What is the real risk to the AI trade?

**Marked to market, 12 October — the split has reached the shares, and the forecasters still have not moved.** In the week to 9 October the companies that buy computing power with their own cash rose (Microsoft +3.4%, Amazon +4.3%, Alphabet +2.4%) while the chain that builds it or borrows to supply it fell: the chip-stock fund -4.3%, TSMC -4.1%, ASML -4.7%, and CoreWeave, the name most reliant on borrowing, -8.4%. A demand scare would have hit the buyers first, because they place the orders; they rose. Across nineteen companies in the AI hardware chain the engine's consensus panel recorded three rating changes between 1 and 9 October, all upgrades (AMD, CoreWeave and Marvell, Marvell's on 8 October, the worst day of the sell-off), and no downgrades. Corporate bond funds rose on the week, so the cost is showing up in the borrowers' shares before it shows up in lending terms. Our risk radar carried this concern as rising on 7 and 11 October; on 12 October it moved off the radar into the brief's lead, because it is now the thesis rather than a scenario.

**Marked to market, 7 October — the funding has moved into debt, and the price of it is visible.** In the first week of October the largest AI build-outs went to the credit market at scale. SpaceX is seeking $40bn of debt financing led by Apollo to buy Nvidia chips; JPMorgan was pitching an 11% yield on a $5bn loan for an AI infrastructure company, with GPU financing reaching asset-backed bonds. Our risk radar moved this concern from stable to rising on 7 October for that reason.

The falsifier has not fired. It asked for new data-centre issuance to clear at stable spreads; an 11% coupon on an infrastructure loan is the opposite of a stable spread, so the 18 August reading has been confirmed rather than reversed. One thing has been added to the position since August: power has become a third cost of compute alongside chips and funding. On 6 October Google signed a 20-year supply deal for extra output from existing nuclear plants, and the companies that sell electricity at market prices rose about five times as much as regulated utilities. Debt has to be serviced from cash flow, so a build-out financed at double-digit yields and paying a scarcity price for power is more exposed to a slower revenue ramp than one financed with equity.

The framing matters because it changes what to watch. If the risk were demand, the tell would be orders and revenue. If the risk is funding, the tell is credit terms — and the two can diverge for a long time.

The stated falsifier for this position was new data-centre and semiconductor issuance clearing at stable spreads. On 18 August the opposite was reported: a Blackstone-backed data-centre landlord priced $3.9bn of high-grade paper at junk-like yields to fund capacity. That is the concession widening, not holding, and it is the single most direct evidence this position has yet had.

The rest of the week points the same way. A single technology borrower drew over $30bn of orders for a jumbo bank loan, an AI developer opened funding talks at a $40bn valuation, and memory costs are already visible in reported margins at scale. Equities fell for a third consecutive session with the chip complex named as the driver, while the thirty-year sat at 5.29% after touching its highest since 2007.

The equity has acquired a bond-market sensitivity that most models of it do not contain, and the research layer still has not followed. Five weeks on from that print the semiconductor bellwether carries the same sixty buy ratings against three sells across seventy-nine analysts, unchanged to the rating — and across the twenty-one largest names in the theme the engine's consensus snapshots record no net downgrade at all over that window, against 23.4% of comparably-covered companies elsewhere seeing at least one rating change. The scale is saturated rather than the analysts inattentive. The price is leading the research, which is the ordinary sequence when the change is in the cost of funding rather than in the order book.

The dispersion this has produced is now the more visible symptom: over the five weeks to 27 August those twenty-one names spanned sixty-three percentage points, with the software monetisers at the top and the capital-equipment suppliers at the bottom. That ordering is what a funding constraint looks like from the equity side — the names carrying the build-out cost are the ones being marked down, and no rating change will announce it.