Opinion · August 2026
The AI-Infrastructure Trade Just Met Its Cost of Capital
Every AI-infrastructure thesis I’ve written this year rests on one implicit assumption: that debt-financed capacity build-out stays affordable long enough for revenue to catch up to the capital spent building it. Tuesday’s selloff is what it looks like when the market starts pricing that assumption as a live risk instead of a footnote.
The mechanism, not just the headline
The Motley Fool’s read on CoreWeave’s 12.10% decline was specific: $9.4 billion in second-quarter capital spending against nearly $30 billion in long-term debt, at a debt-to-equity ratio north of 14, getting repriced as the cost of that debt rises. That’s a financing-cost story, not a demand story — nothing in Tuesday’s reporting suggested AI-infrastructure demand weakened. Nebius’s own numbers, reported the same week, still showed 454% quarterly revenue growth.
Those two facts sit uncomfortably together, and the discomfort is the whole point. Demand can be spectacular and the equity can still fall, because the equity is a residual claim sitting behind $30 billion of debt that has to be serviced first.
A projection: what the debt actually costs
Below is the arithmetic I keep coming back to. I do not know CoreWeave’s blended cost of debt — that is not something I can source reliably — so rather than guess a single number, this runs the sensitivity across a plausible range and lets you see where it starts to bind.
Figure 1 · Illustrative sensitivity
Annual interest burden on $30B of debt, at four coupon levels
Interest = $30B × coupon. Percentages are that figure against $6.23B trailing-twelve-month revenue. Illustrative sensitivity, not a forecast — see the assumptions note below.
| Rate scenario | Annual interest | % of TTM revenue | Revenue needed to hold interest at 25% |
|---|---|---|---|
| 6% average coupon | $1.80B | 28.9% | $7.20B |
| 8% average coupon | $2.40B | 38.5% | $9.60B |
| 10% average coupon | $3.00B | 48.2% | $12.00B |
| 12% average coupon | $3.60B | 57.8% | $14.40B |
Assumptions, stated plainly. This applies a single average coupon to the entire $30B long-term debt balance reported by The Motley Fool, against $6.23B of trailing-twelve-month revenue. It ignores debt maturity schedule, fixed-versus-floating mix, capitalised interest, and any cash interest income — none of which I can source at the required precision. It is a sensitivity showing the shape of the exposure, not a projection of what CoreWeave will actually pay. The useful takeaway is the last column: at a 10% average cost of debt, revenue would need to roughly double from here just to keep interest at a quarter of the top line.
Why gold falling is the detail that matters
Gold and silver falling together with growth stocks — GLD down 1.71%, SLV down 3.58% — is what makes me read this as broader than one sector’s story. A classic flight-to-safety session sends money into gold as equities fall. Tuesday sent money out of both. That pattern looks less like fear of AI infrastructure specifically and more like every asset carrying duration risk getting marked down together.
| Asset type | Why it carries duration | Aug 18 |
|---|---|---|
| Debt-financed AI infrastructure | Long-dated debt funding assets whose revenue arrives years later | −12.10% |
| Long-duration growth equity | Value concentrated in distant cash flows | −4.09% |
| Precious metals | No yield at all — pure opportunity cost against rates | −3.58% |
| Broad index | Blended; mostly shorter-duration cash flows | −0.68% |
CoreWeave, SMH, SLV and SPY respectively. The ordering is the argument: the more of an asset’s value sits in the distant future, the harder it fell.
Correcting the rate claim
I originally framed this as yields “moving higher” on Tuesday. The constant-maturity series says otherwise: the 10-year fell a basis point on the day, after rising nine basis points across the three preceding sessions. The repricing followed the rate move rather than coinciding with it, which is a lag worth noting rather than smoothing over — it suggests positioning adjusting to a rate level, not a same-day reaction to a rate print.
Where this connects to my own book
This is the live version of the argument in the Physical Limits of Compute whitepaper: the real risk in this cycle was never a single bad print, it’s the moment the market starts treating financing cost as a genuine constraint on the buildout rather than a rounding error. It is also the same funding gap quantified in The AI Capex Reality Check — Alphabet generated $0.28 of free cash flow per $1 of Q1 capex, Meta $0.65. Neither self-funds its build-out, and neither carries CoreWeave’s leverage.
Tuesday wasn’t that risk resolved. One session of rate-driven weakness against continued real demand growth isn’t a thesis break. But it’s the clearest single-day illustration yet of the exact mechanism that whitepaper flagged, and worth logging as the day the market started pricing it explicitly instead of implicitly.
What would change my mind
Two things would tell me this is a repricing rather than a regime change: the selloff stopping without follow-through once a fresh headline arrives, and capex guidance from these names holding steady rather than being trimmed. The opposite — continued declines with no new catalyst required — would be evidence the market is reassessing the financing structure itself. That test ran the very next session, and I wrote up the answer in A Second Consecutive Day Tells You More.