that would really switch you around and get you really scared, would it just be that this operating cash flow thing doesn't play out and therefore we just need to debt finance this? Yeah, if the operating cash flow does not continue to accelerate, that would be negative. And that, to some degree, is going to be a function of how Anthropic, OpenAI, GrokCursor, and open source do. If there was a pretty dramatic contraction in GPU prices that was kind of sustained, the market would react to that instantly. That would be worrisome if it started to get to be really easy to get GPUs. I mean, have you heard anyone say they have too many GPUs? Like not a single person. It's not like it's the opposite. It sounds like a drug market or Yeah it really does It just wild But yeah I mean I think there a long list of pretty obvious things If the sum of these labs plateaus or starts to decline that really negative unless it just because open source tokens are net growing the pie and taking share. And I do really think the future is multi-model, particularly for the AI natives. They're going to want to take an open source model. It's got all these inference clouds have gotten really good at supervised fine tuning and reinforcement learning. So you can take your data, customize an open source model, and then get something that you can put behind a router. And the router routes it to often first your model and then quad frontier model, whatever, quad grok checks it. And you can, in a lot of cases, get slightly better outcomes at half the cost. But again, that half the cost, I think a lot of people hear that They're like, that's bad for AI demand. It's actually not at all because the cost the user pays is just a function of the margin on the tokens. And you're literally just shifting tokens from really expensive tokens with like 90 gross margins to tokens with maybe let call it a 30 gross margin And that where the savings are coming from but the tokens cost the same amount of compute to produce And then also all these things are kind of happening on different cycle times. All these big public companies are like, oh my God, my AI spend is 20X, I've burned my budget in three months. So they set up a router and that actually cuts their AI spend, but it doesn't really impact, It may actually increase the amount of tokens that they are generating just by shifting them to these cheaper open source tokens. And that's just more compute. So a company getting smarter about which model to use for which task, that may lead to a stabilization of their spend or even a decline. But it actually has nothing to do with the amount of GPU compute hours they are effectively consuming behind these model layers of this router. The GPU computer hours probably are going up as you shift to these cheaper tokens you can use more of. That's happening to like a cutting edge of public companies. And then you have this whole wave of AI natives. They're leaning into this so hard and they're not hiring humans. They're just putting it mostly into tokens. They're not slowing down. [30:00] And then you have companies on the East coast of America who have like barely adopted AI companies, broadly speaking, you know, not on the coast who maybe are as cutting and then Europe, who's just trying to figure out how to regulate it. Before using it. Yeah, so just like there's kind of these differential waves of adoption all happening at the same time. But the thought I can't get out of my mind is like, I think I said it maybe last time, but just you access it out. Like 500,000 people in the world, 250,000 maybe are using agentic AI and we're in an acute compute shortage. There's seven or eight billion people on the planet. What happens when we go from 500,000 to 100 million, you know, to 500 million? It is interesting. You know, a lot of people, I do think it's like helpful to post on X to see the pushback. And a lot of people are saying, we accept your argument that hyperscalers are under earning and it's compute reprices. Their operating cashflow is going to accelerate and maybe we could fund this. But like, where's that operating cashflow going to come from? Where is the customer? And kind of definitionally it has to either come from faster economic growth through productivity kind of such as comments like either we going to start growing 10 or we not or labor substitution And for sure I think in a lot of these AI natives, you're seeing labor substitution, but not because they're firing people. They're just not hiring nearly as many humans. The gross profit dollars per FTE and A16Z, Iconic, a bunch companies that have done this work, they're vertical, particularly relative to past generations of startups. And then it is interesting. Are you doing any surveys of your companies and their token spend relative to labor spend? Oh, yeah. I mean, it's tokens as a percent of total comp spend or something like this. What are the ranges you've seen? I mean, like in the really pilled companies, like it gets really high, 20 percent, 25 percent. Our friend Dylan Patel at his company, So he's an ASI maxi, but he's at 30%. That's probably the highest one I've heard. I've actually heard of 50. And there's $25 trillion in knowledge work. Let's take your 20% number. That's $5 trillion. And that either comes out of labor substitution or faster economic growth. And we really really really want as humans to come from faster economic growth One interesting thing I heard this morning from one of the great leading technology CEOs that founded several companies If you look at the founder and controlled companies and adjust for some of the COVID era overhiring nobody really laying people off. These are the people that would probably be most quick to adopt AI to become more efficient or whatever. They're not really doing, jack aside, huge scale layoffs, which probably tells you something about where they think there will be lots of opportunity to still have people plus. 100%. Well, the bull case, you've seen church from cognition, ramp and stripe that the companies that are spending the most on AI are growing meaningfully faster. Yeah. I love that cognition index. The cognition index is wild. All the skeptics will point out rightfully. It's not really controlling for industry, but then if like you dig down into it, I think one of them gave an example of, I forget if it was a plumber or an HVAC contractor, but like everybody who's a blue collar workers doing great because of AI, by the way, something that I think we should touch on, and we can do it now or later, is just everybody is signing these LTAs. So everything's at a shortage. If there's weakness, it's just because we can't energize the gigawatts fast enough. The gigawatts are going to get energized, like regulatory policies moving in a good way The turbine manufacturers the diesel jet manufacturers you ripping turbines off old airplanes and reconditioning them and then repurposing them There crazy things happening Capitalism is very, very good at this. But I do think one of the most important questions of the market and like a transition of the market that I got wrong is we are shifting, particularly for memory more than anything else, from crushing numbers in the short term to their trading short-term upside for these, what they call supply chain agreements, long-term agreements, LTAs. There's many flavors, but customer prepays, there's a floor and a ceiling. And this comes back to the point about labor, because a lot of people after firing too many people during COVID were really reluctant to lay people off. They talked about labor hoarding, if you remember a few years ago. You remember this? Yes. Let's just think about the game theory of breaking an LTA. So there's four companies that matter at scale. There's Amazon with their tradiums. There's Google with their TPUs. There's AMD. And then there's NVIDIA, who's like much bigger than everybody else combined. Let's just say it's 2027. And it's very important to realize as memory is... [35:00] More memory you put with flop for a given unit of compute, the more tokens you get out. It's the single most important thing you could do to increase token output per unit of compute. And then that obviously, definitionally, actually lowers costs, which is why the demand hasn't responded at all negatively. There's been no elasticity just because it's the access that is dominating all others. and this is at some level like a giant Game of Thrones or IMPERS between these companies. Okay, it's 2027 or 28. You're vaguely tempted to break one of these LTAs and try and get a lower price. But to a large degree, market shares, I think for the next several years, are going to be determined by supply chain allocations and kind of what you have pre-purchased. So if you break the LTA, this is assuming we're not in a severe oversupply situation. The game theory even holds in a severe oversupply situation. If you break your LTA, and then in the next two or three years, for any reason, leverage shifts back to the memory, guys, you out of business It over Let just say Google breaks an LTA There an oversupply I making this up at 28 29 They break their LTAs Well if they breaking their LTAs it probably means your oversupply prices are coming down and then capacity naturally contracts. Well, what do you think is going to happen to Google's allocations? And then this is a cyclical industry and oversupply is followed by undersupply. What do you think they think is going to happen to their allocations next time. So I just think given that this is the axis around which kind of everything is revolving, you might blow up your entire business and your franchise by breaking an LTA. And that was never the case before. Apple, who cares? They don't have a competitor. They're overwhelmingly the largest purchaser. This is going back three, four, five years. They know they can do whatever they want with no consequences because their volume is so big that even if they like super screw Hydex, Micron will of course take them. This is just different. You have at least four players. Did you have all the startups? You're an investor and etched. If you break an LTA, they just say, okay, fine, great. You broke the price agreement. We going to break the volume agreement and screw you We going to give the volume to your competitor You just lost share you know NVIDIA dominance the current environment the extent to which it favors NVIDIA it is a little hard for me to understand why it's trading at such a low multiple. In other words, if you need to be able to finance the chips, and you do, nothing's more financeable than an NVIDIA GPU. Nothing. If you need to get land and power, well, they're doing a very good job of playing that chess game and matchmaking. And then they've rolled out this really clever new business model, which I would describe as kind of like a credit wrapper with a revenue share if GPU prices are above the floor. Yeah, this could lead to them having a really giant cloud business effectively through royalties really quickly. And it is another way of kind of alleviating this cash flow mismatch. Like, hey, we're making all the cash. this isn't really vendor financing because they're not loading them the money somebody else is loading the gpu buyer the money they're still making equity investments but it's not like you're just putting money into someone that some of that money was used to buy your chips even though nvidia said that they write into all their equity investments that the money can be used to buy nvidia chips but obviously money is fungible and um funny thing makes no sense sense Yeah But you know I think at some level it probably makes everybody feel better What would you do if you were the member, like if you were the CEO of Hynix? I'd do the exact same thing NVIDIA is doing right now. Which is? I would be going to the buyers of GPUs, Tradiums, and whoever and say, I'll participate in the NVIDIA credit wrapper. Now, their business is just inherently less stable and predictable. but in some way, and maybe they just put up some cash up front, so it's like they're not on the hook. I'm just making this up. But do something. Because you have money now, and credit markets are revolting. I'm sure our friends at Blackstone and Apollo are suggesting some variant of this to the memory companies. But hey, we will put up some amount of money from our cash flow today, and then it's gone. It's surety that makes the person who's extending the debt feel better. But we want some sort of a cut of the ongoing revenues as well. That is 100% what I would do. And it's almost like a logical extension of the LTAs where they're trading upside for durability. Here, [40:00] you can effectively get a royalty on recurring revenues. And that is what NVIDIA is doing. And I do think that is very misunderstood. And I think it would serve NVIDIA well to really explain this. One, they're really bullish on AI. Essentially, every time they haven't taken an equity stake in something, it's been a mistake. They've taken equity stake in everything, essentially, except the memory companies that for a long while, Anthropik, that they took an equity stake in Anthropic. But like, why not if you have cash flow and you're bullish on AI and Jensen because he sees every lab, he knows all the advances, you know, like all these continual learning labs, you know, safe super intelligence is now working with them. He sees everything and like what he sees makes him bullish. So what have some equity upside and then to have a revenue share and you're generating hundreds of billions of dollars of free cash flow and helping to kind of bridge what is clearly kind of a gap, at least given everybody's got free cash flow negative, until the operating cash flow accelerates enough that you can internally fund this. It very opportunistic in a good way and it significantly increases their revenue per gigawatt And then it also strengthens their competitive position You and I we both have startups but okay that great Use that startup chip What prices are they paying at Taiwan Semi Higher than NVIDIA and all these guys. What prices are they paying for HBMD RAM? Higher. Can you finance those chips easily at the same rate as NVIDIA? No. And so it's always like there's a real burden, particularly if you use HBMD RAM, you're in the crosshairs of this. Unless like actually they made really different architectural choices. Everything that's happening is actually pretty good for him. By the way, going back to game theory, Anthropic, if they had been as aggressive on compute as OpenAI had been, they would have run away with it. Now OpenAI is back in the game. I think Grok is in the game. those are the companies on the Pareto frontier. And they have the compute. Do you think after watching that, anyone is going to let off the gas? Right. It was, I think, four months ago that Dario was talking about how it was a really thoughtful commentary, but he's like, it's really, really hard because if you buy too much compute, you could go bankrupt at the scale of these things. But if you don buy enough you could lose Well OpenAI just got back into the game and now SpaceX is in the game in a big way with Glock 4 and Cursor After watching that from a game theory perspective, is anybody going to back off anytime soon, especially if it can be funded out of operating cashflow? Have you met anyone in your travels out here that you would say is like way more bullish than you? And if so, what do they believe that you don't? I mean, essentially everyone out here is more bullish than me, man. You know, I read this thing that Dorkesh wrote and I was like – The 3x compute price thing or whatever? Yeah, well, he was – I forget what it was. No, no, it was like 15x or something. Yeah, but just basically that renting an H100 for a year would cost $250,000. And that's 15x the current spot or something. Exactly. Like, wow, you know, that was just like – That wasn't in my book. That wasn't in my – forget my like Bayesian probability space of expected outcomes. that wasn't even in my considered but dismissed his totally unlikely outcomes for cash he's very smart guy he's very plugged in then he pointed out that margins on compute are going up the amount of compute is going up and inference margins going up and if you multiply those three that how you getting this crazy acceleration into some of the labs plus open source although the margins on open source are not really going up I look at what happening in the stock market and I feel like a foolish optimist And then when I talk to people, whether it's people at the labs, anyone in this ecosystem, I'm like bearish relative to essentially everyone. Just a strange state of affairs. What do you make of the DUV news out of China where I've seen reactions really along a spectrum of like, this is the equivalent of like what ASML had in 2001 or something. Or like, no, this is actually the first bit of news in a new story for how we should think about the global supply of cutting edge compute. I think both could be true. Make an analogy. Like, let's just say a DUV machine was a jet turbine. And now an EUV machine is like a warp drive. to UV machines like a propeller plane, UVs like a jet turbine. They didn't have it before, and now they allegedly do. And that is like a phase transition. You've gone from liquid to solid. Now, that solid, that jet engine, prop plane, whatever, is 25 years behind. But still, it's important, and I don't think it should be dismissed, but I also... [45:00] So, you know, it's kind of funny. You just see this in the stock market. The stock market massively overreacts. And then if this ever hits ASML's orders, maybe it hits it in five years. And like the market has forgotten about it, got worried about it, forgotten about it, got worried about it, forgotten about it multiple times along the way. So I do think that was probably an overreaction, but we shouldn't dismiss that either. And if you're China, like this is really important to you. You know, there are some reports that like an EV machine had been smuggled into China. I mean, what a feat of espionage because those things are like huge. I don't know if that's true. You know, there's some noise about it. But, you know, China, they're really, really good. They're really, really smart. They work really hard. They see this as super important for them as a country. But are they going to go from the year 2001 to 2026 or even 2030? It's a learning by doing. And you can't accelerate the doing. You can't teleport into the future. You actually have to go through those learning cycles. Is it significant? Yes Did the market overreact Probably It very hard as an American to really understand what is happening in China and have total conviction and clarity For better or worse we are decoupling. That is a process that has been set in motion. And at this point, it almost feels like self-reinforcing on each side. That's unfortunate. We are where we are. They're not going to stop. Neither are we. Any commentary on like every other company in America? Like I feel like right now it is 10 companies, couple private. Not last month. Everything but AI was vertical. And I do think open source getting closer to the frontier and companies like Fireworks making it really easy to customize a model such that you can get, in some cases, better than Frontier Performance for a meaningfully lower cost. That is a godsend for the software industry. And it's also a godsend for all these AI natives. You know, it's like our friend Vishria, I think he said two years ago, I've never seen more companies go from being founded to like $50 million a year in revenue and generating cash flow in like whatever it is, nine months. And it's hard to know if any of them are durable because a lot of people would dismiss them as chat GPT wrappers Well now with open source you generated some data that unique to your use case whatever your vertical you're going after as a wrapper is. Fireworks did come out with a really cool product called Nexus. And if you're using Cloud Code, OpenAI Codex, GrokBuild, it is literally three lines of code, like 20 words. And Fireworks ingests your data. They can RL a model, and then there's a router that sends the query, and they've had amazing results. And this is kind of the solution for every AI native. That's why you saw Harvey before it was acquired, Cursor, lean so heavily into this. Harvey, Lagora, all of them. Because if you can go from just using one, two or three frontier models to using those frontier models for whatever it is, 30 to 60 percent of your token consumption and then use your own RL model, all of a sudden you're not a rapper. You're way more defensible. I was so interested by that cursor thing that came out. I think it was cursor where it's sort of like AI is speed running, like what we learned amongst humans which is you could use the frontier model to plan and then farm out tasks to the dumber models and it 15 times more efficient or whatever the metric was And it may be and this is like super ironic lower margin open source tokens that are just a little bit behind the frontier. We have friends who believe that once a frontier model hits RSI, it will actually have a dramatically lower cost to serve at every level of intelligence by kind of distilling this. And then there's no place for open source. And I would say that's like a anthropic, open AI, Grok, maximalist view. We shouldn't dismiss anything. Anything is possible. We want to be very humble. I particularly want to be humble after the month I've had. But that doesn't seem that likely to be. One, because there are so many of these AI natives that have actually generated a decent amount of domain-specific proprietary data. And before open source had this moment and these inference clouds and these routers really developed, you kind of didn't have a choice. Like whatever the terms of service were, you accepted them. But if you can now get off that treadmill, that gives you a degree of [50:00] independence, maybe durability, safety. But going back to your point, it may be that these cheaper tokens massively inflate the value of the most cutting edge frontier tokens. Because if today, if you have, I'm going to make this up, 120 IQ open source models, and they're really cheap to run, doesn't that make a 160 IQ model that can orchestrate them more valuable? We talked last time about how I've been really surprised that so much of the economic returns have accrued to the frontier. That is changing with what we're seeing with these inference clouds, Together, modal, base 10, they're all working in a very cash efficient way. What's shocking about those business models is they're growing almost as fast as the frontier labs in the early days, but burning very little cash. It's pretty extraordinary to go back to silly SaaS metrics like the rule of 40 perspective, these are crazy numbers. Do you think there's a lot of instruction in just the distribution of pay inside of an organization? Like the CEO makes X times more than the median person at a company and maybe that frontier tokens versus open source tokens Yeah Something simple Yeah it may be that what we discussed last time where you know frontier tokens like the pie is growing really, really fast. They may continue to capture the overwhelming majority of economic value, but not all of it the way they have been. And open source tokens might be the majority of tokens processed. Again, going back, that's great for infrastructure demand because a token is a token. And it takes the same amount of flops, watts, space, cooling to make. What's the worst thing that could happen in AI? Is it regulatory? I think regulatory has to be the biggest risk. It's the most obvious risk. And so that was kind of one reason I was excited to be here this week. I want to be scared. You know, I like I don't want to feel like a lunatic watching these stocks get cheaper, thinking the expected forward returns are going up. While it feels like the on-the-ground fundamentals have pretty materially improved in July relative to even June, but I still come away thinking like regulation, it just has to be the biggest risk. Like you just can ignore New York making a data center moratorium We living in this weird post post political world And I mean I think the AI industry it has done a terrible job of PR And I do think I think it at least realizes that now. Maybe if not fixed it, it realizes it. Yeah. But like the political narrative, I think amongst a lot of ordinary Americans is like data centers, they're going to raise your electricity prices. They're going to take all your water, And then they're going to take your job. The reality is, given the deals that are being cut now, when a data center goes in, electricity prices actually generally go down for everyone around there because of behind the meter deals. This is that data center pledge that Trump asked people to sign. Generally, the data center developer used to be they just had to get the police department or the fire departments new trucks and new cars and new body armor or whatever. Now it's like, we're going to build you a hospital, a school, a new police station, and a fire station, and we're going to lower your power bills. How does that sound? And by the way, the jobs are ongoing because it turns out that you kind of need these plumbers, electricians, HVAC contractors, data centers are in a lot of ways the best thing to happen for blue wages in my lifetime And yet you have the Democrats who ostensibly represent the blue collar workers taking those jobs away It just kind of wild how what is the phrase Like a lie could go around the world. Faster than truth gets out of bed. Yeah. Yeah. Faster than truth gets out of bed. But an author made a mistake in a book. It overestimated the amount of water usage in data centers by 10,000 X. Not a little bit, like not one order of magnitude, not two orders of magnitude, not three she's admitted that mistake many times i was completely wrong it's like been super debunked like the popeye effect you ever hear that example no you know popeye spinach the reason was same deal in an academic book they placed the decimal two things wrong so spinach does not have more iron than everything else it was just this one source and then that propagated people still say it has more iron i literally had i thought it had more iron i mean that's like 80 years ago that's wild I literally thought spinach had more iron. That's amazing. It'd be crazy. Yeah, you learn something new every day. Yeah, same thing though. Yeah, it's the same thing. And it's just, so somebody just needs to tell the truth. I feel like the industry, geez, maybe if nobody else is going to do it, like I'll do it. There needs to be some sort of foundation. Maybe it's a pack that runs ads during the final four, during NFL games, during college football games. [55:00] games. Here's the virtues. World Series. Here's what a data center does. Your power, a data center that signed this pledge in your community, your power prices are going to go down. They're almost certainly going to contribute to the community in a material way. You're going to see a massive influx of super high playing blue collar jobs that are going to persist. And I think a lot of people thought that they were one time and they're just not like there's for sure a spike. And then that moves to the next data center. But there is an ongoing need for RMA and then upgrades at these data centers and technology is changing. So you're going to have more jobs. You're going to have cheaper power. You're going to have a wealthier community. There's going to be no impact on water, no impact on the environment. But it's easy to build the data center 10 miles out of town. You know, that story needs to be told along with we heard a story, I think we talked about it last time, about how AI is increasingly really saving lives, curing rare diseases. I think it was at ASCO this year. The vibe was like, this is the most scientific breakthroughs we've ever seen at a single conference. And for sure, some of that is due to AI. And so we need to tell those stories Like you know if you have a sick child sick parent a sick loved one like AI meaningfully increases the odds of them recovering Everybody needs to tell this. And I think people out here, all of this is so blindingly obvious to them. They can't process that this is a true but wildly divergent view from most Americans. The industry really needs to tell its story better. New York, it just feels like is the first of many. And even in some of these deep red states, they're super pro-growth. They're just like, hey, you guys are not doing a good job telling your story. We can't tell your story. If you tell your story, though, we can retell it. But like, you're the experts. If you do not speak your own truth, no one else will. What have we missed? I do think something that is missing from all of this conversation about compute is what is going to happen when you put these sram based accelerators that are not constrained by hbmd ram that are often made on older nodes that are not competing with like the latest greatest gpus when you disaggregate inference people talk about pre-fill and decode but decode is two parts attention and feed forward network and like the ultimate holy grail is if you could do pre on one chip that probably doesn have HBM DRAM do the attention on a super high powered chip with HBM DRAM and then do the feed forward network on one of these SRAM chips But like the ROI on adding these SRAM accelerators to the existing installed base of compute and new compute. But like what we're seeing is you do better. You just can't beat SRAM in particular for that feed forward network. And no matter how much you try and get the ratio of compute to HBM DRAM to SRAM on the chip correct, the workloads are always changing and there's different workloads. being able to disaggregate into these three parts, this is going to be really, really positive for the ROI on AI. For some reason, I just thought of a funny question, which I love their framing of Game of Thrones versus all these people. Can you imagine a player that is not currently on everyone's mind becoming relevant at like the major Game of Thrones scale? Like that could be like Micron all of a sudden, someone that becomes as important as Anthropic, OpenAI, Microsoft, Amazon. So like a dark horse. Like Leap Boo is probably a dark horse lit at fireworks She is an absolute killer I think our friend Scott Wu cognition is kind of You here to that one Yes. I think those are the most obvious names. What about SpaceX? What's it been like watching that be digested by public markets, at least initially? Do you think the market understands it as a company, the most important new company to be public? It doesn't really feel like it does. The fundamentals have gotten better since an IPO, like ROK 4.5, the Cursor acquisition. Cursor has clearly accelerated meaningfully. They've shown over the last three years they can bring on more compute faster than anyone at lower prices. And now we know that they can, even adjusting for the spot first contract gap, their big advantage was they came into the market, hit those spot highs. And in a strange way, like one of the more bullish things for compute is they put a vast amount of compute into the market overnight. And it wasn't even really a blip. It was like the market just utterly absorbed it. The freight trade didn't slow down at all. A sub stack writer will fund to AI. They think that SpaceX is going to try and bring on eight gig [60:00] of compute over the next 18 months. So eight gigawatts over the next 18 months. I will never bet against Elon, but I mean, that would be a truly incredible feat. Rates have gone up since they signed those last contracts, not down. And they're monetizing at something like 50 billion a gig. And consensus estimates for next year are 73 billion. So forget Starlink V3, forget Starlink Direct to sell. Grok, 4.5 and cursor. I think that the sum of that probably hits a $10 billion ARR pretty quickly. Forget all of that. Forget the core base Starlink business. If they bring out anywhere near that, the consensus estimate is 73 billion, and that's eight gigs at 50 billion a gig. And obviously, that would not all be lit up at the beginning of 27. And it seems very implausible to me. like I almost don't believe the funder report. But to this day, the only companies that have brought on more than 500 megawatts of power in a year are the hyperscalers, CoreWeave, Crusoe, and SpaceX And SpaceX has kind of brought on the most the fastest at the lowest cost And then people do actually really like their clusters But again it kind of like the market is going to need to see that That would not be the market's interpretation of SpaceX today. No, no. And it does feel like, you know, there's this big New York hedge fund short case on it. And I think they think the spot price for compute is going to go down 90%. You're going to bring on all this compute. It's not going to generate, you know, nearly as much revenue as you think. Maybe, but I also want to be really clear. Like I've seen Elon's companies do really impressive things. The funder AI report of 8 gigawatts in 18 months. I mean, I'm just quoting that because it's public. It's available to everyone. I think one of Elon's phrases is we specialize in making the impossible late. I've never heard that. That's great. You know, there's kind of a lot of truth to that. Yeah, yeah. but I just think very little is built in from my perspective to that stock for the amount of compute that they might be able to bring on. And again, I don't think it's anywhere near eight, and it's going to be really hard. And energizing these GPUs is really hard, but they been good at it And it doesn feel like that in estimates or really in people thinking I thinking about that funny meme that says SpaceX the data center company Oh 100 Yes absolutely And then I would also just say like I did spend a lot of time at Starbase and orbital compute feels more real every day. Pretty cool to see that Starship landing the other day. Pretty cool to see the Starship landing and that it's, you know, it is funny. Our friends at Benchmark, they funded StarCloud. And I don't know, last time StarCloud is an orbital compute company that like SpaceX is kind of partnering with. They're going to, I think, let them use the Starlink laser technology, which is really important for orbital compute. But I do think that's like kind of a good sanity check. Last time I checked, the benchmark guys were pretty smart and they're not coming from the Elon ecosystem at all. And they chose to fund an orbital compute company, a decent valuation without the internal launch costs that SpaceX gets. To me, that's a good like, hey, am I crazy? And it's like, well, maybe I'm crazy and maybe Elon's crazy. And maybe Benchmark is also crazy. And maybe the SpaceX engineers are also crazy that, man, that just doesn seem that probable to me Should we say whose offices we in Yeah we sitting in the Benchmark office Yes this is their famous table for their famous dinners So thank you Benchmark Thank you Benchmark for this episode Yes, thanks, Eric. And Sheila, we should thank them all. Eric coordinated for me, so he gets a special shout out. Thank you, Eric. Thank you, all the partners. Thank you, Eric. But I mean, we will see where all of these stocks are in a year. And the great thing is, time will tell. People are going to be right or wrong. The future is probabilistic, but it's an exciting moment. Well, if we keep doing this on the model release cycle, I'll see you in a couple of weeks. Yeah, it's crazy. You're going to benchmark. It's always a blast to do with you. If you enjoyed this episode, visit Colossus.com. You'll find every episode of this podcast complete with hand edited transcripts. You can also subscribe to Colossus, our quarterly print, digital, and private audio publication featuring in-depth profiles of the founders, investors, and companies that we admire most. Learn more at Colossus.com slash subscribe. Thank you. [65:00] pound over time. That's true in investing and just as true in how you run your company. Your spending system is your capital allocation strategy. Ramp makes it smarter by default. Better data, better decisions, better economics over time. See how at ramp.com slash invest. As your business grows, Vanta scales with you, automating compliance and giving you a single source of truth for security and risk. Learn more at vanta.com slash invest. The best AI and software companies from OpenAI to Cursor to Perplexity use WorkOS to become enterprise ready overnight, not in months. Visit WorkOS.com to skip the unglamorous infrastructure work and focus on your product. Ridgeline is redefining asset management technology as a true partner, not just a software vendor. They've helped firms 5x in scale, enabling faster growth, smarter operations, and a competitive edge. Visit RidgelineApps.com to see what they can unlock for your firm. Every investment firm is unique and generic AI doesn't understand your process. Rogo does. It's an AI platform built specifically for Wall Street, connected to your data, understanding your process and producing real outputs. Check them out at rogo.ai slash invest.