Citadel Securities: AI Trade Unraveling, Capital Expenditure Becomes a Two-Way Risk
nashnova research
Citadel Securities' fixed-income sales head Noshad Shah warns the recent AI sell-off signals investors no longer assume massive spending converts to profit; he frames capex as a two-way risk — cutting raises doubts, adding more raises different doubts.
Why is capex a "two-way risk"?
Cut spending → the market questions whether the AI investment cycle can last. Increase spending → the market questions financing capacity and ultimate returns. This means → big tech is trapped either way.
Meta's Q3 results illustrate the bind: revenue grew 26% year-on-year, but capex including finance-lease principal hit $34.9 billion — exceeding operating cash flow of $32.8 billion.
In plain terms = Meta's cash earnings could not cover what it spent on AI. Quarterly free cash flow came in at negative $2.1 billion, and full-year guidance implies another $49–59 billion in Q4 alone.
AI usage is rising — why isn't spending keeping up?
Shah cites data from enterprise spending platform Ramp: AI usage kept growing, but token spending was nearly flat from August to September. This means → falling prices and efficiency gains absorbed the volume increase.
Spending hit a new high in early October, but growth slowed. Shah cautions that Ramp data covers only its client base, not the broader economy — a slowdown in dollar spending does not equal a matching drop in usage.
DeepSeek's release of DeepInfer adds further uncertainty: software-level improvements boost effective output on existing hardware. In plain terms = even if cheaper inference drives more usage, how much new compute the future actually requires is anyone's guess.
Cheaper AI is good for users — but what about suppliers?
Shah's core point: cheaper AI improves customer economics and widens the range of viable applications. But for suppliers, profitability depends on the race between how fast costs fall vs. how fast prices fall, and whether higher volumes compensate.
This reflects a deeper tension — a technology can create enormous economic value while delivering disappointing returns to some of its producers.
In plain terms = users saving money does not mean compute sellers are making money. Railroads and fiber optics followed a similar script.
How strong is the political headwind?
A Reuters/Ipsos poll this month: 69% of registered Democratic voters support limiting data-center construction; among Republicans the figure is 43%. This means → opposition crosses party lines — it is not a single-party issue.
November midterm elections could bring more congressional scrutiny. State and local races directly affect permitting and development conditions.
Even without a blanket federal restriction, added compliance requirements, local building limits, and lengthy approvals make deployment slower and costlier. For leveraged projects, delays mean extra interest expense before revenue arrives — and by the time capacity comes online, the competitive landscape may have shifted.
Why are cybersecurity incidents adding to the risk?
This week's security news: AI-assisted researchers found a decade-old critical vulnerability in widely used PHP software. Amazon identified exploitation attempts and warned that AI is shortening the window from vulnerability disclosure to working exploit.
Shah's concern: legitimate security worries could become a pretext for sweeping restrictions. This means → the economic cost of the political response may far exceed the vulnerability itself.
This reflects yet another layer of risk to AI return timelines — not a failure of the technology, but a policy environment around it that could slow the entire commercialization pace.
What is the central question?
Shah's ultimate test: can capex sustain growth in a high-cost, high-political-resistance environment and convert into verifiable returns within the promised timeline?
In plain terms = the AI trade's logic chain is "spend big → build capacity → generate revenue." Every arrow now carries a question mark — is there enough money, can it be built, when does the revenue arrive?
Investors are no longer asking "does AI have a future?" They are asking "who pays, and when does the cash show up?" That is the core logic shift behind the current sell-off.
市场有风险，内容仅供研究参考，不构成投资建议。