Ok I Admit It. I Underestimated SpaceXAI (by a LOT)

以下是内容的中文翻译： SpaceX的早期投资者加文·贝克（Gavin Baker）继续他的播客之旅，讨论xAI工具以及更广阔的埃隆·马斯克（Elon Musk）帝国。a16z播客中的这些片段重点介绍了SpaceX和xAI的几项重大发展和预测。 **xAI的Grokbot和代理主导地位：** 加文·贝克赞扬了xAI最近发布的Grokbot，指出它能够在短短数秒（7-12秒）内创建复杂的AI代理，例如播客摘要工具、Substack摘要工具和X情绪追踪器，而传统“云代码”则需要数小时。他将Grokbot描述为“简直是奇迹般的出色”、“用户友好”、“直观”且“快速发展”，并预测xAI有望在“代理市场”中占据重要份额。播客主持人甚至认为Grokbot可以取代他的整个工作流程，从内容创作、声音克隆到视频制作、上传和排程。他敦促所有SpaceX投资者安装并使用Grokbot，指出他曾“严重且故意低估了”xAI在模型层的成功，但现在预计将获得巨大的市场份额。他指出，xAI几年前才从零开始，但“凭借Grokbot代理，一举超越了整个行业。”据报道，Grokbot使用率很高，“服务器偶尔会超载。” **轨道数据中心（Starmind）和成本经济学：** 讨论深入探讨了轨道数据中心，据报道这是SpaceX正在专注的业务。目前在地球上启动一吉瓦AI计算的成本约为500亿美元。然而，通过利用太空，这一成本可以减少约150亿美元（消除了地面电力、冷却和劳动力成本），使轨道成本降至每吉瓦约350亿美元。这是一个“非常重要的事情”，只有SpaceX具备这种能力，前提是星舰实现快速可靠的重复使用。播客强调地面成本是通胀的，援引电工现在“年收入远超50万美元”，这是由于AI数据中心的需求所致。相比之下，轨道计算，特别是结合星舰重复使用，预计将是通缩的，发射成本可能降至“10亿美元以下”。埃隆·马斯克（Elon Musk）和黄仁勋（Jensen Huang）“共同设计了一个鲁本机架”（reuben rack）用于轨道计算，计划于2027年第四季度发射（如果延误则为2028年第二季度）。演讲者设想“Starmind”将成为SpaceX第二个“印钞的巨大全球星座”，将AI计算能力租借给OpenAI或Anthropic等公司，这些公司由于缺乏自己的发射能力，将被迫寻求SpaceX的服务。 **SpaceX的强劲估值和多元化机遇：** 播客强调，即使忽略了主要的潜在收入来源，SpaceX的长期估值依然非常强劲。主持人的20年估值模型，即使假设在“代理层”（Grokbot）几乎没有实质性成功，并且AI模型“对公司完全没有实质性影响”，仍然认为SpaceX是“绝对划算的买卖”和“非常便宜的便宜货”。他表示，“AI基础设施”（地球上和地球外的计算能力）目前“支撑着”他对SpaceX的长期估值。其他重大机遇包括星链移动（Starlink Mobile），预计这是一个“8000-9000亿美元营收”的市场，为“近两万亿美元”的宽带和移动市场做出贡献。SpaceX还拥有“快速增长的AI年度经常性收入基础”（AI ARR base），并进行了“非常激进的计算投资”。即使xAI“在自家AI应用上失误”，SpaceX仍然拥有“庞大的基础设施业务”。估值表明，即使删除SpaceX内部“整个数万亿美元的机遇”，“也不会显著影响公司的公允价值。” **星舰和未来的发射基础设施：** 讨论谈到了“路易斯安那星际基地”（Starbase Louisiana）的开发，预计将实现“每年数千次发射”。有预测称，未来将在中东、欧洲、日本和韩国等全球多地建设星际基地。播客区分了中国用于捕获火箭的“临时拼凑系统”（十多年前在SpaceX的Reddit子版块上被提出）与星舰的完全可重复使用性，后者助推器和星舰被捕获、堆叠、加注燃料并再次发射。星舰发射台正在设计中，可能实现“每天两次以上”的发射，这预示着一个雄心勃勃的发射频率。 播客总结道，尽管成功无法保证，但xAI从无到有、迎头赶上，然后凭借Grokbot等代理“超越”整个行业的速度是一个“极其有希望的迹象”，使其未来的主导地位“越来越有可能”。

Gavin Baker, an early SpaceX investor, continues his podcast tour discussing xAI tools and the broader Elon Musk empire. These clips from the a16z podcast highlight several significant developments and predictions for SpaceX and xAI. **xAI's Grokbot and Agent Dominance:** Gavin Baker praises xAI's recently released Grokbot, noting its ability to create sophisticated AI agents like podcast summarizers, Substack summarizers, and X sentiment trackers in mere seconds (7-12 seconds), compared to hours with traditional "cloud code." He describes Grokbot as "fucking amazing," "user-friendly," "intuitive," and "rapidly evolving," predicting that xAI could capture a significant share of the "agent market." The podcast host even suggests Grokbot could replace his entire workflow, from content creation and voice cloning to video production, uploading, and scheduling. He urges all SpaceX investors to install and utilize Grokbot, noting that he had "massively and intentionally underestimated" xAI's success in the model layer, but now expects massive market share gains. He notes that xAI, starting from scratch just a few years ago, has "leapfrogged the whole industry with agents with Grokbot." Grokbot is reportedly experiencing high usage, with "servers overloaded every once in a while." **Orbital Data Centers (Starmind) and Cost Economics:** The discussion delves into orbital data centers, a venture SpaceX is reportedly focused on. The current cost to bring a gigawatt of AI compute online on Earth is approximately $50 billion. However, by leveraging space, this cost can be reduced by about $15 billion (eliminating terrestrial power, cooling, and labor expenses), bringing the orbital cost to around $35 billion per gigawatt. This is a "very big deal," with only SpaceX having the capacity, provided Starship achieves rapid and reliable reusability. The podcast highlights that terrestrial costs are inflationary, citing electricians now making "well in excess of half a million US dollars per year" due to the demand from AI data centers. In contrast, orbital compute, especially with Starship reusability, is expected to be deflationary, with launch costs potentially falling "under a billion." Elon Musk and Jensen Huang have "co-designed a reuben rack" for orbital compute, scheduled to launch in Q4 2027 (or Q2 2028 if delayed). The speaker envisions "Starmind" as SpaceX's second "gigantic global constellation that prints money," leasing AI compute to companies like OpenAI or Anthropic, who, lacking their own launch capabilities, would be forced to seek SpaceX's services. **SpaceX's Robust Valuation and Diversified Opportunities:** The podcast emphasizes that SpaceX's long-term valuation remains incredibly strong, even when ignoring major potential revenue streams. The host's 20-year valuation model, which assumes "almost zero meaningful success in the agents layer" (Grokbot) and "no meaningful impact on the company at all" from AI models, still views SpaceX as an "absolute steal" and "dirt cheap bargain." He states that the "AI infrastructure" (compute on and off Earth) is currently "carrying" his long-term SpaceX valuation. Other significant opportunities include Starlink Mobile, projected as an "$800-900 billion of revenue" market, contributing to a "close to two trillion" market for broadband and mobile. SpaceX also has a "rapidly growing AI ARR base" and has made "very aggressive compute investments." Even if xAI "fumbles the ball with their first party AI applications," SpaceX still possesses a "massive infrastructure business." The valuation suggests that deleting "entire multi-trillion dollar opportunities" within SpaceX "doesn't meaningfully move the fair value of the company." **Starship and Future Launch Infrastructure:** The discussion touches upon the development of "Starbase Louisiana," which is expected to enable "thousands of launches a year." There are predictions for multiple global Starbases in locations like the Middle East, Europe, Japan, and South Korea. The distinction is made between China's "jury rig system" for catching rockets (suggested on the SpaceX subreddit over 10 years ago) and Starship's full reusability, where the booster and Starship are caught, stacked, fueled, and sent back up. Starship pads are being engineered for potentially "more than two a day" launches, suggesting a highly ambitious launch cadence. The podcast concludes that while success is not guaranteed, the speed at which xAI has progressed from non-existence to catching up and then "leapfrogging" the industry with agents like Grokbot is an "extremely promising sign," making future dominance "increasingly more likely."