The AI capex debate has moved from whose math is right to a race between two clocks — resolves 2027-07-24

The debate: whether the huge sums the big cloud providers spend each year on AI computing power are justified, with bulls arguing from four separate calculations that they are still underbuying and bears arguing that too few companies actually pay for the results.

Why it matters: if the money does not come back, the first damage lands on the cloud providers’ free cash flow and on the bond market, and then on orders across the supply chain.

Who is on each side: the underbuild camp, the broken-conversion camp (companies use AI but do not earn from it) and the funding-structure camp (once cash flow turns negative, the financing window decides the outcome first), and SecondSource gives a layered ruling rather than naming one winner.

SecondSource’s reading: the supply side has delivered, with Nvidia’s data-center business at an annualized revenue near $300 billion, three to four years ahead of consensus, but selling shovels is not the same as end customers paying for productivity.

Conversion is a curve tiered by price: plans above $250 a month have settled at 70% gross retention (customers who renew, not counting upsells) and 85% net revenue retention (revenue kept, including upsells).

Plans under $50 a month keep only 23%, and in ordinary companies the gains that reach the income statement sit between 5% and 20%.

What broke first was the bulls’ own financial discipline: capex now takes 90% to 100% of operating cash flow, the research group Epoch projects combined free cash flow touching zero in the third quarter of 2026, and debt is taking over.

So the debate is no longer open-ended waiting for conversion to prove itself, but the speed at which premium-tier conversion spreads against the speed at which the financing window closes.

BG2 Pod (Brad Gerstner and Bill Gurley)
“I think it increases the chance that we go over…”

By July 2027: any two hyperscalers cutting capex guidance in the same year voids underbuild; high-tier NRR below 100% or GRR below 60% (baselines 85%/70%) weakens tiering; a separately reported AI revenue line outside coding strengthens conversion; debt markets absorbing negative free cash flow for 12 months without spreads widening weakens the funding clock.

What would overturn it

Reversal runs through the same tests: two hyperscalers cutting capex guidance, or high-tier retention breaking its baselines, would overturn the tiered reading; debt markets funding negative free cash flow without spreads widening would break the funding-clock leg.

Correction to the settlement rule

Two wording fixes on 2026-08-03; the claim and thresholds were not changed. First, the "what would overturn it" line used to say only "see the reversal condition on another tree card", which gave readers nothing; it now states a concrete rule (if the high-tier retention check and the AI-revenue-curve check both trigger while the financing window has not closed, the tiered reading no longer holds; a single trigger only weakens its own leg). Second, an internal extraction-process tag in the call was removed and the body of the sentence kept word for word. No result had been seen. The original wording stays in our ledger and in the public mirror's version history.