Third Correction, Third Seller: $2 Trillion Wiped From the Nasdaq-100 as Levered AI Credit, Not the Dip Buyer, Did the Selling

The selloff ran three consecutive weeks and each week had a different trigger. Week one was a credit event, week two was a memory and semiconductor event, and week three was an earnings event. The Nasdaq Composite fell 2.14% on October 16 on AI-debt anxiety, closed at 25,955.34 on October 23, and then lost another 2.86% on October 30 to close at 25,437.00. By the end of October the Composite sat 7.8% below its October 6 record close of 27,599.79.

CoreWeave is the cleanest case study because its capital structure makes the equity a thin, levered claim on a contracted but debt-funded asset base. Recourse debt rose to $31.4 billion from $20.7 billion in six months, non-recourse debt rose to $3.7 billion from $0.6 billion, and total debt reached $51.6 billion against $6.4 billion of cash. Over the same trailing twelve months the company burned $13.7 billion of free cash flow while issuing $16.8 billion of net debt and $5.8 billion of stock.

That is the mechanism that ends dip-buying. When a company funds a build with bonds and leases, a credit repricing does not merely lower the multiple; it changes who owns the downside. Puts the equity below the debt whenever the asset base cannot earn its cost of capital. Goldman Sachs estimates roughly a third of 2026 AI capital spending is debt-financed, rising to about 35% in 2027, with hyperscaler issuance alone running at $194 billion year to date against $108 billion for all of 2025.

A dip buyer fails when there is no cash on the sidelines, and in late September there was none. Volatility-control funds were running equity allocations at the 98th percentile since2010, roughly88% of assets for a typical fund targeting 10% volatility, with strategy assets estimated at $300 billion to $500 billion. Trend-following funds were also positioned at an 82nd percentile.

The useful question after a third correction is not whether the AI trade is over. It is which AI exposures were financed and which were paid for out of cash flow. The balance sheets answer it directly: operating cash flow covered quarterly capital spending 1.47 times at Microsoft and 0.11 times at CoreWeave. That single ratio separates the companies that can keep building from the ones that must keep borrowing.

The cruelest detail is Alphabet. It has $265.1 billion of cash and investments against $125.1 billion of borrowings excluding leases, and its trailing free cash flow is still positive at $19.7 billion. Yet it posted the biggest one-day percentage drop on record and told investors to expect 2027 capital spending to rise significantly from the $195 billion to $205 billion guided for 2026. Prices the funding plan rather than the earnings, which tells you the market is now discounting the marginal dollar of 2027 to 2029 spending, not this quarter's cash.

The upstream of the AI build is not being financed by the credit market, which is exactly why it should not behave like the downstream. TSMC runs trailing capital spending at roughly half its operating cash flow and stays in net cash, while Micron runs capital spending at about a third of operating cash flow with a net cash balance sheet and a trailing free cash flow yield near 5.9%. SK hynix, the memory supplier behind high-bandwidth memory, carries net cash with trailing capital spending at 28% of operating cash flow and a free cash flow yield around 9.1%.

Nvidia sits at the top of that chain and is the strongest counterexample to the credit scare: trailing operating cash flow of $134.4 billion against $7.4 billion of capital spending, and $57.4 billion of stock repurchases in the same period. Funds its own growth from customer cash rather than from the bond market, which is why it belongs with the funded buyers rather than with the financed ones.

The payoff from a credit-led shakeout is consolidation. Gartner forecasts that by2030 more than 80% of neocloud providers will cease independent operations, and Goldman Sachs expects project finance of $300 billion in 2027 on top of roughly $250 billion of direct hyperscaler issuance. Hands the distressed assets to the buyers with cash, which is the one part of the AI build that gets cheaper rather than more expensive in a credit shock.

The number that explains the credit repricing is not a multiple. To cover announced capital spending, financing costs and a 12% return on invested capital through 2030, AI data-center owners would need $8.5 trillion of cumulative revenue, against total AI-attributable revenue forecast at no more than $3.6 trillion. On that arithmetic, more than $3.5 trillion of announced AI infrastructure spending is exposed to impairment by the end of 2030 and more than $4 trillion of asset value is exposed to write-down through 2032.

The synthesis is simple and it is not a call on AI demand. The Nasdaq-100's third correction of 2026 was a repricing of who pays for the build and who holds the risk when the returns disappoint. Shifts the burden of proof from price to balance sheet: companies that fund capex from operating cash flow can keep building and can buy assets cheaply, while companies that fund capex with bonds are now valued as leveraged credit rather than as growth stories.