What Happens When the AI Boom Runs Out of Money

以下是这段内容的中文翻译： 这位演讲者首先提出，如果美国在AI竞赛中取得“决定性胜利”，尤其是在AI赋予其军事优势这种虚构场景下，那将是“极具问题”的。他认为，面对美国的这种主导地位，中国基于博弈论的最优回应将是摧毁台积电，这凸显了全球对台湾半导体制造的依赖。这种脆弱性也揭示了对中国制造更广泛但被低估的依赖，这种依赖由于成本劣势，在非冲突情况下难以轻易解决。 他描述了当前的AI平衡对美国有利，OpenAI和Anthropic处于领先地位，谷歌、Meta和Grok紧随其后，而中国公司则落后6-9个月。然而，他质疑这种差距的可持续性，尤其是随着AI的自我提升，这可能导致“加速”或“爆发”，从而使追赶变得极其困难。 一个关键的财务问题是巨额投资与收入产生之间的“时间错配”。他将其与铁路时代进行比较，指出AI所需的资本支出是天文数字，目前远远超过了回报。他强调了谷歌最近的股权发行，暗示它正朝着伯克希尔·哈撒韦的模式发展，即通过高利润的传统业务（如谷歌搜索，被比作“时思糖果”）来资助那些利润率较低但绝对利润潜力巨大的新兴业务（如AI，被比作“BNSF铁路”）。 谈到AI的能力，演讲者对其经济影响“超级看好”，但对其在“可验证领域”（如编码和数学）之外的“泛化能力”则“不那么看好”。他怀疑AI在没有更多数据的情况下，能否将其能力转化为“不可验证领域”（如人类思想或情感），这可能暗示了诸如Neuralink之类的技术。尽管如此，他认为仅在可验证领域的经济机会就“巨大”，即使AI不再进一步改进，他称自己为“不情愿的加速主义者”。 讨论随后转向商业模式，对比了聚合理论中“零边际成本”的分发与AI实际存在的推理成本。他批评微软的企业AI定价可能“隐患重重”，因为它从可预测的按人头授权转变为基于使用量的成本，这可能挑战既定的预算流程，并促使客户更严格地审视产品价值。他还感叹硅谷普遍不愿将广告作为消费者商业模式，提到了OpenAI错失的机会，并赞扬Meta的广告平台是一种“社会积极面”。 在计算供应链方面，演讲者观察到内存和晶圆制造商（如台积电）历来都很保守，导致了当前的短缺。他认为，台积电不愿过度投资产能，将风险转移给了主要的科技公司，这些公司现在由于计算能力不足而面临“错失的收入”。这种稀缺性最终“挽救了英特尔”和三星的逻辑代工雄心，因为科技巨头现在即便痛苦也乐于分散芯片制造。 分析具体公司时，亚马逊凭借其“先为自己构建，再出售给他人”的策略（AWS、物流、Graviton/Trainium等定制芯片）脱颖而出，使其核心业务对AI具有韧性。苹果虽然似乎置身于AI竞争之外，但受益于其强大的生态系统和客户访问权限，可能利用设备端AI来避免推理成本。然而，如果在一个环境AI世界中它过于以手机为中心，它可能会陷入“微软陷阱”。 在这些前沿AI公司中，OpenAI和Anthropic受到“信仰”和“信念”的驱动，而Meta则受益于马克·扎克伯格的创始人活力，利用其广告主导地位进行AI内容生成和广告匹配。另一方面，微软则采用了90年代“IBM的策略”，旨在成为帮助企业集成AI的中间件提供商，优先考虑稳定性与集成而非前沿创新。 最后，讨论了英伟达“不自然的”利润率。演讲者认为，英伟达对AI“新云”的资金支持实际上起到了“降价”的作用，因为它承担了未来计算需求的风险。他认为超大规模云服务商（谷歌、亚马逊）是英伟达最大的长期威胁，因为它们拥有更低的资本成本，并能够开发和销售自己的通用芯片。他总结道，AI热潮中最关键的长期回报是“能源富裕”的潜力，这将是人类的颠覆性益处。

The speaker opens by positing that a definitive U.S. "win" in the AI race, particularly in a fantastical scenario where AI grants military superiority, would be "very problematic." He suggests that China's game theory optimal response to such U.S. dominance would be to destroy TSMC, highlighting the global dependency on Taiwan's semiconductor manufacturing. This vulnerability underscores a broader, underappreciated reliance on China for manufacturing that isn't easily fixed outside of a conflict due to cost disadvantages. He describes the current AI equilibrium as favorable to the U.S., with OpenAI and Anthropic leading, Google, Meta, and Grok chasing, and Chinese companies staying 6-9 months behind. However, he questions the sustainability of this gap, especially as AI improves itself, potentially leading to an "acceleration" or "takeoff" that could make catching up extremely difficult. A crucial financial concern is the "timing mismatch" between massive investment and revenue generation. Comparing it to the railroad era, he notes the astronomical capital expenditure required for AI, which currently far outstrips returns. He highlights Google's recent equity issuance, suggesting it's moving towards a Berkshire Hathaway model where high-margin legacy businesses (like Google Search, likened to "Seas Candies") fund ventures with lower margins but vastly larger absolute profit potential (like AI, likened to "BNSF railways"). Regarding AI's capabilities, the speaker is "super bullish" on its economic impact but "less bullish" on its generalizability beyond "verifiable domains" like coding and math. He expresses skepticism about AI's ability to translate this into "unverifiable domains" (like human thought or emotion) without more data, possibly hinting at technologies like Neuralink. Nonetheless, he believes the economic opportunity in verifiable domains alone is "massive," even if AI doesn't improve further, labeling himself a "reluctant accelerationist." The discussion then shifts to business models, contrasting aggregation theory's "zero marginal cost" distribution with AI's very real inference costs. He criticizes Microsoft's enterprise AI pricing as potentially "fraught" because it shifts from predictable per-headcount licensing to usage-based costs, which can challenge established budgetary processes and encourage customers to scrutinize product value. He also laments Silicon Valley's general reluctance to embrace advertising as a consumer business model, citing OpenAI's missed opportunity and praising Meta's ad platform as a "societal positive." On the compute supply chain, the speaker observes that memory and fab manufacturers (like TSMC) have historically been conservative, leading to current shortages. He argues that TSMC's reluctance to over-invest in capacity has transferred risk to major tech companies, who now face "foregone revenue" due to insufficient compute. This scarcity is what ultimately "saved Intel" and Samsung's logic foundry ambitions, as tech giants are now incentivized to diversify chip manufacturing despite the pain. Analyzing specific companies, Amazon stands out for its "build for them, then sell to others" strategy (AWS, logistics, custom chips like Graviton/Trainium), making its core business resilient to AI. Apple, while seemingly sitting out the AI race, benefits from its strong ecosystem and customer access, potentially leveraging on-device AI to avoid inference costs. However, it risks a "Microsoft trap" if it remains too phone-centric in an ambient AI world. Among the frontier AI companies, OpenAI and Anthropic are driven by "religion" and "belief," while Meta benefits from Mark Zuckerberg's founder energy, leveraging its advertising dominance for AI content generation and ad matching. Microsoft, on the other hand, adopts an "IBM playbook" of the 90s, aiming to be the middleware provider that helps enterprises integrate AI, prioritizing stability and integration over frontier innovation. Finally, NVIDIA's "unnatural" profit margins are discussed. The speaker argues that NVIDIA's financial backing of AI "neoclouds" effectively acts as a "price cut," as NVIDIA assumes risk for future compute demand. He sees hyperscalers (Google, Amazon) as NVIDIA's biggest long-term threat due to their lower cost of capital and ability to develop and sell their own commodity chips. The most crucial long-term payoff from the AI boom, he concludes, is the potential for "power abundance," which would be a transformative benefit for humanity.