The AI Selloff Doesn't Match the Data | Top AI Investor Explains

以下是这段内容的中文翻译： 该发言人将七月份的AI股票市场描述为“一个月内的2022年”，许多股票下跌了40-60%，尽管他认为其底层基本面持续改善。他指出自己一直在对负面量化指标进行“压力测试”，但并未发现任何问题；相反，从GPU供应和租用价格，到DRAM现货价格和token增长，所有指标都在加速。 他认为市场负面反应的原因在于对Anthropic、OpenAI等私营实体以及Fireworks、Base 10、Modal等开源推理云缺乏了解，而这些实体都呈现出大规模的加速增长。市场错误地将半导体现金流的增长与超大规模云服务商自由现金流的持平解读为负面信号，却忽略了旧款GPU价格“垂直飙升”的事实，这暗示现有合同被严重低估。超大规模云服务商（微软、Meta、亚马逊）的运营现金流已经在加速增长，即便像Rubin这样的新芯片尚未完全部署，合同也未重新定价。 在他看来，有几个市场事件被误读了： 1. **Meta对外租借算力**：这被视为利空（产能过剩），但这很可能是一种战略举动，效仿SpaceX以高溢价销售算力的成功模式，可能预示着未来资本支出的增加而非削减。 2. **开源模型引发的“恐慌”**：由于向更便宜的开源token转变而导致“硅数据token指数”下跌，这被视为负面。然而，就所需算力而言，“一个token就是一个token”。开源模型的兴起，即使会将利润从前沿模型转移，最终也会增加整体算力需求，从而利好英伟达等基础设施提供商。 3. **中国的DUV设备**：导致半导体设备股遭到抛售，但该技术仍比尖端水平落后几十年。 该发言人主要的量化担忧是实际收益率上升和信用利差扩大，如果AI的建设严重依赖债务，这将是一个问题。然而，他反驳说，目前超大规模云服务商对新芯片（Blackwell、Rubin）的变现能力估算远低于其潜力（例如低于两代前的Ampere芯片）。如果这些芯片即使以低于当前Blackwell芯片的折扣价变现，运营现金流也可能达到2万亿美元，抵消数千亿美元的信贷需求。随着合同价格向上重新定价，已部署的算力基础也将改善信用比率。 他强调，英伟达目前的远期市盈率处于10年来的最低水平（排除了解放日和DeepSeek等V型底部），这表明市场认为他们“过度盈利”严重。然而，GPU租用价格却在飙升（例如，B200在七个月内从每小时约2美元涨到约4美元），这表明超大规模云服务商实际上在现有合同上“盈利不足”。 展望未来，他认为“持续学习”和“样本高效学习”是潜在的创新，可能会暂时减少训练需求，但训练只占整体算力的一小部分。他认为最大的风险仍然是监管反弹，这源于关于数据中心对电力、水资源和就业影响的误导信息。他倡导AI行业更好地宣传其益处，例如救生应用和高薪蓝领工作。 最后，他指出新兴的“黑马”参与者，并强调市场可能低估了像SpaceX这样的公司，SpaceX展示了快速扩展算力基础设施的卓越能力。尽管最近市场动荡，他仍保持谨慎乐观，并指出硅谷的大多数人甚至比他还要乐观。

The speaker describes July as a "2022 in a month" for AI stocks, with many names down 40-60% despite what he sees as continuously improving underlying fundamentals. He notes that he's been "pressure testing" for any negative quantitative metrics but has found none; instead, every metric from GPU availability and rental pricing to DRAM spot prices and token growth is accelerating. He attributes the market's negative reaction to a lack of visibility into private entities like Anthropic, OpenAI, and open-source inference clouds (Fireworks, Base 10, Modal), which are all showing massive acceleration. The market mistakenly interprets rising semiconductor cash flow alongside flat hyperscale free cash flow as a negative, missing the fact that older GPU prices are going "vertical," implying existing contracts are significantly undervalued. Hyperscaler operating cash flow (Microsoft, Meta, Amazon) is already accelerating, even before new chips like Rubin are fully deployed and contracts are repriced. Several market events were, in his view, misinterpreted: 1. **Meta renting out compute:** Seen as bearish (excess capacity), but was likely a strategic move mirroring SpaceX's success in selling compute at a premium, potentially signaling future CapEx increases rather than cuts. 2. **Open-source models causing a "freak out":** The dip in a "silicon data token index" due to a mix shift towards cheaper open-source tokens was seen as negative. However, a "token is a token" in terms of compute required. Open source's rise, even if it shifts margin away from frontier models, ultimately increases overall compute demand, benefiting infrastructure providers like NVIDIA. 3. **China's DUV machine:** Led to a sell-off in semi-cap, but the technology is still decades behind the cutting edge. The speaker's primary quantitative concern is the rise in real yields and widening credit spreads, which would be problematic if the AI build-out heavily relied on debt. However, he counters this by arguing that hyperscalers are currently modeled to monetize new chips (Blackwell, Rubin) at rates far below their potential (like Ampere, two generations older). If these chips monetize even at a discount to current Blackwells, operating cash flow could reach $2 trillion, negating hundreds of billions in credit demand. The installed base of compute, as contracts reprice upwards, will also improve credit ratios. He highlights that NVIDIA is currently trading at its lowest forward PE in 10 years (excluding V-bottoms like Liberation Day and DeepSeek), indicating the market thinks they are significantly "over-earning." Yet, GPU rental prices are surging (e.g., a B200 from ~$2 to ~$4 per hour in seven months), showing hyperscalers are actually *under-earning* on current contracts. Looking ahead, he considers "continual learning" and "sample-efficient learning" as potential innovations that could temporarily reduce training demand, but training is a small percentage of overall compute. He believes the biggest risk remains regulatory backlash, stemming from misinformation about data centers' impact on power, water, and jobs. He advocates for the AI industry to better communicate its benefits, such as life-saving applications and high-paying blue-collar jobs. Finally, he points to emerging "dark horse" players and underscores that the market likely undervalues companies like SpaceX, which has demonstrated an exceptional ability to rapidly scale compute infrastructure. He remains cautiously optimistic, noting that most people in Silicon Valley are even more bullish than he is, despite the recent market turmoil.