More Trillion Dollar IPOs, Anthropic $3T, Zuck's Price War, China Ends Open Source?, Trump Accounts

以下是播客内容的中文摘要，包括每位嘉宾的重点： ### **Jason Calacanis (主持人)** * **开场白：** 欢迎各位嘉宾，提到Freeberg正在度假，Brad在东北部，Jason在巴黎参加Raise会议进行采访，Chamath则在工作。 * **AI/IPO 更新：** 介绍了“万亿美元IPO热潮”，提到了SpaceX的业绩表现以及Anthropic和OpenAI的秘密提交上市申请。他询问Chamath，OpenAI专注于消费者市场是否是一个错误，而Anthropic在企业级市场则表现强劲。 * **Token成本与投资回报率：** 强调他个人使用开源模型和动态路由将token成本降低了95%的经验，这从根本上改变了行为（例如，每小时运行代理）。提到Uber和DoorDash的CTO公开讨论了如何管理token支出并优化投资回报率。 * **Meta与垂直化模型：** 提到马克·扎克伯格宣布推出Muse Spark 1.1，表明Meta正进军提供低成本代理模型。观察到领先模型（如Lovable和11 Labs）的主要客户正在开发自己的垂直化AI模型以减少依赖的趋势。 * **中国/AI主权：** 提到了路透社关于中国可能限制其顶级AI模型海外访问的报道，并向Sacks询问其影响。 * **特朗普账户：** 介绍了Brad的“特朗普账户”倡议，强调其在应用商店的成功和广泛影响。他表达了对该倡议的强烈支持，称其为“你能做的最美国化的事情”和“一份美丽的礼物”，可以团结国家并为实现美国梦提供一条道路，将其与社会主义意识形态进行对比。他赞扬Joe Gebbia在开发该软件方面的作用。 ### **Chamath Palihapitiya** * **AI/IPO 更新：** * 认为Anthropic和OpenAI都是出色的公司，但市场需求和价格将是其IPO的关键。 * 指出OpenAI的现金消耗率远高于Anthropic，而Anthropic可能“意外地实现了盈利”。 * 透露他自己公司的token成本“每45天翻一番”，而生产力提升“最多只有5%”，暗示正在达到“渐近线”。建议公司在投资回报率讨论变得更加激烈之前“现在就退出”。 * 认为企业收入比消费者收入“更脆弱”，因为企业客户会要求明确的投资回报率。 * 从联合国AI委员会的观察中发现，没有哪个国家愿意屈服于一个闭源的美国模型，这推动了对主权AI的需求。 * 建议如果开源模型“有99%的好”（就像“可乐/百事可乐”的比较），成本将推动用户放弃领先模型。 * 预测一系列与AI支出相关的“盈利未达预期”将迫使公司削减成本，使他们考虑更便宜的替代方案。 * **中国/AI主权：** 警告能源供应是AI的主要瓶颈，指出美国需要“相当于三个加州的能源”，以及台湾容易受到液化天然气封锁的影响。 ### **Brad Gerstner** * **AI/IPO 更新：** * 称SpaceX的IPO是“教科书式的”，以1.75万亿美元的估值筹集了750亿美元，为Anthropic和OpenAI树立了先例。 * 认为Anthropic（传闻收入1000亿美元）和OpenAI（传闻年运行收入700亿美元）在未来6-9个月内上市的几率“非常高”。Altimeter公司将是这两家公司的买家。 * 强调这些公司是“复合型企业”，多年来收入增长超过30%，而非“一夜暴富的计划”。 * 强调AI拥有前所未有的总潜在市场（TAM）——“地球上每一个小型、中型、大型公司”。 * 承认虽然最初的AI支出可能是实验性的，但在生命科学和产品创新（例如，黄仁勋用AI设计芯片）等领域的突破使AI不可或缺。 * 尽管token价格降低了90%，他认为“经济价值份额”正在向领先实验室增加，而非开源。 * 认为对于“高端工作负载”（例如，替代每小时200美元的工程师），便宜和昂贵模型（例如，3美元与15美元）之间的成本差异“无关紧要”，如果更昂贵的模型是“万无一失”的。 * 质疑领先模型和开源模型之间的“智能差距”会趋同的假设；他认为由于递归式自我改进，这种差距可能会“扩大”。 * **中国/AI主权：** 表示华盛顿在AI领域领先中国的问题上立场一致，并将采取措施打击外国实体对美国模型的盗用。认为中国的举动是“下棋”，对其自身的伤害大于对美国的伤害。 * **特朗普账户：** * 将“特朗普账户”描述为一项“四年任务”，最终在7月4日发布了该应用程序，成为下载量第一的应用程序。 * 宣布与纽约证券交易所/纳斯达克在椭圆形办公室共同敲钟，总统呼吁为所有5000万至7000万美国儿童“自动创建”账户。 * 报告称，在最初的24小时内，创建了超过150万个账户，存款超过10亿美元。 * 宣布这将是历史上“最大的直接慈善平台”，目标是在12个月内达到1000亿美元，避免了“慈善中间人”。 * 针对“特朗普精神失调症（TDS）”的批评，他表示各种政治倾向的父母都在使用它。强调总统致力于确保“没有孩子掉队”。 * 将“依赖华盛顿”模式与账户的“独立于政府”的财富积累方法进行对比，称其为“抵制更多社会主义的解药”。 * 预测这可能将美国股权所有权从50%提高到70-75%。 * 解释了访问规则：18岁时，25%的资金可以用于购房/创业/上大学，其余则转入个人退休账户（IRA）（提前取款有罚款）。 * 强调了教育意义，让孩子们对金钱感到兴奋。指出25个州正在向账户注入资金。 * 认为这是“国家的变革者”，一个与社会保障同样重要的“新平台”，实现了“普遍所有权”。 * 建议扩大该计划，为18岁以上的成年人设立补充账户，不属于社会保障范畴。 * 强调了从出生开始早期复利的强大力量（“长坡上的小雪球”）。 ### **David Sacks** * **AI/IPO 更新：** * 承认企业*确实希望*摆脱闭源模型，原因是成本和“AI主权”，但往往缺乏像Coinbase或DoorDash那样构建复杂路由层的技术能力。 * 指出开源模型在*企业支出*中的份额（而非使用量）今年实际上从19%下降到11%，这表明领先实验室正在占据更多的市场份额。 * 解释了开源模型适用于*成熟、明确*的用例，可以在后期进行训练，但*不成熟、处于发现阶段*的用例仍然需要最强大的通用智能（领先模型）。 * 强调了“模型互换性”的技术挑战——抽象化内存、上下文和历史记录，以便轻松地在不同模型之间切换。 * **中国/AI主权：** * 重申了他之前的观点，即禁止开源模型会损害美国。建议中国限制对其模型的访问以损害美国将是合乎逻辑的。 * 观察到中国模型（如阿里巴巴、智谱）的模式是先开放以“追赶”，然后闭源以获取价值，这与山姆·奥特曼的策略如出一辙。 * 提出美国赢得AI竞赛的“最好情况”是中国出现自己的“末日论社区”，从而阻碍其自身进步。 * 担心虽然美国最高层领导人希望赢得AI竞赛，但下层官僚机构或国会的影响可能会导致“笨拙的”决策，适得其反。 * **特朗普账户：** * 称赞“特朗普账户”是一个“令人惊叹的慈善平台”和一个“巨大的新慈善平台”，是民粹主义的“解药”。 * 强调其作为“中产阶级家庭规划”工具的价值，称其为“有史以来最伟大的事情之一”，因为它具有税收优惠（每年最高5000美元的供款，免税复利，雇主供款，转入罗斯个人退休账户）。 * 计算出在18岁时存入20万至30万美元，到60岁时可能增值到1000万美元以上。 * 指出它填补了一个“市场空白”，提供了一种从出生就开始的类似个人退休账户的储蓄工具，而通常这种工具在获得收入之前（通常在22岁之后）是无法使用的。

Here's a summary of the podcast, including each guest's main points: ### **Jason Calacanis (Host)** * **Introduction:** Welcomes guests, mentions Freeberg is on vacation, Brad is in the Northeast, Jason is in Paris for interviews at the Raise conference, and Chamath is working. * **AI/IPO Update:** Introduces the "trillion-dollar IPO rush," noting SpaceX's performance and the confidential filings of Anthropic and OpenAI. Asks Chamath about OpenAI's consumer focus potentially being a mistake compared to Anthropic's enterprise strength. * **Token Costs & ROI:** Highlights his own experience with 95% token cost reduction using open-source models and dynamic routing, which fundamentally changes behavior (e.g., hourly agent runs). Mentions Uber and DoorDash CTOs' public discussions about managing token spend and optimizing for ROI. * **Meta & Verticalized Models:** Notes Mark Zuckerberg's announcement of Muse Spark 1.1, indicating Meta's move into providing low-cost agentic models. Observes a trend of major customers (like Lovable and 11 Labs) of frontier models developing their own verticalized AI models to reduce reliance. * **China/AI Sovereignty:** Raises the Reuters report about China potentially restricting overseas access to their top AI models, asking Sax about the implications. * **Trump Accounts:** Introduces Brad's initiative, "Trump Accounts," highlighting its app store success and the broad impact. Expresses strong support for the initiative, calling it "the most American thing you can do" and a "beautiful gift" that could unify the country and provide a path to the American dream, contrasting it with socialist ideologies. Praises Joe Gebbia for his role in developing the software. ### **Chamath Palihapitiya** * **AI/IPO Update:** * Believes Anthropic and OpenAI are great businesses, but market appetite and price will be key for their IPOs. * Notes Open AI's high cash burn compared to Anthropic, which might be "accidentally profitable." * Reveals his own company's token costs are "doubling every 45 days" while productivity gains are only "5% max," suggesting an "asymptote" is being hit. Advises companies to "get out now" before the ROI discussion becomes more intense. * Argues enterprise revenue is "more brittle" than consumer, as enterprise clients will demand clear ROI. * From a UN commission on AI, observes that no country wants to be subjugated to a closed-source American model, driving demand for sovereign AI. * Suggests that if open-source models are "99% as good" (like a "Coke/Pepsi" comparison), cost will drive adoption away from frontier models. * Predicts that a series of "earnings misses" related to AI spend will force companies to cut costs, making them consider cheaper alternatives. * **China/AI Sovereignty:** Warns that energy supply is a major bottleneck for AI, citing the US needing "three entire California's worth of energy" and Taiwan's vulnerability to LNG blockades. ### **Brad Gerstner** * **AI/IPO Update:** * Calls SpaceX's IPO "textbook," achieving 75 billion raised at a 1.75 trillion valuation, and setting a precedent for Anthropic and OpenAI. * Believes Anthropic (rumored 100B revenue) and OpenAI (rumored 70B run rate) have a "very high" chance of going public in the next 6-9 months. Altimeter would be a buyer of both. * Emphasizes that these companies are "compounders" with 30%+ revenue growth for years, not "get rich quick schemes." * Highlights AI's unprecedented Total Addressable Market (TAM) – "every single small, medium, large company on the planet." * Acknowledges that while initial AI spending can be experimental, breakthroughs in fields like life sciences and product innovation (e.g., Jensen Huang designing chips with AI) make AI indispensable. * Despite 90% token price reductions, believes the "share of economic value" is increasing for frontier labs, not open source. * Argues that for "premium workloads" (e.g., replacing a $200/hour engineer), the cost difference between cheap and expensive models (e.g., $3 vs. $15) is "irrelevant" if the more expensive one is "bulletproof." * Challenges the assumption that the "intelligence gap" between frontier and open-source models will converge; suggests it might "extend" due to recursive self-improvement. * **China/AI Sovereignty:** States that Washington is united in wanting to stay ahead of China in AI and will take steps against distillation of US models by foreign entities. Believes China's moves are "chess playing" that hurt them more than the US. * **Trump Accounts:** * Describes the "Trump Accounts" as a "four-year mission" culminating in the app's launch on July 4th, becoming the #1 downloaded app. * Announces a joint bell-ringing at NYSE/NASDAQ from the Oval Office, where the President called for "auto-creating" accounts for all 50-70 million US kids. * Reports over 1.5 million accounts created and over $1 billion in deposits in the first 24 hours. * Declares it will be the "largest direct philanthropic platform" in history, aiming for $100 billion in 12 months, avoiding the "charitable middleman." * Addresses the "TDS" criticism by stating parents across the political spectrum are using it. Stresses the President's commitment to ensuring "no child is left behind." * Contrasts the "dependent on Washington" model with the "independent of government" wealth-building approach of the accounts, calling it an "antidote to more socialism." * Predicts it could increase US equity ownership from 50% to 70-75%. * Explains the access rules: at 18, 25% can be withdrawn for home/business/college, the rest rolls into an IRA (with penalties for early withdrawal). * Emphasizes the educational aspect, getting kids excited about money. Notes 25 states are adding funds to the accounts. * Considers it a "game changer for the country," a "new platform" as significant as Social Security, enabling "universal ownership." * Suggests expanding the program to include supplemental accounts for adults over 18, outside of Social Security. * Highlights the power of early compounding from birth ("small snowball on a long hill"). ### **David Sacks** * **AI/IPO Update:** * Acknowledges enterprises *want* to diversify away from closed models due to cost and "AI sovereignty" but often lack the technical capability to build sophisticated routing layers like Coinbase or DoorDash. * Points out that open-source's share of *enterprise spending* (not usage) actually decreased from 19% to 11% this year, suggesting frontier labs are capturing more wallet share. * Explains that open models are great for *mature, defined* use cases where post-training is possible, but *immature, discovery-phase* use cases still require the most powerful general intelligence (frontier models). * Highlights the technical challenge of "model fungibility" – abstracting away memory, context, and history to easily swap between models. * **China/AI Sovereignty:** * Reiterates his previous point that banning open models would harm the US. Suggests China restricting access to its models would be logical to harm the US. * Observes a pattern of Chinese models (like Alibaba, Zipu) initially going open to "catch up" and then going closed to capture value, mirroring Sam Altman's strategy. * Proposes the "best thing" for the US to win the AI race is if China develops its own "doomer community" to hinder its own progress. * Worries that while top US leadership wants to win the AI race, lower-level bureaucracy or congressional influence could lead to "ham-fisted" decisions that are counterproductive. * **Trump Accounts:** * Hails the "Trump Accounts" as an "amazing philanthropic platform" and a "tremendous new philanthropic platform" that's an "antidote" to populism. * Emphasizes its value as a "middle-class family planning" tool, calling it one of the "greatest things ever" for its tax advantages (up to $5k/year contributions, tax-free compounding, employer contributions, Roth IRA rollovers). * Calculates that $200-300k at age 18 could become $10M+ by age 60. * Notes that it fills a "market gap" by providing an IRA-like savings vehicle from birth, which typically isn't available until one has earned income (usually after age 22).