# hyperscaler capex — X 热门讨论 (2026-09-19 01:56 UTC)

## @MF_Camillus (LBM_LXXVIII) · 09-18 18:20 · ♥34 ↻2 💬0 Perfect summary of the extraordinary ROIIC behind AI capex

Recall the definition of ROIIC (Return on Incremental Invested Capital):

ROIIC = (NOPAT_t - NOPAT_t-3y) / (IC_t-1y - IC_t-4y)

Aggregate hyperscaler ROIIC peaked close to 40% in Q1 2026, versus an estimated cost of capital of only about 8%

Consensus expects ROIIC to fall as the current capex wave hits the denominator, but only to around 23% in Q3 2027, before recovering toward roughly 35% by 2030

Look at those numbers again and consider how implausible it is that 25–75bps adjustments in policy rates are the binding constraint on the AI buildout...

That does not mean monetary policy is ineffective, of course. It means that when incremental returns remain this far above the cost of capital, relatively modest changes in the risk-free rate are unlikely to materially alter the investment decision at the margin

The obvious implication is that the weight of tighter monetary policy will be felt elsewhere in the economy > 引用 @mjmauboussin: We recently published a new report called "To Free or Not to Free (Cash Flow)." Free cash flow (FCF) is a topic of interest for the hyperscalers. The main premise is that negative free cash flow is fine as long as the returns on investment are sufficient. In the report, we cover 5 areas: 1. We offer a working definition of FCF and examine where some leading technology companies (5 hyperscalers + 3 others) stand today using results and consensus forecasts; 2. We examine whether the acceleration of investment spending changes each company’s position within the life cycle; 3. We calculate return on incremental invested capital (ROIIC) for each of the companies; 4. We look at the relationship between free cash flow and ROIC; 5. We look at how consensus estimates have changed for sales, earnings before interest and taxes (EBIT), and capital expenditures. These drivers provide a sense of what determines ROIIC.

We use only results and consensus forecasts (which are subject to a lot of change as point 5 shows).

https://t.co/u39Y2l7Y5H https://x.com/MF_Camillus/status/2101013592451531231