The Fed Better Have AI's Back Or There Will Be Problems | Dan Krausz on “Liquidity Trap Door” In AI Infrastructure and Why Policymakers Are Inclined to Support Financial Conditions

Dan Krausz sees the AI hardware selloff chiefly as a liquidity problem, not evidence that demand for AI is failing. He says demand has not yet been the constraint: companies are finding uses for the technology, while financing determines how quickly they can build and adopt it. He identifies three possible routes to an AI downturn—weak demand, policy obstructing data centers, or insufficient financing liquidity—and considers the third the more immediate risk. Hyperscalers once funded capital spending from accumulated cash; greater reliance on markets now makes the buildout more exposed to available capital. AI capex may be near its peak rate of increase, he says, but not its peak level.

Liquidity, in Krausz’s definition, is the global availability of capital, influenced by fiscal deficits and their financing, central-bank rates and balance sheets, bank lending, private credit, and the dollar. As Treasury and AI-related issuance has grown while liquidity has weakened, borrowers compete for capital. In his funding “waterfall,” Treasuries and investment-grade hyperscalers get financed before weaker credits; Google and Meta should still be able to borrow, but their borrowing leaves less for others. Rising rates are also reducing excess liquidity—the capital left after funding issuance. He cites emerging-market currencies peaking in February as an early sign of pressure, and notes that nonresidential investment as a share of GDP is near its highest level.

Krausz sees policymakers balancing the risk of an AI bubble against the cost of restraining investment and damaging the broader economy. They view AI as crucial to competitiveness and potentially disinflationary, so he expects them to avoid tightening financial conditions enough to derail infrastructure spending. Treasury has discussed more short-term issuance and buybacks, which Krausz sees as possible liquidity support; he describes policymakers as putting “trampolines” under a liquidity trap door to avert a damaging market shock. He would be surprised if government liquidity policies ultimately tripped up AI investment, but says persistently rising inflation could force officials to withdraw support. He expects the Fed to keep tightening at a measured pace while inflation remains too high, and warns that tightening too quickly could become self-defeating.

The timing is central to his outlook. Liquidity weakness may take six to nine months to materially affect the economy, and if liquidity returns sooner, markets could recover before significant economic damage occurs. In that case, he says, the AI selloff could be about 70% retraced. If liquidity remains tight, AI spending could slow materially by the back half of next year; Krausz says that point has not yet been reached. He expects material economic slowing through next year, including weaker non-AI cyclical growth and consumption, but not necessarily a recession: strong capex and a fiscal deficit of about 6% should cushion the slowdown. His base case is that the economy and earnings remain okay if liquidity does not tighten further; another squeeze could produce a “trapdoor” correction of 10% or more.

Market positioning may have amplified the selloff and subsequent rotation. Goldman’s AI-at-risk index rose more than 50% from its lows during the decline. Krausz says funds had been long AI-related stocks and short perceived AI losers, so unwinding those trades could force buying in the shorted names. He sees the rally in non-AI stocks partly as short covering, not proof of stronger fundamentals, and expects that mechanical lift could fade next year as disruption concerns return. He distinguishes liquidity-sensitive AI infrastructure from disruption risk, which depends more on adoption and businesses using AI; their returns need not remain negatively correlated. Nvidia and Broadcom can support infrastructure with their balance sheets, but that capacity is finite and may not reassure markets if liquidity tightens.

The long-term economic payoff remains uncertain. Krausz says AI could raise revenue or reduce costs, but revenue effects will vary by industry and cost savings are too early to judge while firms run AI alongside existing employees. He cautions that company anecdotes are not yet evidence of economy-wide gains, and measured productivity may not show improvements for one or two years. As adoption spreads, initial margin benefits could fade if AI becomes table stakes and competition passes productivity gains to consumers; lower-cost Chinese models could also pressure providers such as Anthropic. He sees no conclusive evidence yet about AI’s employment effects. Better labor productivity is needed to support growth, debt and deficits, and the case for AI capex, but the gains could come at employees’ expense.