According to Woofun AI, the US stock market experienced rare shocks on October 22. The Nasdaq index closed down about 2.6%, the Philadelphia Semiconductor Index fell sharply by about 6.7%, Meta (META.US) stock price fell about 7%, and the Dow remained almost flat. On the face of it, this is a retreat from crowded AI transactions, but the pricing logic has fundamentally changed: in the past two years, the market focused on “how many chips to buy,” but now the focus is on “who will pay for AI.”
The shift was triggered by Meta's earnings report the day before, and market attention quickly shifted from the demand side to the financing side.
Meta's third-quarter earnings report showed adjusted earnings per share of $7.40 (excluding $9.89 billion in legal fees), and GAAP was only $3.83. However, free cash flow changed from positive $10.63 billion to negative $2.12 billion in the same period last year, showing a net quarterly outflow for the first time in more than a decade. The 2026 capital expenditure guidance was adjusted to $135 billion to $145 billion, up the lower end from the previous $130 billion. The Board approved a 1:1 distribution of non-voting Class C shares in January 2027. Capital expenditure for the quarter was approximately $35 billion, and the increase in advertising revenue did not cover server and data center expenses. The Class C stock design preserves Zuckerberg's control while reserving room for future fundraising. The market sell-off was not about stock splits, but rather about concerns about the potential supply of equity. The valuation is based on the assumption of sustainable cash flow growth. Changes in financing methods directly affect discount rates, not just quarterly profits.
According to data compiled by Woofun AI, there are significant differences between the upstream and downstream industries in attributing the decline. Nvidia (NVDA.US) CEO Hwang In-hoon said in a CNBC interview that Neocloud demand has “exploded”. Currently, the bottleneck is working capital. Capital demand far exceeds supply, but “this problem will be solved.” The statement pointed to liquidity restrictions rather than the disappearance of orders, but was interpreted by the market as circumstantial evidence of financing bottlenecks. If upstream admits that customers lack money, the pace of chip orders will depend on the progress of financing rather than the intensity of demand.
Meanwhile, Bernstein (AB.US) analyst Madison Rezaei lowered the CoreWeave (CRWV.US) target price from $74 to $51 to maintain its underperforming market rating, and estimated that it would require around $85 billion in financing to fulfill the promised order. The market reaction was more immediate: CoreWeave bond yields rose to around 15.8%, and the 5-year credit default swap spread was over 1100 basis points. When the marginal cost of borrowing exceeds 15%, the rate of expansion of the rental card business is determined by creditor quotas, not market demand.
Wong In-hoon believes that demand is worry-free and that the funding problem is solvable, while Bernstein sees rising capital costs as a sign of a decline in effective demand. The October 22 market marked the market's first large-scale trend towards the latter. Chip stocks fell more than cloud vendors, with Nvidia down about 4% and CoreWeave about 10%. Lam Research (LRCX.US) intensified mood swings. The median revenue guidance for the December quarter was slightly higher than analysts' agreed expectations, but fell short of buyers' private expectations. At the same time, it also suggested that Chinese customers' revenue share declined and supply chain bottlenecks could delay shipments.
Despite Lam's strong demand for AI, global fab equipment spending is expected to exceed US$200 billion in 2027, and record gross margin and operating profit margins for the quarter, and the stock price still fell by about 10%. The Lam guide can only explain the retracement of equipment stocks themselves; it cannot explain the sector's overall decline of 6.7%, but as a concrete example, it clearly shows how financing concerns are transmitted from cloud vendor balance sheets to the visibility of equipment orders. The market's sensitive reaction to 'margins falling short of expectations' confirms the high dependence of valuations on capital expenditure slopes.
From a macro perspective, Meta became the first visible sample of changes in capital structure. Morgan Stanley (MS.US) estimates that capital expenses related to hyperscale cloud vendors plus neo-cloud may reach 1.58 trillion US dollars next year, and the free cash flow gap including leases may reach 540 billion US dollars. Goldman Sachs (GS.US) estimates that the capital expenditure of leading cloud vendors will be around $800 billion in 2026. The data difference between the two investment banks is nearly double, but they all indicate that the scale of AI construction has exceeded the ability of a single company to cover endogenous cash. The investment structure is shifting from endogenous cash flow to external capital, and the market is beginning to price dilutions and interest spreads.
Current evidence does not support 'AI demand peak' or 'financing window closed'. Although CoreWeave bonds are expensive, there are still buyers, and Meta Class C has only reserved channels for additional issuance. The key observation is when the 'expected dilution' is converted into an 'achieved dilution'. If leading companies actually issue large amounts of equity or new types of debt, the valuation pressure will change from discount assumptions to actual share capital supply. If credit spreads in the technology sector continue to widen, companies outside of Neo-Cloud are also required to pay higher risk premiums, and financing restrictions will no longer be a case. Conversely, if capital expenditure continues to rise next quarter and order revenue follows, this round of adjustments is more likely to be a phased repricing. What was left on October 22 was not an “end of AI” conclusion, but rather a new equation: demand continues, but behind every GPU, the question of the source of funding must first be answered.