美国财报季拉开帷幕，人工智能（AI）的看涨派与看跌派正日益固守己见。市场出现了若干怪象、异象与反常现象，凸显了这种极端分化。
US earnings season gets underway, AI bulls and bears are becoming more entrenched in their positions. Several market quirks, curiosities and anomalies have popped up that underscore these extremes.
这不足为奇，因为在AI问题上，投资者与分析师很大程度上是在“盲飞”。该技术尚处于起步阶段，演进速度极快，潜在能力几乎难以想象。而为这场革命提供资金的成本？也几乎难以想象。
This shouldn't come as a surprise, because, when it comes to AI, investors and analysts are largely flying blind. The technology is in its infancy, is evolving at breakneck speed, and has potential powers that are almost unfathomable. And the cost of funding this revolution? It's almost unfathomable too.
对更广泛市场的影响同样巨大。据摩根大通私人银行的Michael Cembalest称，AI相关股票——他将其范围缩小至41只——自2022年11月ChatGPT推出以来，贡献了标普500指数（.SPX）75%的回报、80%的盈利增长和90%的资本支出增长。
The implications for the broader market are also enormous. According to JPMorgan Private Bank's Michael Cembalest, AI-related stocks, a group he narrows down to 41 names, have accounted for 75% of S&P 500 (.SPX) returns, 80% of earnings growth and 90% of capital spending growth since ChatGPT launched in November 2022.
超大规模云服务商、“新云”厂商、芯片制造商、软件供应商以及所有AI相关企业的业绩与展望，将为华尔街和全球市场迈入明年定下基调。以下是背景情况：
The results and outlooks from hyperscalers, “neoclouds”, chipmakers, software providers and all the AI-adjacent firms will set the tone for Wall Street and global markets going into next year. Here's the backdrop:
股债缺口 俗话说，债券投资者因看空而获利，股票投资者因看多而获利。这或许有助于解释，为何用于为美国创纪录的AI投资融资的许多债券交易价格如此疲软，而AI相关股票却大多持续上涨。在最新一波市场紧张情绪中，容易让人忘记标普500指数和纳斯达克指数（.IXIC）仅两周前还处于历史高位。
Equity-bond gap As the old saying goes, bond investors are paid to be bearish and equity investors are paid to be bullish. Perhaps that helps explain why many of the bonds issued to finance America's record-breaking AI investment are trading at such weak levels even as AI-related shares mostly continue to move higher. Amid the latest wave of market jitters, it's easy to forget that the S&P 500 and Nasdaq (.IXIC) were at record highs only two weeks ago.
这些紧张情绪当然意义重大。费城半导体指数（.SOX）较6月高点下跌25%，今年早些时候标普500软件指数（.SPLRCIS）曾较历史高点下跌35%。但随着第三季财报季开始，芯片指数今年以来上涨55%，软件股已攀升回年内正区间。
Those jitters have been meaningful, of course. The Philadelphia Semiconductor Index (.SOX) is down 25% from its June peak, and earlier this year the S&P 500 software index (.SPLRCIS) was down 35% from its record high. But the third-quarter earnings season begins, the chip index is up 55% year to date and software stocks have climbed back into positive territory for the year.
“新云”公司 CoreWeave (CRWV.O) 的股价今年下跌 13%，而其 2031 年 10 月到期的债券自 4 月发行以来暴跌逾 20%，交易价格跌至面值的 80 美分以下，进入困境债区间。
While “neocloud” company CoreWeave (CRWV.O) shares are down 13% this year, its October 2031 bond has plunged more than 20% since its launch in April and is trading in distressed territory below 80 cents on the dollar.
与此同时，衡量债券违约保险成本的信用违约互换 (CDS) 点差最近飙升至历史高位，涉及的 AI 借款方包括 CoreWeave 和 Oracle (ORCL.N)，甚至连万亿美元级巨头 Amazon (AMZN.O)、Alphabet (GOOGL.O) 和 Meta (META.O) 也不例外。或许有人该提醒投资者，债务在偿付顺序上仍优先于股权。CoreWeave：股价惊魂？
Meanwhile, credit default swap spreads, which measure the cost of insuring against bond default, have recently soared to record highs for AI borrowers including CoreWeave and Oracle (ORCL.N), and even for trillion-dollar giants Amazon (AMZN.O), Alphabet (GOOGL.O) and Meta (META.O). Someone might want to remind investors that debt is still senior to equity. CoreWeave: stock and awe?
CoreWeave 处于最新市场动荡的中心。这家获得 Nvidia (NVDA.O) 支持的 AI 基础设施公司亏损经营、债务高企，且面临飙升的借贷成本。但华尔街分析师对其前景的看法截然不同。
CoreWeave is at the center of the latest market maelstrom. The Nvidia-backed (NVDA.O) AI infrastructure firm is loss-making, heavily indebted and faces soaring borrowing costs. But Wall Street analysts have wildly different views about its prospects.
汇丰 (HSBC) 分析师本周发布了一份极度悲观的报告，预测到 2030 年的平均回报率将低于 CoreWeave 的借贷成本。
Analysts at HSBC this week published an extremely gloomy note, forecasting average returns through 2030 below CoreWeave’s cost of borrowing.
“公司在这种回报水平下投入越多，价值毁灭越严重，”他们写道。汇丰建议投资者减持该股，给出的目标价仅为 39 美元，比周二收盘价 62 美元左右低约 37%。
"The more the company invests with this return profile, the more the value destruction," they wrote. Investors should cut their holdings of the stock, HSBC advised, assigning it a target price of only $39, some 37% below Tuesday’s close around $62.
看来结论显而易见，对吧？也许并非如此。看着同样的数据，高盛 (Goldman Sachs) 股票策略师本周将该股评级从“中性”上调至“买入”。他们预计未来 24 个月每兆瓦年度经常性收入可能上升，并认为近期收益率飙升并不代表公司的真实借贷成本，因为其大部分融资是资产支持的。尽管他们将目标价从 139 美元下调至 106 美元，但这仍意味着约 70% 的上涨空间。
Seems pretty clear-cut, right? Maybe not. Looking at the same data, equity strategists at Goldman Sachs this week raised their recommendation on the stock to a “buy” from “neutral.”They reckon annual recurring revenue per megawatt is likely to rise over the next 24 months, and argue that the recent surge in yields doesn't represent the firm's true cost of borrowing because most of its financing is asset-backed. Even though they cut their price target to $106 from $139, that still represents upside of around 70%.
象征性姿态 代币（token）支出的增长正在放缓，但代币的使用量仍在持续飙升。代币是AI模型在训练和推理过程中处理的数据单位，本质上反映了市场对“算力”需求的强弱。阿波罗的托尔斯滕·斯洛克表示，价格增长放缓或 outright 下跌将表明，除非企业能够将其转化为“更高的营收增长、更大的利润率和大幅的成本节约”，否则它们不愿为这些算力付费。而这些好处尚未兑现。
Token gesture Growth in token spending is slowing, yet token usage continues to soar. Tokens are units of data processed by AI models during training and inference, and they essentially reflect the strength of demand for "compute". Slowing growth or outright declines in prices would indicate companies are reluctant to pay for that compute unless they can turn it into "higher top-line growth, bigger profit margins or substantial cost savings," according to Apollo's Torsten Slok. These benefits have yet to materialize.
但代币使用量仍在强劲而稳定地上升。所以需求依然存在。只是支出似乎已经见顶，因为代币价格已经下跌。这提醒投资者，OpenAI和Anthropic等大型AI模型开发商的收入可能不会像人们所希望的那样继续惊人地增长。AI股票对有关未来AI使用以及——显而易见——收入的预期高度敏感。投资者能无视两者之间的差距多久？
But token usage is still rising strongly and steadily. So demand is still there. It's just that spending appears to have topped out because token prices have fallen, which is a reminder to investors that revenue at the big AI model developers like OpenAI and Anthropic may not continue to rise as spectacularly as hoped. AI stocks are highly sensitive to expectations about future AI use and, obviously, revenue. How long can investors ignore the gap between the two?
冷存储 大量资金正被用于解决算力短缺问题，但显然其中一些算力根本无法启用。字面意义上的无法启用。能源供应链的瓶颈，包括电力不足和电网接入不完善，意味着许多AI芯片正闲置在仓库中并不断贬值。
Cold storage Vast sums are being spent to address the shortage of computing power, but apparently some of that power simply cannot be switched on. Quite literally. Energy supply-chain chokepoints, including insufficient electricity and inadequate grid connections, mean many AI chips are sitting and depreciating in warehouses.
就在本周，微软（MSFT.O）退出了与英伟达支持的Nscale合作、由后者锚定英国一个旗舰数据中心的交易，此前电网官员表示，他们预计在2030年代之前无法为该站点供电。摩根士丹利分析师现在估计，到2028年美国数据中心开发商将面临30%至40%的电力缺口，相当于芯片总需求的34%。电力需求正在上升，但新增电力无法足够快地并网。
Only this week, Microsoft (MSFT.O) pulled out of a deal to anchor a flagship data center in Britain with Nvidia-backed Nscale after electricity grid officials said they did not expect to supply power to the site until the 2030s. Morgan Stanley analysts now reckon US data center developers will face a 30% to 40% power shortfall through 2028, equivalent to 34% of total chip demand. Demand for power is rising, but it can't come online quickly enough.
收入难题 归根结底，AI投资押注能否成功，取决于借来并投入这项技术的巨额资金能否带来回报。第三季度财报不会给出确定答案，但其中的每一个暗示和指引都会被密切关注。这些数字令人咋舌。
Revenue conundrum Ultimately, the AI investment bet hinges on whether the extraordinary sums being borrowed and spent on the technology will pay off. Third-quarter earnings won't give a definitive answer, but every nod and steer will be seized upon. The sums are eye-watering.
麦肯锡和标普全球估计，仅美国六大超大规模云服务商在2025年至2030年间，就将在数据中心及其他AI相关资本支出上花费超过7万亿美元。
McKinsey and S&P Global estimate that the top six US hyperscalers alone will spend more than $7 trillion on data centers and other AI-related capital expenditure between 2025 and 2030.
哥伦比亚大学的Stijn Van Nieuwerburgh估计，从去年到2032年，为AI建设融资的成本将达到10.3万亿美元。这类支出高度集中——根据一项研究，仅2%的公司就占了总支出的50%以上。
Columbia University's Stijn Van Nieuwerburgh reckons the cost of financing the AI buildout from last year to 2032 will be $10.3 trillion. This spending is highly concentrated — just 2% of firms account for over 50% of total spending, according to one study.
Van Nieuwerburgh估计，如果这些公司希望从这笔支出中获得保守的10%回报，那么到2032年，年收入需要达到约3.55万亿美元。一些分析师认为，要让这笔投资获得回报，年收入需要超过4万亿美元。这超过了2032年预期GDP的10%。这是一个非常非常高的门槛。
If these firms want a conservative 10% return on that outlay, Van Nieuwerburgh estimates that annual revenue will need to be around $3.55 trillion by 2032. Some analysts suggest annual revenue of over $4 trillion will be needed to make the investment pay off. That's over 10% of expected GDP in 2032. That’s a very, very high bar.