What Anthropic’s $2 Trillion IPO Means For SpaceX

这份记录讨论了隐藏在SpaceX内部的潜在数万亿美元业务，主要集中在其人工智能计算能力以及其自身的人工智能模型上。 以下是提及的每一个新闻要点的总结： * **Anthropik的首次公开募股（IPO）和估值：** * 据《金融时报》报道，Anthropik的目标首次公开募股（IPO）估值高达**2万亿美元**。 * 这将**打破SpaceX隐含的、创纪录的1.75万亿美元IPO估值**（作为目前最高的私有公司估值）。 * 《金融时报》周三晚（录音前）报道称，投资者预计Anthropik将在**10月**上市，距离录音日期还有6-8周。 * Polymarket数据显示，Anthropik今年进行IPO的可能性为**80%**。 * Brad Gerstner，一位“第五位好友”，也暗示IPO会很快发生。 * Anthropik的年化经常性收入（ARR）预计到年底将达到**1000亿至1200亿美元**。 * 2万亿美元的估值将是其销售额的**16到20倍**，并指出这仅是SpaceX和Palantir首次公开募股（IPO）市盈率的一小部分。 * Anthropik的收入在过去一年中增长了**10倍**。 * Polymarket显示（截至录音前几天），Anthropik到年底拥有顶级AI模型的可能性为**67%**。 * 新细节（通过Kate Rooney，Tech Check）表明，Anthropik的年化收入到**7月底已超过650亿美元**。 * 这比去年同期增长了**7倍**。 * 这些信息是在最近一个周末的更新中与现有投资者分享的。 * Anthropik在5月的年化经常性收入（ARR）约为**470亿美元**。 * 去年全年，Anthropik的收入不到**100亿美元**。 * 他们实现更高收入增长的最大瓶颈是**获得更多计算资源**。 * **SpaceX与Anthropik的人工智能计算交易：** * Anthropik于**2026年5月6日**与SpaceX签署了一项AI计算协议（注：此日期看似未来日期，很可能是原始来源中的笔误，但报告中如此说明）。 * Anthropik同意每月支付**12.5亿美元**，持续到2029年中。 * 任何一方均可提前**90天通知**取消此交易。 * Anthropik获得了**SpaceX Colossus 1数据中心全部容量的独家使用权**，该中心容量约为300兆瓦。 * SpaceX AI目前拥有约**1.4吉瓦**的在线计算能力，大约是提供给Anthropik容量的五倍。 * 该计算能力在**当月（5月）上线**，主要用于**推理（inference）**，而非训练（training）。 * 交易的**立竿见影的实际效果**（与宣布日同日生效）： * Anthropik将Claude Code的Pro、Max、Team和基于席位的企业计划的**五小时速率限制翻倍**。 * 他们**取消了高峰时段限制**，并**大幅提高了API费率**。 * 这使得Anthropik能够将现有需求转化为付费使用，解除了容量限制。 * Anthropik的年化经常性收入（ARR）从4月的**300亿美元**增长到2026年5月的**470亿美元**（同样是未来日期），到2026年7月底达到**650亿美元**。 * 这意味着在大约两个月内，年化经常性收入（ARR）增加了约**180亿美元**。 * SpaceX的计算能力充当了一个关键的短期桥梁，缓解了瓶颈并促进了持续的快速增长。 * **SpaceX的人工智能计算业务（“隐藏”的2万亿美元机遇）：** * SpaceX AI拥有独特的能力，能够迅速上线巨量AI计算资源，业内人士估计其速度比其他任何公司**快6-8倍**。 * 他们可以以**低得多的总成本**做到这一点。 * 在最近的第二季度财报电话会议中，SpaceX披露其AI计算资本支出（capex）的回收期**不到一年**，这被认为是荒谬的（相比之下，典型的回收期为5年、10年甚至20-25年）。 * 每家AI公司都将对AI计算资源（用于推理和训练）拥有近乎无限的需求。 * SpaceX的AI计算业务本身在未来几年内每年可能增长到**数千亿美元**。 * SpaceX正计划在**未来12-24个月内**上线“巨大”的计算能力。 * 他们优先考虑**冷却和热管理**，而不是GPU采购。 * 其他主要公司（如Anthropik、Google）向SpaceX支付费用，表明他们自身无法复制这种速度和成本效益。 * **星智（轨道数据中心）：** SpaceX认为他们将在**未来12-24个月内通过星舰（Starship）**发射这些数据中心。 * 星舰（Starship）的进展“非常顺利”。 * 轨道数据中心因其发射能力而对SpaceX来说是独一无二的。 * 地球上的数据中心扩张变得越来越困难和昂贵（“我的后院不要数据中心”的心理战）。 * SpaceX拥有显著的议价能力，在计算协议中要求**90天通知取消条款**，以便为其自身模型重新获得容量或利用飙升的现货价格。 * 演讲者表示：“谁控制了AI计算，谁就控制了宇宙。” * **SpaceX的人工智能模型业务（Grok——真正隐藏的2万亿美元秘密）：** * 潜在的**2万亿美元机遇**是AI模型本身，而不仅仅是计算业务。 * SpaceX最近完成了对**Cursor**的收购。 * SpaceX AI Grok层面的创新速度“惊人”。 * 提到了Grok 4.5、4.6和4.7“即将推出”。 * **Grok机器人“刚刚上线”。** * 官方消息称，**Grok 4.6“已经赶上了行业内其他产品，几乎达到了智能前沿，但成本却低得多。”** * Grok被描述为**“单位智能成本”第一**。 * 演讲者个人20年的SpaceX估值模型*目前*显示，SpaceX AI模型（包括Cursor）占SpaceX总利润的比例从未超过约**3.5%**，通常更接近2%。 * 尽管他相信其真实潜力巨大，但他**有意地在其模型中“低估和轻视”了这一数字**。 * 低估的原因包括不希望假定SpaceX所有业务都“完全主导”市场，以及AI模型领域“现在判断赢家还为时过早”的论点。 * 然而，他断言，在未来几十年里，**AI这个“蛋糕”将增长许多个数量级**。 * 如果Grok能保持适度的竞争力，仅此一项就可能价值**超过SpaceX目前总市值（低于2万亿美元）**。 * 他可以从其估值模型中删除整个AI模型部分，而这“几乎不会影响”他对公司公允估值的估计。 * **总体结论/协同效应：** * SpaceX可能“公司内部就拥有一个完整的Anthropik”。 * 即使SpaceX AI在AI模型层面没有取得实质性成功，它仍将通过提供地面和轨道AI计算资源，间接受益于其他AI公司的成功。 * 演讲者总结道：**“他们不可能输。”**

This transcription discusses the potential multi-trillion dollar businesses hidden within SpaceX, focusing primarily on its AI compute capabilities and its own AI models. Here's a summary of every single news point mentioned: * **Anthropik's IPO and Valuation:** * Anthropik is reportedly targeting a **$2 trillion IPO valuation**, according to the Financial Times. * This would **break SpaceX's implied record-setting $1.75 trillion IPO** (as the current highest private valuation). * The FT reported on Wednesday night (pre-recording) that investors expect Anthropik to go public in **October**, which is 6-8 weeks from the recording date. * Polymarket data shows an **80% chance of Anthropik IPO-ing this year**. * Brad Gerstner, a "fifth bestie," also signaled an IPO sooner rather than later. * Anthropik's annualized run rate (ARR) is projected to end the year between **$100 billion and $120 billion**. * A $2 trillion valuation would be **16 to 20 times sales**, which is noted as a fraction of SpaceX and Palantir's IPO multiples. * Anthropik's revenue growth has been **10x over the last year**. * Polymarket indicates a **67% chance that Anthropik will have the top AI model by the end of the year** (as of a couple of days prior to recording). * New details (via Kate Rooney, Tech Check) suggest Anthropik's annualized revenue **topped $65 billion by the end of July**. * This is a **7x year-over-year increase**. * This intel was shared with existing investors in an update over a recent weekend. * Anthropik's ARR in May was around **$47 billion**. * For all of last year, Anthropik brought in **less than $10 billion**. * Their biggest bottleneck to more revenue growth is **access to more compute**. * **SpaceX's AI Compute Deal with Anthropik:** * Anthropik signed an AI compute deal with SpaceX on **May 6th, 2026** (note: this date appears to be futuristic, likely a typo in the original source, but reported as stated). * Anthropik agreed to pay **$1.25 billion per month**, ongoing until mid-2029. * The deal is **cancelable by either party on 90 days notice**. * Anthropik gained **exclusive access to the full capacity of SpaceX's Colossus 1 data center**, which is around 300 megawatts. * SpaceX AI currently has roughly **1.4 gigawatts of compute online**, which is about five times the capacity provided to Anthropik. * The compute capacity came online **within the month** (May) and was primarily for **inference**, not training. * **Immediate concrete effects** of the deal (effective same day as announcement): * Anthropik **doubled Claude Code's five-hour rate limits** for Pro, Max, Team, and Seat-based enterprise plans. * They **removed peak hours limits** and **substantially raised API rates**. * This allowed Anthropik to convert existing demand into paid usage, unlocking capacity constraints. * Anthropik's ARR grew from **$30 billion in April** to **$47 billion in May 2026** (same futuristic date) to **$65 billion by the end of July 2026**. * This represents an increase of roughly **$18 billion in ARR in approximately two months**. * The SpaceX compute acted as a critical near-term bridge, relieving bottlenecks and enabling continued rapid growth. * **SpaceX's AI Compute Business (The "Hidden" $2 Trillion Opportunity):** * SpaceX AI has a unique ability to rapidly bring online enormous amounts of AI compute, estimated by industry insiders to be **6-8 times faster than anyone else**. * They can do so at a **much lower total cost**. * In their recent Q2 earnings call, SpaceX disclosed that the payback period for their AI compute capex is **less than a year**, which is described as absurd (compared to typical 5, 10, or 20-25 year payback periods). * Every AI company will have near-infinite demand for AI compute (for inference and training). * The SpaceX AI compute business itself could grow to **many hundreds of billions of dollars per year in the next few years**. * SpaceX is looking at bringing "enormous" amounts of capacity online in the **next 12-24 months**. * They are prioritizing **cooling and thermal management** ahead of GPU acquisition. * Other major companies (Anthropik, Google) paying SpaceX suggests they cannot replicate this speed and cost efficiency themselves. * **StarMind (Orbital Data Centers):** SpaceX believes they'll be launching these in the **next 12-24 months on Starship**. * Starship progress is "going very well." * Orbital data centers are unique to SpaceX due to their launch capabilities. * Earth-based data center expansion is becoming harder and more expensive ("no data centers in my backyard" PSYOP). * SpaceX has significant leverage, requesting a **90-day notice cancellation clause** in compute deals to regain capacity for its own models or capitalize on surging spot prices. * The speaker states, "He who controls the AI compute controls the universe." * **SpaceX's AI Models Business (Grok - The *Real* Hidden $2 Trillion Secret):** * The potential **$2 trillion opportunity** is the AI models themselves, not just the compute business. * SpaceX recently closed on the acquisition of **Cursor**. * The pace of innovation within the Grok layer of SpaceX AI is "phenomenal." * Mentions Grok 4.5, 4.6, and 4.7 "coming soon." * **Grok bots have "just gone live."** * It is official that **Grok 4.6 has "caught up to the rest of the industry almost at the frontier of intelligence, but for a fraction of the cost."** * Grok is described as **number one on cost per intelligence**. * The speaker's personal 20-year SpaceX valuation model *currently* shows SpaceX AI models (including Cursor) representing **never more than around 3.5% of total SpaceX profits**, typically closer to 2%. * The speaker **intentionally downplays and underestimates** this figure in his model, despite believing its true potential is vast. * Reasons for downplaying include not wanting to assume "complete and utter domination" across all SpaceX ventures and the "still too early to call winners" argument in AI models. * However, he asserts that the **AI "pie" will grow many orders of magnitude** in the coming years and decades. * If Grok can maintain moderate competitiveness, this element alone could be worth **more than SpaceX's current entire market capitalization (less than $2 trillion)**. * He can delete the entire AI models section from his valuation model, and it "barely moves the needle" on his estimated fair valuation for the company. * **Overall Conclusion/Synergy:** * SpaceX could have an "entire Anthropik inside the company." * Even if SpaceX AI has no meaningful success in the AI models layer, it will still benefit indirectly from the success of other AI companies by providing Earth-based and orbital AI compute. * The speaker concludes, **"They cannot lose."**